[ Team LiB ]

• Table of Contents
Process Control: Modeling, Design, and Simulation

By B. Wayne Bequette

Publisher: Prentice Hall PTR

Pub Date: December 26, 2002
ISBN: 0-13-353640-8
Pages: 769

Master process control hands on, through practical examples and MATLAB(R)
simulations

This is the first complete introduction to process control that fully integrates software
tools—enabling professionals and students to master critical techniques hands on, through
computer simulations based on the popular MATLAB environment. Process Control: Modeling,
Design, and Simulation teaches the field's most important techniques, behaviors, and control
problems through practical examples, supplemented by extensive exercises—with detailed
derivations, relevant software files, and additional techniques available on a companion Web site.
Coverage includes:

Fundamentals of process control and instrumentation, including objectives, variables, and

block diagrams

Methodologies for developing dynamic models of chemical processes

Dynamic behavior of linear systems: state space models, transfer function-based models,
and more

Feedback control; proportional, integral, and derivative (PID) controllers; and closed-loop
stability analysis

Frequency response analysis techniques for evaluating the robustness of control systems

Improving control loop performance: internal model control (IMC), automatic tuning, gain
scheduling, and enhancements to improve disturbance rejection

Split-range, selective, and override strategies for switching among inputs or outputs
 Control loop interactions and multivariable controllers

An introduction to model predictive control (MPC)

Bequette walks step by step through the development of control instrumentation diagrams for an
entire chemical process, reviewing common control strategies for individual unit operations, then
discussing strategies for integrated systems. The book also includes 16 learning modules
demonstrating how to use MATLAB and SIMULINK to solve several key control problems, ranging
from robustness analyses to biochemical reactors, biomedical problems to multivariable control.

[ Team LiB ]
[ Team LiB ]

• Table of Contents
Process Control: Modeling, Design, and Simulation

By B. Wayne Bequette

Publisher: Prentice Hall PTR

Pub Date: December 26, 2002
ISBN: 0-13-353640-8
Pages: 769

Copyright
PRENTICE HALL INTERNATIONAL SERIES IN THE PHYSICAL AND CHEMICAL ENGINEERING SCIENCES
About Prentice Hall Professional Technical Reference
Preface
Background
Textbook Goals
Chapters
Learning Modules
Textbook Web Page
Acknowledgments

Chapter 1. Introduction
Section 1.1. Introduction
Section 1.2. Instrumentation
Section 1.3. Process Models and Dynamic Behavior
Section 1.4. Control Textbooks and Journals
Section 1.5. A Look Ahead
Section 1.6. Summary
Student Exercises

Chapter 2. Fundamental Models

Section 2.1. Background
Section 2.2. Balance Equations
Section 2.3. Material Balances
Section 2.4. Constitutive Relationships
Section 2.5. Material and Energy Balances
 Section 2.6. Form of Dynamic Models
Section 2.7. Linear Models and Deviation Variables
Section 2.8. Summary
Suggested Reading
Student Exercises
Appendix 2.1: Solving Algebraic Equations
Appendix 2.2: Integrating Ordinary Differential Equations

Chapter 3. Dynamic Behavior

Section 3.1. Background
Section 3.2. Linear State Space Models
Section 3.3. Introduction to Laplace Transforms
Section 3.4. Transfer Functions
Section 3.5. First-Order Behavior
Section 3.6. Integrating System
Section 3.7. Second-Order Behavior
Section 3.8. Lead-Lag Behavior
Section 3.9. Poles and Zeros
Section 3.10. Processes with Dead Time
Section 3.11. Padé Approximation for Dead Time
Section 3.12. Converting State Space Models to Transfer Functions
Section 3.13. MATLAB and SIMULINK
Section 3.14. Summary
References
Student Exercises

Chapter 4. Empirical Models

Section 4.1. Introduction
Section 4.2. First-Order + Dead Time
Section 4.3. Integrator + Dead Time
Section 4.4. Discrete-Time Autoregressive Models
Section 4.5. Parameter Estimation
Section 4.6. Discrete Step and Impulse Response Models
Section 4.7. Summary
References
Student Exercises
Appendix 4.1: Files Used to Generate Example 4.4
Appendix 4.2

Chapter 5. Introduction to Feedback Control

Section 5.1. Motivation
Section 5.2. Development of Control Block Diagrams
Section 5.3. Response to Setpoint Changes
Section 5.4. PID Controller Algorithms
Section 5.5. Routh Stability Criterion
Section 5.6. Effect of Tuning Parameters
Section 5.7. Response to Disturbances
Section 5.8. Open-Loop Unstable Systems
Section 5.9. SIMULINK Block Diagrams
Section 5.10. Summary
 References
Student Exercises

Chapter 6. PID Controller Tuning

Section 6.1. Introduction
Section 6.2. Closed-Loop Oscillation-Based Tuning
Section 6.3. Tuning Rules for First-Order + Dead Time Processes
Section 6.4. Direct Synthesis
Section 6.5. Summary
References
Student Exercises

Chapter 7. Frequency-Response Analysis

Section 7.1. Motivation
Section 7.2. Bode and Nyquist Plots
Section 7.3. Effect of Process Parameters on Bode and Nyquist Plots
Section 7.4. Closed-Loop Stability Concepts
Section 7.5. Bode and Nyquist Stability
Section 7.6. Robustness
Section 7.7. MATLAB Control Toolbox: Bode and Nyquist Functions
Section 7.8. Summary
Reference
Student Exercises

Chapter 8. Internal Model Control

Section 8.1. Introduction to Model-Based Control
Section 8.2. Practical Open-Loop Controller Design
Section 8.3. Generalization of the Open-Loop Control Design Procedure
Section 8.4. Model Uncertainty and Disturbances
Section 8.5. Development of the IMC Structure
Section 8.6. IMC Background
Section 8.7. The IMC Structure
Section 8.8. The IMC Design Procedure
Section 8.9. Effect of Model Uncertainty and Disturbances
Section 8.10. Improved Disturbance Rejection Design
Section 8.11. Manipulated Variable Saturation
Section 8.12. Summary
References
Student Exercises
Appendix 8.1: Derivation of Closed-Loop Relationships for IMC

Chapter 9. The IMC-Based PID Procedure

Section 9.1. Background
Section 9.2. The Equivalent Feedback Form to IMC
Section 9.3. IMC-Based Feedback Design for Delay-Free Processes
Section 9.4. IMC-Based Feedback Design for Processes with a Time Delay
Section 9.5. Summary of IMC-Based PID Controller Design for Stable Processes
Section 9.6. IMC-Based PID Controller Design for Unstable Processes
Section 9.7. Summary
References
 Student Exercises

Chapter 10. Cascade and Feed-Forward Control

Section 10.1. Background
Section 10.2. Introduction to Cascade Control
Section 10.3. Cascade-Control Analysis
Section 10.4. Cascade-Control Design
Section 10.5. Cascade IMC
Section 10.6. Feed-Forward Control
Section 10.7. Feed-Forward Controller Design
Section 10.8. Feed-Forward Control in the IMC Structure
Section 10.9. Summary of Feed-Forward Control
Section 10.10. Combined Feed-Forward and Cascade
Section 10.11. Summary
References
Student Exercises—Cascade Control
Student Exercises—Feed-Forward Control
Student Exercises—Feed-Forward and Cascade

Chapter 11. PID Enhancements

Section 11.1. Background
Section 11.2. Antireset Windup
Section 11.3. Autotuning Techniques
Section 11.4. Nonlinear PID Control
Section 11.5. Controller Parameter (Gain) Scheduling
Section 11.6. Measurement/Actuator Selection
Section 11.7. Implementing PID Enhancements in Simulink
Section 11.8. Summary
References
Student Exercises

Chapter 12. Ratio, Selective, and Split-Range Control

Section 12.1. Motivation
Section 12.2. Ratio Control
Section 12.3. Selective and Override Control
Section 12.4. Split-Range Control
Section 12.5. Simulink Functions
Section 12.6. Summary
References
Student Exercises

Chapter 13. Control-Loop Interaction

Section 13.1. Introduction
Section 13.2. Motivation
Section 13.3. The General Pairing Problem
Section 13.4. The Relative Gain Array
Section 13.5. Properties and Application of the RGA
Section 13.6. Return to the Motivating Example
Section 13.7. RGA and Sensitivity
Section 13.8. Using the RGA to Determine Variable Pairings
 Section 13.9. MATLAB RGA Function File
Section 13.10. Summary
References
Student Exercises
Appendix 13.1: Derivation of the Relative Gain for an n-Input–n-Output System
Appendix 13.2: m-File to Calculate the RGA

Chapter 14. Multivariable Control

Section 14.1. Background
Section 14.2. Zeros and Performance Limitations
Section 14.3. Scaling Considerations
Section 14.4. Directional Sensitivity and Operability
Section 14.5. Block-Diagram Analysis
Section 14.6. Decoupling
Section 14.7. IMC
Section 14.8. MATLAB tzero, svd, and LTI Functions
Section 14.9. Summary
References
Student Exercises
Appendix 14.1

Chapter 15. Plantwide Control

Section 15.1. Background
Section 15.2. Steady-State and Dynamic Effects of Recycle
Section 15.3. Unit Operations Not Previously Covered
Section 15.4. The Control and Optimization Hierarchy
Section 15.5. Further Plantwide Control Examples
Section 15.6. Simulations
Section 15.7. Summary
References
Student Exercises

Chapter 16. Model Predictive Control

Section 16.1. Motivation
Section 16.2. Optimization Problem
Section 16.3. Dynamic Matrix Control
Section 16.4. Constraints and Multivariable Systems
Section 16.5. Other MPC Methods
Section 16.6. Matlab
Section 16.7. Summary
References and Relevant Literature
Student Exercises
Appendix 16.1: Derivation of the Step Response Formulation
Appendix 16.2: Derivation of the Least Squares Solution for Control Moves
Appendix 16.3

Chapter 17. Summary

Section 17.1. Overview of Topics Covered in This Textbook
Section 17.2. Process Engineering in Practice
Section 17.3. Suggested Further Reading
 Section 17.4. Notation
Student Exercises

Module 1. Introduction to MATLAB

Section M1.1. Background
Section M1.2. Matrix Operations
Section M1.3. The MATLAB Workspace
Section M1.4. Complex Variables
Section M1.5. Plotting
Section M1.6. More Matrix Stuff
Section M1.7. For Loops
Section M1.8. m-Files
Section M1.9. Summary of Commonly Used Commands
Section M1.10. Frequently Used MATLAB Functions
Additional Exercises

Module 2. Introduction to SIMULINK

Section M2.1. Background
Section M2.2. Open-Loop Simulations
Section M2.3. Feedback-Control Simulations
Section M2.4. Developing Alternative Controller Icons
Section M2.5. Summary
Additional Exercises

Module 3. Ordinary Differential Equations

Section M3.1. M ATLAB ode—Basic
Section M3.2. M ATLAB ode—Options
Section M3.3. SIMULINK sfun (.mdl Files)
Section M3.4. SIMULINK sfun (.mdl Files)—Advanced
Section M3.5. Summary

Module 4. MATLAB LTI Models

Section M4.1. Forming Continuous-Time Models
Section M4.2. Forming Discrete-Time Models
Section M4.3. Converting Continuous Models to Discrete
Section M4.4. Converting Discrete Models to Continuous
Section M4.5. Step and Impulse Responses
Section M4.6. Summary
Reference
Additional Exercises

Module 5. Isothermal Chemical Reactor

Section M5.1. Background
Section M5.2. Model (Chapter 2)
Section M5.3. Steady-State and Dynamic Behavior (Chapter 3)
Section M5.4. Classical Feedback Control (Chapters 5 and 6)
Section M5.5. Internal Model Control (Chapter 8)
Reference
Additional Exercises
Module 6. First-Order + Time-Delay Processes
Section M6.1. Motivation
Section M6.2. Closed-Loop Time-Domain Simulation
Section M6.3. Bode Analysis
Section M6.4. Ziegler-Nichols Tuning
Section M6.5. IMC-Based PID Control
Section M6.6. Summary
References
Additional Exercises
Appendix M6.1

Module 7. Biochemical Reactors

Section M7.1. Background
Section M7.2. Steady-State and Dynamic Behavior
Section M7.3. Stable Steady-State Operating Point
Section M7.4. Unstable Steady-State Operating Point
Section M7.5. SIMULINK Model File
Reference
Additional Exercises

Module 8. CSTR
Section M8.1. Background
Section M8.2. Simplified Modeling Equations
Section M8.3. Example Chemical Process—Propylene Glycol Production
Section M8.4. Effect of Reactor Scale
Section M8.5. For Further Study: Detailed Model
Section M8.6. Other Considerations
Section M8.7. Summary
References
Additional Exercises
Appendix M8.1

Module 9. Steam Drum Level

Section M9.1. Background
Section M9.2. Process Model
Section M9. Feedback Controller Design
Section M9.4. Feed-Forward Controller Design
Section M9.5. Three-Mode Level Control
Appendix M9.1: S IMULINK Diagram for Feed-Forward/Feedback Control of Steam Drum Level
Appendix M9.2: S IMULINK Diagram for 3-Mode Control of Steam Drum Level

Module 10. Surge Vessel Level Control

Section M10.1. Background
Section M10.2. Process Model
Section M10.3. Controller Design
Section M10.4. Numerical Example
Section M10.5. Summary
Reference
Additional Exercises
Appendix M10.1: The SIMULINK Block Diagram
Module 11. Batch Reactor
Section M11.1. Background
Section M11.2. Batch Model 1: Jacket Temperature Manipulated
Section M11.3. Batch Model 2: Jacket Inlet Temperature Manipulated
Section M11.4. Batch Model 3: Cascade Control
Section M11.5. Summary
Reference
Additional Exercises

Module 12. Biomedical Systems

Section M12.1. Overview
Section M12.2. Pharmacokinetic Models
Section M12.3. Intravenous Delivery of Anesthetic Drugs
Section M12.4. Blood Glucose Control in Diabetic Patients
Section M12.5. Blood Pressure Control in Post-Operative Patients
Section M12.6. Critical Care Patients
Section M12.7. Summary
References
Additional Exercises

Module 13. Distillation Control

Section M13.1. Description of Distillation Control
Section M13.2. Open-Loop Behavior
Section M13.3. SISO Control
Section M13.4. RGA Analysis
Section M13.5. Multiple SISO Controllers
Section M13.6. Singular Value Analysis
Section M13.7. Nonlinear Effects
Section M13.8. Other Issues in Distillation Column Control
Section M13.9. Summary
References
Additional Exercises
Appendix M13.1

Module 14. Case Study Problems

Section M14.1. Background
Section M14.2. Reactive Ion Etcher
Section M14.3. Rotary Lime Kiln Temperature Control
Section M14.4. Fluidized Catalytic Cracking Unit
Section M14.5. Anaerobic Sludge Digester
Section M14.6. Drug Infusion System
Section M14.7. Suggested Case Study Schedule
Section M14.8. Summary
Additional Exercises

Module 15. Flow Control

Section M15.1. Motivating Example
Section M15.2. Flowmeters
Section M15.3. Control Valves
 Section M15.4. Pumping and Piping Systems
Section M15.5. Summary
References
Additional Exercises

Module 16. Digital Control

Section M16.1. Background
Section M16.2. PID Controllers
Section M16.3. Stability Analysis for Digital Control Systems
Section M16.4. Performance of Digital Control Systems
Section M16.5. Discrete IMC
Section M16.6. Summary
References
Additional Exercises
Appendix M16.1: SIMULINK .mdl File for Example M16.2
Appendix M16.2: SIMULINK .m and .mdl Files for Example M16.3

About the Author

[ Team LiB ]
[ Team LiB ]

Copyright
Library of Congress Cataloging-in-Publication Data

Bequette, B. Wayne, 1957–

Process control : modeling, design, and simulation / by B. Wayne Bequette.

p. cm. — (Prentice-Hall international series in the physical and chemical engineering sciences)

ISBN 0-13-353640-8 (paper : alk. paper)

1. Process control—Mathematical models. 2. Manufacturing processes—Mathematical models. I.

Title. II. Series.

TS156.8 .B46 2003

670.42'7—dc21

2002032953

Editorial/production supervision: Argosy

Full-service production manager: Anne R. Garcia

Cover design director: Jerry Votta

Cover design: Talar Agasyan-Boorujy

Manufacturing buyer: Maura Zaldivar

Publisher: Bernard Goodwin

Editorial assistant: Michelle Vincenti

Marketing manager: Dan DePasquale

©2003 Pearson Education, Inc.

Publishing as Prentice Hall Professional Technical Reference

Upper Saddle River, New Jersey 07458

Prentice Hall books are widely used by corporations and government agencies for training,
marketing, and resale.

For information regarding corporate and government bulk discounts please contact:

Corporate and Government Sales (800) 382-3419 or

[email protected]

Other product or company names mentioned herein are the trademarks or registered trademarks
of their respective owners.
For MATLAB product information, please contact:

The MathWorks, Inc.

3 Apple Hill Drive

Natick, MA 01760-2098 USA

Tel: 508-647-7000

Fax: 508-647-7101

E-mail: [email protected]

Web: www.mathworks.com

All rights reserved. No part of this book may be reproduced, in any form or by any means,
without permission in writing from the publisher.

Printed in the United States of America

10 9 8 7 6 5 4 3 2 1

Pearson Education LTD.

Pearson Education Australia PTY, Limited

Pearson Education Singapore, Pte. Ltd.

Pearson Education North Asia Ltd.

Pearson Education Canada, Ltd.

Pearson Educación de Mexico, S.A. de C.V.

Pearson Education—Japan

Pearson Education Malaysia, Pte, Ltd.

Dedication
To Pat, Brendan, and Eileen and my parents, Bill and Ayleen Bequette

[ Team LiB ]
[ Team LiB ]

PRENTICE HALL INTERNATIONAL

SERIES IN THE PHYSICAL AND
CHEMICAL ENGINEERING SCIENCES
NEAL R. AMUNDSON, SERIES EDITOR, University of Houston

ADVISORY EDITORS

ANDREAS ACRIVOS, Stanford University

JOHN D AHLER, University of Minnesota

H. SCOTT FOGLER, University of Michigan

THOMAS J. H ANRATTY , University of Illinois

JOHN M. PRAUSNITZ, University of California

L. E. SCRIVEN, University of Minnesota

BALZHISER, SAMUELS, AND ELIASSEN Chemical Engineering Thermodynamics

BEQUETTE Process Control: Modeling, Design, and Simulation

BEQUETTE Process Dynamics

BIEGLER, GROSSMAN, AND WESTERBERG Systematic Methods of Chemical Process Design

BROSILOW AND JOSEPH Techniques of Model-based Control

CROWL AND LOUVAR Chemical Process Safety: Fundamentals with Applications, 2nd edition

CONSTANTINIDES AND M OSTOUFI Numerical Methods for Chemical Engineers with MATLAB Applications

CUTLIP AND SHACHAM Problem Solving in Chemical Engineering with Numerical Methods

DENN Process Fluid Mechanics

DOYLE Process Control Modules: A Software Laboratory for Control Design

ELLIOT AND LIRA Introductory Chemical Engineering Thermodynamics

F OGLER Elements of Chemical Reaction Engineering, 3rd edition

HIMMELBLAU Basic Principles and Calculations in Chemical Engineering, 6th edition

HINES AND M ADDOX Mass Transfer

KYLE Chemical and Process Thermodynamics, 3rd edition

PRAUSNITZ, L ICHTENTHALER, AND DE AZEVEDO Molecular Thermodynamics of Fluid-Phase Equilibria, 3rd

edition

PRENTICE Electrochemical Engineering Principles

SHULER AND K ARGI Bioprocess Engineering, 2nd edition

STEPHANOPOULOS Chemical Process Control

TESTER AND M ODELL Thermodynamics and Its Applications, 3rd edition

TURTON, BAILIE, WHITING, AND SHAEIWITZ Analysis, Synthesis, and Design of Chemical Processes, 2nd
edition

WILKES Fluid Mechanics for Chemical Engineering

[ Team LiB ]
[ Team LiB ]

About Prentice Hall Professional

Technical Reference
With origins reaching back to the industry's first computer science publishing program in the
19605, Prentice Hall Professional Technical Reference (PH PTR) has developed into the leading
provider of technical books in the world today. Formally launched as its own imprint in 1986, our
editors now publish over 200 books annually, authored by leaders in the fields of computing,
engineering, and business.

Our roots are firmly planted in the soil that gave rise to the technological revolution. Our
bookshelf contains many of the industry's computing and engineering classics: Kernighan and
Ritchie's C Programming Language, Nemeth's UNIX System Administration Handbook,
Horstmann's Core Java, and Johnson's High-Speed Digital Design.

PH PTR acknowledges its auspicious beginnings while it looks to the future for inspiration. We
continue to evolve and break new ground in publishing by providing today's professionals with
tomorrow's solutions.

[ Team LiB ]
[ Team LiB ]

Preface
Background

Textbook Goals

Chapters

Learning Modules

Textbook Web Page

Acknowledgments
[ Team LiB ]
[ Team LiB ]

Background
There are a variety of courses in a standard chemical engineering curriculum, ranging from the
introductory material and energy balances course, and culminating with the capstone process
design course. The focus of virtually all of these courses is on steady-state behavior; the rare
exceptions include the analysis of batch reactors and batch distillation in the reaction engineering
and equilibrium stage operations courses, respectively. A concern of a practicing process
engineer, on the other hand, is how to best operate a process plant where everything seems to
be changing. The process dynamics and control course is where students must gain an
appreciation for the dynamic nature of chemical processes and develop strategies to operate
these processes.

[ Team LiB ]
[ Team LiB ]

Textbook Goals
The major goal of this textbook is to teach students to analyze dynamic chemical processes and
develop automatic control strategies to operate them safely and economically. My experience is
that students learn best with immediate simulation-based reinforcement of basic concepts. Rather
than simply present theory topics and develop analytical solutions, this textbook uses "interactive
learning" through computer-based simulation exercises (modules). The popular MATLAB software
package, including the SIMULINK block-diagram simulation environment, is used. Students,
instructors, and practicing process engineers learning new model-based techniques can all benefit
from the "feedback" provided by simulation studies.[1]
[1]It should be noted that I am not a proponent of a solely "simulation-based" control education, where
students iteratively adjust parameters in a JavaScript simulation until acceptable responses are obtained. I
wish for students to obtain the classic mode of understanding as analyzed so well by Robert Pirsig in Zen
and the Art of Motorcycle Maintenance (Bantam Books, 1974). This deeper understanding of process control
can be obtained by rigorous analysis and by selected simulations where the student plays a direct role in the
implementation of an algorithm or strategy of choice.

Depending on the goals of the instructor and the background of the students, roughly one chapter
(± 0.5) and one module can be covered each week. Still, it is probably too ambitious to cover the
entire text during a typical 15-week semester, so I recommend that instructors carefully choose
the topics that best meet their personal objectives. At Rensselaer Polytechnic Institute, we teach
the one-semester, 4-credit course in a studio-based format, with students attending two 2-hour
sessions and one 2-hour recitation (which also provides plenty of "catch-up" time) each week.
During these sessions we typically spend 45 minutes discussing a topic, then have the students
spend the remaining hour performing analysis and computer simulation exercises, working in
pairs. During the discussion periods the students face the instructor station at the front of the
room, and during the simulation periods they swivel in their chairs to the workstations on the
countertops behind them. This textbook can also be used in a more traditional lecture-based
course, with students working on the modules and solving homework problems on their own.

[ Team LiB ]
[ Team LiB ]

Chapters
An introduction to process control and instrumentation is presented in Chapter 1. The
development and use of models is very important in control systems engineering. Fundamental
models are developed in Chapter 2, including the steady-state solution and linearization to form
state space models. Chapter 3 focuses on the dynamic behavior of linear systems, starting with
state space models and then covering transfer function-based models in detail. Chapter 4 we
cover the development of empirical models, including continuous and discrete transfer function
models.

Chapter 5 provides a more detailed introduction to feedback control, developing the basic idea of
a feedback system, proportional, integral, derivative (PID) controllers, and methods of analyzing
closed-loop stability. Chapter 6 presents the Ziegler-Nichols closed-loop oscillation method for
controller tuning, since the same basic concept is used in the automatic tuning procedures
presented in Chapter 11. Frequency response analysis techniques, important for determining
control system robustness, are presented in Chapter 7.

In recent years model-based control has lead to improved control loop performance. One of the
clearest model-based techniques is internal model control (IMC), which is presented in Chapter 8.
The PID controller remains the most widely used controller in industry; in Chapter 9 we show how
to convert internal model controllers to classical feedback (PID) controllers.

In Chapter 10 the widely used cascade and feed-forward strategies are developed. Many control
loops suffer from poor performance, either because they were not tuned well originally, or
because the process is nonlinear and has changed operating conditions. Two methods of dealing
with these problems, automatic tuning and gain scheduling, are presented in Chapter 11. The
phenomenon of reset windup and the development of antireset windup strategies are also
presented in Chapter 11.

Many control strategies must be able to switch between manipulated inputs or select from several
measured outputs. Split-range, selective and override strategies are presented in Chapter 12.
Process units contain many control loops that generally do not operate independently. The effects
of these control-loop interactions are presented in Chapter 13. The design of multivariable
controllers is developed in Chapter 14.

The development of the control instrumentation diagram for an entire chemical process is
challenging and remains somewhat of an art. In Chapter 15 recycle systems are shown to cause
unique and challenging steady-state and dynamic control problems. In addition, an overview of
corporate-wide optimization and control problems is presented. Model predictive control (MPC) is
the most widely applied advanced control strategy in industry. The basic step response model-
based MPC method is developed in Chapter 16. This is followed by a discussion of the constrained
version of MPC, and enhancements to improve disturbance rejection.

[ Team LiB ]
[ Team LiB ]

Learning Modules
The chapters are followed by a series of learning modules that serve several purposes; some
focus on the software tools, while others focus on particular control problems. The first two
provide introductions to MATLAB and SIMULINK, the simulation environment for the modules that
follow. The third module demonstrates the solution of ordinary differential equations using MATLAB
and S IMULINK, while the fourth shows how to use the MATLAB Control Toolbox to create and convert
models from one form to another. The modules that follow focus on a particular unit operation, to
provide a detailed demonstration of various control system design, analysis or implementation
techniques. Module 5 develops a simple isothermal CSTR model that is used in a number of the
chapters. Module 6 details the robustness analysis of processes characterized by first-order +
deadtime (FODT) models.

Module 7 presents a biochemical reactor with two possible desired operating points; one stable
and the other unstable. The controller design and system performance is clearly different at each
operating point. The classic jacketed CSTR with an exothermic reaction is studied in Module 8.
Issues discussed include recirculation heat transfer dynamics, cascade control, and split-range
control. Level control loops can be tuned for two different extremes of closed-loop performance,
as shown in Module 9 (steam drum, requiring tight level control) and Module 10 (surge drum,
allowing loose level control to minimize outflow variation). Challenges associated with jacketed
batch reactors are presented in Module 11. Some motivating biomedical problems are presented
in Module 12. Challenges of control loop interaction are demonstrated in the distillation application
of Module 13. Module 14 provides an overview of several case study problems in multivariable
control. Here the students can download SIMULINK .mdl files for the textbook web page and
perform complete modeling and control system design. These case studies are meant to tie
together many concepts presented in the text. Issues particular to flow control are discussed in
Module 15, and digital control techniques are presented in Module 16.

[ Team LiB ]
[ Team LiB ]

Textbook Web Page

MATLAB and SIMULINK files, as well as additional learning material and errata, can be found on the
textbook web page:

www.rpi.edu/dept/chem-eng/WWW/faculty/bequette/books/Process_Control/

or

www.rpi.edu/~bequeb/Control

[ Team LiB ]
[ Team LiB ]

Acknowledgments
A few acknowledgments are in order. First of all, Professor Jim Turpin at the University of
Arkansas stimulated my interest in process dynamics and control when I took his course as an
undergraduate. As a neophyte process engineer for American Petrofina I had the opportunity to
serve as a process operator during two work-stoppages. A newfound respect for control loop
interaction led me to graduate study at the University of Texas, where Professor Tom Edgar
provided the "degrees of freedom" for me to explore a range of control topics. Collaborations at
Merck, Inc., led to the presentation of modeling and control of batch reactors in Module 11.
Research sponsored by the Whitaker Foundation and the National Science Foundation resulted in
material presented in Modules 12 and 14.

My own graduate students have served as teaching assistants in the dynamics and control
courses, and have provided me with valuable feedback on various versions of this textbook. In
particular, Lou Russo, now at ExxonMobil, helped me understand what works and what does not
work in the classroom and in homework assignments. He certainly had a major positive impact on
the education of many Rensselaer undergraduates.

Professor Robert Parker at the University of Pittsburgh classroom tested this textbook, and made
a number of valuable suggestions. In addition, Brian Aufderheide (now at the Keck Graduate
Institute) critiqued Chapter 16.

My colleagues at Rensselaer have promoted an environment that provides an optimum mix of

teaching and research; our department has published four textbooks during the past two years.
Various educational initiatives at Rensselaer have allowed me to develop an interactive learning
approach to dynamics and control. In particular, the Control Engineering Studio environment
gives me immediate feedback on the level of practical understanding on a particular topic and
allows me to give immediate feedback to students. A Curriculum Innovation grant from P&G led to
the development of experiments and learning modules for the dynamics and control course, and
for other courses using the Control Engineering Studio classroom.

Various Troy and Albany establishments have served to "gain schedule" my personal regulatory
system and allowed me to obtain a better understanding of the pharma cokinetics and
pharmacodynamics of caffeine and ethanol. The Daily Grind (www.dailygrind.com; you won't find
a better coffee roaster in Seattle) in Albany provided beans for the many espressos that "kick
started" numerous sections of this textbook. Group meetings at the Troy Pub & Uncle Sam
Brewery (www.troypub.com; try the Harwood Porter the next time you are in town) led to many
interesting education and research[2] discussions (not to mention political and other topics).
[2]The important interplay of research and education should not be overlooked. Seemingly innocuous
problems assigned in the control class have led to interesting graduate research projects. Similarly,
graduate research results have been brought into the undergraduate classroom.

Naturally, completing this text would have been a struggle without the support of my wife, Pat
Fahy, and the good sleeping habits of my kids, Brendan and Eileen. They have done their best to
convince me that not all systems are controllable.

[ Team LiB ]
[ Team LiB ]

Chapter 1. Introduction
The goal of this chapter is to provide a motivation for, and an introduction to, process control and
instrumentation. After studying this chapter, the reader, given a process, should be able to do the
following:

Determine possible control objectives, input variables (manipulated and disturbance) and
output variables (measured and unmeasured), and constraints (hard or soft), as well as
classify the process as continuous, batch, or semicontinuous

Assess the importance of process control from safety, environmental, and economic points
of view

Sketch a process instrumentation and control diagram

Draw a simplified control block diagram

Understand the basic idea of feedback control

Understand basic sensors (measurement devices) and actuators (manipulated inputs)

Begin to develop intuition about characteristic timescales of dynamic behavior

The major sections of this chapter are as follows:

1.1 Introduction

1.2 Instrumentation

1.3 Process Models and Dynamic Behavior

1.4 Control Textbooks and Journals

1.5 A Look Ahead

1.6 Summary

[ Team LiB ]
[ Team LiB ]

1.1 Introduction
Process engineers are often responsible for the operation of chemical processes. As these
processes become larger scale and/or more complex, the role of process automation becomes
more and more important. The objective of this textbook is to teach process engineers how to
design and tune feedback controllers for the automated operation of chemical processes.

A conceptual process block diagram for a chemical process is shown in Figure 1-1. Notice that the
inputs are classified as either manipulated or disturbance variables and the outputs are classified
as measured or unmeasured in Figure 1-1a. To automate the operation of a process, it is
important to use measurements of process outputs or disturbance inputs to make decisions about
the proper values of manipulated inputs. This is the purpose of the controller shown in Figure 1-
1b; the measurement and control signals are shown as dashed lines. These initial concepts
probably seem very vague or abstract to you at this point. Do not worry, because we present a
number of examples in this chapter to clarify these ideas.

Figure 1-1. Conceptual process input/output block diagram.

The development of a control strategy consists of formulating or identifying the following.

1. Control objective(s).

2. Input variables— classify these as (a) manipulated or (b) disturbance variables; inputs may
change continuously, or at discrete intervals of time.

3.
 1.

2.

3. Output variables— classify these as (a) measured or (b) unmeasured variables;

measurements may be made continuously or at discrete intervals of time.

4. Constraints— classify these as (a) hard or (b) soft.

5. Operating characteristics— classify these as (a) continuous, (b) batch, or (c) semicontinuous
(or semibatch).

6. Safety, environmental, and economic considerations.

7. Control structure— the controllers can be feedback or feed forward in nature.

Here we discuss each of the steps in formulating a control problem in more detail.

1. The first step of developing a control strategy is to formulate the control objective(s). A
chemical-process operating unit often consists of several unit operations. The control of an
operating unit is generally reduced to considering the control of each unit operation
separately. Even so, each unit operation may have multiple, sometimes conflicting
objectives, so the development of control objectives is not a trivial problem.

2. Input variables can be classified as manipulated or disturbance variables. A manipulated

input is one that can be adjusted by the control system (or process operator). A disturbance
input is a variable that affects the process outputs but that cannot be adjusted by the
control system. Inputs may change continuously or at discrete intervals of time.

3. Output variables can be classified as measured or unmeasured variables. Measurements

may be made continuously or at discrete intervals of time.

4. Any process has certain operating constraints, which are classified as hard or soft. An
example of a hard constraint is a minimum or maximum flow rate—a valve operates
between the extremes of fully closed or fully open. An example of a soft constraint is a
product composition—it may be desirable to specify a composition between certain values to
sell a product, but it is possible to violate this specification without posing a safety or
environmental hazard.

5. Operating characteristics are usually classified as continuous, batch, or semicontinuous

(semibatch). Continuous processes operate for long periods of time under relatively constant
operating conditions before being "shut down" for cleaning, catalyst regeneration, and so
forth. For example, some processes in the oil-refining industry operate for 18 months
between shutdowns. Batch processes are dynamic in nature—that is, they generally operate
for a short period of time and the operating conditions may vary quite a bit during that
period of time. Example batch processes include beer or wine fermentation, as well as many
specialty chemical processes. For a batch reactor, an initial charge is made to the reactor,
and conditions (temperature, pressure) are varied to produce a desired product at the end
of the batch time. A typical semibatch process may have an initial charge to the reactor, but
feed components may be added to the reactor during the course of the batch run.

Another important consideration is the dominant timescale of a process. For continuous

processes this is very often related to the residence time of the vessel. For example, a
vessel with a liquid volume of 100 liters and a flow rate of 10 liters/minute would have a
residence time of 10 minutes; that is, on the average, an element of fluid is retained in the
vessel for 10 minutes.

6. Safety, environmental, and economic considerations are all very important. In a sense,
economics is the ultimate driving force—an unsafe or environmentally hazardous process will
 6.

ultimately cost more to operate, through fines paid, insurance costs, and so forth. In many
industries (petroleum refining, for example), it is important to minimize energy costs while
producing products that meet certain specifications. Better process automation and control
allows processes to operate closer to "optimum" conditions and to produce products where
variability specifications are satisfied.

The concept of "fail-safe" is always important in the selection of instrumentation. For

example, a control valve needs an energy source to move the valve stem and change the
flow; most often this is a pneumatic signal (usually 3–15 psig). If the signal is lost, then the
valve stem will go to the 3-psig limit. If the valve is air-to-open, then the loss of instrument
air will cause the valve to close; this is known as a fail-closed valve. If, on the other hand, a
valve is air-to-close, when instrument air is lost the valve will go to its fully open state; this
is known as a fail-open valve.

7. The two standard control types are feed forward and feedback. A feed-forward controller
measures the disturbance variable and sends this value to a controller, which adjusts the
manipulated variable. A feedback control system measures the output variable, compares
that value to the desired output value, and uses this information to adjust the manipulated
variable. For the first part of this textbook, we emphasize feedback control of single-input
(manipulated) and single-output (measured) systems. Determining the feedback control
structure for these systems consists of deciding which manipulated variable will be adjusted
to control which measured variable. The desired value of the measured process output is
called the setpoint.

A particularly important concept used in control system design is process gain. The process gain is
the sensitivity of a process output to a change in the process input. If an increase in a process
input leads to an increase in the process output, this is known as a positive gain. If, on the other
hand, an increase in the process input leads to a decrease in the process output, this is known as
a negative gain. The magnitude of the process gain is also important. For example, a change in
power (input) of 0.5 kW to a laboratory-scale heater may lead to a fluid temperature (output)
change of 10°C; this is a process gain (change in output/change in input) of 20°C/kW. The same
input power change of 0.5 kW to a larger scale heater may only yield an output change of 0.5°C,
corresponding to a process gain of 1°C/kW.

Once the control structure is determined, it is important to decide on the control algorithm. The
control algorithm uses measured output variable values (along with desired output values) to
change the manipulated input variable. A control algorithm has a number of control parameters,
which must be "tuned" (adjusted) to have acceptable performance. Often the tuning is done on a
simulation model before implementing the control strategy on the actual process. A significant
portion of this textbook is on the use of model-based control, that is, controllers that have a
model of the process "built in."

This approach is best illustrated by way of example. Since many important concepts, such as
control instrumentation diagrams and control block diagrams, are introduced in the next
examples, it is important that you study them thoroughly.

Example 1.1: Surge Tank

Surge tanks are often used as intermediate storage for fluid streams being transferred between
process units. Consider the process flow diagram shown in Figure 1-2, where a fluid stream from
process 1 is fed to the surge tank; the effluent from the surge tank is sent to process 2.

Figure 1-2. Tank level problem.

There are obvious constraints on the height in this tank. If the tank overflows it may create safety
and environmental hazards, which may also have economic significance. Let us analyze this
system using a step-by-step procedure.

1. Control objective: The control objective is to maintain the height within certain bounds. If it
is too high it will overflow and if it is too low there may be problems with the flow to process
2. Usually, a specific desired height will be selected. This desired height is known as the
setpoint.

2. Input variables: The input variables are the flow from process 1 and the flow to process 2.
Notice that an outlet flow rate is considered an input to this problem. The question is which
input is manipulated and which is a disturbance? That depends. We discuss this problem
further in a moment.

3. Output variables: The most important output variable is the liquid level. We assume that it is
measured.

4. Constraints: There are a number of constraints in this problem. There is a maximum liquid
level; if this is exceeded, the tank will overflow. There are minimum and maximum flow
rates through the inlet and outlet valves.

5. Operating characteristics: We assume that this is a continuous process, that is, that there is
a continuous flow in and out of the tank. It would be a semicontinuous process if, for
example, there was an inlet flow with no outlet flow (if the tank was simply being filled).

6. Safety, environmental, and economic considerations: These aspects depend somewhat on

the fluid characteristics. If it is a hazardous chemical, then there is a tremendous incentive
from safety and environmental considerations to not allow the tank to overflow. Indeed, this
is also an economic consideration, since injuries to employees or environmental cleanup
costs money. Even if the substance is water, it has likely been treated by an upstream
process unit, so losing water owing to overflow will incur an economic penalty.

Safety considerations play an important role in the specification of control valves (fail-open
or fail-closed). For this particular problem, the control-valve specification will depend on
which input is manipulated. This is discussed in detail shortly.

7. Control structure: There are numerous possibilities for control of this system. We discuss
first the feedback strategies, then the feed-forward strategies.

Feedback Control
The measured variable for a feedback control strategy is the tank height. Which input variable is
manipulated depends on what is happening in process 1 and process 2. Let us consider two
different scenarios.

Scenario 1

Process 2 regulates the flow-rate F 2. This could happen, for example, if process 2 is a steam
generation system and process 1 is a deionization process. Process 2 varies the flow rate of water
(F2) depending on the steam demand. As far as the tank process is concerned, F2 is a "wild"
(disturbance) stream because the regulation of F2 is determined by another system. In this case
we would use F 1 as the manipulated variable; that is, F 1 is adjusted to maintain a desired tank
height.

The control and instrumentation diagram for a feedback control strategy for scenario 1 is shown
in Figure 1-3. Notice that the level transmitter (LT) sends the measured height of liquid in the
tank (hm) to the level controller (LC). The LC compares the measured level with the desired level
(hsp, the height setpoint) and sends a pressure signal (Pv) to the valve. This valve top pressure
moves the valve stem up and down, changing the flow rate through the valve (F1). If the
controller is designed properly, the flow rate changes to bring the tank height close to the desired
setpoint. In this process and instrumentation diagram we use dashed lines to indicate signals
between different pieces of instrumentation.

Figure 1-3. Feedback control strategy 1. The level is measured and the
inlet flow rate (valve position) is manipulated.

A simplified block diagram representing this system is shown in Figure 1-4. Each signal and device
(or process) is shown on the block diagram. We use a slightly different form for block diagrams
when we use transfer function notation for control system analysis in Chapter 5. Note that each
block represents a dynamic element. We expect that the valve and LT dynamics will be much
faster than the process dynamics. We also see clearly from the block diagram why this is known
as a feedback control "loop." The controller "decides" on the valve position, which affects the inlet
flow rate (the manipulated input), which affects the level; the outlet flow rate (the disturbance
input) also affects the level. The level is measured, and that value is fed back to the controller
[which compares the measured level with the desired level (setpoint)].

Figure 1-4. Feedback control schematic (block diagram) for scenario 1.

F1 is manipulated and F 2 is a disturbance.
Notice that the control valve should be specified as fail-closed or air-to-open, so that the tank will
not overflow on loss of instrument air or other valve failure.

Scenario 2

Process 1 regulates flow rate F 1. This could happen, for example, if process 1 is producing a
chemical compound that must be processed by process 2. Perhaps process 1 is set to produce F1
at a certain rate. F 1 is then considered "wild" (a disturbance) by the tank process. In this case we
would adjust F 2 to maintain the tank height. Notice that the control valve should be specified as
fail-open or air-to-close, so that the tank will not overflow on loss of instrument air or other valve
failure.

The process and instrumentation diagram for this scenario is shown in Figure 1-5. The only
difference between this and the previous instrumentation diagram (Figure 1-3) is that F2 rather
than F1 is manipulated.

Figure 1-5. Feedback control strategy 2. Outlet flow rate is

manipulated.

The simplified block diagram shown in Figure 1-6 differs from the previous case (Figure 1-4) only
because F 2 rather than F1 is manipulated. F 1 is a disturbance input.

Figure 1-6. Feedback control schematic (block diagram) for scenario 2.

F2 is manipulated and F 1 is a disturbance.
Feed-Forward Control

The previous two feedback control strategies were based on measuring the output (tank height)
and manipulating an input (the inlet flow rate in scenario 1 and the outlet flow rate in scenario 2).
In each case the manipulated variable is changed after a disturbance affects the output. The
advantage to a feed-forward control strategy is that a disturbance variable is measured and a
manipulated variable is changed before the output is affected. Consider a case where the inlet
flow rate can be changed by the upstream process unit and is therefore considered a disturbance
variable. If we can measure the inlet flow rate, we can manipulate the outlet flow rate to maintain
a constant tank height. This feed-forward control strategy is shown in Figure 1-7, where FM is the
flow measurement device and FFC is the feed-forward controller. The corresponding control block
diagram is shown in Figure 1-8. F 1 is a disturbance input that directly affects the tank height; the
value of F1 is measured by the FM device, and the information is used by an FFC to change the
manipulated input, F2.

Figure 1-7. Feed-forward control strategy. Inlet flow rate is measured

and outlet flow rate is manipulated.

Figure 1-8. Feed-forward control schematic block diagram.

The main disadvantage to this approach is sensitivity to uncertainty. If the inlet flow rate is not
perfectly measured or if the outlet flow rate cannot be manipulated perfectly, then the tank height
will not be perfectly controlled. With any small disturbance or uncertainty, the tank will eventually
overflow or run dry. In practice, FFC is combined with feedback control to account for uncertainty.
A feed-forward/feedback strategy is shown in Figure 1-9, and the corresponding block diagram is
shown in Figure 1-10. Here, the feed-forward portion allows immediate corrective action to be
taken before the disturbance (inlet flow rate) actually affects the output measurement (tank
height). The feedback controller adjusts the outlet flow rate to maintain the desired tank height,
even with errors in the inlet flow-rate measurement.

Figure 1-9. Feed-forward/feedback control strategy. The inlet flow

rate is the measured disturbance, tank height is the measured output,
and outlet flow rate is manipulated.

Figure 1-10. Feed-forward/feedback control schematic block diagram.

Discussion of Level Controller Tuning and the Dominant Timescale

Notice that we have not discussed the actual control algorithms; the details of control algorithms
and tuning are delayed until Chapter 5. Conceptually, would you prefer to tune level controllers
for "fast" or "slow" responses?

When tanks are used as surge vessels it is usually desirable to tune the controllers for a slow
return to the setpoint. This is particularly true for scenario 2, where the inlet flow rate is
considered a disturbance variable. The outlet flow rate is manipulated but affects another process.
In order to not upset the downstream process, we would like to change the outlet flow rate
slowly, yet fast enough that the tank does not overflow or go dry.

Related to the controller tuning issue is the importance of the dominant timescale of the process.
Consider the case where the maximum tank volume is 200 gallons and the steady-state operating
volume is 100 gallons. If the steady-state flow rate is 100 gallons/minute, the "residence time"
would be 1 minute. Assume the inlet flow rate is a disturbance and outlet flow rate is manipulated
(Figure 1-5). If the feed flow rate increased to 150 gallons/minute and the outlet flow rate did not
change, the tank would overflow in 2 minutes. On the other hand, if the same vessel had a
steady-state flow rate of 10 gallons/minute and the inlet flow suddenly increased to 15
gallons/minute (with no change in the outlet flow), it would take 20 minutes for the tank to
overflow. Clearly, controller tuning and concern about controller failure is different for these two
cases.

The first example was fairly easy compared with most control-system synthesis problems in
industry. Even for this simple example we found that there were many issues to be considered
and a number of decisions (specification of a fail-open or fail-closed valve, etc.) that needed to be
made. Often there will be many (and usually conflicting) objectives, many possible manipulated
variables, and numerous possible measured variables.

It is helpful to think of common, everyday activities in the context of control, so you will become
familiar with the types of control problems that can arise in practice. The following activity is just
such an example.

Example 1.2: Taking a Shower

A common multivariable control problem that we face every day is taking a shower. A simplified
process schematic is shown in Figure 1-11. We analyze this process step by step.

1.
1. Control objectives: Control objectives when taking a shower include the following:

a. to become clean

b. to be comfortable (correct temperature and water velocity as it contacts the body)

c. to "look good" (clean hair, etc.)

d. to become refreshed

To simplify our analysis, for the rest of the problem we discuss how we can satisfy the
second objective (to maintain water temperature and flow rate at comfortable values).
Similar analysis can be performed for the other objectives.

2. Input variables: The manipulated input variables are hot-water and cold-water valve
positions. Some showers can also vary the velocity by adjustment of the shower head.
Another input is body position—you can move into and out of the shower stream.
Disturbance inputs include a drop in water pressure (say, owing to a toilet flushing) and
changes in hot water temperature owing to "using up the hot water from the heater."

3. Output variables: The "measured" output variables are the temperature and flow rate (or
velocity) of the mixed stream as it contacts your body.

4. Constraints: There are minimum and maximum valve positions (and therefore flow rates) on
both streams. The maximum mixed temperature is equal to the hot water temperature and
the minimum mixed temperature is equal to the cold water temperature. The previous
constraints were hard constraints—they cannot be physically violated. An example of a soft
constraint is the mixed-stream water temperature—you do not want it to be above a certain
value because you may get scalded. This is a soft constraint because it can physically
happen, although you do not want it to happen.

5. Operating characteristics: This process is continuous while you are taking a shower but is
most likely viewed as a batch process, since it is a small part of your day. It could easily be
called a semicontinuous (semibatch) process.

6. Safety, environmental, and economic considerations: Too high of a temperature can scald
you—this is certainly a safety consideration. Economically, if your showers are too long,
more energy is consumed to heat the water, costing money. Environmentally (and
economically), more water consumption means that more water and wastewater must be
treated. An economic objective might be to minimize the shower time. However, if the
shower time is too short, or not frequent enough, your clothes will become dirty and must
be washed more often—increasing your clothes-cleaning bill.

7. Control structure: This is a multivariable control problem because adjusting either valve
affects both temperature and flow rate. Control manipulations must be "coordinated," that
is, if the hot-water flow rate is increased to increase the temperature, the cold-water flow
rate must be decreased to maintain the same total flow rate. The measurement signals are
continuous, but the manipulated variable changes are likely to be discrete (unless your
hands are continuously varying the valve positions).

Feedback control: As the body feels the temperature changing, adjustments to one or both
valves is made. As the body senses a flow rate or velocity change, one or both valves are
adjusted.

Feed-forward control: If you hear the toilet flush, you move your body out of the stream to
 avoid the higher temperature that you anticipate. Notice that you are making a manipulated
variable change (moving your body) before the effect of an output (temperature or flow
rate) change is actually detected.

Figure 1-11. Process schematic for taking a shower.

Some showers may have a relatively large time delay (or dead time) between when a
manipulated variable change is made and when the actual output change is measured. This could
happen, for example, if there was a large pipe run between the mixing point and the shower head
(this would be considered an input time delay). Another type of time delay is measurement dead
time, for example if your body takes a while to detect a change in the temperature of the stream
contacting your body.

Notice that the control strategy used has more manipulated variables (two valve positions and
body movement) than measured outputs (total mixed-stream flow rate and temperature).

In the shower example, the individual taking the shower served as the controller. The
measurements and manipulations for this example are somewhat qualitative (you do not know
the exact temperature or flow rate, for example). Most of the rest of the textbook consists of
quantitative controller design procedures, that is, a mathematical model of the process is used to
develop the control algorithm.

This chapter has covered the important first step of control system development—identifying
seven basic steps in analyzing a process control problem. We have used simple examples with
which you are familiar. As you learn about more chemical and environmental processes, you
should get in the habit of thinking about them from a process systems point of view, just as you
have with these simple systems.

[ Team LiB ]
[ Team LiB ]

1.2 Instrumentation
The example level-control problem had three critical pieces of instrumentation: a sensor
(measurement device), actuator (manipulated input device), and controller. The sensor measured
the tank level, the actuator changed the flow rate, and the controller determined how much to
vary the actuator, based on the sensor signal.

There are many common sensors used for chemical processes. These include temperature, level,
pressure, flow, composition, and pH. The most common manipulated input is the valve actuator
signal (usually pneumatic).

Each device in a control loop must supply or receive a signal from another device. When these
signals are continuous, such as electrical current or voltage, we use the term analog. If the
signals are communicated at discrete intervals of time, we use the term digital.

Analog
Analog or continuous signals provided the foundation for control theory and design and analysis. A
common measurement device might supply either a 4- to 20-mA or 0- to 5-V signal as a function
of time. Pneumatic analog controllers (developed primarily in the 1930s, but used in some plants
today) would use instrument air, as well as a bellows-and-springs arrangement to "calculate" a
controller output based on an input from a measurement device (typically supplied as a 3- to 15-
psig pneumatic signal). The controller output of 3–15 psig would be sent to an actuator, typically
a control valve where the pneumatic signal would move the valve stem. For large valves, the 3-
to 15-psig signal might be "amplified" to supply enough pressure to move the valve stem.

Electronic analog controllers typically receive a 4- to 20-mA or 0- to 5-V signal from a

measurement device, and use an electronic circuit to determine the controller output, which is
usually a 4- to 20-mA or 0- to 5-V signal. Again, the controller output is often sent to a control
valve that may require a 3- to 15-psig signal for valve stem actuation. In this case the 4- to 20-
mA current signal is converted to the 3- to 15-psig signal using an I/P (current-to-pneumatic)
converter.

Digital
Many devices and controllers are now based on digital communication technology. A sensor may
send a digital signal to a controller, which then does a discrete computation and sends a digital
output to the actuator. Very often, the actuator is a valve, so there is usually a D/I (digital-to-
electronic analog) converter involved. Indeed, if the valve stem is moved by a pneumatic actuator
rather than electronic, then an I/P converter may also be used.

In the past few decades, digital control-system design techniques that explicitly account for the
discrete (rather than continuous) nature of the control computations have been developed. If
small sample times are used, the tuning and performance of the digital controllers is nearly equal
to that of analog controllers.

Techniques Used in This Textbook

Most of the techniques used in this book are based on analog (continuous) control. Although
many of the control computations performed on industrial processes are digital, the discrete
sample time is usually small enough that virtually identical performance to analog control is
obtained. Our understanding of chemical processes is based on ordinary differential equations, so
it makes sense to continue to think of control in a continuous fashion. We find that controller
tuning is much more intuitive in a continuous, rather than discrete, framework. In Chapter 16 we
spend some time discussing techniques that are specific to digital control systems, namely model
predictive control (MPC).

Control Valve Placement

In Example 1.1 and in most of the examples given in this textbook, we use a simplified
representation for a control valve and signal. It should be noted that virtually all control valves
are actually installed in an arrangement similar to that shown in Figure 1-12. When the control
valve fails, the adjacent block valves can be closed; the control valve can then be removed and
replaced. During the interim, the bypass valve can be adjusted manually to maintain the desired
flow rate. Generally, these control valve "stations" are placed at ground level for easy access,
even if the pipeline is in a piperack far above the ground.

Figure 1-12. Typical control valve arrangement. When the control

valve needs to be taken out of service, the two block valves are closed
and the control valve is removed. The bypass valve can then be
manually adjusted to control the flow.

[ Team LiB ]
[ Team LiB ]

1.3 Process Models and Dynamic Behavior

Thus far we have mentioned the term model a number of times, and you probably have a vague
notion of what we mean by model. The following definition of a model is from the McGraw-Hill
Dictionary of Scientific and Technical Terms:

"A mathematical or physical system, obeying certain specified conditions, whose behavior is
used to understand a physical, biological, or social system to which it is analogous in some
way."

In this textbook, model will be taken to mean mathematical model. More specifically, we develop
process models. A working definition of process model is

a set of equations (including the necessary input data to solve the equations) that allows us
to predict the behavior of a chemical process.

Models play a very important role in control-system design. Models can be used to simulate
expected process behavior with a proposed control system. Also, models are often "embedded" in
the controller itself; in effect the controller can use a process model to anticipate the effect of a
control action. We can see from Example 1.1 that we at least need to know whether an increase
in the flow rate will increase or decrease the tank level. For example, an increase in the inlet flow
rate increases the tank level (positive gain), while an increase in the outlet flow rate leads to a
decrease in the tank level (negative gain). In order to design a controller, then, we need to know
whether an increase in the manipulated input increases or decreases the process output variable;
that is, we need to know whether the process gain is positive or negative.

An example of a process model is shown next. A number of other examples are developed in
Chapter 2.

Example 1.3: Liquid Surge Vessel Model

In the development of a dynamic model, simplifying assumptions are often made. Also, the model
requirements are a function of the end-use of the model. In this case, we are ultimately
interested in designing a controller and in simulating control-system behavior. Since we have not
covered control algorithms in depth, our objective here is to develop a model that relates the
inputs (manipulated and disturbance) to measured outputs that we wish to regulate.

For this process, we first assume that the density is constant. The model we develop should allow
us to determine how the volume of liquid in the vessel varies as a function of the inlet and outlet
flow rates. We will list the state variables, parameters, and the input and output variables. We
must also specify the required information to solve this problem (see Figure 1-2). The system is
the liquid in the tank and the liquid surface is the top boundary of the system. The following
notation is used in the modeling equations:

F 1 = inlet volumetric flow rate (volume/time);

F 2 = outlet volumetric flow rate;

V = volume of liquid in vessel;

h = height of liquid in vessel;

 r = liquid density (mass/volume);

A = cross-sectional area of vessel.

Here we write the balance equations based on an instantaneous rate-of-change,

Equation 1.1

where the total mass of fluid in the vessel is Vr, the rate of change is dVr/dt, and the density of
the outlet stream is equal to the density of the vessel contents

Equation 1.2

Notice the implicit assumption that the density of fluid in the vessel does not depend on position
(the perfect mixing assumption). This assumption allows an ordinary differential equation (ODE)
formulation. We refer to any system that can be modeled by ODEs as lumped parameter systems.
Also notice that the outlet stream density was assumed to be equal to the density of fluid in the
tank. Assuming that the density of the inlet stream and fluid in the vessel are equal, this equation
is then reduced to[1]
[1]It might be tempting to the reader to begin to directly write a "volume balance" expression, which looks
similar to Equation (1.3). We wish to make it clear that there is no such thing as a volume balance and
Equation (1.3) is only correct because of the constant density assumption. It is a good idea to always write
a mass balance expression, such as Equation (1.2), before making assumptions about the fluid density,
which may lead to Equation (1.3).

Equation 1.3

In Equation (1.3) we refer to V as a state variable, and F1 and F2 as input variables (even though
F2 is an outlet stream flow rate). If density remained in the equation, we would refer to it as a
parameter.

In order to solve this problem we must specify the inputs F 1(t) and F2(t) and the initial condition
V(0). Direct integration of Equation (1.3) yields

Equation 1.4
If, for example, the initial volume is 500 liters, the inlet flow rate is 5 liters/second and the outlet
flow rate is 4.5 liters/second, we find

Example 1.3 provides an introduction to the notion of states, inputs, and parameters. Consider
now the notion of an output. We may consider fluid volume to be a desired output that we wish to
control, for example. In that case, volume would not only be a state, it would also be considered
an output. On the other hand, we may be concerned about fluid height, rather than volume.
Volume and height are related through the constant cross-sectional area, A

Equation 1.5

Then we have the following modeling equations,

Equation 1.6

where V is a state, F1 and F2 are inputs, h is an output, and A is a parameter. We could also
rewrite the state variable equation to find

or

Equation 1.7

where fluid height is now the state variable. It should also be noted that inputs can be classified
as either manipulated inputs (that we may regulate with a control valve, for example) or
disturbance inputs. If we desired to measure fluid height and manipulate the flow rate of stream
1, for example, then F 1 would be a manipulated input, while F2 would be a disturbance input.

We have found that a single process can have different modeling equations and variables,
depending on assumptions and the objectives used when developing the model.

The liquid level process is an example of an integrating process. If the process is initially at steady
state, the inlet and outlet flow rates are equal (see Equation 1.3 or 1.7). If the inlet flow rate is
suddenly increased while the outlet flow rate remains constant, the liquid level (volume) will
increase until the vessel overflows. Similarly, if the outlet flow rate is increased while the inlet flow
rate remains constant, the tank level will decrease until the vessel is empty.

In this textbook we first develop process models based on fundamental or first-principles analysis,
that is, models that are based on known physical-chemical relationships, such as material and
energy balances, as well as reaction kinetics, transport phenomena, and thermodynamic
relationships. We then develop empirical models. An empirical model is usually developed based
on applying input changes to a process and observing the response of measured outputs. Model
parameters are adjusted so that the model outputs match the observed process outputs. This
technique is particularly useful for developing models that can be used for controller design.

[ Team LiB ]
[ Team LiB ]

1.4 Control Textbooks and Journals

There are a large number of undergraduate control textbooks that focus on control-system design
and theory. The following books include an introduction to process modeling and dynamics, in
addition to control system design.

Coughanowr, D.R., Process Systems Analysis and Control, 2nd ed., McGraw-Hill, New York
(1991).

Luyben, M.L., and W.L. Luyben, Essentials of Process Control, McGraw-Hill, New York (1997).

Luyben, W.L., Process Modeling Simulation and Control for Chemical Engineers, 2nd ed., McGraw-
Hill, New York (1990).

Marlin, T.E., Process Control: Designing Processes and Control Systems for Dynamic
Performance, 2nd ed., McGraw-Hill, New York (2000).

Ogunnaike, B.A., and W.H. Ray, Process Dynamics, Modeling and Control, Oxford, New York
(1994).

Riggs, J.B., Chemical Process Control, Ferret Publishing, Lubbock, Texas (1999).

Seborg, D.E., T.F. Edgar, and D.A. Mellichamp, Process Dynamics and Control, Wiley, New York
(1989).

Smith, C.A., and A.B. Corripio, Principles and Practice of Automatic Process Control, 2nd ed.
Wiley, New York (1997).

Stephanopoulos, G., Chemical Process Control, Prentice Hall, Englewood Cliffs, NJ (1984).

Svrcek, W.Y., D.P. Mahoney, and B.R. Young, A Real-Time Approach to Process Control, Wiley,
Chichester (2000).

The following books are generally more applied, with specific control applications detailed.

Levine, W.S. (ed.), The Control Handbook, CRC Press, Boca Raton, FL (1996).

Liptak, B.G., and K.Venczel (eds.), Instrument Engineers Handbook, Process Control Volume,
Chilton Book Company, Radnor, PA (1985).

Luyben, W.L., B.D. Tyreus, and M.L. Luyben, Plantwide Process Control, McGraw-Hill, New York
(1999).

Schork, F.J., Deshpande, P.B., and K.W. Leffew, Control of Polymerization Reactors, Marcel
Dekker, New York (1993).

Shinskey, F.G., Distillation Control, McGraw-Hill, New York (1977).

Shunta, J.P., Achieving World Class Manufacturing Through Process Control, Prentice Hall, Upper
Saddle River, NJ (1995).

The following sources often provide interesting process control problems and solutions.
Advances in Instrumentation and Control (ISA Annual Conference)

American Control Conference (ACC) Proceedings—yearly

Chemical Engineering Magazine (McGraw-Hill)—monthly

Chemical Engineering Progress—monthly

Control Engineering Practice (an IFAC Journal)

Hydrocarbon Processing (petroleum refining and petrochemicals)—monthly

Instrumentation Technology (InTech, an instrumentation industry magazine)—monthly

IEEE Control Systems Magazine—bimonthly

ISA (Instrument Society of America) Transactions

TAPPI Journal (pulp and paper industry) —monthly

The following sources tend to be more theoretical but often have useful control-related articles.

American Institute of Chemical Engineers (AIChE) Journal

Automatica (Journal of the International Federation of Automatic Control, IFAC)

Canadian Journal of Chemical Engineering

Chemical Engineering Communications

Chemical Engineering Research and Design

Chemical Engineering Science

Computers and Chemical Engineering

Conference on Decision and Control (CDC) Proceedings—yearly

Industrial and Engineering Chemistry Research (I&EC Research)

IEEE Transactions on Automatic Control

IEEE Transactions on Biomedical Engineering

IEEE Transactions on Control System Technology

International Federation of Automatic Control (IFAC) Proceedings

International Journal of Control

International Journal of Systems Sciences

Journal of Process Control

Proceedings of the IEE (part D, Control Theory and Applications)

[ Team LiB ]
[ Team LiB ]

1.5 A Look Ahead

Chapter 2 develops fundamental models based on material and energy balances, while Chapter 3
covers dynamic analysis. Chapter 4 shows how to develop empirical models from plant tests.
Chapter 5 is an introduction to feedback control and provides the first look at quantitative control-
system design procedures.

The best way to understand process control is to work many problems. In particular, it is
important to use simulation for complex problems. A numerical package that is particularly useful
for control-system analysis and simulation is MATLAB; the S IMULINK block-diagram simulator is
particularly useful. If you are not familiar with M ATLAB/ SIMULINK, we recommend that you work
through the M ATLAB and SIMULINK tutorials (Modules 1 and 2). Simply reading the tutorials will not
give you much insight into the use of MATLAB; you must sit at a computer, work through the
examples, and try new ideas that you have.

[ Team LiB ]
[ Team LiB ]

1.6 Summary
You should now be able to formulate a control problem in terms of the following:

Control objective

Inputs (manipulated or disturbance)

Outputs (measured or unmeasured)

Constraints (hard or soft)

Operating characteristics (continuous, batch, semibatch)

Safety, environmental, and economic issues

Control structure (feedback, feed forward)

You should also be able to sketch control and instrumentation diagrams, and control block
diagrams. In addition, you should be able to recommend whether a control valve should be fail-
open or fail-closed.

The following terms were introduced in this chapter:

Actuator

Air-to-close

Air-to-open

Algorithm

Control block diagram

Control valve

Controller

Deadtime or time-delay

Digital

Fail-closed

Fail-open

Gain

Integrating process

Model
 Process gain

Process and instrumentation diagram

Sensor

Setpoint

The abstract notions of states, inputs, outputs, and parameters were introduced and are covered
in more detail in Chapter 2. The examples used were as follows:

1.1 Surge Tank

1.2 Taking a Shower

1.3 Liquid Surge Vessel Model

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Discuss the following problems (a–g) in the context of control:

i. Identify control objectives;

ii. Identify input variables and classify as (a) manipulated or (b) disturbance;

iii. Identify output variables and classify these as (a) measured or (b) unmeasured;

iv. Identify constraints and classify as (a) hard or (b) soft;

v. Identify operating characteristics and classify as (a) continuous, (b) batch, (c)
semicontinuous (or semibatch);

vi. Discuss safety, environmental, and economic considerations;

vii. Discuss the types of control (feed forward or feedback).

Measurements and manipulated variables can vary continuously or may be sampled

discretely.

Select from the following:

a. Driving a car

b. Choose one of your favorite activities (skiing, basketball, making a cappuccino,

etc.)

c. A stirred tank heater

d. Beer fermentation

e. An activated sludge process

f. A household thermostat

g. Air traffic control

2: Literature Review. The process control research literature can be challenging to read,
with unique notations and rigorous mathematical analyses. Find a paper from one of
the magazines or journals listed in Section 1.4 that you would like to understand by
the time you have completed this textbook. You will find many articles to choose from,
so use some of the following criteria for your selection:

The process is interesting to you (do not choose mainly a theory paper)

The modeling equations and parameters are in the paper (make certain the
equations are ordinary differential equations and not partial differential
equations)
 There are plots that you can verify (eventually) through simulation (the plots
should be based on simulation results)

The control algorithm is clearly written

The objectives of the paper are reasonably clear to you

Provide the following:

i. A short (one paragraph) summary of the overall objectives of the paper; why are
you interested in the paper?

ii. A short list of words and concepts in the paper that are familiar to you.

Suggested Topics (choose one):

a. Fluidized catalytic cracking unit (FCCU)— petroleum refining

b. Reactive ion etching— semiconductor manufacturing

c. Rotary lime kiln— pulp and paper manufacturing

d. Continuous drug infusion— biomedical engineering and anesthesia

e. Anaerobic digester— waste treatment

f. Distillation— petrochemical and many other industries

g. Polymerization reactor— plastics

h. pH— waste treatment

i. Beer production— food and beverage

j. Paper machine headbox— pulp and paper manufacturing

k. Batch chemical reactor— pharmaceutical production

3: Instrumentation Search. Select one of the following measurement devices and use
Internet resources to learn more about it. Determine what types of signals are input to
or output from the device. For flow meters, what range of flow rates can be handled by
a particular flow meter model?

a. Vortex-shedding flow meters

b. Orifice-plate flow meters

c. Mass flow meters

d. Thermocouple-based temperature measurements

e. Differential pressure (delta P) measurements

f. Control valves

g. pH
 f.

g.

4: Work through the Module 1: Introduction to Matlab.

5: A process furnace heats a process stream from near ambient temperature to a desired
temperature of 300°C. The process stream outlet temperature is regulated by
manipulating the flow rate of fuel gas to the furnace, as shown below.

a. Discuss the objectives of this control strategy.

b. What is the measured output?

c. What is the manipulated input?

d. What are possible disturbances?

e. Is this a continuous or batch process?

f. Is this a feed-forward or feedback controller?

g. Should the control valve fail-open or fail-closed? For the strategy you chose, is
the valve gain positive or negative? Why?

h. Discuss safety, environmental, and economic issues.

i. Draw the control block schematic diagram and label all signals and blocks on the
diagram.

6: A fluidized catalytic cracking unit (FCCU) produces a significant portion of the gasoline
produced by a typical petroleum refinery. A typical FCCU processes 30,000 Bbl/day (1
Bbl = 42 gallons) of heavy gas oil from a crude oil distillation unit, producing roughly
 15,000 Bbl/day of gasoline, along with streams of other components. The value of
gasoline alone produced by this unit is on the order of $500,000/day, so any
improvement in yield and energy consumption owing to improved control can have a
significant economic impact.

Question: A control system revamp for a 30,000 Bbl/day FCCU is estimated to cost $2
million. It is expected that the implementation of advanced control schemes will result
in an economic increase of 2% in the value of products produced. Based on the value
of gasoline alone, how many days will it take to pay back the control system
investment?

7: Furnaces are often used to heat process streams to temperatures above 400°F. A
typical fired furnace may have a heat duty of 100 x 106 Btu/hour, requiring roughly
1667 scfm (standard cubic feet per minute) of natural gas (methane has a fuel value
of approximately 1000 Btu/scf). The cost of this fuel gas is on the order of $5/1000
scf, yielding an annual fuel cost of $4.4 million/year. Excess combustion air is needed
to assure complete combustion; however, too much excess air wastes energy (the
heated air simply goes out the exhaust stack). Too little excess air leads to incomplete
combustion, wasting energy and polluting the atmosphere with unburned
hydrocarbons. It is important, then, to deliver an optimum amount of combustion air
to the furnace. With the large flow rates and high temperatures involved, maintaining
safe operation is also very important. The control system must be designed so that
excess combustion air is maintained, no matter what is happening to the fuel gas flow
rate. A fired furnace control system clearly needs to satisfy safety, environmental, and
economic criteria.

Question: An advanced control scheme is estimated to save 2% in energy costs, for a

fired furnace with a heat duty of 100 x 10 6 Btu/hour. If it is desired to have a 2-year
payback period on this control system investment, what is the maximum investment
allowable?

8: Consider the surge vessel process in Example 1.3. If the steady-state volume is 500
liters, and the steady-state inlet and outlet flow rates are 50 liters/minute, find how
the liquid volume varies with time if the inlet flow rate is Fi(t) = 50 + 10 sin(0.1t),
while the outlet flow rate remains constant at 50 liters/minute.

9: The human body is composed of many innate feedback and feed-forward control loops.
For example, insulin is a hormone produced by the pancreas to regulate the blood
glucose concentration. The pancreas in a type I (insulin dependent) diabetic has lost
the ability to produce significant insulin. An insulin-dependent diabetic must monitor
her/his blood glucose (accurate blood glucose strips have been on the market for
years) and provide insulin injections several times per day. It is particularly important
to use knowledge of the meal characteristics to determine the amount of insulin
necessary to compensate for the glucose.

a. Discuss the actions taken by a type I diabetic in terms of the formulation of a

control problem. State the objectives and list all variables, etc.

b. It is desirable to form an automated closed-loop system, using a continuous

blood glucose measurement and a continuous insulin infusion pump. Draw a
"process and instrumentation" diagram and the corresponding control block
diagram.

10: Consider the following three heat exchanger control instrumentation diagrams. For
each diagram (a, b, and c), the objective is to maintain a desired cold stream outlet
temperature. Since the cold stream exiting the exchanger is fed to a reactor, it is
important that the stream temperature never be substantially higher than the setpoint
value. Please answer the two basic questions about each strategy, then the final
question (part d).

a. Basic cold stream temperature control strategy.

Is the process gain relating the manipulated flow rate to the measured
temperature positive or negative?

Should the control valve should be fail-open or fail-closed? Why?

b. Temperature control using hot stream bypass strategy.

Is the process gain relating the manipulated flow rate to the measured
temperature positive or negative?

c.
 Should the control valve should be fail-open or fail-closed? Why?

c. Temperature control using cold stream bypass strategy.

Is the process gain relating the manipulated flow rate to the measured
temperature positive or negative?

Should the control valve should be fail-open or fail-closed? Why?

d. Which strategy (a, b, or c) will have the fastest dynamic behavior? Why?

11: During surgery it is important for an anesthesiologist to regulate a patient's blood

pressure to a desired value. She does this by changing the infusion rate of vasoactive
drugs to the patient. In addition to the effect of manipulated vasoactive drugs, blood
pressure is affected by the level of anesthetic given to the patient. Discuss actions
taken by an anesthesiologist in the context of feedback control. Sketch a control block
diagram for an automated system that measures blood pressure and manipulates the
infusion rate of a vasoactive drug.

[ Team LiB ]
[ Team LiB ]

Chapter 2. Fundamental Models

In this chapter, a methodology for developing dynamic models of chemical processes is
presented. After studying this chapter, the reader should be able to do the following:

Write balance equations using the integral or instantaneous methods

Incorporate appropriate constitutive relationships into the equations

Determine the state, input, and output variables and the parameters for a particular model
(set of equations)

Determine the necessary information to solve a system of dynamic equations

Linearize a set of nonlinear equations to find the state space model

The major sections of this chapter are as follows:

2.1 Background

2.2 Balance Equations

2.3 Material Balances

2.4 Constitutive Relationships

2.5 Material and Energy Balances

2.6 Form of Dynamic Models

2.7 Linear Models and Deviation Variables

2.8 Summary

[ Team LiB ]
[ Team LiB ]

2.1 Background

Reasons for Modeling

There are many reasons for developing process models. Improving or understanding chemical
process operation is a major overall objective for developing a dynamic process model. These
models are often used for (i) operator training, (ii) process design, (iii) safety system analysis, or
(iv) process control.

Operator training: The people responsible for the operation of a chemical manufacturing process
are known as process operators. A dynamic process model can be used to perform simulations to
train process operators, in the same fashion that flight simulators are used to train airplane pilots.
Process operators can learn the proper response to upset conditions, before having to experience
them on the actual process.

Process design: A dynamic process model can be used to properly design chemical-process
equipment for a desired production rate. For example, a model of a batch chemical reactor can be
used to determine the appropriate size of the reactor to produce a certain product at a desired
rate.

Safety: Dynamic process models can also be used to design safety systems. For example, they
can be used to determine how long it will take, after a valve fails, for a system to reach a certain
pressure.

Process control: Feedback control systems are used to maintain process variables at desirable
values. For example, a control system may measure a product temperature (an output) and
adjust the steam flow rate (an input) to maintain that desired temperature. For complex systems,
particularly those with many inputs and outputs, it is necessary to base the control-system design
on a process model. Also, before a complex control system is implemented on a process, it is
normally tested by simulation.

It should be noted that no single model of a process exists, since a model only approximates the
process behavior. The desired accuracy and resulting complexity of a process model depends on
the final use of the model. Usually more-complex models will require much more data and effort
to develop than simplified models, since more model parameters will need to be determined. The
focus of this textbook is on process control, so model development is provided with this in mind.

Lumped Parameter System Models

The models developed in this textbook are known as lumped parameter systems models. These
models consist of initial-value ordinary differential equations, often based on a perfect mixing
assumption. The models have the form
where x is the vector of state variables, the vector of state variables derivatives with respect to
time equal to dx/dt, u the vector of input variables, p the vector of parameters, y the vector of
output variables, and, f(x,u,p) and g(x,u,p) the vectors of functions.

State variables are variables that naturally appear in the derivative term of ordinary differential
equation models. Common states resulting from overall material balance equations include total
mass, volume, level for liquid-phase processes, and pressure for gas-phase processes.
Component compositions are the most common states that arise from component material
balances. Temperature is the most common state arising from an energy balance modeling
equation.

This state-variable representation seems very abstract at this juncture, and it generally takes
students some time to become comfortable with it. The easiest way is to work through some
simple examples to begin to associate the notion of states, parameters, inputs, and outputs with
the physical variables associated with chemical processes. Throughout the text we use matrix and
vector notation; you may wish to review any standard linear algebra book to become familiar with
this notation. A concise review is also provided in the M ATLAB module (Module 1).

[ Team LiB ]
[ Team LiB ]

2.2 Balance Equations

In this text, we are interested in dynamic balances that have the form

This equation is deceptively simple because there may be many in and out terms, particularly for
component balances. The in and out terms would then include the generation and conversion of
species by chemical reaction, respectively. The rate of mass accumulation in a system has the
form dM/dt, where M is the total mass in the system. Similarly, the rate of energy accumulation
has the form dE/dt, where E is the total energy in a system. If Ni is used to represent the moles
of component i in a system, then dNi/dt represents the molar rate of accumulation of component i
in the system.

When solving a problem, it is important to specify what is meant by system. In some cases the
system may be microscopic in nature (a differential element, for example), while in other cases it
may be macroscopic in nature (the liquid content of a mixing tank, for example). Also, when
developing a dynamic model we can take one of two general viewpoints. One viewpoint is based
on an integral balance, while the other is based on an instantaneous balance. Integral balances
are particularly useful when developing models for distributed parameter systems, which result in
partial differential equations; the focus in this text is on ordinary differential equation-based
models. Another viewpoint is the instantaneous balance where the time rate of change is written
directly.

Integral Balances
An integral balance is developed by viewing a system at two different "snapshots" in time.
Consider a finite time interval, Dt, and perform a material balance over that time interval,

The mean-value theorems of integral and differential calculus are then used to reduce the
equations to differential equations. For example, consider the system shown in Figure 2-1, where
one boundary represents the mass in the system at time t, while the other boundary represents
the mass in the system at t + Dt.

Figure 2-1. Conceptual material balance problem.

An integral balance on the total mass in the system is written in the form

Mathematically this is written as

Equation 2.1

where M represents the total mass in the system, while and represent the mass rates
entering and leaving the system, respectively. We can write the right-hand side of Equation (2.1),
using the mean-value theorem of integral calculus, as

Equation 2.2

where 0 < a < 1. Substituting the right-hand side of Equation (2.2) into Equation (2.1), we find

Equation 2.3
By dividing Equation (2.3) by Dt, and using the mean value theorem of differential calculus (0 < b
< 1) for the left-hand side,

Equation 2.4

Equation 2.5

and by substituting Equation (2.5) into Equation (2.4),

Equation 2.6

and taking the limit as Dt goes to zero, we find

Equation 2.7

Representing the
r total mass
r as M = V , r as Fin in and asr Fout , where V is the volume, is
the mass density (mass/volume), and F is a volumetric flow rate (volume/time), we obtain the
equation

Equation 2.8

Note that we have assumed that the system is perfectly mixed, so that the density of material
leaving the system is equal to the density
r of
r material in the system ( out = ).

Instantaneous Balances
Here we write the dynamic balance equations directly, based on an instantaneous rate-of-change
Equation 2.7

which can also be written as

Equation 2.8

This is the same result obtained using an integral balance. Although the integral balance takes
longer to arrive at the same result as the instantaneous balance method, the integral balance
method is probably clearer when developing distributed parameter (partial differential equation-
based) models.

Steady State
At steady state, the derivative with respect to time is zero, by definition, so from Equation (2.7),

Equation 2.9

or from Equation (2.8),

Equation 2.10

Steady-state relationships are often used for process design and determination of optimal
operating conditions.

[ Team LiB ]
[ Team LiB ]

2.3 Material Balances

The simplest modeling problems consist of material balances. In this section we use two process
examples to illustrate the modeling techniques used. Recall that a model for a liquid surge vessel
was developed in Chapter 1 (Example 1.3).

Example 2.1: Gas Surge Drum

Surge drums are often used as intermediate storage capacity for gas streams that are transferred
between chemical process units. Consider a drum depicted below (Figure 2-2), where qi is the
inlet molar flow rate and q is the outlet molar flow rate. A typical control problem would be to
manipulate one flow rate (either in or out) to maintain a desired drum pressure. Here we develop
a model that describes how the drum pressure varies with the inlet and outlet flow rates.

Figure 2-2. Gas surge drum.

Let V = volume of the drum and n = the total amount of gas (moles) contained in the drum.

Assumption: The pressure-volume relationship is characterized by the ideal gas law, PV = nRT,
where P is pressure, T is temperature (absolute scale), and R is the ideal gas constant.

The rate of accumulation of the mass of gas in the drum is described by the material balance

Equation 2.11

where MW represents the molecular weight. Assuming that the molecular weight is constant, we
can write

Equation 2.12

From the ideal gas law, since V, R, and T are assumed constant,
so

which can be rewritten

Equation 2.13

To solve this equation for the state variable P, we must know the inputs q i and q, the parameters
R, T, and V, and the initial condition P(0). Once again, although q is the molar rate out of the
drum, we consider it to be an input in terms of solving the model.

It should be noted that just like the liquid level process discussed in Example 1.3, this is an
integrating system. For example, if the inlet molar flow rate increases while the outlet flow rate
stays constant, then the pressure increases without bound. In reality, an increase in pressure
would most likely cause an increase in outlet molar flow rate (owing to the increased driving force
for flow out of the drum). Indeed, we model that case now.

Outlet Flow as a Function of Gas Drum Pressure

Consider the case where the outlet molar flow rate is proportional to the difference in gas drum
pressure and the pressure in the downstream header piping, Ph. Let b represent a flow coefficient.
If the flow/pressure difference relationship is linear, then

Equation 2.14

So the dynamic modeling equation is

Equation 2.15

At steady state, dP/dt = 0, so we find the steady-state relationship

Equation 2.16

where we use the subscript s to indicate a steady-state solution. Solving explicitly for Ps, we find

Equation 2.17

which is a linear relationship. An increase in qis will lead to an increased value of Ps. This type of
system is known as self-regulating, since a change in an input variable eventually leads to a new
steady-state value of the output variable. Contrast self-regulating systems with integrating
systems that do not achieve a new steady state (the output "integrates" until a vessel overflows
or a tank overpressures).

The modeling equations for Examples 1.3 and 2.1 were based on writing an overall material
balance. In the case of a liquid vessel we found that either liquid volume or height could serve as
an appropriate state variable. For the gas drum we found that pressure was the most appropriate
state variable.

Liquid level and gas pressure vessels represent inventory problems, which are integrating by
nature. If there is an imbalance in the inlet and outlet flow rates, the inventory material (liquid or
gas) can easily increase or decrease beyond desirable limits. It is the independence of the flow
rates that can cause this problem. Notice, however, that a feedback controller can be designed to
regulate the inventory levels (liquid volume or gas pressure). A feedback controller manipulates a
stream flow rate to maintain a desired inventory level.

There are many control loops that a process engineer must consider at the design stage of a
process. Because of the critical nature of inventory loops, these must receive the highest level of
consideration. In Chapter 15, we find that inventory loops must be closed before other loops are
considered.

The next example illustrates the use of modeling for reactor design.

Example 2.2: An Isothermal Chemical Reactor

Ethylene oxide (A) is reacted with water (B) in a continuously stirred tank reactor (CSTR) to form
ethylene glycol (P). Assume that the CSTR is maintained at a constant temperature and that the
water is in large excess. The stoichiometric equation is

Here we develop a model (Figure 2-3) to find the concentration of each species as a function of
time.

Figure 2-3. Isothermal stirred tank reactor.

Overall Material Balance

The overall mass balance (since the tank is perfectly mixed) is

Equation 2.18

Assumption: The liquid-phase density, r, is not a function of concentration. The vessel liquid (and
outlet) density is then equal to the inlet stream density, so

and we can write Equation (2.18) as (notice this is the same result as Example 1.3)

Equation 2.19

Component Material Balances

It is convenient to work in molar units when writing component balances, particularly if chemical
reactions are involved. Let C A and CP represent the molar concentrations of A and P
(moles/volume). The component material balance equations are

Equation 2.20a

Equation 2.20b

where r A and rP represent the rate of generation of species A and P per unit volume, and CAi
represents the inlet concentration of species A. Since the water is in large excess its concentration
does not change significantly, and the reaction rate is first order with respect to the concentration
of ethylene oxide,

Equation 2.21

where k is the reaction rate constant and the minus sign indicates that A is consumed in the
reaction. Each mole of A reacts with a mole of B (from the stoichiometric equation) and produces
one mole of P, so the rate of generation of P (per unit volume) is

Equation 2.22

Expanding the left-hand side of Equation (2.20a),

Equation 2.23

Combining Equations (2.19), (2.20a), and (2.23), we find

Equation 2.24

Similarly, the concentration P can be written as

Equation 2.25

This model consists of three differential equations (2.19, 2.24, 2.25) and, therefore, three state
variables (V, CA, and CP). To solve these equations we must specify the initial conditions [V(0),
CA(0), and CP(0)], the inputs (F i, F, CAi ) as a function of time, and the parameter (k).

The state, input, and parameter vectors are

Equation 2.26

Using state-variable notation, the model has the form

Equation 2.27

or,

Equation 2.28

Simplifying Assumptions

The reactor model presented in Example 2.2 has three differential equations. Often other
simplifying assumptions are made to reduce the number of differential equations to make them
easier to analyze and faster to solve. Assuming a constant volume (dV/dt = 0), perhaps owing to
a feedback controller, reduces the number of equations by one.

The resulting differential equations (since we assumed dV/dt = 0, F = Fi) are

Equation 2.29

Steady-State Solution

At steady state, we find the following relationships (where the subscript s represents a steady-
state solution):

Equation 2.30

Notice that the concentrations are a function of the space velocity (F s/V), which has units of
inverse time. The space velocity can be thought of as the number of reactor volumes that
"change over" per unit time. It is inversely related to the fluid residence time (V/Fs), which has
units of time and can be thought of as the average time that an element of fluid spends in the
reactor.

The concept of conversion is important in chemical-reaction engineering. The conversion of

reactant A is defined as the fraction of the feed-stream component that is reacted.

Equation 2.31

So, from Equations (2.30) and (2.31), we find that the conversion is related to the space velocity,

Equation 2.32

Notice that the conversion is a function of the dimensionless term kV/F s, which is known as the
Damkohler number. The Damkohler number is the ratio of the characteristic residence time to the
characteristic reaction time and is widely used by chemical-reaction engineers to understand
reactor behavior. Two different chemical-reaction systems can have the same conversion if their
Damkohler numbers are the same. A system with a large rate constant and low residence time
can have the same conversion as a system with a small rate constant and high residence time.
Numerical Example Using an Experimentally Determined Rate Constant

Laboratory chemists have determined that the reaction rate constant at 55°C is k = 0.311 min -1.
Here we find the steady-state concentrations of ethylene oxide (A) and ethylene glycol (P) as a
function of the steady-state space velocity and residence time. The plots in Figure 2-4 all illustrate
the same basic concept. On the left-hand plots, the independent variable is the space velocity,
while the right-hand plots have residence time as the independent variable. The top plots have
concentrations as the dependent variables, while the bottom plots have conversion as the
dependent variable. At low space velocities (large residence times) there is nearly complete
conversion of ethylene oxide to ethylene glycol, while at high space velocities (low residence
times) there is little conversion.

Figure 2-4. Steady-state relationships for ethylene glycol reactor.

Design Objective

It is desired to produce 100 million pounds per year of ethylene glycol. The feed-stream
concentration is 0.5 lbmol/ft3 and an 80% conversion of ethylene oxide has been determined to
be reasonable. What volume of reactor should be specified to meet the production rate
requirement? Since process plants often have a shutdown period every 18 months or so, assume
350 days/year of operation.

The design flow-rate calculation is shown below. Since 80% of the ethylene oxide is converted to
ethylene glycol, the ethylene glycol concentration is 0.4 lbmol/ft3 [see Equation (2.32)]. Since the
molecular weight is 62 lb/lbmol, the mass concentration is 24.8 lb/ft3.

The operating flow rate is

Solving Equation (2.32) for reactor volume, we find that the required volume is 102.9 ft3 or 769
gallons. It should be noted that reactors of this size range can be purchased in standard sizes.
Most likely the engineer would have a choice of 750-gallon or 1000-gallon models and would
choose the 1000-gallon model for expansion capability. Larger scale reactors (greater than
roughly 10,000–20,000 gallons) are usually special orders involving on-site construction (or off-
site with rail or truck delivery). For the remainder of this problem we assume that the reactor is
operated with a volume of 769 gallons, regardless of its maximum capacity.

Dynamic Response

Assume that a control strategy will be specified to maintain the desired ethylene glycol
concentration in the reactor by manipulating the reactor feed flow rate. In order to design the
controller, it is important to understand the dynamic response between an input change and the
observed output(s). A step change of 5% in the space velocity (F/V) yields the responses in the
ethylene oxide and ethylene glycol concentrations shown in Figure 2-5. An increase in the space
velocity (corresponding to a decrease in residence time) results in a decrease in the conversion of
A to P. We also see that it takes roughly 10 minutes for the reactor concentrations to achieve new
steady-state values. These simulations were performed by integrating differential Equations
(2.29) using the techniques presented in Section 2.6.

Figure 2-5. Response of ethylene oxide and ethylene glycol

concentrations to a step change in space velocity of 5% (from F/V =
0.0778–0.0817 min-1).
Examples 2.1 and 2.2 illustrate the use of material balances to develop models. In the gas drum
example, the state variable of interest was the drum pressure. In the isothermal ethylene glycol
reactor, the state variables of interest were the concentrations of ethylene oxide and ethylene
glycol. Material balance equations are rarely adequate to develop most models of interest. In
Section 2.5, we review the development of energy balance models, where temperature is often a
state variable. First, however, we cover the basic idea of constitutive relationships in Section 2.4.

[ Team LiB ]
[ Team LiB ]

2.4 Constitutive Relationships

Examples 2.1 and 2.2 required more than simple material balances to define the modeling
equations. These required relationships are known as constitutive equations; several examples of
constitutive equations are shown in this section.

Gas Law
Process systems containing a gas will often need a gas-law expression in the model. The ideal gas
law is commonly used to relate pressure (P), molar volume ( ), and temperature (T):

Equation 2.33

The van der Waal's P T relationship contains two parameters (a and b) that are system specific:

Equation 2.34

For other gas laws, see a thermodynamics text, such as Smith, Van Ness, and Abbott (2001).

Chemical Reactions
The rate of reaction per unit volume (mol/volume*time) is usually a function of the concentration
of the reacting species. For example, consider the reaction A + 2B C + 3D. If the rate of the
reaction of A is first order in both A and B, we use the expression

Equation 2.35

where r A is the rate of reaction of A (mol A/volume time), k the reaction rate constant, CA the
concentration of A (mol A/volume), and CB the concentration of B (mol B/volume).

Reaction rates are normally expressed in terms of generation of a species. The minus sign
indicates that A is consumed in the reaction above. It is good practice to associate the units with
all parameters in a model. For consistency in the units for rA, we find that k has units of (vol/mol
B * time). Notice that 2 mol of B react for each 1 mol of A. Then we can write
Usually, the reaction rate coefficient is a function of temperature. The most commonly used
representation is the Arrhenius rate law,

Equation 2.36

where k(T) is the reaction rate constant, as a function of temperature, k0 the frequency factor or
preexponential factor, E the activation energy (cal/gmol), R the ideal gas constant (1.987
cal/gmol K), and T the absolute temperature scale (K or R.)

The frequency factor and activation energy can be estimated based on data of the reaction
constant as a function of reaction temperature. Taking the natural logarithm of the Arrhenius rate
law, we find

Equation 2.37

and we see that k0 and E can be found from the slope and intercept of a plot of (ln k) vs. (1/T).

Equilibrium Relationships
The relationship between the liquid- and vapor-phase compositions of component i, when the
phases are in equilibrium, can be represented by

Equation 2.38

where yi is the vapor-phase mole fraction of component i, xi the liquid-phase mole fraction of
component i, and Ki the vapor/liquid equilibrium constant for component i.

The equilibrium constant is a function of composition and temperature. The simplest assumption
for the calculation of an equilibrium constant is to use Raoult's law. Here,

Equation 2.39
where the pure component vapor (saturation) pressure often has the following form:

Equation 2.40

Often, we will see a constant relative volatility assumption made, to simplify vapor/liquid
equilibrium models. In a binary system, the relationship often used between the vapor and liquid
phases for the light component is

Equation 2.41

where x is the liquid-phase mole fraction of light component, y the vapor-phase mole fraction of
light component, and a the relative volatility (a > 1).

Heat Transfer
The rate of heat transfer through a vessel wall separating two fluids (a jacketed reactor, for
example) can be described by

Equation 2.42

where Q is the rate of heat transfer from hot to cold fluid, U the overall heat transfer coefficient, A
the area for heat transfer, and DT the difference between hot and cold temperatures.

At the design stage the overall heat transfer coefficient can be estimated from correlations; it is a
function of fluid properties and velocities. The individual film heat transfer coefficients (h i and ho),
the metal conductivity (k, and thickness, Dx), and a fouling factor (f) can be used to determine
the overall heat transfer coefficient from the relationship

Equation 2.43

The individual film coefficients are a strong function of fluid properties and velocities. The overall
heat transfer coefficient is often estimated from experimental data.
Flow Through Valves
The flow through valve is often described by the relationship

Equation 2.44

where F is the volumetric flow rate, Cv the valve coefficient, x the fraction the valve is open (0

x 1), DPv the pressure drop across the valve, s.g. the specific gravity of the fluid, and f(x) the
flow characteristic (varies from 0 to 1, as a function of x).

Three common valve characteristics are (i) linear, (ii) equal-percentage, and (iii) quick-opening.
For a linear valve, f(x) = x. For an equal-percentage valve, f(x) = a x - 1. For a quick-opening
valve, . These flow characteristics are plotted in Figure 2-6.

Figure 2-6. Valve flow characteristics. The sensitivity or valve "gain" is

related to the slope of the curve.

Notice that for the quick-opening valve, the sensitivity (or "gain") of flow to valve position is high
at low openings and low at high openings; the opposite is true for an equal-percentage valve. The
sensitivity of a linear valve does not change as a function of valve position. The equal-percentage
valve is commonly used in chemical processes because of desirable characteristics when installed
in piping systems where a significant piping pressure drop occurs at high flow rates. Knowledge of
these characteristics will be important when developing feedback control systems. Flow control is
discussed in detail in Module 15.

[ Team LiB ]
[ Team LiB ]

2.5 Material and Energy Balances

Section 2.3 covered models which consist of material balances only. These are useful if thermal
effects are not important, where system properties, reaction rates, etc. do not depend on
temperature, or if the system is truly isothermal (constant temperature). Many chemical
processes have important thermal effects, so it is necessary to develop material and energy
balance models. One key is that a basis must always be selected when evaluating an intensive
property such as enthalpy.

Review of Thermodynamics
Developing correct energy balance equations is not trivial and the chemical engineering literature
contains many incorrect derivations. Chapter 5 of the book by Denn (1986) points out numerous
examples where incorrect energy balances were used to develop process models.

The total energy (TE) of a system consists of internal (U), kinetic (KE), and potential energy (PE),

Equation 2.45

where the kinetic and potential energy terms are

For most chemical processes where there are thermal effects, the kinetic and potential energy
terms can be neglected, because their contribution is generally at least an order of magnitude less
than that of the internal energy term.

When dealing with flowing systems, we usually work with enthalpy. Total enthalpy is defined as

Equation 2.46

The heat capacity is defined as the partial derivative of enthalpy with respect to temperature, at
constant pressure. The heat capacity, on a unit mass basis, is

Equation 2.47
where the overbar indicates that the enthalpy is on a unit mass basis. We make use of this
relationship in the following example.

The goal of the following example is severalfold:

develop a model consisting of both material and energy balances

illustrate the steady-state effect of the input on the output

illustrate the effect of process "size" (magnitude of flow rate, for example)

illustrate dynamic behavior

Example 2.3: Heated Mixing Tank

Consider a perfectly mixed stirred-tank heater, with a single feed stream and a single product
stream, as shown below. Assuming that the flow rate and temperature of the inlet stream can
vary, that the tank is perfectly insulated, and that the rate of heat added per unit time (Q) can
vary, develop a model (Figure 2-7) to find the tank liquid temperature as a function of time.

Figure 2-7. Stirred-tank heater.

Material Balance
Neglecting changes in density due to temperature, we find

Equation 2.48

Energy Balance

Here we neglect the kinetic and potential energy contributions,

Equation 2.49

We write the total work done on the system as a combination of the shaft work (WS) and the
energy added to the system to get the fluid into the tank and the energy that the system
performs on the surroundings to force the fluid out.

Equation 2.50

This allows us to write Equation (2.49) as

Equation 2.51

and since H = U + pV, and , we can rewrite Equation (2.51) as

Neglecting pressure*volume changes, we find

Equation 2.52
We must remember the assumptions that went into the development of Equation (2.52).

The kinetic and potential energy effects were neglected.

The change in the pV term was neglected.

The total enthalpy term is

and assuming no phase change, we select an arbitrary reference temperature (Tref) for enthalpy

Equation 2.53

Often we assume that the heat capacity is constant, or calculated at an average temperature, so

Equation 2.54a

Equation 2.54b

We now write the energy balance (2.52) in the following fashion,

Expanding the derivative term and assuming that the density is constant, we have

or from Equation (2.48)

Canceling common terms gives

Equation 2.55

but Tref is a constant, so d(T - Tref)/dt = dT/dt. Also, neglecting WS (which is significant only for
very viscous fluids), we can write

Equation 2.56

which yields the two modeling equations

Equation 2.57

Equation 2.58

In order to solve this problem, we must specify the parameters r and cp, the inputs Fi, F, Q, and
Ti (as a function of time), and the initial conditions V(0) and T(0).

Steady-State Behavior and the Effect of Scale (Size)

The steady-state solution can be found by setting the derivative terms in Equations (2.57) and
(2.58) to 0. The resulting relationship between the manipulated power and the outlet temperature
is

Equation 2.59
where the subscript s is used to indicate a steady-state value. Notice that for a given steady-state
flow rate, the relationship between heater power and outlet temperature is linear. Also, the
volume of the vessel has no effect on the steady-state temperature (volume has a solely dynamic
effect).

Here we consider a stream of water entering a stirred-tank heater at 20°C, at three possible flow
rates: 1 liter/minute (espresso machine), 10 liters/minute (household shower), and 100
liters/minute (small car wash). The outlet temperature as a function of heater power [Equation
(2.59)] is plotted in Figure 2-8 for each of the three cases. As expected from Equation (2.59), the
curves are linear. The lower flow rate operation has a high sensitivity (slope) of temperature to
power, while the higher flow rate operation has a low sensitivity. This makes physical sense,
because a given change in power will have a much larger affect on a low flow rate than a high
flow rate stream.

Figure 2-8. Outlet temperature as a function of heater power. The

slope is the sensitivity (also known as the "gain") of the output with
respect to the input.

This sensitivity is also known as the process gain and is defined as the partial derivative of the
output with respect to the input, evaluated at steady state.

Equation 2.60
It is clear from Equation (2.60) that larger flow rates will have smaller gains (slopes or
sensitivities). It is often useful to work with scaled variables. For example, if a scaled steady-state
input is defined as

Equation 2.61

then all three input flow rates have the same steady-state sensitivity of the output to the scaled
input. This is shown in Figure 2-9 and the following equation:

Equation 2.62

Figure 2-9. Outlet temperature as a function of scaled heater power.

The discussion thus far has centered on the steady-state behavior of stirred-tank heaters, and we
found that the volume had no effect. The volume has a major impact, however, on the dynamic
behavior of a stirred-tank heater.
Dynamic Behavior

Volume has a considerable effect on the dynamic behavior of this process. The response of the
temperature to a step change in the scaled heat input is shown in Figure 2-10, as a function of
the residence time (V/F). As expected, longer residence times have a slower response time than
shorter residence-time systems. These curves were obtained by integrating Equation (2.58) for
the three different residence times (with V assumed constant). The initial steady-state values are
T = Ti = 20°C and Qscaled = 0. At t = 0, Qscaled is stepped from 0° to 1°C.

Figure 2-10. Response of temperature to step change in scaled heat

addition rate.

[ Team LiB ]
[ Team LiB ]

2.6 Form of Dynamic Models

The dynamic models derived in this chapter consist of a set of first-order (only first derivatives
with respect to time), nonlinear, explicit, initial-value, ordinary differential equations. A
representation of a set of first-order differential equations is

Equation 2.63

where xi is a state variable, ui is an input variable, and pi is a parameter. The notation is used
to represent dxi/dt. Notice that there are nx equations, nx state variables, nu inputs, and np
parameters.

Also included in these models is a set of algebraic equations, relating states, inputs, and
parameters to output variables.

Equation 2.64

State Variables
A state variable is a variable that arises naturally in the accumulation term of a dynamic material
or energy balance. A state variable is a measurable (at least conceptually) quantity that indicates
the state of a system. For example, temperature is the common state variable that arises from a
dynamic energy balance. Concentration is a state variable that arises when dynamic component
balances are written.

Input Variables
An input variable is a variable that normally must be specified before a problem can be solved or
a process can be operated. Inputs are normally specified by an engineer, based on knowledge of
the process being considered. Input variables typically include flow rates of streams entering or
leaving a process (notice that the flow rate of an outlet stream might be considered an input
variable!). Compositions or temperatures of streams entering a process are also typical input
variables. Input variables are often manipulated (by process controllers) in order to achieve
desired performance.

Parameters
A parameter is typically a physical or chemical property value that must be specified or known to
mathematically solve a problem. Parameters are often fixed by nature, that is, the reaction
chemistry, molecular structure, existing vessel configuration, operation, and so forth. Examples
include density, viscosity, thermal conductivity, heat transfer coefficient, and mass-transfer
coefficient. When designing a process, a parameter might be "adjusted" to achieve some desired
performance. For example, reactor volume may be an important design parameter.

Output Variables
An output variable is often a state variable that is measured, particularly for control purposes.
Very often the measured outputs are simply a subset of the state variables. Other times the
outputs are a nonlinear function of the states (or even inputs).

Vector Notation
The set of differential and algebraic Equations (2.63) and (2.64) can be written more compactly in
vector form.

Equation 2.65

where x is the vector of state variables, u the vector of input variables, p the vector of
parameters, and y the vector of output variables.

Steady-State Solutions
Notice that dynamic models (2.63) can also be used to solve steady-state problems, since

That is,

Equation 2.66
for steady-state processes, resulting in a set of algebraic equations. In this case, all inputs and
parameters would be specified, leaving the nx state values to be solved for; that is, nx equations
in nx unknowns must be solved. The focus of this text is not on the development of numerical
methods, so we briefly cover the basic idea in Appendix 2.1. Note that differential equation
solvers can also be used to solve for the steady state of stable systems, by simply integrating
from an initial value for the states for a long period of time, until a steady state is reached.

Numerical Integration
Here we briefly consider numerical methods to integrate ordinary differential equations.

Equation 2.67

if the derivative term is approximated (where k represents a time index) as

Equation 2.68

The explicit Euler integration technique involves specifying the integration step size, Dt,

Equation 2.69

and marching sequentially from one time step to another. This approach is illustrated in Appendix
2.2. In practice, more accurate integration routines using a variable step size are used. For more
details on how to use MATLAB integration routines, see Module 3.

[ Team LiB ]
[ Team LiB ]

2.7 Linear Models and Deviation Variables

Consider the stirred-tank heater model (Example 2.3), when the volume, flow rate, and inlet
temperature are constant at their steady-state values (indicated by the subscript s):

Equation 2.70

Deviation Variable Formulation

Control engineers like to think in terms of "deviation variables," that is, perturbations from a
steady-state operating condition. The reader should show that if we define the following deviation
variables

Equation 2.71

then Equation (2.70) can be written in the form

Equation 2.72

or

Equation 2.73

where the new parameters that appear are

Notice that the process gain is the same as the sensitivity shown in Equation (2.60) and the time
constant, in this case, is the same as the residence time. Equation (2.73) is one of the most
widely used models to describe the dynamic behavior of chemical processes.

Linearization of Nonlinear Models

The material and energy balance models that describe the behavior of chemical processes are
generally nonlinear, while commonly used control strategies are based on linear systems theory.
It is important, then, to be able to linearize nonlinear models for control system design and
analysis purposes. The method that we use to form linear models is based on a Taylor series
approximation to the nonlinear model. The Taylor series approximation is based on the steady-
state operating point of the process.

One State Variable

Consider a single variable function (equation)

Equation 2.74

The value of this function can be approximated using a Taylor series expansion of the form

where the subscript s is used to indicate the point of linearization (usually the steady-state
operating point). The quadratic and higher order terms are neglected, resulting in the following
approximate equation:

Equation 2.75

Since the steady-state operating point is chosen as the point of linearization, then [by definition of
a steady state, f(xs) = 0]

Equation 2.76

and since xs is a constant value, we can write the following form

Equation 2.77

or, dropping the "approximately equal" notation

where x' = x - xs represents a deviation variable, and is the derivative of the function
evaluated at the steady-state value.

One State and One Input

Consider now the following single-state, single-input equation,

Equation 2.78

The value of this function can be approximated using a Taylor series expansion of the form

where the subscript s is used to indicate the point of linearization (usually the steady-state
operating point). The quadratic and higher order terms are neglected, resulting in the following
approximate equation

Equation 2.79
Since the steady-state operating point is chosen as the point of linearization, then [by definition of
a steady state, f(xs, us) = 0]

where x' = x - xs represents a deviation variable, and and are the

derivatives of the function with respect to the state and input, evaluated at the steady-state
value.

Output Variable

Consider now the expression for an output variable

Equation 2.80

A Taylor series expansion about the state and input yields (after neglecting higher order
derivatives)

Equation 2.81

and since ys = g(xs,us)

where y' = y - ys, x' = x - xs, and u' = u - us represent deviation variables, and and

are the derivatives of the function with respect to the state and input, evaluated at
the steady-state value.

These basic ideas are illustrated in the following example.

Example 2.4: A Second-Order Reaction
Consider a CSTR with a single, second-order reaction. The modeling equation, assuming constant
volume and density is

Here the state variable is CA and the input variable is F. A Taylor series expansion performed at
the steady-state solution yields

Now, consider the concentration of A to be the output variable

so

and the state space model is

where the state, input, and output (in deviation variable form) are

For the following parameters,

A steady-state operating point is

and the partial derivatives are

and the linear model is

Generalization

Consider the general nonlinear model with nx states, ny outputs, nu inputs, and np parameters

The elements of the linearization matrices are defined as

where ij subscripts refer to the ith row and jth column of the corresponding matrix. For example,
element Bij refers to the effect of the jth input on the ith state derivative.

The linear state space form is

where the deviation variables are defined as perturbations from their steady-state values

In future chapters we normally drop the prime (') notation for deviation variables and assume
that a state space model is always in deviation variable form.

Example 2.5: Jacketed Heater

Consider the jacketed stirred-tank heater shown in Figure 2-11. A hot fluid circulated through the
jacket (which is assumed to be perfectly mixed), and heat flow between the jacket and vessel
increases the energy content of the vessel fluid. The rate of heat transfer from the jacket fluid to
the vessel fluid is

Figure 2-11. Jacketed stirred-tank heater.

where U is the overall heat transfer coefficient and A is the area for heat transfer. Assuming that
the volume and density are constant, F i = F. Energy balances on the vessel and jacket fluids
result in the following equations.

Here the outputs are the vessel and jacket temperatures, which are also the states; the inputs
are the jacket flow rate, feed flow rate, feed temperature, and jacket inlet temperature. If the
outputs, states, and inputs, in deviation variable form, are

Then, the linearized model is

Similarly, the reader should show that

Exercise 8 is a numerical example for this problem.

[ Team LiB ]
[ Team LiB ]

2.8 Summary
We have focused primarily on the development of ordinary differential equation models that
describe the dynamic behavior of processes where perfect mixing can be assumed. The models
have the form

where the states, inputs, outputs, and parameters are x, u, y, and p, respectively.

States appear in the "accumulation" (derivative with respect to time) term, inputs can be either
manipulated or disturbance variables, outputs are often a subset of the states, and parameters
are often physical intensive variables, such as density or kinetic rate parameters. The modeling
equations are based on material (component or total) and energy balances. States are often
concentrations (from the component balances), volume (total material balance on liquids),
pressure (total material balances on gases), or temperature (energy balances).

The steady-state solution is , or f(x,u,p) = 0 which can be found numerically, using M ATLAB
functions such as fzero and fsolve (which require the optimization toolbox) as shown in the
MATLAB tutorial module.

To integrate the differential equations numerically, we use a variant of Euler integration, which is

The ODE module provides a tutorial on the use of MATLAB to integrate ordinary differential
equations.

An understanding of dynamic behavior is obtained by using linear state-space models, which have
the form

where the prime notation (') is used to indicate a deviation variable. In the next chapter we
further analyze linear state space models to obtain an understanding of dynamic behavior. These
state space models are converted to Laplace transfer function models, which are used later for
control system design.

For the processes studied, some important characteristics were discussed. For example, residence
time (volume/flow rate) is often a good indicator of the relative "speed" of the process dynamics,
particularly for nonreacting systems. For chemical reactors, the Damkohler number is an
important parameter, since it is the ratio of a characteristic reaction time to residence time.

The concept of process gain is extremely important for process control design. Process gain is the
sensitivity of the process output to a manipulated input. That is, process gain is the ratio of the
long-term (steady-state) change in process output to the change in process input.

Many process systems can be represented as single-state models, where the output is the state.
The resulting first-order model has the form

where kp is the process gain and tp is the process time constant. The dynamic behavior of first-
order processes will be discussed in the next chapter.

The examples used in this chapter were

2.1 Gas Surge Drum

2.2 An Isothermal Chemical Reactor

2.3 Heated Mixing Tank

2.4 A Second-Order Reaction

2.5 Jacketed Heater

[ Team LiB ]
[ Team LiB ]

Suggested Reading

A nice introduction to chemical engineering calculations is provided by Felder, R. M., and R.

Rousseau, Elementary Principles of Chemical Processes, 2nd ed. Wiley, New York (1986).

Excellent discussions of the issues involved in modeling a mixing tank, incorporating density
effects and energy balances, is provided in the following two books: Denn, M. M., Process
Modeling, Longman, New York (1986); Russell, T.R. F., and M. M. Denn, Introduction to Chemical
Engineering Analysis, Wiley, New York (1971).

An introduction to chemical reaction engineering is Fogler, H. S., Elements of Chemical Reaction

Engineering, 3rd ed., Prentice Hall, Englewood Cliffs, NJ (1999).

An excellent textbook for an introduction to chemical engineering thermodynamics is Smith, J. M.,

H. C. Van Ness, and M. M. Abbott, Chemical Engineering Thermodynamics, 6th ed., McGraw-Hill,
New York (2001).

A more detailed discussion of process dynamics is provided in the textbook Bequette, B.W.,
Process Dynamics: Modeling, Analysis and Simulation, Prentice Hall, Upper Saddle River, NJ
(1998).

The reaction of ethylene oxide and water to form ethylene glycol is discussed in Fogler, H. S.,
Elements of Chemical Reaction Engineering, 2nd ed., Prentice Hall, Upper Saddle River, NJ
(1992).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider the gas drum in Example 2.1. Often the outlet flow relationship is actually
nonlinear, with the form

so the modeling equation is

Discuss whether this is now a self-regulating process. Also, sketch the steady-state
input (qis) -output (Ps) curve, based on the flow coefficient, b = 1 mol · s -1 · atm-1/2 and
a constant header pressure, Ph = 1 atm.

2: Consider the heated mixing tank example, which had the modeling equations

For steady-state inlet and outlet flow rates of 100 liters/minute, a liquid volume of 500
liters and inlet and outlet temperatures of 20° and 40°C, respectively:

a. Find the steady-state heating rate, Q.

b. Consider a step inlet temperature change from 20° to 22°C. Use Euler integration
with a integration step size of 0.5 minutes to find the vessel temperature response
for the first 2 minutes. Compare this with MATLAB ode45 (see the ODE module).

3: Often liquid surge tanks (particularly those containing hydrocarbons) will have a gas
"blanket" of nitrogen or carbon dioxide to prevent the accumulation of explosive vapors
above the liquid, as depicted below.
 Develop the modeling equations with gas pressure and liquid volume as the state
variables. Let qf and q represent the inlet and outlet gas molar flow rates, Ff and F the
liquid volumetric flow rates, V the constant (total) volume, Vl the liquid volume, and P
the gas pressure. Assume the ideal gas law. Show that the modeling equations are

and state any other assumptions.

4: Consider the ethylene glycol problem (Example 2.2). If two 800-gallon reactors are
placed in series, what volumetric flow rate is necessary to produce 100 million lb/year of
ethylene glycol? What percentage savings is this compared with using a single 800-
gallon reactor?

5: Consider the ethylene glycol problem (Example 2.2). Solve the dynamic equations using
ode45 (see the ODE module), for a step change in space velocity from 0.0778 to 0.0817
min-1. Compare your plots with those shown in Figure 2-4.

6: Semibatch reactors are operated as a cross between batch and continuous reactors. A
semibatch reactor is initially charged with a volume of material, and a continous feed of
reactant is started. There is, however, no outlet stream. Develop the modeling
equations for a single first-order reaction. The state variables should be volume and
concentration of reactant A.

7: A stream contains a waste chemical, W, with a concentration of 1 mol/liter. To meet

Environmental Protection Agency and state standards, at least 90% of the chemical
must be removed by reaction. The chemical decomposes by a second-order reaction
with a rate constant of 1.5 liter/(mol hr). The stream flow rate is 100 liter/hour and two
available reactors (400 and 2000 liters) have been placed in series (the smaller reactor
is placed before the larger one).

a. Write the modeling equations for the concentration of the waste chemical. Assume
constant volume and constant density. Let

Cw1 = concentration in reactor 1, mol/liter

 Cw2 = concentration in reactor 2, mol/liter

F = volumetric flow rate, liter/hr

V1 = liquid volume in reactor 1, liters

V2 = liquid volume in reactor 2, liters

k = second-order rate constant, liter/(mol hr)

b. Show that the steady-state concentrations are 0.33333 mol/liter (reactor 1) and
0.09005 mol/liter (reactor 2), so the specification is met.

(Hint: You need to solve quadratic equations to obtain the concentrations.)

c. Linearize at steady state and develop the state space model (analytical) of the
form

where

d. Show that the A and B matrices are

(also, show the units associated with each coefficient).

e. Assuming that each state is an output, show that the C and D matrices are

f. Find the eigenvalues of A using the MATLAB eig function, and find the eigenvalues
by hand, by solving det (lI - A) = 0.

g. Solve a for the nonlinear differential equations, using ode45. Compare the linear
and nonlinear variables on the same plots (make certain you convert from
deviation to physical variables for the linear results) for a step change in the flow
rate from 100 liters/hour to 110 liters/hour. Assume the initial concentrations are
the steady-state values (0.3333 and 0.09005). Compare the linear and nonlinear
responses of the reactor concentrations. Is the removal specification still obtained?

h. Would better steady-state removal of W be obtained if the order of the reaction

vessels was reversed? Why or why not? (Show your calculations.)

8: Consider Example 2.5, the stirred-tank heater example. Read and work through the
example. Use the following parameters and steady-state values:

Fs = 1 ft3/min rcp = 61.3 Btu/(°F · ft3) rjcpj = 61.3 Btu/(°F · ft3)

Tis = 50°F Ts = 125°F V = 10 ft 3

Tjis = 200°F Tjs = 150°F Vj = 2.5 ft3

a. By solving the steady-state equations, verify that the following values obtained for
UA (overall heat transfer coefficient * area for heat transfer) and F js (steady-state
jacket flow rate) are correct:

b. Find the values of the matrices in the state space model.

c. Write a function file, heater.m (example shown below), to be used with ode45
(Module 3) to solve the two nonlinear ordinary differential equations.

d. First, verify that the steady-state state variable values are correct by simulating
the process with no change in the jacket flow rate.

e. Now, perform simulations for small and large step changes in the jacket flow rate.
Use the MATLAB step command to solve for the linear state space model. Realize
that the step results are based on deviation variables and for a unit step change in
input one (jacket flow rate), and convert the linear states to physical variable
form.

We now consider the important issue of scale-up: Pilot plants are small-scale
(intermediate between lab-scale and full-size manufacturing-scale) chemical
processes, used to understand process operating behavior before the
manufacturing process is designed. Here we discover the effect of vessel scale on
the heat transfer removal capability of a vessel.

f. Consider now a vessel that can handle 10 times the throughput of the previous
vessel (that is, 10 ft3/min rather than 1 ft3/min). Assume that the same residence
time is maintained (V/F = 10 minutes), so the volume of the new vessel is 100 ft3.
Assume that the heat transfer coefficient (U) remains constant, but that the heat
transfer area changes. Assume that the vessel can be modeled as a cylinder, with
the height (L) = 2*diameter (D). Find the value of UA for the larger vessel.

g. For the larger vessel, find the new steady-state value of jacket temperature that
must be used to maintain the vessel temperature at T = 125°F. Also, find the new
steady-state jacket flow rate. Hint: Solve the two modeling equations at steady
state to obtain these values. Discuss the effect of process scale-up on the
operating conditions. How large can the vessel become before the jacket
temperature is too high (it approaches the inlet jacket temperature)?

h. For the larger vessel, find the new state space model, assuming that the jacket
volume is 0.25 times the vessel volume. Calculate the eigenvalues of the A matrix.
How do they compare with the smaller vessel?

i. Find the step response for the nonlinear and linear (state space) systems for a
step increase of 0.1 ft3/min in jacket flow rate. How do these compare with the
smaller vessel?

j.
 i.

j. Find the step response for the nonlinear and linear (state space) systems for a
step increase of 10% in the jacket flow rate. How do these compare with the
smaller vessel?

Example m-file to be used by ode45

function xdot = heater(t,x,flag,delFj);

%
% Dynamics of a stirred tank heater
% (c) 1994 - B.W. Bequette
% 8 July 94
% 23 Jan 01 - revised for jacket step change in argument list
%
% x(1) = T = temperature in tank
% x(2) = Tj = temperature in jacket
% delFj = change in jacket flowrate
% F = Tank flowrate
% Tin = Tank inlet temperature
% Tji = Jacket inlet temperature
% V = Tank volume
% Vj = Jacket volume
% rhocp = density*heat capacity
% rhocpj = density*heat capacity,jacket fluid
%
% parameter and steady-state variable values are:
%
F = 1;
Fjs = 1.5;
Ti = 50;
Tji = 200;
V = 10;
Vj = 2.5;
rhocp = 61.3;
rhocpj= 61.3;
UA = 183.9;
%
% step change in jacket flow
%
Fj = Fjs + delFj;
T = x(1);
Tj= x(2);
%
% odes
%
dTdt = (F/V)*(Ti - T) + UA*(Tj - T)/(V*rhocp);
dTjdt = (Fj/Vj)*(Tji - Tj) - UA*(Tj - T)/(Vj*rhocpj);
xdot = [dTdt; dTjdt];
%
% generate simulations using
% [t,x] = ode45('heater',tspan,x0,[],delFj);
% where tspan = [0 30], for example
To make the runs go faster, you may wish to generate and run the following script file:

%
% runs stirred tank heater example - 23 Jan 01
% step changes to stirred tank heater
%
% make certain you enter the step change value (delFj)
% before running this file
% also, generate function file with the following first line:
% function xdot = heater(t,x,flag,delFj);
%
options = odeset('RelTol',1e-6,'AbsTol',[1e-6 1e-6]);
%
[t,x] = ode45('heater',tspan,x0,options,delFj);
%
figure(1)
subplot(2,1,1),plot(t,x(:,1))
title('nonlinear, using ode45')
xlabel('time, min')
ylabel('temp, deg F')
subplot(2,1,2),plot(t,x(:,2))
xlabel('time, min')
ylabel('jacket temp, deg F')
%
% state space model with unit step change
a = [-0.4 0.3;1.2 -1.8];
b = [0;20];
c = [1 0;0 1];
d = 0;
sys = ss(a,b,c,d);
[ylin,tlin] = step(sys);
%
figure(2)
subplot(2,1,1),plot(tlin,ylin(:,1))
title('linear state space, unit step, deviation')
xlabel('time, min')
subplot(2,1,2),plot(tlin,ylin(:,2))
xlabel('time, min')
%
% scale deviation by delFj
ylinscale = ylin*delFj;
% plot in physical variable form
figure(3)
subplot(2,1,1),plot(tlin,125+ylinscale(:,1))
title('linear state space, physical magnitude')
xlabel('time, min')
subplot(2,1,2),plot(tlin,150+ylinscale(:,2))
xlabel('time, min')
% compare linear and nonlinear on same plot
figure(4)
subplot(2,1,1),plot(t,x(:,1),tlin,125+ylinscale(:,1),'--')
 title('nonlinear vs. linear')
xlabel('time, min')
ylabel('temp, deg F')
legend('nonlinear','linear')
subplot(2,1,2),plot(t,x(:,2),tlin,150+ylinscale(:,2),'--')
xlabel('time, min')
ylabel('jacket temp, deg F')
legend('nonlinear','linear')

[ Team LiB ]
[ Team LiB ]

Appendix 2.1: Solving Algebraic Equations

Fortunately, the MATLAB fsolve function is easy to use for solving algebraic equations. For a
simplified presentation, we use the form

Equation A.1

obtained from Equation (2.66) with a fixed p and u. The most commonly used numerical
techniques are related to Newton-Raphson iteration. The "guess" for iteration k + 1 is determined
from the value at iteration k, using

Equation A.2

where f[x(k)] is the vector of function evaluations at iteration k, and J(k) is the Jacobian matrix

Equation A.3

The ij element of the Jacobian represents the partial derivative of equation i with respect to
variable j. If analytical derivatives are not available, elements of the Jacobian are obtained by
perturbation of the state variable, requiring n + 1 function evaluations for an n-equation system
of equations. Various quasi-Newton techniques provide approximations to the Jacobian and do not
require as many function evaluations, reducing computational time.

In practice, the Jacobian matrix in Equation (A.2) is not inverted. Rather, a set of linear algebraic
equations is solved for x(k+1),

Equation A.4

In this text we do not focus on the solution of algebraic equations. See the text by Bequette
(1998) for more details on these techniques.

[ Team LiB ]
[ Team LiB ]

Appendix 2.2: Integrating Ordinary Differential

Equations
The Euler integration method often requires very small integration step sizes to obtain a desired
level of accuracy. Note that x is a vector of n state variables at each time step. For example, if
there are two states (two differential equations)

we have left the inputs and parameters out of the function variable list for convenience. An
example is shown next.

Example 2.2: An Isothermal Chemical Reactor, continued

Consider the ethylene glycol reactor problem in Example 2.2. We use k1 to represent the reaction
rate constant, so that it is not confused with the time-step index. The Euler integration algorithm
results in the following two equations, where k represents the time-step index.

For C A(0) = 0.1 and CP(0) = 0.4, and a space velocity (F/V) = 0.0816 min-1 we find, for the first
time step,

Substituting the numerical values for an integration step size of 0.5 minutes, we find
resulting in the concentration values at t = 0.5 minutes

Values can be obtained at future times by continuing to march forward in time. MATLAB has a suite
of routines for integrating differential equations. These are covered in Module 3.

[ Team LiB ]
[ Team LiB ]

Chapter 3. Dynamic Behavior

The goal of this chapter is to understand dynamic behavior. We begin by working with linear state
space models, often obtained by linearizing a nonlinear model, such as those developed in
Chapter 2. We then introduce Laplace transforms. The main advantage to Laplace transforms is
that they allow us to analyze behavior exhibited by linear differential equations by using simple
algebraic manipulations. Laplace transforms are used to create transfer function models, which
are the basis for many control system design techniques.

After studying this chapter, the reader should be able to:

Apply the initial and final value theorems of Laplace transforms

Understand first-order, first-order + dead time and integrating system step responses

Understand second-order under-damped behavior

Understand the effect of pole and zero values on step responses

Convert state space models to transfer functions

The major sections of this chapter are as follows:

3.1 Background

3.2 Linear State Space Models

3.3 Introduction to Laplace Transforms

3.4 Transfer Functions

3.5 First-Order Behavior

3.6 Integrating System

3.7 Second-Order Behavior

3.8 Lead-Lag Behavior

3.9 Poles and Zeros

3.10 Processes with Dead Time

3.11 Padé Approximation for Dead Time

3.12 Converting State Space Models to Transfer Functions

3.13 MATLAB and SIMULINK

3.14 Summary
[ Team LiB ]
[ Team LiB ]

3.1 Background
One of the major goals of this chapter is to obtain an understanding of process dynamics. Process
engineers tend to think of process dynamics in terms of the response of a process to a step input
change. Assume that the process is initially at steady state, then apply a step change to an input
variable. The majority of chemical processes will exhibit one of the responses shown in Figure 3-
1. In this plot, we assume that a positive step increase has been made to the input variable of
interest. The solid curves are examples of "positive gain" processes; that is, the process output
increases for an increase in the input. The dashed lines are those of negative gain processes. The
curves in Figure 3-1a show a monotonic change in the output; this behavior is generally known as
overdamped. The curves in Figure 3-1b are indicative of "integrating" processes; a prime example
is a liquid surge vessel, where the level continues to change when there is an imbalance in the
inlet and outlet flow rates. The curves in Figure 3-1c are known as underdamped or oscillatory
responses. This type of behavior may occur in exothermic chemical reactors or biochemical
reactors. It more often occurs in processes that are under feedback control, particularly if the
controller is poorly tuned. The behavior shown in Figure 3-1d is known as "inverse response" and
is seen in steam drums, distillation columns, and some adiabatic plug flow reactors.

Figure 3-1. Common responses of process outputs to step changes in

process inputs. Assuming a positive step change, the solid curves are
illustrative of "positive gain" processes, and the dashed curves are
indicative of "negative gain" processes. (a) Overdamped or first order,
(b) integrating, (c) underdamped or oscillatory, and (d) inverse
response.

In the sections that follow, we discuss the characteristics of process models that lead to each of
the behaviors shown in Figure 3-1.
[ Team LiB ]
[ Team LiB ]

3.2 Linear State Space Models

In Chapter 2 we developed fundamental models, which were normally nonlinear in nature. We
then developed state space models that were based on linearizing the fundamental models at a
steady-state solution. This led to the notion of a perturbation or deviation variable, which is
simply the perturbation of a variable from its steady-state value.

State space models have the following form, where the states (x), inputs (u), and outputs (y) are
all perturbation or deviation variables

Equation 3.1

Recall that in matrix notation, the first subscript refers to the row and the second subscript refers
to the column. When matrices multiply vectors, each row corresponds to a particular output of the
multiplication, while the column corresponds to a particular input of the operation. Consider the C
matrix, which relates the states to the outputs. Element cij relates the effect of state xj on output
yi.

The shorthand notation for Equation (3.1) is

Equation 3.2

It is important to always check for dimensional consistency in matrix operations. In a matrix-

vector operation y = Cx, the number of rows in C must be equal to the number of elements in y.
Also, the number of columns in C must be equal to the number of elements in x.

Stability
One of the first basic concepts that we need to cover is the notion of stability. Consider a process
where one or more states have been perturbed from the steady-state solution or operating point.
The process is stable if after a period of time, the variables return to the steady-state values. This
means that the state variables, since they are deviation variables, return to zero.

Numerically, we can determine the stability of a state space model by finding the eigenvalues of
the state space A matrix. Remember that the A matrix is simply the matrix of derivatives of the
dynamic modeling equations with respect to the state variables.

If all of the eigenvalues are negative, then the system is stable; if any single eigenvalue is
positive, the system is unstable. A system with all but one eigenvalues negative and with one
eigenvalue equal to zero is called an integrating system and is characteristic of processes with
liquid levels or gas drum pressures that can vary.

Examples of unstable systems are shown in Figure 3-2. If an eigenvalue is real and positive, the
system response is that shown in the top curves. If there are complex conjugate eigenvalues,
with positive real portions, the system oscillates (with ever increasing amplitude), as shown at the
bottom.

Figure 3-2. Unstable responses. (a) Monotonic and (b) oscillatory.

Mathematically, the eigenvalues of the A matrix are found from the roots of the characteristic
polynomial

Equation 3.3

where l is known as an eigenvalue, and I is the identity matrix. For a state space model with n
states, A is an n x n matrix, and there will be n solutions (eigenvalues) of Equation (3.3). There
are analytical solutions for two- and three-state systems; the two-state solution is shown below.

In two-state systems,

The determinant can be found by

and the eigenvalues are found as the two solutions (roots) to

Equation 3.4

The roots can be found using the quadratic formula

Equation 3.5

It is easy to show that if and a11 + a22 < 0 and a11a22 >a12a21, the roots (eigenvalues) are
negative and the system is stable. A more general method of qualitatively checking the stability,
known as the Routh stability criterion, is shown in Chapter 5.

Example 3.1: Exothermic CSTR

Models for an exothermic, CSTR are detailed in Module 8. For a two-state representation, the first
state is the concentration and the second state is the reactor temperature. For a particular
reactor with two different operating conditions, the A matrix is (the time unit is hours)

Operating condition 1 Operating condition 2

and the eigenvalues for operating condition 1 can be found using the following steps

with the solutions [using the quadratic formula (3.5)]

Since both eigenvalues are negative, operating condition 1 is stable.

The reader should show that the eigenvalues of A2 are

where the positive eigenvalue indicates that operating condition 2 is unstable.

MATLAB Eigenvalue Function

The M ATLAB eig command can be used to quickly find eigenvalues of a matrix. The reader should
use the MATLAB command window to verify the following results for the second operating
condition:

» a2 = [-1.8124 -0.2324;9.6837 1.4697];

» eig(a2)
ans =
-0.8366
0.4939

Again, the positive eigenvalue indicates that the second operating condition is unstable.

Generalization
Notice that a solution of a second-order polynomial was required to find the eigenvalues of the
two-state example; this resulted in two eigenvalues. For the general case of an n x n matrix,
there will be n eigenvalues. It is too complex to find these analytically for all but the simplest low-
order systems. The simplest way to find eigenvalues is by using existing numerical analysis
software; for example, in MATLAB the eig function can be used to find eigenvalues.

The values of the eigenvalues are related to the "speed of response," and the eigenvalue unit is
inverse time. If the unit of time used in the differential equations is minutes, for example, then
the eigenvalues have min-1 as the unit. For stable systems (where all eigenvalues are negative),
the larger magnitude (more negative) eigenvalues are faster.

For matrices that are 2 x 2 or larger, some eigenvalues may occur in complex conjugate pairs. In
this case, the stability is determined by the sign of the real portion of the complex number. As
long as all real portions are negative, the system is stable.

[ Team LiB ]
[ Team LiB ]

3.3 Introduction to Laplace Transforms

Most control system analysis and design techniques are based on linear systems theory. Although
we could develop these procedures using the state space models, it is generally easier to work
with transfer functions. Basically, transfer functions allow us to make algebraic manipulations
rather than working directly with linear differential equations (state space models). To create
transfer functions, we need the notion of the Laplace transform.

The Laplace transform of a time-domain function, f(t), is represented by L[f(t)] and is defined as

Equation 3.6

The Laplace transform is a linear operation, so the Laplace transform of a constant (C) multiplying
a time-domain function is just that constant times the Laplace transform of the function,

Equation 3.7

The Laplace transforms of a few common time-domain functions are shown next.

Exponential Function
Exponential functions appear often in the solution of linear differential equations. Here

The transform is defined for t > 0 (we also use the identity that ex+y = exey)

So we now have the following relationship:

Equation 3.8
Derivatives
This will be important in transforming the derivative term in a dynamic equation to the Laplace
domain (using integration by parts),

so we can write

Equation 3.9

For an nth derivative, we can derive

Equation 3.10

n initial conditions are needed: f(0),..., f (n – 1) (0)

One reason for using deviation variables is that all of the initial condition terms in Equation (3.10)
are 0, if the system is initially at steady-state.

Time Delays (Dead Time)

Time delays often occur owing to fluid transport through pipes, or measurement sample delays.
Here we use q to represent the time delay. If f(t) represents a particular function of time, then f(t
– q) represents the value of the function q time units in the past.
We can use a change of variables, t* = t – q, to integrate the function. Notice that the lower limit
of integration does not change, because the function is defined as f(t) = 0 for t < 0.

Equation 3.11

So the Laplace transform of a function with a time delay (q) is simply e–qs times the Laplace
transform of the nondelayed function.

Step Functions
Step functions are used to simulate the sudden change in an input variable (say a flow rate being
rapidly changed from one value to another). A step function is discontinuous at t = 0. A "unit"
step function is defined as

and using the definition of the Laplace transform,

so

Equation 3.12

Similarly, the Laplace transform of a constant, C, is

Equation 3.13

Pulse
Consider a pulse function, where a total integrated input of magnitude P is applied over tp time
units, as shown in Figure 3-3.

Figure 3-3. Pulse function.

The function is f(t) = P/t p for 0 < t < tp and f(t) = 0 for t > t p. The Laplace transform is

Equation 3.14

Impulse
An impulse function can viewed as a pulse function, where the pulse period is decreased while
maintaining the pulse area, as shown in Figure 3-4. In the limit, as tp approaches 0, the pulse
function becomes (using L'Hopital's rule)

Equation 3.15

Figure 3-4. Concept of an impulse function.

If we denote a unit impulse as f(t) = d, then the Laplace transform is

Equation 3.16

Examples of common impulse inputs include a "bolus" (shot or injection) of a drug into a
physiological system, or dumping a bucket of fluid or bag of solids into a chemical reactor.

Other Functions
It is rare for one to derive the Laplace transform for a function; rather a table of known
transforms (and inverse transforms) can be used to solve most dynamic systems problems.

Table 3-1 presents solutions for most common functions. If you desire to transform a function
from the time domain to the Laplace domain, then look for the time-domain function in the first
column and write the corresponding Laplace domain function from the second column. Similarly, if
your goal is to "invert" a Laplace domain function to the time domain, then look for the Laplace
domain function in the second column and write the corresponding time-domain function from the
first column. This notion of the inverse Laplace transform can be written

Equation 3.17

For example
Initial- and Final-Value Theorems
The following theorems are very useful for determining limiting values in dynamics and control
studies. The long-term behavior of a time-domain function can be found by analyzing the Laplace
domain behavior in the limit as the s variable approaches zero. The initial value of a time-domain
function can be found by analyzing the Laplace domain behavior in the limit as s approaches
infinity.

Table 3-1. Laplace Transforms for Common Time-Domain Functions

Time-domain function Laplace domain function

f(t) F(s)

d(t) (unit impulse) 1

C (constant)

f(t - q) e-qsF(s)

tn

sF(S) - f(0)

Snf(S) - Sn-1f(0) - L - sf (n-2)(0) - f(n-1)(0)

e-at

te-at

1-e-t/t

sin wt
 Time-domain function Laplace domain function

cos wt

e–at sin wt

e–atcos wt

where ,

The final-value theorem is

Equation 3.18a

The initial-value theorem is

Equation 3.18b

It should be noted that these theorems only hold for stable systems.

Example 3.2: Application of Initial- and Final-Value Theorems

Find the long-term and short-term behavior of the time-domain function, y(t), using the final- and
initial-value theorems on the Laplace domain function Y(s) (we see later that this arises from a
step input applied to a second-order process):
The long-term behavior, y(t ), is found using the final-value theorem,

The short-term behavior, y(t 0), is found using the initial-value theorem,

The reader should verify that the time-domain function, y(t), can be found by applying Table 3-1
to find

and that the values of y(t ), and y(t 0), are consistent with the final- and initial-value
theorems.

General Solution Procedure

To obtain analytical solutions for differential equation-based models, the general procedure is
composed of several steps.

1. Start with an nth order linear differential equation

and known initial conditions

2. Transform each element of the differential equation to the Laplace domain,

3.
 3. Use algebraic manipulations to solve for the transformed variable.

where N(s) and D(s) are polynomials in s.

4. Perform a partial fraction expansion to isolate the individual elements.

where a, b, and so forth are the roots of D(s). If roots are repeated the partial fraction
expansion has the form (for an example where three roots all have a value of a).

5. Invert each element back to the time domain to find the final solution for y(t).

For the case of repeated roots (say 3 that have a value of a), the solution has the form

It should be noted that chemical process systems are rarely described by an nth order differential
equation. Usually, a set of n first-order differential equations is transformed to a single nth order
equation, as shown in Example 3.3.

Example 3.3: Second-Order Differential Equation

Consider the following state space model of an isothermal CSTR (Module 5).

Equation 3.19
where x1 and x2 represent the concentrations of two components (in deviation variable form) in
an isothermal reactor; the initial conditions are x1(0) = x2(0) = 0. Solve for the output
(concentration of component 2) response to a unit step input.

1. The reader should show (see exercise 6) that this can be arranged to a second-order
differential equation,

with the initial conditions y(0) = dy(0)/dt = 0, and the input is initially u(0)= 0.

2. Taking the Laplace transform of each element,

and substituting back into the differential equation, we find[1]

[1]
Alternatively, one could take the Laplace transform of each equation in (3.19) and use algebraic
substitution to find this second-order equation in s.

3. Solving for Y(s), we find

Here we consider a unit step input, U(s) = 1/s.

4. Performing a partial fraction expansion using the roots of the denominator polynomial,

Equation 3.20
 Solve for C1 by multiplying by s,

Setting s = 0 to find

Similarly, solve for C2 by multiplying Equation (3.20) by s + 2.4053 and setting s = –2.4053

Also, solve for C3 by multiplying Equation (3.20) by s + 2.2376 and setting s = –2.2376

which yields

5. Inverting each element back to the time domain,

Equation 3.21

Although it is nice to have an analytical solution, it is generally more pleasing to engineers to

view a plot of a variable as it changes with time, as shown in Figure 3-5. In this plot we
notice that the concentration initially decreases, before increasing to a new steady-state
value.

Figure 3-5. Response of the concentration as a function of time.

See Equation (3.21).
In Example 3.3 we solved for the coefficients (C1, C2, C3) by selecting values of s to minimize the
multiplications performed during each step. An alternative is to solve for three equations in three
unknowns by using various values of s (say s = 1, 2, and 3, for example) in Equation (3.20).

Differential equations textbooks, such as Boyce and DiPrima (1992) present many examples of
applications of Laplace transforms to solve differential equations. In practice, it is rare for process
engineers to seek analytical solutions to differential equations; it is far easier to solve these
numerically. The primary goal of this section was to provide background material to understand
the transfer functions and dynamic responses shown in the next section.

[ Team LiB ]
[ Team LiB ]

3.4 Transfer Functions

A transfer function relates inputs to outputs in the Laplace domain. In particular, the Laplace
domain relationship between a manipulated input and output is called a process transfer function.

Consider an nth order differential equation,

Equation 3.22a

Since we are assuming that the model is based on deviation variables, and that the system is
initially at steady state, the initial conditions are

Equation 3.22b

Taking the Laplace transform of each term,

Solving for Y(s), we find

The ratio of polynomials is called the transfer function. When it relates a manipulated input to an
output it is commonly called a process transfer function. In general, we will use gp(s) to represent
the process transfer function.
Equation 3.23

In this case the process transfer function is

Equation 3.24

The roots of the numerator polynomial are known as zeros, and the roots of the denominator
polynomial are call poles. The significance of poles and zeros are discussed in detail in Section
3.9.

We have used capital letters to distinguish Laplace domain variables from the time-domain
variables. In the rest of this text we generally use lowercase letters for all input and output
variables. If the argument is s, then we assume that we are referring to the Laplace domain.

Transfer functions are often used in block diagrams. For example, the relationship between an
input and output is shown as

In the rest of this chapter we study the dynamic behavior of some commonly used transfer
functions. Our focus is on Step Responses, since process engineers often apply step changes to
understand dynamic behavior.

[ Team LiB ]
[ Team LiB ]

3.5 First-Order Behavior

Many chemical processes can be modeled as first-order systems. The differential equation for a
linear first-order process is often written in the following form:

Equation 3.25

This can also be written as

where the parameters (tp and kp) and variables (y and u) have the following names: tp is the
process time constant (units of time), kp the process gain (units of output/input), y the output
variable, and u the input variable. Taking the Laplace transform of each term (notice that we are
now using lower-case variables to represent the Laplace domain input and output), and assuming
that the initial condition is y(0) = 0,

So the Laplace transform of Equation (3.25) can be written

or solving for y(s) we find a first-order transfer function

Equation 3.26

Step Response
Consider the case where the output is initially zero (steady state in deviation variable form), and
the input is suddenly step changed by an amount Du. The Laplace transform of the input is

Equation 3.27

So Equation (3.26) can be written

Equation 3.28

Using a partial fraction expansion and inverting to the time domain, you should find (see Exercise
1)

Equation 3.29

Here the notion of a process gain is clear. After a substantial amount of time (t >> tp), we find,
from Equation (3.29),

Equation 3.30

That is,

and, since y(0) = 0, we can think of y(t ) as Dy, so

Equation 3.31

We can think of the process time constant as the amount of time it takes for 63.2% of the
ultimate output change to occur, since when t = tp,
Remember that this holds true only for first-order systems.

Impulse Response
Consider now an impulse input of magnitude P, which has units of the input*time; if the input is a
volumetric flow rate (volume/time), then the impulse input is a volume. The output response is

You should find that the time domain solution is

which has an immediate response of Pkp/tp followed by a first-order decay with time.

Example 3.4: Stirred-Tank Heater

Recall that an energy balance on a constant-volume stirred-tank heater (Example 2.3) yielded

where the subscript s is used to indicate that a particular variable remains at its steady-state
value. Defining the following deviation variables

The equation can be written in the form

or

where the parameters of this first-order model are

The gain and time constant are clearly a function of scale. A process with a large steady-state
flow rate will have a low gain, compared to a process with a small steady-state flow rate. This
makes physical sense, since a given heat power input will have a larger effect on the small
process than the large process. Similarly, a process with a large volume-to-flow rate ratio is
expected to have a slow response compared to a process with a small volume-to-flow rate ratio.

Consider a heater with a constant liquid volume of Vs = 50 liters and a constant volumetric flow
rate of Fs = 10 liters/minute. For liquid water, the other parameters are rcp = 1 kcal/ liter°C. The
process gain and time constant are then

Step Response

If a step input change of 10 kW is made, the resulting output change is [from Equation (3.29)]

A plot of the step input and the resulting output are shown in Figure 3-6.

Figure 3-6. Step response of a stirred-tank heater, characteristic of a

first-order system. Deviation variables.
Remember that the inputs and outputs in this expression are in deviation variable form. If the
steady-state values are (for an inlet temperature of 20°C)

Then, the physical temperature response is

Impulse Response

If an impulse input of 30 kJ is made, you should be able to show that the temperature changes
immediately by 0.143°C (see Exercise 15).

[ Team LiB ]
[ Team LiB ]

3.6 Integrating System

We found in the previous chapter that material balances on liquid surge vessels or gas drums
often yielded models with the following form:

Equation 3.32

In the Laplace domain, this is

Equation 3.33

Consider an integrating process initially at steady state, with y(0) = 0.

Step Response
If a step input change of Du is made at t = 0,

and we find the time-domain value

Equation 3.34

That is, the output ramps with a constant slope of kDu.

Impulse Response
If an impulse input of magnitude P is made at t = 0,
then the output immediately changes to a new steady-state value of

Example 3.5: Tank-Height Problem

The mathematical model for a liquid surge tank is (see Example 1.3)

where h is the liquid height, A is the constant cross-sectional area of the tank, F1 is the inlet flow
rate, and F2 is the outlet flow rate. Assume that the outlet flow rate remains constant at a steady-
state value of F2s. Defining the output and input in deviation variable form as

For a constant cross-sectional area of 10 m2, the model is

Step Response

For a step input change of 0.25 m3/min, the output response is

which is shown in Figure 3-7. If the steady-state height is 2 meters, then the height as a function
of time is

Figure 3-7. Step response of a liquid surge tank. Deviation variables.

Impulse Response

For an impulse input of 1 m3, the output response is

which makes sense, because the cross-sectional area is 10 m3.

[ Team LiB ]
[ Team LiB ]

3.7 Second-Order Behavior

Second-order models arise from systems that are modeled with two differential equations (two
states). In this section we separately consider transfer functions that do not have "numerator"
dynamics and those that do.

Pure Second-Order Systems

Consider a linear second-order ODE, with constant parameters

Equation 3.35

This is often written in the form

Equation 3.36

where (obviously a0 0)

and the parameters are as follows: k is the gain (units of output/input), z the damping factor
(dimensionless), and t the natural period (units of time).

The second-order model shown in Equation (3.35) or (3.36) generally arises by changing a set of
two first-order equations (state-space model) to a single second-order equation. For a given
second-order ODE, there are an infinite number of sets of two first-order (state-space) models
that are equivalent.

Taking the Laplace transform of Equation (3.36),

Equation 3.37

where y(s) indicates the Laplace transformed variable.

Assuming initial conditions are zero, that is , we find

Equation 3.38

The characteristic equation of the second-order transfer function is t2s2 + 2zts + 1. We can find
the roots (known as the poles) by using the quadratic formula

which yields the following values for the roots:

Equation 3.39

The following analysis assumes that z >0 and t > 0. This implies that the real portions of p1 and
p2 are negative and, therefore, the system is stable. The three possible cases are shown in Table
3-2.

Step Responses

Now, we consider the dynamic response of second-order systems to step inputs (u(s) = Du/s),

Equation 3.40

where Du represents the magnitude of the step change.

Case 1: Overdamped (z > 1)

For z > 1, the denominator polynomial, t 2s2 + 2zts + 1, can be factored into the form
where the time constants are

Table 3-2. Characteristic Behavior of Second-Order Transfer Functions

Case Damping factor Pole location Characteristic behavior

1 z>1 Two real, distinct roots Overdamped

2 z=1 Two real, equal roots Critically damped

3 z<1 Two complex conjugate roots Underdamped

We can derive the following solution for step responses of overdamped systems,

Equation 3.41

Note that, as in the case of first-order systems, we can divide by kDu to develop a dimensionless
output. Also, the dimensionless time is t/t and we can plot curves for dimensionless output as a
function of z. This is done in Figure 3-8, which includes the critically damped case, as discussed
next. Most chemical processes exhibit overdamped behavior.

Figure 3-8. Step response of a second-order overdamped system.

Case 2: Critically damped (z = 1)

The transition between overdamped and under damped is known as critically damped. We can
derive the following for the step response of a critically damped system

Equation 3.42

Notice that the main difference between overdamped (or critically damped) step responses and
first-order step responses is that the second-order step responses have an S shape with a
maximum slope at an inflection point, whereas the first-order responses have their maximum
slope initially.

The initial behavior for a step change is really dictated by the relative order of the system. The
relative order is the difference between the orders of the numerator and denominator polynomials
in the transfer function. If the relative order is 1, then output response has a nonzero slope at the
time of the step input; the step response of a system with a relative order greater than 1 has a
zero slope at the time of the step input.

Case 3: Underdamped (z < 1)

For z < 1, we find [from Equation (3.39)] that the poles are complex,

Equation 3.43a
where the real and imaginary contributions are

Equation 3.43b

We can derive the following step response for an underdamped system,

Equation 3.44a

where

Equation 3.44b

Again, dividing Equation (3.44a) by kDu, we can produce the dimensionless plot shown in Figure
3-9.

Figure 3-9. Step response of a second-order underdamped system as a

function of the damping factor (z). Interpolate between the curves for
the behavior of other damping factor values.
A number of insights can be obtained from Figure 3-9 and from an analysis of the step response
equations. For z < 1, the ratio of the imaginary portion to the real portion of the pole [from
Equation (3.39)] is

As the imaginary/real ratio gets larger, the response becomes more oscillatory (z becomes
smaller). We also notice that a decreasing ratio corresponds to a larger negative value for the real
portion. As the real portion becomes larger in magnitude (more negative), the response becomes
faster. We use these insights to interpret pole/zero plots in Section 3.9.

Underdamped Step Response Characteristics

The following common measures of underdamped second-order step responses are shown in
Figure 3-10, and defined below: (1) rise time, (2) time to first peak, (3) overshoot, (4) decay
ratio, and (5) period of oscillation.

Figure 3-10. Step response characteristics of underdamped second-

order processes.
Rise time is the amount of time it takes to first reach the new steady-state value. Time to first
peak is the time required to reach the first peak. Overshoot is the distance between the first peak
and the new steady state. This is usually expressed as the overshoot ratio, as shown in Figure 3-
10. Decay ratio is a measure of how rapidly the oscillations are decreasing. A b/a ratio of 1/4 is
commonly called "quarter wave damping." Period of oscillation is the time between successive
peaks.

Second-Order Systems with Numerator Dynamics

The previous discussion involved pure second-order systems, where the relative order (difference
between the denominator and numerator polynomial orders) was two.

Consider now a second-order system with numerator dynamics with the gain/time constant form

Equation 3.45

which is relative order one. The reader should show that the pole-zero form is
where the parameters are

The gain/time constant form has the following time-domain response to a step input (see Exercise
4):

Equation 3.46

The reader should show that, if tn = t2, the response is the same as a first-order process.

Example 3.6: Illustration of Numerator Dynamics

Consider the following input-output relationship:

The unit step responses are shown in Figure 3-11. Notice that negative numerator time constants
(corresponding to positive zeros) yield a step response which initially decreases before increasing
to the final steady state. This type of response is known as inverse response and causes tough
challenges for process control systems. Positive zeros are often caused by two first-order transfer
functions, with gains of opposite sign, acting in parallel (see Exercise 5).

Figure 3-11. Step responses of a second-order system with numerator

dynamics.
Notice also that a numerator time constant that is greater than the denominator time constant
causes overshoot before settling to the final steady state. Also notice that the inverse response
becomes "deeper" as the numerator zero (–1/tn) approaches a value of 0 from the positive side.

[ Team LiB ]
[ Team LiB ]

3.8 Lead-Lag Behavior

Lead-lag transfer functions have the same order numerator polynomial as denominator
polynomial. This occurs when the input has a direct effect on the output variable. In terms of the
state space model (3.2), this means that D 0.

Consider a lead-lag transfer function where the numerator and denominator polynomials are first
order:

Equation 3.47

For a step input of magnitude Du, the output response is

Equation 3.48

A dimensionless output can be defined by dividing Equation (3.48) by kpDu, and t/tp is a natural
dimensionless time. The responses to a step input at t = 0 are shown in Figure 3-12. Notice that
there is an immediate response that is equal to the tn/tp ratio. This can also be found by applying
the initial value theorem to Equation (3.47"/>) for a step input change (see Exercise 3).

Figure 3-12. Step responses of the lead-lag example.

It is rare for processes to exhibit lead-lag behavior, but many controllers exhibit such behavior.

[ Team LiB ]
[ Team LiB ]

3.9 Poles and Zeros

There are a number of different ways to represent process transfer functions. The "polynomial"
form is

Equation 3.49

The values of s that cause the numerator of Equation (3.49) to equal zero are known as the
"zeros" of the transfer function. The values of s that cause the denominator of Equation (3.49) to
equal zero are known as the "poles" of the transfer function.

The "pole-zero" form is

Equation 3.50

and complex poles (or zeros) must occur in complex conjugate pairs.

The "gain-time constant" form is the one that we use most often for control system design.

Equation 3.51

where tni is a numerator time constant and tpi is a denominator time constant. This form is
normally used when the roots (poles) of the denominator polynomial are real.

Example 3.7: Comparison of Various Transfer Function Forms

Consider a transfer function with the following gain-time constant form
The polynomial form is

The gain-polynomial form is

and the pole-zero form is

The zero is 1/10, and the poles are –1/3 and –1/15.

Notice that the zero for Example 3.7 is positive. A positive zero is called a right-half-plane (RHP)
zero, because it appears in the right half of the complex plane (with real and imaginary axes).
RHP zeros have a characteristic inverse response, as shown in Figure 3-11 for tn = -10 (which
corresponds to a zero of +0.1).

Also notice that the poles are negative (left-half-plane), indicating a stable process. RHP (positive)
poles are unstable. Recall that complex poles will yield an oscillatory response. A pole-zero plot of
the transfer function in Example 3.7 is shown in Figure 3-13 [the pole locations are (-1/3,0) and
(-1/15,0) and the zero location is (1/10,0), with the coordinates (real,imaginary)]. For this
system, there is no imaginary component and the poles and zeros lie on the real axis.

Figure 3-13. Pole-zero location plot for Example 3.7 (x, poles; o, zero).

As poles move further to the left they yield a faster response, and increasing the magnitude of the
imaginary portion makes the response more oscillatory. This behavior is summarized in Figure 3-
14. Recall also that a process with a pole at the origin (and none in the RHP) is known as an
integrating system; that is, the system never settles to a steady state when a step input change
is made.

Figure 3-14. Effect of pole-zero location on dynamic behavior (x,

poles; o, zero). As poles become more negative, the response is faster.
As the imaginary/real ratio increases, the response becomes more
oscillatory.

Multiple RHP zeros cause multiple "changes in direction"; for example, with two RHP zeros, the
step response, initially going in one direction, switches direction, then switches back to the initial
direction.

[ Team LiB ]
[ Team LiB ]

3.10 Processes with Dead Time

Many processes have a delayed response to a process input, either due to transport lags (such as
flow through pipes) or high-order effects. In Section 3.3.4 we found that a time-delay of q time
units had a transfer function of e–qs.

A first-order process combined with a time-delay has the transfer function

Consider a process with kp = 1, tp = 10, and q = 5. A unit step input change at t = 0 yields the
response shown in Figure 3-15. We see that the time-delay shifts the response of the output.

Figure 3-15. Response of a first-order + dead time process.

[ Team LiB ]
[ Team LiB ]

3.11 Padé Approximation for Dead Time

As discussed in the previous section the transfer function for a pure time delay is e–qs, where q is
the time delay. Some control system design techniques require a rational transfer function; the
Padé approximation for dead time is often used in this case.

A first-order Padé approximation is

A second-order Padé approximation is

Example 3.8: Application of the Padé Approximations for Dead Time

Consider the first-order + dead time transfer function, where the time-delay dominates the time
constant

The first-order Padé approximation yields the transfer function

and the second-order Padé approximation yields

A comparison of the step responses of g(s), g1(s), and g2(s) is shown in Figure 3-16. Notice that
the first-order approximation has an inverse response, while the second-order approximation has
a "double inverse response." The reader should find that there is a single positive zero for g1(s),
and there are two positive, complex-conjugate zeros of the numerator transfer function of g2(s).

Figure 3-16. Comparison of first-order + dead time response with first

and second-order Padé approximations for dead time.

Most ordinary differential equation numerical integrators require pure differential equations (with
no time delays). If you have a system of differential equations that has time delays, the Padé
approximation can be used to convert them to delay-free differential equations, which can then be
numerically integrated.

One of the many advantages to using SIMULINK is that time delays are easily handled so that no
approximation is required.

[ Team LiB ]
[ Team LiB ]

3.12 Converting State Space Models to Transfer

Functions
A general state space model can be converted to transfer function form, using the following steps.
Starting with the state space model

Take the Laplace transform of each term, assuming zero initial conditions

Solving for x(s), then y(s) (it should be noted that often D = 0)

where G(s) is a transfer function matrix. For example, the transfer function relating input j and
output i is

or in matrix form (with m inputs and r outputs)

Example 3.9: Isothermal CSTR

Consider the isothermal CSTR shown in Example 3.3 and Module 5. The state space model is
The first input (u1, manipulated) is the dilution rate (F/V), the second input (u2, disturbance) is
the feed concentration (C Af), and the output is the concentration of the intermediate component,
B. The eigenvalues of A (obtained by solving det(lI – A) = 0) are –2.4048 and –2.2381 min-1.

The sequence of steps used to find the transfer function matrix is

and multiplying,

Also,

The manipulated input-output process transfer function for the reactor is

and the disturbance input-output transfer function is

The transfer function poles (–2.2381 and –2.4048) are equal to the eigenvalues of the A matrix.
Also, the positive zero (1/0.3549) in g11(s) yields the inverse response shown in Figure 3-5.

We see that it is straightforward to convert state space models to transfer function models. An n-
state system results in transfer functions that have a denominator polynomial that is nth order in
s, that is, with n poles. Sometimes the resulting transfer functions can be factored into lower
order transfer functions because of pole-zero cancellation (a value of a pole is equal to a value of
a zero). An example of pole-zero cancellation is shown in Exercise 13.

[ Team LiB ]
[ Team LiB ]

3.13 MATLAB and SIMULINK

MATLAB and SIMULINK are ideal for analyzing and simulating dynamic behavior. The Control Toolbox
uses LTI (linear, time invariant) objects to represent dynamic models. These objects can be state
space or transfer function models; details on how to use these are presented in Module 4.

Here we first present some useful MATLAB functions not covered in other modules.

conv (Convolution) and roots

Often it is necessary to multiply polynomials in s together. Although the following multiplication is
quite simple: (5s + 1)(3s + 1) = 15s2 + 8s + 1, a more complex multiplication, such as (2s2 +
1.5s + 1)(3s3 + 8s2 + 4s + 1) could be quite time consuming. The following conv (convolve)
command makes this straightforward. Also, the roots of the resulting polynomial are easily
calculated using the roots command.

» g = conv([2 1.5 1],[3 8 4 1])

g =
6.0000 20.5000 23.0000 16.0000 5.5000 1.0000
» p = roots(g)
p =
-2.1095
-0.3750 + 0.5995i
-0.3750 - 0.5995i
-0.2786 + 0.2836i
-0.2786 - 0.2836i

step

First, define an object, such one of the transfer functions in Example 3.6.

» num = [-10 1]; den = [45 18 1];

» sysc = tf(num,den)

Transfer function:
-10 s + 1
-----------------
45 s^2 + 18 s + 1
then plot the step response (use help step to view other options)

» step(sysc)

A number of other useful functions, such as pole and tzero are shown in Module 4.

Simulink
It is highly recommended that you interactively work through the initial sections of Module 2 to
understand how SIMULINK can be used to simulate dynamic systems.

[ Team LiB ]
[ Team LiB ]

3.14 Summary
The main objective of this chapter was to develop an understanding of dynamic process behavior.
Although Laplace transforms can be used to obtain analytical solutions to differential equations,
we will not be using them for this purpose during the rest of this textbook. The concept of a
transfer function is very useful for control system design and analysis. In future chapters we find
that transfer functions allow the use of algebra rather than differential equations when analyzing
control systems composed of a number of components (controller, valve, process, sensor, etc.).

The stability of a process is determined by the eigenvalues of the state space A matrix, or the
poles of the Laplace transfer function: if all are negative, the system is stable. Complex
eigenvalues (poles) correspond to underdamped (oscillatory) behavior, characterized by damping
factors less than 1. Dynamic responses are also affected by the values of the zeros (roots of the
numerator polynomial in s) of a transfer function. If zeros are positive (in the right-half-plane),
step responses have a characteristic inverse response.

A first-order + dead time model is the most common process representation. Time-delays are
represented by an irrational term (e–qs) in the Laplace domain. First and second-order Padé
approximations are sometimes used for controller design; the approximations lead to right-half-
plane zeros and inverse response.

The initial and final value theorems of Laplace transforms can be used to find the limiting behavior
of transfer functions without having to obtain a full solution. The process gain is a measure of the
long-term change in a process output for a given input change.

State space models are easily converted to Laplace transfer function models. The use of LTI
objects (see Module 4) allows easy interconversion of models in MATLAB.

[ Team LiB ]
[ Team LiB ]

References

For more details on analytical mathematical methods (including Laplace transforms) to solve
differential equations, see: Boyce, W., and R. DiPrima, Ordinary Differential Equations and
Boundary Value Problems, 5th ed., Wiley, New York (1992).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Solve for the time-domain output of a first-order transfer function to a step input
change.

2: A second-order process with one pole at the origin has the transfer function

Find the output as a function of time, for a unit step input change. Sketch the expected
behavior.

3: Use the initial value theorem to find the immediate response of a lead-lag transfer
function to a step input change at t = 0.

Also, use the final value theorem to find the long-term response of a lead-lag transfer
function to a step input change.

4: For the following second-order process with numerator dynamics, solve for the time-
domain output response to a step input change of magnitude Du at t =0.

For kp = 1°C/Lpm, t1 = 3 min, t2 = 15 min, tn = 20 min find the peak temperature

and the time that it occurs.

5: Consider an input-output transfer function that mimics two first-order processes in

parallel

where
 If the gain of g1 is positive and the gain of g2 is negative, find the conditions
(relationship between gains and time constants for the two transfer functions) that
cause a right-half-plane zero (resulting in inverse response to a step input change) in
gp(s).

6: Consider the state space model

Find the second-order differential equation in y. Hint: first solve for x 1 from the second
equation, then take the derivative and substitute into the first equation.

7: Consider the following state-space model

Which has the following input-output transfer function relationship

For a unit step change in the input, u(s) = 1/s:

i. Find the output at t=0 and as t approaches infinity, using the initial and final
value theorems.

ii. Find the time domain solution, y(t)

iii. Sketch the time domain behavior of y(t)

iv. Are your results for i, ii, and iii consistent?

8: As a process engineer with the Complex Pole Corporation, you are assigned a unit with
an exothermic chemical reactor. In order to learn more about the dynamics of the
process, you decide to make a step change in the input variable, the coolant
temperature, from 10°C to 15°C. Assume that the reactor was initially at a steady
state. You obtain the following plot for the output variable, which is reactor
temperature (notice that the reactor temperature is in °F). Use Figure 3-9 to help
answer the following questions.
 i. What is the value of the process gain? (show units)

ii. What is the value of t? (show units)

iii. What is the value of z? (show units)

iv. What is the decay ratio?

v. What is the period of oscillation? (show units)

vi. Write the second-order transfer function.

Hint: Use Figure 3-9 to assist you.

9: Match the transfer functions with the responses to a unit step input, shown in the
figure.

i.

ii.

iii.

iv.

v.
 v.

10: Consider the following state space model:

i. Find the transfer function model.

ii. Find (analytically) the time-domain output response to a unit step input change.
Sketch the expected response to a unit step input change.

11: As a process engineer, you decide to develop a first-order + time delay model of a
process using a step test. The process is initially at steady state, with an input flow
rate of 5 gpm and an output of 0.75 mol/L. You make a step increase of 0.5 gpm at
3:00 p.m. and do not observe any changes until 3:07 p.m. At 3:20 p.m., the value of
the output is 0.8 mol/L. You become distracted and do not have a chance to look at
the output variable again, until you leave for happy hour at a local watering hole at
6:30 p.m. You note that the output has ceased to change and has achieved a new
steady-state value of 0.85 mol/L. What are the values of the process parameters, with
units? Show your work.

12: Use the initial and final value theorems of Laplace transforms to determine the initial
and final values of the process output for a unit step input change to the following
transfer functions.

i.
 i.

ii.

iii.

13: Consider the following state space model for a biochemical reactor. Since there are two
states (the A matrix is 2 x 2) we expect that the process transfer function will be
second-order. Show that pole-zero cancellation occurs, resulting in a first-order
transfer function. Find the values of the gain and time constant.

14: Match the transfer functions with the responses to a unit step input, shown in the
figure.

i.

ii.

iii.

iv.

v.
 v.

15: Consider Example 3.4. For an impulse input of 30 kJ, find the value of the vessel
temperature immediately after the impulse input is applied.

16: Consider the following second-order transfer function

For a unit impulse input, find the output response as a function of time. What is the
peak change and when does it occur?

[ Team LiB ]
[ Team LiB ]

Chapter 4. Empirical Models

The previous two chapters have focused on (i) the development of fundamental models based on
material and energy balances (Chapter 2), and (ii) understanding dynamic behavior, with an
emphasis on linear systems (Chapter 3). In this chapter we discuss the development of empirical
models, that is, developing models based on plant tests. For continuous control-system design
(most of the focus of this text), a study of Sections 4.1–4.3 will be sufficient. The discrete models
developed in Sections 4.4 and 4.5 will be useful when model predictive control (MPC) techniques
are presented in Chapter 16, and when digital control is studied in Module 16.

After studying this chapter the reader should be able to:

Develop continuous first-order and integrator + dead time models from step tests

Estimate parameters for discrete-time autoregressive models based on input-output data

Calculate poles and zeros of discrete-time models

Develop finite step and impulse response models

The major sections of this chapter are as follows:

4.1 Introduction

4.2 First-Order + Dead Time

4.3 Integrator + Dead Time

4.4 Discrete-Time Autoregressive Models

4.5 Parameter Estimation

4.6 Discrete Step and Impulse Response Models

4.7 Summary

[ Team LiB ]
[ Team LiB ]

4.1 Introduction
For many processes there is not enough time, or it is not worth the effort, to develop fundamental
process models. Particularly if your main interest is tuning a specific control loop, it is likely that
you will develop a transfer function-based model by performing a plant test. The most common
plant test is to make a step change in the manipulated input (controller output) and observe the
measured process output response. Then a model is developed to provide the best match
between the model output and the observed plant output.

There are a number of important issues in developing an input-output model. Foremost is the
selection of the proper input and output variables. For many processes this is not trivial, because
a particular manipulated input may affect a number of measured outputs. Similarly, a measured
output may be affected by a number of manipulated inputs. In this chapter we assume that the
manipulated input and measured output have already been selected. The important discussion of
the selection of a particular input to be "paired" with a particular output is postponed until
Chapter 13.

It is common to base an input-output model on step responses. In this procedure we first bring
the process to a consistent and desirable steady-state operating point, then make a step change
in the input variable. An important decision is the magnitude of the step change to make.

If the step change in input is too small, the measured output may not change enough to
develop a reliable model. This is particularly true if the measured output is "noisy." Clearly,
the magnitude of step input must be enough so that the output "signal-to-noise" ratio is
high enough to obtain a good model.

If the step change in the input is too large, the output variable may change too much and
produce product that is "off-specification." This is not desirable because of the severe
economic penalty (i.e., the plant loses too much money while the step test is being
performed). Also, if the step input change is too large, nonlinear effects may dominate. That
is, the operating condition may become significantly different than the desired condition.

Clearly there is a trade-off here. The input must be changed enough to observe a change in the
output variable (it must increase above the noise level), yet not so much that the output variable
change is too great (causing an economic penalty).

As a process engineer conducting a plant step test, you will usually have some basic knowledge of
(or experience with) the input-output pair under consideration. Observations of the measured
output with time will provide an estimate of the standard deviation or variance of the
measurement noise. A rough estimate of the process gain will enable you to select an input
change magnitude so that the output change is "above the noise level." An estimate of the
process gain can often be obtained through steady-state material and energy balances.

In the following sections we show how to estimate parameters for some common simple models.
By far the most commonly used model, for control-system design purposes, is the first-order +
time-delay model discussed in the next section.

[ Team LiB ]
[ Team LiB ]

4.2 First-Order + Dead Time

Recall that a first-order + dead-time process, represented by the transfer function relationship

Equation 4.1

has the following output response to a step input change,

Equation 4.2

where the measured output is in deviation variable form. The three process parameters can be
estimated by performing a single step test on the process input.

The gain is found as simply the long-term change in process output divided by the change in
process input. Also the time delay is the amount of time, after the input change, before a
significant output response is observed. There are several easy ways to estimate the time
constant for this type of model.

Time for 63.2% Approach to New Steady State

In Equation (4.2), if we set t = tp + q, we find y(t = tp + q) = kpDu(1 - exp(–1)) = 0.632kpDu.
Now, since kpDu is simply the long-term change in the process output, then t63.2% = tp + q is the
amount of time it takes for the output to make 63.2% of its ultimate change. This method of
estimating the time constant is illustrated in Figure 4-1.

Figure 4-1. Estimating the time constant from a response to a step

input.
Example 4.1: Numerical Application of 63.2% Method
Consider the response to a step input change at t = 1 minute, shown in Figure 4-2. The measured
output is shown as the noisy solid curve, and a "best fit" first-order + time-delay model is shown
as the dashed curve. Here we find the process gain from

Figure 4-2. Example of the step response with measurement noise.

The time delay is observed to be 1 minute, and the time constant is the time it takes (after the
time delay) for the output to change by 0.632(-2) = -1.3°C. That is, the time when y = 23.7°C.
In this case,

and the process transfer function is

Maximum Slope Method

We can also find (see Exercise 1) that the maximum slope of the output response to a step input
change at t = 0 occurs at t = q and is

So the time constant can be estimated from

A major disadvantage is that the slope estimate may not be accurate if there is significant
measurement noise. This approach is particularly useful for higher-order responses that are
approximated as a first-order + time-delay response, as shown in the following example.

Example 4.2: Maximum Slope Method

The response of a high order process is shown in Figure 4-3. Here, the dashed line is drawn
through the maximum slope of the output response. The slope is

Figure 4-3. Example of the slope method for estimation of time

constant.

Since the process gain is kp = Dy/Du = 6°C/2 gpm = 3°C/gpm, and the time-delay is 8-1 = 7
min, the first-order + time-delay transfer function is estimated as
Two-Point Method for Estimating Time Constant
A technique similar to the first one, but using two points on the output response, is shown in
Figure 4-4. Here the time required for the process output to make 28.3% and 63.2% of the long-
term change is denoted by t28.3% and t63.2%, respectively. The time constant and time delay can
be estimated from (see Exercise 2)

Figure 4-4. Example of the two-point method for estimation of time

constant.

Limitation to FODT Models

The primary limitation to using step responses to identify first-order + dead-time transfer
functions is the amount of time required to assure that the process is approaching a new steady
state. That is, the major limit is the time required to determine the gain of the process. For large
time constant processes it is often desirable to use a simpler model that does not require a long
step test time. The integrator + dead-time model shown next will often be good enough for
control-system design.

[ Team LiB ]
[ Team LiB ]

4.3 Integrator + Dead Time

An integrator + dead-time process has the input-output transfer function relationship

Equation 4.3

and the output response to a step input

Equation 4.4

with the response shown in Figure 4-5. The gain, k, can be found from the slope by solving

Figure 4-5. Estimating parameters for an integrator + time-delay

model from a step input test.
and the time delay is clearly identified by the time required for a change in output.

Integrator + dead-time models are good for describing the behavior of "integrating processes,"
such as vessel liquid levels or gas drum pressures. They can provide a good short-term
approximation to the step response behavior of a first-order + dead-time process. Consider the
step response of a distillation column, which has a fairly long time constant. We see from Figure
4-6 that it takes roughly 5–6 hours to obtain good estimates for a FODT model.

Figure 4-6. Distillation column overhead composition response to a

step change of 0.01 kmol/min in reflux rate at t = 10 minutes.
The reader should be able to show that the transfer function is approximately

where the gain has units of mol%/(kmol/min) and the time unit is minutes.

The initial response of Figure 4-6 is "blown-up" in Figure 4-7. Notice that fewer than 20 minutes
(the step change is made at t = 10 minutes) is required to obtain a satisfactory integrator + time-
delay model. Clearly, plant operators (and managers) will be much more willing to tolerate a 20-
minute test than a 4- to 5-hour test.

Figure 4-7. Distillation column overhead composition response to a

step change of 0.01 kmol/min in reflux rate at t = 10 minutes. Focus is
on the short-term response.
The reader should be able to show that the estimated transfer function is approximately

[ Team LiB ]
[ Team LiB ]

4.4 Discrete-Time Autoregressive Models

Most of the models developed and used in this text are based on continuous time. In the most
general case of fundamental models, these are nonlinear ordinary differential equations. State
space models consist of sets of linear ordinary differential equations, where the states and inputs
are in deviation variable form. Transfer functions are used to denote a linear relationship between
inputs and outputs. All these previous models assume that the manipulated input and the
measured output signals are continuously available.

In practice, manipulated input changes are made at discrete time intervals and measured outputs
are available at discrete sample times. The physical inputs and outputs are continuous; only the
changes and measurements occur at discrete times. Since the focus of this textbook is on
controller design based on continuous models, the topic of discrete models may be skipped on a
first reading. Discrete models are primarily useful for model predictive control (Chapter 16) and
digital control design and analysis (Module 16).

In this section we cover discrete autoregressive models, in Section 4.5 parameter estimation of
discrete autoregressive models is covered, and in Section 4.6 we cover finite step and impulse
response models. It should be noted that the M ATLAB Control Toolbox has routines that convert
continuous-time models to discrete and vice versa. The details of the conversion are beyond the
scope of this text, but are summarized in the appendix; read Module 4 for explanations on how to
use these routines.

Introduction to Autoregressive Models

Discrete autoregressive models assume that an output at the current time step is a function of
the outputs and inputs at previous time steps. The general form is

Equation 4.5

where k represents the current time step, k - 1 the previous time step, and so forth. The notation
y(k) is used to indicate the value of the output at step k. If all of the outputs are brought to the
left hand side, Equation (4.5) can be written

Equation 4.6

This form is primarily used to lead to the notion of discrete input-output transfer functions
covered next.
Z-Transforms
Similar to the use of Laplace transforms for continuous systems, Z-transforms are used for
discrete systems.

The Z-transform of a discrete variable, y(k), is

Equation 4.7

The so-called backwards shift operator (z-1) is represented by

Equation 4.8

so Z[y(k - 2)] = z-2y(z), and so forth. Taking the Z-transform of each term in Equation (4.6), we
find

Equation 4.9

Solving for y(z) yields the discrete transfer function relationship

Equation 4.10

where the discrete transfer function is

Equation 4.11

The discrete transfer function is the discrete-time analogy to the continuous-time transfer
function based on the Laplace transform. Similar to continuous (Laplace) transfer functions,
discrete (z-transform) transfer functions can be written in several forms. Multiplying Equation
(4.11) by zn/zn, we find
Equation 4.12

For most process systems there is not an immediate effect of the input on the output, so b0 = 0.

Poles/Zeros of Discrete Models

Recall that the stability of a continuous input-output (transfer function) model is determined by
the values of the poles (the roots of the denominator polynomial in s). If the poles are negative,
then the model is stable. Similarly, the stability of a discrete input-output model is determined by
the poles of the denominator polynomial in z. For a discrete-time model, if all poles have a
magnitude less than 1 (i.e., they are "inside the unit circle") the model is stable. If any pole has a
magnitude greater than 1, the model is unstable.

The polynomial forms of Equations (4.11) and (4.12) can be factored into gain-pole-zero form

Equation 4.13

where zi and pi are the zeros and poles, respectively.

Example 4.3: Discrete Poles and Stability

Consider the simple first-order model

which has the transfer function

The pole is found by solving for the roots of the denominator polynomial. In this case

To illustrate the importance of pole values, we will study the following values for a1:
yielding the corresponding pole values

Also, realize that the modeling equation can be written

For simplicity, we will assume no input change, so u(k) = 0. Also, let the initial value of the output
be y(0) = 1. The first few values of y(k) for each value of a1 are shown in the table below.

a1 0.5 -0.5 1.5 -1.5

p -0.5 0.5 -1.5 1.5
y(1) -0.5 0.5 -1.5 1.5
y(2) 0.25 0.25 2.25 2.25
y(3) -0.125 0.125 -3.375 3.375
y(4) 0.0625 0.0625 5.0625 5.0625
Characteristic Oscillatory, Monotonic, Oscillatory, Monotonic,
behavior stable stable unstable unstable

As expected, when the magnitude of the pole is less than 1.0, the process is stable; when the
magnitude is greater than 1.0, the process is unstable. One thing that is unusual for this first-
order discrete example is that the output can oscillate even with a constant (in this case, 0) input
value. For continuous-time systems there must be at least two states (the process must be at
least second-order) for oscillation to occur. This is because the condition for oscillation in
continuous-time systems is that the eigenvalues must be complex; for this to happen, there must
be at least two eigenvalues, each complex-conjugates of the other. Discrete time systems can
oscillate, as long as a pole has a negative value; if the negative pole has a magnitude less than
one, it is a stable oscillation.

Final and Initial Values Theorems for Discrete Systems

Recall that the process gain of a continuous system could be determined by setting s = 0.
Similarly, the gain of a discrete-time system can be found by setting z = 1.

Final Value Theorem for Discrete-Time Systems

The formal statement of the final value theorem is (where Dt is the sample time)
For a discrete-time input-output model

subject to a unit step input

we find

so, applying the final value theorem

so we can find the long-term behavior of the process output, subject to a unit step input, simply
by setting z = 1 in gp(z).

Initial Value Theorem for Discrete-Time Systems

Similarly, the initial value theorem is

[ Team LiB ]
[ Team LiB ]

4.5 Parameter Estimation

Often when discrete linear models are developed, they are based on experimental system
responses rather than on converting a continuous model to a discrete model. The estimation of
parameters for discrete dynamic models is no different from linear regression (least squares)
analysis.

The measured inputs and outputs are the independent variables, and the dependent variables are
the outputs. For simplicity, consider the following model where two previous values of the input
and output are used to predict the next value. This model has four parameters

Equation 4.14

For the system of N data points we can write the following

Notice that k = 1 is an arbitrary starting point for calculating the output, and that two previous
data points for the output and the input are needed to start the algorithm. To make it clear that
we are performing a model-based prediction of yk, we will use the notation k to indicate a model
prediction

Equation 4.15

or, using matrix-vector notation

Equation 4.16

where
Equation 4.17

The objective is to choose a set of parameters (a1, a2, b1, b2) to minimize the square of the
residuals (the differences between the model outputs and measured outputs).

Equation 4.18

The sum of the squares of the residuals can be written in matrix notation as

Equation 4.19

where Y is the vector of measured outputs, Y = [y(1) . . . y(N)]T.

The solution to this optimization problem is (Ljung, 1996)

Equation 4.20

Example 4.4: Process Identification

Although the response to a step input has often been used for process identification, here we
apply a pseudo-random binary sequence of inputs (the input changes randomly between two
values) to estimate parameters in a discrete time model. The input and output responses are
shown in Figure 4-8; the open circles represent the measured outputs obtained with a sample
time of 0.25 minutes. The variables are in deviation form.

Figure 4-8. Output response to a pseudo-random binary input

sequence. Open circles represent output data obtained with a sample
time of 0.25 minutes.
The output and input data as a function of the discrete time index are:

k -1 0 1 2 3 4 5
y 0 -0.0889 0.0137 0.1564 0.4618 0.1771 0.3446
u 1.0000 1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000
k 6 7 8 9 10 11 12
y 0.2171 -0.1558 0.0485 -0.1879 -0.1123 0.0463 0.2003
u 1.0000 -1.0000 1.0000 1.0000 1.0000 1.0000 -1.0000
k 13 14 15 16 17 18 19
y 0.5007 0.3846 -0.0172 0.1513 -0.1162 0.1134 0.0502
u -1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000 -1.0000

The F (input-output data) matrix, Q parameter vector, and discrete transfer functions are shown
below.
The M ATLAB and SIMULINK .m and .mdl files used to generate this example are shown in the
Appendix. It should be noted that the discrete transfer function can be converted to a continuous
transfer function using d2c to find

This is virtually identical to the continuous transfer function used to generate the original data.

This example illustrated the response of a perfectly modeled system (no measurement noise).
The approach can also be applied to a system with arbitrary inputs, and with noisy
measurements. See Exercise 8 for the same process, with output data corrupted by measurement
noise. To verify model parameter estimates it is common to use a portion of experimental data to
estimate parameters, then another portion of the data in "simulation" mode to see how well the
model predicted outputs match the data.

The data was analyzed in a batchwise fashion, that is, all of the data were collected before the
parameter estimation was performed. For on-line (real-time) estimation and control, we would
prefer to re-estimate parameters each time we obtain a new data point. This can be done by
using a "moving horizon" of past data points (each time a new data point is collected, the oldest
one is thrown out), or by using recursive identification techniques. Recursive techniques are
commonly applied in adaptive control, where the model is updated at each time step and a control
calculation is performed based on the updated model.

[ Team LiB ]
[ Team LiB ]

4.6 Discrete Step and Impulse Response Models

The common model predictive control techniques presented in Chapter 16 are based on step or
impulse responses.

Step Response Models

Finite step response (FSR) models are obtained by making a unit step input change to a process
operating at steady state. The model coefficients are simply the output values at each time step,
as shown in Figure 4-9. Here, si represents the step response coefficient for the ith sample time
after the unit step input change. If a nonunit step change is made, the output is scaled
accordingly.

Figure 4-9. Illustration of step response parameter identification.

The step response model is the vector of step response coefficients,

where the model length n is long enough so that the coefficient values are relatively constant
(i.e., the process is close to a new steady state).
Impulse Response Models
Another common form of model is known as a finite impulse response (FIR). Here, a unit pulse is
applied to the manipulated input, and the model coefficients are simply the values of the outputs
at each time step after the pulse input is applied. As shown in Figure 4-10, hi represents the ith
impulse response coefficient.

Figure 4-10. Illustration of step response parameter identification.

There is a direct relationship between step and impulse response models:

Figure 4-11 illustrates how impulse response coefficients can be obtained from step responses.
The impulse response coefficients are simply the changes in the step response coefficient at each
time step. Similarly, step response coefficients can be found from impulse responses; a step
response coefficient is the sum of the impulse response coefficients to that point.

Figure 4-11. Obtaining impulse response models from step response

data.
[ Team LiB ]
[ Team LiB ]

4.7 Summary
The first sections of this chapter illustrated the development of first-order + time-delay and
integrator + time-delay models based on step input changes. These models will be sufficient for
many simplified controller designs. Higher order models can easily be developed by "fitting"
(adjusting) parameters to obtain a best match between model predictions and the actual
measured process output responses.

Linear regression (least squares) techniques can be used to estimate parameters in discrete time
autoregressive models. Usually, parameters are estimated based on a subset of the available
experimental data. The model parameters are then verified by applying the model to a different
subset of the data. If the model predictions match the measured outputs reasonably well, then
the discrete model is usually acceptable for discrete control system design and analysis.

The stability of a discrete transfer function model is determined from the values of the poles of
the denominator polynomial. If the poles are less than 1 in magnitude, the discrete model is
stable. Module 16 covers classical discrete controller design procedures. Discrete step and impulse
response models are often used in model predictive control (Chapter 16).

[ Team LiB ]
[ Team LiB ]

References

An overview of system identification techniques is provided by Ljung: Ljung, L., System

Identification, Chapter 58 ( pp. 1033–1054) in The Control Handbook, W.S. Levine (ed.), CRC
Press (1996).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider a first-order + time-delay transfer function

which has the following response to a step input change at t = 0:

Show that the maximum rate-of-change (slope) of the output occurs at t = q . Also, find this slope.

2: Consider the response of a first-order + time-delay transfer function to a step input change. Find the
value of the output (y ) at the following times, as a fraction of the long-term output change

Find the time constant and time-delay based on the values of t 1 and t 2 .

3: Consider the following step response. Estimate parameters for a first-order + time-delay model using
the three techniques shown in Section 4.3 . Include the units for each parameter.
4: Consider the following step response. Estimate the parameters for an integrator + time-delay model,
including the units for each parameter.

5: Consider the following continuous state space model:

 a. Find the continuous transfer function model (do this analytically).

b. For a sample time of 0.25, find the discrete state space and transfer function models (use M ATLAB
; see Module 4 ).

c. Compare the step responses of the continuous and discrete models (use MATLAB ). What do you
observe?

6: Consider the discrete-time model

using z -transforms, find the corresponding model represented as equations (4.11 ), (4.12 ) and (4.13
). Discuss the stability of the process.

7: Consider a unit step input change made at k = 0, resulting in the output response shown in the plot
and table below.

k -1 0 1 2 3 4 5 6 7 8 9 10
y 0 0 0.1044 0.3403 0.6105 0.8494 1.0234 1.1244 1.1616 1.1531 1.1184 1.0746
u 0 1 1 1 1 1 1 1 1 1 1 1
k 11 12 13 14 15 16 17 18 19 20
y 1.0336 1.0023 0.9828 0.9744 0.9742 0.9790 0.9860 0.9929 0.9985 1.0022
u 1 1 1 1 1 1 1 1 1 1
 Estimate the parameters for a discrete linear model with the form

Compare the step response of the estimated model with the data. Use MATLAB to convert the discrete
model to a continuous model. Compare the step responses of the discrete and continuous models.

8: Consider Example 4.4 with measurement noise on the output variable, as shown below. Estimate the
discrete model parameters based on this data. How do the parameters compare with those of Example
4.4 ? Compare the step responses of the two models.

The input/output data are

k -1 0 1 2 3 4 5
y 0.0741 -0.0857 -0.0399 0.1663 0.4065 0.1521 0.3910
u 1.0000 1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000
k 6 7 8 9 10 11 12
y 0.2284 -0.2569 0.0910 -0.1737 -0.1260 0.0668 0.1958
u 1.0000 -1.0000 1.0000 1.0000 1.0000 1.0000 -1.0000
k 13 14 15 16 17 18 19
y 0.4976 0.3724 0.0119 0.0927 -0.0528 0.1357 0.0580
u -1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000 -1.0000
 9: Consider the continuous state space model (where the time unit is minutes)

a. Find the eigenvalues and the transfer function (use MATLAB for these calculations, if desired)

b. Using a sample time of 3 minutes, find the discrete state space model and the discrete transfer
function. Refer to the Appendix for the form of the discrete state space model, and Module 4 to
understand how to use MATLAB for these computations.

[ Team LiB ]
[ Team LiB ]

Appendix 4.1: Files Used to Generate Example 4.4

Figure A4-1. Simulink File est_vdv.mdl.

% use PRBS to estimate a discrete model of the vdv reactor

%
T = 0:0.25:5; % time vector (row)
T = T'; % time vector (column)
seed = [3;0.5]; % seed for noise generation
randn('state',seed); % using the same seed each time
uu = 0.5 1 0.25*randn(length(T),1); % mean of 0.5 with variance
% of 0.25
U = 2*round(uu)-1; % creates PRBS with -1 and 1 values
%
sim('est_vdv'); % runs simulation of linear van de vusse
% diagram
figure(1); % plot input-output data
subplot(2,1,1),plot(tp,yp,'k',t,y,'ko');
xlabel('time, min'), ylabel('y')
title('PRBS estimation example')
subplot(2,1,2),plot(tp,up,'k'); xlabel('time, min'),ylabel('u')
axis([0 5 -1.1 1.1])
% % generate phi matrix for estimation
for j = 4:22;
phi(j-3,:) = [y(j-2) y(j-3) u(j-2) u(j-3)];
end
%
theta = inv(phi'*phi)*phi'*y(3:21) % estimate parameters
%
num = [theta(3) theta(4)]; % numerator of discrete transfer function
den = [1 -theta(1) -theta(2)]; % denominator of discrete transfer function
sysd = tf(num,den,0.25) % create discrete tf object
tzero(sysd) % calculate zeros
pole(sysd) % calculate poles
syszpk = zpk(sysd) % zero-pole-k form

[ Team LiB ]
[ Team LiB ]

Appendix 4.2
Discrete time models can be obtained directly from continuous time models, as summarized in
this appendix. The MATLAB Control Toolbox can be used for these conversions, as shown in Module
4.

Consider a continuous state space model of the following form (where we have assumed D = 0)

Equation A4.1

A similar discrete state space model has the following form

Equation A4.2

Assuming that the sample time is Dt, and that the input u is held constant between time tk and
tk+1, (A4.1) can be integrated to yield

which has the solution (although the matrix exponential is not an intuitive concept, the
computation is readily performed by MATLAB)

which is usually written in the following form

where
The stability of this discrete state space model is determined by the eigenvalues of F, which must
have a magnitude less that 1.0 for stability.

The discrete transfer function is, similar to the continuous time case

[ Team LiB ]
[ Team LiB ]

Chapter 5. Introduction to Feedback

Control
The objective of this chapter is to introduce the Laplace domain analysis of feedback control
systems. After studying this chapter, the reader should be able to do the following:

Convert a process and instrumentation diagram (P&ID) to a control block diagram

Find the controller transfer function for standard feedback controllers

Find the closed-loop transfer function for a given process and controller transfer function

Determine the stability of a closed-loop system

Find the transfer function relationship between any two signals on a block diagram

The major sections of this chapter are as follows:

5.1 Motivation

5.2 Development of Control Block Diagrams

5.3 Response to Setpoint Changes

5.4 PID Controller Algorithms

5.5 Routh Stability Criterion

5.6 Effect of Tuning Parameters

5.7 Response to Disturbances

5.8 Open-Loop Unstable Systems

5.9 SIMULINK Block Diagrams

5.10 Summary

[ Team LiB ]
[ Team LiB ]

5.1 Motivation
Consider the level control problem first described in Chapter 1 and shown in Figure 5-1. A level
measurement device (usually a differential pressure cell) senses the level and sends a signal to
the controller. Notice that the controller is comparing the tank height with the desired setpoint
and sending a controller output (pressure signal) to the valve, which changes the valve position
and therefore the volumetric flow rate of stream 1, F 1. These signals are shown as dashed lines
on the figure.

Figure 5-1. Level control, inlet flow rate manipulated. The

instrumentation signals are shown as dashed lines. The level
transmitter sends a signal to the level control, which sends a signal to
the control valve.

The question is, what is the control algorithm? How does the controller change the flow rate to
the process? You can probably think of a number of possible algorithms.

On-Off Control
An on/off controller is similar to a controller used in a typical home heating unit. If the liquid level
is too high (higher than the desired setpoint), the valve is closed and if the level is too low (lower
than the desired setpoint) the valve is fully opened. The problem with an on-off strategy is that
the valve will constantly be opening and closing, as shown in Figure 5-2, causing wear and tear on
the valve.

Figure 5-2. On-off control of tank height.

Notice that an on-off controller is based on the error between the setpoint (desired height, h s p)
and the measured output (measured height). If the error (h sp – hm) is greater than a certain
positive value (d, 1/2 the dead band), the valve is turned off; if the error is less than a certain
negative value (–1/2 the dead band), the valve is turned on. Assume that an increase in valve top
pressure increases the flow through the valve [this is known as an air-to-open (or fail-closed)
valve]. A fail-closed valve is used to make certain that the tank does not overflow in the event of
a control valve failure. The algorithm could be stated

Equation 5.1

We see that the controller behavior is periodic, as shown in Figure 5-2.

The next option is to make the signal to the valve proportional to the difference between the
desired level and the actual level. This is known as a proportional controller (or P-controller) and
is discussed below.

Proportional Control
The following equation suggests a control action that is proportional to the error (difference
between setpoint and measurement),

Equation 5.2

where b is the bias term, kc is the proportional gain (a tuning parameter), and the difference
between the setpoint and the measured output is the error

Equation 5.3

If the error is 0 at steady state, notice that the bias term is equal to the steady-state valve-top
pressure.

Now we discuss the adjustment of the tuning parameter (proportional gain, kc). Consider the case
where the error is positive (setpoint is greater than the measured output). This indicates that the
flow rate entering the tank should be higher and therefore the valve-top pressure should increase.
The proportional gain, kc, must then be positive, for the control strategy shown in Figure 5-1. As
kc is increased, more control action is requested for a given amount of error.

Now consider the control strategy shown in Figure 5-3. A similar proportional control algorithm is

Equation 5.4

Figure 5-3. Level control with outlet flow rate manipulated.

In this case we assume that the valve is fail-open (or air-to-close), since we desire that the tank
not overflow in the event of a control valve (or instrument air) failure. Now we rationalize the
adjustment of the tuning parameter. Consider the case where the error is positive (height
setpoint is greater than the measured height). This indicates that the flow rate leaving the tank
should be lower and that the pressure to the valve should increase. The proportional gain, kc,
must also be positive for the system in Figure 5-3.

Relationship between kpkv and kc

Recall that the process gain is simply the steady-state change in output for a steady-state change
in process input.

For the first strategy, the process gain is positive; that is, an increase in F 1 causes an increase in
h.
Since the input to the valve is the valve-top pressure (Pv1) and the output of the valve is flow
rate, the valve gain is

The overall gain between valve-top pressure and tank height is then

For the second strategy, the process gain is negative (an increase in F2 causes a decrease in h),

and the valve gain is also negative. The gain between valve-top pressure and tank height is

Notice that a process with a positive gain (kp1kv1 or kp2kv2) requires a controller proportional gain
that is also positive. Similarly, a process with a negative gain would require a controller
proportional gain that is also negative.

Valve Gains
Most control valves are designed to operate with signals between 3 and 15 psig. A fail-closed
valve will be fully closed at 3 psig and fully open at 15 psig. For example, if the maximum flow
rate through the valve is 120 gallons per minute (gpm), then the average valve gain is

Similarly, a fail-open valve of the same size would have a negative gain (-10 gpm/psig).

[ Team LiB ]
[ Team LiB ]

5.2 Development of Control Block Diagrams

Control block diagrams are used to analyze the dynamic effect of feedback control loops. All
dynamic elements in a control loop are combined, using their Laplace transfer function
representation. In this section we use the control strategy from Figure 5-1 (where the inlet flow
rate is manipulated) for illustration. Block diagrams are based on Laplace domain signals, which
are assumed to be in deviation variable (perturbations from steady state) form. The block
diagram for Figure 5-1 has the following components:

Level controller (relates error to controller output)

Valve (relates controller output signal to flow-through valve)

Process (relates manipulated input to process output)

Disturbance (relates the disturbance input to the process output)

Sensor (measures tank level)

Controller Transfer Function

Notice first that the level controller compares the measured tank level (hm) to the desired tank
level (hs p). The desired tank level is known as the setpoint, and the difference between the
setpoint and the measured process output is the error.

From the proportional control law (algorithm) for this system [Equation (5.2)], realizing that b =
Pv1s (steady-state pressure to valve), we find the controller input-output relationship

where the controller input signal is the error, which is the difference between the setpoint and
measured process output. We write the equation [where c(s) is the controller output, the
pressure to the valve]

which is the transfer function form for a proportional-only controller.

Equation 5.5

A block diagram for the controller is shown in Figure 5-4, where gc(s) = kc. For convenience when
analyzing block diagrams, the comparator is shown outside the controller transfer function block.
The block diagram uses r(s) to represent the setpoint, and ym(s) to represent the measured
process output.

Figure 5-4. Block diagram relationship for controller transfer function.

The orientation for the transfer function representation of the
controller is slightly different from that of the physical system.

Valve Transfer Function

The block diagram for the valve is shown in Figure 5-5. The input signal is the pressure to the
valve top and the output is the flow rate of fluid through the valve.

Figure 5-5. Block diagram of valve and the physical system.

Process Transfer Function

This process block diagram is shown in Figure 5-6. The input to the block is the flow rate to the
tank, and the output is the tank level.

Figure 5-6. Block diagram of the manipulated input effect on the

process.
Disturbance Transfer Function
This disturbance (outlet flow rate) block diagram is shown in Figure 5-7. The disturbance input is
the flow rate from the tank (which may be due to changes in a downstream control valve not
shown), and the output is the tank level.

Figure 5-7. Block diagram of the disturbance input effect on the

process.

Measurement (Sensor) Transfer Function

This measurement block diagram is shown in Figure 5-8. The input to the block is the tank level
and the output is the actual measurement of tank level.

Figure 5-8. Block diagram of sensor.

Control Block Diagram

We can combine Figures 5-4 to 5-8 to obtain the feedback control system block diagram shown in
Figure 5-9. This is similar to the conceptual block diagram developed in Chapter 1 (Figure 1-4).
There are two externally supplied signals, the setpoint (desired tank height) and the load
disturbance (flow rate to downstream process).

Figure 5-9. Control system block diagram.

We often assume that the output variable can be perfectly measured, and that process input
(usually a flow rate) is directly manipulated; in this case we do not include the valve and
measurement transfer functions in the closed-loop block diagram. Equivalently, we can "lump" the
valve and measurement dynamics into the process transfer function, again allowing us to neglect
the valve and measurement transfer functions. Similarly, the measurement device can be lumped
into the disturbance transfer function.

[ Team LiB ]
[ Team LiB ]

5.3 Response to Setpoint Changes

A simplified [assuming gm(s) = gv(s) = 1] feedback block diagram is shown in Figure 5-10. This
simplification is equivalent to lumping the measurement and valve dynamics into the process
transfer function. By convention, we denote the output of the controller block as u(s), which is
also the manipulated input. Notice that we are focusing on setpoint changes only, so we have not
included the disturbance block. In Section 5.7 we will focus on disturbances.

Figure 5-10. Simplified control block diagram. Disturbances are

neglected and valve and measurement dynamics are lumped into the
process transfer function.

One objective of block diagram analysis is to find the output response to a setpoint change. A
critical aspect is to determine the closed-loop stability. The key is to use block diagram
manipulation to find the relationship between the setpoint and the output.

In Figure 5-11, we have absorbed the closed-loop diagram of Figure 5-10 into a single transfer
function block, to provide the closed-loop relationship between setpoint and process output.

Figure 5-11. External view of Figure 5-10.

Our goal is to obtain the output y(s) as a function of the setpoint, r(s). Here, we combine the
process input-output relationship, y(s) = gp(s) u(s), with controller relationship, u(s) = gc(s) e(s),
to find

Also, since the error is defined as e(s)= r(s) - y(s), we can write
which can be solved to find

We refer to the relationship between r(s) and y(s) as the closed-loop transfer function, gCL (s),

Equation 5.6

If all of the poles of gCL (s) are stable, then the closed-loop system is stable. The denominator of
gCL (s) is also known as the characteristic equation.

Equation (5.6) has been derived for the simple diagram shown in Figure 5-10. Please realize that
more-complex block diagrams will have more-complex closed-loop transfer functions. You should
be able to derive the closed-loop transfer function for any block diagram. For example the closed-
loop transfer function for Figure 5-9 (neglecting disturbances) is

Possible Problems with Offset Using Proportional Controllers

A problem with proportional controllers is that there is generally offset when a setpoint change is
made. By offset, we mean that the actual process output will not be able to achieve the desired
setpoint change. For simplicity, consider a first-order process, gp(s) = kp/(tps + 1), with a
proportional controller, gc(s) = kc, which leads to the following closed-loop transfer function

We recognize that this is a first-order transfer function, where kCL = kpkc/(1 + kpkc)and tCL =
tp/(1 kpkc). For closed-loop stability, the requirement is that tCL > 0 or tp/(1 + kpkc) > 0. Since tp
> 0 (the process is open-loop stable), we have the requirement that 1 + kpkc > 0 or
for stability. Now consider a step setpoint change of magnitude R. You should be able show that
the output response will be

Notice that if kCL is not exactly equal to 1, there will be "offset" in the closed-loop response, as
shown in Figure 5-12. That is, the process output will not exactly match the desired setpoint,
even in the limit as time goes to infinity. Also, as a practical matter, it is important that kpkc > 0.
Notice that if –1 < kpkc < 0, then kCL is negative; although the closed-loop is stable, the process
output actually moves in the opposite direction as the setpoint change.

Figure 5-12. Illustration of offset in system with a proportional

controller.

Offset can also be shown by the final value theorem of Laplace transforms.

or the fractional offset is 1/(1 + kpkc). For a given kp, the offset will be reduced as kc gets larger
and larger. Also, an increase in kc decreases tCL , speeding up the closed-loop response.

Example 5.1: Offset with Proportional (P) Control of a First-Order

Process
Consider the process transfer function for a stirred-tank heater, where the output is temperature
(°C), the manipulated input is heater power (kW), and the timescale is minutes.
Assume a unit step temperature setpoint change [r(s) = 1/s]. The stability condition was kpkc >
–1, so the closed-loop should be stable as long as kc > –1. Closed-loop responses for various
values of kc (-0.5, 1, 5, and 10 kW/°C) are shown in Figure 5-13. As expected, increasing kc
speeds up the response and reduces the offset. Also, it is obvious that the controller gain should
be greater than zero (that is, it should be the same sign as the process gain), otherwise the
response is in the opposite direction; although this may be stable, it is certainly not desirable.

Figure 5-13. Closed-loop response of a first-order system under

proportional-only control. Effect of proportional gain on offset, speed
of response, and manipulated input action.

Thus far we have not considered the effect of control action. It is easy to derive the following
relationship between r(s) and u(s) from Figure 5-10:

For this example,

which we recognize is a lead-lag response. The control (manipulated) variable response for this
example is also shown in Figure 5-13 for various values of kc. As kc increases, the magnitude and
speed of the manipulated variable response increases. For a particular physical system, too high
of a value of kc may cause the manipulated variable to hit a constraint. Consider a step setpoint
change of magnitude R [that is r(s) = R/s]. From the definition of a proportional controller, or
from the initial value theorem, we can show that

If Rmax represents a maximum setpoint change and umax represents a maximum allowable
manipulated variable change, then clearly kc must be less umax/Rmax to avoid manipulated
variable saturation.

Notice that increasing the controller gain reduces the offset and speeds up the response. In the
limit of an infinite controller gain, there is no offset and the dynamic closed-loop response is
instantaneous; a clear limitation is that the manipulated input may hit a constraint (in this
example, the heater power will be limited to some maximum value). In practice, there are time
delays or other model uncertainties and that may cause the closed-loop to be unstable with high
controller gains. It is necessary, then, to devise other control strategies to remove offset. These
strategies are detailed in Section 5.4.

Please note that there is one type of process where proportional control does not yield offset
when a setpoint change is made. These are integrating processes, where one pole of the process
has a value of zero; for example,

The reader should use the final value theorem of Laplace transforms to show that this process will
not have offset when a step setpoint change is made, under proportional control.

[ Team LiB ]
[ Team LiB ]

5.4 PID Controller Algorithms

Here we continue to work in deviation variable form. Recall that a proportional-only controller
could lead to offset between the desired setpoint and the actual output. This happened because
the controller output (process input) and process output came to new equilibrium values before
the error went to zero. The next step is to add a term where the controller output is proportional
to the integral of the error, in addition to the term that is proportional to the error. This is known
as proportional-integral (PI) control. As long as there continues to be an error signal to the
controller, the controller output (process input) will continue to change. Therefore, the integral of
the error forces the error signal to zero. Notice that this type of controller implicitly accounts for
the past system response. A proportional-only controller does not account for this "process
history."

PI Control
A PI controller has two terms (and two tuning parameters), one proportional to the error (kc) and
the other proportional to the integral of the error (kI).

This is more commonly written in the form

Equation 5.7

where tI is known as the integral time. Both the proportional gain and integral time are tuning
parameters that can be adjusted by the control instrument technician or process operator. Notice
that kI = kc/tI. The transform of an integral is

Taking the Laplace transform of Equation (5.7), we find as representation for a PI controller

Equation 5.8
and the controller transfer function for PI control is

Equation 5.9

The following example is an application of PI control to a first-order process.

Example 5.2: First-Order Process with a PI Controller

A first-order process with a PI controller has the closed-loop transfer function

which can be rearranged to the form

Notice that there is no offset, because the closed-loop gain is 1. The quadratic formula can be
used to show that for a second-order characteristic polynomial, if all the coefficients are positive,
the roots are negative (stable). In this case, we have the conditions

Assuming that the process is open-loop stable (tp > 0) and realizing that tI > 0 by definition, we
see that from the first condition,
that is, kp and kc must be the same sign. The second condition requires that 1 + kpkc > 0, or

which is less restrictive than the first condition. The first condition, kpkc > 0, then must be met for
closed-loop stability of this system.

We must realize that there are two controller tuning parameters, kc and tI, that can be adjusted
to give a desired closed-loop response. The process parameters are not available for adjustment,
since they are physical parameters based on the process under consideration.

We have already found the conditions for stability but have not discussed the expected dynamic
behavior of a stable closed-loop system. If some poles of the closed-loop system are complex, the
response will consist of damped oscillations (underdamped), and if all the poles are real, the
response will be critically damped or overdamped. The reader has the opportunity (see Exercise
16) to show that the closed-loop behavior will be underdamped if the following condition is
satisfied

The reader should also show that if the integral time, tI, is equal to the process time constant, tp,
the closed-loop response will be first order with a closed-loop time constant of tp/(kpkc). These
results can be illustrated by simulation for a first-order process with the following parameter
values (see Example 5.1)

Simulations for a setpoint change of 1°C are shown in Figure 5-14, for tI = 0.25 minutes and tI =
5 minutes. In both cases the proportional gain is kc = 5 kW/°C. For the smaller integral time
there is oscillatory performance in both the controlled output (temperature) and manipulated
input (heater power). This oscillatory performance is often viewed as unfavorable to process
operators. An oscillatory closed-loop system is usually more sensitive to parameter uncertainty;
that is, if the process operating conditions changed the closed-loop system could go unstable.

Figure 5-14. Closed-loop response of a first-order system under PI-

control. Effect of integral time on offset and speed of response. The
proportional gain, kc, is 5 kW/°C in both cases.
In Example 5.2, a first-order process with PI control, we showed that the controller gain and the
process gain must be the same sign for closed-loop stability. This result holds true for any open-
loop stable process with integral control. Just remember that the requirement kpkc > 0 is
necessary and sufficient for stability of PI control of a first-order system only, because the closed-
loop transfer function is second-order and positive coefficients for the second-order transfer
function are necessary and sufficient for stability. For higher order, open-loop stable systems,
kpkc > 0 is necessary for closed-loop stability but not sufficient. The Routh array (Section 5.5)
must be checked for the sufficient conditions.

PID Control
The next level of controller complexity is to add a term that accounts for the current rate of
change (derivative) of the error. Using knowledge of the derivative of the error allows the
controller to "predict" where the future error is heading and compensate for it. The time-domain
representation of PID control is

Equation 5.10

which is often written (kD = kctD)

Equation 5.11
where tD is called the derivative time. In the Laplace domain, Equation (5.11) is written

Equation 5.12

Equation 5.13

It turns out that the "ideal" PID algorithm is not physically realizable, that is, no instrument can
take a perfect derivative, so most "practical" PID algorithms have one of the following forms

Equation 5.14a

or

Equation 5.14b

where tF + atD. A typical value for a is 0.1 or less. Even if the ideal PID controller of Equation
(5.13) could be implemented, it would not be desirable. A step setpoint change, for example,
would cause the derivative of the error to be infinity and cause a spike in the manipulated
variable; this behavior is known as derivative "kick." For this reason the derivative action is often
based on the process output rather than the error. That is, Equation (5.11) is implemented as

Equation 5.15
The transfer function representation is

Equation 5.16

Again, since an ideal derivative cannot be implemented, the following filtered derivative of the
output can be used

Equation 5.17

Proportional Band
We have been using a controller formulation based on proportional gain. Some industrial
controllers use proportional band, rather than proportional gain. The proportional band is the
range of error that causes the controller output (manipulated input) to change over its full range.
The proportional band is related to the proportional gain by

Equation 5.18

as shown in Module 15.

[ Team LiB ]
[ Team LiB ]

5.5 Routh Stability Criterion

The stability of the system is determined from the values of the roots (poles) of its characteristic
equation. Finding roots is easy for first- and second-order equations (and not too hard for third)
since there is an analytical solution for the roots of polynomials through third order. If the
polynomial is fourth order or higher, the roots must be determined numerically. There is a method
for determining whether any of the roots are positive (unstable) without actually calculating the
roots (Routh, 1905). This method involves an analysis of the coefficients of the characteristic
polynomial by setting up the Routh array. The test of the coefficients in the Routh array is called
the Routh stability criterion.

The Routh (pronounced like truth) stability criterion is based on a polynomial equation that has
the following form

The necessary condition for all the roots to be negative is that all ai > 0 (you can multiply all
coefficients by –1 if needed). The sufficiency test requires the Routh array

where the elements of rows 3 and higher have the form

If all the coefficients in the first column of the Routh array are positive, then the sufficient
condition for stability is satisfied. If all the coefficients are not positive, then we can determine the
number of unstable (positive) roots by the number of changes in the sign of the coefficients as we
move down the first column. An example of the use of the Routh array to determine limits of a
tuning parameter for closed-loop stability is shown next.

Example 5.3: Third-Order Process with a P-Only Controller

Find the upper bound on proportional gain (for P-only control) for closed-loop stability of the
following process

The closed-loop transfer function is

We must use the Routh stability criterion on the characteristic polynomial. The coefficients are

The Routh array is

Inserting our values into the Routh array,

we see from the necessary condition that 1 + kc > 0, or kc > 0. The sufficient condition that must
be checked is b1 > 0

which leads to kc < 10 for stability. The range of tuning parameters for stability is then
As a practical matter, however, the controller gain should be the same sign as the process gain
(recall that in Example 5.1 a controller gain with the wrong sign led to a process output response
in the opposite direction of the setpoint change). Simulation results for a unit step setpoint
change are shown in Figure 5-15. As kc is increased, the closed-loop response becomes more
oscillatory. The closed-loop system will lose stability at kc = 10.

Figure 5-15. Response of the third-order process to a unit setpoint

change at t = 0. Effect of proportional gain. The system loses stability
at kc = 10.

Indeed, in Table 5-1 we show the relationship between kc and the closed-loop poles. Notice that
one pole stays real and becomes more negative (faster) as kc increases. The other two poles
become complex and the magnitude of the imaginary portion becomes larger as kc increases,
indicating that the magnitude of oscillation is increasing. At kc = 10, the complex poles cross from
the left-half plane to the RHP, indicating that the closed-loop system is unstable for kc > 10.

Figure 5-16 is a plot of the roots of the characteristic equation, as a function of kc. It shows
vividly the results presented in Table 5-1. The closed-loop poles start as three real, distinct poles
at kc = 0. Two become complex, and move toward the RHP as kc is increased. When kc > 10, the
two complex poles move into the RHP; at kc = 10, a continuous oscillation is formed.

Figure 5-16. Roots of the characteristic equation, as a function of kc. x

marks the three poles of the open-loop system (kc = 0). This type of
plot is often called a root-locus plot.
[ Team LiB ]
[ Team LiB ]

5.6 Effect of Tuning Parameters

We have shown how to use the closed-loop characteristic equation to determine the closed-loop
stability of processes. Except for P-only control, we have not discussed the effect of controller
tuning parameters on the feedback system performance (closed-loop response).

Our discussion here is based on open-loop stable processes. To better understand the effect of
controller tuning parameters, you should experiment with a number of transfer functions. You can
also determine the closed-loop poles as a function of the tuning parameters, for a particular
process.

Table 5-1. Closed-Loop Poles as a Function of Proportional Gain for

Example 5.3[a]

kc p1 p2 p3
0 –1.0000 –0.5000 –0.3333

0.1000 –1.0432 –0.3951–0.1402j –0.3951+0.1402j

0.2500 –1.0926 –0.3704–0.2313j –0.3704+0.2313j
0.5000 –1.1549 –0.3392–0.3184j –0.3392+0.3184j

1.0000 –1.2452 –0.2940–0.4257j –0.2940+0.4257j

3.0000 –1.4612 –0.1861–0.6493j –0.1861+0.6493j
5.0000 –1.5991 –0.1171–0.7821j –0.1171+0.7821j

7.0000 –1.7054 –0.0640–0.8819j –0.0640+0.8819j

9.0000 –1.7938 –0.0198–0.9637j –0.0198+0.9637j
10.0000 –1.8333 0.0000–1.0000j 0.0000+1.0000j
11.0000 –1.8704 0.0185–1.0339j 0.0185+1.0339j

13.0000 –1.9384 0.0525–1.0959j 0.0525+1.0959j

15.0000 –2.0000 0.0833–1.1517j 0.0833+1.1517j

[a] The poles (p1,p2,p3) are the roots of the characteristic equation, 6s3 + 11s2 + 6s + 1 + kc.

Effect of Controller Gain

An increase in controller gain (kc) will speed up the closed-loop response. Except for first- and
second-order processes with P-only control, too large a value of controller gain will destabilize the
system.
Integral Time
An increase in integral time (tI) tends to "slow down" the closed-loop response, while decreasing
the integral time speeds up the response. Too small an integral time can cause the closed-loop
system to be unstable.

Derivative Time
In this chapter we have not explicitly shown the effect of derivative action—examples are given in
the next chapter. An increase in derivative time (tD) tends to help stabilize the closed-loop
response. A disadvantage of derivative action is that it causes the control system to be sensitive
to measurement noise.

The next chapter presents some standard procedures for controller tuning.

[ Team LiB ]
[ Team LiB ]

5.7 Response to Disturbances

Consider the block diagram in Figure 5-17, which has both setpoint and load disturbance signals.
It is easy to derive the following closed-loop relationship:

Figure 5-17. Control block diagram including a load disturbance.

In our previous analysis we have been assuming that there is no load disturbance entering the
process; that is, our concern has been with setpoint responses. In this section we analyze
disturbance rejection (or the response to load disturbances) in detail.

Assume that there is no setpoint change, that is, r(s) = 0,

Notice that a disturbance does not affect the stability of a system (as long as the disturbance
transfer function itself is stable), because the closed-loop characteristic equation [1 + gp(s)gc(s)]
is the same for setpoint tracking or disturbance rejection.

Example 5.4: First-Order Process and Load Transfer Functions with P-

Only Control
Consider a first-order process and disturbance with proportional control
The effect of the load disturbance on the output is

We see that this system is stable as long as kc is tuned to satisfy 1 + kpkc > 0. Since we are
dealing with deviation variables and we have made no setpoint change, we desire for y to stay as
close to 0 as possible.

Consider a step load disturbance of l(s) = Dl/s. Using the final-value theorem, we find that

which will be a small value as long as the disturbance magnitude (Dl) is low, the disturbance gain
(kd) is low, or the controller gain (kc) is high. Since it is undesirable to have offset, we would use
a PI controller, in practice, to remove the offset (see Exercise 24).

[ Team LiB ]
[ Team LiB ]

5.8 Open-Loop Unstable Systems

The examples presented thus far in this chapter have assumed that the process is open-loop (that
is, without control) stable. For these systems we can always "detune" a controller enough so that
the closed-loop system is stable. Indeed, if we simply let kc = 0, then we have stability because
the closed-loop transfer function is equal to the open-loop transfer function. Open-loop unstable
systems are much tougher to control, because the control must be tuned "tightly" enough to
stabilize the system, yet not so tightly that the closed-loop system will be unstable. Some
exothermic chemical reactors and biochemical reactors are operated at open-loop unstable steady
states (that is, at least one eigenvalue of the A matrix is positive; this is equivalent to at least one
pole of the process transfer function being positive).

Example 5.5: First-Order Open-Loop Unstable Process with P-Only

Control
Consider a first-order open-loop unstable process (see Module 7 on bioreactor control for an
example),

where tu > 0. We use this notation to make it clear that the process is open-loop unstable, since
the pole is positive.

Here again we use a proportional controller

The closed-loop transfer function is

which can be written

Notice that for stability we require that

which means that (since – tu < 0), 1 + kpkc < 0, or

Recall that an open-loop stable first-order process under P-only control (Example 5.1) required
that kpkc > –1 for closed-loop stability (and kpkc > 0 for reasonable performance). For open-loop
stable processes and controllers with integral action, kpkc > 0 is required for stability; this means
that the process gain and controller gain must be the same sign. It is interesting that this open-
loop unstable process needs the controller and process gain to be different signs.

Not all open-loop unstable processes can be stabilized by P-only control. For example, one of the
processes in Exercise 10 cannot be stabilized by P-only control.

[ Team LiB ]
[ Team LiB ]

5.9 SIMULINK Block Diagrams

SIMULINK is a natural environment for simulating closed-loop systems. It is recommended that you
read Module 2, Introduction to SIMULINK, and reproduce the simulation results shown in Example
5.3. Generate the block diagram shown in Figure 5-18 and run the simulations for various
controller proportional gains.

Figure 5-18. SIMULINK diagram for Example 5.3.

Recall that the closed-loop characteristic equation for Example 5.3 is

For kc + 10, the MATLAB roots command can be used to find that the process is on the verge of
instability, with the following values for the poles:

» roots([6 11 6 11])
ans =
-1.8333
-0.0000 + 1.0000i
-0.0000 - 1.0000i

Notice that this analysis can be performed for several values of kc, to construct Table 5-1.

[ Team LiB ]
[ Team LiB ]

5.10 Summary
The objective of this chapter was to provide an introduction to feedback control. We showed how
to convert an instrumentation and control diagram to a control block diagram. Block diagram
manipulations were used to develop a closed-loop transfer function. Transfer function
representations for standard PID controllers were developed. The Routh array was used to test for
the necessary and sufficient conditions for closed-loop stability based on the characteristic
equation of the closed-loop transfer function. We noted that proportional controllers exhibit offset
for setpoint and disturbance changes, motivating the development of controllers with integral
action. Open-loop stable processes generally require that the controller gain be the same sign as
the process gain, while open-loop unstable processes often require that the controller gain be the
opposite sign of the process gain.

It is suggested that you work some of the problems in the student exercises section. It is
particularly important to be able to find the closed-loop relationship (transfer function) between
any two signals on a control block diagram.

What has not been discussed in this chapter is the issue of robustness. We have presented
methods of analysis that assume that the process transfer function is exactly known. In reality, a
process model is only an approximate representation of the dynamic behavior of the process. A
set of tuning parameters that is stable under one operating condition may be unstable under
another. Chapter 7 presents tuning methods that account for uncertainty in process parameters.
Also, the model-based methods developed in Chapters 8 and 9 allow for detuning parameters
based on model parameter uncertainty.

Common terms used in this chapter are as follows:

Characteristic eqn: denominator polynomial for closed-loop analysis

Fail-closed: valve closes on loss of instrument air (air-to-open)

Fail-open: valve opens on loss of instrument air (air-to-close)

Manipulated input: also called the controller output (CO)

Necessary condition: must be satisfied for a chance of stability
Offset: difference between setpoint and long-term process output
Process output: also called the process variable (PV)

Routh stability: stability test based on the closed-loop characteristic equation

Sufficient condition: guarantees stability (if necessary condition is satisfied)

The tuning parameters are as follows:

kc: proportional gain

tI: integral time

tD: derivative time

PB: proportional band (PB + 100/kc)

The block diagram (Laplace domain) variables are as follows:

c(s): controller output

e(s): error

r(s): setpoint (reference signal)

u(s): manipulated variable

y(s): process output

ym(s): measured output

The transfer functions are as follows:

gc(s): controller
gv(s): valve
gp(s): process

gm(s): measurement
gCL (s): closed-loop

The controller types are as follows:

P: proportional
PI: proportional-integral

PID: proportional-integral-derivative

The relationships are as follows:

gc(s) = kc P control

PI control

"Ideal" PID control

"Real" PID control (series)

"Real" PID control (parallel)

Ideal derivative of output rather than error

Real derivative of output rather than error

 Closed-loop transfer function (CLTF)

CLTF simplified

CLCE = 1 + gp(s)gv(s)gc(s)gm(s) Closed-loop characteristic equation

CLCE = 1+ gp(s)gc(s) Simplified CLCE

[ Team LiB ]
[ Team LiB ]

References

Routh, E.J., Dynamics of a System of Rigid Bodies, Part II, MacMillan, London (1905).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider a household heating system with on-off control. Normally there is a dead
band of 2°F; that is, the temperature must drop to 1°F below the setpoint before the
heater kicks on, and it must go 1°F above the setpoint before the heater kicks off.
Clearly the thermostat/heater has periodic behavior with periods where the heater is
on, followed by periods where it is off. Discuss the effect of the dead band on this
periodic behavior. Sketch the expected heater (on-off) and temperature profiles (as a
function of time) as the dead band is changed.

2:
a. Derive the closed-loop transfer functions relating L(s) and r(s) (the disturbance
and setpoint, respectively) to ym(s) (the measured output) for the following
control block diagram. This block diagram is for a system that has significant
measurement and valve dynamics.

Use the following form: ym(s) = gCL (s) r(s) + gCL1(s) L(s) [that is, find gCL (s) and
gCL1(s)].

b. Also derive the closed-loop transfer functions relating L(s) and r(s) to u(s) (the
input to the process).

3: Derive the closed-loop transfer function between L(s) and y(s) for the following control
block diagram (this is known as a feed forward/feedback controller).
 How would you check the stability of the closed-loop system? Will the stability of this
system be any different than that of the standard feedback system? Why?

4: Derive the closed-loop transfer function between r 1(s) and y1(s) for the following
control block diagram (this is known as a cascade control system). Hint: Find a single
block to represent the "inner" feedback loop.

How would you check the stability of the closed-loop system?

5: Consider the following control instrumentation diagram for a heat exchanger.

a.
 a. Construct a control block diagram, labeling all signals and transfer functions.
Make certain that you include valve and measurement sensor dynamics. Also
include at least one important disturbance input.

b. Should the control valve should be fail-open or fail-closed. Why?

c. Based on whether the control valve is fail-open or fail-closed, is the process gain
relating the manipulated valve position (or valve top pressure) to the measured
temperature positive or negative?

6: Consider a first-order open-loop unstable process that has the following transfer
function

Find the range of values of parameters for a PI controller that stabilize this process.

7: Calculate the offset to a step setpoint change due to P-only control, for the following
process:

Do you have an interesting result? Explain.

8: A process has the following transfer function:

Using a P-only controller, find the range of the controller gain that will yield a stable
closed-loop system.

9: Consider the following open-loop unstable process

For a PI controller, find the range of stabilizing controller proportional gains (kc) for an
integral time constant of tI = 2.

10: Consider the two following unstable second-order processes:

 Assume that you can apply a P-only feedback controller to each of these processes.
What are the bounds on kc (the controller proportional gain) that will assure stability of
the closed-loop system for each process?

11: Consider the open-loop unstable process transfer function

a. Find the range of kc for a P-only controller that will stabilize this process.

b. It turns out that kc = 2 will yield a stable closed-loop (if this is not in your range
from part (a), you should check your results) for this system. In practice there is
a measurement lag in the feedback loop. Assuming a first-order lag on the
measurement, find the maximum measurement time constant which is allowed
before the system (with kc = 2) is destabilized.

12: Consider a process with the following transfer function:

Can a PI controller satisfy the necessary condition for stability of this process?

13: Consider the following process with a RHP zero:

For P-only control, find the bound on the proportional gain to assure closed-loop
stability of this process.

14: A PI controller is used on the second-order process

The process parameters are kp = 1, t = 2, z = 0.7. The tuning parameters used are kc
= 5, tI = 0.2.

Is the process closed-loop stable?

15: Consider an isothermal CSTR that has the following reaction scheme
This system exhibits a maximum in the concentration of B vs. dilution rate (F s/V,
where F s is the steady-state flow rate and V is the constant reactor volume) curve
shown in Figure 5-19. In this problem we consider two possible operating points, which
we call case 1 and case 2.

Figure 5-19. Steady-state relationship between input (dilution

rate) and output (concentration of B).

Case 1. The steady-state dilution rate of Fs/V = 4/7 min–1 is on the left side of the
peak shown in Figure 5-19. The transfer function relating the input (dilution rate) to
the output (concentration of B) is

Case 2. The steady-state dilution rate of Fs/V = 2.8744 min–1 is on the right side of
the peak in Figure 5-19. The transfer function relating the input to the output is

a. Draw a process and instrumentation diagram for this system.

b. Find the process gain, poles, and zero for the two transfer functions. Are the
gains consistent with the steady-state input/output curve?

c.
 b.

c. For a P-only controller, find the range of controller gains that will stabilize this
system at both operating points (case 1 and case 2).

16: Show that, for a stable first-order process, a PI controller will have an oscillatory
response if

17: Compare the effect of kc on the closed-loop stability requirements for the following
processes under P-only control:

18: Consider a second order process (without numerator dynamics),

Determine the conditions (range of values) for the PI tuning parameters to assure
closed-loop stability.

19:
It is easy to show for the second-order system that there is no
upper bound on kc for P-only control, so that offset can be minimized by making kc
very large. Now consider the more practical case, where there is a measurement lag.
How low must tm be to assure that the maximum offset for a setpoint change is 5%?
(Hint: Consider the offset of a P-only control system, but also make certain that the
closed-loop system satisfies the Routh stability criterion.)

20: A process is known to have the following input-output transfer function

In the process of tuning a P-only feedback controller for this process, the operator
discovers a steady-state offset of 2/5 for a unit setpoint change, when the proportional
gain is 1. She also finds that the closed-loop system goes unstable when the
proportional gain is greater than 2.

a. What is the value of the process gain, k?

b. What is the value of the numerator time constant, t?

c.
 a.

b.

c. What type of instability occurs when the proportional gain is 2? That is, does the
closed-loop system oscillate owing to simultaneous crossing of two complex poles
across the imaginary axis? Or does a single real pole cross the imaginary axis?

21: An open-loop unstable chemical reactor is known to have a process transfer function
with the following form, where the manipulated input is the coolant flow rate
(liters/minute) and the measured process output is the reactor temperature (°C):

When a P-only controller, with a value kc = -3 lpm/°C, is used, the response to a unit
setpoint change is shown below.
When a PI controller is used, the closed-loop system becomes unstable, as shown
below, when the tuning parameter values are kc = -1.333 lpm/°C and tI = 2 minutes.

What are the values for the process parameters, kp and tu? Hints: Can the final-value
 theorem be used? Can the Routh stability criterion be used?

22: Consider the following chemical process. The concentration of propylene glycol exiting
the adiabatic reactor is measured. The output of the concentration controller is a
current signal, which is converted to a pressure signal. The pressure signal actuates a
control valve that changes the flow rate of steam to a heat exchanger, which changes
the temperature of the reactor inlet stream.

a. Construct the control block diagram, clearly labeling all signals and blocks.

b. Find the closed-loop transfer function from the concentration setpoint to the
measured output.

23: Consider a PID controller with a special form that eliminates "derivative kick":

This controller is implemented in block diagram form as shown below.

Derive the closed-loop transfer function relating a setpoint change, r, to the process
output, y.
 24: Consider Example 5.4. If a PI controller is used, rather than P-only, show that there
will not be any offset to a step load disturbance. Also, show the conditions on the
tuning parameters for closed-loop stability.

[ Team LiB ]
[ Team LiB ]

Chapter 6. PID Controller Tuning

In Chapter 5 we were given a process transfer function and a controller transfer function (P, PI,
PID, etc.) and asked to find the closed-loop response (or stability). In this chapter we provide an
overview of classical methods for tuning PID controllers. We also provide an introduction to
model-based techniques, using the direct synthesis approach.

After studying this chapter, the reader should be able to do the following:

Understand the different forms of a PID controller

Tune PID controllers using the classical, Ziegler-Nichols, and Cohen-Coon methods

Derive controllers based on a process model and desired closed-loop response

The major sections of this chapter are as follows:

6.1 Introduction

6.2 Closed-Loop Oscillation-Based Tuning

6.3 Tuning Rules for First-Order + Dead Time Processes

6.4 Direct Synthesis

6.5 Summary

[ Team LiB ]
[ Team LiB ]

6.1 Introduction
Consider the standard feedback block diagram shown in Figure 6-1, where disturbance inputs
have been neglected. Transfer functions (and block diagrams) are used to analyze the behavior of
control systems, because the algebraic expressions are easy to manipulate.

Figure 6-1. Feedback control block diagram.

We noted in Chapter 5 that the closed-loop transfer function could be used to determine, for
example, the range of controller gains that assure closed-loop stability.

PID Controller Forms

PID controller algorithms were developed in Chapter 5. Here we provide a concise review of the
algorithms in common use.

P-Only Control

The proportional only algorithm is

Equation 6.1a

which has the following transfer function relationship between error and controller output:

Equation 6.1b
PI-Control

The PI algorithm is

Equation 6.2a

which has the following transfer function relationship between error and controller output:

Equation 6.2b

PID Control

The ideal PID algorithm is

Equation 6.3a

which has the following transfer function relationship between error and controller output:

Equation 6.3b

In practice it is impossible to perfectly differentiate the error signal, so the following Laplace
transfer function approximations are often used for "real" PID control (where a = 0.1 is common):

Equation 6.4a

or
Equation 6.4b

A problem with taking the derivative of the error is that step setpoint changes cause the
derivative to become unbounded and result in a "spike" in the manipulated variable action. In
most practical PID controllers, then, the derivative of the measured process output is used.

Equation 6.5a

The transfer function representation is

Equation 6.5b

which is usually implemented in the form of

Equation 6.5c

where the measured process output has been "filtered" to minimize noise problems.

[ Team LiB ]
[ Team LiB ]

6.2 Closed-Loop Oscillation-Based Tuning

A PID controller has three tuning parameters. If these are adjusted in an ad hoc fashion, it may
take a while for satisfactory performance to be obtained. Also, each tuning technician will end up
with a different set of tuning parameters. There is plenty of motivation, then, to develop an
algorithmic approach to controller tuning. The first widely used method for PID tuning was
published by Ziegler and Nichols in 1942.

Ziegler-Nichols Closed-Loop Method

The Ziegler-Nichols closed-loop tuning technique was perhaps the first rigorous method to tune
PID controllers. The technique is not widely used today because the closed-loop behavior tends to
be oscillatory and sensitive to uncertainty. We study the technique for historical reasons, and
because it is similar to commonly used automatic tuning ("autotune") techniques covered in
Chapter 11.

The closed-loop Ziegler-Nichols method consists of the following steps.

1. With P-only closed-loop control, increase the magnitude of the proportional gain until the
closed-loop is in a continuous oscillation. For slightly larger values of controller gain, the
closed-loop system is unstable, while for slightly lower values the system is stable.

2. The value of controller proportional gain that causes the continuous oscillation is called the
critical (or ultimate) gain, kcu . The peak-to-peak period (time between successive peaks in
the continuously oscillating process output) is called the critical (or ultimate) period, Pu.

3. Depending on the controller chosen, P, PI, or PID, use the values in Table 6-1 for the tuning
parameters, based on the critical gain and period.

Tyreus and Luyben have suggested tuning parameter rules that result in less oscillatory responses
and that are less sensitive to changes in the process condition. Their rules are shown in Table 6-
2.

Table 6-1. Ziegler-Nichols Closed-Loop Oscillation Method Tuning

Parameters

Controller type kc tI tD

P-only 0.5 kcu — —

PI 0.45 kcu Pu/1.2 —
PID 0.6 kcu Pu/2 Pu/8

Table 6-2. Tyreus-Luyben Suggested Tuning Parameters Based on the

 Ziegler-Nichols Closed-Loop Oscillation Tuning Method

Controller type kc tI tD
PI kcu /3.2 2.2 Pu —
PID kcu /2.2 2.2 Pu Pu/6.3

Example 6.1: Third-Order Process

Consider the third-order system used in Example 5.3, with a time unit of minutes

Recall that a P-only controller caused the closed-loop to be on the verge of instability (continuous
oscillation) when the value of the controller gain was kc = 10. This is the critical proportional gain,
kcu . From the response shown in Figure 6-2 we see that the ultimate period (period of oscillation)
is Pu = 6.2 minutes.

Figure 6-2. Response to unit step setpoint change at t = 0; kcu = 10.

Notice that Pu = 6.2 minutes.

The closed-loop responses for Ziegler-Nichols tuning for P, PI, and ideal PID controllers (based on
Table 6-1) are shown in Figure 6-3. Notice that a P-only controller has offset, as expected. Also,
all the responses are quite oscillatory; this is one of the major disadvantages to the Ziegler-
Nichols tuning method. It typically results in more oscillatory behavior than would be allowable in
a typical process plant. The tuning parameters are also not very robust, that is, they are very
sensitive to process uncertainty. If the process conditions change, then the control system may
become unstable.

Figure 6-3. Response to unit step setpoint change. Comparison of

Ziegler-Nichols P, PI, and PID tuning rules. Notice the "spike" in the
manipulated input for the PID controller.

The closed-loop responses for Tyreus-Luyben tuning for P (assumed to be the Ziegler-Nichols
value), PI, and PID controllers are compared in Figure 6-4. The Tyreus-Luyben parameters result
in less oscillatory responses and will be less sensitive to uncertainty.

Figure 6-4. Response to unit step setpoint change. Comparison of

Tyreus-Luyben P, PI, and PID tuning rules. Notice the "spike" in the
manipulated input for the PID controller.
Note that the PID controllers simulated in Figures 6-3 and 6-4 are based on an ideal derivative of
the error. A problem is that this results in a "spike" in the manipulated input when a step setpoint
change is made. Also, an ideal derivative controller will be more sensitive to measurement noise.

[ Team LiB ]
[ Team LiB ]

6.3 Tuning Rules for First-Order + Dead Time Processes

The previous tuning rules were based on tests that forced a process into a continuous oscillation.
Obvious disadvantages to the techniques are that the system is forced to the edge of instability,
and it may take a while to iteratively adjust the controller to obtain a continuous oscillation. In
this section we present tuning rules based on process models that have been obtained through
the open-loop step tests presented in Chapter 4.

Ziegler-Nichols Open-Loop Method

Ziegler and Nichols also proposed tuning parameters for a process that has been identified as
integrator + time-delay based on an open-loop process step response,

Since first-order + time-delay processes have a maximum slope of k = kp/tp at t = q for a unit
step input change, these same rules can be used for first-order + time-delay processes,

Their recommended tuning parameters, which should give roughly quarter-wave damping, are
shown in Table 6-3. We see a potential problem for systems with a low time-delay/time-constant
ratio, since this causes the proportional gain to become very large. Similarly, the integral time
tends to be low, causing oscillatory behavior.

Cohen-Coon Parameters
The method developed by Cohen and Coon (1953) is based on a first-order + time-delay process
model. A set of tuning parameters was empirically developed to yield a closed-loop response with
a decay ratio of 1/4 (similar to the Ziegler-Nichols methods). The tuning parameters as a function
of the model parameters are shown in Table 6-4.

A major problem with the Cohen-Coon parameters is that they tend not to be very robust; that is,
a small change in the process parameters can cause the closed-loop system to become unstable.

Table 6-3. Ziegler-Nichols Open-Loop Tuning Parameters

 Controller type kc tI tD

P-only — —

PI 3.3q —

PID 2q 0.5q

Table 6-4. Cohen-Coon Tuning Parameters

Controller type kc tI tD

P-only — —

PI —

PID

[ Team LiB ]
[ Team LiB ]

6.4 Direct Synthesis

Consider the standard feedback block diagram shown in Figure 6-1. Recall that we could
determine the performance or stability of the closed-loop system from the closed-loop transfer
function

Equation 6.6

In the direct synthesis procedure, we select a desired closed-loop response, gCL (s), and, based on
the known process, gp(s), find the controller gc(s) that will yield this response. Solving Equation
(6.6) for gc(s), we obtain

Equation 6.7

By now you should be able to perform block diagram manipulations to find the relationship
between any external signal [such as a setpoint change, r(s)] and any other signal on the control
block diagram. For example, it is important to analyze the manipulated variable action required
for a setpoint change to make certain that it is not too rapid or that it does not violate
constraints. From Figure 6-1 we can easily derive the effect of the setpoint change on the
manipulated variable action,

Equation 6.8

The direct synthesis procedure then consists of specifying the desired closed-loop transfer
function (e.g., first-order response, second-order underdamped, etc.), gCL (s), using Equation
(6.7) to find the feedback controller, and considering the manipulated variable response (usually
tested by simulation).

The real question is: How do we specify the desired closed-loop response?

It turns out that we are not limited by the desired closed-loop response, if the system is minimum
phase (the process does not have RHP zeros or time delays—the terminology becomes clear in
Chapter 7). In the following section we present the direct synthesis method for minimum-phase
systems, and cover non-minimum-phase systems in the subsequent section.
Direct Synthesis for Minimum-Phase Processes
It seems fairly natural to specify a desired closed-loop response that is first order, since we
understand the characteristics of a first-order response.

Equation 6.9

For a specified first-order response, there is only one tuning parameter, l, since we desire a
closed-loop gain of 1 (we want the process output to equal the setpoint as the closed-loop system
goes to a new steady state); small values of l results in fast responses, while large values result
in slow responses. One could also argue for a desired closed-loop response that is second order
and underdamped, which would lead to specifying two parameters.

Example 6.2: Direct Synthesis For a First-Order Process

Consider a first-order process

Equation 6.10

Assume that a first-order closed-loop response is specified. If we desire a fast closed-loop

response, we make l small; for a slower (more "robust") response, we make l large. Solving for
gc(s) from Equation (6.7), we find

Equation 6.11

which can be written (by multiplying by tp/tp)

Equation 6.12

Recall that the form of a PI controller is

Equation 6.13

so that our direct synthesis controller for a first-order process is simply a PI controller, where

Notice that there was only one "tuning parameter" for this direct synthesis example. The desired
closed-loop time constant, l, was the only adjustable parameter. Given the first-order transfer
function parameters (kp and tp) and the desired closed-loop time constant (l), we found that the
direct synthesis controller was PI, but that we only needed to "adjust" one PI parameter (kc). This
is a nice result, because once we find the process time constant (tp), we can set the integral time
(tI) equal to the time constant, and tune kc on-line until we achieve a desired response. Tuning a
single controller parameter is much easier than tuning two or three.

Numerical Example

Consider the following first-order process, with a time constant of 10 minutes and a process gain
of 2 %/%

The output and manipulated variable responses for l = 1, 5, and 10 min (tI = 10; kc = 5, 1, and
0.5, respectively) are shown in Figure 6-5. As expected, the manipulated variable response for
the faster desired closed-loop time constant is much more aggressive than for the slower closed-
loop time constants. Notice that setting the closed-loop time constant equal to the open-loop time
constant (l = tp = 10) results in a single step change in the manipulated variable action.

Figure 6-5. Response to a unit step setpoint change.

The same type of procedure shown in Example 6.2 can be used if a second-order closed-loop
response is specified. See Exercise 13.

Direct Synthesis for Nonminimum-Phase Processes

This section presents examples for nonminimum-phase processes, that is, processes that have
time delays or RHP zeros. The general technique remains the same for these processes; however,
there is a restriction on the type of closed-loop response that can be specified. The next example
provides the motivation for specifying different desired closed-loop responses for systems with
time delays.

Example 6.3: First-Order + Dead Time Example

Consider the following process transfer function:

If the desired closed-loop response is first order, gCL (s) = 1/ls + 1, the resulting controller is

which is a PI controller with an additional term (e3s). This additional term is not physically
realizable because it requires knowledge of future errors to obtain the current control action. This
is clearly impossible. Perhaps this is shown more clearly in the time domain

where the control action at time t depends on the error at time t + 3, which is clearly impossible
to implement.

The next example shows how a system with a RHP zero must have a modified desired closed-loop
response.

Example 6.4: Process with a RHP Zero

Consider a process with the following transfer function, where the time unit is minutes

Equation 6.14

The direct synthesis procedure with gCL (s) = 1/ls + 1 yields the controller

which is unstable because of the RHP pole. The RHP pole in the controller is due to the inversion
of the process zero. This inversion occurs because of the specification of a first-order closed-loop
response. The output and manipulated variable responses are shown in Figure 6-6 for l = 5
minutes. Notice that the inverse response does not appear in the output variable but that the
manipulated variable is unbounded (unstable). This is often called internal instability.

Figure 6-6. Response to a unit step setpoint change when the

controller is unstable.
Clearly, the manipulated variable for any physical system will eventually hit a constraint, making
the closed-loop system effectively open-loop, since the controller is no longer functioning.

It is easy to understand the behavior shown in Figure 6-6. On a short timescale, the process
appears to the controller to have a negative gain, since a step change in the manipulated variable
yields an output that initially dips before going in the positive direction. The controller is acting on
this effective "negative gain" to continually force the manipulated variable in a negative direction.
The reader should verify that once the manipulated variable hits a lower bound, the process
output will begin to decrease.

Specifying a desired first-order closed-loop response for a system with a RHP zero resulted in an
unstable controller and unbounded manipulated variable action. A stable controller can be
obtained if the desired closed-loop response has the same RHP zero as the process.

Reformulation of the Desired Response

Including the right-half-plane zero desired closed-loop transfer function specified as

using the direct synthesis procedure, you should find the following controller,
which is an ideal PID controller cascaded with a first-order filter. The parameters are

The output and input responses for a step setpoint change with l = 5 minutes are shown in Figure
6-7. The closed-loop system has inverse response (as specified by the closed-loop transfer
function) and reasonable manipulated variable action.

Figure 6-7. Responses to a unit step setpoint change at t = 0 minute,

when RHP zero is in the desired response (l = 5).

[ Team LiB ]
[ Team LiB ]

6.5 Summary
The closed-loop Ziegler-Nichols method (Table 6-1) was shown to lead to oscillatory closed-loop
behavior, which is one of the major disadvantages to the approach. A major advantage to the
approach is that a process model is not needed. Tyreus-Luyben parameters (Table 6-2) were
shown to be less oscillatory and are generally recommended over the Ziegler-Nichols parameters.

The open-loop Ziegler-Nichols (Table 6-3) and the Cohen-Coon (Table 6-4) methods also tend to
lead to oscillatory closed-loop behavior; we have covered these approaches primarily for historical
reasons.

In the direct synthesis design procedure, a desired closed-loop response (or closed-loop transfer
function) is specified. A feedback controller is then synthesized to obtain that response. We have
seen that a PID-type controller often results. Some problems can arise if the process has a time
delay or inverse response. In these cases, the desired closed-loop response must also have a time
delay or inverse response. We develop a more transparent method for designing controllers for
first-order + time-delay processes in Chapters 8 and 9.

[ Team LiB ]
[ Team LiB ]

References

The original Ziegler-Nichols parameters are developed in the following paper: Ziegler, J. G., and
N. B. Nichols, "Optimum Settings for Automatic Controllers," Trans. ASME, 64, 759–768 (1942).

The Cohen-Coon parameters are developed in the following paper: Cohen, G. H., and G. A. Coon,
"Theoretical Considerations of Retarded Control," Trans. ASME, 75, 827 (1953).

The Tyreus-Luyben parameters are presented in the following textbook: Luyben, M. L., and W. L.
Luyben, Essentials of Process Control, McGraw-Hill, New York (1997).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider the gas pressure problem shown below. The objective of this problem is to
understand (via simulation) how the tuning parameters for a PI controller affect the
stability and speed of response for setpoint changes or disturbances.

The relationship between the manipulated valve position (u in deviation variables) and
the pressure in the second tank (y in deviation variables) is (with a time unit of
minutes)

Assume that the dynamic behavior of the pressure measurement/transmitter is

characterized by a first-order lag with a time constant of 9 seconds:

Also assume that the dynamic behavior of the control valve is also characterized by a
first-order lag with a time constant of 6 seconds,

where c(s) is the output from the controller and u(s) is the valve position. For P-only
control, find the value of kc (via simulation) that causes a closed-loop to go unstable.
Call this value kcu , and call Pu the period of oscillation (time between peaks) when the
system goes unstable. These values are used in the Ziegler-Nichols closed-loop
oscillation method.

a.
 a. Show that the value of kc sightly greater than that you obtained causes at least
one root of the closed-loop characteristic equation [gCL (s)] to be positive. Find
the P and PI tuning parameters on the Ziegler-Nichols closed-loop oscillations
method.

b. Compare the response of the two different controllers (P vs. PI), for step setpoint
changes of 1 psig in the desired output (y).

The closed-loop block diagram is shown below.

2: Consider the following first-order + time-delay process:

Perform simulations for the process output and manipulated input for unit step
changes in the setpoint.

a. Compare the closed-loop step response of this process using P-only control based
on (i) Ziegler-Nichols closed-loop oscillations method, (ii) Ziegler-Nichols open-
loop method, (iii) Cohen-Coon, and (iv) Tyreus-Luyben tuning.

b. Compare the closed-loop step response of this process using PI control based on
(i) Ziegler-Nichols closed-loop method, (ii) Ziegler-Nichols open-loop method, (iii)
Cohen-Coon, and (iv) Tyreus-Luyben tuning.

c. Compare the closed-loop step response of this process using PID control based
on (i) Ziegler-Nichols closed-loop method, (ii) Ziegler-Nichols open-loop method,
(iii) Cohen-Coon, and (iv) Tyreus-Luyben tuning.

Comment on the results for all of these tuning methods.

3: Consider the process transfer function for the Van de Vusse reactor (Module 5).
 Find the Ziegler-Nichols controller parameters for P, PI, and PID controllers for this
process, based on the closed-loop oscillation method. Compare the responses of all
three controllers to a step setpoint change.

4: Most PID controller design procedures assume that a perfect derivative controller is
used. For the process transfer function used in problem 3 above, and the Ziegler-
Nichols closed-loop method, compare the responses of (i) ideal PID, (ii) real PID, and
(iii) PID with ideal derivative action on the process output, rather than the error.

5: Show that the following state space representation of a controller,

has the following transfer function representation:

where

6: Consider the PID algorithm

where yf is a "filtered" value of the process output. Assuming that a first-order filter is
used, with a time constant of tf , write the modeling equations (differential and
algebraic) to simulate the behavior of this controller.

7: Apply (simulate) the Tyreus-Luyben parameters for PI and PID controllers to the
following process

Compare these results with Cohen-Coon. Which do you recommend for implementation
on a real process?
 8: Find the feedback controller for an integrating process, gp(s) = kp/s, assuming a
desired first-order response using the direct synthesis method.

Answer: It is a P-only controller, with kc = 1/kpl

9: Find the feedback controller for an integrating process, gp(s) = kp/s, assuming a

desired second-order response

Partial Answer: The controller is a first-order lag.

10: Show that the direct synthesis procedure for the following process, assuming a desired
first-order response, yields a PID controller

Find the PID tuning parameters if a closed-loop time constant of 5 minutes is desired.

11: For a second-order system with numerator dynamics,

find a controller that gives a first-order closed-loop response. (Hint: It will be a PID
with a first-order lag.)

12: Consider the following first-order process:

If the desired closed-loop response to a setpoint change is second order with the
following form,

find the feedback controller required, where a and l are adjustable tuning parameters
(they are both positive). What type of controller is this? If the controller is PID form
(perhaps with a lag), find each of the tuning parameters (kc, tI, tD, tF). Show that l >
0.5a is required for the controller to be stable.

13: Consider a first-order process with a desired closed-loop response that is second order.
Use the direct synthesis procedure with the following specified closed-loop transfer
function (which is critically damped),
 to derive the controller. Perform simulations for several values of l and compare and
contrast the closed-loop results with those shown in Figure 6-5.

[ Team LiB ]
[ Team LiB ]

Chapter 7. Frequency-Response Analysis

The objective of this chapter is to develop frequency response techniques for control system
analysis and design. After studying this chapter, the reader should be able to do the following:

Substitute s = jw, and find Bode or Nyquist plots for a transfer function

Given a Bode plot, construct a Nyquist plot

Understand the concepts of "all-pass" and "nonminimum phase"

Understand the effect of gains, time constants, RHP zeros, and time delays on amplitude
ratios and phase angles

Find the gain and phase margins for a given process and controller

Tune a controller to achieve a given gain or phase margin

Determine the amount of dead time or gain uncertainty that can be tolerated before a
closed-loop system becomes unstable

Compare the robustness of PI and PID controller tuning techniques

The major sections of this chapter are as follows:

7.1 Motivation

7.2 Bode and Nyquist Plots

7.3 Effect of Process Parameters on Bode and Nyquist Plots

7.4 Closed-Loop Stability Concepts

7.5 Bode and Nyquist Stability

7.6 Robustness

7.7 MATLAB Control Toolbox: Bode and Nyquist Functions

7.8 Summary

[ Team LiB ]
[ Team LiB ]

7.1 Motivation
A control system must satisfy desired performance characteristics (closed-loop time constant,
minimal overshoot, etc.) for nominal operating conditions. By nominal, we normally mean that
the plant is perfectly described by the model used for control-system design. In reality, a model is
never perfect, so controllers must be designed to be robust (to remain stable even when the true
plant characteristics are different from the model).

So far we have analyzed the closed-loop characteristic equation (1 + gcgp) for stability, either by
directly calculating the roots or by using the Routh array to verify that the roots (closed-loop
poles) are negative. There are two problems with this approach:

1. Although the closed-loop characteristic equation may be stable for the nominal plant, the
values of the poles do not indicate how much uncertainty can be tolerated before the system
becomes unstable.

2. The closed-loop characteristic equation is irrational if there is a time delay, which is

characterized by a e–qs term.

The best method for dealing with both of these problems is frequency response analysis. Our
motivation for covering frequency response analysis is primarily that it provides a measure of the
robustness of the controller tuning; that is, it provides a measure of the amount of model
uncertainty that can be tolerated before the control system will become unstable. We consider the
ability to determine the stability of systems with irrational transfer functions (time delays) to be
less important, because a higher order Padé approximation can provide an adequate
representation of the time delay, for analysis purposes.

To understand the concept of frequency response, consider the well-insulated, perfectly mixed
tank with constant flow in and out (perfect level control), shown in Figure 7-1. Assume that the
inlet temperature varies in a sinusoidal fashion between a minimum of 20°C, a maximum of 30°C,
with an average of 25°C. The amplitude of the inlet temperature sine wave is then 5°C.

Figure 7-1. Perfectly mixed tank with varying inlet temperature.

This tank has a volume of 100 gallons, with a flow rate of 10 gallons/minute. You can easily show
that this system is modeled by a first-order transfer function, with a gain of 1°C/°C and a time
constant of 10 minutes. The responses to different frequencies in the inlet temperature are shown
in Figure 7-2. Notice that at low-frequency (0.01 rad/minute) forcing (top plot), the outlet
temperature (dashed line) has approximately the same amplitude as the inlet temperature (solid
line), with very small shift in the peak time. At an intermediate frequency (0.1 rad/minute)
forcing, we begin to see a decrease in the amplitude of the outlet temperature, as well as a "lag"
in the peak time (center plot). At a high frequency (1 rad/minute), the outlet temperature does
not vary substantially (bottom plot).

Figure 7-2. Response of outlet temperature to various inlet

temperature frequencies. Inlet temperature (solid), outlet
temperature (dashed). Top, 0.01 rad/minute; center, 0.1 rad/minute;
bottom, 1 rad/minute. Note the different timescales.

Please note that radians/time is the common unit used for frequency response analysis in process
control. You may be more comfortable thinking in terms of cycles/time. Since there is one cycle
for each 2p radians, then a frequency (f) in cycles/time can be converted to a frequency, w, in
radians/time by

Equation 7.1

Also, the period P ("peak-to-peak time") corresponds to one cycle and can be found by

Equation 7.2
The input-output behavior shown in Figure 7-2 makes physical sense. At low frequencies the inlet
temperature changes on a much slower timescale than the characteristic time of the process, so
the outlet temperature is virtually identical to the inlet temperature. At high frequencies, the inlet
temperature changes so rapidly that the "thermal mass" of the process does not allow the outlet
temperature to change significantly.

In terms of dynamic behavior, this type of system is known as a "low-pass filter" that is, low-
frequency inlet forcing is directly felt at the output. High input frequencies are effectively filtered
by the process and not felt at the output.

There are two important measures that can be obtained from these time-domain plots. One is the
amplitude ratio, which is the amplitude of the output sine wave divided by the input sine wave.
The second measure is the phase shift of the output sine wave relative to the input sine wave, as
illustrated in Figure 7-3. The phase shift, or phase angle, is

Equation 7.3

Figure 7-3. Illustration of phase shift or phase angle, f = –360°(DP/P).

where P is the period, and DP is the lag between an input peak and an output peak.

For example, if the output were exactly out of phase with the input (output at a minimum when
the input is a maximum), the phase angle would be –180° or p radians. If the output lagged the
input by a full cycle, the phase angle would be –360°, and it is possible for the phase angle to
exceed this value (particularly for processes with a time delay). The phase angle in the bottom
plot of Figure 7-2 is approximately –90° (0.5p radians).

Rather than generating numerous time-domain plots like Figure 7-2, each at a different
frequency, the equivalent results can be displayed concisely on a Bode plot, as shown in Figure 7-
4. The amplitude ratio plot is log-log, while the phase angle plot is semilog.

Figure 7-4. Bode plot of the first-order example. Amplitude ratio and
phase shift are plotted as a function of frequency. The circles
correspond to the three frequencies used on the mixing tank.
[ Team LiB ]
[ Team LiB ]

7.2 Bode and Nyquist Plots

In the previous section the basic concept of frequency response was presented. For linear
systems, an input forcing with a certain frequency results in an output with the same frequency,
shifted by the phase angle. There are two ways to perform frequency response analysis: (i)
conversion of the input signal to the Laplace domain, partial fraction expansion of the output, and
inversion to the time domain, or (ii) substitution of s = jw in the transfer function.

Method 2, substitution of s = jw in the transfer function, is the easiest approach. There are also
two ways to display the frequency response information: Bode plots and Nyquist plots. The
following first-order example is used to illustrate these concepts.

Example 7.1: First-Order System

Consider a first-order process transfer function

We use two methods to find the output, given a sine forcing on the process input.

1. Laplace Transform Method

Consider a first-order system, with an input that is forced with a sine wave of magnitude A and a
frequency of w (rad/time). The time and Laplace domain representations of the input are

The Laplace domain output is

Using a partial fraction expansion,

Solving for the coefficients and inverting to the time domain (see Exercise 1), we find

The exponential term decays with time, so the first term can be neglected at long times. Also, a
trignometric identity can be used to find that

Equation 7.4

The ss subscript is used to indicate the long-term behavior of the output, after the initial
transients have decayed. Notice that the output approaches 0 at high frequencies. This is
consistent with the results shown in Figure 7-2. The phase angle also approaches –90° at high
frequencies.

The amplitude ratio is the magnitude of the output divided by the input.

Equation 7.5

Notice that there is a natural dimensionless frequency, wtp, resulting from this analysis.

An easier way to generate frequency response plots is to simply set s = jw in the transfer
function. We illustrate this procedure, again with a simple first-order example.

2. Substitution of s = jw Method

Substitute s = jw into the transfer function

Notice that the result is a complex number (real and imaginary portions) that changes as a
function of frequency. Since the magnitude of a complex number Re + jIm, can be represented by
, then the magnitude (amplitude) of the transfer function (notice that the amplitude at
zero frequency is simply the steady-state process gain) is

Equation 7.6a

Also, the phase angle of a complex number can be represented by tan–1(Im/Re), so

Equation 7.6b

Notice that these are exactly the relationships shown with method 1 [Equations (7.4) and (7.5)].

Also notice that one can simply plot the imaginary values as a function of the real values to obtain
a Nyquist plot. Such a plot is shown in Figure 7-5 for this system. The curve begins close to (kp,0)
at low frequencies, and approaches (0,0) through the first quadrant at high frequencies. That is, it
starts with an amplitude of 1 at a phase angle of 0, and ends with an amplitude of 0 at a phase
angle of –90°.

Figure 7-5. Nyquist plot for a first-order system. The circles

correspond to the three frequencies (1 = 0.01 rad/minute, 2 = 0.1
rad/minute, 3 = 1 rad/minute) used on the mixing tank.
Generalization

To perform a frequency-response analysis, we simply substitute s = jw in the transfer function.

Bode plots consist of amplitude ratio and phase angle plots as a function of frequency. The
amplitude ratio plots are log-log, while the phase angle plots are semilog. Nyquist plots consist of
a curve of the real and imaginary components of the transfer function, as a function of frequency.

The steps to generate the plots are as follows.

1. Generate a vector of frequencies, w, linearly spaced on a logarithmic scale.

2. Substitute s = jw into gp(s).

3. Find the real and imaginary components, Re(w ) + jIm(w).

4. For Bode plots, at each frequency point find the amplitude ratio and phase angle

a. Plot amplitude ratio as a function of frequency on a log-log plot.

b. Plot phase angle as a function of frequency on a semilog plot.

5. For Nyquist plots, plot Im(w) vs. Re(w).

 b.

5.

The results for several common transfer functions are shown in Table 7-1.

Table 7-1. Nyquist and Bode Values for Common Transfer Functions

g(s) Re(w) Im(w) AR f f(w 0) f(w )

tan–1(–tpw) 0 –90°

e–qs cos(qw) –sin(qw) 1 –qw 0 –

0 –90° –90° –90°

ts + 1 1 tw tan–1(tw) 0 90°

kp kp 0 kp 0 0 0

Complex Transfer Functions

Consider now a transfer function that is composed of a number of transfer functions

Equation 7.7

and substituting s = jw, we can find the magnitude at each frequency

Equation 7.8

The amplitude ratio and phase angle can be found at each frequency from

Equation 7.9

Equation 7.10
That is, at each frequency the amplitude ratios add on a logarithmic scale, while the phase angles
add on a linear scale.

Example 7.2: First-Order + Time Delay

Consider a first-order + time-delay process, written as three independent terms

The amplitude ratio and phase angle can be found from Equations (7.9) and (7.10), and Table 7-1

Notice that the time delay has no effect on the amplitude ratio (compared with the pure first-
order process), but it has a great effect on the phase angle, particularly at high frequencies. We
can also think of the effect of the changes in each of the parameters.

Gain Change

A change in the gain merely shifts the amplitude ratio up or down (additive on a log scale), with
no change in the phase angle.

Time Delay Change

A change in the time delay changes the phase angle, but not the amplitude ratio. An increase in
the time delay increases the phase lag (makes it more negative) at any given frequency.

Time Constant Change

A change in the time constant affects both the amplitude ratio and phase angle. An increase in
the time constant tends to decrease the amplitude ratio and increase the phase lag (make it more
negative) at any given frequency.

[ Team LiB ]
[ Team LiB ]

7.3 Effect of Process Parameters on Bode and Nyquist

Plots
In this section we stress the effect of process parameters and the concepts of "all-pass" and
"nonminimum phase."

Effect of Process Order

Consider the following first-, second-, and third-order transfer functions.

The Bode plot is shown below in Figure 7-6. The higher order transfer functions have amplitude
ratios that decrease more rapidly with frequency, and the phase angles asymptotically approach
larger (more negative) values (-180° and -270° for the second- and third-order processes).

Figure 7-6. Bode plots. Comparison of first-, second-, and third-order

systems with a gain of 1 and time constants of 1 minute.
The corresponding Nyquist plot is shown in Figure 7-7. The first-order system stays in the first
quadrant, asymptotically approaching the origin along the imaginary axis. The second-order
system goes into the second quadrant and asymptotically approaches the origin along the real
axis from the negative direction. The third-order system passes into the third quadrant and
asymptotically approaches the origin along the positive imaginary axis.

Figure 7-7. Nyquist plots. Comparison of first-, second-, and third-

order systems with a gain of 1 and time constants of 1 minute. As the
frequency increases, the curves go from (1,0) to (0,0).

Concepts of "All-Pass" and "Nonminimum Phase"

Consider the following four transfer functions,

where the first transfer function is a pure gain, the second and third are Padé approximations
(first and second order, respectively), and the fourth is a pure time delay. The Bode plots are
shown in Figure 7-8. Notice that all these transfer functions have an amplitude that is constant
(does not vary with frequency). A transfer function with a constant amplitude ratio is called all-
pass because all frequencies are directly transferred from input to output with no change in
magnitude.
 Figure 7-8. Bode plots. Comparison of different processes with an
amplitude ratio of 2.

Also, notice that the phase angle is different for each transfer function. A transfer function is
"minimum phase" if there is no other transfer function that has the same amplitude ratio, with a
larger phase angle. In this case, the first transfer function (gain with no dynamics) is minimum
phase, while all of the others are non-minimum phase. The high frequency asymptotes of the
first- and second-order Padé approximations are –180° and –360°, respectively. The high-
frequency phase angle of the time delay is unbounded. The corresponding Nyquist plots are
shown in Figure 7-9; each curve begins at (2,0). The constant gain remains at (2,0). The first-
order Padé transfer function ends at (-2,0), which is –180°. The second-order Padé transfer
function ends at (2,0), which is –360°. The time delay wraps an infinite number of times.

Figure 7-9. Nyquist plots. Comparison of different processes with an

amplitude ratio of 2. For each individual Nyquist plot, the frequency
increases as the curve moves clockwise from the (2,0) point.
Frequency Response Introductory Summary
We have introduced Bode and Nyquist plots. You should be able to develop Nyquist plots given
Bode plots and vice versa. Both analysis techniques are used for control-system tuning, as
presented in the next sections.

For stable systems (poles < 0) with no time delays, we can state the following general results. Let
m be the order of the numerator polynomial and n the order of the denominator polynomial (the
relative order is n–m).

If all process zeros are negative (no RHP zeros), the high-frequency phase lag is
(n–m)*(–90°)

Each RHP zero adds –90° of phase lag (left-half-plane zeros add +90° of phase lead) at high
frequencies

Gains change the amplitude ratio plot but not the phase angle plot

We also can state the following about systems with time delays:

Time delays change the phase-angle plot but not the amplitude ratio plot

Processes with time delays have a phase-angle that continuously decreases with frequency;
this corresponds to a continuous "wrapping" of the origin on a Nyquist plot

[ Team LiB ]
[ Team LiB ]

7.4 Closed-Loop Stability Concepts

Sections 7.1–7.3 introduced frequency response analysis, including the construction of Bode and
Nyquist plots. In this section we illustrate how Bode and Nyquist plots can be used to determine
the stability of a closed-loop system. We also show how to use these plots to tune controllers for
robustness, that is, to be able to tolerate a given level of model uncertainty. The tests used
involve information about the controller and process in open-loop. The open-loop analysis that we
use in this chapter is different from that used in model-based control chapters.

Consider the case where the loop has been opened, as shown in Figure 7-10. We introduce the
concept of Bode stability using the following thought experiment. Assume that the setpoint
applied to the "open-loop" system in Figure 7-10 is a sine wave. Also assume that the controller
has been tuned so that the output lags the setpoint by 180° and has the same amplitude as the
setpoint, as shown in Figure 7-11.

Figure 7-10. Feedback control loop has been "opened."

Figure 7-11. Process output (y) lags setpoint (r) by 180°.

Realize that –y(t) is exactly 180° out of phase with y(t), which means that –y(t) is equal to r(t).
Now consider the case where the setpoint signal, r(t), is suddenly stopped and simultaneously the
loop is closed, as shown in Figure 7-12. This means that the error signal will simply be –y(t),
which is identical to r(t). Since it is identical to r(t), then every signal on the control loop diagram
remains the same. The output continues to oscillate with the same frequency and magnitude as
before the loop was closed. We refer to this control loop as nominally stable.

Figure 7-12. Feedback loop has been "closed."

Consider the case where the output is exactly out of phase (f = –180°) with the setpoint in the
open-loop, but the amplitude of the output is less than the setpoint. Therefore the amplitude ratio
(|y|/|r|) is less than 1. Then, if the loop is closed, the output decreases each time "around the
loop" and the system is stable. In contrast, consider the case where the output is exactly out of
phase with the setpoint when the control loop is open, but the amplitude of the output is greater
than the setpoint. Then, if the loop is then closed, the output increases each time "around the
loop" and the system in unstable.

The thought experiment that we have just performed is the basis for the Bode stability criterion
stated in the next section.
[ Team LiB ]
[ Team LiB ]

7.5 Bode and Nyquist Stability

In the previous section we used a thought experiment based on the open-loop (controller +
process) response to a sinusoidal setpoint input. We found that if the amplitude ratio of the open-
loop transfer function [gc(s)gp(s)] is less than 1 when the phase angle is -180°, then the system
will be stable when the loop is closed. This requirement is also known as the Bode stability
criterion. Notice that we are assuming no valve and measurement effects [or lumping them into
gp(s)]. If these effects are included, then gc(s)gp(s)gv(s)gm(s) should be used in the analysis.

Bode Stability Criterion

Assume that the process is stable [gp(s) has no poles in the RHP] and the controller is stable
[gc(s) has no poles in the RHP]. Also assume that the phase angle (f) crosses –180° only once on
a Bode plot. The system will be closed-loop stable if and only if the amplitude ratio of gcgp is less
than 1 at the crossover frequency.

Equation 7.11

where wco is known as the crossover frequency, which is defined as the frequency where the
phase angle is –180°.

The requirements placed by the Bode stability criterion are not that restrictive for most chemical
processes. For processes where the phase angle crosses –180° more than once, it is necessary to
use the Nyquist stability criterion.

Nyquist Stability Criterion

A system will be closed-loop stable if a Nyquist plot of gc(s)gp(s) does not encircle the critical
point (-1,0). Examples of closed-loop stable and unstable systems are shown in Figure 7-13.

Figure 7-13. Illustration of Nyquist stability criterion. (Left) Stable

system, critical point is not encircled. (Right) Unstable system, critical
point is encircled.
The Bode and Nyquist stability criteria indicate whether a closed-loop system will be stable or not,
but they do not indicate how "close to instability" a system is. That is, the Bode and Nyquist
stability criteria alone do not indicate the robustness of a feedback system to model uncertainty.
Gain and phase margins, defined below, will provide an indication of the robustness of a feedback
system.

Gain Margin
Let ARco represent the amplitude of gc(s)gp(s) at the crossover frequency (wco, where f = -180°).
The gain margin is defined as

Equation 7.12

and the Bode stability requirement is that ARco < 1, so for stability

Equation 7.13

The gain margin is the multiplicative factor that the gain can be increased by before the system
becomes unstable.

Phase Margin
Let fpm represent the phase angle when the amplitude of gc(s)gp(s) is 1. This occurs at a
frequency wpm. The phase margin is defined as

Equation 7.14
The closed-loop system will be stable if f > -180° when AR = 1, so the stability requirement is

Equation 7.15

The phase margin is the amount that the phase angle can decrease (in magnitude, at wpm) before
the system becomes unstable.

We have developed the Bode and Nyquist stability criteria and defined the gain and phase
margins. Next we use an example to illustrate the techniques.

Example 7.3: Nonminimum-Phase Process

Consider the Van de Vusse reactor (Module 5) controlled using a P-only controller.

where time unit is minutes. The Bode plots for gp(s)gc(s) with kc = 2.5 are shown in Figures 7-14
and 7-15.

Figure 7-14. Bode plot for Example 7.3. Calculation of the gain margin.
 Figure 7-15. Bode plot for Example 7.3. Phase margin calculation.

We can use the following steps and the Bode plot in Figure 7-14 to calculate the gain margin.

1. Find the frequency where the phase angle for gpgc = –180°. Here, wco = 4.3 rad/min.

2. At this frequency, find the amplitude ratio for gpgc. Here, AR = 0.60.

3. The gain margin is 1/AR. Here, the gain margin = 1/0.60 = 1.67.

We can use the following steps and the plot in Figure 7-15 to find the phase margin.

1. Find the frequency where the gpgc amplitude ratio is 1.0. Here, wpm = 2.1 rad/minute.

2. At this frequency, find the gpgc phase angle. Here, fpm = –121°.

3. The phase margin is f – (–180). Here, the phase margin = –121 + 180 = 59°.

The Nyquist plot is shown in Figure 7-16. The unit circle is plotted on this diagram so that it is
easy to determine where the amplitude is equal to one. Notice that the Nyquist diagram is
consistent with the Bode diagram, since the amplitude ratio is 0.60 when the phase angle is
–180°, and the phase angle is –121° when the amplitude ratio is 1.0.

Figure 7-16. Nyquist plot for Example 7.3. The unit circle (dashed line)
is included for convenience. The points for the gain margin and phase
margin calculations are shown as the open circle and star,
 respectively.

[ Team LiB ]
[ Team LiB ]

7.6 Robustness
We have noted throughout this text that it is important to tune controllers to be robust. By
robust, we mean that a controller can tolerate a certain amount of change in the process
parameters without causing the feedback system to go unstable. The gain and phase margins
presented in Section 7.5 quantify the amount of uncertainty that can be tolerated.

To illustrate the relationship between gain and phase margins and the model parametric
uncertainty that can be tolerated, we consider the following first-order + time-delay process.

The nominal process model is

Equation 7.16

Let Dk and Dq represent the additive uncertainty in the process gain and dead time, respectively
(notice that we are neglecting any uncertainty in the time constant). The actual process is then

Equation 7.17

Notice that we could also use the representation

Equation 7.18

where . The open-loop transfer function used in the frequency response analysis is

Equation 7.19

Notice that the gain and time-delay uncertainties are clearly separated from the nominal model
and controller e in this equation. Since 1 + is a constant term, it will only affect the
amplitude ratio. Also notice that e–Dqs will only affect the phase angle. This means that we can
think of the gain margin as being related to the amount of gain uncertainty that can be tolerated,
while the phase margin is related to the amount of dead-time uncertainty that can be tolerated.
The change in phase angle, Df, due to the change in deadtime, Dq, can be found from

Equation 7.20

These concepts are shown more clearly in the first-order + dead-time learning module (Module
6).

[ Team LiB ]
[ Team LiB ]

7.7 MATLAB Control Toolbox: Bode and Nyquist

Functions
The M ATLAB Control Toolbox has special functions to generate Bode and Nyquist plots, and to
calculate gain and phase margins. Here we use Example 7.3 to illustrate the use of the MATLAB
bode, nyquist, and imargin functions that are available in the Control Toolbox. For information,
simply enter help bode, and so forth, in the MATLAB command window.

First, define the process transfer function (gp).

» vdvtf = tf(0.5848*[-0.3549 1],[0.1858 0.8627 1])

Transfer function:
-0.2075 s + 0.5848
-------------------------
0.1858 s^2 + 0.8627 s + 1

Next, generate the open-loop transfer function (gcgp) for a controller gain (kc) of 2.5.

» kc = 2.5;

» gcgp = kc*vdvtf

Transfer function:
-0.5189 s + 1.462
-------------------------
0.1858 s^2 + 0.8627 s + 1

Now, use the bode function to generate magnitude and phase angles as a function of frequency.
Also, use imargin to perform the gain margin (Gm) and phase margin (Pm) calculations at their
respective frequencies (Wcg and Wcp).

»[mag,phase,w] = bode(gcgp);

[Gm,Pm,Wcg,Wcp] = imargin(squeeze(mag),squeeze(phase),w)
Gm =
1.6680

Pm =
59.0905

Wcg =
4.3064

Wcp =
2.0956

Figure 7-14 is generated using the following sequence of commands.

subplot(2,1,1),loglog(w,squeeze(mag),[0.1 Wcg Wcg],

[1/Gm 1/Gm min(mag)],'--')
subplot(2,1,2),semilogx(w,squeeze(phase),[min(w) Wcg Wcg],
[-180 -180 0],'--')

Figure 7-15 is generated using the following sequence of commands.

subplot(2,1,1),loglog(w,squeeze(mag),[0.1 Wcp Wcp],

[1 1 min(mag)],'--')
subplot(2,1,2),semilogx(w,squeeze(phase),[0.1 Wcp Wcp],
[-180+Pm -180+Pm 0],'--')

The Nyquist plot, Figure 7-16, is generated using the following sequence of steps. First, define an
appropriate frequency range.

w1 = logspace(-2,2,300);

Then use the nyquist function to generate real and imaginary vectors.

[regcgp,imgcgp] = nyquist(gcgp,w1);

Generate circles for convenience when interpreting the Nyquist plot.

circlex = cos(0:0.2:2*pi);
circley = sin(0:0.2:2*pi);
phaseptx = cos((-180+Pm)*pi/180);
phasepty = sin((-180+Pm)*pi/180);
plot(squeeze(regcgp),squeeze(imgcgp),circlex,circley,'--',
[-1/Gm],[0],'o', phaseptx,phasepty,'*')
axis('square')
xlabel('real')
ylabel('imaginary')

[ Team LiB ]
[ Team LiB ]

7.8 Summary
In this chapter, we developed the Bode and Nyquist frequency response techniques for control-
system design. Gain and phase margins indicate the amount of uncertainty a feedback system
can tolerate before becoming unstable. The gain margin is related to the change in gain (process
or controller) that can be tolerated, while the phase margin is related to the change in time delay
that can be tolerated. From a practical perspective, we recommend a gain margin of at least 2.5
and phase margin of at least 60° to assure robustness.

The basic concepts covered were as follows:

Amplitude ratio (AR): Ratio of the output to input amplitudes

Phase angle (f): Peak-to-peak shift between input and output curves

All-pass: A transfer function with constant amplitude at all frequencies

Nonminimum phase: Generally, transfer functions with time delays and RHP zeros

Bode stability criterion: Amplitude ratio of gcgp < 1 at –180°.

Nyquist stability criterion: The critical point (–1,0) is not encircled by gcgp.

Gain margin: Multiplicative factor tolerated by gcgp before losing stability

Phase margin: Phase-angle error tolerated by gcgp before losing stability

[ Team LiB ]
[ Team LiB ]

Reference

The following text contains an advanced discussion of control system design for robustness:
Morari, M., and E. Zafiriou, Robust Process Control, Prentice-Hall, Englewood Cliffs, NJ (1989).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Find the time-domain solution for the output of a first-order process, with an input sine
wave of magnitude A at frequency w.

2: The peak day length (in terms of hours of daylight) occurs on 21 June, while the peak
average daily temperature occurs on 21 July (30 days later). Assuming that the
relationship between daylight hours and average daily temperature can be modeled by a
first-order transfer function, estimate the time constant of Earth's weather system
(provide the units). Also, what are the units of the "process gain"?

3: Apply the Ziegler-Nichols method to a second-order + dead time process

where kp = 1.5°F/gpm, q = 5 minutes, t1 = 3 minutes, t2 = 15 minutes.

a. Use a Bode plot to find critical (ultimate) proportional gain (kcu ) for a P-only
controller that will make the closed-loop system nominally stable (gain margin = 1,
phase margin = 0). Also, find the ultimate period.

b. Find the gain and phase margins for a Ziegler-Nichols P-only controller. How much
gain uncertainty can be tolerated? How much dead-time uncertainty can be
tolerated?

c. Find the gain and phase margins for a Zielger-Nichols PI controller. How much gain
uncertainty can be tolerated? How much dead-time uncertainty can be tolerated?

d. Find the gain and phase margins for a Zielger-Nichols PID controller. How much
gain uncertainty can be tolerated? How much dead-time uncertainty can be
tolerated?

e. Would you recommend the Ziegler-Nichols tuning if you desired the following gain
and phase margins: gain margin, > 2.5; phase margin, > 60°?

f. Compare the closed-loop responses, for step setpoint changes, of the controllers
obtained in b–d. Show both output and manipulated input responses. Recommend
a set of tuning parameters.

4: For the open-loop system [go1(w) = g pgc(w)] below, find the gain and phase margins.
Also, show your work on the diagram. The unit circle is plotted for convenience.
5: Consider a pH control problem that has the transfer function

A PI controller with kc = 4.35 and tI = 69 yields the following Bode plot. Find the
controller gain that will provide a gain margin of 2.
6: The Bode plots for a particular system are shown below. You know that there is no gain
uncertainty in this process, however, when you implement the tuning parameters, the
feedback system is unstable. At least how much additional dead time was not included
in your model?
7: A first-order + time delay process is controlled by a PI controller. The Nyquist plot for kc
= 4.9688 and tI = 12.5 is shown below. Keeping tI constant, find the proportional gain
that will yield a gain margin of 3.
8: A process under P-only control has the following Bode plot, for kc = 1. What is the
ultimate gain (kcu ) and period (Pu)? If the Ziegler-Nichols recommended controller gain
for P-only control is used, what will be the phase margin? Recall that Ziegler-Nichols
recommended kc = 0.5 kcu for P-only control.
 9: Consider the following Bode plot for a system under PI control.

a. Sketch the Nyquist diagram.

b. How much can the controller gain be increased if a gain margin of 2 is desired?

c. How much can the deadtime increase before the system goes unstable?

[ Team LiB ]
[ Team LiB ]

Chapter 8. Internal Model Control

In Chapter 6 we presented several methods for tuning PID controllers and developed a model-
based procedure (direct synthesis) to synthesize a controller that yields a desired closed-loop
response trajectory. In this chapter, we first develop an "open-loop control" design procedure that
then leads to the development of an internal model control (IMC) structure. There are a number
of advantages to the IMC structure (and controller design procedure), compared with the classical
feedback control structure. One is that it becomes very clear how process characteristics such as
time delays and RHP zeros affect the inherent controllability of the process. IMCs are much easier
to tune than are controllers in a standard feedback control structure.

After studying this chapter, the reader should be able to do the following:

Design internal model controllers for stable processes (either minimum or nonminimum
phase)

Sketch the closed-loop response if the model is perfect

Derive the closed-loop transfer functions for IMC

Design IMC for improved disturbance rejection

The major sections of this chapter are as follows:

8.1 Introduction to Model-Based Control

8.2 Practical Open-Loop Controller Design

8.3 Generalization of the Open-Loop Control Design Procedure

8.4 Model Uncertainty and Disturbances

8.5 Development of the IMC Structure

8.6 IMC Background

8.7 The IMC Structure

8.8 The IMC Design Procedure

8.9 Effect of Model Uncertainty and Disturbances

8.10 Improving Disturbance Rejection Design

8.11 Manipulated Variable Saturation

8.12 Summary

[ Team LiB ]
[ Team LiB ]

8.1 Introduction to Model-Based Control

In the previous chapters we focused on techniques to tune PID controllers. The closed-loop
oscillation technique developed by Ziegler and Nichols did not require a model of the process.
Direct synthesis, however, was based the use of a process model and a desired closed-loop
response to synthesize a control law; often this resulted in a controller with a PID structure. In
this chapter we develop a model-based procedure, where a process model is "embedded" in the
controller. By explicitly using process knowledge, by virtue of the process model, improved
performance can be obtained.

Consider the stirred-tank heater control problem shown in Figure 8-1. We can use a model of the
process to decide the heat flow (Q) that needs to be added to the process to obtain a desired
temperature (T) trajectory, specified by the setpoint (Tsp). A simple steady-state energy balance
provides the steady-state heat flow needed to obtain a new steady-state temperature, for
example. By using a dynamic model, we can also find the time-dependent heat profile needed to
yield a particular time-dependent temperature profile.

Figure 8-1. Stirred-tank heater.

Assume that the chemical process is represented by a linear transfer function model, and that it is
open-loop stable. The input-output relationship is shown in Figure 8-2, where u is the input
variable (heat flow) and y is the output variable (temperature).

Figure 8-2. Open-loop process system.

When the process is at steady state, and there are no disturbances, then the inputs and outputs
are zero (since we are using deviation variables). Consider a desired change in the output y; we
refer to the desired value of y as the setpoint, which is represented by r. We wish to design an
open-loop controller, q(s), so that the relationship between r(s) and y(s) has desirable dynamic
characteristics (fast response without much overshoot, no offset, etc.). The open-loop control
system is shown in Figure 8-3 (we may also wish to think of this as a feed-forward controller,
based on setpoint). We use q(s) to represent the open-loop controller transfer function, to
emphasize that it is a different type of controller than the feedback controllers of Chapters 5 and
6.

Figure 8-3. Open-loop model-based control system.

Using block diagram analysis, we find the following relationship between the setpoint and the
output

Equation 8.1

Static Control Law

The simplest controller will result if q(s) is a constant. Let kq represent this constant. As an
example, consider a first-order process, gp(s) = kp/(tps + 1). Then the relationship between r(s)
and y(s) is

To obtain a desirable response, kq = 1/kp; offset will result otherwise. We can see this from the
final-value theorem. Consider a step setpoint change, of magnitude R.

Equation 8.2

From the final-value theorem

and for no offset, we require that kq = 1/kp. We can also find the time-domain solution to
Equation (8.2)
Again, we can see that kq = 1/kp is necessary for offset-free performance. Notice also that the
speed of response is the same as the time constant of the open-loop process. In order to "speed
up" the response, we must use a dynamic control law, as developed in the next section.

Dynamic Control Law

Better control can be obtained if the controller, q(s), is dynamic rather than static. Indeed, we
find that if

Equation 8.3

then the relationship between r(s) and y(s) is

That is, we have perfect control, since the output perfectly tracks the setpoint! For a first-order
process, the controller is

Although this is mathematically possible, perfect control is unachievable in practical application.

Consider the signals in and out of the control block, shown in Figure 8-3. Since the transfer
function relationship between r(s) and u(s) is, for this example,

Equation 8.4

the differential equation that corresponds to Equation (8.4) is

From a practical point of view, it is impossible to take an exact derivative of r(t), particularly if a
discontinuous step setpoint change is made.

Here we use the inverse Laplace transform to solve Equation (8.4) for u(t), when there is a step
change in r.

A table of Laplace transforms can be used to find the time-domain solution

where d(t) is the impulse function, which has infinite height, infinitesimal width, and unit area.
Since this is hard to understand conceptually, you probably realize that it is impossible to
implement exactly. Think about how you would approximate it.

The Bottom Line. Equation (8.4) is representative of a controller that is not physically
realizable. Notice that the equation has a numerator polynomial, but not a denominator
polynomial. We can state the following general result.

Physical Realizability. For a controller to be physically realizable, the order of the denominator
of the controller transfer function [q(s)] must be at least as great as the order of the numerator.

Definition. A transfer function that satisfies this condition is proper. If the order of the
denominator is greater than the order of the numerator, then the transfer function is strictly
proper. If the order of the numerator is equal to the order of the denominator, then the transfer
function is often called semiproper.

[ Team LiB ]
[ Team LiB ]

8.2 Practical Open-Loop Controller Design

In order to design a physically realizable controller for the first-order process, there must be a
denominator polynomial in the controller. Define a first-order filter as the following transfer
function,

Equation 8.5

where l is a filter tuning parameter that has units of time. We see shortly that we can think of l
as a desired time constant for the output response; if a faster response is desired, then l is
decreased. Now, let the controller be designed in the following fashion to make q(s) proper.

Equation 8.6

which yields the following controller for a first-order process,

Equation 8.7

Equation (8.7) is a lead/lag controller. Since the order of the denominator is at least as great as
the order of the numerator, the design is physically realizable. The response of the output
variable, y, is

Equation 8.8

in this case,

Equation 8.9
That is, there will be a first-order response with a time constant of l. Contrast this with the static
control law [q(s) = kq = 1/kp], which will yield a first-order response with a time constant of tp.
As long as l < tp, the dynamic controller will have a faster response than the static controller.

Response of Manipulated and Output Variables to Step Setpoint

Changes
We have already seen that the output response to a setpoint change is first order [e.g., Equation
(8.9)]. We can also find the manipulated variable response to a setpoint change. Here we
consider the dynamic controller [q(s) = f(s)/gp(s)]

Equation 8.10

Notice that this is a lead/lag controller. For a step setpoint change of magnitude R,

Equation 8.11

The time-domain response for the manipulated input is

Equation 8.12

The output response is

Equation 8.13

Figure 8-4 is a plot of y(t) and u(t), as a function of as a function of l, for a process with tp = 10
minutes and kp = 1 %/%. Notice that the manipulated input action occurs immediately at the
time of the setpoint change. You should show that the magnitude is consistent with the time-
domain solution [Equation (8.12)], and with the Laplace domain solution [Equation (8.11)] using
the initial value theorem.
 Figure 8-4. Response to step setpoint change, as a function of l, for a
first-order process with a time constant of 10 minutes.

Issues in Dynamic Controller Design

We have already seen that a controller needs to be proper to be implemented in a practical
situation. We found that a dynamic controller could be designed by using a process model inverse,
cascaded with a filter transfer function, to make the controller proper.

Question 8.2.1: Can the process model inverse always be used for controller design?

The answer to question 8.2.1 is shown clearly in the following example.

Example 8.1: Inverse Response System

Consider the following transfer function,

Equation 8.14

where b is a positive real number, indicating a positive zero (yielding inverse response) in the
process transfer function. If we use the previous design procedure for a dynamic open-loop
controller, we find
Equation 8.15

and if we let f(s) = 1/(ls + 1), we obtain

Equation 8.16

Notice that the zeros of the process model become the poles of the controller, when the model
inverse is used for control-system design. This creates an unstable controller and the possibility of
unbounded, manipulated variable action. Therefore, if a process has a RHP zero, this zero must
be factored out before using the model inverse for the controller design.

Answer 8.2.1: The process model cannot simply be inverted to form the controller. It must be
factored so that the resulting controller is stable and realizable.

Look at Equation (8.16). If we simply take out the unstable pole, we have

Equation 8.17

Notice that this is not acceptable, because q(s) is not proper. To make it proper, we can simply
increase the order of the filter.

Equation 8.18

where we have used a second-order filter, f(s) = 1/(ls + 1)2, for the controller design. What kind
of output response will be achieved?

Equation 8.19
Notice that the output will exhibit an inverse response when there is a step setpoint change.

This is an important result—process inverse response behavior cannot be removed by any stable
control system.

Example 8.2: Numerical Example of an Inverse Response System

Consider the following transfer function, which has a RHP zero at 1/9 min-1:

Compare the responses (input and output) under open-loop control, for the two cases: (i) when
the RHP pole is removed from the controller and (ii) when the RHP pole is not removed from the
controller (i.e., the controller is unstable)

(i) When the unstable controller pole is removed, we have from Equation (8.18)

which yields the following manipulated and output variable response:

For a unit step change in setpoint, the output and manipulated variable responses are shown in
Figure 8-5 for l = 5 minutes. Notice that the inverse response appears in the output variable, and
that the manipulated input changes immediately when the setpoint change is made. The reader
should be able to show that the manipulated input change is consistent with that predicted by the
initial value theorem.

Figure 8-5. Unit step setpoint response when controller is stable (l = 5

minutes).
(ii) When the unstable controller pole is not removed, we have from Equation (8.16)

Equation 8.20

which is an unstable controller. This controller yields the manipulated variable response

Equation 8.21

The output response is

Equation 8.22

The output and manipulated variable responses are shown in Figure 8-6 for l = 5 minutes. Notice
that the inverse response does not appear in the output variable, but the manipulated variable is
unbounded (unstable). This is often called internal instability.

Figure 8-6. Unit step setpoint response with unstable controller (note
scale of response).

This last example illustrated an important point that is not often mentioned in textbooks. If only
the process output is calculated or plotted, then the engineer may overlook the fact that a
manipulated variable is becoming unbounded. Good output performance does not ensure feasible
manipulated variable action.

When designing the open-loop control system of Figure 8-3, it is critical that both the controller
and the process be stable. If the process is unstable, an open-loop controller cannot be used. The
proper control-system design procedure for open-loop unstable systems is covered in Chapter 9.

[ Team LiB ]
[ Team LiB ]

8.3 Generalization of the Open-Loop Control Design

Procedure
So far we have used two example processes: (i) first-order and (ii) second-order with a RHP zero.
Now is the time to generalize our results to any process system.

First of all, we have been using the inverse of the process transfer function, gp(s), in our control
law. We must realize that the actual process transfer function is never known exactly. We now
use two transfer function representations of the process. One is considered the process
(sometimes we use the term plant), which is never known exactly. The other is considered the
process model, which we do know exactly.

Equation 8.23

Equation 8.24

Recall that when the process had a RHP zero, we performed a factorization, so that the RHP zero
did not form a RHP pole in the controller. In general, we factor the process model in the following
fashion,

Equation 8.25

where contains the invertible elements and contains the noninvertible elements. One
easy way to remember this is to recall that if RHP zeros are "inverted" (the RHP is the positive
side of the complex plane), they become RHP poles, which are unstable.

Question 8.3.1: Is dead time invertible?

Answer 8.3.1: The reader should verify that dead time is not invertible. If there is dead time in
the process, it cannot be removed by any physically realizable controller.

Our controller design will consist of inverting the invertible part of the process model and
cascading this with a filter that is of high enough order to make the controller proper.

Equation 8.26
In the next section, we see that there are a number of ways to factor the same transfer function
into invertible and noninvertible parts.

Controller Factorization
There are a number of ways to factor the same transfer function,

The most common method, known as an all-pass factorization, is shown as method 2 in the
example below. The reason for the term all-pass was illustrated in Chapter 7, frequency-response
techniques.

Example 8.3: Factorization Techniques

Consider a process model with inverse response characteristics (a RHP zero)

Equation 8.27

where is positive.

Method 1: Simple factorization

The simple factorization approach is to simply place the RHP zeros in the noninvertible part of the
process model .

Equation 8.28

Method 2: All-pass factorization

The all-pass factorization places the RHP zero in the noninvertible part of the process model, but
it also places a pole at the reflection of the RHP zero:

Equation 8.29
Generally, Method 2 (all-pass) is used because the resulting controlled system minimizes the
integral squared error (ISE).

Comparison of Output Responses for Different Controller

Factorizations
Here we compare the different controlled variable responses for methods 1 and 2.

Method 1

Equation 8.30

where we have used a second-order filter to make q(s) realizable.

Now, since y(s) = q(s) gp(s) r(s),

Equation 8.31

Assuming a perfect model

Equation 8.32

which is the same as the result obtained in Equation (8.19).

Method 2

Equation 8.33
Again, since y(s) = q(s) gp(s) r(s),

Equation 8.34

and assuming a perfect model

Equation 8.35

which is slightly different from the response shown in Equation (8.32). The responses will be
equivalent only if we set l = b (and there is generally no reason to do that).

Summary of Controller Design Procedure

The controller design procedure has been generalized to the following steps.

1. Factor the process model into invertible and noninvertible portions

Method 1 uses a simple factorization

Method 2 uses an all-pass formulation

2. Invert the invertible portion of the process model (the "good stuff") and cascade with a filter
that makes the controller q(s) proper.

The factorization is used only for control-system design. The actual model still contains the RHP
zeros and time delays.
Remember that the factorization was performed so that the controller would be stable. For linear
systems, if the controller is stable and the process is stable, then the overall controlled system is
stable. This is true simply because if two transfer functions are stable, then the transfer functions
cascaded together (multiplied) are stable. This is a nice result because in a standard feedback
control formulation, the controller and the process can each be stable, yet the feedback system
may be unstable.

[ Team LiB ]
[ Team LiB ]

8.4 Model Uncertainty and Disturbances

Thus far we have assumed that the model is perfect when designing the open-loop control
system. In practice, it is impossible to have a perfect model. Often, the process gain can be off by
20–50% (or even more, for highly nonlinear processes). The next example shows the problems
associated with a purely open-loop control strategy when there are uncertainties or disturbances.

Example 8.4: First-Order Process with Model Uncertainty

The model and process are represented by the following transfer functions:

There are no noninvertible elements, so the control design procedure results in the following:

Notice that the response will be second order with numerator dynamics if the model is not perfect.
Also, there will be offset when there is a step change in setpoint.

Consider the following numerical example, where the model gain and time constant are 1 %/%
and 10 minutes, respectively, but where the actual process has a gain of 1.2 %/% and a time
constant of 8 minutes. If the tuning parameter, l = 4 minutes, the relationship between the
setpoint and the output is

Whereas the response with a perfect model would be y(s) = r(s)/(4s+1). The response with
model uncertainty is compared with the perfect model response in Figure 8-7 for a unit step
change in setpoint. The manipulated input value is the same whether the process model is perfect
or in error, since there is no feedback from the measured output to correct for this error.

Figure 8-7. Comparison of open-loop controller responses.

Some form of feedback is needed to correct for model uncertainty (as well as any disturbances
entering the process). The method that we develop to account for model uncertainty and
disturbances is known as internal model control (IMC). In the next section, we begin to develop
the IMC structure.

[ Team LiB ]
[ Team LiB ]

8.5 Development of the IMC Structure

Consider the open-loop control structure shown in Figure 8-3.

We now consider a process model that receives the same manipulated variable signal as the
actual process (Figure 8-8). We can now subtract the difference between the process output
(actually measured) and the process model output (model predicted) to determine the model
error. This is shown in Figure 8-9. We must also realize that disturbances can enter the system,
as shown in Figure 8-10.

Figure 8-8. Process model in parallel with the actual process.

Figure 8-9. Calculating model error.

Figure 8-10. Incorporating the process disturbance.

Notice that Figure 8-10 shows the calculation of model uncertainty (which includes unmeasured
disturbances). This information can now be used by the controller, to compensate for the model
uncertainty. This creates a feedback system, as shown in Figure 8-11.

Figure 8-11. Feedback structure derived from the open-loop structure.

Figure 8-11 is also known as the IMC structure, which is discussed in depth in the next sections.

[ Team LiB ]
[ Team LiB ]

8.6 IMC Background

The main advantage to IMC is that it provides a transparent framework for control-system design
and tuning. As we show in this chapter, the IMC control structure can be formulated in the
standard feedback control structure. For many processes, this standard feedback control
structure will result in a PID controller (sometimes cascaded with a first-order lag). This is
pleasing because we can use standard equipment and algorithms (i.e., PID controllers) to
implement an "advanced" control concept.

The IMC design procedure is exactly that of the open-loop "control" design procedure developed
in Section 8.3. Remember that a factorization of the process model was performed so that the
resulting controller would be stable. If the controller is stable and the process is stable, then the
overall controlled system is stable. This is true simply because if two transfer functions are stable,
then the transfer functions cascaded together (multiplied) are stable. This is a nice result because
in a standard feedback control formulation, the controller and the process can each be stable, yet
the feedback system may be unstable. The restriction of this design procedure is that the process
must be stable. An extension to unstable processes is covered in the Chapter 9.

Although the IMC design procedure is identical to the open-loop "control" design procedure, the
implementation of IMC results in a feedback system. Thus, IMC is able to compensate for
disturbances and model uncertainty, while open-loop "control" is not. Note that the internal model
controller must be detuned to assure stability if there is model uncertainty.

[ Team LiB ]
[ Team LiB ]

8.7 The IMC Structure

The IMC structure is shown in Figure 8-12. The distinguishing characteristic of this structure is the
process model, which is in parallel with the actual process (plant). Note that (~) is generally used
to represent signals associated with the model. Other literature sources may use a subscript
(such as m) to represent the model. Figure 8-13 illustrates that both the controller and model
exist as computer computations; it is convenient to treat them separately for design and analysis.

Figure 8-12. The internal model control structure.

Figure 8-13. The IMC strategy. The dotted line indicates the
calculations performed by the model-based controller.

A list of transfer function variables shown in the IMC block diagram are given below.
d(s)= disturbance

estimated disturbance
=

gp(s)= process
process model
=

q(s)= internal model controller

r(s)= setpoint

= modified setpoint (corrects for model error and disturbances)

u(s)= manipulated input (controller output)

y(s)= measured process output

= model output

Notice that the feedback signal is

Equation 8.36

The signal to the controller is

Equation 8.37

Consider now some limiting cases.

Perfect Model, No Disturbances

If the model is perfect and there are no disturbances (d(s) = 0), then the feedback
signal is zero. The relationship between r(s) and y(s) is then

Equation 8.38

Notice that this is the same relationship that we get for an open-loop control system design.

Why is this nice? If the controller, q(s), is stable and the process, gp(s), is stable, then the closed-
loop system is stable. We developed a design procedure that yields a stable, physically realizable
controller in Section 8.3.

Recall that a standard feedback controller could actually destabilize a process if we did not
correctly choose the tuning parameters. An analysis of the poles of the closed-loop transfer
function must be performed to determine the stability of standard feedback controllers.

Perfect Model, Disturbance Effect

If the model is perfect and there is a disturbance, then the feedback signal is

This illustrates that feedback is needed because of unmeasured disturbances entering a process.

Model Uncertainty, No Disturbances

If there are no disturbances [d(s) = 0] but there is model uncertainty , which is

always the case in the real world, then the feedback signal is

This illustrates that feedback is needed because of model uncertainty.

The closed-loop relationship is

Recapitulating, the reasons for feedback control include the following:

Unmeasured disturbances

Model uncertainty

Faster response than the open-loop system (with a static controller)

Closed-loop stability of open-loop unstable system

The primary disadvantage of IMC is that it does not guarantee stability of open-loop unstable
systems. The procedure detailed in Chapter 9 handles these systems.

[ Team LiB ]
[ Team LiB ]

8.8 The IMC Design Procedure

The IMC design procedure for SISO systems is identical to the design procedure that we
developed for open-loop controller design earlier. The assumption we are making is that the
model is perfect, so the relationship between the output, y, and the setpoint, r, is given by
Equation (8.1). Model uncertainty is handled by adjusting the "filter factor" for robustness
(tolerance of model uncertainty) and speed of response. The IMC design procedure consists of the
following four steps.

1. Factor the process model into invertible ("good stuff") and noninvertible ("bad stuff"—time
delays and RHP zeros) elements (generally, an all-pass factorization will be used).

Equation 8.39

This factorization is performed so that the resulting controller will be stable.

2. Form the idealized IMC controller. The ideal internal model controller is the inverse of the
invertible portion of the process model ("good stuff").

Equation 8.40

3. Add a filter to make the controller proper. A transfer function is proper if the order of the
denominator polynomial is at least as high as the numerator polynomial.

Equation 8.41

If it is most desirable to track step setpoint changes, the filter transfer function usually has
the form

Equation 8.42

and n is chosen to make the controller proper (or semiproper). If it is most desirable to track
ramp setpoint changes (often used for batch reactors or transition control problems), then
 4. Adjust the filter-tuning parameter to vary the speed of response of the closed-loop system.
If the l is "small," the closed loop system is "fast," if l is "large," the closed-loop system is
more robust (insensitive to model error).

The student should note that the factorization performed in Equation (8.39) is used only for
controller design. The process model that is simulated in parallel with the process is still the full
model, .

If the process model is perfect, then we can easily calculate what the output response to a
setpoint change will be. Substituting Equation (8.40) into Equation (8.38), we find

Equation 8.43

If the model is perfect, then

Equation 8.44

and we can substitute Equation (8.44) into Equation (8.43) to find

Equation 8.45

which yields

Equation 8.46

Equation (8.46) indicates that the bad stuff must appear in the output response. That is, if the
open-loop process has a RHP zero (inverse response), then the closed-loop system must exhibit
inverse response. Also, if the process has dead time, then dead time must appear in the closed-
loop response. Please remember that Equation (8.46) only holds for the case of a perfect model.

The most common process model is a first-order plus time-delay transfer function. The design
procedure for this system is shown in the next example.
Example 8.5: First-Order + Dead Time Process
Consider a first-order + time-delay model:

Using the four-step design procedure, first factor out the noninvertible elements,

Then form the idealized IMC controller,

and add a filter to make the controller proper

Once again, the controller is of lead-lag form. Finally, adjust l for response speed and robustness.
The closed-loop response (assuming a perfect model) to a setpoint change is

For a step setpoint change of magnitude R,

Numerical Example

Consider a first-order + time-delay process with a time constant of 10 seconds, a time delay of 5
seconds, and a gain of 1%/%. The output and manipulated variable responses for various values
of l are shown in Figure 8-14.
 Figure 8-14. Response as a function of l for IMC of a first-order +
time-delay process.

Notice that the IMC procedure has effectively "compensated" for the time delay. The time delay
was not removed, but the controller does not "expect" results from a manipulated variable move
until "after" the time delay because the model is integrated "in-parallel" with the process.

Example 8.6: Second-Order with an RHP Zero

Consider the following transfer function, which has an RHP zero at 1/9 min-1:

Using all-pass factorization of the model, we find

Forming the idealized controller,

Add the filter to make the controller semiproper:

For a perfect model, the output response is

and the manipulated variable response is

The output and manipulated responses for a step setpoint change is shown in Figure 8-15 as a
function of l. Notice that as l is decreased, the inverse response becomes more pronounced—the
inverse response can not be removed by a stable controller.

Figure 8-15. Input and output response as a function of l when all-

pass filter is used.
It should be noted that the initial-value theorem can be used to determine the manipulated
variable action that occurs immediately after a step setpoint change.

We have shown how to factor models with time delays and RHP zeros and have used several
examples to illustrate the effect of the tuning parameter (or filter time constant), l, on the output
response to setpoint changes if the model is perfect. The IMC design procedure is identical to the
open-loop control design procedure presented in Section 8.3. Although the control-system design
procedure is the same, the implementation of IMC is much different, since it incorporates
feedback to compensate for model uncertainty or disturbances.

[ Team LiB ]
[ Team LiB ]

8.9 Effect of Model Uncertainty and Disturbances

For controller design purposes, we assumed that the model was perfect. In practice there is
model uncertainty. Model uncertainty will be handled by detuning the filter parameter, l, for
robustness.

The Effect of Model Uncertainty on Setpoint Response

Consider a setpoint change only; then

We can see immediately that if the model is perfect , then we have the open-loop
"control" formulation. When the model is not perfect, then the closed-loop response is much more
complicated and can even be unstable if the filter [f(s)] is not detuned sufficiently.

Disturbance Rejection
Consider a disturbance only; then

Equation 8.47

Assume a load disturbance transfer function of the following form

Indeed, often the load disturbance is really a disturbance to the process input, so the disturbance
transfer function has dynamic behavior similar to the process transfer function. It turns out that a
controller that is tightly tuned and works well for setpoint changes may fail miserably on handling
disturbance rejection. This is shown by the following example.

Example 8.7: First-Order Process

Here we consider a first-order process model transfer function, with a time unit of minutes
For simplicity, we assume a perfect model ( ) and that the disturbance transfer
function is the same as the process ( ). The IMC controller design procedure leads to
the response for a step setpoint change shown in Figure 8-16, with l = 2 min. Although the
response is unrealistically fast (l is too small compared with the process time constant), we wish
to show that good setpoint tracking does not necessarily lead to good disturbance rejection.

Figure 8-16. Setpoint response for Example 8.7, with l = 2 min.

Disturbance rejection results are shown in Figure 8-17. Notice that there is a very slow response
to the step load disturbance. This result is perhaps the greatest criticism of the standard IMC
design procedure; it tends to lead to slow responses to load disturbances that occur at the
process input. An improved IMC procedure for disturbance rejection is presented in the next
section.

Figure 8-17. Disturbance response for Example 8.7, with l = 2 min.

[ Team LiB ]
[ Team LiB ]

8.10 Improved Disturbance Rejection Design

We can achieve improved disturbance rejection by specifying a different IMC filter design
procedure. We perform the same factorization of the process model as before. The difference is in
the selection of the IMC filter transfer function. Here, rather than using a filter of the form

we use a filter with the form

Equation 8.48

where g is selected to achieve good disturbance rejection. In practice, g will be selected to cancel a
slow disturbance time constant. Consider the closed-loop transfer function for disturbance
rejection

Equation 8.49

In the case of a perfect model, this results in

Equation 8.50

The controller, using the new filter form, is

Equation 8.51

so the output response is

Equation 8.52

Although it is hard to tell from this general expression, g should be selected to cancel the slow
time constant associated with gd(s). This is shown by way of an example for a first-order process.

Example 8.8: First-Order Process, Improved Disturbance Rejection

Design
For a first-order process model,

The IMC controller, using the proposed filter design, is

Assuming a perfect model and load disturbance transfer function equal to the process transfer
function, Equation (8.52) becomes
If we select l2/(2l - g) to cancel the process model time constant, ,we find

Numerical Example

Here we consider again the first-order process model transfer function from Example 8.7,

and assume that the disturbance transfer function is equal to the process transfer function (this is
equivalent to assuming that the load disturbance occurs at the process input).

which leads to the closed-loop results for a unit step disturbance, shown in Figure 8-18.

Figure 8-18. Improved disturbance response for Example 8.8, with l =

2 min. A comparison with the standard IMC filter design.
It should be noted that the improved disturbance rejection design does not lead to detrimental
performance for setpoint changes. In Figure 8-19, the new filter design also leads to faster
performing setpoint responses. The major disadvantage is the increased manipulated variable
action. In practice, the l value should be detuned for improved robustness.

Figure 8-19. Setpoint response for Example 8.8, with l = 2 min. A

comparison of design for improved disturbance rejection with the
standard IMC filter design.

[ Team LiB ]
[ Team LiB ]

8.11 Manipulated Variable Saturation

The reader should note that if the actual manipulated variable signal is used by the model, then
manipulated variable saturation is not a problem because the system becomes open-loop and the
model prediction is the same as the process output (if there are no disturbances and the model is
perfect). In conventional PID controllers, special precautions must be taken to prevent reset or
integral windup from occurring when manipulated variables hit constraints. This phenomenon is
discussed more fully in Chapter 11.

The IMC block diagram for the case where the actual manipulated variable value that is
implemented on the process is also used by the model is shown in Figure 8-20. Integral "wind-up"
will not be a problem with this implementation. Figure 8-21 shows an IMC implementation where
the manipulated variable value implemented on the process is constrained, while the process
model uses the unconstrained value. Integral wind-up may be a problem in this case. It is very
important when implementing IMC to use the actual manipulated input to the model in the control
law.

Figure 8-20. The actual manipulated variable value is "measured" and

used by the process model. (Correct implementation)

Figure 8-21. The manipulated variable value used by the process

model is different from the value actually implemented on the process.
(Undesirable implementation)
[ Team LiB ]
[ Team LiB ]

8.12 Summary
The IMC design procedure is exactly the same as the open-loop control design procedure. Unlike
open-loop control, the IMC structure compensates for disturbances and model uncertainty. The
IMC tuning (filter) factor, l, is used to detune for model uncertainty. It should be noted that the
standard IMC design procedure is focused on setpoint responses, but good setpoint responses do
not guarantee good disturbance rejection, particularly for disturbances that occur at process
inputs. A modification of the IMC design procedure was developed to improve input disturbance
rejection. Tolerance of model uncertainty is called robustness and was discussed in more detail in
Chapter 7.

Like open-loop control, the disadvantage compared with standard feedback control is that IMC
does not handle integrating or open-loop unstable systems. In Chapter 9, we develop a
procedure, based on IMC, to form a standard feedback control system that can handle open-loop
unstable systems. We call this IMC-based PID.

Terms
All-pass: Property of a transfer function (detailed in Chapter 7).

Factorization: Factoring a transfer function into invertible (good stuff) and noninvertible (bad
stuff) portions.

Minimum phase: No RHP zeros or time delays.

Nonminimum phase: Contains RHP zeros and/or time delays.

Physical realizability: A controller can be physically implemented if it is proper.

Proper: A transfer function is proper if the order of the denominator is at least as great as
the order of the numerator. If they are exactly the same order, the transfer function is
semiproper. If the order of the denominator is greater than the numerator, the transfer
function is strictly proper.

Summary of Internal Model Control System Design Procedure

1. Develop a process model.

2. Factor the process model into invertible (good) and noninvertible (bad) portions,

3.
 usually using an all-pass factorization.

3. Invert the invertible portion of the process model (the good stuff) and cascade with a filter
that makes the controller q(s) proper.

For a focus on step setpoint changes, the following form is often used:

Although not discussed in detail in the chapter, for good tracking of ramp setpoint changes,
the following form for the filter should be used:

For good rejection of step input load disturbances, the form used is

where g is selected to cancel the slow process time constant.

4. Implement in the form of Figure 8-20 to handle constraints on the manipulated input.

[ Team LiB ]
[ Team LiB ]

References

The improved IMC disturbance rejection procedure was suggested by Braatz and co-workers:
Horn, I. G., J. A. Arulandu, C. J. Gombas, J. G. VanAntwerp, and R. D. Braatz, "Improved Filter
Design in Internal Model Control," Ind. Eng. Chem. Res., 35(10), 3437–3441 (1996).

A complete IMC design procedure is presented in the following: Morari, M., and E. Zafiriou, Robust
Process Control, Prentice-Hall, Upper Saddle River, NJ (1989).

[ Team LiB ]
[ Team LiB ]

Student Exercises
Problems 1–5 require analytical solutions, Problems 6 and 8 are SIMULINK problems.

1: Design an IMC for the following process model:

Assume that the all-pass factorization technique is used. Assuming the model is perfect,
sketch the transient response of y(t) to a step setpoint change in r(t).

2: Consider the following IMC block diagram. Find the closed-loop transfer functions
relating the load [l(s)] and the setpoint [r(s)] to the output [y(s)]. Assume that the
model is not perfect.

3: A biochemical reactor with several components has the following linearized model, which
relates the biomass concentration to the dilution rate:

a. Design an IMC, q(s), for this process, assuming that (i) an all-pass filter is used,
and (ii) the controller is strictly proper (order of the denominator is greater than

b.
 a.

the order of the numerator).

b. Qualitatively sketch the type of closed-loop response to a setpoint change that you
would expect if the model is perfect.

4: A process has the following relationship between the input and output:

Design an IMC controller for this process. Let q(s) be semiproper (order of denominator
is equal to the order of numerator) and use an all-pass factorization. Find the closed-
loop transfer function assuming a perfect model. Sketch the closed-loop response to a
step setpoint change.

5: Consider the closed-loop response for IMC when the model is not perfect. Show that
there is no offset for a setpoint change for the following process model and actual
process transfer functions. Also, find the minimum value of l that assures closed-loop
stability for this system.

6: Design an IMC for the following process:

where , , , , and b = 6 min.

a. For this part, do not factor out the right-half-plane zero to form the controller.
Perform simulations to show that the manipulated variable grows unbounded when
a step setpoint change is made, although good performance of the output is
achieved.

For the following parts, assume that the "all-pass" factorization technique is used.

b. Assuming the model is perfect, plot the transient response of y(t) to a unit step
setpoint change in r(t) as a function of l. (Show your control loop diagram.)

c. Discuss how dead-time uncertainty degrades the closed-loop performance (show

several curves on one plot) and show how the sensitivity depends on the filter
constant, l.

d. Show how gain uncertainty degrades the closed-loop performance (show several
curves on one plot) and show how the sensitivity depends on the filter time

e.
 d.

constant, l.

e. Now, assume that there is no setpoint change but that there is a step disturbance.
Assume that the disturbance transfer function is equal to the process transfer
function. Compare and contrast the response of the output with the open-loop
control response.

For the following parts, assume that the all-pass factorization technique is not used.

f. Assuming the model is perfect, plot the transient response of y(t) to a unit step
setpoint change in r(t) as a function of l. (Show your control loop diagram.)

g. Discuss how dead-time uncertainty degrades the closed-loop performance (show

several curves on one plot) and show how the sensitivity depends on the filter
constant, l.

h. Show how gain uncertainty degrades the closed-loop performance (show several
curves on one plot) and show how the sensitivity depends on the filter time
constant, l.

i. Now, assume that there is no setpoint change, but that there is a step disturbance.
Assume that the disturbance transfer function is equal to the process transfer
function. Compare and contrast the response of the output with the open-loop
control response.

7: Packed-bed reactors often exhibit "wrong way" (inverse response) behavior. You are
responsible for the control-system design for a packed-bed reactor that has the
following step response behavior, where a step decrease in steam valve position was
made at t = 5 minutes.

You have developed the following process model (the time unit is minutes):
 a. What are the units for the process gain?

b. Design an IMC for this process. Use the all-pass factorization for the RHP zero, and
assume that q(s) is semiproper (numerator order in s is equal to the denominator
order in s).

c. Assuming a perfect model, plot qualitatively how the temperature will respond to a
step setpoint change of 1°C.

d. It is desirable to make certain that the control valve position, immediately after a
10°C setpoint change, does not move more than 25%. What is the smallest value
of l that you can use? Show your work.

8: Consider the following isothermal chemical reactor, where the dilution rate (feed flow
rate per unit volume of reactor) is manipulated to achieve a desired concentration of
product.

The process model is

which has a RHP zero (inverse response) and a time delay (due to the concentration
measurement device).

Design an IMC for this process.

a. For this part, do not factor out the RHP zero to form the controller. Perform
simulations to show that the manipulated variable grows unbounded when a step
setpoint change is made, although good performance of the output is achieved.

For the following parts, assume that the all-pass factorization technique is used.

b.
 b. Assuming the model is perfect, plot the transient response of y(t) to a unit step
setpoint change in r(t) as a function of l. (Show your control loop diagram.)

c. Discuss how dead-time uncertainty degrades the closed-loop performance (show

several curves on one plot) and show how the sensitivity depends on the filter
constant, l.

d. Show how gain uncertainty degrades the closed-loop performance (show several
curves on one plot) and show how the sensitivity depends on the filter time
constant, l.

e. Now, assume that there is no setpoint change but that there is a step disturbance.
Assume that the disturbance transfer function is equal to the process transfer
function. Compare and contrast the response of the output with the open-loop
control response.

For the following parts, assume that the all-pass factorization technique is not used.

f. Assuming the model is perfect, plot the transient response of y(t) to a unit step
setpoint change in r(t) as a function of l. (Show your control loop diagram.)

g. Discuss how dead-time uncertainty degrades the closed-loop performance (show

several curves on one plot) and show how the sensitivity depends on the filter
constant, l.

h. Show how gain uncertainty degrades the closed-loop performance (show several
curves on one plot) and show how the sensitivity depends on the filter time
constant, l.

i. Now, assume that there is no setpoint change but that there is a step disturbance.
Assume that the disturbance transfer function is equal to the process transfer
function. Compare and contrast the response of the output with the open-loop
control response.

[ Team LiB ]
[ Team LiB ]

Appendix 8.1: Derivation of Closed-Loop Relationships

for IMC
Refer again to Figure 8-12. Recall that the feedback signal is

and the signal to the controller is

The controller output or process input is

which can be rearranged to

Now, we realize that

and we find

or
which can be written

[ Team LiB ]
[ Team LiB ]

Chapter 9. The IMC-Based PID Procedure

In Chapter 8 we developed a transparent framework for control-system design: the IMC
structure. One nice thing about the IMC procedure is that it results in a controller with a single
tuning parameter, the IMC filter (l). For a system that is "minimum phase," l is equivalent to a
closed-loop time constant (the "speed of response" of the closed-loop system). Although the IMC
procedure is clear and IMC is easily implemented, the most common industrial controller is still
the PID controller. The purpose of this chapter is to show that the IMC block diagram can be
rearranged to the form of a standard feedback control diagram. We find that the IMC law, for a
number of common process transfer functions, is equivalent to PID-type feedback controllers.

After studying this chapter, the reader should be able to do the following:

Design an internal model controller, then find the equivalent feedback controller (IMC-based
PID) in standard form; derive and use the results presented in Table 9-1

Use an approximation for time delays in order to find a PID-type control law; compare the
performance of various IMC-based PID controllers for time-delay processes; use Table 9-2

Use Table 9-3 to find PID-type controllers for unstable processes

The major sections of this chapter are as follows:

9.1 Background

9.2 The Equivalent Feedback Form to IMC

9.3 IMC-Based Feedback Design for Delay-Free Processes

9.4 IMC-Based Feedback Design for Processes with a Time Delay

9.5 Summary of IMC-Based PID Controller Design for Stable Processes

9.6 IMC-Based PID Controller Design for Unstable Processes

9.7 Summary

[ Team LiB ]
[ Team LiB ]

9.1 Background
As we show in this chapter, the IMC structure can be rearranged to the standard feedback
structure.

Question: Why do we care about IMC, if we can show that it can be rearranged into the standard
feedback (PID) structure?

Answer: Because the process model is explicitly used in the control-system design procedure.
The standard feedback structure uses the process model in an implicit fashion; that is, PID tuning
parameters are often "tweaked" based on a transfer function model, but it is not always clear how
the process model affects the tuning decision. In the IMC formulation, the controller, q(s), is
based directly on the "good" part of the process transfer function. The IMC formulation generally
results in only one tuning parameter, the closed-loop time constant (l, the IMC filter factor). The
IMC-based PID tuning parameters are then a function of this closed-loop time constant. The
selection of the closed-loop time constant is directly related to the robustness (sensitivity to
model error) of the closed-loop system. Also, for open-loop unstable processes, it is necessary to
implement the IMC strategy in standard feedback form, because the IMC suffers internal stability
problems.

The reader should realize that the IMC-based PID controller presented in this chapter will not give
the same results as the IMC strategy when there are process time delays, because the IMC-based
PID procedure uses an approximation for dead time, while the IMC strategy uses the exact
representation for dead time.

[ Team LiB ]
[ Team LiB ]

9.2 The Equivalent Feedback Form to IMC

In this section we derive the standard feedback equivalence to IMC by using block- diagram
manipulation. Begin with the IMC structure shown in Figure 9-1; the point of comparison between
the model and process output can be moved as shown in Figure 9-2.

Figure 9-1. IMC structure.

Figure 9-2. Cosmetic change in IMC structure.

Figure 9-2 can be rearranged to the form of Figure 9-3.

Figure 9-3. Rearrangement of IMC structure.

The arrangement shown inside the dotted line of Figure 9-3 is shown in Figure 9-4.

Figure 9-4. Inner loop of the rearranged IMC structure shown in Figure
9-3.

Figure 9-4 can be rearranged to the form of Figure 9-5.

Figure 9-5. Equivalent block to Figure 9-4.

Notice that r(s) – y(s) is simply the error term used by a standard feedback controller. Therefore,
we have found that the IMC structure can be rearranged to the feedback control (FBC) structure,
as shown in Figure 9-6. This reformulation is advantageous because we find that a PID controller
often results when the IMC design procedure is used. Also, the standard IMC block diagram
cannot be used for unstable systems, so this feedback form must be used for those cases.

Figure 9-6. Standard feedback diagram illustrating the equivalence

with IMC. The feedback controller, gc(s), contains both the internal
model, , and the internal model controller, q(s).
Now, we can use the IMC design procedure to help us design a standard feedback controller. The
standard feedback controller is a function of the internal model, , and internal model
controller, q(s), as shown in Equation (9.1).

The standard feedback controller which is equivalent to IMC is

Equation 9.1

We refer to Equation (9.1) as the IMC-based PID relationship because the form of gc(s) is often
that of a PID controller. The IMC-based PID procedure is similar to the IMC procedure of the
previous chapter, with some additional steps. One major difference is that the IMC-based
procedure will, many times, not require that the controller be proper. Also, the process dead time
will be approximated using the Padé procedure, in order to arrive at an equivalent PID-type
controller. Because of the Padé approximation for dead time, the IMC-based PID controller may
not perform as well as IMC for processes with significant time delays.

The IMC-Based PID Control Design Procedure

The following steps are used in the IMC-based PID control system design.

1. Find the IMC controller transfer function, q(s), which includes a filter, f(s), to make q(s)
semiproper or to give it derivative action [the order of the numerator of q(s) is one order
greater that the denominator of q(s)]. Notice that this is a major difference from the IMC
procedure. Here, we may allow q(s) to be improper, in order to find an equivalent PID
controller. For integrating or unstable processes, or for better disturbance rejection, a filter
with the following form will often be used

2. Find the equivalent standard feedback controller using the transformation

Equation 9.1
 Write this in the form of a ratio between two polynomials.

3. Show this in PID form and find kc, tI, tD. Sometimes this procedure results in an ideal PID
controller cascaded with a first-order filter, with a filter time constant (tF):

Equation 9.2

4. Perform closed-loop simulations for both the perfect model case and cases with model
mismatch. Adjust l considering a tradeoff between performance and robustness (sensitivity
to model error). Initial values for l will generally be around 1/3 to 1/2 the dominant time
constant.

[ Team LiB ]
[ Team LiB ]

9.3 IMC-Based Feedback Design for Delay-Free

Processes
The IMC-based PID procedure, like the IMC procedure from Chapter 8, can be based on either
good setpoint tracking or improved disturbance rejection. Here, we first consider design for
setpoint tracking.

Focus on Setpoint Tracking

The procedure outlined in Section 9.2 is illustrated by way of two examples: a first-order process
and a second-order process. For clarity we drop the (~) notation on all of the process model
parameters.

Example 9.1: IMC-Based PID Design for a First-Order Process

Find the PID-equivalent to IMC for a first-order process

Step 1. Find the IMC controller transfer function, q(s), which includes a filter to make q(s)
semiproper.

Equation 9.3

Step 2. Find the equivalent standard feedback controller using the transformation

Equation 9.4

Recall that the transfer function for a PI controller is

 Equation 9.5

Step 3. Rearrange Equation (9.4) to fit the form of Equation (9.5), to find how the model
parameters and l are related to the PI controller parameters. Multiplying Equation (9.4) by
tp/tp, we find

Equation 9.6

Equating terms in Equations (9.5) and (9.6), we find the following values for the PI tuning
parameters

Equation 9.7

The IMC-based PID design procedure for a first-order process has resulted in a PI control law. The
major difference is that there are no longer two degrees of freedom in the tuning parameters (kc,
tI)—the IMC-based procedure shows that only the proportional gain needs to be adjusted. The
integral time is simply set equal to the process time constant. Notice that the proportional gain is
inversely related to l, which makes sense. If l is small (closed loop is "fast"), the controller gain
must be large. Similarly, if l is large (closed loop is "slow"), the controller gain must be small.
Also notice that the same results were obtained using the direct synthesis method—a specified
first-order closed-loop response for a first-order process leads to a PI controller with the
parameter values shown in Equation (9.7).

This procedure can be used to develop the equivalent PID (plus lag, in some cases) controller for
a number of other transfer functions, as shown in Table 9-1. In the next example, we derive the
PID controller for a second-order process.

Example 9.2: IMC-Based PID Design for a Second-Order Process

Find the PID equivalent to IMC for a second-order process

Step 1. Find the IMC controller transfer function, q(s)—here we allow q(s) to be improper
because we wish to end up with an ideal PID controller.
 Equation 9.8

Step 2. Find the equivalent standard feedback controller using the transformation

Equation 9.9

Recall that the transfer function for an ideal PID controller is

Equation 9.10

Step 3. Rearrange Equation (9.9) to fit the form of Equation (9.10), to find how the model
parameters and l are related to the PID controller parameters. Multiplying Equation (9.9)
by (t1 + t2)/(t1 + t2), we find

Equation 9.11

Equating terms in Equations (9.10) and (9.11), we find the following relationships:

which is shown as row C in Table 9-1.

The IMC-based PID controllers for first- and second-order processes, as designed above, have
good setpoint tracking characteristics. Although setpoint tracking may be quite good, disturbance
rejection can be poor, particularly for input disturbances. Next, we apply the IMC-based PID
procedure to develop good disturbance rejection.

Focus on Disturbance Rejection

For improved disturbance rejection, we use an IMC filter with the form

where g is selected to achieve good disturbance rejection. In practice, it is selected to cancel a

slow disturbance time constant. Consider the closed-loop transfer function for disturbance
rejection in the case of a perfect model (as developed in Section 8.10):

The internal model controller, using the new filter form, is

so the output response is

Equation 9.12

For a first-order process and an input disturbance, in Chapter 8 (Example 8.8), we found that the
internal model controller

Using the procedure in Equation (9.12) results in a numerator filter term

The reader should show that the IMC-based PID procedure leads to a PI controller with the
parameter values

and

with the requirement that l < 2tp.

Integrating Processes
For integrating processes, a filter with the following form is used:

The reader should be able to derive the controllers shown in elements I and K of Table 9-1.

Summary for Delay-Free Processes

The results for the previous examples are shown in Table 9-1. To develop a more complete
understanding of this procedure, the reader should derive the parameters for some of the other
relationships shown in Table 9-1. The reader should also note that the same results would be
obtained using the direct synthesis approach if the proper desired closed-loop transfer function is
specified. The proper desired closed-loop transfer function appears clearly in the IMC procedure.

Notice that Table 9-1 is for process transfer functions that do not have a time delay. The following
section develops PID tuning relationships for a first-order + time-delay process.

[ Team LiB ]
[ Team LiB ]

9.4 IMC-Based Feedback Design for Processes with a

Time Delay
In order to arrive at a PID equivalent form for processes with a time delay, we must make some
approximation to the dead time. (Please note that we do not use an approximation for dead time
when implementing the "pure" IMC strategy developed in Chapter 8. We are doing this only for
the IMC-based PID procedure presented in this chapter; this is only done to yield a PID-type
controller.) We use either a first-order Padé approximation for dead time or neglect dead time
entirely and use the first-order transfer function results.

First-Order + Dead Time

Since first-order + dead time is the most common representation of chemical process dynamics,
the PID equivalent form developed here is useful for a large number of process control loops. We
develop the procedure by way of example.

Example 9.3: IMC-Based PID Design for a First-Order + Dead Time

Process
Find the PID controller which approximates IMC for a first-order + time-delay process

Equation 9.13

Step 1. Use a first-order Padé approximation for dead time ,

Equation 9.14

Step 2. Factor out the noninvertible elements (this time do not make the bad part all-
pass),

Equation 9.15
 Step 3. Form the idealized controller,

Equation 9.16

Step 4. Add the filter—this time we will not make q(s) proper because a PID controller will
not result. We use the "derivative" option, where we allow the numerator of q(s) to be one
order higher than the denominator (NOTE: This is done only so that we will obtain an ideal
PID controller).

Equation 9.17

Now, find the PID equivalent. Recall that

Equation 9.18

We can expand the numerator term to find

Equation 9.19

We can multiply Equation (9.19) by (t p + 0.5)/(q/tp + 0.5q) to find the PID parameters

Equation 9.20
The IMC-based PID controller design procedure has resulted in a PID controller, when the process
is first-order + dead time. Remember that a Padé approximation for dead time was used in this
development, meaning that the filter factor (l) cannot be made arbitrarily small; therefore, there
will be performance limitations to the IMC-based PID strategy that do not occur in the IMC
strategy. Rivera et al. (1986) recommend that l > 0.8q because of the model uncertainty due to
the Padé approximation.

In Example 9.3 the all-pass formulation was not used. The reader should show that the use of an
all-pass in the factorization will lead to a PID controller in series with a first-order lag. The
parameters, in this case, are shown as the first entry in Table 9-2. Morari and Zafiriou (1989)
recommend l > 0.25q for the PID + lag formulation. The third and fourth entries neglect the time
delay in forming the PI controller.

Summary of PI(D) Control of First-Order + Time-Delay Processes

Different assumptions are used to derive the PI and PID controllers shown in Table 9-2. A zeroth-
order Padé approximation leads to a PI controller while a first-order Padé approximation leads to
PID + lag (when an all-pass filter is used) and PID (when the all-pass is not used) controllers.
Generally, the PID + lag controller will be easier to tune for robustness and will certainly be less
sensitive to noise than the PID controller. The PID + filter perfomance is almost identical to the
"pure" IMC. This shows that the powerful IMC framework can be used to design PID-type
controllers that can be implemented in industrial processes using existing (PID) control
equipment.

Integrator + Dead Time

For processes where the time constant is dominant, the step response behavior can be
approximated as integrator + dead time, as characterized by the following transfer function.

Equation 9.21

Here, assume that a Taylor series approximation for dead time is used. Also, the special filter
form for integrating systems is used.
Using the IMC-based PID procedure,

the reader should show that a PI controller results with the following parameters:

Equation 9.22

Gain + Dead Time

For processes where the time delay is dominant, the step response behavior can be approximated
as gain + dead time, as characterized by the following transfer function.

Equation 9.23

Using a second-order Padé approximation for the time delay,

the reader should find that a PID + filter controller results, with

Equation 9.24

with the condition that .

[ Team LiB ]
[ Team LiB ]

9.5 Summary of IMC-Based PID Controller Design for

Stable Processes
We have shown several examples where the IMC design procedure could be used to develop an
equivalent PID-type control law. For stable processes with no time delay, the IMC-based PID
procedure gives exactly the same feedback performance as does IMC. For stable processes with a
time delay, the IMC-based PID procedure will not give exactly the same performance as IMC,
because a Padé approximation for dead time is used in the controller design.

We want this following point to be clear to the reader. For process transfer functions without time
delays, the IMC-based PID controller will yield exactly the same performance as does IMC. This
will occur if no approximation has to be made in the process model to find a feedback form that is
equivalent to PID. If an approximation (such as Padé) is made in the IMC-based PID strategy, and
this approximation is not made in the IMC strategy, then the performance will not be the same.

It should also be noted that although the standard IMC filter form, f(s) = 1/(ls + 1)n, leads to
good setpoint tracking, it generally does not lead to good rejection of disturbances that have
dynamics similar to the process transfer function. It is generally better to use a filter of the form
f(s) = (gs + 1)/(ls + 1)n.

Table 9-1 provides a summary of the PID tuning parameters for systems without a time delay.
Table 9-2 summarizes the PID tuning parameters for stable processes with a time delay. Notice
that there are minimum recommended values for l shown, since there is inherent model
uncertainty due to the Padé approximation. Selecting l near this value will put the closed-loop
system on the edge of stability, so most often your initial tuning values for l will be significantly
larger than these minimum values.

Table 9-1. Ideal PID Tuning Parameters for Open-Loop Stable and
Integrating Processes[a].

gp(s) gCL (s) kc tI tD tF

A tp

B[b]

C t1 + t2

D 2zt

E[c]. 2zt
 gp(s) gCL (s) kc tI tD tF

F[c],[d] 2zt

G[d] 2zt

I[e] 2l

J t

K[e] 2l + t

[a]Parameters for other process transfer functions are given in Rivera et al. (1986) and Morari and Zafiriou
(1989)

[b]
The controller is designed for improved input disturbance rejection; . Notice that we desire g
> 0, which leads to l < 2tp.

[c] The controller is PID + lag,.

[d] It is assumed that b > 0 (inverse response, RHP zeros).

[e]The controllers are designed for ramp setpoint changes. This also generally leads to better input
disturbance rejection.

Table 9-2. PID Tuning Parameters for Stable Time-Delay Processes[a]

gp(s) kc tI tD tF Notes[b]

A 1

B 2

C tp 3
 gp(s) kc tI tD tF Notes[b]

D 4

E 2l + q 5

F 2l + q 6

G ke–qs 7

H ke–qs 8

[a] Based on the ideal PID controller transfer function: .

[b]IMC-based PID is based on a first-order Padé approximation, unless otherwise noted. 1, With an all-pass
factorization and semiproper q(s) recommended l > 0.25q. 2, Without an all-pass factorization and
improper q(s); recommended l > 0.8q. 3, Time-delay neglected; recommended l > 1.7q. 4, Time-delay
neglected; effective time constant increased by 0.5q; recommended l > 1.7q. 5, Using a Taylor series
approximation for time delay (-qs + 1). 6, With an improper q(s). 7, Using a first order Padé approximation
for time delay. 8, Using a second order Padé approximation for time delay. In all cases it is recommended
that l > 0.2tp.

[ Team LiB ]
[ Team LiB ]

9.6 IMC-Based PID Controller Design for Unstable

Processes
The IMC procedure must be modified for unstable processes. Rotstein and Lewin (1991) have
used the procedure developed by Morari and Zafiriou (1989) to find IMC-based PID controllers for
unstable processes. The modification to the procedure shown in Sections 9.2 and 9.3 is to use a
slightly more complicated filter transfer function.

1. Find the IMC controller transfer function, q(s), which includes a filter, f(s), to make q(s)
semiproper. An additional requirement is that the value of f(s) at s = pu (where pu is an
unstable pole) must be 1. That is,

Morari and Zafiriou (1989) recommend a filter transfer function that has the form

where n is chosen to make q(s) proper (usually semiproper). A value of g is found that
satisfies the filter requirement f(s = pu) = 1.

2. Find the equivalent standard feedback controller using the transformation

Write this in the form of a ratio between two polynomials.

3. Show this in PID form and find kc, tI, tD. Sometimes this procedure results in a PID
controller cascaded with a lag term (tF):

4. Perform closed-loop simulations for both the perfect model case and cases with model
mismatch. Choose the desired value for l as a tradeoff between performance and
robustness.

Example 9.4 illustrates this procedure for a first-order unstable process.

Example 9.4: IMC-Based PID Design for a First-Order Unstable
Process
Find the IMC-based PID controller for a first-order unstable process

Equation 9.25

where tu is given a positive value. The pole, pu, is 1/tu, which is positive, indicating instability.

Step 1. Find the IMC controller transfer function, q(s),

Equation 9.26

Note that we have selected a second-order polynomial filter to make the controller, q(s),
semiproper. Now we solve for g so that f(s = pu) = 1.

so

Solving for g, we find

Equation 9.27

Step 2. Find the equivalent standard feedback controller using the transformation
 After a lengthy bit of algebra,

Equation 9.28

Step 3. This is in the form of a PI controller, where

Equation 9.29

As a numerical example, consider . The closed loop output responses for various
values of l are shown in Figure 9-7a, while the manipulated variable responses are shown in
Figure 9-7b. Notice that we do not achieve the nice overdamped-type of closed-loop output
responses that we were able to obtain with open-loop stable processes. The reader should show
that the closed-loop relationship for this system is

which has overshoot if g > l [this is always the case for this system; see Equation 9.27)].

In Figure 9-7 we notice that the closed-loop response had overshoot. A response to a setpoint
change, without overshoot, can be obtained by including a setpoint filter, as shown in Figure 9-8.
The setpoint filter is

Figure 9-7. Responses for a step setpoint change, for various values of
l. (a) Output; (b) manipulated input.
 Figure 9-8. FBC with a setpoint filter.

which yields the following closed-loop response for a setpoint change.

The results for several unstable process transfer functions are shown in Table 9-3. See Rotstein
and Lewin (1991) for a discussion of the effect of dead time and model uncertainty on the control
of unstable processes.

Summary of IMC-Based PID Controller Design for Unstable Processes

A major tuning consideration by the student is that there are both upper and lower bounds on l to
assure stability of an unstable process. This is in contrast to stable processes, where the closed
loop is guaranteed to be stable under model uncertainty, simply by increasing l to a large value
(detuning the controller).

[ Team LiB ]
[ Team LiB ]

9.7 Summary
We have shown how the IMC procedure can be used to design PID-type feedback controllers. If
the process has no time delay and the inputs do not hit a constraint, then the IMC-based PID
controllers will have the same performance as does IMC. If there is dead time, then the IMC-
based PID controllers will not perform as well as IMC because of the Padé approximation for
deadtime.

Table 9-3. PID Tuning Parameters for Unstable Processes

gp(s) gCL(s) kc tI tD Notes[a]

A g 1

B g + tp 1

C g 1, 2

[a] 1, . 2, PI cascaded with a lead-lag filter, .

It is interesting to note that the IMC-based PID controllers for all of the transfer functions shown
in Table 9-1 could have been designed using the direct synthesis method of Chapter 6. The key
issue in the direct synthesis method is the specification of the closed-loop response characteristic.
If the process has a RHP zero, then the specified closed-loop response must also have a RHP
zero. The IMC-based PID procedure provides a clear method for handling this. Also, note that the
standard IMC filter design results in good setpoint response performance, but other filter designs
must be used for good input disturbance rejection.

The IMC design method of Chapter 8 was modified to handle unstable processes. The standard
IMC structure cannot handle unstable processes, so the controller for an unstable process must
be implemented in standard PID feedback form.

For a good example application of the IMC-based PID procedure, work through Module 7 on
Biochemical Reactors.

IMC-Based PID Procedure Summary

1. Find the IMC controller transfer function, q(s), which includes a filter, f(s). The controller,
q(s), may be semiproper or even improper to give the resulting PID controller derivative
action. If a process model has a time delay, use an approximation, such as a first- or
second-order Padé.
 For good setpoint tracking, a filter with the following form is generally used

For improved disturbance rejection, or for integrating and open-loop unstable processes, a
filter with the following form is used:

For disturbance rejection, g is selected so that the term inside the brackets cancels a slow
pole in the distubance transfer function [often gd(s) = gp(s) for input disturbances],

For unstable processes, a value of g is found that satisfies the filter requirement f(s = pu) =
1.

2. Find the IMC-based PID controller using the transformation

Write this in the form of a ratio between two polynomials.

3. Show this in PID form and find kc, tI, tD. Sometimes this procedure results in an ideal PID
controller cascaded with a first-order filter, with a filter time constant (tF):

4. Perform closed-loop simulations for both the perfect model case and the cases with model
mismatch. Adjust the value of l on-line as a tradeoff between performance and robustness
(sensitivity to model error).

[ Team LiB ]
[ Team LiB ]

References

The following paper derives IMC-based PID controllers for "real PID" control: Chien, I.-L., and P.
S. Fruehauf, "Consider IMC Tuning to Improve Controller Performance," Chem. Eng. Prog.,
86(10), 33–41 (1990).

The IMC procedure is fully developed in the following monograph: Morari, M., and E. Zafiriou,
Robust Process Control, Prentice-Hall, Englewood Cliffs, NJ (1989).

Initial IMC-based tuning parameters were presented in the following paper: Rivera, D. E., M.
Morari, and S. Skogestad, "Internal Model Control. 4. PID Controller Design," Ind. Eng. Chem.
Proc. Des. Dev., 25, 252–265 (1986).

The following paper discusses the IMC-based tuning rules for open-loop unstable systems:
Rotstein, G. E., and D. R. Lewin, "Simple PI and PID Tuning for Open-Loop Unstable Systems,"
Ind. Eng. Chem. Res., 30, 1864–1869 (1991).

[ Team LiB ]
[ Team LiB ]

Student Exercises
Exercises 1, 19, and 20 are SIMULINK-based problems.

1: Use the IMC-based PID design procedure to find the PID controller for a second-order
transfer function

Hint: First design the IMC, assuming that you will allow the controller to be improper;
that is, the order of the numerator is one higher than the denominator. What order
filter do you find?

a. Find the PID parameters, kc, tI, tD, as a function of the process model
parameters, kp, t, z, and the filter factor, l.

b. For a perfect model, plot the closed-loop response of y to a step setpoint change
in r (as a function of l). (Show your control block diagram.)

c. Discuss the effect of uncertainty in z (show plots) on closed-loop performance, as

a function of l.

2: Compare the response of the following first-order + time-delay process, using IMC and
IMC-based PID. Discuss the effect of l on the closed-loop stability for both systems.

Do you find that there is a minimum l required for the stability of the IMC-based PID
strategy? How does this relate to the recommendations in Table 9-2? Is there a
minimum l required for the stability of the IMC strategy?

3: In Example 9.3, the all-pass formulation was not used for the first-order + time-delay
process with a Padé approximation for dead time. Show that the use of an all-pass in
the factorization [and semi-proper q(s)] leads to a PID controller in series with a first
order lag.

4: Use the following inverse response process,

 a. Find the IMC-based PID controller if no all-pass is used and the controller is
improper.

b. Find the IMC-based PID controller (which must be cascaded with a first-order lag)
if an all-pass is used and the IMC controller, q(s), is semiproper.

5: For the following integrating process,

show that the IMC-based PID procedure results in a P controller if

and in a PI controller if .

6: The following process transfer function represents the relationship between boiler
feedwater flow rate and steam drum level:

a. Use the all-pass factorization method to show that the IMC-based PID controller
is PD with a first-order lag.

b. Use and a non-"all-pass" factorization to find a PID controller.

7: Use the IMC-based feedback controller design procedure to design a PID controller for
the following process.

Assume that the IMC filter is . Find kc, tI and tD—these will be a function
of the process parameters and l (show all work). What is the closed-loop transfer
function? Sketch the expected response of the output to a unit step setpoint change.

8: Consider the following first-order + time-delay process

 Find the tuning parameters for the IMC-based PID controller (no lag). What is the
maximum value of kc that you would recommend? Why?

9: Consider Example 9.2, where q(s) was allowed to be improper. This yielded a PID
controller. Now assume that q(s) is forced to be proper (actually semiproper). Find the
resulting feedback controller. Elaborate on the control structure.

10: A stack gas scrubber has the following relationship between the fresh feedwater flow
rate and the SO2 concentration in the water leaving the scrubber:

Use the IMC-based PID procedure to find the PID controller (plus first-order lag) for
this process.

Hint: Let the IMC q(s) be semiproper (order of denominator = order of numerator).
Which controller tuning parameters does l affect?

11: Use the IMC-based feedback controller design procedure to design a PID-type
controller for the following process. Assume that q(s) is semiproper and tn > 0.

Find the PID tuning parameters (assuming ideal PID) as a function of the process
parameters and l (show all work). What is the closed-loop transfer function? Sketch
the expected response behavior.

12: Consider the closed-loop response for IMC when the model is not perfect. Show that
there is no offset for a setpoint change for the following model and process transfer
functions. Also, find the minimum value of l that assures closed-loop stability when

13: Show that the IMC strategy cannot be implemented on an unstable process and must
be implemented in standard feedback form.

14: A styrene polymerization reactor is operated at an open-loop unstable point and has
the following input-output model
 Design the IMC-based PI controller for this system. Sketch the expected response for a
setpoint change.

15: A vinyl-acetate polymerization reactor is operated at an open-loop unstable point and

has the following input-output model, for control of output temperature (°C) by
manipulating the jacket temperature (°C). The timescale is minutes.

a. Design the IMC-based PID controller for this system; that is, find kc, tI, and tD.
Show units for all parameters.

b. What is the closed-loop transfer function? Sketch the expected response for a
step setpoint change.

16: Derive g for the first three elements in Table 9-3. The value of the filter must be one at
the location of the unstable pole.

17: A reactor is operated at an open-loop unstable point and has the following input-output
model:

a. Design the IMC-based PID controller (perhaps with a lead-lag filter) for this
system, that is, find the tuning parameters.

b. What is the closed-loop transfer function? Sketch the expected response for a
step setpoint change.

18: Show that a PI controller cannot stabilize the process

if tu < tp. (Hint: Use the Routh stability criterion.)

The IMC-based PID controller (entry B in Table 9-3) can handle this process.

19: Consider a first-order + dead-time process with the following transfer function (use
this to represent the process in all analyses and simulations)

Now assume two different PI design procedures for this process: The model is first-
order + dead time with the known parameter values; and the model is integrator +
dead time where k = 1/10 = 0.1.

a.
 a. For a small time delay (0.5 minutes), compare the performance of the two PI
control strategies for step setpoint changes. Discuss the effect of l. Also compare
the performance of the two PI control strategies for step input load disturbance
changes (let the load disturbance transfer function be equal to the process
transfer function). Discuss the effect of l.

b. For a larger time delay (5 minutes), compare the performance of the two PI
control strategies for step setpoint changes. Discuss the effect of l. Also compare
the performance of the two PI control strategies for step input load disturbance
changes (let the load disturbance transfer function be equal to the process
transfer function). Discuss the effect of l.

c. Summarize your results and make recommendations.

20: Consider a first-order + dead time process with the following transfer function (use
this to represent the process in all analysis and simulations):

Compare the setpoint performances of PI, improved PI, PID, and PID + filter tuning
rules from Table 9-2 (rows A–D). Discuss the effect of the closed-loop time constant,
l; for what values is unstable or poor performance obtained?

21: Consider the following process model with a RHP zero

Use the IMC-based PID design procedure to find a PI controller. Assume that an all-
pass factorization is used for the RHP zero. Also, assume that q(s) is semiproper. Use
an IMC filter with the form

and solve for g to give exactly a PI controller (no other terms).

Find the proportional gain and integral time as a function of the model parameters (k,
b) and l.

[ Team LiB ]
[ Team LiB ]

Chapter 10. Cascade and Feed-Forward

Control
The objective of this chapter is to develop feed-forward- and cascade-control system design
procedures. The emphasis of both of these strategies is on rejecting disturbances. After studying
this chapter, the reader should be able to do the following:

Given an instrumentation diagram for a cascade-control strategy, develop the corresponding

control block diagram using either a series or parallel structure

Tune cascade controllers; tune the secondary loop first; then tune the primary controller
with the secondary loop closed

Given an instrumentation diagram for feed-forward/feedback control, develop the

corresponding control block diagram

Develop the closed-loop transfer functions for feed-forward/feedback control

Design feed-forward/feedback controllers. Understand physical realizability limitations due to

time delays and RHP zeros

The major sections of this chapter are as follows:

10.1 Background

10.2 Introduction to Cascade Control

10.3 Cascade-Control Analysis

10.4 Cascade-Control Design

10.5 Cascade IMC

10.6 Feed-Forward Control

10.7 Feed-Forward Controller Design

10.8 Feed-Forward Control in the IMC Structure

10.9 Summary of Feed-Forward Control

10.10 Combined Feed-Forward and Cascade

10.11 Summary

[ Team LiB ]
[ Team LiB ]

10.1 Background
Thus far in this textbook we have emphasized control-system design for single input–single output
(SISO) processes; that is, processes with one output (measured) variable and one manipulated
variable. We have also focused on setpoint responses, primarily because it is then easier to tune a
feedback controller. In particular, it is easier to tune on-line for setpoint changes because we do
not know when a disturbance is going to enter a system. We have developed a number of control-
system design procedures that are based on a desired response to a setpoint change.

In practice, disturbance rejection is very important. The primary disadvantage to feedback-only

control is that a disturbance must be "felt" by the output variable before there is a control-system
response. The purpose of this chapter is to show how to use multiple measurements to improve
the response to a disturbance. In Sections 10.2 through 10.5, we study cascade control. In
cascade control, multiple output measurements are used to improve the response of the most
important (primary) output to a disturbance. In the last part of the chapter (Sections 10.6 to
10.8) we cover feed-forward control, where the measurement of a disturbance is directly used to
improve the system response to the disturbance. In Section 10.10 we combine feed-forward and
cascade control.

[ Team LiB ]
[ Team LiB ]

10.2 Introduction to Cascade Control

Cascade control involves the use of multiple measurements and a single manipulated input. As a
motivating example, consider the temperature-control problem shown in Figure 10-1, where a
fired furnace is used to heat a process fluid stream. The outlet temperature is controlled by
manipulating the valve position of the fuel gas control valve. Clearly, disturbances in the fuel gas
header pressure (upstream of the valve) will end up changing the fuel gas flow rate, and,
therefore, the outlet temperature. Also, any problems with the control valve, such as stiction or
hysteresis (see Module 15), will affect the fuel gas flow rate.

Figure 10-1. Feedback control of process outlet temperature.

Cascade to Flow Control

The best way to compensate for disturbances directly affecting the fuel gas flow rate is to cascade
the temperature controller to a fuel gas flow controller, as shown in Figure 10-2. Here, the output
of the temperature controller is the setpoint to the fuel gas flow controller. The output of the flow
controller is the pressure to the control valve, which changes the valve position and, therefore,
the flow rate. Any change in the fuel gas header pressure will be "felt" immediately by the flow
measurement[1], allowing the flow controller to take immediate corrective action. A block diagram
of this strategy is shown in Figure 10-3.
[1]Note that the flowmeter has been placed upstream of the control valve. This is common practice because
a length of straight-run piping is needed upstream of the flowmeter for reasonable accuracy. Also, the
pressure upstream of the valve is relatively constant (making the fluid density relatively constant),
 compared to the downstream pressure.

Figure 10-2. Cascade control of process outlet temperature.

Figure 10-3. Cascade control block diagram for the fired heater. Series
transfer function form.
In this strategy, the temperature controller is known as the primary, master, or outer-loop
controller, while the flow controller is the secondary, slave, or inner-loop controller. The dynamics
of the flow control loop are very fast, making the flow controller easy to tune. The temperature-
control loop is much slower, so the primary loop can be effectively tuned as if the flow controller
response is instantaneous.

For cascade control strategies where flow control is the inner loop, a control block diagram where
the secondary and primary process transfer functions are in series (as in Figure 10-3) is natural.
The primary disturbance (process inlet temperature) directly affects the primary output (process
outlet temperature), with no direct effect on the secondary output (fuel gas flow rate, which is
also the manipulated input for the primary process). The secondary disturbance (fuel gas header
pressure) directly affects the secondary output, so the control block diagram shown in Figure 10-3
is relatively straightforward to generate. This is known as a series representation for the block
diagram that is useful when the "secondary" process (such as the control valve or fuel gas flow
rate) are naturally separated from the "primary" process (the heater and process fluid). It should
be noted that other primary disturbances, such as process fluid flow rate, could easily be added to
the diagram. See Exercise 9 for a numerical study of this process.

An example of the improved performance achievable with cascade control is shown in Figure 10-
4, where a 1-psig disturbance in fuel gas header pressure occurs at t = 0 minutes. In the
cascade-control strategy (Figure 10-2), the flow controller (secondary controller) rejects the
header pressure disturbance before it is even "felt" by the process outlet temperature, since there
is a primary process time-delay of one minute. The standard feedback controller (Figure 10-1)
has poor performance due to the process time delay. It should be noted that setpoint changes
and responses to a primary disturbance (process feed temperature) are the same for both control
strategies (see Exercise 9 for a numerical study).

Figure 10-4. Response to a disturbance in fuel gas header pressure of

1 psig. Comparison of cascade control with standard feedback control.
Reactor Temperature Cascade Control
As another example of cascade control, consider the stirred-tank reactor shown in Figure 10-5. In
the standard feedback-control strategy the temperature of the reactor is measured and the jacket
flow rate is manipulated. If there is a disturbance in the jacket feed temperature, it will affect the
jacket temperature, which will affect the reactor temperature. In the cascade-control strategy
(Figure 10-6) the temperature of the reactor is measured and compared with the desired reactor
temperature. The output of this reactor temperature controller is a setpoint to the jacket
temperature controller. The jacket temperature controller manipulates the jacket flow rate. Notice
that two measurements (reactor temperature and jacket temperature) are made but only one
manipulated variable (jacket flow rate) is ultimately adjusted.

Figure 10-5. Stirred-tank reactor. Standard feedback control.

 Figure 10-6. Stirred-tank reactor. Cascade control.

In this strategy, the reactor temperature controller is the primary (master or outer loop)
controller, while the jacket temperature controller is the secondary (slave or inner loop)
controller. This is effective because the jacket temperature dynamics are normally significantly
faster than the reactor temperature dynamics. An inner-loop disturbance, such as jacket feed
temperature, will be "felt" by the jacket temperature before it has a significant effect on the
reactor temperature. This inner-loop (secondary) controller then adjusts the manipulated variable
before a substantial effect on the primary output has occurred.
Notice that an appropriate block diagram representation of cascade control is less obvious here.
Jacket flow rate certainly has a direct effect on the jacket temperature, which then affects the
reactor temperature; one would think that a series representation would again be appropriate.
Notice, however, that the jacket and reactor temperatures interact in a way that is different than
the primary and secondary processes of the fired heater. Here, the reactor and jacket
temperatures are directly coupled by the heat transfer between them. A change in the reactor
temperature causes a change in jacket temperature and vice versa; a simple input-output
transfer function block diagram does not truly capture this behavior. A more appropriate
representation is shown in Figure 10-7, where a state space model is used to represent the
interactions, and the temperatures are simply viewed as two outputs from the process. See
Exercise 10 for a numerical study of a similar stirred-tank heater problem.

Figure 10-7. Cascade control block diagram for a jacketed chemical

reactor. State space form for the reactor.

[ Team LiB ]
[ Team LiB ]

10.3 Cascade-Control Analysis

There are two common ways to represent a cascade-control system in transfer function form:
series cascade and parallel cascade. The series cascade representation shown in Figure 10-8 is
the most common, so it will be used in the analysis performed in this section. Students interested
in analysis using the parallel structure can work Exercise 11.

Figure 10-8. Series cascade structure.

We use the following algebraic manipulations to understand the effect of the secondary (inner-
loop) on the primary (or outer) loop. Notice that the secondary output can be written

Equation 10.1

The secondary closed-loop transfer function can be defined as

Equation 10.2

Further analysis yields

Equation 10.3
After tuning the inner loop, we can use the following transfer function to design the outer-loop
controller.

Equation 10.4

and the closed-loop relationship for a primary setpoint change is

Equation 10.5

where it is clear that the secondary closed-loop transfer function affects the primary control loop.
Notice that if the secondary control loop is much faster than the primary loop, so that gc2CL = 1
(on a relative time scale to the primary control loop), then the closed-loop transfer function for
the primary loop is

[ Team LiB ]
[ Team LiB ]

10.4 Cascade-Control Design

Tuning the cascade controller consists of two steps. First, tune the inner-loop controller [gc2 (s)]
based on the secondary process [gp2 (s)] and use the inner-loop/closed-loop transfer function
[Equation (10.2)] to find an effective process transfer function [Equation(10.4)] for tuning the
outer-loop controller [gc1 (s)].

The inner-loop control system design can be based on IMC-based PID or any other procedure,
such as frequency-response gain and phase margins. Once the inner-loop is tuned, the effective
outer-loop transfer function can be used to tune the outer-loop controller. Again, any control
design procedure can be used.

Figure 10-9. Series cascade structure using a closed-loop transfer

function for the inner-loop.

Rules of Thumb for Cascade Control

Cascade control can be successfully used to reject secondary-process disturbances when the
primary process has a much larger time constant and a time delay or RHP zero while the
secondary process has a small time constant and little or no dead time or nonminimum-
phase (RHP zeros) behavior.

The most common cascade-control loop involves a flow controller as the inner loop. This
type of loop easily rejects disturbances in fluid stream pressure, either upstream or
downstream of the valve.

The inner loop (secondary, slave loop) in a cascade-control strategy should be tuned before
the outer loop (primary, master loop). After the inner loop is tuned, and closed, the outer
loop should be tuned using knowledge of the dynamics of the inner loop.

Cascade control is widely used on chemical processes, and should not be considered an
"advanced control strategy."

There is little or no advantage to using cascade control if the secondary process is not
 significantly faster than the primary process dynamics. In particular, if there is much dead
time in the secondary process, or if there is an RHP zero, it is unlikely that cascade control
will be much better than standard feedback control.

Cascade control can be easily combined with other forms of control, such as feed-forward
control.

Any of your favorite methods can be used for tuning the inner and outer loops. Indeed, the
IMC strategy can be used on either or both of the inner and outer loops.

[ Team LiB ]
[ Team LiB ]

10.5 Cascade IMC

IMC can also be implemented in a cascade structure. Frequently, the primary controller is IMC,
while the secondary controller is PID; this is particularly true when the secondary control loop
involves flow control, as shown in Figure 10-10. Notice that the measured flow implemented on
the process is also used on the model (recall that the model output is obtained by integrating the
model differential equations). If the control valve hits a constraint and cannot achieve the desired
flow rate (flow setpoint), then the process model "knows" this and a realistic model output is
predicted. This minimizes a problem known as "reset windup," which is sometime associated with
the manipulated inputs hitting a constraint (this will be discussed further in Chapter 11).

Figure 10-10. Cascade IMC outer-loop, PID flow-control inner-loop.

The internal model controller should be designed based on a model that includes the flow control
loop dynamics

The reader should also be able to develop a control block diagram for cascade control where both
the primary and secondary loops use IMC (see Exercise 12).

[ Team LiB ]
[ Team LiB ]

10.6 Feed-Forward Control

When a disturbance affects a process under feedback control it is necessary for a measured
process output to change before corrective action is taken to change the manipulated input. It
would be preferable to have a sensor that measures the disturbance and adjusts the manipulated
input before the process output changes. Consider the furnace system studied earlier in this
chapter and shown in Figure 10-11. One possible disturbance is the process fluid flow rate; if the
flow rate increases 20%, then the required heat duty should also increase by 20%. The purpose
of the feed-forward control strategy shown is to immediately change the fuel gas flow rate (for
simplicity we assume that the fuel gas flow rate is directly manipulated in this diagram) when a
change in the process flow rate is sensed.

Figure 10-11. Feed-forward control of temperature based on the

measured process fluid flow rate (disturbance input).

For this furnace example, the steady-state energy balance can be used to find the fuel gas flow
rate
where F = process stream flow rate, r = density, cp = heat capacity, T = desired outlet
temperature, Tf = feed temperature,
e = furnace efficiency, and DHc = heat of combustion of fuel
gas. A steady-state feed-forward controller computes the fuel gas flow rate based on the
measured process flow rate (and possibly the process feed temperature). Sometimes it is
important to also consider dynamic effects, as shown in several examples that follow.

The performance of feed-forward control is limited by model uncertainty. Without a perfect

process model the controller cannot completely compensate for disturbances and there will
generally be offset. In practice, feed-forward control is combined with feedback control, as shown
in Figure 10-12. Here the feed-forward controller compensates for process flow rate disturbances,
while the feedback controller compensates for other disturbances and model uncertainty. The
corresponding block diagram is shown in Figure 10-13.

Figure 10-12. Feed-forward/feedback control of temperature based on

the measured process fluid flow rate (disturbance input).

Figure 10-13. Feed-forward/feedback control block diagram for the

furnace example.
Consider the feed-forward/feedback block diagram shown in Figure 10-13. We can develop the
closed-loop transfer function by using the following steps:

Rearranging, to solve for y(s), we find

or

Equation 10.6

Equation (10.6) is the closed-loop relationship for feed-forward/feedback control. Notice that a
(stable) feed-forward controller does not affect the closed-loop stability, since the feed-forward
controller transfer function does not appear in the closed-loop characteristic equation
(denominator polynomial).

[ Team LiB ]
[ Team LiB ]

10.7 Feed-Forward Controller Design

Remember that we are working in deviation variables. Assume that we are not making a setpoint
change, so r(s) = 0. Since we do not want the output variable to change, that is equivalent to
saying that we desire y(s) = 0. From Equation (10.6) we can see that the only way to assure y(s)
= 0 is for the following relationship to hold

Equation 10.7

In this expression all the transfer functions are determined for a particular system, except for the
feed-forward controller, gcf(s). Solving Equation (10.7) for gcf(s), we find

Equation 10.8

If we make the simplifying assumption that the disturbance measurement has no dynamics, then

Equation 10.9

Notice that Equation (10.9) requires the inverse of the process model. Clearly, there will be
problems if the process model has RHP zeros, or if the process time delay is greater than the
disturbance time delay. The use of Equation (10.9) as a feed-forward design equation is shown in
Examples 10.1–10.4.

Example 10.1: First-Order Process and Disturbance Transfer

Functions
Consider the first-order process and disturbance transfer functions

Equation 10.10
The feed-forward control law, from Equations (10.9) and (10.10), is

Equation 10.11

which is simply a lead-lag controller. Most control systems have feed-forward controllers that are
of lead-lag form.

Numerical Example
Here we study the case where kd = kp = 1 and td = tp = 5 minutes. If we use the IMC-based PI
procedure for feedback control, we find that tI = 5 and kc = 2 for l = 2.5 minutes. We compare
the disturbance rejection of feedback-only with feed-forward/feedback control in Figure 10-14 for
a unit step change in the load disturbance at t = 0. The feed-forward controller is a static
controller with kcf = 1 for this case. Notice that the feed-forward controller implements an
immediate change in the manipulated input to counteract the disturbance.

Figure 10-14. Comparison of feed-forward/feedback with feedback-

only for numerical example.

Example 10.2: First-Order + Dead Time Process and Disturbance

Transfer Functions
Consider the transfer functions

Equation 10.12

The feed-forward control law, from Equations (10.9) and (10.12), is

Equation 10.13

which is simply a lead-lag feed-forward controller with a time delay. The main thing that we notice
about Equation (10.13) is that qd qp is required for the controller to be realizable. If qp > qd,
we cannot have a time-delay term in the feed-forward controller and must accept the fact that we
cannot have perfect feed-forward control, even if the models are perfect.

Example 10.3: Process Higher Order than Disturbance Transfer

Function
Consider the following transfer functions:

Equation 10.14

The feed-forward controller is

Equation 10.15

Since Equation (10.15) is improper, it is not physically realizable. We could simply add a first-
order filter, perhaps something like 1/(s + 1), to make the feed-forward controller second order
over second order. Perhaps the simplest thing to do is to make the numerator term first order.
Most likely, we would use 18s + 1 to approximate the numerator to arrive at the lead-lag
controller

Equation 10.16
Numerical Simulation
Assume that an IMC-based PID controller is used for feedback control. With l = 9, we find (from
Table 9-1) that kc = 1, tI = 18, and tD = 2.5. Figure 10-15 compares the closed-loop response of
the simplified feed-forward/feedback [Equation (10.16)] with feedback-only to a unit step change
in load disturbance at t = 0. Although the feed-forward/feedback control is not perfect, the
performance is substantially better than feedback only control. This performance is achieved by
the initial manipulated variable action that is characteristic of lead-lag feed-forward controllers
with a numerator lead time constant that is greater than the denominator lag time constant (see
the lead-lag transfer function behavior in Chapter 3).

Figure 10-15. Comparison of feed-forward/feedback with feedback-

only.

Example 10.4: Process Has Inverse Response, Disturbance Does Not

Consider the process and disturbance transfer functions

Equation 10.17
The feed-forward controller is

Equation 10.18

Notice that Equation (10.18) has a RHP pole and is therefore unstable. We must factor RHP zeros
from the process transfer function before designing a feed-forward controller. In all likelihood, we
would neglect the dynamic part and use only a static feed-forward controller, as shown in
Equation (10.19) below.

Static Feed-Forward Control

Often, we will be satisfied to neglect the differences in the dynamics between a process and a
disturbance, and to simply design a static feed-forward compensator,

Equation 10.19

This will get us the majority of the performance improvement over feedback-only control.

In the next section, we show feed-forward control in the IMC structure.

[ Team LiB ]
[ Team LiB ]

10.8 Feed-Forward Control in the IMC Structure

Feed-forward control can also be implemented with an IMC structure, as shown in Figure 10-16,
where we have neglected the measurement dynamics for clarity. The feed-forward control design
is

Equation 10.20

Figure 10-16. Feed-forward control in the IMC structure.

Again, noninvertible elements of qf(s) must be removed before the feed-forward controller can be
implemented.

The closed-loop transfer relationship (the derivation is left as an exercise for the reader) is

[ Team LiB ]
[ Team LiB ]

10.9 Summary of Feed-Forward Control

We have developed feed-forward control based on models of the disturbance and process transfer
functions. If the disturbance and process models are not available, feed-forward control can still
be implemented as a tunable gain plus lead-lag controller. You should consider the following
important issues when implementing feed-forward control:

The feed-forward controller must be physically realizable and stable, as noted by the next
three items.

The disturbance time delay must be greater than the process time delay for perfect
feed-forward compensation. If the process time delay is greater than the disturbance
time delay, the feed-forward controller will have no time-delay compensation, and
perfect control cannot be achieved.

If the process has an RHP zero, it must be factored out before designing the feed-
forward controller. Otherwise, the feed-forward controller will be unstable.

If the process is higher order than the disturbance, fast time constants probably must
be neglected in designing the feed-forward controller.

Often a static feed-forward controller will have good performance, particularly if the
disturbance dynamics are the same timescale as the process dynamics.

A feed-forward controller does not change the closed-loop stability of the feedback system,
assuming the feed-forward controller is stable. Also, a feed-forward controller does not
change the setpoint response of a closed-loop system.

Feed-forward control can be implemented with either the classical feedback (PID-type) or
IMC structure.

[ Team LiB ]
[ Team LiB ]

10.10 Combined Feed-Forward and Cascade

In Section 10.6, a fired heater was used to provide motivation for feed-forward control. In the
example it was assumed that the fuel gas flow was directly manipulated. In practice a flow
controller would be used, resulting in the feed-forward, feedback, and cascade-control strategy
shown in Figure 10-17.

Figure 10-17. Feed-forward, feedback, and cascade control of a fired

heater.

Notice that each "mode" is used to reject a different type of disturbance. The feed-forward
controller is used to reject feed flow rate disturbances. The fuel gas flow controller is used to
reject disturbances in the fuel gas header pressure. The temperature controller plays a role in
compensating for all disturbances. See Exercise 19 for a numerical example.

A common process that combines feed-forward/feedback and cascade control is a steam drum.
See Module 9 for a case study of steam drum level control.

[ Team LiB ]
[ Team LiB ]

10.11 Summary
Disturbance rejection provides the motivation for cascade and feed-forward control. Cascade
control uses a secondary process measurement to detect and correct for a disturbance before it
affects the primary process output. Feed-forward control uses a direct measurement of the
disturbance to change the manipulated input before the disturbance affects the output. It is
common to combine feed-forward/feedback with cascade control, as shown in the furnace control
block diagram, Figure 10-17.

Although some texts refer to cascade and feed-forward control as "advanced control," these
techniques are commonly used in industry. Almost every control system that has a flow rate as a
manipulated variable will have a flow controller as an inner loop in a cascade strategy.

[ Team LiB ]
[ Team LiB ]

References
A number of feed-forward control examples are presented in the following sources:

Badavas, P.C., "Feedforward Methods for Process Control Systems," Chemical Engineering,
103–108 (October 15, 1984).

Stephanopoulos, G., Chemical Process Control, Prentice Hall (1984).

[ Team LiB ]
[ Team LiB ]

Student Exercises—Cascade Control

1: Derive the closed-loop transfer function relating the primary setpoint to the primary
process output for a cascade-control system.

2: Derive the closed-loop transfer function relating a secondary disturbance to the

primary output for a cascade-control system.

3: For the following process, develop the double cascade-control loop diagram, where the
jacket flow controller is a "tertiary controller." Use the state space form similar to
Figure 10-6, where jacket flow rate is an input to the state space representation of the
reactor.

4: Consider the following instrumentation diagram for a chemical reactor. Csp represents
a concentration setpoint.
 Draw the control block diagram for this system. Label every transfer function and
signal on the diagram. What type of control strategy would you call this? What is the
primary measured variable for this system? What is the primary control variable for
this system? Which controller is probably tuned to be the slowest? Which controller is
probably tuned to be the fastest? What is the gain for device D1 if a current signal
(4–20 mA) is being converted to a pneumatic signal (3–15 psig) (show units)?

5: For a particular biochemical reactor, the concentration of the microorganism (bug) is

controlled by varying the pH in the reactor. The pH is controlled by manipulating the
flow of caustic solution to the reactor. The bugs grow by "eating" a waste chemical in
the feed stream.

a. Complete the instrumentation diagram shown below. Draw the control block
diagram associated with the bug concentration control strategy.

b. What type of control strategy is this (for the bug concentration control)?

c. Identify the control objective and possible disturbance variables. Name the
disturbance variables that each controller is meant to reject.
 c.

6: As a budding young engineer in the semiconductor device manufacturing industry,

your first assignment involves a chemical vapor deposition reactor. Before you
obtained this position, the process did not have many feedback control loops. Your
goal is to design a control system that can control the growth rate of a thin film on a
wafer. A spectroscopic ellipsometer is available to measure the growth rate of this film.
The growth rate is a function of the composition [mole fraction of TMG (trimethyl
gallium)] of the inlet feed stream. The composition of the inlet feed stream is a
function of the flow rate of hydrogen that bubbles through a bottle of TMG.

Draw the following control instrumentation diagram. Now, draw the corresponding
control block diagram, in order to be able to analyze the control loops and tune the
controllers. Label all controllers, measurement devices, and signals on the block
diagram.
7: Consider the following depiction of the bottom portion of a distillation column. It is
desired to control the composition of the bottom stream at a certain value (setpoint)
by manipulating the steam to the reboiler. Composition is a slow variable, and the
measurement device can have significant delays; it is known that measuring and
controlling the temperature of a tray in the column can improve disturbance rejection.
The temperature is regulated by the pressure of the steam in the reboiler, which is a
function of the steam flow to the reboiler. Show the proper signals for a cascade type
of control system. Also, develop the corresponding block diagram, showing clearly the
disturbances that each loop is meant to reject.
8: A common type of chemical reactor is a packed-bed reactor, which is assumed to
operate adiabatically (no heat loss). The objective of the following reactor control
strategy is to maintain a desired reactor outlet concentration by manipulating the
temperature of the inlet stream to the reactor. The reactor inlet temperature is
maintained by manipulating the steam flow to a feed preheat exchanger. Connect the
measurements and control devices on the control instrumentation diagram and draw
the corresponding control block diagram. Label all controllers and signals on both the
control instrumentation and control block diagrams. Draw appropriate disturbances on
the control block diagram.

9: As a young process engineer, you have been given the responsibility for a catalytic
reforming unit at a refinery that, for some reason, is just now getting involved with
"advanced control." The catalytic reforming unit produces roughly 25% of the gasoline
 produced by the refinery. The throughput of the unit is 25,000 barrels per day
(Bbl/day) and the value is roughly $30/Bbl. The gross operating income of the unit is
then roughly $1 million/day. You feel certain that tighter control can improve the yield
of the unit, increase the value of the product, and reduce operating costs. The
reforming unit consists of three fuel gas-fired furnaces and three catalytic reactors
with a number of additional heat exchangers and separation columns. You feel that
you must prove that advanced control can work on part of the process before you
attempt to implement it on the entire process. You decide that the place to start is on
the temperature control of the process fluid leaving the first furnace.

A simplified process instrumentation diagram for the furnace that heats the process
stream for the first reactor is shown in Figure 10-1. The "old school" operations
personnel have been controlling the outlet temperature of the process stream by
simply manipulating the fuel gas valve directly. You realize that this method allows fuel
gas header pressure changes to affect the fuel gas flow rate, which will affect the
process fluid outlet temperature.

You wish to show the benefits of cascade control for this system. Use your practical
engineering knowledge, a number of discussions with the process operators
responsible for the reforming unit, an investigation of the unit log books, and
determine the following.

Relevant information:

a. For each increase of 1 psig to the fuel gas valve actuator, the fuel gas flow rate
increases by 83.3 scfm (standard cubic feet per minute).

b. For each increase of 1 scfm in fuel gas flow, there is an increase in the outlet
reformate stream temperature of 0.36°F.

c. For a change in the upstream fuel gas header pressure of 1 psig, there is a
corresponding change of 13.5 scfm in the fuel gas flow rate.

d. For a change in the inlet process stream of 1°F, there is a change of 1°F in the
outlet temperature.

e. You assume that the time constant associated with the control valve is roughly 6
seconds (0.1 minutes).

f. The process time constant associated with the furnace is approximately 5

minutes and there is an additional 1 minute time delay.

Design a cascade-control strategy and compare results (using S IMULINK) with the
standard feedback-control strategy for two types of disturbances: fuel gas header
pressure (±5 psig), and process stream inlet temperature disturbances (±25°F).

10: Consider a stirred-tank heater that has the following process model:
 with the parameter and variable values

V/F = 30 minutes

UA = 24 kcal/min °C

Vrcp = 850 kcal/°C

Vjrjcpj = 120°C

Tjf = 120°C

T = 40°C

Tf = 20°C

a. Find the state space model for these conditions, if the inputs are F j, Tjf, and Tf .

b. Show that the linear transfer function model has the form of the following block
diagram and find all of the process and disturbance transfer functions. Is a zero
of the secondary process transfer function related to the pole of the primary
process transfer function?

c. If changes in the vessel temperature are slow and can be neglected so that only
the second (jacket temperature) equation is considered, find the secondary
transfer function, gp2 (this is first-order). How does this compare with part b?

11: Consider a cascade control block diagram based on a parallel process representation
(note that the disturbances are excluded here). Show that the closed-loop transfer
function for the inner-loop is the same as for the series representation. Derive the
closed-loop transfer function for the outer-loop as a function of the inner-loop closed-
loop transfer function. Also, relate gp to gp1 and gp2 .
 12: Develop a block diagram that has the IMC structure on both the primary and
secondary loops.

[ Team LiB ]
[ Team LiB ]

Student Exercises—Feed-Forward Control

13 Use physical reasoning to discuss why the disturbance dead time must be greater than
the process dead time to even have the possibility of perfect feed-forward disturbance
compensation.

14

Let and .

Design a feed-forward controller, gcf(s), for this system. Elaborate on its expected
performance.

15 A mixing vessel is used to maintain a desired pH level in a stream flowing to a waste

treatment plant. The pressure to the valve on an acid stream is used as the
manipulated variable. Most of the variability in pH is due to waste stream 1, which is a
caustic stream. It is desirable to implement a feed-forward controller to reject the pH
disturbances due to this stream.

The following data are relevant to this problem. Without control, a change in the inlet
pH of 0.5 leads to a change of 0.25 pH in the outlet stream. The time delay is 10
minutes and the time constant is 30 minutes. A change in the acid stream valve-top
pressure of 1 psig leads to a change in pH in the outlet stream of 0.4 pH. The time
constant is 25 minutes and the time delay is 7.5 minutes.

Design a feed-forward controller for this process. Show units.

16 Derive the following closed-loop relationship for IMC with feed-forward control,
 and show that there is no offset for a step load disturbance.

[ Team LiB ]
[ Team LiB ]

Student Exercises—Feed-Forward and Cascade

17 Consider again the reactor control system from Exercise 8. If there is also a
measurement of the concentration of the reactor feed stream (before it enters the heat
exchanger), show how it can be used in a feed-forward-control strategy (using the
previously developed cascade type of strategy). Show this on both the control
instrumentation diagram below and the control block diagram.

18 Design a feed-forward/cascade control strategy for the following level control problem.
Draw the instrumentation directly on the figure. Draw a control block diagram, labeling
all the signals on the diagram

19 Consider the heater shown in Exercise 9. Now consider implementing feed-forward and
cascade control when process flow disturbances can occur, as shown below.
 For a change in the inlet process flow rate of 1,000 Bbl/day, there is a change of –10°F
in the outlet temperature. Also, the dynamic behavior can be described by a first-order
lag of 5 minutes and a one minute time delay. Use the information from Exercise 9 to
design a feed-forward/feedback/cascade-control strategy, and perform simulations for
a step feed flow rate disturbance of 2,000 Bbl/day.

[ Team LiB ]
[ Team LiB ]

Chapter 11. PID Enhancements

Commercial PID controllers generally have many features beyond the simple control algorithm. All
have antireset windup capability and many have automatic tuning and gain scheduling functions.
The objective of this chapter is to provide an overview of the common techniques used. The major
sections of this chapter are as follows:

11.1 Background

11.2 Antireset Windup

11.3 Autotuning Techniques

11.4 Nonlinear PID Control

11.5 Controller Parameter (Gain) Scheduling

11.6 Measurement/Actuator Selection

11.7 Implementing PID Enhancements in SIMULINK

11.8 Summary

[ Team LiB ]
[ Team LiB ]

11.1 Background
The focus of the techniques developed thus far in this textbook has been on linear systems,
although it is recognized that chemical processes are inherently nonlinear. Fortunately, when
processes are operated close to a desired operating point, the linear-system approaches generally
work well. Also, when the main nonlinearity is due to the control valve or flow-pressure drop
characteristics, cascade-control approaches work quite well. The inner-loop flow controller has
fast dynamics, and the nonlinear effects will usually not propagate to the outer-loop.

In this chapter we cover several different approaches to handling nonlinear effects. The first
nonlinear effect is due to manipulated input saturation. When manipulated inputs hit a constraint,
the phenomena of reset (or integral) windup can occur. Approaches for antireset windup (ARW)
are presented in Section 11.2.

When chemical processes change operating points, the dynamic input-output characteristics often
change, requiring a control system retuning. Automatic tuning procedures can be used to retune
controllers. So-called autotune procedures are developed in Section 11.3.

Sometimes it is desirable to have a controller gain that is a function of the error. If small errors
can be tolerated, then a controller with a proportional gain that is a function of the absolute value
of the error can often be used. The basic idea of nonlinear PID is introduced in Section 11.4. If the
process gain changes as a function of operating condition, then "gain scheduling" approaches can
be used to vary the controller tuning parameters as a function of the operating condition. Gain
scheduling is presented in Section 11.5. Sometimes process input-output behavior can be made
more linear by proper measurement/actuator selection. Such techniques are discussed in Section
11.6.

[ Team LiB ]
[ Team LiB ]

11.2 Antireset Windup

Reset (Integral) Windup

The phenomenon of reset windup occurs when a controller with integral action operates with the
manipulated variable at a constraint for a period of time. When the manipulated input hits the
constraint, the integral term in a PID controller continues to accumulate the error, requesting
more and more manipulated variable action; the manipulated input cannot increase, however,
since it is already constrained. Owing to this integral windup, the manipulated variable may stay
saturated longer than is necessary. A block diagram for a process with manipulated input
constraints is shown in Figure 11-1a, where v represents the unconstrained input from the
controller, while u represents the actual constrained input implemented on the process. The
"saturation" element detailed in Figure 11-1b, has the following mathematical description

Equation 11.1

Figure 11-1. Representation of a process with a constrained input.

The effect of reset windup is illustrated in the following example.

Example 11.1: Illustration of Reset Windup
Consider the first-order process,

where the time unit is minutes. Here we compare two processes: where the manipulated input is
unconstrained, and where the manipulated input is constrained between –1 and 1. Assuming
deviation variable form and an initial steady-state output of 0, we know that the constrained
process cannot achieve a steady-state output change of greater than 0.5. Here we tune the PI
controller for a closed-loop response time of 0.25 minutes. The controller gain and integral time
parameter values are kc = 8, tI = 1 min.

Figure 11-2 compares the input and output responses for a setpoint change of 0.4 in both the
constrained and unconstrained control systems. The process output for the constrained system
exhibits overshoot, while the unconstrained system has the expected first-order closed-loop
response. The overshoot is due to reset windup. Notice the constrained manipulated input stays
saturated for 2 minutes, although the output is above the setpoint after 1.5 minutes (after the
setpoint change).

Figure 11-2. Comparison of unconstrained and constrained responses

to a step setpoint change.

A SIMULINK block diagram for this problem is shown in Section 11.7. For small setpoint changes
(say, ±0.1) you should find that reset windup is not a problem. You should show that as the
setpoint change approaches 0.5 (remember, this is the maximum steady-state change that can
be handled by the process, owing to the input constraints), the reset windup effect becomes
worse. See Exercise 2.
Antireset Windup (ARW) Techniques
There are a number of ways that have been developed to minimize problems with reset windup.
Two ways are compared with the standard PI algorithm in Figure 11-3. Figure 11-3a shows the
standard PI algorithm in block diagram form, with a saturation element to represent the
manipulated input constraints. It is clear that there is no feedback to the algorithm to prevent the
unconstrained input v from changing due to integration of the error signal. The "back-calculating"
ARW strategy is shown in Figure 11-3b. Here, the difference between the ideal (unconstrained)
input v and the actual (constrained) input u is fed-back to the integrator. The resetting time, tr, is
an adjustable parameter that is normally set equal to the integral time, tI. The first-order
feedback method of implementing integral action is shown in Figure 11-3c. For a resetting time
equal to the integral time, the two antiwindup strategies (Figures b and c) yield the same closed-
loop performance.

Figure 11-3. Various forms of PI control. (a) Standard PI control

without windup compensation. (b) Reset back-calculating. (c) First-
order feedback. (b, c) Antiwindup strategies.
Example 11.2: Illustration of Reset Windup Compensation (ARW)
A comparison of the antiwindup approach with the saturating PI strategy is shown in Figure 11-4,
for the system given in Example 11.1. Note that the controller with antiwindup remains saturated
for a much shorter time. Clearly, the antiwindup strategy is preferred.

Figure 11-4. Comparison of antiwindup (ARW) and uncompensated

responses for system in Example 11.1. The overshoot in the
uncompensated system is due to reset (integral) windup.

Reset windup can be particularly problematic in cascade-control systems, particularly when batch
processes are involved. This is illustrated by the cascade-control block diagram shown in Figure
11-5. The outer loop (primary control loop) of a cascade strategy is normally tuned for much
slower dynamics than is the inner loop (secondary control loop). If the inner loop performs more
slowly than expected, due to constraints on the manipulated input for example, the performance
of the outer loop controller may suffer. In this case external windup protection can be provided,
as shown in the following example.

Figure 11-5. Series representation of a cascade-control system with

saturation limits on the manipulated input.
Example 11.3: Cascade Control of a CSTR
A physical example (temperature control of a jacketed CSTR) is shown in Figure 11-6. The output
of the primary controller (reactor temperature) is the setpoint to the inner-loop controller. The
manipulated variable for the secondary loop is the jacket flow rate. If the flow rate becomes
constrained, or if the jacket temperature controller is not tuned tightly, the jacket temperature
will not closely match its setpoint, causing problems with the integral action in the primary loop.

Figure 11-6. Cascade-control applied to a CSTR.

For simplicity, let the primary and secondary processes be represented by the following first-order
transfer functions,

and assume IMC-based PI tuning, where the secondary process dynamics were neglected in the
design of the primary controller. Closed-loop time constants of 10 and 2 minutes were used for
the primary and secondary loops, respectively. The manipulated input is assumed to saturate at
±3. The primary and secondary tuning parameters are then
We first perform a simulation with antiwindup on the secondary control loop, since the jacket flow
rate is constrained. The response to a unit setpoint change is shown in Figure 11-7. Notice that
there is still some windup involved, causing overshoot of the primary process output. This is
because the output of the primary controller is a setpoint to the secondary controller. The
secondary controller does not track its setpoint very well because of the saturation of the
manipulated input. This is shown clearly in the middle plot of Figure 11-7.

Figure 11-7. Performance of antiwindup secondary controller. The

primary controller output (jacket temperature setpoint) differs
substantially from the secondary measurement (jacket temperature).

Improved performance can be obtained using an external input to the primary controller. The
basic idea is illustrated in Figure 11-8 for the reactor temperature (master) controller, where the
controller output is the jacket temperature setpoint. Here the jacket temperature (actual
manipulated input) is compared with the jacket temperature setpoint (ideal manipulated input). If
the jacket temperature is not equal to its setpoint, some of the integral action is "removed" by
corrective feedback action. The improved cascade control, with antiwindup (Figure 11-3c) on the
secondary loop and external resetting antiwindup (Figure 11-8) on the primary loop, has excellent
performance, as shown in Figure 11-9. There is now no overshoot in the primary process output
(reactor temperature). The external reset windup protection is shown as a signal between the
measured jacket temperature and the reactor temperature controller, on the control
instrumentation diagram in Figure 11-10.
Figure 11-8. Reactor temperature controller with external variable
reset.

Figure 11-9. Performance of CSTR cascade control with antiwindup

protection on both controllers. The external ARW control strategy of
Figure 11-8 is implemented on the reactor temperature controller.

Figure 11-10. Cascade-control applied to a CSTR. The reactor

temperature controller has external antireset windup (ARW)
protection based on a measurement of the jacket temperature.
It should be noted that antiwindup protection can also be provided by using the "velocity form" in
a digital control system. This is discussed in Module 16.

[ Team LiB ]
[ Team LiB ]

11.3 Autotuning Techniques

Many process control systems have an automatic tuning (autotune) feature. The operator can
simply push the autotune button and have the controller tune itself, that is, determine the values
of the tuning parameters. This is done not continuously (an approach known as adaptive control)
but only when the operator feels that the current set of tuning parameters is not performing well.

The most common method of automatic tuning uses a relay switch to create what is essentially an
on-off controller. The resulting oscillatory behavior is analyzed to determine the proper controller
settings.

Control Block Diagram

The basic control loop for a controller with an autotune relay switch is shown in Figure 11-11. For
normal operation, the switch is connected to the PID controller, as shown in Figure 11-11a. When
the autotune function is operating, the switch is set to the output of the relay block, as shown in
Figure 11-11b.

Figure 11-11. Control block diagram for an autotune relay switch. (a)
Switch set to PID Control. (b) Switch set to relay function for
automatic tuning.
Notice that the relay block represents a nonlinear function. As drawn, it represents the following:

Equation 11.2

Assuming that the input is symmetrical with a steady-state value of 0 (in deviation variable form),
the magnitude of the input change allowed is h, so u max = h, and umin = –h. Notice that this
assumes that the process gain is positive. If the process gain is negative, then switch the
inequalities shown in Equation 11.2.

In autotune mode, the closed-loop system oscillates and the manipulated variable action is on-off
(or bang-bang), as shown in Figure 11-12. In this example, h = 0.05.

Figure 11-12. Oscillations in autotune mode.

There are two parameters that result from this autotune test. One is the period (time between
successive peaks), P, and the other is the amplitude of the process output, a. The period has
units of time; the ultimate frequency can be found from

Equation 11.3

And the ultimate gain can be found from the amplitude

Equation 11.4
In the results shown in Figure 11-12, the period is 16.4 min and the process output amplitude is
0.019. The manipulated input magnitude is 0.05, yielding the following values for kcu and wu:

The behavior is very similar to that obtained from the Ziegler-Nichols closed-loop cycling method.
The Ziegler-Nichols results for the same process are shown in Figure 11-13.

Figure 11-13. Ziegler-Nichols closed-loop test. The controller

proportional gain is kc = 3.83.

The Ziegler-Nichols test yields an ultimate gain of 3.83 and an ultimate period of 17.13 min;
contrast these with the relay values of 3.35 and 16.4 min.

Relay Deadband (Hysteresis)

If the ideal relay is implemented, there can be problems if there is process or measurement noise.
To handle this, a deadband (or ehysteresis) with a magnitude of is added to the relay switch, as
shown below (again, a positive process gain is assumed). The magnitude of the deadband is
selected to be at least two to three times the standard deviation of the measurement noise:
Equation 11.5

Also, if the error is in between the minimum and maximum limits, then the control action stays at
the current value.

Controller Tuning
Once the process parameters (ultimate gain and frequency) are identified from the relay test, a
controller tuning procedure must be used. An obvious method is to use the Ziegler-Nichols values,
since the autotune test is similar to the Ziegler-Nichols test. These controller tuning parameters
are not robust, however, and result in closed-loop systems that may go unstable if the process
conditions change.

An alternative is to use a model-based procedure, such as IMC-based PID. A problem is that the
ultimate gain and frequency must be converted to model parameters. Since only two parameters,
ultimate gain and frequency, are obtained from the relay test, a two-parameter process model
must be developed. The classical first-order + time-delay model cannot be used because it has
three parameters. Two options are integrator + time delay (11.6) and gain + time delay (11.7):

Equation 11.6

Equation 11.7

The integrator + time-delay model is generally a very good approximation to a first-order + time-
delay process if the time constant is much larger than the time delay. The gain + time-delay
model is generally a good approximation if the time delay is much larger than the time constant
(such as for liquid flows with significant pipe volume delays).

Integrator + Time Delay

For the integrator + time-delay model, the dead time and process gain can be found from the
relay parameters in the following fashion:

Equation 11.8
The IMC-based PI tuning parameters can be found from (see Chapter 9)

Equation 11.9

Gain + Time Delay

For the gain + time-delay model, the dead time and process gain can be found from the relay
parameters in the following fashion

Equation 11.10

The IMC-based PI tuning parameters can be found from

Equation 11.11

In both cases, the IMC filter-factor (l) is adjusted to provide a tradeoff between nominal
performance and robustness to model uncertainty.

Comparison of Model-Based PI Controllers

In the example shown in Figure 11-12, the period and output amplitudes were 16.4 min and
0.019 for a control magnitude of 0.05. The IMC-based PI controllers obtained under the two
different model-based assumptions are compared in Table 11-1.

The responses of the two model-based PI controllers are shown in Figure 11-14. The l value was
adjusted to achieve 25% overshoot in both cases. For more robust performance, the l values
must be increased.

Figure 11-14. Comparison of closed-loop responses for the example

process. IMC-based PI tuning parameters based on two different
models (integrator + time-delay and gain + time-delay) identified
from a relay test.
[ Team LiB ]
[ Team LiB ]

11.4 Nonlinear PID Control

The fixed-parameter PID controller is ubiquitous in the process industries. The ideal PID controller
has the form (see Chapter 5)

Equation 11.12

where it is understood that any bias associated with steady-state operation is absorbed into the
integral term. Other forms for PID control are presented in Chapter 5.

Often it is useful to have a nonlinear controller even if the process is linear. A common example is
level control of a surge drum where the primary objective is to maintain a reasonably consistent
outlet flow under disturbances in the inlet flow rate, subject to high and low drum level
constraints. In this situation, some small nominal controller gain can be used when the level is
close to the setpoint, and as the error increases, the gain can increase. A common choice is to
make the proportional gain a function of the absolute value of the error,

Table 11-1. Example Parameters for Model-Based Controllers

Model Integrator + time delay Gain + time delay

k 4(0.019)/(16.4)0.05 = 0.093 0.019/0.05 = 0.38

q 16.4/4 = 4.1 min 16.4/2 = 8.2 min

kc (2l+4.1)/[0.093(l +4.1) 2] = 1.831 for l = (8.2/2)/(0.38l) = 10.79/l = 1.269 for l =

5 min 8.5 min
tI 2l +4.1 = 14.1 for l = 5 min 8.2 min

Equation 11.13

This means that the controller output is effectively proportional to the square of the error, as
shown below for a P-only nonlinear controller.

Equation 11.14
An alternative to making the controller gain a continuous function of the error is to form a
piecewise linear controller (sometimes called a gap controller), where (often kc,high = kc,low)

Equation 11.15

An example of nonlinear PID control is given in Module 10 (Surge Vessel Level Control).

[ Team LiB ]
[ Team LiB ]

11.5 Controller Parameter (Gain) Scheduling

Perhaps the most widely mentioned technique for controlling nonlinear processes is gain
scheduling. Let a represent the scheduling variable, which will often be the measured output, y. A
gain-scheduled PID controller is then represented by

Equation 11.16

Notice that the primary difference between this algorithm and the nonlinear PID controllers of the
previous section is that a is a general scheduling variable, whereas the controllers in the previous
section were based on error as the scheduling variable. Although all of the controller parameters
can be scheduled, controller gain is commonly scheduled because processes are often
characterized as a changing gain with relatively constant dynamics. Typically, the controller gain,
kc, is varied to keep kckp constant, which then keeps the stability margin constant. If the process
gain can be characterized as a function of the scheduling parameter, kp(a), then the controller
gain can be scheduled as

Equation 11.17

where the subscript 0 indicates a nominal operating point.

An important step in developing a gain-scheduled controller is determining the proper scheduling

variable. Rules of thumb such as scheduling on a "slow variable" or using a variable that "captures
the nonlinearities" are often used. It is most common, then, to schedule based on either the
setpoint or measured output, since the process input (manipulated variable) will vary more
rapidly than the output. When there are additional measurements, then a measured or inferred
"auxiliary" variable can be used for scheduling. The four basic steps in developing a parameter-
scheduled controller are as follows:

1. Develop a linear process model for a set of operating conditions

2. Design linear controllers for each operating condition (model)

3. Develop a schedule for the controller parameters

4. Implement the parameter-scheduled controller on the nonlinear plant

Note that steps 1 and 2 will often be combined by using some form of closed-loop tuning at
discrete operating points. This is particularly true if the autotune feature of commercial controllers
 4.

is used; the use of three operating points appears to be common.

There are several different options for the scheduling of the controller parameters.

Switch parameters at discrete values of the scheduling variable

Interpolate parameters as a function of the scheduling variable

Vary parameters continuously with the scheduling variable

A major advantage of a parameter-scheduled controller is that linear control system design

procedures can be used, and at least the local control system behavior can be understood. A
problem can occur in steps three and four, where parameter scheduling is developed and
implemented. If the dynamic effect of the scheduling variable is not included in the design
process, then even the local linearized closed-loop behavior may not be predicted correctly.

A general representation of a gain-scheduled controller is shown in Figure 11-15, where a

represents the scheduling variable. The double lines on the process and controller blocks indicate
that these are nonlinear.

Figure 11-15. Parameter-scheduled control strategy.

An example application of gain scheduling is distillation. Since the process gain is a function of the
product composition, it makes sense to schedule the controller gain as a function of composition.
This approach was taken by Tsogas and McAvoy (1985) in the control of distillation composition,
with distillate flow rate as the manipulated variable. The scheduling algorithm used was

Equation 11.18

Notice that this is a "one-way" algorithm. Since the process gain increases with lower purity,
maintaining a constant controller gain speeds up the response when the distillate is less pure.

[ Team LiB ]
[ Team LiB ]

11.6 Measurement/Actuator Selection

Perhaps the best way to assure that a nonlinear control system is well maintained is through
design of the physical control devices (actuators and measurements). For example, in Module 15
(flow control), we show that the installed characteristic of an equal-percentage valve makes the
valve position/flow rate relationship linear over a wider range of operating conditions than other
valve types.

In distillation control, defining a process output as the logarithm of the product composition can
be used to linearize the input-output relationship. For example,

Equation 11.19

where xD is the distillate composition. The new setpoint variable, r, can be found as the logarithm
of the desired distillate composition, xDsp, by

Equation 11.20

and the controller output (reflux flow rate, for example) would be a function of r - y.

[ Team LiB ]
[ Team LiB ]

11.7 Implementing PID Enhancements in Simulink

SIMULINK has a number of standard features that make it easy to perform simulations for the
control strategies developed in this chapter. Saturation and relay blocks, for example, can be
found in the Discontinuities SIMULINK library in MATLAB version 6.5 (Release 13).

Constrained vs. Unconstrained

The simulation results in Figure 11-2 were generated using the S IMULINK .mdl diagram shown in
Figure 11-16.

Figure 11-16. Comparison of constrained and unconstrained processes

to a step setpoint change.

ARW
The antiwindup results presented in Figure 11-4 were generated using the S IMULINK .mdl diagram
shown in Figure 11-17. The cascade results of Figures 11-7, and 11-9 were generated using the
SIMULINK .mdl diagrams shown in Figures 11-18 and 11-19.

Figure 11-17. Comparison of ARW strategies with "classical" PID.

Figure 11-18. SIMULINK cascade-control simulation with antiwindup on
both controllers.

Figure 11-19. Unmasked view of anti-windup strategies used for the

cascade-control simulations.
Autotuning
The autotuning simulation shown in Figure 11-12 was generated using Figure 11-20.

Figure 11-20. SIMULINK .mdl diagram for autotuning simulation.

[ Team LiB ]
[ Team LiB ]

11.8 Summary
A number of enhancements to standard PID controllers have been presented in this chapter. All
commercial controllers implement ARW methods of one form or another. It is particularly
important for cascade controllers to use external reset windup protection.

A number of controllers also have autotune features, but some of these are proprietary. We have
presented the approach that is most widely presented in the literature. Using a relay function in
the closed loop provides a closed-loop oscillation similar to the Ziegler-Nichols closed-loop
oscillation method. There are several of advantages to the relay-based approach: the magnitude
of the oscillations is bounded by proper selection of the manipulated input constraints, and the
time required to develop several periods of the oscillation is much shorter than the time it takes
to laboriously change a proportional gain to create a closed-loop oscillation.

One of the primary ways of creating a nonlinear PID controller is to make the controller gain a
function of the absolute value of the error. This approach works well for processes where small
deviations from setpoint are acceptable, such as surge vessel level control. Gain scheduling is a
approach that can be used when the model and controller gain can be characterized as a function
of the operating condition.

The proper selection of measurements and actuators can help linearize the input-output
relationship. Equal percentage valves often make flow systems appear linear over a wider range.
Similarly, using the logarithm of distillation product compositions can make the input-output
behavior more linear.

[ Team LiB ]
[ Team LiB ]

References

A nice presentation of antireset windup is presented in: Astrom, K. J., and L. Rundqwist,
"Integrator Windup and How to Avoid It," in Proceedings of the 1989 American Control
Conference, Pittsburgh, pp. 1693–1698 (1989).

The following paper presents model-based (integrator + time delay and gain + time delay) PI
approaches for tuning SISO and two-input, two-output processes: Friman, M., and K. V. Waller,
"Autotuning of Multiloop Control Systems," Ind. Eng. Chem. Res., 33(7), 1708–1717 (1994).

A nice summary of PID design and implementation is presented in the following book: Kiong, T.
K., W. Quing-Guo, and H. C. Chieh with T. J. Hagglund, Advances in PID Control, Springer Verlag,
London (1999).

The following paper reviews gain scheduling approaches to process control: Bequette, B. W.,
"Practical Approaches to Nonlinear Control: A Review of Process Applications," in Nonlinear Model-
based Process Control, NATO ASI Series, Ser. E, Vol. 353, pp. 3–32 (R. Berber and C. Kravaris,
Eds.), Kluwer, Dordrecht (1998).

Distillation gain-scheduling control has been studied in the following: Tsogas, A., and T. J.
McAvoy, "Gain Scheduling for Composition Control of Distillation Columns," Chem. Eng. Commun.,
37, 275–291 (1985).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Verify that the two block diagrams below result in the same PI control algorithm
transfer function.

2: Use the SIMULINK diagram presented in Figure 11-16 to verify the results presented in
Example 11.1. Show that, as the setpoint change approaches 0.5 the integral windup
effect becomes worse.

3: Verify the results of Example 11.2 using a SIMULINK simulation. Also, discuss the effect
of the resetting time, tr, on the ARW results.

4: Consider the process furnace temperature cascade control system shown in Figure 10-
2, where the output of the temperature control loop is a setpoint for the fuel gas flow
controller. Draw a signal, similar to Figure 11-10, that shows that an external antireset
strategy is used for the temperature controller.

5: Consider Example 11.3. Use S IMULINK to show that cascade control without antireset
windup protection yields significant overshoot for a setpoint change.

6: Show that the relay procedure for an integrator + time-delay process will exhibit the
following period and output amplitude:

7: Show that the relay procedure for a gain + time-delay process will exhibit the following
period and output amplitude:
 8: The example process used in this module is first order + time delay, with a gain of 1, a
time constant of 10, and a time delay of 5. Compare the Ziegler-Nichols PI control
performance when the critical gain and period are obtained using two different
methods: relay-based cycling, and Ziegler-Nichols closed-loop cycling. Also, compare
the Tyreus-Luyben tuning parameters obtained using the two methods. Does the
relay-based method generally yield more conservative results?

9: For an integrator + time-delay model, find the IMC-based PI parameters, assuming an

IMC filter with the form

where g is selected to yield a PI controller.

10: For a gain + time-delay model, find the IMC-based PI parameters when the all-pass
method is not used.

11: For a first-order + time-delay process, with a time constant of 10 minutes and a time
delay of 0.5 minutes, use the autotune procedure. How close are the critical gain and
ultimate period to the values obtained by the Ziegler-Nichols closed-loop oscillation
method? What are the process parameters for an integrator + time-delay model?
Based on these parameters, simulate the closed-loop behavior of a PI controller, with
various values of l. What l values do you recommend? Why?

12: For a first-order + time-delay process, with a time constant of 0.5 minutes and a time
delay of 10 minutes, use the autotune procedure. How close are the critical gain and
ultimate period to the values obtained by the Ziegler-Nichols closed-loop oscillation
method? What are the process parameters for a gain + time-delay model? Based on
these parameters, simulate the closed-loop behavior of a PI controller, with various
values of l. What l values do you recommend? Why?

[ Team LiB ]
[ Team LiB ]

Chapter 12. Ratio, Selective, and Split-

Range Control
After reading this chapter the reader should be able to understand and develop ratio, selective
and override control, and split-range control strategies. Ratio control is similar to feed-forward
control, while selective and override control choose between alternative outputs or inputs. Split-
range control chooses between two or more control valves to implement a manipulated input.

The major sections of this chapter are as follows:

12.1 Motivation

12.2 Ratio Control

12.3 Selective and Override Control

12.4 Split-Range Control

12.5 SIMULINK Functions

12.6 Summary

[ Team LiB ]
[ Team LiB ]

12.1 Motivation
In Chapter 10 we discussed feed-forward and cascade-control strategies. The purpose of these
control strategies was to take advantage of additional measurements to improve the disturbance
rejection capability. In this chapter, we also develop strategies that use more than one
measurement and, in addition, manipulate more than one input. Ratio control, presented in
Section 12.2, is very similar to feed-forward control. In Section 12.3, we discuss selective and
override control, which involves the use of "selectors" to decide the proper control action to take.
In Section 12.4, split-range control is presented as a way to use more than one valve to regulate
a process.

[ Team LiB ]
[ Team LiB ]

12.2 Ratio Control

Ratio control is similar to feed-forward control, since both typically involve the measurement of a
stream flow rate. Ratio control is often used in component blending problems. Consider the
control strategies shown in Figure 12-1, where stream A is a "wild stream" (disturbance) and it is
desirable to maintain the flow rate of stream B at a constant ratio to stream A. In Figure 12-1a
the F B/FA ratio is calculated and sent to a ratio controller. In Figure 12-1b the flow rate of stream
A is measured and multiplied by desired ratio to determine the setpoint for the stream B flow
rate.

Figure 12-1. Ratio control. Comparison of two alternatives.

Alternative 12–1b is favored over 12–1a since it results in a linear input-output relationship. In
Figure 12-1a the steady-state relationship between the manipulated input (valve position, related
to F B) and the measured output (the ratio, F B/FA) is

Equation 12.1

and the steady-state process gain is

Equation 12.2

Clearly, the process gain varies as a function of the flow rate of stream A. This means the ratio
controller would need to be tuned differently depending the stream A flow rate. In contrast, the
relationship between the manipulated input (valve position, related to FB) and the measured
output (FB) for the flow controller in Figure 12-1b is

Equation 12.3

so the gain is

Equation 12.4

which is constant.

The ratio could be specified to maintain an excess of component B in a reaction, for example. A
common use of this would be the supply of combustion air to a furnace. Stream A could represent
the fuel gas, for example, and stream B would represent the combustion air.

The blend stream could also have a composition or product property measurement and controller;
the output of this controller would be cascaded to the desired ratio of stream B flow to stream A
flow. In Exercise 1 you have the opportunity to sketch this control strategy.

[ Team LiB ]
[ Team LiB ]

12.3 Selective and Override Control

Often a controller needs to select from a set of possible measurements for control. For example,
consider the packed-bed reactor shown in Figure 12-2. It is important to maintain the "hot spot"
temperature at a certain value to minimize catalyst degradation. In this strategy, the "high
selector" (represented as the block with >) chooses the highest of three bed temperatures to
send to the temperature controller. In algorithmic form, let T1, T2, and T3 represent the three
temperature measurements, and T represent the temperature measurement used by the
controller, then

Figure 12-2. High-temperature selector for a packed-bed reactor.

Equation 12.5

When manipulating the fuel gas flow to a fired heater (furnace), it is important to make certain
that a minimum air/fuel ratio is maintained. High/low selectors can be used to assure this, as
shown in Figure 12-3; here the A/F boxes indicate the air/fuel ratio. To understand how this
strategy works, first consider a positive change in the fuel flow setpoint. Notice that the
combination of high and low selectors will force the combustion air flow rate to increase before
the fuel flow increases. Also, if a decrease in fuel flow setpoint is made, the fuel flow must
decrease before the combustion air flow. This strategy makes certain that excess combustion air
is always available.

Figure 12-3. Use of high/low selectors for fuel gas and air flow control
of a fired heater. Notice the simplified representation for the air flow;
ordinarily a damper inside an air duct would be adjusted to vary the
air flow.
In practice this strategy is just one part of the overall furnace control strategy. Usually, the fuel
flow setpoint will be the output of another controller (e.g., process fluid temperature or steam
system pressure), and the air/fuel ratio might be the output of a furnace stack gas O2 or CO
measurement and control strategy.

[ Team LiB ]
[ Team LiB ]

12.4 Split-Range Control

A split-range strategy is often used in situations where one or more valves may be used,
depending on the operating scenario. An example application is a batch reactor. Batch reactors
must often be heated from ambient temperature to a desired operating temperature. Particularly
if the reaction is exothermic, then cooling will be needed to maintain the desired reactor
temperature. In the split-range control strategy shown in Figure 12-4, if the jacket temperature
controller output is between 0 and 50%, the cold glycol valve is open; if the jacket temperature
controller output is between 50 and 100%, the hot glycol valve is open. For safety reasons we
wish the cold glycol valve to fail-open and the hot glycol valve to fail-closed. The diagram in
Figure 12-5 more clearly shows the relationship between valve positions and the jacket
temperature controller output.

Figure 12-4. Batch reactor temperature control. The jacket

temperature controller has a split-range output, where the cold glycol
valve is open during "cooling mode" and the hot glycol valve is open
during "heating mode."

Figure 12-5. Depiction of the split-range controller action.

The implementation is as follows. Let u represent the output of the jacket temperature controller
(0–100%). Also, let vc and vh represent the cold and hot glycol valve positions, respectively. The
split-range algorithm can be stated as

Equation 12.6

Sometimes there is a small deadband, where the cold glycol is open between 0 and 45%, while
the hot glycol is open between 55 and 100%. This adds the equivalent of a small time delay to the
control strategy but avoids situations where both cold and hot valves are open owing to small
miscalibrations.

Example 12.1: 1000 Liter Stirred-Tank Heater

Here we show results for a 1000 liter jacketed, continuous stirred-tank heater, operated at 85%
of capacity. For clarity, we consider only the jacket temperature controller, which is split-ranged
as shown in Figure 12-5. Ordinarily this jacket temperature controller would be the secondary
(inner-loop or slave controller) where the output of a vessel temperature controller (outer-loop or
primary controller) is the setpoint to the jacket temperature controller.

The modeling equations are

where the jacket make-up flow rate (F j ; of either cold or hot glycol) is the process input.
Numerical values are shown in Exercise 5.

The responses for setpoint changes in jacket temperature are shown in Figure 12-6. A positive
setpoint change is made at t = 0 minutes, followed by a negative change at t = 10 minutes. The
middle plot shows the jacket temperature controller output and the bottom plot shows the
individual valve positions. All variables are in deviation form. When the controller output is
positive, the hot valve is activated; when the controller output is negative, the cold valve is
activated.

Figure 12-6. Response to step setpoint changes in jacket temperature.

All variables are in deviation form.

[ Team LiB ]
[ Team LiB ]

12.5 Simulink Functions

SIMULINK has a number of built-in functions to perform ratio, selective, and split-range control.
Examples are shown here.

Product Can be used for ratio control

MinMax Can be used for high or low selection
Switch Can be used for switching between alternatives
Saturation Can be used for split-range control, and controller saturation

An example SIMULINK block diagram to implement a split-range jacket temperature controller from
Example 12.1 is shown in Figure 12-7. Since this has been formulated in deviation variable form,
the output of the jacket temperature controller ranges from –50 to 50%. Saturation elements are
used for the cold and hot valve positions.

Figure 12-7. SIMULINK implementation of a split-range strategy for a

jacket temperature controller.

[ Team LiB ]
[ Team LiB ]

12.6 Summary
Ratio control has many similarities to feed-forward control. Usually a "wild stream" flow rate is
measured and another stream is manipulated to maintain a constant ratio to the wild stream.
Selective and override strategies are often used to improve the safety of a process system. A
prime example is a fired heater where it is important for combustion air to "lead" the fuel gas flow
on increasing loads, but to "lag" fuel gas flow on decreasing loads. Split-range control allows the
use of two or more valves as manipulated inputs, depending on the operating condition.

[ Team LiB ]
[ Team LiB ]

References

Ratio control is discussed in: Myron Jr., T. J., "Feedback Methods for Process Control Systems,"
Chemical Engineering, 233–238 (Nov. 14, 1983).

A number of selective control examples are shown in: Kurth, T. G., "How to Use Feedback Loops
to Meet Process Conditions," Chemical Engineering, 77–83 (April 30, 1984).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider the ratio control strategy shown in Figure 12-1b. If the blend stream has a
concentration measurement and controller, construct the control instrumentation
diagram where the output of the concentration controller is the desired ratio of stream B
to stream A flows.

2: Consider the high-temperature select control problem shown in Figure 12-2. Modify the
control instrumentation diagram to include a cascade-control strategy, where the output
of the reactor temperature controller is the setpoint to a reactor inlet temperature
controller. The reactor inlet temperature controller then manipulates the flow to the
preheat exchanger.

3: Consider a gas-phase reactor where a valve on the output stream is normally regulated
to maintain a desired outlet composition. Design a selector-based control strategy that
switches to pressure control if the pressure is too high.

4: Consider a waste treatment system. A pH controller may need to adjust the pH by

manipulating either basic or acidic stream flow rates. Design a split-range control
strategy to do this.

5: Consider Example 12.1, where input 1 is the cold glycol make-up flow rate, and input 2
is the hot glycol make-up flow rate. The steady-state temperature is 20°C, which is also
the vessel feed temperature. At steady-state the cold and hot glycol make-up flow rates
are 0. The parameter values are

fov = 1/30; % minutes^-1

ua = 24 ; % kcal/min deg C
vrhocp = 850 ; % kcal/deg C
vrhocpj = 250 ; % kcal/deg C
Tjfcold = -20 ; % deg C
Tjfhot = 100 ; % deg C
Tjs = 20 ; % deg C
vj = 250 ; % liters

and the state space model matrices are

 Solve for the numerical values of all matrices. Design a jacket temperature controller
and implement the SIMULINK diagram shown in Figure 12-7. Discuss your jacket
temperature controller design and tuning procedure. Notice that the jacket temperature
output has a range of 100%, constrained between –50% and +50%. The cold valve is
fully open at –50% and fully closed at 0%. The hot valve is fully closed at 0% and fully
open at +50%,

6: Develop a SIMULINK diagram of a cascade control strategy, where Figure 12-7 represents
the inner-loop (jacket temperature) control system. The output of the primary (outer-
loop) controller that you develop should be the setpoint for the jacket temperature
controller. Discuss the design and implementation (tuning) of both controllers. Use the
parameter values from Exercise 5.

[ Team LiB ]
[ Team LiB ]

Chapter 13. Control-Loop Interaction

Most process unit operations have a number of manipulated inputs, each of which affects some
measured output, and it is not always clear which input should be "paired" with which output for
control-system design. This general pairing problem is known as control structure selection and is
often more important than the actual controller design and tuning method used. After studying
this chapter, the reader should be able to do the following:

Use the relative gain array (RGA) to determine the proper input-output pairings for multiple
single input–single output controllers

Understand that the RGA gives insight about process sensitivity

Discuss the implications of the RGA for failure sensitivity

The major sections of this chapter are as follows:

13.1 Introduction

13.2 Motivation

13.3 The General Pairing Problem

13.4 The Relative Gain Array

13.5 Properties and Application of the RGA

13.6 Return to the Motivating Example

13.7 RGA and Sensitivity

13.8 Using the RGA to Determine Variable Pairings

13.9 MATLAB RGA Function File

13.10 Summary

[ Team LiB ]
[ Team LiB ]

13.1 Introduction
Thus far in this textbook we have focused primarily on SISO control systems. Most chemical
processes, however, have many manipulated variables affecting many output variables, as shown
in Figure 13-1. We generally assume that a desired output variable is also measured. In some
cases (composition for example), an output variable cannot be measured and must be "inferred"
from other measurements (such as temperatures and flow rates).

Figure 13-1. General multivariable process.

Although most processes are multivariable, SISO control loops are formed by selecting a
measured output that is most strongly affected by a particular manipulated input. This is done for
two reasons: SISO design techniques are easy to understand, and hardware and software are
readily available for SISO controllers. We refer to control systems that use several SISO
controllers on a multivariable process as MVSISO (multivariable single input–single output)
controllers. The selection of which measured output to pair with which manipulated input is known
as variable "pairing" and is the focus of much of this chapter.

[ Team LiB ]
[ Team LiB ]

13.2 Motivation
A specific example of a multivariable control problem is shown in Figure 13-2. Here, two different
streams are blended together. The objective is to manipulate the individual stream flow rates to
meet specifications on the output flow rate and the output composition. Let the outputs be
represented by y2 (total flow) and y1 (composition), and the inputs by u1 (flow rate of stream 1)
and u 2 (flow rate of stream 2). How should our outputs and inputs be paired? If we assume that
the output-input pairings are y1-u1 and y2-u2, the instrumentation diagram is shown in Figure 13-
3a. If we assume that the output-input pairings are y1-u2 and y2-u1, the instrumentation diagram
is shown in Figure 13-3b.

Figure 13-2. Blending system.

Figure 13-3. Blending system alternative closed-loop pairings. (a) u1-

y1, u 2-y2 pairing. (b) u1-y2, u 2-y1 pairing.
We can see potential problems in the operation of the control strategy in either Figure 13-3a or b.
For discussion purposes, consider Figure 13-3a. A setpoint change in output 1 (y1, concentration)
will cause the flow controller to change input 1 (u1, stream 1 flow). The change in input 1 will also
affect output 2 (y2, total flow). Once output 2 is disturbed, the flow controller will cause a change
in input 2. Input 2 will disturb the concentration, causing the concentration controller to change
input 1, and we are back to where we started. The effect of one control loop on the other is
referred to as control loop interaction. Although it is natural to think of these interactions as
occurring in a sequential fashion, they actually happen simultaneously.

Example 13.1: Whiskey Blending

Consider now a very specific problem to see whether physical intuition can be used to determine
the proper pairing. Assume that an enterprising control student has decided that there is a profit
to be made by blending a stream of 80 proof (40% ethanol) whiskey with pure water to produce a
product stream that is 60 proof (30% ethanol). Assume that it is important to meet both quality
and production-rate requirements, and that you can measure the alcohol content and total flow
rate accurately. Assume that the 80 proof stream flow rate is the first manipulated input (u1) and
that the dilution water flow rate is the second manipulated input (u2). The blended product flow
rate is output 2 and the blend composition is output 1.

Your intuition is probably to pair output 1 (alcohol measurement) with input 1 (80 proof whiskey
stream), since there is no alcohol in the water stream; this is the strategy shown in Figure 13-3a.
We find later that this is not the best pairing, indicating that the best control structure is often
counterintuitive. We first see the problem by observing the behavior of control loop 1 (y1-u1
pairing) when control loop 2 (y2-u2 pairing) is open (or under "manual" control). The response to
an ethanol concentration setpoint change is shown in Figure 13-4. Observe that the closed-loop
behavior of ethanol concentration is quite good, but that the total flow rate deviates, since the
flow control loop (control loop 2) is not closed.

Figure 13-4. Response to an ethanol setpoint change; loop 1 has

ethanol concentration as the measured output (y1) and whiskey flow
rate as the manipulated input (u1). Loop 2 is open.

Assume that it is desirable to maintain the total blend rate at a constant value. This requires that
loop 2 be closed. Loop 2 has blend flow as the measured process output (y2) and the dilution
water flow rate as the manipulated input (u2). Assume that it has been tuned independently from
loop 1, that is, loop 1 is open. The response to a setpoint change in flow rate is shown in Figure
13-5. Notice that the flow-rate response is quite good but that the ethanol concentration deviates
from the desired value, because loop 1 is open.

Figure 13-5. Response to total blend flow rate setpoint change; loop 2
has blend flow rate as the measured output (y2) and dilution water
flow rate as the manipulated input (u 2). Loop 1 is open.
We have seen that the individual control loops have quite good performance. Since we desire to
control both the concentration and the flow rate of the blend stream, we need to have both loops
closed simultaneously. The response to a setpoint change in the total blend flow rate with both
loops closed is shown in Figure 13-6. Notice that closed-loop system is now unstable! What
characteristic of this control strategy caused the instability to occur when both loops were closed?
Would this behavior have occurred if a different variable pairing had been used? That is, would
there be problems if loop 1 paired the ethanol measurement (y1) with dilution water flow (u2) and
loop 2 paired the total blend flow rate (y2) with the whiskey flow (u 1)? We find the answer to this
question in the sections that follow.

Figure 13-6. Response to a total blend flow rate setpoint change with
both loops closed. Loop 2 pairs blend flow rate (y2) and dilution water
flow rate (u2), while loop 1 pairs ethanol concentration (y1) with
whiskey flow (u1).
[ Team LiB ]
[ Team LiB ]

13.3 The General Pairing Problem

Given a general multivariable process diagram, how do we select the pairings for a control
structure that consists of multiple SISO loops? In the following sections, we use matrix vector
notation, where y(s) is a vector of n outputs and u(s) is a vector of m inputs,

Equation 13.1

Equation 13.2

where each output can be represented by

Equation 13.3

To form a MVSISO system, we must ask the following:

Which output yi should be paired with which input uj to form an SISO control loop?

Two Input–Two Output Processes

Let us consider the relationship between yi and uj under a number of conditions. For simplicity,
we will deal primarily with 2 x 2 (two input–two output) systems. The open-loop input-output
relationships are

Equation 13.4
which is written in matrix form as

Equation 13.5

and we can think of the input output relationships as shown in Figure 13-7.

Figure 13-7. Input-output block diagram for a two input–two output

process.

The corresponding feedback control system, if the pairings are u1-y1 and u2-y2, is shown in Figure
13-8.

Figure 13-8. Feedback block diagram for a two input–two output

process where output y 1 is paired with input u1.
Input 1–Output 1 Dynamic Behavior
For consistency in notation, we refer to loop 1 as the y1-u1 pairing and loop 2 as the y2-u2 pairing.
We wish to consider how control loop 1 should be designed. We have focused on the use of input-
output models for control system design. In the case of loop 1, we need to know how input 1
affects output 1.

Now, consider how u1 affects y1 if (a) loop 2 is open and if (b) loop 2 is closed.

(a) Loop 2 is open (loop 1 is also open). If the second input (u2) is constant, then the outputs
depend only on the first input (u 1), as shown in Figure 13-9.

Figure 13-9. Block diagram with loop 2 open (loop 1 is open as well).

The input-output relationship is

If we designed a SISO feedback control system for loop 1, we would use g 11(s) as the process
transfer function for control system design.

(b) Loop 2 is closed. If the second loop (y2-u2) is closed, then the outputs depend on input 1 (u1),
as shown in Figure 13-10. Notice that there is an additional effect of input 1 on input 2 owing to
the action of controller 2. Input 1 affects output 2, which then changes input 2 through the
second controller. This change in input 2 then has an additional effect on output 1; this effect can
be positive or negative.

Figure 13-10. Block diagram of input 1–output 1 relationship with loop

2 closed.

Thus, the relationship between u1(s) and y1(s) is no longer just g11(s). Indeed, we can derive the
relationship in the fashion:

Equation 13.6

but (assuming no setpoint change in loop 2, r2 = 0),

Equation 13.7

also

Equation 13.8
Combining Equations (13.7) and (13.8),

Equation 13.9

and solving for y2(s),

Equation 13.10

Substituting this into Equation (13.7),

Equation 13.11

and substituting this result into Equation (13.6), we find

Equation 13.12

That is, y1(s) = g11,eff(s) u1(s), where

Equation 13.13

g11,eff(s) is the effective input-output relationship between u1 and y1, with loop 2 closed. This
means that g11,eff(s) should be used as the process transfer function to design controller 1 if loop
2 is closed. Unless g12(s) or g21(s) = 0, if the second control loop is closed, then the first control
loop must be designed differently than if the second control loop were open. Naturally, the
relationship between u1 and y1 is a function of the controller gc2 when the second loop is closed.

Example 13.1, continued

Consider the expected effect of the second loop in the blending example posed earlier. In Figure
13-11 we compare the response of ethanol concentration due to a change in whiskey flow rate,
under two conditions: with the flow controller (loop 2) open (solid), and with the flow controller
closed (dashed). Notice the tremendous amplification of the effect of the whiskey flow rate. Here,
we rationalize the effect of the flow loop being closed. When the whiskey flow rate is increased, it
causes the total blend flow rate to increase. The total blend flow-rate controller then cuts back on
the dilution water flow rate, which causes the ethanol concentration to increase even more. The
flow rate controller has caused the whiskey stream to have roughly four times the effect on the
ethanol concentration, compared to when the flow control loop is open.

Figure 13-11. Response of ethanol concentration to a step increase in

whiskey flow rate; loop 2 open (solid curve) and loop 2 closed (dashed
curve).

Steady-State Effective Gain

Comment:

What we would really like is a method to determine the relationship between u1 and y1 with loop 2
closed, without knowing gc2 (s).

We can easily determine this relationship for a limiting case, the steady state. For the steady
state, we simply let s 0. Assuming integral action is used in the controller, we find that as s
0, gc2 (s) . Also, from Equation (13.13) and gc2 (0) as s 0, we find the
following relationship (where 0 represents s = 0, the steady state; do not confuse this with t = 0,
the initial condition):

Equation 13.14
This can also be determined from the steady-state relationship

Assuming perfect control of y2, that is, y2 = 0,

Equation 13.15

Equation 13.16

and we can then find the following relationships:

Equation 13.17

We now have a steady-state effective gain relationship between u1 and y1 with the loop between
u2 and y2 closed.

[ Team LiB ]
[ Team LiB ]

13.4 The Relative Gain Array

Ed Bristol (a control engineer with Foxboro) developed a heuristic technique to predict possible
interactive effects between control loops when multiple SISO loops are used. He was motivated to
develop a dimensionless measure of interaction by some boiler control problems he was working
on. His simple measure of interaction is known as the relative gain, or the Bristol relative gain
array (RGA). Since then, many papers have been written to provide a more rigorous theoretical
basis for the technique.

Two-Inputs and Two-Outputs

The relationship between u1 and y1 at steady state (s = 0), with no change in u2, is

Equation 13.18

This is referred to as the gain between u1 and y1 with all other loops open (that is, u2 is constant
for this 2 x 2 example).

The steady-state relationship between u1 and y1 at steady state, with y2 maintained constant at
its setpoint (using a controller with integral action), is

Equation 13.19

Notice that Equation (13.19) is the gain between input 1 and output 1, assuming output 2 is
constant.

Definition of the Relative Gain

The relative gain (lij) between input j and output i is defined in the following fashion:

Equation 13.20
By convention, the symbol used for the relative gain is lij. Please do not confuse it with the l that
was used for the IMC filter or for the l that is commonly used to represent the eigenvalues of a
matrix. The relative gain will normally be shown with two subscripts, distinguishing it from
eigenvalues or the IMC filter factor.

Relative Gain Between Input 1 and Output 1 for a Two Input–Two

Output System
The relative gain between input 1 and output 1 is then [see Equation (13.23)]

We see that the relative gain between u1 and y1 for a 2 x 2 system is

Equation 13.21

Remember that Equation (13.21) is only true for the two input–two output case.

The RGA
The RGA is simply the matrix that contains the individual relative gain as elements, that is L =
{lij} . For a 2 x 2 system, the RGA is

Equation 13.22
Question: What value do we desire for the relative gain?

Think of it this way. If we want to use a u1-y1 pairing for a SISO controller, we do not want it to
matter whether the other loops in the system are closed or not. This tells us that we desire a
relative gain close to 1.0. That is,

Answer: We desire l11 1.0 if we wish to pair output 1 with input 1.

The l11 calculation for a 2 x 2 system yielded . The more general result is (see
Appendix 13.1 for a derivation)

Equation 13.23

where element of G–1(0). Notice that this is not the same as matrix multiplication. Here,
the individual elements of the two matrices are being multiplied together.

We check the general results derived above by applying them to a 2 x 2 (two output, two input)
system,

and recalling our rules for inverting a 2 x 2 matrix,

Multiplying the individual elements of G(0) and (0), as in Equation (13.23), we find

Equation 13.24

[ Team LiB ]
[ Team LiB ]

13.5 Properties and Application of the RGA

Sum of Rows and Columns Property of the RGA

Notice that Equation (13.24) yields the following relationships

This property for 2 x 2 systems is general for n x n systems. That is, each row of the RGA sums to
1.0 and each column of the RGA sums to 1.0.

Then, for a 2 x 2 system, only one relative gain must be calculated to find the entire array,

Equation 13.25

For a 3 x 3 system, only four of the nine elements would need to be calculated. In practice, they
are all calculated simultaneously using (13.23).

Use of RGA to Determine Variable Pairing

It is desirable to pair output i and input j such that l ij is as close to 1 as possible.

Example 13.2: RGA for Variable Pairing

Consider a relative gain array,

We would pair y1 with u1 and y2 with u2 in this case.

However, if

we would pair y1 with u2 and y2 with u1.

Two Input–Two Output Systems

It is easy to show the two possible cases for the relative gain array of a 2 x 2 system. If there are
an odd number of positive elements in G(0), then

That is, for an odd number of positive elements in the process gain matrix, all the relative gains
will be between 0 and 1. If there are an even number of positive elements in G(0), then

This means that for an even number of positive elements in the process gain array, all the relative
gains will be outside the 0–1 range.

Implications for the Sign of a Relative Gain

First of all, we can state that if l11 was negative, we would not want to pair input 1 with output 1.
Let us use the following reasoning. Assume that the open-loop gain between input 1 and output 1
(g11) is positive; this implies that the controller gain is also positive. A negative relative gain tells
us that the gain between input 1 and output 1 with output 2 perfectly controlled (that is, loop 2
closed) is negative. This means that if loop 2 is closed, we need a negative controller gain on loop
1. However, we noted before that if loop 2 is open, we need a positive controller gain on loop 1. It
is not desirable to have a control system where you have to change the sign of the controller gain
depending on whether the other loops are open or closed, so the bottom line is the following:

We do not want to pair output i and input j if lij is negative.

The results for a 2 x 2 system indicate that if l11 > 1 then l12 = l21 < 0. Also, if 0 < l11 < 1,
then 0 < l12 < 1. Let us think about the differences between having a l11 greater than 1 and
having a l11 less then 1.

If l 11 > 1, then .

This tells us that the gain between u1 and y1 is larger with loop 2 open than with loop 2 closed.

If we design the u1-y1 controller with loop 2 open, then close loop 2, we expect somewhat
sluggish behavior, since the effective process gain will be lower than the u1-y1 gain.

However,

If we design the u1-y1 controller with loop 2 closed, then if we have to open loop 2 for some
reason, we could have destabilizing behavior, since the u1-y1 gain will be higher than when
the controller was tuned.

If 0 < l11 < 1, then .

This tells us that the gain between u1 and y1 is smaller with loop 2 open than with loop 2 closed.
 If we design the u1-y1 controller with loop 2 open, then close loop 2, we expect more
aggressive behavior, since the effective process gain will be higher than the u 1-y1 gain.

However,

If we design the u1-y1 controller with loop 2 closed, then if we have to open loop 2 for some
reason, we expect sluggish behavior, since the u1-y1 gain will be lower than when the
controller was tuned.

Midchapter Summarizing Remarks

You must be careful when tuning SISO loops in a multiple input–multiple output (MIMO) system,
because the other loops in the system can greatly affect the gain between the input-output
variables under consideration.

Also, note that we have only been concerned with static (steady-state) effects. Dynamic
interaction has a major effect when control systems are tuned. Our experience is that one rarely
reverses a pairing decision based on the steady-state relative gain, when one considers the
dynamic effects.

The primary result from a dynamic interaction analysis is to recognize additional interaction
affects above the steady-state ones and either accept the degradation in control system
performance, or use a "true" multivariable control system design technique rather than separate
SISO controllers. Sometimes the interaction and other problems will be so severe that only one
loop can be closed.

[ Team LiB ]
[ Team LiB ]

13.6 Return to the Motivating Example

Consider Example 13-1, where 40% ethanol (stream 1) is being blended with pure water (stream
2) to produce a product of 30% ethanol. Here we wish to use the RGA to help us develop the
control structure (pair inputs and outputs).

We first develop the steady-state model, then linearize to find the process gain matrix, and finally
calculate the RGA to decide on variable pairing. For simplicity we neglect density differences
between the streams and components.

Total Material Balance

Neglecting density differences between the streams and components, the total blend product
volumetric flow rate is equal the sum of the volumetric flow rates of the two feed streams,

Component Material Balance on Ethanol

Assume a binary system with z1 = volume fraction of ethanol in stream 1. Let z = volume fraction
of ethanol in the blend stream. Since there is no ethanol in the water stream, we find

Solving these equations for the volume fraction of ethanol in the blend stream,

Let the volume fraction of ethanol be output 1 and total flow rate output 2. Also, inputs 1 and 2
are the flow rates of the whiskey and water feedstreams, respectively. The following steady-state
values were used in Example 13.1: F 1 = 3 gpm, F2 = 1 gpm, z = 0.3 mole fraction ethanol, and F
= 4 gpm.

The steady-state input-output gains are

which yields the process gain matrix

The relative gain relating input 1 to output 1 is

This value of 0.25 indicates that the effective gain relating input 1 to output 1 with loop 2 closed is
four times the value of the gain with loop 2 open. This explains the result shown in Figure 13-11.

The RGA is then

which indicates that the blend total flow (y2) should be paired with the whiskey flow rate (u1) and
the blend ethanol concentration (y1) should be paired with the water flow (u2). Notice that the
largest component stream flow rate then controls the total flow.

[ Team LiB ]
[ Team LiB ]

13.7 RGA and Sensitivity

The previous sections have introduced the concept of the RGA, to help determine which input
should be paired with which output to form a multivariable control system composed of a number
of single input–single output controllers (we refer to this as a MVSISO system). The purpose of
this section is to discuss the relationship between relative gain and controller tuning. It should
become clear that the RGA gives valuable insight into the failure sensitivity of a control system.
We also present numerous examples to illustrate the important concepts.

Failure Sensitivity
When tuning a set of SISO controllers to form a MVSISO strategy, it is important to consider the
failure sensitivity of the system. For example, consider a two input–two output system. If loop 1
is tuned to operate well when loop 2 is closed, we must also consider what happens when loop 2
is opened. It is possible for loop 2 to be opened in a number of ways. The most obvious way is for
an operator to put loop 2 on manual control (i.e., open loop). Another way is if the second
manipulated variable hits a constraint, say if the valve goes all the way open or closed—this
means that the second input can no longer affect the second output, which is equivalent to being
open loop.

Relative Gain as a Perturbation of the Nominal Process

Consider a system where loop one has output 1 paired with input 1. The relative gain is defined as

Equation 13.26

Using k11 to represent the steady-state gain between input 1 and output 1,

Equation 13.27

From Equation (13.27), we can represent the effective gain as

Equation 13.28
where

We can really represent the effect of the relative gain by comparing the two block diagrams
shown in Figures 13-12a and b.

Figure 13-12. Process gain between input 1 and output 1 when loop 2
is open versus when loop 2 is closed. (a) Process gain between input 1
and output 1 when loop 2 is open. (b) Effective process gain between
input 1 and output 1 when loop 2 is closed.

Comparing Figures 13-12a and b, we see that the effective process gain between input 1 and
output 1 with loop 2 is closed is simply a perturbation of the open-loop gain between input 1 and
output 1 by a multiplicative factor of 1/l11.

Relative Gains Between 0 and 1

If a relative gain is less than 1, this means that k11,eff > k11. If a controller for loop 1 is based on
k11 (implicitly assuming that loop 2 is open), then when loop 2 is closed, the controller gain for
loop 1 (kc1 ) should be detuned approximately by a factor of l11, otherwise the control system will
respond too rapidly and there will be a chance of instability. If, on the other hand, the controller
for loop 1 is tuned based on loop 2 being closed, then kc1 can probably be increased by a factor of
1/l11 when loop 2 is opened. The problem is, if the kc1 is kept at this new larger value when loop
2 is closed again, there is again a potential for instability.

This analysis indicates that for safety purposes, if the relative gain is less than 1, then loop 1
should be tuned based on loop 2 closed.

Relative Gains Greater than 1

If a relative gain is greater than 1, this means that k11,eff < k11. If a controller for loop 1 is based
on k11 (implicitly assuming that loop 2 is open), then when loop 2 is closed, the controller gain for
loop 1 (kc1 ) can be increased approximately by a factor of l 11. The problem is, if loop 2 is then
opened again, the controller gain will be too high, and there is a chance of instability. If, on the
other hand, the controller for loop 1 is tuned based on loop 2 being closed, then kc1 should be
decreased by a factor of 1/l11 when loop 2 is opened.

This analysis indicates that for safety purposes, if the relative gain is greater than 1, then loop 1
should be tuned based on loop 2 open.

Negative Relative Gains

Think about the ramifications of having a negative relative gain. If l11 is negative, this means that
the gain between u1 and y1 with loop 2 closed (k11,eff) has the opposite sign compared with the
gain between u1 and y1 with loop 2 open (k11). This has the following ramifications. First of all,
assume that the process is open-loop stable. For open-loop stable processes that have a
controller with integral action, the controller gain must have the same sign as the process gain. If
the controller for loop 1 is designed with loop 2 open, it will have a controller gain with a certain
sign. Then, when loop 2 is closed, it must have an opposite sign in order to be stable! This is
certainly not a desirable situation and is unacceptable in practice. A controller for loop 1 designed
based on g11 (loop 2 open) will be destabilized when loop 2 is closed. If the controller for loop 1 is
designed based on g11,eff (loop 2 closed), the controller will be destabilized when loop 2 is opened.
This leads to the following statement regarding failure sensitivity.

Failure Sensitivity

A control strategy that is based on pairing on a negative relative gain is failure sensitive and is
unacceptable. Never select input-output pairings based on a negative relative gain.

Implication for Subsystems

Many control strategies have more than two control loops. Consider a case where there are three
control loops, for example. If the variable pairings have been performed so that the corresponding
relative gains are positive (and hopefully close to 1.0), then we need to consider the failure
sensitivity of all possible controller failures. For example, if loop 1 fails, it is important then to
make certain that the RGA of the subsystem with loop 1 removed still satisfies the requirement of
positive relative gains on all paired inputs and outputs.

Sensitivity to Model Uncertainty

Large Relative Gains (>>1)

Large relative gains correspond to a high sensitivity to uncertainty. This is shown in the following
example.

Example 13.3: Model Uncertainty and the RGA

Consider a steady-state process model with the corresponding RGA:

The actual process could very easily have a minor error of 5% in the gain relating u1 to y2, as
shown in the actual process gain matrix and RGA:

Notice that the model requires that input 1 be paired with output 1, while the relative gain for the
actual plant (model + uncertainty) indicates that input 1 must be paired with output 2. In this
case, we have no choice but to use only one control loop. In this case we would use physical
insight and dynamic considerations to select the loop to close.

Sensitivity to Model Uncertainty

A large relative gain (the order of 25 and higher) corresponds to an extreme sensitivity to model
uncertainty and indicates that some loop should be opened; that is, not all outputs should be
under feedback control. One process that can have multivariable control, even with a large
relative gain, is distillation, where nonlinear effects dominate.

[ Team LiB ]
[ Team LiB ]

13.8 Using the RGA to Determine Variable Pairings

In the previous section, we showed how to tune controllers, assuming that a decision on the
pairing had already been established. The purpose of this section is to give some examples to
illustrate how inputs and outputs should be paired to form MVSISO control loops.

Example 13.1, continued

Consider the whiskey-blending problem, which had the steady-state process gain matrix and
RGA:

indicating that the output-input pairings should be y1-u2 and y2-u1. In order to achieve this
pairing, we could use the block diagram shown in Figure 13-13.

Figure 13-13. Control block diagram for 1-2/2-1 pairing.

Notice that the difference between r2 and y2 (the error in output 2) is used to adjust u1, using a
PID controller (gc1 ), hence, we refer to this pairing as y2-u1. Similarly, the difference between r 1
and y1 (the error in output 1) is used to adjust u2, using a PID controller (gc1 ); hence, we refer to
this pairing as y1-u2. This corresponds to the physical diagram shown in Figure 13-3b.

An alternative is to rearrange (renumber) our outputs or inputs so that the natural pairing is y1-u1
and y2-u2 (in the renumbered variables). For example, we could renumber the outputs. Let
and , where * indicates the newly redefined variable. This corresponds to the control and
instrumentation diagram shown in Figure 13-14.

Figure 13-14. Control instrumentation diagram for redefined variables.

We can write

or

The reader should verify that the new RGA is

Notice that redefining the outputs is equivalent to switching the rows of the process gain matrix
related to the corresponding outputs (in this case, switching rows 1 and 2 is equivalent to
exchanging outputs 1 and 2). The new RGA is found by making the same transformation of rows
in the RGA, as was performed on the process gain matrix.

Similarly, redefining the inputs is equivalent to switching the columns of the process gain matrix
related to the corresponding inputs. The new RGA is found by making the same transformation of
columns in the RGA, as was performed on the process gain matrix.

Generalization
We have shown, by way of example, that outputs can be renumbered by switching the rows in
the process gain matrix, and inputs can be renumbered by switching columns in the process gain
matrix. Each time we perform this exchange of variables, we do not need to recalculate the
relative gain array—we simply find the new RGA by performing the same row and column
exchanges on the RGA that we performed on the process gain matrix. By performing this
exchange (renumbering) of variables, we can always rearrange our inputs and outputs such that
loop 1 has a pairing of y1 and u1 (in the newly defined variables). For this reason, we will often
assume that a y1-u1 pairing has been made and discuss the effect of l 11 on controller tuning.

Example 13.4: A Three Input–Three Output System

Consider the following RGA for a system with three inputs and three outputs:

How would you choose input-output pairings for this process?

Solution:

1. First of all, we know that we do not want to pair on a negative relative gain.

2. Second, we do not want to pair with a relative gain of 0 because that means that the
particular input does not have an effect on the particular output, when all of the other loops
are open.

Look at row 3 in the relative gain array; this corresponds to output 3. We would not pair y3 with
u3 because of the 0 term. We also would not pair y3 with u1 because of the –3 term. This means
that y3 must be paired with u 2. Let us indicate this choice in our RGA,

Now we have eliminated the third output (row) and the second input (column) from our selection
process. From the first row (output), we see that we would not pair y1 with u3 because of the –1
term. We also cannot pair y1 with u2 because we have already paired u2 with y3 (hence, the
circle). Our only choice is to pair y1 with u1. Let us indicate this choice in our RGA,

Now, we have eliminated outputs 1 and 3 (rows 1 and 3) as well as inputs 1 and 2 (columns 1
and 2). We have no choice but to pair y2 with u3, and we make that choice below:

Notice also, that we can renumber our outputs, such that pairings of the new variables occur on
the diagonal. For example, if we define and , this is equivalent to switching the
second and third rows to obtain

which is based on the new output vector

Example 13.5: A Four Input–Four Output Distillation Column (Alatiqi

and Luyben, 1986)
The steady-state gain input–output relationship is

The RGA (using the m-file rga.m shown in Appendix 13.2) is

Now, let us systematically choose our pairings. Looking at column 3, we see only one relative gain
that is not negative, therefore, we must pair y2 with u3. Looking at row 4, we notice that there is
only one relative gain that is not negative, therefore, we must pair y4 with u1. Looking at column
2, there are 2 nonnegative relative gains, but we have already used output 2, so we must pair y3
with u2. This leaves y1 and u4, which, fortunately, has a favorable relative gain (0.8546). If the
y1-u4 relative gain had not been favorable, we would have been forced to drop output 1 and input
4 and simply have three control loops for our system.

[ Team LiB ]
[ Team LiB ]

13.9 MATLAB RGA Function File

The M ATLAB function file, rga.m, given in Appendix 13.2 can be used to the find the RGA of any
square (number of inputs = number of outputs) gain matrix. For example, consider the wet
grinding circuit of Exercise 10.

First, the gain matrix is entered

» g = [119 153 -21; 370 767 -50; 903 -667 -1033]

g =
119 153 -21
370 767 -50
903 -667 -1033

Then the RGA is calculated using the call to the function file:

» lambda = rga(g)
lambda =
3.6449 -1.9131 -0.7318
-2.3616 2.9581 0.4035
-0.2833 -0.0450 1.3283

[ Team LiB ]
[ Team LiB ]

13.10 Summary
The purpose of this chapter was to develop a technique (the RGA) to assist us with the decision of
how to pair input and output variables to form multiple SISO control loops. The relative gain
between output i and input j is defined as

and the RGA is a matrix of relative gains. Some of the important things that we learned about the
RGA are summarized below.

The rows and columns of the RGA sum to 1

Do not pair on a negative or zero relative gain

Try to pair on relative gains close to 1

If you cannot pair on positive relative gains, or if the relative gains are high (>25), then the
number of closed-loop control loops must generally be reduced; that is, one or more loops
must be placed on manual control

Switching rows on the process gain and relative gain matrices is equivalent to renumbering
the outputs; switching columns on the process gain and relative gain matrices is equivalent
to renumbering the inputs

After selecting variable pairings, check the relative gain array for all subsystems to make
certain that they are not sensitive to controller failure

[ Team LiB ]
[ Team LiB ]

References

A nice summary of the features of the relative gain array are presented in the following:
Grosdidier, P., M. Morari, and and B. Holt, "Closed-Loop Properties from Steady-State Gain
Information," Ind. Eng. Chem. Fund., 24(2), 221–235 (1985).

The following sources were used for some of the RGA examples in this chapter: Alatiqi, I. M., and
W. L. Luyben, "Control of a Complex Sidestream Column/Stripper Distillation Configuration," Ind.
Eng. Chem. Proc. Des. Dev., 25(3), 762–767 (1986).

Doukas, N., and W. L. Luyben, "Control of Sidestream Separating Ternary Mixtures," Instrum.
Technol. (INTECH), pp. 43–48 (June, 1978).

Exercise 12 is presented in the following: Reeves, D. E., and Y. Arkun, "Interaction Measures for
Nonsquare Decentralized Control Structures," AIChE J., 35(4), 603–613 (1989).

The grinding circuit problem (Exercise 10) was presented in the following: Hulbert, D. G., and E.
T. Woodburn, "Multivariable Control of a Wet Grinding Circuit," AIChE J., 29, 186 (1983).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: A 3 x 3 system has the following steady-state gain matrix and RGA:

How would you pair the variables for an MVSISO strategy and why?

2: A two input–two output process has the following RGA:

Assume that the MVSISO system is tuned for stable feedback control with both loops
closed.

Case 1— u 1 is paired with y1 (loop 1) and u2 is paired with y2 (loop 2). What happens if
loop 2 is opened? Why? Must loop 1 be retuned? How?

Case 2— u 1 is paired with y2 (loop 1) and u2 is paired with y1 (loop 2). What happens if
loop 2 is opened? Why? Must loop 1 be retuned? How?

3: A system has the following process gain matrix:

How would you pair variables for your control loops? Why?

4: A three input–three output process has the following RGA:

How would you pair variables for your control loops? Why?

5: Consider the following process gain matrix and RGA:

 a. Can three SISO control loops be used?

b1. If your answer to a is yes, what are the recommended pairings [will
these pairings assure stability if one of the loops must be opened (think
about the 2 x 2 subsystems)]? Reorder the inputs and outputs such that the
paired variables are on the diagonal of the new G(0) matrix. Show the new
G(0) matrix and the new L matrix.

b2. If your answer to a is no, do you recommend a reduction to two loops?

If so, what are your recommendations for input-output pairings for this
system?

6: Consider the following steady-state gain matrix for a process with three inputs and
four outputs. [This is for a Benzene-Toluene distillation column, with a side draw—see
Doukas and Luyben (1978) for details.] Assuming all the outputs are of equal
importance, which input-output pairings would you use for a MVSISO control system?

7: For a 2 x 2 system, show that a y1-u2 and y2-u1 pairing yields the following result for
l12.

where

8: Consider the steady-state process gain matrix

which yields the following RGA:

 Is it possible to pair outputs and inputs to form three separate single-loop controllers?
If it can be done, do it. If it cannot be done, explain why and suggest another
multivariable control strategy.

9: Your boss has told you to develop a control strategy for the following stirred-tank
heater, which has two inputs (rate of heat addition, Q, and flow rate to tank, F i), and
two outputs (tank height and temperature). She also gave you the following transfer
function matrix, assuming that you knew which output corresponded to which
measured variable and which input corresponded to which manipulated variable.

Use physical reasoning to determine which output corresponds to which measurement

(i.e., is y1 the tank height or the tank temperature) and which input corresponds to
which manipulated variable (i.e., is u1 the heat addition rate or the flow rate to the
tank?). If two SISO loops are used to control this process, how should the variables be
paired? Renumber the outputs such that the output-input pairings are on the diagonal
of the transfer function matrix. Show the new transfer function matrix.

10: The objective of a wet grinding circuit is to grind large solid material to form granular
material. Solids are fed with a water stream to a ball mill where the solids are ground.
The slurry effluent from the ball mill is fed to a sump which is used as surge capacity.
The slurry is pumped to a cyclone separator, which separates water from solid material
and returns water to the ball mill. A schematic diagram is shown below.
 Consider the following process model (Hulbert and Woodburn, 1983)

Which has the following RGA

a. What pairing is suggested by the RGA for this system? Why?

b. Connect the appropriate input and output variables with controllers, on the
diagram above, to form a control instrumentation diagram.

c. Discuss the failure sensitivity of this three input–three output system. That is, if
any of the control loops failed, would the resulting two input–two output system
have the correct pairing?

11: A forced circulation evaporator is used to concentrate dilute liquors by evaporating

solvent from a feed stream. Feed is mixed with recirculating liquor and pumped into
the tube side of a vertical heat exchanger. Steam enters the shell side of the vertical
heat exchanger and condenses on the outside of the tubes. The liquor is partially
vaporized as it passes through the tube side of the heat exchanger and passes into the
separator. Vapor from the separator is condensed by a cooling water exchanger, while
a portion of the liquid stream from the separator is withdrawn as product and the rest
 is recirculated to the heat exchanger. A schematic process and instrumentation
diagram is shown below.

It seems fairly clear that it is desirable to control the level in the separator (y1) by
manipulating the product flow rate (u1). It is particularly important to keep the level
under feedback control, since is has integrating action (a pole is at the origin). With
the level controlled by manipulating the product flow rate, the following steady-state
input-output relationship is found for the other variables

How should the other loops be paired? Why?

12: Consider the following system with two outputs and three inputs (Reeves and Arkun,
1989). This model is for a mixing tank with three input streams. The outputs are the
concentration and tank height, while the three inputs are stream flow rates:

To use a MVSISO strategy, we must remove one of the inputs (set it to a constant
value, rather than using it for feedback control).

13: As a promising young engineer, you are placed in charge of a gasoline blending facility.
This facility blends three refinery streams together to form a final product stream that
must meet road octane (RO) and Reid vapor pressure (RVP, psi) specifications. Two
 streams are available for manipulation. You are given the following data

Flow 1 Flow 2 RO RVP Total flow

10 Bbl/min 20 Bbl/min 90.000 10.000 60 Bbl/min

11 20 90.164 10.164 61

10 21 90.164 9.918 61

Let Flow 1 be the first input and Flow 2 be the second input. RO is the first output and
RVP is the second output.

a. Calculate the RGA.

b. How would you pair the input-output variables for this process? (Why?)

c. Complete the control instrumentation diagram shown below.

d. Draw a control block diagram for this process (with the proper pairings). (Are
there any cascade loops?)

14: Consider the whiskey blending problem. A new market for 70 proof whiskey (0.35
volume fraction ethanol) has developed. Assume that the feedstreams are still 80
proof whiskey and pure water, and that the desired production rate of 70 proof
whiskey is 4 gpm. Find the new gain matrix and relative gain array. What are the
implications for controller pairing?

[ Team LiB ]
[ Team LiB ]

Appendix 13.1: Derivation of the Relative Gain for an n-

Input–n-Output System
The relative gain between input j and output i is defined as

Output i is the following steady-state function of all of the inputs:

The relationship between input j and output i with all other inputs constant is

For any desired output vector, y, we can find the corresponding input vector, u, from

Defining (0)=G-1(0)=K-1

The relationship between all outputs and input j is

The relationship between output i and input j with all other outputs constant is
and

The relative gain between input j and output i is

Equation 13.23

where element of G–1(0).

What Equation (13.23) tells us is that the ij element of the RGA is found by multiplying the ij
element of the steady-state process gain matrix by the ji element of the inverse of the steady-
state process gain matrix.

[ Team LiB ]
[ Team LiB ]

Appendix 13.2: m-File to Calculate the RGA

function lambda = rga(g)

%
% calculate the relative gain array for a
% square gain matrix of any size
%
% first, check the dimensions to see if the
% process gain matrix is square
%
[mrow ncol]=size(g);
%
% if the gain matrix is not square, let the
% user know
%
if mrow ~= ncol;
disp('needs to be square, buddy')
end;
else
%
% the matrix ghat is simply the inverse of the
% gain matrix. ghatt is the transpose of ghat
%
ghat = inv(g);
ghatt=ghat';
%
% perform an element by element multiplication
% of g and ghatt to find the relative gain array
%
lambda = g.*ghatt;
end

[ Team LiB ]
[ Team LiB ]

Chapter 14. Multivariable Control

After studying this chapter, the reader should be able to do the following:

Understand the performance limitations of processes with RHP transmission zeros

Understand the effect of "direction" in a multivariable system

Perform matrix transfer function block-diagram analyses and understand that the order of
multiplication is very important

Design a decoupling controller

Understand the basic multivariable IMC design

The major sections of this chapter are as follows:

14.1 Background

14.2 Zeros and Performance Limitations

14.3 Scaling Considerations

14.4 Directional Sensitivity and Operability

14.5 Block-Diagram Analysis

14.6 Decoupling

14.7 IMC

14.8 MATLAB tzero, svd, and LTI Functions

14.9 Summary

[ Team LiB ]
[ Team LiB ]

14.1 Background
Chapter 13 presented the problem of control-loop interaction, and a method (RGA) to help
determine the best input-output pairing to form a set of SISO control loops. The goal of this
chapter is severalfold. One is to understand the inherent performance limitations of a
multivariable system due to RHP transmission zeros. Also, multivariable systems exhibit a
"directional sensitivity" that can be detected by performing a singular value decomposition (SVD).
Use of the SVD requires that the gain matrix be properly scaled, so that the scaled inputs and
outputs have roughly the same ranges of expected values.

Two forms of multivariable control are presented in this chapter: decoupling and IMC. The more
commonly used multivariable technique in practice, model predictive control (MPC), is the topic of
Chapter 16.

[ Team LiB ]
[ Team LiB ]

14.2 Zeros and Performance Limitations

Recall that RHP zeros severely limit the performance of SISO processes. For example, a model
inverse can not be implemented as the controller, because this would cause the controller to be
unstable. A similar problem exists for multivariable processes; RHP transmission zeros limit MIMO
performance.

SISO Zeros
For a SISO process transfer function with the form

Equation 14.1

the zeros are the roots of the numerator polynomial, N(s). The inverse of the process transfer
function is

Equation 14.2

That is, the zeros of g(s) are the poles of g–1(s). A RHP zero then results in an inverse transfer
function with a RHP pole, which is unstable. Some model-based control strategies, such as IMC,
make use of a model inverse for control system design. In Chapter 8, we found that the RHP
zeros must be factored out, and that only a portion of the model could be inverted for controller
design.

Multivariable Transmission Zeros

The transmission zeros of a multivariable transfer function matrix are the values of s that cause
the input-output matrix to lose rank. That is, the transmission zeros are the values of s that
cause

Equation 14.3

Similar to single variable systems, the multivariable transmission zeros become the poles of the
inverse system. RHP transmission zeros for a particular transfer function matrix then become RHP
poles for the inverse transfer function matrix (that is, the inverse is unstable). Again, a
multivariable controller that simply tries to invert the process model will be unstable.

We illustrate these concepts with the following example. Here we use the symbol z to represent
the zero.

Example 14.1: Calculation of Transmission Zeros

The calculation of transmission zeros are illustrated for two different systems (A and B).

System A

Consider the transfer function matrix, where the time unit is minutes

The determinant of G(s) is

and solving for det[G(s)] = 0,

We can also see that the rank of the G matrix drops to 1 when s = 3 min-1,

since the first column of the matrix is now a multiplicative factor (3) of the second column. Note
that although the zero of the g11 element is -3 min-1, the matrix transmission zero is +3 min -1
(that is, a RHP transmission zero)!

The inverse[1] of the transfer function matrix is

[1] Recall that the inverse of a 2 x 2 matrix, , is calculated by

which has an unstable pole of p = +3, as expected.

It is possible for an individual element to have a RHP (positive) zero while the matrix transmission
zero is in the left-half-plane (negative); this is illustrated by the following transfer function matrix.

System B

Consider the transfer function matrix

which differs from system A by the RHP zero in the g11(s) transfer function. The determinant of
G(s) is

and solving for det[G(s)] = 0,

We can also see that the rank of the G matrix drops to 1 when s = -3,

since the first column of the matrix is (like system A) a multiplicative factor (3) of the second
column. Note that although the zero of the g11 element is +3 (RHP), the matrix transmission zero
is -3 min-1 (that is, a left-half-plane transmission zero).

It is left to the reader to show that the inverse of this matrix has a stable pole at -3 min-1.

An example of a physical process with a RHP transmission zero is shown in Figure 14-1. This is a
laboratory-scale system developed by Johansson (2000) to study the effect of operating condition
on the location of the transmission zeros. The inputs are the voltages to the pumps (v1, v2) and
the outputs are the voltage measurements of the height of liquid in tanks 1 and 2 (y1, y2). The
flow from pump 1 is split between tanks 1 and 4, while the flow from pump 2 is split between
tanks 2 and 3. Depending on the actual flow splits, it is possible to obtain a positive RHP
transmission zero. Numerical details are presented in Example 14.2.

Figure 14-1. Quadruple tank control problem.

Although this is a "toy" problem, you can see how some processes with recycle might have similar
behavior.

Example 14.2: Quadruple Tank Problem (Johansson, 2000)

This tank system can be operating at two operating points. The first one has two transmission
zeros that are in the left-half plane (negative), while the second has a RHP transmission zero.

Operating Point 1

Operating point 1 has the following steady-state inputs and outputs:

The result is the process transfer function matrix

where the gains are volts/volt, and the timescale is seconds. The transmission zeros are

which are both in the left-half-plane (so the matrix inverse will be stable).

Operating Point 2

Operating point 2 has the following steady-state inputs and outputs

This operating point has the transfer function matrix

The transmission zeros are

Notice that there is a RHP transmission zero, so the matrix inverse will be unstable. In Exercise 1,
you will find that the feedback control performance of this operating point is severely limited
compared with the first operating point.

The M ATLAB routine for calculating transmission zeros is discussed in Section 14.7.

[ Team LiB ]
[ Team LiB ]

14.3 Scaling Considerations

To properly use the SVD method presented in Section 14.4, it is necessary to scale the process
transfer function matrix such that expected ranges of all scaled inputs and outputs are
approximately the same.

An output scaling matrix can be found to assure that the scaled outputs cover the same range.
Let (*) represent the scaled variables. A diagonal scaling matrix for the output variable is

Equation 14.4

Similarly, a diagonal scaling matrix for the input variable is

Equation 14.5

Since the input-output relationship is

Equation 14.6

we find

Equation 14.7

Also,

Equation 14.8

so the scaled input-output relationship is

Equation 14.9
yielding the scaled transfer function matrix

Equation 14.10

This is illustrated by the following example.

Example 14.3: Mixing Tank

The mixing tank shown in Figure 14-2 mixes two streams together (plus a possible disturbance
stream). It is desirable to control both the level (y1) and the temperature (y2) in the tank. The
flow rates of the hot (u1) and cold (u2) streams can be manipulated. The flow rate out of the tank
is proportional to the square root of the height of liquid in the tank ( ). The residence time
at steady state is 10 minutes. The steady-state flow rates and temperatures are as follows:

Figure 14-2. Mixing tank.

Stream Temperature Flow, liters/minute

Hot (manipulated) 60°C 25

Cold (manipulated) 10°C 25
Cold (disturbance) 10°C 0
Outlet 35°C 50

The cross-sectional area of the tank is 0.581 m2. The steady-state tank height is then 0.86 m,
with a steady-state volume of 0.5 m3 or 500 liters (assume that the tank is half full at steady
state). The input-output transfer function relationship is (see Exercise 3).

Equation 14.11

The time constant unit is minutes, and the units of the gains are

It is interesting that the height time constant is twice the temperature time constant. Notice that
if the height unit was centimeters and the temperature unit was degrees Celsius, the transfer
function matrix would be

Equation 14.12

If, in addition, the time units were seconds, the input-output relationship would be

Equation 14.13

The point of this is that it is very important to understand the units of the variables you are
dealing with, and that the parameters can vary (numerically) by orders of magnitudes depending
on those units.

In this example, the expected range of the manipulated variables is roughly the same, since we
assume that the control valves are designed to handled flows roughly twice the nominal steady-
state values. That is, we expect the manipulated inputs to be bounded between 0 and 50
liters/minute.

Notice that the outputs do not cover the same range. For one thing, the units are different (m for
height, and °C for temperature).

The steady-state value of the tank height is 0.86 m, with a minimum value of 0 m and a
maximum height of 1.72 m. We desire, however, for the height to stay between 20 and 80% of
the full tank level. That is, the height should range from 0.344 to 1.376 m. Since we work in
deviation variables, and it is convenient to cover a range of -1 to 1, then it is natural to scale by
0.516 m (this is found from 0.86 to 0.344). We can then define a scaled y1 variable as y*1, where

Similarly, it is convenient to scale the second output. The approximate range of temperatures is
from 10° to 60°C, with a steady-state value of 35°C. Again, since we are working with deviation
variables and it is natural to consider variations of -1 to 1 in the scaled second output, we use
25°C (obtained from 35°C - 10oC) as the scaling factor. The second scaled output is then

And the scaled output vector is related to the dimensional output vector as

Equation 14.14

Then from Equation (14.11),

Equation 14.15

The gains indicate that the outputs will vary approximately the same order of magnitude (there is
roughly a factor of 3 difference in the gains).

Although not important in this particular example, the inputs can be scaled as well. Since the
steady-state manipulated input flow rates are each 25 liters/minute, and we can assume that the
flows each range from 0 to 50 liters/minute, it is convenient to scale each input by a factor of 25
liters/minute.

Equation 14.16
or

Equation 14.17

and we can write [combining Equations (14.15) and (14.17)]

Equation 14.18

Now it is clear from the gain matrix that a full manipulated variable move will cause the desired
tank height range to be violated, while the temperature range will be satisfied.

When dealing with dynamic systems, it is always important to know the units of all parameters
and variables, and to understand expected operating ranges for all inputs and outputs. It is
convenient to scale the inputs and outputs such that the expected operating range is -1 to +1; for
one reason, it is easier to compare responses on the same set of plots. Also, some analysis tools
(such as the SVD presented in Section 14.4) can be interpreted only on properly scaled systems.

[ Team LiB ]
[ Team LiB ]

14.4 Directional Sensitivity and Operability

In a number of examples in this textbook, we have noted that the design of a process can limit
the ability to control the process. For example, if the "gain" for an input-output pairing is not large
enough, then there is a limit to how large a setpoint change can be made before the manipulated
input saturates (hits a constraint, such as a valve being fully open or closed). For multivariable
systems, the effect of a particular input on a particular output is influenced by the other inputs.
For example, consider the mixing tank of Example 14.3. A change in the hot-stream flow rate can
significantly affect the temperature if the cold-stream flow rate remains constant. If the cold-
stream flow rate changes simultaneously, however, the change in the hot-stream flow rate may
not change the temperature. There are two ways of exploring these multivariable effects. One is
to apply a SVD to the steady-state process gain matrix; this is an inherently linear procedure. An
alternative is to "map out" an operating window or the possible range of process outputs that
corresponds to a range of possible process inputs. This can be done with either the linear model
or a fundamental nonlinear model.

In this section, we first present the SVD technique, followed by the operating window technique.

SVD
The SVD can be used to predict the directional sensitivity of a process. The process gain matrix is
decomposed into three matrices,

Equation 14.19

where U is the left singular vector matrix, S the diagonal matrix of singular values, and V the right
singular vector matrix. The left and right singular vector matrices are both orthonormal matrices;
that is, each column of the matrix is orthogonal to all other columns and the columns each are
unit length. The diagonal singular value matrix is ordered so that the largest singular value is in
the (1,1) position. Note that the standard notation for SVD is to use U to represent the left
singular vector matrix. Please do not confuse this with the u vector commonly used to represent
the vector of manipulated inputs.

The left singular vector matrix indicates the strongest and weakest output directions, while the
right singular vector matrix indicates the strongest and weakest input directions. The ordered
matrix of singular values provides the "magnitude" of the strongest and weakest directions. The
ratio of the minimum and maximum singular values is known as the condition number of the
matrix. Large condition numbers are indicative of ill-conditioned systems, which effectively lose a
degree of freedom (not all inputs are truly independent).

This analysis has the following control-related implications. It is relatively easy to make a setpoint
change in the strongest output direction (indicated by the first column of the left singular vector
matrix), and it is relatively hard to make a setpoint change in the weakest output direction
(indicated by the last column of the left singular vector matrix).

Here we use the mixing tank example to illustrate the use of SVD analysis.
Example 14.3, continued
The singular value decomposition of the mixing tank gain matrix (steady state, s = 0) is (see
Section 14.8 for the MATLAB calculation)

The first column of the U (left singular) matrix indicates that output 1 (liquid height) is the
strongest output direction, while the second column indicates that output 2 (temperature) is the
weakest input direction. The first column of the V (right singular) matrix indicates that the
strongest input direction is to increase (or decrease) both inputs by the same amount, while the
second column indicates that changing both inputs in opposite directions (decreasing one,
increasing the other) results in the weakest effect. The matrix of ordered singular values (S)
indicates that the strongest direction has 3.33 times (2.3575/0.7071) the effect of the weakest
direction.

These results have the following physical interpretation. A simultaneous increase in both the hot
(input 1) and cold (input 2) flow rates will increase the tank height (output 1) but will not change
the outlet temperature (output 2). Similarly, a decrease in the hot flow rate simultaneously with
an increase in the cold flow rate will not change the tank height but will result in a temperature
decrease.

First, consider a simultaneous increase in both flow rates (notice that the "magnitude" of the input

vector is and refer to this as input vector a:

The output for this particular input can be found from

That is, the scaled height increases by 2.3575, but there is no change in temperature. This can be
viewed in Figure 14-3, where the input vector is in the direction (u 1,u2) = (0.71,0.71) and the
output vector is in the direction (y1,y2) = (2.36,0).

Figure 14-3. Input vector; output vector.

Now, consider a decrease in hot flow rate simultaneously with an increase in cold flow rate (again,

notice that the "magnitude" of the input vector is ) and refer to this as
input vector b:

The output can be found from

That is, the scaled temperature decreases by 0.7071, but there is no change in height. Notice that
we can construct Figure 14-4, where the "unit circle" on the left represents all possible input
vectors with a length of 1. The input vectors ua and ub are also shown. The ellipse on the right
represents the corresponding set of outputs (obtained from y = Gu). Notice that a is the longest
output vector and b is the shortest. That is, a unit input in the a direction results in an output in
the a direction shown, with a length of the largest singular value. Similarly, a unit input in the b
direction results in the output shown, with a length of the smallest singular value. All other inputs
result in output vectors with lengths between the smallest and largest singular values.

Figure 14-4. Input to output mapping.

A minor disadvantage of the SVD analysis is the assumption that all possible input vectors are the
same length. For this example, you could argue that it is possible for both the hot and cold flows
to be at their maximum values of 1. A slightly more realistic interpretation of the input-output
mapping is then shown in Figure 14-5.

Figure 14-5. Input to output mapping.

The output space, in physical (rather than scaled) variables, is shown in Figure 14-6. Notice that
the linear analysis predicts negative height values, which clearly cannot be obtained in practice.

Figure 14-6. Linear output space in physical coordinates.

Operating Window
This approach is similar to the "rectangular" mapping shown in the previous example. The idea is
to map the effect of a region of process inputs onto a region of process outputs. Here, simulations
are performed using the steady-state nonlinear model, with various values for the inputs.
Example 14.3, continued
The nonlinear steady-state modeling equations result in the two steady-state output equations

Here it is understood that the steady-state values are being used, so the subscript s has been
neglected. The flow coefficient is under these conditions. For the flows
bounded between 0 and 50 liters/minute each, we find the output region shown in Figure 14-7.
The vessel will overflow for tank heights greater than 1.72 m.

Figure 14-7. Output operating window based on nonlinear model. The

input space is a rectangle with flows between 0 and 50 liters/min.

Notice that some high and low temperatures cannot be obtained at heights between 0.86 and
1.72 m; this is also predicted by the "linear rectangle" analysis shown earlier. At low tank heights,
however, the entire temperature range is feasible; this is not predicted by the previous linear
analysis and is a nonlinear effect.

We have shown that the SVD provides insight about the strong and weak input and output
directions. The primary limitation is the assumption of a linear model. Linear models are suitable
for perturbations from a nominal operating point and are generally very useful for control system
design. After all, the purpose of a controller is to maintain the process close to a desired operating
point. However, linear models usually do not predict the behavior of a nonlinear system over the
entire range of operating conditions. To explore a "window" of operating conditions, the nonlinear
model should be used. If operating the process at a different operating point is desired, the model
can be linearized at that point for controller design and analysis. Another limitation to the analysis
presented in this section is that, like the RGA, it is based solely on steady-state information.
Dynamics certainly play an important role, but this steady-state analysis provides a quick
screening tool to understand physical limitations to steady-state operating conditions.

[ Team LiB ]
[ Team LiB ]

14.5 Block-Diagram Analysis

Consider the multivariable block diagram shown in Figure 14-8. Each of the variables shown on
this diagram is a vector and the blocks represent matrix transfer functions. That is, for two
process outputs, two manipulated inputs, and one disturbance, the corresponding vectors are
given below.

Figure 14-8. Multivariable block diagram.

Setpoints, outputs, manipulated inputs, and the disturbance input are

The controller, process, and disturbance matrices are

Notice that if SISO controllers are used, the controller transfer function matrix is

where gc1 (s) and gc2 (s) will usually be PID-type controllers. The relationship between inputs (both
manipulated and disturbance) and outputs is
Since the error signal is e(s) = r(s) - y(s), we can derive (where I is the identity matrix)

and using matrix inversion to solve for y(s), we find

Equation 14.20

This last equation is the closed-loop transfer function relationship for multivariable systems.
Notice that the order of multiplication is very important. In general, GP(s)Gc(s) Gc(s)Gp(s).
This is clear if, for example, Gp is a 4 x 3 matrix and Gc is a 3 x 4 matrix.

Recall that the SISO closed-loop transfer function is

For SISO systems, the order of multiplication does not matter; for MIMO systems, it is crucial.

[ Team LiB ]
[ Team LiB ]

14.6 Decoupling
We have seen the problems with control-loop interaction; these occur because a manipulated
input affects more than one controlled output. One approach to handling this problem is known as
decoupling. The idea is to develop "synthetic" manipulated inputs that affect only one process
output each. This approach is illustrated in Figure 14-9.

Figure 14-9. Decoupling control strategy—synthetic inputs and process

outputs.

The relationship between the synthetic input vector and the process output vector is

Equation 14.21

For a two input–two output process,

Equation 14.22
where Gp(s)D(s) is a 2 x 2 transfer function matrix. There are a number of possible choices for
the "target" Gp(s)D(s) matrix. Two common ones lead to ideal decoupling and simplified
decoupling.

Ideal Decoupling
One possible choice for Gp(s)D(s) is

Equation 14.23

Solving for D(s), we find

Equation 14.24

where we use the (~) notation to make it clear that these calculations are performed based on a
process model. The relationship between the synthetic inputs and process outputs, y(s) =
GP(s)D(s)u*(s), that results is

Equation 14.25

and we see that each synthetic input affects only one process output (assuming a perfect model).

The main advantage is that independent SISO tuning parameters can be used for each control
loop. A major disadvantage is that the resulting decoupler is the inverse of the process transfer
function matrix. There are numerous problems with this approach, including the fact that the
decoupler will be unstable if there are RHP transmission zeros associated with the process transfer
function matrix. Also, this type of decoupler is known to be extremely sensitive to model error.
Simplified Decoupling
An alternative approach, known as simplified decoupling, is shown for a 2 x 2 example in Figure
14-10. Here, the decoupling matrix is restricted to the form

Equation 14.26

Figure 14-10. Simplified decoupling control strategy—synthetic inputs

and process outputs.

Here, we specify a decoupled response and the decoupler with the structure in Equation (14.26),

Equation 14.27

and we can solve four equations in four unknowns to find

Equation 14.28
Notice that the decoupling elements (d12 and d21) are the same as the feed-forward control
design presented in Chapter 10. The effect of the synthetic inputs are being "fed-forward" to
minimize control loop interaction. Similar to feed-forward control, some factorization of the
decoupling elements must be performed to make certain that they are stable and physically
realizable.

The feedback control diagram for simplified decoupling is shown in Figure 14-11. Whether ideal or
simplified decoupling is being implemented, the result is a true multivariable strategy, as shown in
Figure 14-12.

Figure 14-11. Feedback control using simplified decoupling.

Figure 14-12. Illustration of multivariable control using decoupling.

Static Decoupling
In practice, static decoupling (based on steady-state gains) is used much more often than
dynamic decoupling, since the dynamic version may not be physically realizable, or the dynamic
parameters may be too uncertain. In this case, simply set s = 0 in the decoupling equations.

[ Team LiB ]
[ Team LiB ]

14.7 IMC
The IMC design procedure for MIMO systems is similar to the design procedure developed in
Chapter 8 for SISO systems. The IMC block diagram is shown in the Figure 14-13, where the
blocks now represent matrix transfer functions and the inputs and outputs are vectors. We first
use an example to illustrate the multivariable IMC design procedure, then follow with the general
procedure.

Figure 14-13. IMC.

Example 14.1, continued

Here we consider system A from Example 14.1. Recall that system A had a RHP transmission zero
(z = 3 min-1)

There are a number of ways to factor this matrix. The easiest way is to place the RHP
transmission zero on the diagonal of the "bad (noninvertible) matrix."
Since , we can solve for the "good matrix" from , and find that

Where the controller is

Assume a perfect model with no disturbances. For l1 = l2 = 0.333 min, in Figure 14-14 we find
the response to a step setpoint change in output 1. Notice that the output response is decoupled,
although both manipulated inputs are changed. Also, although the original g11(s) transfer function
did not have a RHP zero, the multivariable system has inverse response behavior in the closed
loop. It should be noted that, in general, the IMC filter factors (l1 and l2) should be tuned to have
different values.

Figure 14-14. Response to a setpoint change in output 1, with a

diagonal factorization matrix and l1 = l2 = 0.333 min.
The reader should show that a setpoint change in output 2 also leads to inverse response
behavior in output 2.

A major disadvantage to the diagonal factorization is that inverse response appears in all output
setpoint responses. It is also possible to perform a factorization that places all the inverse
response behavior in one of the output variables, resulting in good closed-loop performance in the
other outputs. See Holt and Morari (1985) for more details.

The general multivariable IMC procedure is as follows.

1. Factor the process model into invertible and noninvertible elements.

Equation 14.29

The difficulty is that there are many ways to factor the process transfer function matrix. One
way is to place the RHP transmission zero on the diagonal.

Equation 14.30
 Although this is perhaps the simplest method, it usually does not result in the best
performance. Other factorization methods are beyond the scope of this text. The motivated
reader should consult Morari and Zafiriou (1989) for more details.

2. Form the idealized controller.

Equation 14.31

3. Add a filter to make all elements in the controller matrix proper.

Equation 14.32

where F(s) is normally a diagonal matrix.

Equation 14.33

4. The filter factors are adjusted to vary the response and robustness characteristics.

In the limit of a perfect model ( ), the response will be

Equation 14.34
 and where, as in the SISO case, the "bad stuff" must appear in the output response.

[ Team LiB ]
[ Team LiB ]

14.8 MATLAB tzero, svd, and LTI Functions

The M ATLAB Control Toolbox has a routine (tzero) for calculating the transmission zeros of a
multivariable model. MATLAB can also be used to perform a SVD analysis (using the function svd).

Transmission Zero Calculation

Here we illustrate the use of tzero by way of an example.

Example 4.2, continued

Operating Point 1 (minimum phase)

Consider first the minimum-phase operating point. First, set all of the individual matrix elements
as follows:

g11 = tf([2.6],[62 1]);

g12 = tf([1.5],conv([62 1],[23 1]));
g21 = tf([1.4],conv([30 1],[90 1]));
g22 = tf([2.8],[90 1]);

Then define the transfer function matrix

G = [g11 g12;g21 g22]

Transfer function from input 1 to output...

2.6
#1: --------
62 s + 1

1.4
#2: --------------------
2700 s^2 + 120 s + 1

Transfer function from input 2 to output...

1.5
#1: -------------------
1426 s^2 + 85 s + 1
 2.8
#2: --------
90 s + 1

The command to evaluate transmission zeros is

z = tzero(G)

z =
-0.01733947114395
-0.05947212305895

and we see that both zeros are in the left-half-plane, so the model inverse will be stable.

Operating Point 2 (nonminimum phase)

Consider now the nonminimum-phase operating point. Set all of the individual matrix elements as
follows:

g11 = tf([1.5],[63 1]);

g12 = tf([2.5],conv([39 1],[63 1]));
g21 = tf([2.5],conv([56 1],[91 1]));
g22 = tf([1.6],[91 1]);

Then define the transfer function matrix

G2 = [g11 g12;g21 g22]

Transfer function from input 1 to output...

1.5
#1: --------
63 s + 1

2.5
#2: --------------------
5096 s^2 + 147 s + 1

Transfer function from input 2 to output...

2.5
#1: --------------------
2457 s^2 + 102 s + 1
 1.6
#2: --------
91 s + 1

and calculate the transmission zeros

z2 = tzero(G2)

z2 =
0.01300046557077
-0.05649863406894

The RHP (positive) zero indicates that the matrix inverse will be unstable.

SVD
Here we illustrate the use of svd to perform a SVD, by way of an example.

Example 14.3, continued

The M ATLAB svd function is straightforward to use. First, enter the scaled G matrix, then issue the
svd command, which yields the following results:

» G_scaled = [1.667 1.667;0.5 -0.5];

» [U,S,V] = svd(G_scaled)

U =
1.0000 0.0000
0.0000 -1.0000

S =
2.3575 0
0 0.7071

V =
0.7071 -0.7071
0.7071 0.7071

LTI Objects in SIMULINK Block Diagrams

Here we illustrate the use of the LTI feature of the Control Toolbox to perform multivariable
simulations. First, the process and controller objects are generated using the LTI tf (transfer
function) command. Then the block diagram is constructed and the LTI objects from the Control
Toolbox are placed in specific blocks on the diagram.

Example 14.1, continued

Here, we use Example 14.1, at the nonminimum phase (system A) operating point, for
illustration. First, define the process using the LTI objects.

% define the process

g11 = tf([1 3],[1 3 2]);
g12 = tf([2],[1 3 2]);
g21 = tf([3],[1 3 2])
g22 = tf([1],[1 3 2]);
G = [g11 g12;g21 g22];
%
% define the controller (L1 and L2 are the tuning parameters)
% set the L1 and L2 values before simulating the .mdl file
q11 = tf(-(2/3)*[0.5 1.5 1],[(1/3)*L1 ((1/3)+L1) 1]);
q12 = tf((4/3)*[0.5 1.5 1],conv([L2^2 2*L2 1],[(1/3) 1]));
q21 = tf(2*[0.5 1.5 1],[(1/3)*L1 (1/3)+L1 1]);
q22 = tf(-2*conv([0.5 1.5 1],[(1/3) 1]),conv([L2^2 2*L2 1],
[(1/3) 1]));
Q = [q11 q12;q21 q22];

Then, construct the .mdl diagram shown in Figure 14-15, where the multivariable (MV) IMC
controller block contains the Q object, the plant block contains the G object and the model block is
equal to the plant block for this (ideal) example.

Figure 14-15. MV IMC block diagram.

Set the tuning parameter values (L1 and L2) before running the simulation. For a setpoint change
in output 1, the responses are shown in Figure 14-14.

[ Team LiB ]
[ Team LiB ]

14.9 Summary
Transmission zeros are the multivariable equivalent of SISO transfer function zeros. RHP
(positive) transmission zeros indicate that the process inverse will be unstable. A transfer function
matrix may have no individual transfer functions with RHP zeros, and yet have a RHP matrix
transmission zero. Similarly, a transfer function matrix may have an element with a RHP zero,
and yet have no RHP matrix transmission zeros. Processes with RHP transmission zeros have
inherent closed-loop performance limitations; often there is not a significant advantage to full
multivariable control (using decoupling or MV IMC) of these processes.

Multivariable processes exhibit a directional sensitivity, where two different input vector directions
could result in outputs with a significantly different output magnitude. Processes that are ill-
conditioned (high condition number, or ratio of maximum to minimum singular values) will often
be too sensitive for full multivariable control.

Processes with RHP transmission zeros have internal closed-loop performance limitations; often
there is not a significant advantage to full multivariable control (using decoupling or MV IMC) of
these processes.

[ Team LiB ]
[ Team LiB ]

References

The quadruple tank problem (Example 14.2) is presented in the following paper: Johansson, K.
H., "The Quadruple-Tank Process. A Multivariable Laboratory Process with an Adjustable Zero,"
IEEE Trans. Cont. Syst. Tech., 8(3), 456–465 (2000).

Control performance limitations due to RHP transmission zeros is discussed in the following texts
and papers:

Holt, B. R., and M., Morari "Design of Resilient Processing Plants—VI: The Effect of Right-Half-
Plane Zeros on Dynamic Resilience," Chem. Eng. Sci., 40(1), 59–74 (1985).

Morari, M., and E. Zafiriou, Robust Process Control, Prentice Hall, Upper Saddle River, NJ (1989).

Skogestad, S., and I. Postlethwaite, Multivariable Feedback Control: Analysis and Design, Wiley,
New York (1996).

A more detailed discussion of the operating window approach is presented in the following: Marlin,
T. E., Process Control: Designing Processes and Control Systems for Dynamic Performance, 2nd
ed., McGraw Hill, New York (2000).

A measure of controllability, based on the operating window approach, is presented in the

following paper: Vinson, D. R., and C. Georgakis, "A New Measure of Process Output
Controllability," J. Process Control, 10, 185–194 (2000).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider Example 14.2, the quadruple tank problem.

a. Find the RGA for each operating point. Which variable pairing do you recommend
for each operating point?

b. Consider the nonminimum-phase (RHP transmission zeros) operating point

(operating point 2) in Example 14.2. Design independent single-loop controllers
for relatively fast closed-loop responses. (Hint: Approximate the transfer
functions as first order and use PI tuning rules.) Then close both loops
simultaneously using the independent SISO tuning. What do you find? How much
detuning is necessary to obtain good performance with both loops closed?

2: Find the inverse of the process transfer function matrix for system B in Example 14.1.

3: Consider the mixing tank presented in Example 14.3.

Write the dynamic material and energy balances for this system. The material balance
will yield the rate of change of height with respect to time, and the energy balance will
yield the rate of change of temperature with respect to time. Linearize the model at
the steady state to develop the input-output transfer function matrix.

4: Perform simulations on the mixing tank shown in Example 14.3.

a. First, work with the scaled transfer function matrix and make step input changes
in both the strongest and weakest directions. Compare the magnitudes of the
output responses.

b. Now, work with the dimensional variables. Convert the strongest and weakest
input directions to dimensional inputs and compare the effect of step changes on
the dimensional outputs.

5: Consider Example 14.1, controlled using the multivariable IMC procedure. Use the
SIMULINK diagram and LTI objects shown in Section 14.8.3 to perform simulations for
various tuning parameter (l1 and l2 values). Show that decreasing the tuning
parameter values below 0.333 does not significantly speed up the response but does
increase the manipulated input action.

6: Consider Example 14.1, controlled using the multivariable IMC procedure. Use the
SIMULINK diagram and LTI objects shown in Section 14.8. but revise these for the
minimum-phase (system B) operating point. Since there is not a RHP transmission
zero, no factorization of the process transfer function is necessary. Perform simulations
for various tuning parameter (l1 and l2) values. Show that decreasing the tuning
parameter values to approach 0 leads to almost instantaneous control, at the expense
of large manipulated variable action.
 7: Consider Example 14.2 (quadruple tank problem). The modeling equations are
presented in the appendix. Verify that the given steady-state values are correct for
each operating point. Also, linearize and find the state space model and verify the
transfer function matrix that was given for each operating point in Example 14.2.

8: Consider Example 14.2 (quadruple tank problem). Use MATLAB to calculate the
transmission zeros for both operating points.

9: Consider Example 14.2 (quadruple tank problem). Design a multivariable IMC

controller for the second (nonminimum phase) operating point. Use the standard
factorization that places the RHP transmission zero on the diagonal terms of the "bad
stuff" ( ) matrix. Implement the closed-loop controller in SIMULINK and discuss
the effect of the IMC filter factors on setpoint changes in each output (at different
times). Consider the allowable changes in the process outputs and manipulated inputs,
given the steady-state data shown in the appendix.

10: Consider Example 14.3 (mixing tank) with the transfer function matrix shown in
Equation (14.11). Implement the simplified decoupling strategy. Compare this with
static decoupling. Discuss the effect of controller tuning. For setpoint changes of a
reasonable magnitude, make certain that the manipulated input ranges are feasible.

11: Consider a process ("the shower problem") where a cold water stream at 15°C is
mixed with a hot water stream of 50°C. The outputs are the total mixed stream flow
rate and temperature. Assume steady-state flow rates of 0.19 and 0.13 liters/sec for
the cold and hot streams respectively. The steady-state outputs are then 0.32
liters/sec and 29°C. Scale the inputs and outputs to cover appropriate ranges, find the
scaled gain matrix and perform an SVD analysis. What is the condition number? What
are the most sensitive input and output directions (provide a physical interpretation of
this result)? For a rectangular input space where the cold water flow can range from 0
to 0.25 liters/second and the hot water flow can range from 0 to 0.19 liters/sec,
construct the output operating window. Given those input flow limits, is it possible to
operate over a mixed stream flow rate of 0.19 to 0.44 liters/sec, with a mixed stream
temperature between 23 and 34°C?

[ Team LiB ]
[ Team LiB ]

Appendix 14.1
The nonlinear modeling equations for the quad tank process are presented by Johansson (2000).

where

Ai = cross-sectional area of tank i;

ai = cross-sectional area of the outlet hole i;

hi = water height of tank i;

ki = pump coefficient i;
gi = flow splits;
g = gravitational constant

The parameter values are

A1,A3 = 28 cm 3
A2,A4 = 32 cm 3
a1,a3 = 0.071 cm2

a2,a4 = 0.057 cm2

b = 0.5 V/cm
g = 981 cm/s2

And the measured outputs are yi = bhi (for i = 1,2). Note that the maximum tank height is 20
cm.

The following states, inputs and coefficients correspond to the two operating points:

Operating point 1
h1s,h2s = 12.4, 12.7 cm

h3s,h4s = 1.8, 1.4 cm

v1s,v2s = 3.00, 3.00 cm

k1,k2 = 3.33, 3.35 cm3/Vs

g1,g2 = 0.7, 0.6

Operating point 2

h1s,h2s = 12.6, 13.0 cm

h3s,h4s = 4.8, 4.9 cm

v1s,v2s = 3.15, 3.15 cm

k1,k2 = 3.14, 3.29 cm3/Vs
g1,g2 = 0.43, 0.34

[ Team LiB ]
[ Team LiB ]

Chapter 15. Plantwide Control

A process engineer is concerned about more than the control of a single loop or multiple loops
associated with a single-unit operation. Of real concern is the operation of the entire plant. After
studying this chapter, the reader should be able to do the following:

Develop an appreciation for some of the complexities of plantwide control

Understand the control strategies associated with equipment/unit operations that are part of
a typical chemical process plant

Understand the "snowball effect" and how to fix a stream flow in the recycle loop to avoid it

Understand that recycle can significantly change the process dynamics

Understand the control and optimization hierarchy—that important operation and control
decisions at different corporate levels are made on different timescales

The major sections of this chapter are as follows:

15.1 Background

15.2 Steady-State and Dynamic Effects of Recycle

15.3 Unit Operations Not Previously Covered

15.4 The Control and Optimization Hierarchy

15.5 Further Plantwide Control Examples

15.6 Summary

[ Team LiB ]
[ Team LiB ]

15.1 Background
This textbook has emphasized controller design, primarily for SISO control loops. Much attention
has been focused on control algorithms (such as PID or IMC) and tuning [adjusting the PID
parameters, or the IMC "filter factor" (l)]. Some attention has also been paid to control-loop
interaction, particularly for two input–two output systems. Much of what a control or process
engineer does, however, is not related to specific control algorithm and tuning details. A process
engineer must be concerned about the operation of an entire process plant or operating unit, and
not just individual control loops or even unit operations. The issues addressed by a process
engineer are similar to those developed in Chapter 1, but on a much larger scale. At this point,
the plant operating and control objectives are really inseparable, so we repeat from Chapter 1 the
following topics on the development of a control strategy:

What are the major operational objectives of the plant?

What sensors should be paired with what manipulated inputs to form a control structure to
achieve the major objectives?

For any control valve or sensor failure, what steps will a process operator (and control
system) need to take?

Much of this type of discussion occurs during the design [or "retrofit" (redesign)] stage of a
process. Here, process engineers develop and review process flow sheets and process and
instrumentation diagrams (P&ID). Typically, a detailed operational description of the proposed
process and control strategy is developed, with consideration to all possible operating modes that
can be selected by an operator, as well as to all equipment failure modes.

In this chapter, we cannot do justice to the scope of work that is needed to develop a control
strategy and get an automation and control (including the plant!) strategy "up and running." We
attempt to show some of the major issues in "plantwide control" by using some illustrative
examples. In addition to showing examples of the many control loops required by even a simple
process, we provide an overview of the various levels (hierarchy) involved in a chemical
manufacturing organization.

We should note that the term plantwide control is somewhat of a misnomer. Process unit control
is really a more proper term, since an entire chemical process manufacturing plant is usually
composed of a large number of process units. Many of these process units contain reactors and
separators with recycle. In addition, there are recycle flows between many of the process units,
and the units draw from and contribute to many utility streams, such as steam and natural gas.
Although process unit control is a better term, we will continue the tradition of using plantwide
control as the term for regulation of a process unit.

[ Team LiB ]
[ Team LiB ]

15.2 Steady-State and Dynamic Effects of Recycle

A typical chemical process includes a chemical reactor to create products from reactants.
Generally, the larger the reactor the greater the conversion of reactants to products; the main
limit is imposed by equilibrium considerations. Rather than having a chemical reactor that is large
enough to approach equilibrium conversion, it is usually more economical to have a smaller
reactor, with a separator and a recycle stream to improve the overall conversion of reactants to
products. If secondary reactions occur, or inert components can "build up," it is common to
include a purge stream. Recycle systems are a source of challenging control problems, including
the so-called "snowball" effect (sensitivity of recycle flow rates to disturbances) and the
potentially long-time-scale dynamics, although the individual unit operations have reasonably
short time scales.

Steady-State Effects of Recycle

To illustrate recycle control system issues and challenges, we begin with the simple example
shown in Figure 15-1. The goal of this process is to convert the feed component A to product B.
The make-up feedstream (component A) mixes with the recycle stream (also primarily component
A) and is fed to the isothermal CSTR. A first-order reaction occurs; the reaction is incomplete, so
the reactor product stream (which is fed to the distillation column) contains both components A
and B. The distillation column separates this feedstream into two relatively pure product streams.
The distillate product, containing primarily A, is recycled; the bottoms product contains primarily
component B and is sent to storage for sale to a customer. Clearly, the bottoms product must
meet certain specifications; here we assume a composition specification.

Figure 15-1. Recycle flow control and instrumentation diagram.

Since you have been focused primarily on SISO control systems, the control system may seem
somewhat complex to you. The make-up feedstream is under flow control, and mixes with the
recycle stream; this fixes the flow to the isothermal CSTR. The CSTR level controller manipulates
the flow out of the CSTR to the distillation column. Now, consider the distillation column; see
Module 13 for a more detailed discussion of distillation column control. A standard two-product
column has a minimum of 3 controlled outputs (pressure, distillate receiver level, and bottoms
level) and 5 manipulated inputs (cooling water flow, distillate product flow, bottoms flow, reflux
flow, and steam-to-reboiler). In addition, the feed flow rate is an input that may either be a
disturbance or manipulated input, depending on the control strategy. For this particular flow sheet
we use the control variable pairing known as "conventional dual composition control," where the
overhead (distillate) and bottoms products are measured and controlled. Each of the measured
outputs has been paired with a manipulated input; there are no steady-state degrees of freedom.
That is, if the column feed flow rate and composition are known, all other flows and compositions
in the column can be calculated. This allows us to take a simplified view shown in Figure 15-2; it
should be clear that, if the overhead and bottoms product stream compositions are fixed, then the
overall material balance and the component balance on A can be used to solve for the distillate
(recycle) and bottoms product flowrates.

Figure 15-2. Simplified recycle flow control and instrumentation

diagram. Only the loops directly affecting the external flows from the
separator are shown.
Example 15.1 is used to illustrate an inherently steady-state problem (the "snowball effect") that
exists with the proposed control strategy. This is followed by Example 15.2, which illustrates how
recycle systems can appreciably change the dynamic characteristics of a process.

Example 15.1: A Simple Recycle Problem

Consider a single, first-order reaction

and an idealized flow sheet composed of a CSTR and a "perfect separator." An example is shown
in Figure 15-3; a basis of 100 moles/hour of the feedstream (stream 1) is used and the stream
component flow rates are shown directly on the flow sheet. A reactor volume of 200 moles and a
reaction rate constant of 0.6 hr-1 are assumed; this results in a 16% conversion of reactant A to
product B in the CSTR. In addition, the separator perfectly separates component A from B.
Because of the relatively low conversion, a 5:1 recycle/feed stream ratio is required.

Figure 15-3. Idealized flow sheet example with reaction, separation,

and recycle.
A surprising result is obtained if the fresh feed is increased from 100 moles/hour to 110
moles/hour, while maintaining the reactor volume at a constant level and assuming a perfect
separator. The new material balance is shown in Figure 15-4.

Figure 15-4. Idealized flow sheet example with a change in feed flow
rate of 10%.

Notice the tremendous change in the recycle flow rate. A 10% increase in feed flow rate has
resulted in more than doubling the recycle flow rate (from 500 to 1210 moles/hr)! Similarly
consider a 10% decrease in feed flow rate, which results in the stream flow rates shown in Figure
15-5. A 10 moles/time decrease (from the nominal amount of 100 from Figure 15-3) has led to a
decrease in the recycle stream flow rate from 500 to 230 moles/hr. You are encouraged to work
Exercise 1 to verify these results.

Figure 15-5. Idealized flow sheet example with a change in feed flow
rate of -10%.
This extreme sensitivity to small disturbances, illustrated by Example 15.1, has been termed the
"snowball effect" by Luyben (see Plantwide Process Control by Luyben et al., 1999). To minimize
this effect, Luyben et al. suggest that one of the flows within the recycle system be fixed. For
example, consider the strategy shown in Figure 15-6. Here, the flow rate of the mixed stream to
the CSTR is maintained at a constant value by manipulating the feedstream flow. The more
complete control instrumentation diagram is shown in Figure 15-7. Another alternative is to fix
the flow rate of stream 3 (between the CSTR and column) by using a flow controller on stream 3,
and to control the CSTR volume by manipulating the fresh make-up (stream 1) flow rate (see
Exercise 2).

Figure 15-6. Simplified recycle flow control and instrumentation

diagram, with the reactor feed flow rate (stream 2) fixed to minimize
the "snowball effect."

Figure 15-7. Recycle flow control and instrumentation diagram. The

reactor feed flow is fixed to minimize the "snowball effect."
In this section we have covered a relatively simple process with recycle, where there were 7
control loops. A number of more complex flow sheet examples are shown later, in Section 15.5.

Dynamic Effects of Recycle

It is known that recycle systems can exhibit dynamic behavior that is significantly different from
the equivalent process components without a recycle. The reason for this can be found by
analyzing the simple, positive feedback block diagram in Figure 15-8; here, gF(s) and gR(s)
represent the forward and recycle processes, respectively. Notice that the difference between this
diagram and a standard feedback-control diagram is that recycle processes exhibit positive
feedback. Recall that standard feedback control is also known as negative feedback.

Figure 15-8. Illustration of recycle as positive feedback.

The reader should show that the closed-loop transfer function relationship between an input (u)
and the output (y) is

Equation 15.1

It is easy to show that the closed-loop dynamics can be substantially longer than the individual
time constants. This is an interesting result because it indicates that individual unit operations in a
recycle system may be designed for relatively fast behavior, yet the dominant timescale of the
entire recycle system can be substantially longer. This is illustrated by the following example.

Example 15.2: Dynamic Effect of Recycle

To illustrate the positive feedback effect of recycle, we use the following first-order transfer
function to represent the forward process, where the time unit is minutes

Equation 15.2

We also neglect any dynamic behavior of the recycle process and assume that it can be
characterized as a constant gain (kR) between 0 and 1. The resulting input-output relationship is
[from Equations (15.1) and (15.2)]

Equation 15.3

Notice that, as the recycle gain approaches 1, the effective gain and time constant increase
dramatically. This effect is also shown in Figure 15-9, where a unit step input has been applied at
t = 0 min. The response of the forward path (kR = 0) is rapid with a magnitude of 1, while larger
recycle gains cause slower responses with larger magnitude. This is an interesting result, since a
recycle process with no dynamics has caused a tremendous increase in the time constant of the
overall (forward + recycle path) system. For kR = 0.9 with a forward time constant of 1 minute,
the overall dynamic time constant is 10 minutes. In Example 15.6 in Section 15.6 a physically
based recycle system with similar changes in time constant are shown.

Figure 15-9. Effect of recycle gain on response to a step input on a

first-order process.

[ Team LiB ]
[ Team LiB ]

15.3 Unit Operations Not Previously Covered

In this text, we have covered the control of a few unit operations but have not discussed control
structure determination in detail. In this section, we first discuss the two major structures for
maintaining material balance control, and then we discuss the control of some typical unit
operations that occur in a process flow sheet.

Supply-Side vs. Demand-Side Control of Production

Material balance control can be accomplished using two approaches: supply side (direction of
flow) and demand side (direction opposite to flow). A simple surge tank level example is shown in
Figure 15-10. The level controller manipulates the flow out of the vessel in supply-side control,
while the level controller manipulates the feed flow to the vessel in demand-side control.

Figure 15-10. Control in direction of flow (supply side) vs. direction

opposite to flow (demand side).

Using supply-side control, a change in the flow rate setpoint propagates downstream as a
disturbance to all units that follow. In demand-side control, a flow rate setpoint change
propagates upstream to all units that precede the unit. Luyben (1999) has studied demand-side
control structures for complex flow sheets and finds that they often add additional lags and create
control problems, particularly when distillation columns are involved. This causes greater product
variability, so the supply-side structure is generally recommended.

Compressor Control
It is more economical to condense a vapor stream and pump it than to vaporize a liquid stream
and compress it. Gas compressors are often used on recycle streams composed of light
components (hydrogen, methane, etc.). For these systems, it is not economical to condense the
streams and pump them as liquids.
Compressors that operate at a constant speed are usually controlled using one of the strategies
shown in Figure 15-11. Usually the "pinch" method will have lower operating costs than the
"spillback" method. Variable speed control methods are shown in Figure 15-12 for an electric
motor drive and for a steam turbine drive. The operating costs of variable speed compressors are
generally lower than for fixed speed systems, with perhaps a slightly higher initial capital cost.

Figure 15-11. Constant speed control systems.

Figure 15-12. Variable speed control systems.

When compressors are used on recycle streams, Douglas (1988) generally recommends that they
be operated "wide open"; that is, at a constant maximum flow rate. This is because the economic
return from the increased overall yield usually outweighs the incremental operating cost of the
compressor.

Heat Exchangers
Heat exchangers are used in a large number of temperature-control services, so here we focus on
the regulation of the exit temperature of a cold process stream that is exchanging heat with a hot
process stream. A common application would be a feed/effluent exchanger for a chemical reactor.
Three possible strategies are shown in Figure 15-13. The preferred method will often be that
shown in Figure 15-13c, since the dynamics will be much faster than for the other two strategies
shown. A change in the cold bypass flow rate has an almost immediate effect on the cold stream
outlet temperature.

Figure 15-13. Alternative temperature control strategies for a heat

 exchanger.

Adiabatic Plug Flow Reactors

It is important to control the inlet temperature of adiabatic plug-flow reactors, since there is no
method to add/remove energy from the reactor. For an exothermic reaction with a substantial
heat effect, one might consider using the heat from the reactor effluent to preheat the feed to the
reactor, as shown in Figure 15-14a. A problem with this strategy is that it is not very flexible; if
the bypass valve saturates (becomes fully closed, for example), there is a loss of temperature
control. Also, additional equipment will usually need to be added to start up the system. The
alternative shown in Figure 15-14b leads to a more flexible design. Particularly at start up, the
furnace allows the cold feed to be heated to a high enough temperature to initiate the reaction.
The sacrifice is higher energy consumption; the relative increase depends on the steady-state
ratio of the feed/effluent exchanger and furnace duties.

Figure 15-14. Temperature control for an adiabatic plug flow reactor.

[ Team LiB ]
[ Team LiB ]

15.4 The Control and Optimization Hierarchy

Operating Levels
The actual operation of a chemical manufacturing company involves decisions and actions that
occur at a number of levels and timescales, as shown in Figure 15-15. At the corporate level,
allocation decisions are made. Market demand projections, raw material availability, and past
operating costs are used to set long-term plans for corporate-wide optimization. These decisions
are typically made on an infrequent basis—weekly, monthly, or quarterly. Often these decisions
are made using linear programming (LP) techniques. This information is passed to the plant level.

Figure 15-15. Corporate operating levels.

At the next level are plant decisions about allocation among various plant operating units. These
may be made daily, or even more frequently, depending on the plant automation and control
system. For example, steam may be generated by a number of different sources, with different
costs, depending on relative energy and product values. In an oil refinery, different units make
similar components for blending into gasoline (which is produced in many different grades), and
these relative amounts may change depending on energy and product values, operating problems
with certain units, and so forth.

The operating unit level takes information available on a daily basis and over the course of the
day makes hour-by-hour changes, on average, to meet those goals. The unit operation level
involves individual equipment, and desired setpoints may be changed on an hourly (or more
frequent) basis. The regulation on this level occurs on the order of minutes.

The fundamental control level consists primarily of process flow or pressure controllers. The
lowest level generally involves the manipulation of a control valve, or the reading of a sensor,
where the time frame is on the order of 1 second or less. Notice that there is a feedback of
information from the lower levels back up to the higher levels. The basic idea of feedback control
occurs at the unit operation and fundamental control levels. Here, control algorithms are used to
adjust manipulated inputs (controller outputs or control variables) to maintain process outputs
(process variables) at desired values (setpoints).

Petroleum Refining Example

An example of the overall operating problem is illustrated in Figure 15-16, for a typical petroleum
refining company. At the highest level, corporate management decides where to purchase crude
and how to distribute the crude oil to the various refineries in the corporation. At the next level
(level 2), refinery A takes the current and future crude delivery projections and gasoline
production projections and determines the operating conditions for each process unit (level 3,
e.g., the catalytic cracking unit) in the refinery. Setpoints for the unit operations (level 4, e.g.,
distillation) are determined at this point. The unit operations level determines the process flow
rates, such as the distillate or reflux flow rates (level 5). These controllers then determine, for
example, the pressure to the control valves to regulate various flow rates.

Figure 15-16. Corporate operating levels: petroleum refining example.

Discussion
For many plants, much of the feedback provided between different levels is performed manually.
During the past decade, major corporations have moved toward automating more of these
functions, and better integrating the decisions that are made at different levels. The broad term
for this, at the plant level, is computer-integrated manufacturing (CIM). Developments in
information technology and database management are making this task easier. Corporate-wide
technology is known as supply chain management or enterprise optimization. This software links
databases from various suppliers (of raw materials) with corporate scheduling, and with customer
orders (of products).

[ Team LiB ]
[ Team LiB ]

15.5 Further Plantwide Control Examples

Examples and challenges of control of processes with recycle were presented in Section 15.2. In
this section we give the reader more experience with plantwide control strategies by studying two
more examples: MPN and HDA.

Example 15.3: MPN Process and Instrumentation Diagram

Consider the process flow sheet shown in Figure 15-17. Here, methanol is reacted with
acrylonitrile to form MPN. A slight excess (2%) of ACN is supplied and essentially all of the MeOH
is reacted. There is no purge stream, so the excess ACN appears in the MPN product stream. Here
the recycle stream is liquid-phase distillate stream.

Figure 15-17. Preliminary instrumentation diagram for an MPN process

(adapted from Zhan and Grassi, 1999).

The objective is to produce MPN at a specified rate, while satisfying product purity constraints. In
Figure 15-17, there are nine control valves, representing nine degrees of freedom. Also shown are
13 measurements: seven flow, three level, two temperature, and one pressure. Clearly, there are
not enough degrees of freedom to specify a setpoint value for every measured variable. Inventory
variables, such as level and pressure, must be controlled; without feedback, these integrating
variables could easily violate alarm limits. This means that four control valves must be used to
control the three levels and the column pressure.

We consider it an important exercise for the reader to attempt to place nine control loops on
Figure 15-17. Please do this before reading any further.

A Possible Solution to Example 15.3

One possible solution is shown in Figure 15-18.

1. The production rate objective should be satisfied by the methanol flow rate controller, since
there is essentially total conversion of the methanol to MPN.

2. The excess ACN is obtained by specifying the ACN flow rate to be 2% higher than the
methanol flow rate.

3. Exothermic reactor temperatures must be controlled to assure stable reactor operation.

Here, the reactor temperature is controlled by manipulating the cooling water valve.

4. The reactor level (related to residence time) is controlled by manipulating the reactor outlet
flow valve. Notice that there are now only five more loops that can be specified, since there
are a total of nine control valves. The final five loops are associated with the distillation
column; three of these are associated with inventory variables (distillate receiver level,
bottoms level, column pressure).

5. A temperature in the bottom of the column manipulates the steam control valve. Because of
equilibrium relationships, the temperature is related to the product purity.

6. The reflux stream is simply flow controlled. An alternative would be to control a tray
temperature in the top of the column; this would be related to controlling the distillate
purity. Since the distillate stream is recycled to the reactor, it is not considered important to
maintain the stream composition at a given value.

7. The distillate receiver level is controlled by manipulating the distillate (or recycle) flow valve.

8. The column pressure is controlled by manipulating the condensate flow. This essentially
changes the level of condensate in the overhead condenser, changing the effective heat
transfer area, and therefore the condenser duty.

9. The column bottom level is maintained by manipulating the bottom control valve.

Figure 15-18. A typical P&ID for the MPN process.

Problems with This Solution

The main problem with this solution is that the control structure exhibits a high sensitivity in the
recycle flow rate to small disturbances. This so-called snowball effect was discussed in Section
15.2. Luyben et al. (1999) have recommended that at least one control loop in the recycle flow
path be flow controlled in order to minimize the snowball effect.

Alternative Solution 1

An alternative solution [designated plantwide control structure 1 by Zhan and Grassi (1999)] is
shown in Figure 15-19. Here, a flow controller is used to regulate the mixed stream (recycle +
ACN) flow rate by manipulating the distillate (recycle) flow rate; this allows the control structure
to minimize the snowball effect. The distillate level controller manipulates the fresh ACN flow
control valve. The other seven loops are identical to the previous solution shown in Figure 15-18.

Figure 15-19. Plantwide control structure 1 developed by Zhan and

Grassi (1999) for the MPN process.
Other Alternative Solutions

There are some alternative solutions that include flow control of a recycle loop stream. For
example (alternative 2), the reactor outlet stream can be flow controlled. The reactor level can
then be controlled by manipulating the ACN feed-flow valve (see Exercise 3). In addition, a
reactor concentration sensor can be used to improve control (again, see Exercise 3).

Example 15.4: HDA Process

Toluene hydrodealkylation (HDA) is one of the primary processes for the production of benzene. A
simplified process and instrumentation diagram for an HDA process is shown in Figure 15-20. For
clarity, we have omitted a number of details, such as pumps on distillation reflux streams and
furnace combustion air dampers.

Figure 15-20. Preliminary P&ID for an HDA process.

The feed streams are high purity (99.98 mol%) toluene and hydrogen (96 mol% hydrogen, 4
mol% methane). The process operating objective is to produce a desired rate of benzene at a
purity of 99.98 mol%. Because of coking considerations, a 5:1 hydrogen/aromatics ratio must be
maintained at the reactor entrance. The reactor inlet pressure is to be maintained at just under
500 psig (the pressure of the hydrogen feed stream); the pressure drop between the feed stream
mixing point and the flash drum is roughly 35 psi. The minimum reactor inlet temperature is
1150°F, while the maximum outlet temperature is 1300°F. The reactor exit stream must be
immediately quenched to 1150°F to minimize secondary reactions.

The primary reaction is an irreversible reaction of toluene and hydrogen to produce benzene and
methane, while the secondary reaction is the reversible reaction of benzene to form diphenyl and
hydrogen

Equation 15.4

The selectivity decreases as the temperature increases. It is clear that a purge stream is
necessary to eliminate methane from the process system.

We consider it an important exercise for the student to attempt to place control loops on Figure
15-20. Please do this before analyzing the suggested solution shown in Figure 15-21. Use the
following suggestions to guide your control structure development.

Luyben et al. (1999) recommend placing a flow control loop on a stream in the recycle path.
For this process, they suggest placing the recycle toluene stream on flow control.
 Douglas (1988) recommends operating recycle compressors "wide open" because the value
of recovered components is usually higher than the additional compressor power cost.

A tray temperature in each column can be used to "infer" a product composition. Dual
composition control for the columns is not necessary, and this petrochemical plant has had
good success by placing reflux streams under flow control.

Figure 15-21. Plantwide control structure for an HDA process, as

suggested by Luyben et al. (1999).

Notice that there are 23 control degrees of freedom, since there are 23 control valves. It is not
necessary that all these valves be used for feedback control; one or more valves may be held at a
constant valve position.

A Possible Solution to Example 15.4

One possible solution to the HDA problem is shown in Figure 15-21. This solution was developed
by Luyben et al. (1999). The important control structure decisions to note are as follows:

The total toluene flow (recycle + feed) is set to prevent the snowball effect (recall that it is
recommended that one loop in a recycle system be placed on flow control). Notice that this
flow also sets the production rate of benzene.

The makeup (feed) toluene flow is manipulated by the recycle column distillate receiver level
controller. Your intuition may suggest that there would be a significant lag between the feed
flow, through the various unit operations, to the final distillation column. This is not the case,
however, since the total toluene flow is regulated. Any change in toluene feed has an
 immediate effect on the recycle toluene flow and, hence, the distillate receiver level.

The overhead valve from the flash drum is set wide open to allow maximum flow through
the recycle compressor. The flash drum operating pressure is then dictated by pressure drop
through the system, since the pressure is controlled at the reactor entrance.

Since the diphenyl flow rate from the recycle column is so low, the level is controlled by the
reboiler heat duty (steam to the reboiler).

The benzene purity is maintained using a cascade strategy, where the output of the
composition controller is the setpoint to a tray temperature controller. This strategy yields a
faster rejection of feed disturbances and provides satisfactory composition control when the
composition sensor fails or needs recalibration.

It should also be noted that only the basic control strategy is presented. In practice, ratio, feed
forward, and cascade control will be added for improved control. Also, the control techniques
presented in Section 15.3 can be used.

The Art of Process Engineering

At this point, you may be overwhelmed by the complexity of a plantwide control problem. Please
realize that placing controllers on a process flow sheet is an "art," and different process engineers
will often develop different control structures for the same process flow sheet. In practice, you will
not develop these control strategies in isolation. At the process design (or retrofit) stage, a large
number of engineers will be involved in the development of a control structure; this is truly a
team effort, involving participants from the plant, the engineering design firm, and other
contractors involved in automation. An operational description of the process automation system
will evolve over many months of meetings. These discussions will lead to numerous iterations of
the control structure. Experience during and after plant or control-system startup will normally
lead to changes in this structure.

Although computer simulation packages are widely available, a process engineer often gains
insight about operating and design decisions using simple "back of the envelope" calculations, as
shown in the next example.

Example 15.5: HDA "Back of the Envelope" Material Balance

Much insight about the effect of design decisions can be obtained by rough "back of the envelope"
calculations. Here we consider the HDA flow sheet shown in Figure 15-22. The following
assumptions are made to simplify the analysis: (1) only the primary reaction (toluene + hydrogen
= benzene + methane) occurs, (2) a single purge stream of hydrogen and methane, (3) a pure
benzene product stream, (4) pure toluene product and recycle streams.

Figure 15-22. "Back of the envelope" material balance for the

simplified HDA problem.
A basis of 1 mole of toluene make-up (stream 1) is used. Here we use ni to represent a molar
flow rate, where the subscript indicates the stream number. The reader should derive the
following relationships (see Student Exercise 4).

Toluene recycle (stream 10) flow: n10 = (1-XT)/XT, where XT is the single-pass conversion of
toluene in the reactor.

Hydrogen make-up (stream 2) flow: n2 = 1/(0.96-y7,H ), where y7,H is the mole fraction of H2 in
the purge stream (stream 7).

Hydrogen recycle (stream 8) flow:

Equation 15.5

Two of the major operating costs are the hydrogen make-up (raw material) and vapor recycle
(stream 8, compressor costs). We can now quickly calculate the hydrogen make-up flow and
vapor recycle as a function of the toluene conversion and hydrogen purge fraction. For example,
consider a single pass toluene conversion of 0.7. The following table (Table 15-1) is quickly
constructed; additional rows can be generated by considering other toluene conversions (see
Student Exercise 4). Notice that there is a direct trade-off of hydrogen raw material costs with
recycle compressor operating costs. Douglas (1988) finds that the incremental raw material cost
is higher than the incremental recycle compression costs for these types of processes.

Consideration of the secondary reaction in Equation (15.4) results in the simplified flow sheet
shown in Exercise 5.

[ Team LiB ]
[ Team LiB ]

15.6 Simulations
SIMULINK provides a nice modeling environment for simulating very simple process flow sheets. For
rigorous process design and analysis, chemical process flow sheeting packages such as ASPEN
PLUS and HYSYS should be used. Flow sheets can quickly be developed using standard "icons"
that represent unit operations, such as reactors and separators. This next example illustrates the
use of SIMULINK to solve a flow sheet similar to Example 15.1.

Table 15-1. H2 Make-up and Vapor Recycle as a Function of H2 Fraction

in Purge Stream

H2 purge frac, y7,H Toluene conv., XT H2 make-up, n2 Recycle, n8

0.3 0.7 1.515 18.96

0.4 0.7 1.786 13.57
0.5 0.7 2.174 10.11

0.6 0.7 2.778 7.46

Example 15.6: A Simple Recycle System

Here we consider a slightly different steady state than that shown in Figure 15-3. In Figure 15-23
the make-up feedstream (stream 1) has 0.9 mole fraction A. The CSTR has a molar holdup of N =
200 moles and a reaction rate constant of k = 0.6 hr-1. One of our objectives is to compare the
responses of the flow sheets with and without a recycle stream, to disturbances in the fresh feed
(stream 1) composition. A SIMULINK .mdl for the recycle system dynamics is shown in Figure 15-
24, where the mixing point, the CSTR, and the separator are "masked" functions. The CSTR is
"unmasked" in Figure 15-25, to reveal the modeling equation

Equation 15.6

Figure 15-23. Flow sheet example. Steady-state flows shown.

Figure 15-24. SIMULINK diagram (.mdl) for the dynamic flow sheet.
 Figure 15-25. Unmasked view of the CSTR block in Figure 15-24.

Consider a step change in the feed composition of A from 0.9 to 1.0 mole fraction. The results for
the recycle system are compared with the "straight through" system without recycle in Figure 15-
26. Notice that the recycle system has a tremendous change in the separator product stream
(recycle flow) compared to the no-recycle case. As discussed in Section 15.2 (and Example 15.2),
the recycle system has a large gain and time constant compared to the system without recycle.

Figure 15-26. Response to a step disturbance in the feedstream

composition of A (from 0.9 to 1.0) at t = 0 hr. Comparison of recycle
with no-recycle systems. Top plot: distillate product flow rate (n5).
Bottom plot: flow rate of product B (n4).

[ Team LiB ]
[ Team LiB ]

15.7 Summary
This chapter has provided a brief overview of perhaps the most interesting and challenging
aspects of process control. The intuitive placement of controllers on a process with recycle can
often lead to extreme sensitivity to disturbances known as the snowball effect. This effect can be
minimized by fixing the flow rate of one stream in the recycle loop.

Recycle can also significantly increase the dynamic timescale of the process. The specification of a
plantwide control structure is based on years of operational experience with similar plants and will
involve input from a large number of engineers and technicians. If you are interested in a much
more detailed presentation of plantwide control, including many "rules of thumb" and other
suggestions, the book by Luyben et al. (1999) is highly recommended.

Corporations have a large number of levels where decisions need to be made (Section 15.4). The
various levels range from the long-term corporate purchasing and scheduling decisions at the top,
to fundamental control instrumentation at the bottom, as shown in Figures 15-15 and 15-16.

[ Team LiB ]
[ Team LiB ]

References

The MPN process is presented and discussed by in the following: Zhan, Q., and V. G. Grassi,
"Dynamic Matrix Control of a Specialty Chemical Plant with Different Plantwide Control
Structures," in Proceedings of the 1999 American Control Conference, pp. 632–636 (1999).

The best reference for plantwide control is the following monograph: Luyben, W. L., B. D. Tyréus,
and M. L. Luyben, Plantwide Process Control, McGraw Hill, New York (1999).

Plantwide strategies are also discussed in the following papers and textbook:

Ponton, J. W., and D. M. Laing, "A Hierarchical Approach to the Design of Process Control
Systems, Trans. Inst. Chem. Eng., 71(A), 181–188 (1993).

Stephanopoulos, G., Chemical Process Control, Prentice Hall, Englewood Cliffs, NJ (1984).

Problems with "on-demand" control structures are discussed in the following: W. L. Luyben,
"Inherent Dynamic Problems with On-Demand Control Structures," Ind. Eng. Chem. Res., 38,
2315–2329 (1999).

An overview of process systems engineering is provided in the following encyclopedia article:

Bequette, B. W., and L. P. Russo, "Process Design, Simulation, Optimization and Operation," in
the Encyclopedia of Physical Science and Technology, 3rd ed. Vol. 2, pp. 751–766. (R. Matsubara,
ed.), Academic, New York (2002).

The following book provides a nice treatment of the synthesis of chemical process flow sheets to
meet design objectives. Once the process is designed, Douglas recommends that recycle streams
be operated at a constant maximum flow rate, since the resulting yield improvements are worth
the additional energy costs.

Douglas, J. M. Conceptual Design of Chemical Processes, McGraw Hill, New York (1988).

The Tennessee Eastman "challenge problem" (Exercise 6) was presented in the following paper:
Downs, J. J., and E. F. Vogel, "A Plant-Wide Industrial Process Control Problem," Comp. Chem.
Engng., 17, 245–255 (1995).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider Example 15.1. Assume that the isothermal CSTR has perfect level control; that
is, the volume is constant. The CSTR contains 200 moles, and the reaction rate constant
is 0.6 hr -1. Verify the results shown in Figures 15-3, 15-4 and 15-5.

2: Consider Example 15.1. Revise Figure 15-7 to obtain the following alternative control
strategy. Draw the control instrumentation diagram obtained by fixing the flow rate of
stream 3 (between the CSTR and column) by using a flow controller on stream 3, and to
control the CSTR volume by manipulating the fresh make-up (stream 1) flow rate.

3: Consider the MPN example. Alternative 2 is shown here. Now, consider the use of a
reactor concentration sensor. Let the reactor concentration controller manipulate the
methanol feed flow rate. Sketch this control strategy on the diagram. Which controller
setpoint is set to yield a desired production rate?

4: Consider the simplified flow sheet shown in Figure 15-22 (Example 15.5).

a. From material balances find the relationships for all flows as a function of hydrogen
fraction in the purge stream and the single-pass toluene conversion in the reactor.

b. Construct a table similar to Table 15-1.

c. Based solely on the current raw material costs (neglecting capital and energy
 b.

c.
costs) of hydrogen and toluene, can a profit be obtained with a hydrogen purge
concentration of 0.5 mole fraction?

5: Consider the "back of the envelope" calculations for the HDA process (Example 15.5). In
our first pass at material balances, we neglected the diphenyl formation. Now, consider
the selectivity of conversion to toluene and include diphenyl material balances. The
additional separator is shown in the flow sheet here.

a. Find how the raw material, recycle, and benzene product flows change as a
function of toluene conversion, selectivity (fraction of toluene converted to
benzene vs. overall conversion of toluene) and hydrogen purge fraction. Construct
a table similar to Table 15-1.

b. Based solely on the current raw material costs (neglecting capital and energy
costs) of hydrogen and toluene, can a profit be obtained with a hydrogen purge
concentration of 0.5 mole fraction and a selectivity of 0.95?

6: A Tennessee Eastman "challenge problem" was published by Downs and Vogel in 1993
(see the references section of this chapter). A fairly large number of academic studies
have been performed on plantwide control problems. Conduct a literature review on this
topic, read at least two articles, and compare the different control techniques proposed.

7: Most process plants have three steam header pressure levels, approximately 15 psig,
150 psig, and 600 psig. Discuss possible answers to the following:

What determines the pressure level needed for a particular unit operation?

How would you control the pressure of each of the steam headers?

How would you determine the cost of various levels of steam? If some low
pressure steam were being released to the atmosphere to maintain the steam
balance (and the low pressure header at 15 psig), what should be the incremental
cost of this steam to give an economic reward using it for heating needs?
[ Team LiB ]
[ Team LiB ]

Chapter 16. Model Predictive Control

MPC is the class of advanced control techniques most widely applied in the process industries. A
primary advantage to the approach is the explicit handling of constraints. In addition, the
formulation for multivariable systems with time delays is straightforward.

After studying this chapter, the reader should understand the basic ideas behind SISO MPC,
including the following:

Dynamic matrix control (DMC)

Tuning parameters such as prediction and control horizons and control weight

The effect tuning parameters have on the control performance and stability

The ability to explicitly include constraints and handle multivariable processes

The major sections of this chapter are as follows:

16.1 Motivation

16.2 Optimization Problem

16.3 Dynamic Matrix Control

16.4 Constraints and Multivariable Systems

16.5 Other MPC Methods

16.6 MATLAB

16.7 Summary

[ Team LiB ]
[ Team LiB ]

16.1 Motivation
Most of the techniques presented in this textbook have been based on continuous-time models.
The controllers are normally implemented by microprocessors and, therefore, must be converted
to discrete time before implementation. We have focused on continuous time because few
changes in the techniques are needed for the discrete-time case. One advanced control technique
that has solely had discrete-time application is model predictive control (MPC). MPC is by far the
most commonly applied advanced control technique in the chemical process industry, so it is
worth devoting a chapter to it. One major contributor to the success of MPC is the ability to
handle constraints in an optimal fashion. The optimization-based procedure is intuitive and is also
a natural way of handling multivariable systems.

Basic Description
The basic idea behind MPC is shown in Figure 16-1. At each time step, k, an optimization problem
is solved, as illustrated on the top half of the figure. An objective function (usually quadratic)
based on output predictions over a prediction horizon of P time steps is minimized by a selection
of manipulated variable moves over a control horizon of M control moves. Although M moves are
optimized, only the first move is implemented. After uk is implemented, the measurement at the
next time step, yk+1 is obtained, as shown in the bottom half of the figure. A correction for model
error is performed, since the measured output yk+1 will, in general, not be equal to the model
predicted value. A new optimization problem is then solved, again, over a prediction horizon of P
steps by adjusting M control moves. This approach is also known as receding horizon control.

Figure 16-1. Basic concept of MPC.

For clarity, we initially focus on unconstrained SISO systems. Several questions naturally arise.
What type of objective function should be used for optimization? What type of model should be
used to predict the output? How should the model be initialized to predict the future output
values? What is the desired setpoint trajectory? How should constraints be implemented? Some of
these topics are discussed in the next section.

[ Team LiB ]
[ Team LiB ]

16.2 Optimization Problem

The term optimization implies a best value for some type of performance criterion. This
performance criterion is known as an objective function. Here, we first discuss possible objective
functions, then possible process models that can be used for MPC.

Objective Functions
Here, there are several different choices for objective functions. The first one that comes to mind
is a standard least-squares or "quadratic" objective function. The objective function is a "sum of
squares" of the predicted errors (differences between the setpoints and the model-predicted
outputs) and the control moves (changes in control action from step to step).

A quadratic objective function for a prediction horizon of 3 and a control horizon of 2 can be
written

where represents the model predicted output, r is the setpoint, Du is the change in manipulated
input from one sample time to the next, w is a weight for the changes in the manipulated input,
and the subscripts indicate the sample time (k is the current sample time). For a prediction
horizon of P and a control horizon of M, the least-squares objective function is written

Equation 16.1

Another possible objective function is to simply take a sum of the absolute values of the predicted
errors and control moves.

For a prediction horizon of 3 and a control horizon of 2, the absolute value objective function is

which has the following general form for a prediction horizon of P and a control horizon of M:

Equation 16.2
The optimization problem solved is usually stated as a minimization of the objective function,
obtained by adjusting the M control moves, subject to modeling equations (equality constraints),
and constraints on the inputs and outputs.

Equation 16.3

Least-squares formulations are by far the most common objective functions in MPC. Least
squares yields analytical solutions for unconstrained problems and penalizes larger errors
(relatively) more than smaller errors. The absolute value objective function has been used in a
few algorithms because a linear programming (LP) problem results. LPs are routinely solved in
large-scale scheduling and allocation problems. For example, an oil company often uses an LP to
decide how to distribute oil to various refineries and to decide how much and what product to
produce at each plant. This topic was discussed in Chapter 15. The LP approach is not useful for
model predictive control, because the manipulated variable moves often "hop" from one extreme
constraint to another. For the rest of this chapter we use a quadratic (least squares) formulation
for the objective function.

Models
Many different types of models are possible for calculating the predicted values of the process
outputs, which are used in evaluating the objective function. Since the outputs are evaluated at
discrete-time steps, it makes sense to use discrete models for the output prediction. Here, we
review step and impulse response models (introduced in Chapter 4), both of which are used in
common MPC algorithms.

Finite Step Response

FSR models are obtained by making a unit step input change to a process operating at steady
state. The model coefficients are simply the output values at each time step, as shown in Figure
16-2. Here, si represents the step response coefficient for the ith sample time after the unit step
input change. If a non-unit step change is made, the output is scaled accordingly.

Figure 16-2. Illustration of step response parameter identification.

The step response model is the vector of step response coefficients,

Equation 16.4

where the model length N is long enough so that the coefficient values are relatively constant
(i.e., the process is close to a new steady state).

Finite Impulse Response

Another common form of model is a finite impulse response (FIR). Here, a unit pulse is applied to
the manipulated input, and the model coefficients are simply the values of the outputs at each
time step after the pulse input is applied. As shown in Figure 16-3, h i represents the ith impulse
response coefficient.

Figure 16-3. Illustration of impulse response parameter identification.

There is a direct relationship between step and impulse response models:

Equation 16.5

Figure 16-4 illustrates how impulse response coefficients can be obtained from step responses.
The impulse response coefficients are simply the changes in the step response coefficient at each
time step. Similarly, step response coefficients can be found from impulse responses; a step
response coefficient is the sum of the impulse response coefficients to that point. It should be
noted that there are two major limitations to step and impulse response models. They can only be
used to represent open-loop stable processes, and they require a large number of parameters
(model coefficients) compared to state space and transfer function models.

Figure 16-4. Obtaining impulse response models from step response

data.
[ Team LiB ]
[ Team LiB ]

16.3 Dynamic Matrix Control

DMC was developed by Shell Oil Company in the 1960s and 1970s (see Cutler and Ramaker,
1980). It is based on a step response model, which has the form (see Appendix 16.1 for a
derivation)

Equation 16.6

which is written in the form

Equation 16.7

where k is the model prediction at time step k, and uk – N is the manipulated input N steps in
the past.

Note that the model-predicted output is unlikely to be equal to the actual measured output at
time step k. The difference between the measured output (yk) and the model prediction is called
the additive disturbance.

Equation 16.8

The "corrected prediction" is then equal to the actual measured output at step k,

Equation 16.9

Similarly, the corrected predicted output at the first time step in the future can be found from

Equation 16.10
So, for the jth step into the future, we find

Equation 16.11

and we can separate the effects of past and future control moves

Equation 16.12

The most common assumption is that the correction term is constant in the future (this is the
"constant additive disturbance assumption"):

Equation 16.13

Also, realize that there are no control moves beyond the control horizon of M steps, so

In matrix-vector form, a prediction horizon of P steps and a control horizon of M steps, yields

Equation 16.14
which we write using matrix-vector notation

Equation 16.15

In Equation (16.15) the corrected-predicted output response is naturally composed of a "forced

response" (contributions of the current and future control moves) and a "free response" (the
output changes that are predicted if there are no future control moves). The difference between
the setpoint trajectory, r, and the future predictions is

Equation 16.16

which can be written

Equation 16.17

where the future predicted errors are composed of "free response" (E) and "forced response"
(–Sf Duf ) contributions.

The least-squares objective function (16.1) is

Equation 16.18

Notice that the quadratic terms can written in matrix-vector form as

Equation 16.19

and

Equation 16.20

Therefore the objective function can be written in the form

Equation 16.21
subject to the modeling equation equality constraint (16.17)

Equation 16.17

Substituting (16.17) into (16.21), the objective function can be written

Equation 16.22

The solution for the minimization of this objective function is (see Appendix 16.2)

Equation 16.23

Notice that the current and future control move vector (Duf) is proportional to the unforced error
vector (E). That is, a controller gain matrix, K, multiplies the unforced error vector (the future
errors that would occur if there were no control move changes implemented).

Because only the current control move is actually implemented, we use the first row of the K
matrix, and

Equation 16.24

where K1 represents the first row of the K matrix, where

Perhaps it is worth summarizing the steps involved in implementing DMC on a process.

1. Develop a discrete step response model with length N (16.4), based on a sample time Dt.

2. Specify the prediction (P) and control (M) horizons. N P M

3. Specify the weighting on the control action (w = 0 if no weighting).

4. All calculations assume deviation variable form, so remember to convert to/from physical
units.
 4.

The effect of all of these tuning parameters is now discussed for SISO systems.

Model-length and sample-time selection are not independent. The model length should be
approximately the "settling time" of the process, that is, the time required to reach a new steady
state after a step input change. For most systems, the model length will be roughly 50
coefficients. The sample time is usually on the order of one tenth the dominant time constant, so
the model length is roughly the settling time of the process.

Prediction and control horizons differ in length. Usually, the prediction horizon is selected to be
much longer than the control horizon. This is particularly true if the control weighting factor is
selected to be zero. Usually, if the prediction horizon is much longer than the control horizon, the
control system is less sensitive to model error. Often P = 20 or so, while M = 1–3.

Control weighting is often set to zero if the prediction horizon is much longer than the control
horizon. As the control horizon is increased, the control moves tend to become more aggressive
so a larger weight is needed to penalize the control moves.

Example 16.1: First-Order Process

Here we study the first-order process, where the time unit is minutes,

Ordinarily, we would select a sample time of roughly 0.5 minutes, and a model horizon of 50.
Here we select a sample time of 1 minute to study the effect of other tuning parameters. The step
response coefficients generated from this plant are shown in Figure 16-5. We can see that a
model length of at least 20 should be used, since that is when the output is reasonably close to
steady state.

Figure 16-5. Step response coefficients for the first-order example.

Sample time = 1 minute; model length, N = 25.
The importance of prediction horizon is shown clearly in Figure 16-6. In both simulations a control
horizon of 1 is used. A prediction horizon of 1 results in the setpoint being achieved in 1 time step,
while a prediction horizon of 5 yields a much slower response. Notice, however, that P = 1
requires much more control action than P = 5. Although not shown here, the shorter prediction
horizon is more sensitive to model uncertainty; these simulations have assumed a perfect model.

Figure 16-6. Closed-loop responses to a step setpoint change at t = 1

min, first-order example. Sample time = 1 minute; model length, N =
25; control horizon, M = 1; no weighting on input. Comparison of
prediction horizons, P = 5 and P = 1.
The importance of the model length is shown in Figure 16-7. Since a model length of 6 does not
capture the complete dynamics of the process (see Figure 16-5), it effectively results in a model
error and poor performance compared with a model length of 25.

Figure 16-7. Closed-loop responses to a step setpoint change at t = 1

min, first-order example. Sample time = 1 minute; prediction horizon,
P = 5; control horizon, M = 1; no weighting on input. Comparison of
model lengths, N = 25 and N = 6.
From Equation (16.14), for N = 6, P = 5, M = 1, we find (see Figure 16-5 for the step response
coefficients) the following dynamic (Sf ) and past move (Spast) matrices.

also, selecting a control move weight of 0

we calculate the feedback gain matrix from Equation (16.23)

and K1 = K since the control horizon (M) is 1 (the reader should verify these numerical results
using MATLAB), the feedback gain matrix is:

in the simulations shown, the process is initially at steady state when the unit step setpoint
change at t = 1 minute is made. Since deviation variables are used, the values of all previous
manipulated inputs and outputs are 0. The unforced error vector, E, from Equation (16.16) is (we
have transposed it to a row vector to save space)

so the first control move, after the setpoint change is made is

which is consistent with the first control move shown in Figure 16-7.

Example 16.2: Van de Vusse Reactor

Consider the van de Vusse reactor problem covered in Module 5. The continuous state space
model is

where the measured state (output) is the concentration of the second component and the
manipulated input is the dilution rate. The manipulated input-output process transfer function for
the reactor is

Step Response Model

The discrete step response model is shown in Figure 16-8, for a sample time of 0.1 minute. Notice
that a model length of at least 35 is needed to capture the complete dynamic behavior.

Figure 16-8. Step response coefficients for the van de Vusse example.
Sample time = 0.1 minutes, model length, N = 50.
Effect of Prediction Horizon

A sample time of 0.1 minute, a model length of N = 50, and a control horizon of M = 1 are used
in the following simulations. For this particular example, the prediction horizon does not appear to
have an appreciable effect, as shown in Figure 16-9; the setpoint tracking performance is roughly
the same for prediction horizons of P = 10 and 25. There is a lower limit, however, to the length
of the prediction horizon that can be used. A prediction horizon of 7 or less results in an unstable
closed-loop system, as shown in Figure 16-10. This is not due to any model error, since we have
assumed a perfect model in these simulations. If the prediction horizon is too short, the initial
step response coefficients dominate. Since these are negative while the later coefficients are
positive (corresponding to a positive process gain), the prediction is really in error. The effect is
the same as using a PID controller with a controller gain that is the wrong sign.

Figure 16-9. Closed-loop responses to a step setpoint change, van de

Vusse example. Sample time = 0.1 minute; model length, N = 50;
control horizon, M = 1; no weighting on input. Comparison of
prediction horizons, P = 10 and P = 25.
Figure 16-10. Closed-loop responses to a step setpoint change, van de
Vusse example. Sample time = 0.1 minute; model length, N = 50;
control horizon, M = 1; no weighting on input. Comparison of
prediction horizons, P = 7 and P = 8.

[ Team LiB ]
[ Team LiB ]

16.4 Constraints and Multivariable Systems

The dynamic matrix control technique presented in the previous section is based on an
unconstrained optimization of current and future control moves. The combination of a linear
model and a quadratic objective function lead to an analytical solution for the control moves. In
practice, constraints on manipulated inputs (control moves) can be very important. If Equation
(16.23), which is the analytical solution to Equation (16.22), results in an infeasible control action
(e.g., a violation of constraints, such as a flow rate greater than the maximum possible flow),
then obviously the control moves must be "truncated" at the maximum or minimum values. Since
the resulting truncated solutions may not be optimal if the control horizon is greater than 1, it is
important to use a constrained optimization formulation for these problems. Fortunately, dynamic
matrix control is easily formulated to explicitly handle constraints by using quadratic programming
(QP); the method is known as Quadratic Dynamic Matrix Control (QDMC).

Quadratic DMC (QDMC)

QDMC considers constraints on the manipulated inputs. The input constraints can be of the
following form

Equation 16.25

which is suitable for minimum and maximum flowrates, for example. In addition, velocity
constraints that limit the magnitude of the control moves at each sample time have the following
form

Equation 16.26

where ordinarily, Dumin = –Dumax. To use a standard quadratic program (QP), the constraints in
(16.25) need to be written in terms of the control moves, Duk+i. Since the previously
implemented control action (uk–1 ) is known, we can write

Equation 16.27

and so on. Since the manipulated input constraints are enforced over the control horizon of M
steps, (16.25) and (16.27) yield

Equation 16.28
Most standard QP codes use a "one-sided" form

Equation 16.29a

and

Equation 16.29b

which have the form ADuf b.

The velocity constraints are implemented as bounds on the control moves

Equation 16.30

The majority of constrained MPC problems can be solved based on the input constraints
considered above. For completeness, however, we also show how constraints on the process
outputs can be included.

It may be desirable to force the predicted process outputs to be within a range of minimum and
maximum values

Equation 16.31
Here we first rewrite Equation (16.15) in the following form

Equation 16.32

where f, the free response of the "corrected-predicted output" (if no current and future control
moves are made) is

Equation 16.33

so that (16.31) can be written

Equation 16.34

or, in terms of one-sided inequalities

Equation 16.35

Using shorthand matrix-vector notation, the quadratic programming problem is stated as

Equation 16.36

Equation 16.37
where the matrices and vectors in the objective function are

Equation 16.38

and the inequality matrices, A and b, in (16.37) incorporate the matrices in (16.29) and (16.35),
with appropriate dimensionality. For example, if there are no output constraints, then A and b are

Equation 16.39

Futher presentation is beyond the scope of this text. Example m-files for the study of constrained
systems will be placed on the book web page.

Multivariable Systems
In this chapter we have focused on single input–single output processes. One reason that MPC
has been so widely applied is that it naturally handles multivariable systems. The same basic
ideas presented in this chapter hold for the multivariable case, but the formulation for step
response models in a DMC framework is somewhat tedious. We recommend that you use the
Model Predictive Control Toolbox to perform multivariable studies.

[ Team LiB ]
[ Team LiB ]

16.5 Other MPC Methods

MPC is a basic concept or idea that can be implemented in many ways, depending on models used
and assumptions made (about disturbances and so on). With DMC, future output predictions were
composed of two effects: the unforced (or free) response (effect of past control moves) and the
forced response (effect of current and future control moves). Virtually all other model-based
procedures use the same idea. For example, if state space models are used, the free response
can be obtained simply by iteration from the current state values, assuming no input changes.
Here we discuss a number of industrial techniques and current MPC research.

MPC with Industrial Applications

The previous section covered the most widely applied (in industry) MPC technique, DMC. There
are a wide variety of other MPC techniques that have been used in industrial applications; only a
few are discussed here. A nice overview of these approaches is provided by Qin and Badgwell
(1997).

Model predictive heuristic control (MPHC) was developed in France and received initial applications
in the French petroleum refining industry. The software was called IDCOM for "identification and
command." This work is summarized by Richalet et al. (1978). A finite impulse response model is
used, and the desired setpoint trajectory over the prediction horizon is first order.

IDCOM-M, developed by Setpoint, Inc. (now part of ASPEN Technology), differs from IDCOM in a
number of ways. In addition to handling multivariable processes, it includes constraint handling,
where hard constraints can be ranked in terms of priority. Also, rather than considering outputs
over an entire prediction horizon, each output is controlled at a specific point in the future
(termed the coincidence point).

Generalized predictive control (GPC) arose from adaptive control research, where model
parameters are continuously updated to account for changing process dynamics. It is geared
toward SISO processes and has had few major industrial applications. Clarke (1988) provides an
overview of some applications.

MPC Research

Problems with Existing Techniques

Many of the "classical" MPC approaches used in industry have performance limitations, which we
summarize here.

Model structure. The finite step and impulse response models limit applications to open-loop
stable processes and require many model coefficients to describe the response (indeed, even
a first-order system with two parameters may require 50 or more step response
coefficients). Integrating systems have been handled by formulating the derivative of an
integrating output as the controlled output.
 Disturbance assumption. In Chapter 8 we found that the standard IMC design may not
handle input disturbances very well, since it is formulated for output disturbances. Similarly,
DMC uses a "constant output disturbance" assumption. This may not yield good performance
if the real disturbance occurs at the plant input.

Finite horizons. We found in the examples presented in Section 16.3 that control
performance could deteriorate if the prediction or control horizons were not formulated
correctly, even if the model was perfect.

Model type. The step and impulse response models are all linear. For some processes
(exothermic reactors) where the process operating conditions are changed frequently
(different product specifications for each consumer, for example), a single linear model may
not describe the dynamic behavior of the process over the wide range of conditions. Batch
processes also operate over a wide range of conditions. For these systems, better control
performance may be achieved if nonlinear models are used.

Solutions Being Developed

Much academic research has gone into developing approaches to solving these problems. It is
now recognized that there are many advantages to using discrete state space models. State
space models require fewer model parameters than step and impulse response models to describe
process behavior. State space models also provide more flexibility for disturbance estimation.
State estimation techniques can be easily formulated to handle input disturbances, for example.
State estimation also allows unstable systems to be controlled. This is discussed by Muske and
Rawlings (1993), for example.

Infinite horizon approaches can be used to guarantee closed-loop stability. Although it may seem
impossible to solve an optimization problem over an infinite horizon, linear models have analytical
solutions, allowing rapid computations to be performed. This topic is covered by Muske and
Rawlings (1993).

Nonlinear models can be used to improve performance, compared with linear models. Reviews of
these approaches are provided by Bequette (1991) and Henson (1998). An alternative is to
include several linear models to approximate the nonlinear system over a range of operating
conditions. The linear models can then be "scheduled" or "interpolated" so that the best linear
model approximation to the nonlinear process is used at a particular operating condition. An
application to drug infusion during surgery is discussed by Rao et al. (2001).

[ Team LiB ]
[ Team LiB ]

16.6 Matlab
There is an MPC Toolbox available through Mathworks. It is recommended if you are going
perform many MPC simulations, especially if you are working with multivariable systems.

For those without the MPC Toolbox, simple m-files that can be modified for problems of interest
are presented in Appendix 16.3. It is suggested that you download these files from the course
web page and use them for Problems 1 and 2. For each new problem, simply insert the new
continuous state space model at the appropriate location in the dmcsim.m file.

[ Team LiB ]
[ Team LiB ]

16.7 Summary
MPC is the most widely applied advanced control technique in industry. Our presentation focused
on DMC, which has achieved a great deal of success in the petroleum refining and petrochemicals
industries. Most MPC techniques have been based on step or impulse response models, but there
has been a recent trend toward state space models. New mathematical techniques are making it
easier to develop state space models from plant data. Regardless of the model type, the same
basic ideas are used. At the current time step, future process output predictions are based on two
contributions: the free or unforced response (how the outputs will change if no further control
moves are made) and the forced response (the effect of current and future control moves on the
predicted outputs). In DMC, the free response is the effect of the past control moves and the
additive correction term, while the forced response is the effect of the current and future (to be
calculated) control moves.

We studied two simple examples to develop a basic understanding of how to tune SISO model
predictive controllers. When using step or impulse models, it is important to make certain that the
model length (N) is long enough to capture the steady-state change; although not discussed, it is
important to "filter," or smooth, the step response data. We found that it is also important to
make certain that the prediction horizon (P) is long enough to avoid stability problems,
particularly for systems with inverse response behavior. Generally, control horizons (M) are much
shorter than are prediction horizons, yielding more robust performance. A disadvantage of MPC is
that there are many parameters (model length, prediction horizon, control horizon, manipulated
input weighting, and even sample time) that affect the closed-loop performance.

The variables used in this chapter are as follows:

M control horizon (N P M)
P prediction horizon (P M)

N model length (N P)
Sf dynamic matrix
S vector of step response coefficients
Spast matrix, used for effect of past control moves

si step response coefficient

Dt sample time
u manipulated input
Du change in manipulated input (control move)
w weight applied to manipulated in the objective function (often 0 if P >> M)

The following nomenclature is common

DMC dynamic matrix control

FIR finite impulse response

FSR finite step response

MPC model predictive control

[ Team LiB ]
[ Team LiB ]

References and Relevant Literature

There are large number of papers that discuss many different aspects and formulations of MPC.
The basic DMC approach is presented by Cutler and Ramaker (1980). An early academic paper
analyzing DMC is presented by Marchetti et al. (1983). Garcia et al. (1989) review early theory
and applications, while Morari and Lee (1999) summarize recent results and current research
topics. The monograph by Camacho (1999) presents a number of techniques but focuses on GPC.
The textbook by Maciejowski (2002) includes MATLAB files and a number of application case
studies.

Camacho, E.F., and C. Bordons, Model Predictive Control, Springer-Verlag, London (1999).

Clarke, D.W., "Application of Generalized Predictive Control to Industrial Processes," IEEE Control
Syst. Mag., April, pp. 49–55 (1988).

Cutler, C.R., and B.L. Ramaker, "Dynamic Matrix Control—A Computer Control Algorithm," in
Proc. Joint Automatic Control Conference, San Francisco, CA, Paper WP5-B (1980).

Garcia, C.E., and A.M. Morshedi, "Quadratic Programming Solution of Dynamic Matrix Control
(QDMC)," Chem. Eng. Commun., 46, 73–87 (1986).

Garcia, C.E., D.M. Prett, and M. Morari, "Model Predictive Control: Theory and Practice—A
Survey," Automatica, 25, 335–348 (1989).

Maciejowski, J.M., Predictive Control: with Constraints, Prentice Hall, Harlow, England (2002).

Marchetti, J.L., D.A. Mellichamp, and D.E. Seborg, "Predictive Control Based on Discrete
Convolution Models," Ind. Eng. Chem. Proc. Des. Dev., 22(3), 488–495 (1983).

Morari, M., and J.H. Lee, "Model Predictive Control: Past, Present and Future," Comp. Chem. Eng.,
23, 667–682 (1999).

Muske, K., and J.B. Rawlings, "Model Predictive Control with Linear Models," AIChE J., 39(2),
262–287 (1993).

Qin, S.J., and T.A. Badgwell, "An Overview of Industrial Model Predictive Control Technology,"
Proceedings of the Fifth Conference on Chemical Process Control, AIChE Symp. Ser. 316, Vol. 93,
pp. 232–256 (Kantor, J.C., C.E. Garcia, and B. Carnahan, Eds.), (1997).

Richalet, J.A., A. Rault, J.D. Testud, and J. Papon, "Model Predictive Heuristic Control: Applications
to Industrial Processes," Automatica, 14, 413–428 (1978).

Shah, S.L., "A Tutorial Introduction to Constrained Long Range Predictive Control," Pulp Pap.
Can., 96(4), 57–63 (1995).

A focus on nonlinear systems is presented in the following papers and books.

Allgower, F., T.A. Badgwell, J.S. Qin, J.B. Rawlings, and S.J. Wright, "Nonlinear Predictive Control
and Moving Horizon Estimation—An Introductory Overview," Advances in Control: Highlights of
ECC '99, Ch. 12, pp. 391–449, (P.M. Frank, Ed.), Springer-Verlag, Berlin/New York (1999).

Bequette, B.W., "Nonlinear Control of Chemical Processes: A Review," Ind. Eng. Chem. Res., 30,
1391–1413 (1991).

Henson, M.A., "Nonlinear Model Predictive Control: Current Status and Future Directions," Comp.
Chem. Eng., 23, 187–202 (1998).

Lee, J.H., and N.L. Ricker, "Extended Kalman Filter Based Nonlinear Model Predictive Control,"
Ind. Eng. Chem. Res., 33(6), 1530–1541 (1994).

Meadows, E.S., and J.B. Rawlings, "Model Predictive Control," in Nonlinear Process Control, pp.
233–310 (M.A. Henson and D.E. Seborg, Eds.), Prentice Hall, Upper Saddle River, NJ (1997).

Sistu, P.B., and B.W. Bequette, "Model Predictive Control of Processes with Input Multiplicities,"
Chem. Eng. Sci., 50(6), 921–936 (1995).

An application of a multiple model-based MPC approach is discussed in the following paper:

Rao, R., C.C. Palerm, B. Aufderheide, and B.W. Bequette, "Experimental Studies on Automated
Regulation of Hemodynamic Variables," IEEE Engineering in Medicine and Biology Magazine,
20(1), 24-38 (Jan/Feb, 2001).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Reproduce the DMC simulations for the van de Vusse reactor shown in Figures 16-9 and
16-10 using the files shown in Appendix 16.3 (also, you may download these files from
the book web page).

2: Run the DMC simulations for the van de Vusse reactor, for a prediction horizon of 10
(and a model length of 50). Vary the control horizon and discuss the effect; for
example, if M = 5, the manipulated input resembles a "pulse" change. Also, try different
weighting on the manipulated input and discuss the effect.

3: Consider a biochemical reactor operated at a stable operating condition (Module 7), with
the state space model

where the input is the dilution rate (units = min-1) and the output is the biomass
concentration (gmol/L). Modify the DMC simulation code presented in Appendix 16.3 to
simulate the MPC applied to this system. Use a sample time of 0.05 hours and discuss
the effect of model length, prediction horizon, and control horizon on the performance
for setpoint changes of 0.02 gmol/L. Consider the effect of these parameters on the
magnitude of manipulated input changes. What are your recommended tuning
parameters if the maximum manipulated input change is 0.2 min-1?

4: Consider a jacketed chemical reactor, where jacket temperature is the manipulated

input and reactor temperature is the measured output. For the 100 ft 3 propylene glycol
reactor presented in Module 8, the state space model is

Modify the DMC simulation code presented in Appendix 16.3 to simulate MPC applied to
this system. Use a sample time of 0.05 hours and discuss the effect of model length,
prediction horizon, and control horizon (remember that N > P > M) on the performance
for setpoint changes of 5°F. Consider the effect of the tuning parameters on the
manipulated input changes (that is, is the change in jacket temperature reasonable?).

5: Consider a discrete state space model with the following form

 Assume that the state vector is perfectly known at the current time step (k), the model
is perfect, and there are no disturbances. Predict the output up to P time steps in the
future, based on M current and future control moves (Duk, ..., Duk+M-1). Realize that u k
= uk-1 + Duk, etc. Show how the output prediction is naturally composed of free and
forced response contributions.

[ Team LiB ]
[ Team LiB ]

Appendix 16.1: Derivation of the Step Response

Formulation
From superposition of all previous inputs, the output at step k can be found from

However, assuming that all step response coefficients greater than step N are equal to the value
at step N (this is only true for stable systems with a long model horizon), we can write

[ Team LiB ]
[ Team LiB ]

Appendix 16.2: Derivation of the Least Squares Solution

for Control Moves
The objective function is

Expanding the objective function, we find

Unconstrained System

The minimization with respect to the control move vector, when there are no constraints, is

and solving for Du f we find

Equation 16.23

Constrained System

The objective function for a constrained system is no different than the unconstrained

but this is often written in the following form

and since

the objective function in the form of most quadratic programs is

and

The constraints

are shown in Section 16.4.

[ Team LiB ]
[ Team LiB ]

Appendix 16.3
Use the following files to reproduce the van de Vusse simulations shown in Figure 16-9. These are
based on an unconstrained DMC strategy. Files for the constrained formulation, involving qp and
quadprog from the optimization toolbox, are available on the book web page.

dmcsim.m

% dmcsim
%
% 26 Oct 00 - revised 4/5 Aug 01 for Chapter 16 - MPC
% b.w. bequette
% unconstrained DMC simulation, calls: smatgen.m and dmccalc.m
% currently contains van de vusse reactor model and plant
%
% MPC tuning and simulation parameters
n = 50; % model length
p = 10; % prediction horizon
m = 1; % control horizon
weight = 0.0; % weighting factor
ysp = 1; % setpoint change (from 0)
timesp = 1; % time of setpoint change
delt = 0.1; % sample time
tfinal = 6; % final simulation time
noise = 0; % noise added to response coefficients
%
t = 0:delt:tfinal; % time vector
kfinal = length(t); % number of time intervals
ksp = fix(timesp/delt);
r = [zeros(ksp,1);ones(kfinal-ksp,1)*ysp]; % setpoint vector
%
% ----- insert continuous model here -----------
% model (continuous state space form)
%
a = [-2.4048 0;0.8333 -2.2381]; % a matrix – van de vusse
b = [7; -1.117]; % b matrix – van de vusse
c = [0 1]; % c matrix – van de vusse
d = 0; % d matrix – van de vusse
sysc_mod = ss(a,b,c,d); % create LTI "object"
%
% ----- insert plant here -----------
% perfect model assumption (plant = model)
ap = a;
bp = b;
cp = c;
dp = d;
sysc_plant = ss(ap,bp,cp,dp);
%
% discretize the plant with a sample time, delt
%
sysd_plant = c2d(sysc_plant,delt)
[phi,gamma,cd,dd] = ssdata(sysd_plant)
%
% evaluate discrete model step response coefficients
%
[s] = step(sysc_mod,[delt:delt:n*delt]);
%
% generate dynamic matrices (both past and future)
%
[Sf,Sp,Kmat] = smatgen(s,p,m,n,weight);
%
% plant initial conditions
xinit = zeros(size(a,1),1);
uinit = 0;
yinit = 0;
% initialize input vector
u = ones(min(p,kfinal),1)*uinit;
%
dup = zeros(n-2,1);
sn = s(n); % last step response coefficient
x(:,1)= xinit;
y(1) = yinit;
dist(1) = 0;
%
% set-up is done, start simulations
%
for k = 1:kfinal;
%
du(k) = dmccalc(Sp,Kmat,sn,dup,dist(k),r(k),u,k,n);
% perform control calculation
if k > 1;
u(k) = u(k-1)+du(k); % control input
else
u(k) = uinit + du(k);
end
% plant equations
x(:,k+1) = phi*x(:,k)+gamma*u(k);
y(k+1) = cd*x(:,k+1);
% model prediction
if k-n+1>0;
ymod(k+1) = s(1)*du(k) + Sp(1,:)*dup + sn*u(k-n+1);
else
ymod(k+1) = s(1)*du(k) + Sp(1,:)*dup;
end
% disturbance compensation
%
dist(k+1) = y(k+1) - ymod(k+1);
% additive disturbance assumption
% put input change into vector of past control moves
dup = [du(k);dup(1:n-3)];
end
%
% stairs plotting for input (zero-order hold) and setpoint
%
[tt,uu] = stairs(t,u);
[ttr,rr] = stairs(t,r);
%
figure(1)
subplot(2,1,1)
plot(ttr,rr,'--',t,y(1:length(t)))
ylabel('y')
xlabel('time')
title('plant output')
subplot(2,1,2)
plot(tt,uu)
ylabel('u')
xlabel('time')

smatgen.m

function [Sf,Sp,Kmat] = smatgen(s,p,m,n,w)

%
% b.w. bequette
% 28 Sept 00, revised 2 Oct 00
% generates dynamic matrix and feedback gain matrix
% assumes s = step response column vector
% Sf = Dynamic Matrix for future control moves (forced)
% Sp = Matrix for past control moves (free)
% Kmat = DMC feedback gain matrix
% s = step response coefficient vector
% p = prediction horizon
% m = control horizon
% n = model horizon
% w = weight on control input
%
% first, find the dynamic matrix
for j = 1:m;
Sf(:,j) = [zeros(j-1,1);s(1:p-j+1)];
end
%
% now, find the matrix for past moves
%
for i = 1:p;
Sp(i,:) = [s(i+1:n-1)' zeros(1,i-1)];
end
%
% find the feedback gain matrix, Kmat
%
Kmat = inv(Sf'*Sf + w*eye(m))*Sf';

dmccalc.m
function [delu] = dmccalc(Sp,Kmat,sn,delup,d,r,u,k,n)
%
% for use with dmcsim.m
% b.w. bequette
% 2 oct 00
% calculate the optimum control move
%
% first, calculate uold = u(k-n+1)...u(k-n+p)
%
[m,p] = size(Kmat);
uold = zeros(p,1);
for i = 1:p;
if k-n+i>0;
uold(i) = u(k-n+i);
else
uold(i) = 0;
end
end
dvec = d*ones(p,1);
rvec = r*ones(p,1);
y_free = Sp*delup + sn*uold + dvec;
e_free = rvec-y_free;
delu = Kmat(1,:)*e_free;

[ Team LiB ]
[ Team LiB ]

Chapter 17. Summary

A lot of material has been covered in this textbook, yet there are many control-related topics that
have not been covered. The purpose of this chapter is to provide a concise summary of the topics
covered and to provide an overview of what the reader should expect in a process engineering
position. This chapter should spark your memory and help you determine which topics require a
more in-depth review.

After reading this chapter the reader should be able to:

Recall most of the important topics covered in the textbook

"Ace" a final examination in a dynamics and control course

Have a feel for important practical issues in process engineering

The major sections of this chapter are as follows:

17.1 Overview of Topics Covered in This Textbook

17.2 Process Engineering in Practice

17.3 Suggested Further Reading

17.4 Notation

[ Team LiB ]
[ Team LiB ]

17.1 Overview of Topics Covered in This Textbook

In the next few pages we attempt to concisely review the topics covered in this textbook. This
summary should assist you in preparing for a final examination in a process control course, for
example. It also serves the purpose of providing a quick way of helping you determine what
topics you would like to understand better. In the next section these topics are reviewed in the
chronological order presented in the text. In the subsequent section a concise review of the
material presented in the modules is presented.

Chapters
Chapter 1 provided the motivation for process control, while Chapter 2 introduced fundamental
models. Chapter 3 presented dynamic behavior, with a focus on transfer functions. Chapter 4
covered the development of empirical models, including continuous time step responses, as well
as the identification of parameters for discrete-time models. Chapter 5 provided an introduction to
the analysis of closed-loop control strategies, introducing the idea of a block diagram. Chapter 6
presented PID controller tuning techniques, followed by frequency response analysis in Chapter 7.
Many advanced control techniques rely on the use of a process model embedded in the control
strategy; Chapter 8 covers the Internal Model Control strategy. Internal model controllers can
often be rearranged to form a PID controller, as shown in Chapter 9. Improved disturbance
rejection is the main motivation of the cascade and feedforward techniques presented in Chapter
10. Enhancements to PID control, including autotuning, integral windup protection, and nonlinear
approaches are discussed in Chapter 11. Ratio, split-range and selective controllers are presented
in Chapter 12. The interaction of multiple SISO loops, with the RGA as the main analysis tool, is
developed in Chapter 13. In Chapter 14, right-half-plane transmission zeros indicate dynamic
performance limitations; the singular value decomposition provides insight about possible steady-
state performance limitations. The topic of plantwide control is the focus of Chapter 15, followed
by model predictive control in Chapter 16. Specific topics are presented in the pages that follow.

Given a process, be able to do the following:

Identify the control objective and the state, output, manipulated, and disturbance variables

Develop the fundamental nonlinear model

Solve for the steady state

Linearize the nonlinear model about the steady state to find the state-space model

Find the transfer functions relating each input to each output

Convert to different sets of units

Given experimental step responses, the reader should be able to estimate parameters for low-
order transfer function models that match the measured process output.

Given steady-state information, the reader should be able to do the following:

Convert deviation variable results to physical variable results

 Convert physical variable results to deviation variables

Given a process transfer function, be able to

Relate the gain-time constant form and the pole-zero form

Relate poles and time constants

Relate the location of a pole in the complex plane to the speed of response

Consider a first-order + dead time transfer function (Why does the first-order Padé
approximation for dead time provide a good approximation for the response to a step input?
Hint: Think of RHP zeros.)

Use the final- and initial-value theorems of Laplace transforms

Given a physical process with feedback control (a process instrumentation and control diagram),
the reader should be able to do the following:

Draw the block diagram for any type of feedback control (all variables should be clearly
identified on the block diagram

For a given block diagram, the reader should be able to do the following:

Derive the closed-loop transfer functions (for example, relate the setpoint to the output for a
feedback control scheme)

Relate the disturbance to the output for a feedback control scheme

For a given process transfer function, the reader should be able to do the following:

Determine the offset due to proportional feedback control

Perform a closed-loop Ziegler-Nichols (continuous oscillation) test

Compare the performance of Ziegler-Nichols and Tyreus-Luyben suggested tuning

parameter values

Use the Routh stability criterion to determine if a closed-loop system is stable (the
denominator of your closed-loop transfer function will usually be a polynomial with an order
of 2–4

Determine the range of tuning parameter values that will yield a stable feedback system if
the controller tuning parameters are not given

Use the direct synthesis method to find a control algorithm that yields a desired closed-loop
response. Find the corresponding PID parameters if the controller has PID form (sometimes
in series with lead-lag)

Calculate the poles and discuss the response characteristics for a closed-loop transfer
function that results in a first or second-order polynomial

Discuss the response characteristics if you are given the closed-loop poles for a higher order
closed-loop polynomial
Given a Bode diagram of gc(w)gp(w), be able to do the following:

Determine the gain and phase margins

Use that information to calculate the amount of additional deadtime allowable before
instability occurs

Given a Nyquist diagram of gc(w)gp(w), be able to do the following:

Determine if the closed loop system will be stable

Understand the relationship between Bode and Nyquist plots (How is the information from a
Bode plot translated to the Nyquist diagram? How is information of the Nyquist diagram
translated to a Bode plot?)

Understand that the closed-loop Ziegler-Nichols method is equivalent to finding the proportional
controller gain where the phase angle is –180o when the open-loop (gc*gp) amplitude ratio is one.

For a given transfer function the reader should be able to:

Determine the magnitude of the output from the transfer function at low input frequencies
and at high input frequencies

Understand what is meant by the term "nonminimum phase"

Be able to determine if a transfer function is proper, semi-proper, or improper

Be able to handle the complete IMC design procedure for a given process transfer function:

Factor gp(s) into minimum-phase and nonminimum-phase elements {good [gp-(s)] vs. bad
[gp+(s)] stuff}

Invert the minimum phase elements to form the ideal IMC controller

Cascade the ideal IMC controller with a filter that is high enough order to form a physically
realizable IMC controller

Determine what the closed-loop response will look like for an IMC scheme with a perfect
model

Understand the effect of l, the IMC filter time constant, on the closed-loop response (How
does l impact the robustness? A good first guess for tuning l is to use roughly one third to
one half the dominant time constant, depending on the amount of time-delay and model
uncertainty; note that l has units of time)

Know that the IMC-based PID procedure will give the same closed-loop results as the IMC
strategy when there are no time delays (the behavior will be different when there are time
delays because the IMC-based PID procedure uses a Padé approximation for the time delay)

Design a physically realizable feed-forward controller

Derive equivalent transfer functions to analyze a cascade-control strategy as a standard feedback

control strategy

Use ARW techniques (controllers with integral action can exhibit "reset windup" when the
manipulated input becomes constrained)

For multivariable systems, the reader should be able to do the following:

Use the RGA for input-output pairing for MVSISO systems

Renumber the inputs and outputs such that the favorable RGA pairings appear on the
diagonal of the new gain matrix

Realize that you should never pair on a negative or zero relative gain

Find the structure of the control matrix that corresponds to the RGA pairing that has been
selected

Find the closed-loop transfer functions for any multivariable block diagram

Realize that the order of multiplication for matrices is critical

Implement steady-state decoupling

Calculate multivariable transmission zeros

Use a singular value decomposition to understand directional effects

Implement MV IMC, using a diagonal factorization

For a system with more than two inputs and two outputs, the reader should know the following:

To check the RGAs of all reduced-order structures (this is to make certain that when a
control loop is out of service, the reduced order system is not failure sensitive)

Place control loops on a process flow sheet (one heuristic is that one stream in a recycle loop
should be under flow control)

Understand the basic idea behind MPC. For the step response-based technique of DMC,
understand

How the model length (N) is determined

Are longer prediction horizons (P) more robust than shorter ones?

Are shorter control horizons (M) more robust than longer ones?

Modules
The modules in the final section of the text provide detailed application examples to illustrate the
techniques presented in the Chapters. Module 1 reviews MATLAB, while Module 2 covers SIMULINK.
Numerical integration of ordinary differential equations using MATLAB is presented in Module 3,
while Module 4 presents useful functions available in the Control System Toolbox. An isothermal
CSTR with a series/parallel reaction structure (the van de Vusse reaction) is studied in Module 5.
Frequency response techniques are used to analyze the classic first-order + dead time model in
Module 6. Biochemical reactors and the classic exothermic CSTR are convered in Modules 7 and 8,
respectively. Steam and surge drum level control problems differ significantly, as presented in
Modules 9 and 10. Batch reactors are studied in Module 11, followed by biomedical systems in
Module 12. Linear and nonlinear effects in distillation are discussed in Module 13. A set of case
studies is summarized in Module 14; these are particularly useful for open-ended final projects in
a typical process control course. Flow and digital control, Module 15 and Module 16, complete the
text.

[ Team LiB ]
[ Team LiB ]

17.2 Process Engineering in Practice

There is a limit to the amount of material that can be covered in a single textbook with reasonable
length and depth of coverage. In this text we have chosen topics that we consider to form a firm
basis for further study in automation and control. Hopefully you now have a firm foundation for
further study on automation and control-related topics.

Topics Not Covered

Important topics not addressed in this textbook to a significant extent include the following:

Statistical process control (SPC): SPC is used by process engineers and operators to
determine if a process is producing a product that is meeting specified quality targets. SPC
techniques attempt to differentiate between normal statistical deviations in product quality
and deviations due to operating problems (these may be control related)

Advanced digital control techniques: An introduction to the most important advanced control
technique, model predictive control (MPC), was presented in Chapter 16. There are a large
number of important topics in MPC that were not presented. These include the use of
different types of models (linear step response models were covered), different model
updating techniques (output additive disturbance was covered), and different objective
functions (least squares over a finite prediction horizon was covered). Nonlinear models and
adaptive linear models can also be used.

Inferential control: Very often the most important process variable (product quality or
property) cannot be measured on-line. Other measurements must be used to "infer" the
quality variable; one example is to use tray temperature in distillation to infer the distillate
purity.

Noise filtering: In practice, measurements are corrupted by noise. Often "filters" are used to
"average" recent measurements to obtain a better estimate of the actual value.

Specific control system hardware and software: We have chosen not to cover specific
hardware or software for automation projects in detail. Although important in the actual
implementation of a control project, we feel that the role of the process engineer is to
specify control objectives, determine the proper measured variables, etc. Instrumentation
specialists and other project engineers will be involved with many of the detailed
specifications necessary to complete an automation project.

Art and Philosophy of Process Engineering

The majority of students majoring in chemical engineering become process engineers in their first
industrial position. A typical responsibility is to provide technical support for an entire operating
unit (termed a "plant" in Chapter 15). This includes troubleshooting (determining why a product is
suddenly off-specification, for example), and capital project justification (designing new
equipment to reduce energy consumption, minimize waste, maximize profitability, etc.), among
numerous other things. One of the first things that you will do is review process flow diagrams (to
understand the basic process flows) and process and instrumentation diagrams (P&IDs, to
understand the basic control strategy). You will spend a lot of time tracing pipes through
piperacks, to understand physically the layout of the process unit. You will also spend a lot of time
talking to process operators, shift foremen, and the unit manager to better understand basic
operating procedures. An important part of process operation, naturally, is the automation and
control strategy. Each process plant will have an engineer whose primary responsibility is the
basic maintenance of the computer control system. Most often a distributed control system (DCS)
is used. As a process engineer for a particular unit, you will be expected to achieve some level of
knowledge of the control system.

The point of the previous paragraph is that you, as a process engineer, do not operate in
isolation. There are many people involved in the operation of a chemical process, no matter which
industry you are working in. There is an entire history to the operation of a particular process,
and you will not be expected to be an expert overnight. You need to be willing, however, to listen
and learn from the experiences of others. At the same time, you need to realize that not
everything you will be told will be technically correct and that some information is biased (due to
management and labor relation issues, etc.). You will need to achieve a delicate balance between
believing everything you are told and being so skeptical that you believe nothing.

So, will you be deriving equations, manipulating block diagrams, and using the Routh stability
criterion on a daily basis? In all probability, no. But hopefully this textbook has given you some
basic knowledge of dynamics and control that you can use in a broad, qualitative sense. Without
explicitly trying, you will probably begin to think in terms of inputs, outputs, and objectives. You
will begin to develop a feel for the order of magnitude of an effect of an input change on an
output. You will naturally perform "back of the envelope" calculations based on material and
energy balances to understand these relationships (for example, if the steam flow to an
exchanger changes by a certain amount, how much do you expect the process fluid temperature
to change?).

Eventually, as you learn more about the process, you may conceive ideas to improve the control
of the process. There is no better way to develop better control strategies than to achieve a
deeper knowledge of the process. The best "loop-tuner" in the world will not achieve good control
loop performance if a bad input-output pairing is made to begin with.

Rules of Process Operations

We have all heard of Murphy's law—"if something can go wrong, it will go wrong." Similarly with
process operations, it is not a question of if a control element will fail, but rather, when will the
control element fail? Safe processes are designed with the knowledge that most moving
equipment will eventually fail. Also, sensors will fail or need to be recalibrated. As we noted in
Chapter 1, all control valves are designed with a bypass line. When the control valve fails, it can
be taken out of service while the flow is regulated manually through the bypass valve. Similarly,
virtually all pumps have spare pumps placed "in parallel" with them. When one pump fails
(detected as a loss of pressure), the other pump automatically "kicks on" before the process is
affected. This also allows preventive maintenance to be performed on all pieces of moving
equipment.

It should also be noted that the probability of a failure or technical problem occurring between 8
am and 5 pm on Monday–Friday is roughly 27% (you do the math), so 73% of failures and
technical problems occur outside the normal working hours of a process engineer. If the technical
problem is critical to continued process operation (remember, many of these operating units are
producing several $ million/day of product), and you are the process engineer "on-call," you will
likely be called in to help solve the problem. This is usually exciting (sometimes you may wish for
less excitement in your life), sometimes with almost an "emergency room" sense of urgency as
you and many others try and troubleshoot/solve the problem.
[ Team LiB ]
[ Team LiB ]

17.3 Suggested Further Reading

A number of literature sources were suggested in Chapter 1. Now that you have a good basic
background in process control, you may wish to obtain a deeper or more applied knowledge by
studying the following sources.

For books with many practical tips and present numerous application examples, read the
following:

Shinskey, F. G. Feedback Controllers for the Process Industries, McGraw Hill, New York (1994).

Shinskey, F. G., Process Control Systems. Application, Design and Tuning, 4th ed., McGraw Hill,
New York (1996).

The plantwide process control book by Luyben et al. provides numerous unit operation and entire
flow-sheet examples.

Luyben, W. L., B. D. Tyréus, and M. L. Luyben, Plantwide Process Control, McGraw Hill, New York
(1999).

For a more detailed coverage of linear systems analysis and an introduction to nonlinear systems
behavior and analysis, read the following textbook:

Bequette, B. W., Process Dynamics: Modeling, Analysis and Simulation, Prentice Hall, Upper
Saddle River, NJ (1998).

For a nice introduction to robust control system design for multivariable systems, read the
following:

Skogestad, S., and I. Postlethwaite, Multivariable Feedback Control: Analysis and Design, Wiley,
New York (1996).

This textbook focuses on chemical process control. For a discussion of the differences between
process control and other control fields (both practice and education), see the following article:

Bequette, B. W., and B. A. Ogunnaike, "Chemical Process Control Education and Practice," IEEE
Control Syst. Mag., 21(2), 10–17 (2001).

[ Team LiB ]
[ Team LiB ]

17.4 Notation
A Jacobian state space—relates states to the state derivatives

B input matrix state space—relates inputs to state derivatives

C output matrix state space—relates states to outputs

D state space—direct transmittance from inputs to outputs

e error setpoint—measured process output

l load (disturbance)
input

p pole roots of the denominator polynomial of gp(s); equal to the

eigenvalues of A
r setpoint desired value of the output
s Laplace domain when used in frequency-response techniques, set s = jw
variable
x state variable

y output variable measured output (also known as the controlled variable or

process variable)
u manipulated input also known as the controller output
z zero roots of the numerator polynomial of gp(s) equal to the poles of

gp(s) process transfer y(s) = gp(s) u(s)

function
gp(s) = C(sI – A)-1B+D

first-order transfer
function, gp(s)

gc(s) controller transfer

function
gd(s) disturbance transfer
function
gm(s) measurement
transfer function
gv(s) valve transfer
function

kp process gain change in output/change in input

tp process time first-order system time constant
constant
q process time delay

z damping factor used for unfactored second-order systems

kc proportional gain PID controller tuning parameter
tI integral time PID controller tuning parameter
constant

tD derivative time PID controller tuning parameter

constant

tF filter time constant

~ indicates model superscript

"good stuff" portion inverse is used for internal model controller

of process model
"bad stuff" portion of contains time-delays and RHP zeros
process model

q(s) internal model

controller

f(s) IMC filter used to make q(s) physically realizable (proper)

l filter factor, closed- tuning parameter for internal model controllers
loop time constant
l eigenvalue roots of det(lI – A)
lij relative gain relative gain between output i and input j

CLTF closed-loop transfer

function

GM gain margin indication of "gain uncertainty" that can be tolerated

IMC internal model can be either implemented in the IMC structure or rearranged
control to "standard feedback form" to design a PID controller (IMC-
based PID)
MIMO multi-input, multi-
output
MPC model predictive optimization of control moves over a future prediction horizon
control
MVSISO multivariable SISO "decentralized control"—consists of several SISO controllers

NMP nonminimum phase a NMP process has right-half-plane zeros and/or time delays
PM phase margin indication of "phase uncertainty" that can be tolerated (usually
used to find the amount of time delay that can be tolerated)
PID proportional-integral-
derivative (ideal)
PID proportional-integral-
derivative (ideal)

PID proportional-integral-
derivative (real)

RGA relative gain array used for variable pairing and interaction analysis

RHP right half-plane right-half-plane zeros have inverse response

SISO single input–single

output
SVD singular value indicator of directional sensitivity
decomposition

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Why must the process transfer function be factored in the IMC design procedure? (For
example, what happens if the process dead time is not factored out? Also, what
happens if a RHP zero is not factored out?)

2: The main advantage to cascade control is the rejection of inner-loop (secondary)

disturbances rather than outer-loop (primary) disturbances. True or False?

3: Since you are a young process engineer with a good theoretical process-control
background, your boss asks you to work on a problem in the catalytic reforming unit.
The process has been having trouble maintaining yield specifications and your boss
suspects that part of the problem could be due to the control system. You review the
process flow sheets and then proceed to the process unit. Your next step is to do which
of the following:

a. Begin discussing, with the process operator, a recent theoretical article from the
International Journal of Unapplied Incomprehensible Control. You discuss the
various ramifications of pole/zero cancellation, robustness with respect to
parameter uncertainty, and so forth, and relate this theoretical treatise to
problems in the process unit.

b. Grab a cup of coffee and sit down and discuss the most recent local high school
football game with the process operator. After a while, begin to direct the
conversation to the subject of the catalytic reforming process and discuss
possible operating changes that have occurred during the past few weeks.

4: A multivariable process has the following input-output steady-state gain matrix

a. How many process outputs are there?

b. How many manipulated inputs are there?

c. What is the gain relating input 3 to output 2?

5: Consider the first-order + dead-time transfer function

Equation 1
 The tuning parameters for an IMC-based PI controller (not the "improved" version) are
(from Table 9-2 in the text): kc = 1/l, tI = 1. Assuming a Padé approximation for dead
time, the approximate model is

Find the value of kc which brings the approximate closed-loop system to edge of
instability, using the Routh stability criterion.

6: Consider the following process gain matrix and RGA:

a. Can three SISO control loops be used?

b1. If your answer to a is yes, what are the recommended pairings [will these
pairings assure stability if one of the loops must be opened (think about the 2 x 2
subsystems)]? Reorder the inputs and outputs such that the paired variables are
on the diagonal of the new G(0) matrix. Show the new G(0) matrix and the new
L matrix.

b2. If your answer to a is no, do you recommend a reduction to two loops? If so,
what are your recommendations for input-output pairings for this system?

7: The objective of the following process is to produce a product stream with a desired
composition. The composition of the product stream is a function of the inlet
temperature. The inlet temperature is a function of the steam pressure on the shell
side of the feed preheat exchanger. Draw the proper connections of all of the
controllers and the valve shown in the following diagram. Draw the corresponding
block diagram, labeling all signals, controllers, and measurement devices.
8: Consider the process and instrumentation diagrams shown in the surge vessel level
control figure below, where the surge vessels are assumed to be cylindrical.
i. For which orientation can the surge vessel best be modeled as a linear process?
Why?

ii. Choose the vessel that you want to analyze. Assume that the steady-state flow
rate is 100 gallons/minute, and that the vessel has a residence time of 5 minutes
when half-filled with liquid (assume the vessel is one half full at steady state).
Assume a length/diameter ratio of 2:1. Also, assume that the control valve
(which has a linear valve characteristic) is designed to be one half open at the
steady-state operating point. For simplicity, assume that the controller output is
the flow rate of the outlet stream.

a. Determine the proportional gain for a proportional-only controller that will

just maintain the tank height within a range of 20–80% of the vessel
volume, when the inlet stream changes from 100 to 200 gallons/minute.
Remember to show the units for the proportional gain.

b.
 b. Discuss the advantages and disadvantages of having a fail-open valve.

9: An open-loop unstable chemical reactor is known to have a process transfer function

with the following form, where the manipulated input is the coolant flow rate
(liters/minute) and the measured process output is the reactor temperature (oC)

When a proportional-only controller, with a value kc = –3 lpm/oC, is used, the response

to a unit setpoint change is shown below.
 When a PI controller is used, the closed-loop system becomes unstable when the
tuning parameter values are kc = –1.333 lpm/oC and tI = 2 minutes.

Question: What are the values for the process parameters, kp and tu? Hints: Can the
final value theorem be used? Can the Routh stability criterion be used?

10: Type I diabetics must deliver themselves insulin shots several times each day, in an
effort to maintain blood glucose values within a desirable range. Portable infusion
pumps that provide a continuous delivery of insulin are becoming widely available, and
there has been a large effort to develop sensors that provide a continuous blood
glucose measurement. The ultimate goal is to create an "artificial pancreas" by forming
a feedback control system where blood glucose is the measured process variable and
insulin infusion rate is the manipulated input variable.

The Bergman minimal model (with 3 states) has the following input-output third-order
transfer function relationship

The open-loop response (in deviation variables) to a 10m-unit/min step input at t = 0

minutes is shown below (where m-unit means milli-units of insulin). As expected, the
system has integrating action. [Notice that all of these results are for the case of no
meal consumption (no disturbance).]
 a. Show that a proportional-only controller will not have offset to a step closed-loop
setpoint change (Hint: Think final value theorem).

b. For a proportional-only controller, what value of controller gain will cause the
closed-loop system to become unstable? Show the units for the proportional gain.

c. For any stable controller, what is the long-term insulin infusion rate change for a
step glucose setpoint change of –1 mmol/liter? What is surprising or interesting
about this result, compared to those of "non-integrating" systems?

[ Team LiB ]
[ Team LiB ]

Module 1. Introduction to MATLAB

The purpose of this module is to review M ATLAB for those that have used it before and to provide a
brief introduction to MATLAB for those that have not used it before. This is an interactive, "hands-
on" tutorial introduction. After using this tutorial, the reader should be able to do the following:

Enter matrices

Make plots

Write script files

Perform matrix operations

Use M ATLAB functions

Write function files

The sections of this module are as follows:

M1.1 Background

M1.2 Matrix Operations

M1.3 The MATLAB Workspace

M1.4 Complex Variables

M1.5 Plotting

M1.6 More Matrix Stuff

M1.7 For Loops

M1.8 m-Files

M1.9 Summary of Commonly Used Commands

M1.10 Frequently Used MATLAB Functions

[ Team LiB ]
[ Team LiB ]

M1.1 Background
MATLAB is an interactive program for numerical computation and data visualization. It is produced
by MATHWORKS, which has the following URL:

http://www.mathworks.com

You can find answers to frequently asked questions (FAQ) on this homepage. Also, there are a
number of technical notes that give more information on using many M ATLAB features. The
MATHWORKS issues a new release of MATLAB every 1–2 years. At the time of the writing of this
textbook, the most recent release was MATLAB 6.5 (Release 13). This is the first release since
MATLAB 5.2.1 (Release 10) to support the Macintosh (OS 10). The figures presented in this module
are from MATLAB 6.1, while much of the rest of the textbook was generated using MATLAB 5.2.1.

When you start MATLAB, a desktop similar to that shown in Figure M1-1, with three basic work
areas, appears; alternative layouts can be obtained using the "View" pull-down menu. In the
upper left corner, when "Launch Pad" is toggled, you will find the M ATLAB and SIMULINK toolboxes
available with your current MATLAB license. Examples using SIMULINK functions are provided in
Module 2. The right-hand work area is known as the command window; this is where most of
your MATLAB commands will be entered. It is also the focus of this tutorial module. The lower left
work area toggles between the current directory and the history of recent commands that you
have entered in the command window.

Figure M1-1. M ATLAB desktop (Version 6.1, Release 12).

This tutorial provides a brief overview of essential MATLAB commands that are entered in the
command window. You will learn this material more quickly if you use MATLAB interactively as you
are reviewing this tutorial. The M ATLAB commands will be shown in the following font style:

Monaco font

The prompt for a user input is shown by the double arrow (»). MATLAB has an extensive on-line
help facility. For example, type help pi at the prompt,

» help pi
PI 3.1415926535897....
PI = 4*atan(1) = imag(log(-1)) = 3.1415926535897....

so we see that MATLAB has the number p "built in." Notice that although MATLAB reports the
function in uppercase letters, lowercase is actually used when entering the function in the
command window.

[ Team LiB ]
[ Team LiB ]

M1.2 Matrix Operations

Matrix Notation Review

Matrices are ordered arrays of one or more dimensions. A one-dimensional matrix that can be
either a column or a row is known as a vector. In this text, we normally use subscripts to denote
the element numbers. Consider the following vectors of length 3, where c is a column vector and r
is a row vector.

Transposing each vector results in the following, where the superscript 'T' represents a transpose:

A 3 x 2 matrix has three rows and two columns, as shown by the matrix G:

Notice that the ij subscript refers to the element in the ith row and jth column.

Dimensional consistency is very important in matrix operations. Consider the matrix operation

To be dimensionally consistent, the number of columns in G must be equal to the number of rows
in H. When considering matrix multiplication, it is convenient to write the dimensions underneath
the matrices. For a G matrix with mg rows and ng columns, and an H matrix with nh columns,
then the rest of the dimensions must be
The ijth element of F is found in the following fashion:

For the following F and G matrices,

you should be able to calculate that

MATLAB Matrix Operations

The basic entity in M ATLAB is a rectangular matrix; the entries can be real or complex. Commas or
spaces are used to delineate the separate values in a matrix. Consider the following vector, x
(recall that a vector is simply a matrix with only one row or column), with six elements

» x = [1,3,5,7,9,11]
x =
1 3 5 7 9 11

Notice that a row vector is the default in MATLAB (in contrast to the default column vector used by
this and most other textbooks). Instead of commas, we could have used spaces as the delimiter
between columns

» x = [1 3 5 7 9 11]

There is a faster way to enter matrices or vectors that have a consistent pattern. For example,
the following command creates the previous row vector
» x = 1:2:11

where the first element is 1, the increment is 2, up to the final element of 11. Transposing a row
vector yields a column vector (' is the transpose command in MATLAB),

» y = x'
y =
1
3
5
7
9
11

If we want to make x a column vector, we use a semicolon as the delimeter between rows:

» x = [1;3;5;7;9;11]
x =
1
3
5
7
9
11

To make x a row vector again, we use the transpose

» x = x'

Say that we want to create a vector z, which has elements from 5 to 30, by increments of 5. To
suppress the listing of vector elements, we add a semicolon at the end (;):

» z = 5:5:30;

The z vector is generated but not printed in the command window. We can find the value of the
third element in the z vector, z(3), by typing
» z(3)
ans =
15

Matrix multiplication is straightforward. First, enter the two-dimensional matrix (here a semicolon
is used to begin a new row):

»a = [1 2;3 4; 5 6]
a =
1 2
3 4
5 6
» b = [1 2 3;4 5 6]
b =
1 2 3
4 5 6

Remember that the order of multiplication of matrices is important (AB BA, in general)

» c = a*b
c =
9 12 15
19 26 33
29 40 51
» d = b*a
d =
22 28
49 64

Sometimes it is desirable to perform an element-by-element multiplication rather than matrix

multiplication. For example, e(i,j) = a(i,j)*a(i,j) is performed by using the .* command

» e = a.*a
e =
1 4
9 16
25 36
Similarly, element-by-element division, b(i,j)/c(i,j), can be performed using ./

[ Team LiB ]
[ Team LiB ]

M1.3 The MATLAB Workspace

We can view the variables currently in the Workspace by typing

» who
Your variables are:
ans x y z
leaving 621420 bytes of memory free.

More detail about the size of the matrices can be obtained by typing whos.

We can find the size of a matrix or vector by typing

» [m,n]=size(x)
m =
1
n =
6

where m represents the number of rows and n represents the number of columns.

If we do not put place arguments for the rows and columns, we find

» size(x)
ans =
1 6

Since x is a vector, we can also use the length command

» length(x)
ans =
6
It should be noted that MATLAB is case sensitive with respect to variable names. An X matrix can
coexist with an x matrix.

Sometimes it is desirable to clear all of the variables in a Workspace. This is done by simply
entering

» clear

in the command window. More frequently you may wish to clear a particular variable, such as x

» clear x

This is particularly true if you are performing a new calculation of x and the new vector is shorter
than the old vector. If the new vector has length n, then all of the elements of the new x greater
than x(n) will contain values of the previous x vector.

[ Team LiB ]
[ Team LiB ]

M1.4 Complex Variables

Both i and j represent the imaginary number by default:

» i
ans =
0 + 1.0000i
» sqrt(-3)
ans =
0 + 1.7321i

Note that these variables (i and j) can be redefined (as the index in a for loop, for example).

[ Team LiB ]
[ Team LiB ]

M1.5 Plotting
MATLAB has many nice graphics capabilities. Here, we will illustrate standard two-dimensional
plots. First, generate two vectors for plotting

» x = 1:2:11;
» z = 5:5:30;

For a standard solid line plot, simply enter

» plot(x,z)

and Figure M1-2 appears. Axis labels are added by selecting "Insert" and specifying X label or Y
label. Alternatively, in the command window you can use the following commands (perform this):

» xlabel('x')
» ylabel('z')

Figure M1-2. Example plot.

For more plotting options, type

» help plot

If we wish to plot discrete points, using + as a symbol, we can use the following (perform this):

» plot(x,z,'+')

Notice that a new plot appears in the place of the previous one. You can generate a new figure by
entering figure(n), where n is any integer, before entering new plot commands.

Consider now the following equation:

We can plot this for a vector of t values using two simple commands,
» t = 0:1:50;
» y = 4*exp(-0.1*t);

and we can obtain a plot by entering in the command window (perform this)

» figure(2)
» plot(t,y)

Notice that we could shorten the sequence of commands by entering (perform this)

» plot(t,4*exp(-0.1*t))

To perform element-by-element multiplication of two vectors or arrays of the same size, we can
use the .* command. We can plot the function y(t) = t·e –0.1t by using (perform this)

» figure(3)
» y = t.*exp(-0.1*t);
» plot(t,y)
» xlabel('t')
» ylabel('y')

axis('square') will place the plot in a square box, while axis('normal') will change back to a
normal aspect ratio.

You can also explicitly set the upper and lower bounds on the plot with

axis([xlow xhigh ylow yhigh])

For this example, we would use

» axis([0 50 0 4]);

Multiple curves can be placed on the same plot in the following fashion (perform this).

» plot(t,4*exp(-0.1*t),t,t.*exp(-0.1*t),'--')

The subplot command can be used to make multiple plots, as illustrated in Figure M1-3.

» subplot(2,1,1), plot(t,4*exp(-0.1*t)), xlabel('t'),ylabel('y1')

» subplot(2,1,2), plot(t,t.*exp(-0.1*t)), xlabel('t'),ylabel('y2')

Figure M1-3. Use of subplot feature.

Here, subplot(i,j,k) means that there are i "rows" of figures, j "columns" of figures, and the
current plot is the kth figure (counting left to right and top to bottom). The following is an
example of a four plot sequence shown in Figure M1-4:
» t = 0:1:60;
» subplot(2,2,1),plot(t,4*exp(-0.1*t)), xlabel('t')
» subplot(2,2,2),plot(t,t.*exp(-0.1*t)), xlabel('t')
» subplot(2,2,3),plot(t,sin(.25*t)), xlabel('t')
» subplot(2,2,4),plot(t,cos(.25*t)), xlabel('t')

Figure M1-4. Use of subplot feature.

To return to single plots, simply enter subplot(1,1,1).

[ Team LiB ]
[ Team LiB ]

M1.6 More Matrix Stuff

A matrix can be constructed from two or more vectors. If we wish to create a matrix v which
consists of two columns, the first column containing the vector x (in column form) and the second
column containing the vector z (in column form), we can use the following (where x and z are
previously generated row vectors):

» v = [x',z']
v =
1 5
3 10
5 15
7 20
9 25
11 30

If we wished to look at the first column of v, we could use (where : indicates all rows, while the 1
indicates the first column)

» v(:,1)
ans =
1
3
5
7
9
11

If we wished to look at the second column of v, we could use

» v(:,2)
ans =
5
10
15
20
25
30
And we can construct the same plot as before by using the following plot command ('--' gives a
dashed line):

» plot(v(:,1),v(:,2),'--')

[ Team LiB ]
[ Team LiB ]

M1.7 For Loops

A for loop in MATLAB is similar to a DO Loop in FORTRAN. The main difference is that the
FORTRAN DO loop must have an integer index variable; for does not have this restriction. An
example of a for loop that is virtually identical to a DO loop is

» for k = 1:5001;
t(k) = (k-1)*0.01;
y(k) = sin(t(k));
end

Another way of implementing the same loop is to increment t from 0 to 50 in intervals of 0.01:

» k = 0
» for t = 0:0.01:50;
k = k + 1;
y(k) = sin(t);
end

The developers of MATLAB highly recommend that you use the vectorized version of the above for
loops:

t = 0:0.01:50;
y = sin(t);

since the computation time for this method is over 200 times faster than the nonvectorized
methods.

[ Team LiB ]
[ Team LiB ]

M1.8 m-Files
Thus far we have shown the interactive features of MATLAB by entering one command at a time.
One reason that MATLAB is powerful is that it is a language, and programs of MATLAB code can be
saved for later use. There are two ways of generating your own MATLAB code: script files and
function routines.

Script Files
A script file is simply a sequence of commands that could have been entered interactively in the
MATLAB command window. When the sequence is long or must be performed a number of times, it
is much easier to generate a script file.

The following example is for the so-called quadratic map (population model)

where xk represents the value of the population (dimensionless) at the kth time step. We have
titled the file popmod.m. Use the MATLAB text editor to generate this file and save it in a working
directory as popmod.m.

% popmod.m
% population model, script file example
%
clear x,k
n = input('input final time step ');
alpha = input('input alpha ');
xinit = input('input initial population ');
x(1) = xinit;
time(1)= 0;
for k = 2:n+1;
time(k) = k-1;
x(k) = alpha*x(k-1)*(1-x(k-1));
end
plot(time,x)
% end of script file example

Notice that we have used the MATLAB input function to prompt the user for data. Also note that a
percent sign (%) may be used to put comments in a script or function file. Any text after a % is
ignored by the program.

The file is run by simply entering

 » popmod

in the MATLAB command window.

Function Routines
A more powerful way of solving problems is to write MATLAB function routines. Function routines
are similar to subroutines in FORTRAN. Consider the previous example.

function [time,x] = pmod(alpha,xinit,n)

% population model example, pmod.m
clear time; clear x; clear k;
x(1) = xinit;
time(1)= 0;
for k = 2:n+1;
time(k) = k-1;
x(k) = alpha*x(k-1)*(1-x(k-1));
end
% end of function file example

where we generate and save a file titled pmod.m. We can now "run" this function routine (using
alpha = 2.8, xinit = 0.1, n = 30) by typing the following in the MATLAB command window:

» [tstep,xpop]=pmod(2.8,0.1,30);
» plot(tstep,xpop)

This function routine can also be called by other function routines. This feature leads to
"structured programming"; structured programs are easy to follow and debug.

Commonly Used MATLAB Functions

MATLAB has many built-in function routines that you will use throughout this text. The most
commonly used routines are fzero, fsolve, and ode45, used to solve a single nonlinear algebraic
equation, multiple nonlinear algebraic equations, and a set of nonlinear differential equations,
respectively. The optimization toolbox is required to use fzero and fsolve.

Single Nonlinear Algebraic Equation

fzero is used to solve a single nonlinear algebraic equation. As an example, consider

The first step is to generate a function file, which we will call fcn (any name can be used). For
this simple example, the function file can consist of two lines,

function y = fcn(x)
y = x^2 - 2*x -3;

and can be saved as fcn.m. The user must enter the following command in the command window,
where 0 is the initial guess. Notice that fzero calls the fcn routine, using an initial guess for the
solution,

» y = fzero(@fcn,0)

and the following answer is returned

y = -1

We know that there will be two solutions, since this is a quadratic equation. Indeed, an initial
guess of x = 2 yields the following result:

» y = fzero(@fcn,2)
y = 3.0000

Since the function is a polynomial, we can use a function developed for polynomial solutions. The
MATLAB function roots finds the solution to the quadratic equation, using

» roots([1 -2 -3])
ans =
 3
-1

which yields the two solutions expected.

[ Team LiB ]
[ Team LiB ]

M1.9 Summary of Commonly Used Commands

axis axis limits for plots
clear removes all variables from workspace
diary save the text of a MATLAB session
end end of loop
exp exponential function
figure(n) generates a figure window

for generates loop structure

format output display format
function user generated function
gtext place text on a plot
help help function
hold holds current plot and allows new plot to be placed on current plot
if conditional test
length length of a vector
lookfor keyword search on help variables
plot plots vectors
size size of the array (rows, columns)
subplot multiple plots in a figure window
who view variables in workspace
whos view variables in workspace, with more detail (size, etc.)
* matrix multiplication
' transpose
; suppress printing (also end of row, when used in matrices)
.* element-by-element multiplication
./ element-by-element division
: denotes a column or row in a matrix; also creates a vector
% placed before comment statements

[ Team LiB ]
[ Team LiB ]

M1.10 Frequently Used MATLAB Functions

Function Use

eig eigenvalues, eigenvectors

fsolve solve algebraic equations Optimization toolbox

fzero solve a single algebraic equation Optimization toolbox
impulse impulse response Control toolbox
ode45 integrate set of ordinary differential equations

polyfit least-squares fit of a polynomial

ss2tf convert state space to transfer function model Control toolbox

step step response Control toolbox
tf2ss convert transfer function to state-space model Control toolbox

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Plot the following three curves on a single plot and a multiple plots (using the subplot
command): 2 cos(t), sin(t), and cos(t)+sin(t). Use a time period such that two or three
peaks occur for each curve. Use solid, dashed, and + symbols for the different curves.
Use roughly 25–50 points for each curve.

2: a. Calculate the rank, determinant, and matrix inverse of the following matrices (use
help rank, help det, and help inv):

3: Find C·C–1 where

4: Calculate xTx, and calculate xxT where

5: Find the eigenvalues of the matrix

 6: Find the solutions to the equation f(x) = 3x3 + x2 + 5x-6 = 0. Use roots and fzero.

7: Integrate the equations, from t = 0 to t = 5

with the initial condition x1(0) = x2(0) = 1. Use ode45 (Module 3) and plot your results.

[ Team LiB ]
[ Team LiB ]

Module 2. Introduction to SIMULINK

Although the standard M ATLAB package is useful for linear systems analysis, SIMULINK is far more
useful for control system simulation. SIMULINK enables the rapid construction and simulation of
control block diagrams. The goal of the tutorial is to introduce the use of SIMULINK for control-
system simulation. The version available at the time of writing of this textbook is SIMULINK 5.0,
part of Release 13 (including MATLAB 6.5) from MATHWORKS. The version that you are using can
be obtained by entering ver in the M ATLAB command window.

The easiest way to learn how to use S IMULINK is to implement each step of the tutorial, rather than
simply reading it. The basic steps to using SIMULINK are independent of the platform (PC, MAC,
UNIX, ...).

The sections of this module are as follows:

M2.1 Background

M2.2 Open-Loop Simulations

M2.3 Feedback-Control Simulations

M2.4 Developing Alternative Controller Icons

M2.5 Summary

[ Team LiB ]
[ Team LiB ]

M2.1 Background
The first step is to start up MATLAB on the machine you are using. In the Launch Pad window of the
MATLAB desktop, select S IMULINK and then the SIMULINK Library Browser. A number of options are
listed, as shown in Figure M2-1 for SIMULINK 4 (Release 12); SIMULINK 5 (Release 13) has a number
of additional options. Notice that Continuous has been highlighted; this will provide a list of
continuous function blocks available. Selecting Continuous will provide the list of blocks shown in
Figure M2-2. The ones that we often use are Transfer Fcn and State-Space.

Figure M2-1. SIMULINK Library Browser (Version 4.1, Release 12.1).

Figure M2-2. SIMULINK Continuous blocks.

Selecting the Sources icon yields the library shown in Figure M2-3. The most commonly used
sources are Clock (which is used to generate a time vector) and Step (which generates a step
input).

Figure M2-3. SIMULINK Sources.

The Sinks icon from Figure M2-1 can be selected to reveal the set of sinks icons shown in Figure
M2-4. The one that we use most often is the To Workspace icon. A variable passed to this icon is
written to a vector in the MATLAB workspace. The default data method (obtained by "double-
clicking" on the icon after it is placed in a model worspace) should be changed from "structure"
to "array" in order to save data in an appropriate form for plotting.

Figure M2-4. SIMULINK Sinks.

[ Team LiB ]
[ Team LiB ]

M2.2 Open-Loop Simulations

You now have enough information to generate an open-loop simulation. The Clock,simout,step,
and Transfer Fcn blocks can be dragged to a model (.mdl) workspace, as shown in Figure M2-
5a. Renaming the blocks and variables, and connecting the blocks, results in the model shown in
Figure M2-5b. The transfer function studied is the Van de Vusse reactor (see Module 5).

Figure M2-5. Development of an open-loop simulation. (a) Placement

of function blocks. (b) Renaming and connection of blocks.

The s-polynomials in the process transfer function were entered by double clicking on the
transfer function icon and entering the coefficients for the numerator and denominator
polynomials. Notice also that the default step (used for the step input change) is to step from a
value of 0 to a value of 1 at t = 1. These default values can be changed by double clicking the
step icon. The simulation parameters can be changed by going to the Simulation "pull-down"
menu and modifying the stop time (default = 10) or the integration solver method (default
= ode45).
The reader should generate simulations and observe the "inverse response" behavior of the
output with respect to a step input change. Use the subplot command to place the process
output (y) on the top plot, and the manipulated input (u) on the bottom plot. Perform this now. If
desired, change the default simulation stop time by selecting the parameters "pull down" menu.

[ Team LiB ]
[ Team LiB ]

M2.3 Feedback-Control Simulations

The Math icon from Figure M2-2 can be selected, resulting in the functions shown in Figure M2-6.
Additional icons can be found by selecting the Simulink Extras icon shown in Figure M2-1.
Selecting the Additional Linear icon from this group yields the set of icons shown in Figure M2-
7. The most useful icon here is the PID Controller. Any icon can be "dragged" into the
untitled model workspace. In Figure M2-8, we show the preliminary stage of the construction of
a control block diagram, where icons have been dragged from their respective libraries into the
untitled model workspace.

Figure M2-6. SIMULINK Math.

Figure M2-7. SIMULINK Additional Linear.
Figure M2-8. Block diagram for feedback control of the Van de Vusse
CSTR. (a) Preliminary block diagram. (b) Completed block diagram,
with name and parameter changes.
The labels (names below each icon) can easily be changed. The default parameters for each icon
are changed by double clicking the icon and entering new parameter values. Also, connections can
be made between the outputs of one icon and inputs of another. Figure M2-8b shows how the
icons from Figure M2-8a have been changed and linked together to form a feedback-control block
diagram.

It should be noted that the form of the PID control law used by the S IMULINK PID Controller icon
is not the typical form that we use as process-control engineers. The form can be found by double
clicking the icon to reveal the controller transfer function representation

while we normally deal with the following PID structure:

So the numerical values for the following parameters must be entered in the SIMULINK PID
controller:
The s-polynomials in the process transfer function were entered by double clicking on the transfer
function and entering the coefficients for the numerator and denominator polynomials. Notice also
that the default step (used for the step setpoint change) is to step from a value of 0 to a value of
1 at t = 1. These default values can be changed by double clicking the step icon. The simulation
parameters can be changed by going to the Simulation "pull-down" menu and modifying the
start time (default = 0), stop time (default = 10) or the integration solver method (default =
ode45).

The controller tuning parameters of kc = 1.89 and tI = 1.23 are used by entering P = 1.89 and I
= 1.89/1.23 in the default PID Controller block. The following plot commands were used to
generate Figure M2-9:

» subplot(2,1,1),plot(t,r,'--',t,y)
» xlabel('t (min)')
» ylabel('y (mol/l)')
» subplot(2,1,2),plot(t,u)
» xlabel('t (min)')
» ylabel('u (min^-1)')

Figure M2-9. Measured output and manipulated input responses to a

unit step setpoint change.

The curves in Figure M2-9 could be made smoother by selecting the 'refine' box in the
Simulation pull-down menu. Selecting refine = 2 will plot two points for each integration step.
This brief tutorial has gotten you started in the world of SIMULINK-based control block diagram
simulation. You may now easily compare the effect of different tuning parameters or different
formulations of a PID controller ("ideal" vs. "real," for example).

Let's say you generated responses for a set of tuning parameters that we call case 1 for
convenience. You could generate time, input, and output vectors for this case by setting

» t1 = t;
» y1 = y;
» u1 = u;

after running the case 1 values. You could then enter case 2 values, run another simulation, and
create new vectors:

» t2 = t;
» y2 = y;
» u2 = u;

Then, compare case 1 and case 2 results by

» subplot(2,1,1),plot(t1,y1,t2,y2,'--')
» xlabel('t (min)')
» ylabel('y (mol/l)')
» subplot(2,1,2),plot(t1,u1,t2,u2,'--')
» xlabel('t (min)')
» ylabel('u (min^-1)')

which automatically plots case 1 as a solid line and case 2 as a dashed line.

Similarly, you could modify the controller type by placing a transfer function block for the
controller and using a "real PID" transfer function (this only differs when there is derivative
action).

Other Commonly Used Icons

Often you will want to simulate the behavior of systems that have time delays. The Transport
Delay icon can be selected from the Continuous library shown in Figure M2-2. The transport
delay icon is shown in Figure M2-10. Our experience is that simulations can become somewhat
"flaky" if 0 is entered for a transport delay. We recommend that you remove the transport delay
block for simulations where no time delay is involved.

Figure M2-10. Transport Delay icon.

Manipulated variables are often constrained to between minimum (0 flow, for example) and
maximum (fully open valve) values. A saturation icon from the Discontinuities library can be
used to simulate this behavior. The saturation icon is shown in Figure M2-11.

Figure M2-11. Saturation element.

Actuators (valves) and sensors (measurement devices) often have additional dynamic lags that
can be simulated by transfer functions. These can be placed on the block diagram in the same
fashion that a transfer function was used to represent the process earlier.

It should be noted that icons can be "flipped" or "rotated" by selecting the icon and going to the
format "pull-down" menu and selecting Flip Block or Rotate Block. The block diagram of
Figure M2-8 has been extended to include the saturation element and transport delay, as shown
in Figure M2-12.

Figure M2-12. Block diagram with saturation and time-delay elements.

The default data method for the "to workspace" blocks (r,t,u,y in Figure M2-12) must be
changed from "structure" to "array" in order to save data in an appropriate form for plotting.

[ Team LiB ]
[ Team LiB ]

M2.4 Developing Alternative Controller Icons

It was noted earlier that the default SIMULINK PID controller block uses a form different from that
used by most process engineers. It is easy to generate new PID controller blocks as shown below.
The default PID controller icon is shown in Figure M2-13a. This is "unmasked" (by clicking the icon
and selecting 'Look under mask' from the Edit pull-down menu) to yield the diagram shown in
Figure M2-13b. Again, this has the form

Figure M2-13. Default ideal PID controller. (a) Block. (b) Block
unmasked.

while we normally prefer the following PID structure:

Of course, the two algorithms are related by

but it would be much less confusing to work with our standard form.

We have generated a new ideal analog PID, as shown in Figure M2-14a. Notice that there are two
inputs to the controller, the setpoint (r) and the measured output (y), rather than just the error
signal that is the input to the default SIMULINK PID controller. Our new implementation is
unmasked in Figure M2-14b to reveal that

Figure M2-14. Preferred implementation of ideal PID controller. (a)

Block. (b) Unmasked block.

is the algorithm used.

[ Team LiB ]
[ Team LiB ]

M2.5 Summary
SIMULINK is a very powerful block diagram simulation language. Simple simulations, including the
majority of those used as examples in this textbook, can be set up rapidly (in a matter of
minutes). The goal of this module was to provide enough of an introduction to get you started on
the development of open- and closed-loop simulations. With experience, the development of these
simulations will become second nature. It is recommended that you perform the simulations
shown in this module, as well as the additional exercises, to rapidly acquire these simulation skills.

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Compare step responses of the transfer function and state space models for the van de
Vusse reactor.

a. Transfer function: Perform the open-loop simulation for a step input change from 0
to 1 at t = 1, as shown in Figure M2-5.

b. State space: Replace the transfer function block shown in Figure M2-5 with a state
space block. Enter the following matrices in the MATLAB command window:

Show that the resulting step responses of the transfer function and state space models
are identical.

2: Consider the closed-loop SIMULINK diagram for the van de Vusse reactor (Figure M2-8).
Replace the default PID controller with a "Real PID" controller, as shown below in both
the masked and unmasked versions. How do your results compare with ones obtained
using the default PID controller?

3: The following control algorithm uses a filtered derivative of the process output (see
Chapter 5). Develop a SIMULINK implementation of this algorithm, where e(s) = r(s) -
y(s).
[ Team LiB ]
[ Team LiB ]

Module 3. Ordinary Differential Equations

MATLAB provides several powerful approaches to integrating sets of initial-value, ordinary
differential equations. In this module, we will discuss two approaches: using functions called by
various ode routines from a MATLAB command window or script file, and using functions called by a
block from a SIMULINK diagram.

After reviewing this module, the reader should be able to do the following:

Write a basic function file containing differential equations to be integrated by MATLAB ode
routines

Write a more advanced function file, with parameters based through an argument list and
containing differential equations to be integrated by MATLAB ode routines

Write a function (more specifically, an S-function) file containing differential equations to be

integrated by SIMULINK

Generate the SIMULINK block diagram containing an S-function block

The sections of this module are as follows:

M3.1 MATLAB ode—Basic

M3.2 MATLAB ode—Options

M3.3 SIMULINK sfun (.mdl Files)

M3.4 SIMULINK sfun (.mdl Files)—Advanced

M3.5 Summary

[ Team LiB ]
[ Team LiB ]

M3.1 MATLAB ode—Basic

MATLAB has several different routines for numerical integration; all use the same basic call
statements:

[t,x]= ode45('xprime',tspan,x0) % nonstiff, medium order

[t,x]= ode23('xprime',tspan,x0) % nonstiff, low order
[t,x]= ode113('xprime',tspan,x0) % nonstiff, variable order
[t,x]= ode15s('xprime',tspan,x0) % stiff, variable order
[t,x]= ode23s('xprime',tspan,x0) % stiff, low order
[t,x]= ode23t('xprime',tspan,x0) % moderately stiff, trapezoidal
[t,x]= ode23tb('xprime',tspan,x0) % stiff, low order

ode45 uses fourth-order and ode23 uses second-order Runge-Kutta integration. All routines use a
variable integration step size (Dt is not constant). The integration step size is adjusted by the
routine to provide the necessary accuracy, without taking too much computation time.

To use these routines, the user must first generate an m-file to evaluate the state variable
derivatives. The m-file must be named 'xprime.m', and contain the following statement in the first
line of the file

function xdot = xprime(t,x);

where 'xprime' is the name of the function routine (usually selected as something meaningfully
related to the system of equations), 'xdot' is the vector of time derivatives of the states, 't' is
time, and 'x' is the vector of states. 'xdot' must be generated as a column vector.

The command given to perform the integration is

[t,x]= solver('xprime',tspan,x0)

where
solver is any of the integration routines (ode45, etc.)

xprime is a string variable containing the name of the m-file for the state derivatives
tspan is the vector of time values (row vector)
x0 is the initial condition vector for the state variables (column vector).

The arrays that are returned are

t a (column) vector of time

x an array of state variables as a function of time (column 1 is state 1, etc.).

For example, if the time vector has 50 elements and there are three state variables, then the
state variable vector has the 50 rows and three columns. After the integration is performed, if the
student wishes to plot all three variables as a function of time, she/he simply enters

plot(t,x)

If you only want to plot the second state variable, then the command plot(t,x(2,:)) is given.

Example M3.1: Van de Vusse Reaction

Consider the following set of differential equations that describe the van de Vusse reaction
scheme in an isothermal, CSTR.

The parameter values are

The input values used in the following simulation are

The differential equations are placed in a file named vdv_ode.m

 function xdot = vdv_ode(t,x);
%
% Solves the two differential equations modeling
% the van de vusse reaction
% scheme in an isothermal CSTR. The states are the concentration
% of A and B in the reactor.
%
% [t,x] = ode45(vdv_ode,[0 5],x0)
% integrates from t = 0 to t = 5 min, with initial conditions
% ca0 = x0(1) and cb0 = x0(2), and x0 is a column vector
% 16 Jan 99
% b.w. bequette
%
% since the states are passed to this routine in the x vector,
% convert to natural notation

ca = x(1);
cb = x(2);

% the parameters are:

k1 = 5/6; % rate constant for A-->B (min^-1)
k2 = 5/3; % rate constant for B-->C (min^-1)
k3 = 1/6; % rate constant for 2A-->D (mol/(l min))
% the input values are:

fov = 4/7; % dilution rate (min^-1)

caf = 10; % mol/l

% the modeling equations are:

dcadt = fov*(caf-ca) - k1*ca –k3*ca*ca;
dcbdt = -fov*cb + k1*ca – k2*cb;

% now, create the column vector of state derivatives

xdot = [dcadt;dcbdt];

% end of file

In the MATLAB command window, enter the initial conditions and run ode45

» x0 = [2;1.117]
» [t,x] = ode45('vdv_ode',[0 5],x0);
» subplot(2,1,1),plot(t,x(:,1)), xlabel('t'), ylabel('ca')
» subplot(2,1,2),plot(t,x(:,2)), xlabel('t'), ylabel('cb')

The plots are shown in Figure M3-1. Notice that the system converges to the steady-state values
of CA=3, CB=1.117.

Figure M3-1. Transient response to initial conditions perturbed from

 the steady-state values, CA(0) = 2, CB(0) = 1.117.

[ Team LiB ]
[ Team LiB ]

M3.2 MATLAB ode—Options

The M ATLAB integration routines have additional options (such as changing the default tolerances
for integration) that can be set by the user. Also, parameters for the differential equations may be
passed through the argument list to the function m-file.

[t,x]= ode45('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode23('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode113('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode15s('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode23s('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode23t('xprime',tspan,x0,options,P1,P2,...)
[t,x]= ode23tb('xprime',tspan,x0,options,P1,P2,...)

To use the default options (generally recommended) and pass through parameter values, simply
enter [ ] for the options element in the argument list. For example, create an 'xprime.m' m-file
for the differential equation

function xdot = xprime(t,x,P1,P2,...)

...

and enter

[t,x]= ode45('xprime',tspan,x0,[ ],P1,P2,...);

in the MATLAB command window.

A typical application might be to be able to change the magnitude of an input variable. For
example, consider the van de Vusse reactor shown in Example M3.1. The first line of the
vdv_ode.m file could be modified in the fashion

function xdot = vdv_ode(t,x,delstep);

where delstep represents the change in the manipulated input. The file is further modified by
changing the following line

fov = 4/7 + delstep; % dilution rate (min^-1)

so that delstep represents a change from the nominal input value of 4/7 min-1. If the magnitude
of the step change is 0.1, then the line entered in the MATLAB command window is

[t,x]= ode45('vdv_ode',[0 5],x0,[ ],0.1)

[ Team LiB ]
[ Team LiB ]

M3.3 SIMULINK sfun (.mdl Files)

One of the powers of SIMULINK is the ability to construct block diagrams that provide an intuitive
approach to simulating feedback-control systems. Often we use linear transfer functions to
simulate the behavior of linear systems. A block titled "sfun" can obtained from the "nonlinear"
library in SIMULINK and placed in a SIMULINK diagram. When simulations are performed, the
equations in the sfun m-file are solved simultaneously with other blocks in the block diagram. The
sfun file must be given a name with a '.m' extension, while the block diagram is given a '.mdl'
extension.

Example M3.2: Van de Vusse Reaction

The van de Vusse reactor from Example M3.1 is studied here. The block diagram shown in Figure
M3-2 was first generated. Notice that there are two inputs and two outputs shown; mux and
demux blocks are used since inputs and outputs are passed to and returned from S-Functions as
vectors. Notice that the S-Function has been titled vdv_sfun.m. The SIMULINK block diagram
should also be titled and given a .mdl extension.

Figure M3-2. SIMULINK diagram for an open-loop simulation of the van

de Vusse reactor.

The vdv_sfun function file is shown below. The vectors 't', 'x', and 'u' contain time, states, and
inputs, respectively. The 'flag' parameter is passed from S IMULINK, depending on the simulation
status. A value of flag = 0, for example, sets up the simulation with information about the
number of states, and so forth. A value of flag = 1 solves for the derivatives of the states with
respect to time and passes these derivatives in the sys vector back to SIMULINK. A value of flag =
3 passes the values of the outputs back to SIMULINK through the sys vector.

Before running the simulation, double click on the S-Function block to see what parameters or
initial conditions must be set before the simulation. In this case, the initial conditions must be set
in the MATLAB Workspace, using the xi variable. Also, go to the "parameters" pull-down menu in
SIMULINK to set the final time for integration and the differential equation solution method.
For the simulation shown, enter

» x0 = [2;1.117]

to obtain the same plot shown in Figure M3-1.

function [sys,x0,str,ts] = vdv_sfun(t,x,u,flag,x0)

%
% Solves the two differential equations modeling
% the van de vusse reaction
% scheme in an isothermal CSTR. The states are the concentration
% of A and B in the reactor.
%
% Use this as an sfun function in a simulink diagram
%
% 23 Jan 99
% b.w. bequette

% process condition:
% dilution rate (fov) = 4/7;
% feed concentration of a (caf) = 10;
% x1s = cas = 3.0;
% x2s = cbs = 1.117;
% x0 = [x1i;x2i] = [cai;cbi] = initial conditions;

switch flag

case 0 % initialization;

sizes = simsizes;
sizes.NumContStates = 2;
sizes.NumDiscStates = 0;
sizes.NumOutputs = 2;
sizes.NumInputs = 2;
sizes.DirFeedthrough = 0;
sizes.NumSampleTimes = 1;
sys = simsizes(sizes);

str = [];
ts = [0 0];

case 1 % derivatives;

% it is convenient to use common notation for states

ca = x(1);
cb = x(2);
% also, the inputs are

fov = u(1);
caf = u(2);

% parameter values

k1 = 5/6; % rate constant for A-->B (min^-1)

k2 = 5/3; % rate constant for B-->C (min^-1)
k3 = 1/6; % rate constant for 2A-->D (mol/(l min))

% state derivatives

dcadt = fov*(caf-ca) - k1*ca -k3*ca*ca;

dcbdt = -fov*cb + k1*ca - k2*cb;

sys = [dcadt;dcbdt];

case 3 % outputs;

sys = [x(1);x(2)];

case {2, 4, 9}
sys = [];

otherwise
error(['unhandled flag = ',num2str(flag)]);

end

[ Team LiB ]
[ Team LiB ]

M3.4 SIMULINK sfun (.mdl Files)—Advanced

The previous example of the use of S IMULINK and S-functions was quite limited. In the example,
process parameters were imbedded in the S-function (.m file). Parameters can be passed through
argument lists, so that the same files can be used for many different systems by simply changing
parameter values. Some more advanced features are illustrated in the example below.

Example M3.3: Van de Vusse Reactor Extended

The SIMULINK diagram shown in Example M3.2 (Figure M3-2) merely re-created the same
simulation (perturbation of initial conditions from steady state) as the MATLAB ode file from
Example M3.1. A real power of SIMULINK is to create complex block diagrams with additional lags
on sensors and actuators and to apply changes to the input variables. The SIMULINK .mdl file
shown in Figure M3-2 is revised, as shown in Figure M3-3. The "Van de Vusse Reactor" block is
actually a subsystem that contains the S-function to simulate the reactor, and additional lags for
sensors and actuators. Double clicking to "open up" this subsystem yields the diagram shown in
Figure M3-4. Constraints are placed on the input variables to assure that they remain within
physical bounds. Notice that the lags are represented by special transfer function blocks that do
not require the inputs and outputs of the blocks to be in deviation variable form. These special
blocks can be obtained by opening the Blocksets & Toolboxes block on the Library:
simulink, then choosing simulink extras and additional linear.

Figure M3-3. SIMULINK diagram for an open-loop simulation of the van

de Vusse reactor. The simulation diagram of Figure M3-2 has been
extended. The "Van de Vusse Reactor" block is composed of a
"subsystem" shown in Figure M3-4.

Figure M3-4. The subsystem of the block shown in Figure M3-3. Notice
that addition actuator and sensor lags have been placed on the
diagram.
Kinetic parameters are passed into the S-function file by adding the parameters after x0 in the
argument list:

function [sys,x0,str,ts] = vdv_sfun(t,x,u,flag,x0,k1,k2,k3)

The values of the kinetic parameters are then entered in the MATLAB command window before
beginning the simulation.

The simulation diagram of Figure M3-3 is "run" by entering the initial conditions in a MATLAB
command window:

» x0 = [3.0;1.117];

Setting the dilution rate step function to move from 4/7 to 4/7 + 0.1 at t = 1 minute, run the
simulation, then enter the following commands to obtain the plots shown in Figure M3-5.

» subplot(2,1,1),plot(t,ca)
» xlabel('time, min')
» ylabel('ca, mol/liter')
» subplot(2,1,2),plot(t,cb)
» xlabel('time, min')
» ylabel('cb, mol/liter')

Figure M3-5. Responses of the concentration of A and B to a small step

change in the dilution rate (from 4/7 to 4/7 + 0.1 min-1).
Notice the characteristic inverse response shown in bottom plot.

[ Team LiB ]
[ Team LiB ]

M3.5 Summary
We have shown how to write function files to be used by MATLAB and SIMULINK to integrate sets of
ordinary differential equations. It is highly recommended that you work through this tutorial
interactively, generating the suggested .m files and .mdl diagrams. For more information on
options used by the M ATLAB ode routines, enter the following in the command window:

help ode45

[ Team LiB ]
[ Team LiB ]

Module 4. MATLAB LTI Models

The M ATLAB Control Toolbox stores linear time invariant (LTI) models as objects. These objects
can be manipulated (added, multiplied) to form more complex objects. Models can easily be
transformed from one form to another (state space to transfer function, for example). Continuous
models can also be transformed to discrete (and vice versa).

The sections of this module are as follows:

M4.1 Forming Continuous-Time Models

M4.2 Forming Discrete-Time Models

M4.3 Converting Continuous Models to Discrete

M4.4 Converting Discrete Models to Continuous

M4.5 Step and Impulse Responses

M4.6 Summary

The majority of the functions that are used in this textbook are in Section M4.1.

[ Team LiB ]
[ Team LiB ]

M4.1 Forming Continuous-Time Models

Continuous-time models can be specified in one of three forms: state space (ss), transfer
function (tf), and zero-pole gain (zpk).

The commands for forming these models are

sys = ss(a,b,c,d) % state space

sys = tf(num,den) % transfer function
sys = zpk(z,p,k) % zero-pole-gain

where sys is the desired name of the model formed in each case. The a,b,c,d terms are the
state space matrices. The num and den terms are vectors of polynomial coefficients, in decending
order of powers in s. The z and p terms are row vectors of zero and pole locations (roots of
numerator and denominator polynomials, respectively), while k is the gain term (note that this is
not necessarily equal to the gain of the transfer function as we know it).

We use the van de Vusse reactor example from Module 5 to illustrate the use of each. The
following sections illustrate how state space, transfer function, and zero-pole gain models are
directly entered. Then we illustrate how to convert between the different model forms.

State Space
The state space model is

where the parameter values for the van de Vusse CSTR result in

The matrices are each entered in the MATLAB Workspace:

» a = [-2.4048,0;0.8333,-2.2381];
» b = [7;-1.117];
» c = [0,1];
» d = [0];

The state space object command is used (note that we name our system vdv_ss for
convenience):

» vdv_ss = ss(a,b,c,d)

The model is reported in the MATLAB command window:

a =
x1 x2
x1 -2.4048 0
x2 0.8333 -2.2381

b =
u1
x1 7
x2 -1.117

c =
x1 x2
y1 0 1

d =
u1
y1 0

Continuous-time system.

The poles and zeros can be calculated using the pole and tzero commands.

» pole(vdv_ss)

ans =
-2.2381
-2.4048

» tzero(vdv_ss)

ans =
2.8173
Clearly, the system is stable (the poles of the transfer function = the eigenvalues of the A matrix
= negative, indicating stability) and contains a RHP zero, indicating an inverse response.

Transfer Function
The transfer function model is

The numerator and denominator polynomials are entered:

» num = [-1.117,3.1472];
» den = [1,4.6429,5.3821];

The transfer function object command is entered:

» vdv_tf = tf(num,den)

The result is displayed:

Transfer function:
-1.117 s + 3.147
---------------------
s^2 + 4.643 s + 5.382

» pole(vdv_tf)

ans =
-2.4048
-2.2381

» tzero(vdv_tf)

ans =
2.8173
Again, the RHP zero indicates an inverse response.

Zero-Pole Gain
This structure is not often used by process engineers. It is similar to the transfer function form,
but the polynomials have been factored to clearly show the zeros and poles. An example with one
zero and two poles is

Notice that the gain term, k, is not the same as the process gain and is definitely a source of
confusion.

Consider the transfer function with the following zero-pole-gain form (you should be able to write
the previous transfer function in this form):

The M ATLAB command, followed by the result displayed, is

» vdv_zpk = zpk(2.817,[-2.238 -2.405],-1.117)

Zero/pole/gain:
-1.117 (s-2.817)
-------------------
(s+2.238) (s+2.405)

Converting Between Model Types

A model in any one of the standard forms can easily be converted to another standard form.

Converting from State Space Form

As an example, assume that the state space form has been entered as

» a = [-2.4048,0;0.8333,-2.2381];
» b = [7;-1.117];
» c = [0,1];
» d = [0];
» vdv_ss = ss(a,b,c,d)

The transfer function form can found from the state space form using

» vdv_tf = tf(vdv_ss)

and the MATLAB command window displays the following:

Transfer function:
-1.117 s + 3.147
---------------------
s^2 + 4.643 s + 5.382

Also, the zero-pole-gain form can be found from the state space form using

» vdv_zpk = zpk(vdv_ss)

Zero/pole/gain:
-1.117 (s-2.817)
-------------------
(s+2.238) (s+2.405)

An alternative approach is to use ss2tf or ss2zp,

[num,den] = ss2tf(a,b,c,d,iu)
[z,p,k] = ss2zp(a,b,c,d,iu)

where iu is the ith input variable (2, if you want the transfer functions relating input 2 to all of
the outputs). If there is more than one output, then num will be a matrix of elements, where each
row contains the coefficients of the polynomial associated with that output (row 1 is the
numerator polynomial for output 1, etc.).

Converting from Transfer Function to State Space Form

Assume that the transfer function form has been created:

» num = [-1.117,3.1472];
» den = [1,4.6429,5.3821];
» vdv_tf = tf(num,den);

The state space form can be found by using

»vdv_ss1 = ss(vdv_tf)

a =
x1 x2
x1 -4.6429 -1.3455
x2 4 0

b =
u1
x1 1
x2 0

c =
x1 x2
y1 -1.117 0.7868

d =
u1
y1 0

Continuous-time system.

Recall that there are many state space models that yield the same transfer function model. The
state space realization of a transfer function, used by MATLAB, is known as the controllable
canonical form. Notice that this form in this example (vdv_ss1) is different than the original state
space form (vdv_ss). You should show that either state space form results in the same transfer
function.

Multiple Inputs and/or Outputs

The above examples have illustrated LTI objects that are SISO. The extension to MIMO is
straightforward for the state space model (simply increase the dimensions of the matrices
appropriately.

For transfer functions, the easiest way to form a MIMO transfer function matrix is to concatenate
the individual matrix elements. Consider
Enter all the individual transfer functions, then specify the elements in the transfer function
matrix:

» g11 = tf([2],[3 1]);

» g12 = tf([-1],[4 1]);
» g21 = tf([-1],[2 1]);
» g22 = tf([5],[6 1]);
» G = [g11 g12;g21 g22]

MATLAB returns the following:

Transfer function from input 1 to output...

2
#1: -------
3 s + 1

-1
#2: -------
2 s + 1

Transfer function from input 2 to output...

-1
#1: -------
4 s + 1

5
#2: -------
6 s + 1

Input Time Delays

Most chemical processes have significant time delays that can cause control problems. The LTI
objects assume that the time delay is on the inputs. This means that each column of a transfer
function matrix, for example, is assumed to have the same time delay. Consider the previous two
input–two output transfer function matrix, with time delays of 1.5 and 2.5 time units on input 1
and 2, respectively:
Since the delay-free transfer function has already been generated, we can add input delays using
the following:

» set(G,'InputDelay',[1.5 2.5])

The new transfer function properties can be found using

» G

Transfer function from input 1 to output...

2
#1: -------
3 s + 1

-1
#2: -------
2 s + 1

Transfer function from input 2 to output...

-1
#1: -------
4 s + 1

5
#2: -------
6 s + 1

Input delays (listed by channel): 1.5 2.5

Delays to state-space models can be handled in a similar fashion.

[ Team LiB ]
[ Team LiB ]

M4.2 Forming Discrete-Time Models

The formation of discrete-time models is the same as for continuous-time models, except that a
sample time should also be specified. This sample time is simply the next argument in the
function call statement,

sys = ss(a,b,c,d,Ts) % discrete state space

sys = tf(num,den,Ts) % discrete transfer function
sys = zpk(z,p,k,Ts) % discrete zero-pole-gain

where Ts is the sample time.

Discrete State Space Models

A discrete-time state space model has the form

where k represents the discrete-time step index. Consider the discrete state space model

with a sample time of Dt = 0.1 minutes. Please note that, if you need to convert a continuous
state space model to discrete, see Section M4.3.

After the a, b, c and d matrices are entered, the state space object can be created:

» vdv_ssd = ss(a,b,c,d,0.1)

a =
x1 x2
x1 0.78625 0
x2 0.06607 0.79947

b =
 u1
x1 0.62219
x2 -0.07506

c =
x1 x2
y1 0 1

d =
u1
y1 0

Sampling time: 0.1

Discrete-time system.

If you have a discrete-time model but do not know the sample time, enter –1 for Ts.

»vdv_ssd1 = ss(a,b,c,d,-1)

a =
x1 x2
x1 0.78625 0
x2 0.06607 0.79947

b =
u1
x1 0.62219
x2 -0.07506

c =
x1 x2
y1 0 1

d =
u1
y1 0

Sampling time: unspecified

Discrete-time system.

The poles and zeros are found using the same commands as for continuous systems.

» pole(vdv_ssd)
ans =
 0.7995
0.7863
» tzero(vdv_ssd)
ans =
1.3339

The poles are inside the unit circle, indicating stability. The zero is outside the unit circle,
indicating that the model inverse is unstable.

Discrete Transfer Function

The discrete transfer function model is

which represents the following input-output model

The transfer function object is created

» vdv_tfd = tf([-0.07506 0.1001],[1 -1.586 0.6286],0.1)

Transfer function:
-0.07506 z + 0.1001
----------------------
z^2 - 1.586 z + 0.6286

Sampling time: 0.1

» pole(vdv_tfd)
ans =
0.7995
0.7863

» tzero(vdv_tfd)
ans =
1.3339

Again, the poles are inside the unit circle, indicating stability. The zero is outside the unit circle,
indicating that the model inverse is unstable.

Discrete Filter Form

It is common to use the backward shift notation for discrete chemical process models. There is an
additional Control Toolbox LTI object specifically for this.

sys = filt(num,den,Ts) % discrete transfer function, z^-1

Consider the transfer function in backward shift form:

The M ATLAB command and results are

» vdv_ff = filt([0 0.1564 0.2408],[1 -0.3513 0.0307],0.1)

Transfer function:
0.1564 z^-1 + 0.2408 z^-2
-----------------------------
1 - 0.3513 z^-1 + 0.0307 z^-2

Sampling time: 0.1

Note that with the discrete filter form, it is critical to have the leading zero element in the
numerator polynomial.

Converting Between Discrete Model Types

A model in any one of the standard discrete forms can easily be converted to another standard
discrete form.

Converting from State Space Form to Transfer Function Form

As an example, assume that the discrete state space form has been entered as vdv_ssd. The
discrete transfer function is found using tf:

» vdv_tfd1 = tf(vdv_ssd)
Transfer function:
-0.07506 z + 0.1001
----------------------
z^2 - 1.586 z + 0.6286

Sampling time: 0.1

[ Team LiB ]
[ Team LiB ]

M4.3 Converting Continuous Models to Discrete

The Control Toolbox can also easily convert continuous models to discrete, using the c2d
command.

sysd = c2d(sysc,Ts,method) % continuous to discrete

where sysc is the LTI object for a continuous model, sysd is the discrete object to be created, Ts
is the sample time, and method is the type of hold placed on the input variable. We always
assume a zero-order hold ('zoh'); this is also the default if no entry is made for method.

Consider the continuous state space model for the van de Vusse reactor, vdv_ss. It can be
converted to the discrete-time model using the following MATLAB commands:

» vdv_ssd = c2d(vdv_ss,0.1,'zoh')

a =
x1 x2
x1 0.78625 0
x2 0.066067 0.79947

b =
u1
x1 0.62219
x2 -0.075061

c =
x1 x2
y1 0 1

d =
u1
y1 0

Sampling time: 0.1

Discrete-time system.

[ Team LiB ]
[ Team LiB ]

M4.4 Converting Discrete Models to Continuous

Discrete-time models are often developed based on the sampled inputs and outputs, using
parameter estimation. Control system design, however, is usually performed based on continuous
models. It is important, then, to be able to convert discrete models to continuous. Fortunately,
the Control Toolbox makes this easy via the d2c command.

sys = d2c(sysd, method) % discrete to continuous

where sysd is the LTI object for a discrete model, sysc is the continuous object to be created,
and method is the type of hold placed on the input variable. We always assume that a zero-order
hold ('zoh') is used for the discrete model; this is also the default if no entry is made for method.

Consider the discrete transfer function model for the van de Vusse reactor, vdv_tfd. It can be
converted to the continuous-time model using the following MATLAB command:

» vdv_tfc = d2c(vdv_tfd,'zoh')

Transfer function:
-1.117 s + 3.147
---------------------
s^2 + 4.643 s + 5.382

[ Team LiB ]
[ Team LiB ]

M4.5 Step and Impulse Responses

Step responses of LTI objects can be determined or compared in a number of different ways. The
following command compares the step responses of the continuous and discrete-time van de
Vusse models previously entered in the MATLAB command window.

» step(vdv_tfc,vdv_tfd)

The resulting response is shown in Figure M4-1. Notice that a zero-order hold has been applied to
the discrete step response. Also, note that the step command assumes that the time unit is
seconds, which is not normally the case for chemical process systems.

Figure M4-1. Step responses of continuous and discrete van de Vusse

reactor model. Illustration of zero-order hold. Notice the default
labeling of a time unit of seconds.

Alternatively, the following sequence of commands can be used to generate Figure M4-2:

» [yc,tc] = step(vdv_tfc);
» [yd,td] = step(vdv_tfd);
» plot(tc,yc,td,yd,'o')
» xlabel('time')
» ylabel('y')
» title('step response of van de Vusse reactor models')
» legend('continuous','discrete')

Similarly, the impulse responses shown in Figure M4-3 are generated using

» impulse(vdv_tfc,vdv_tfd)

Notice that the tremendous difference in the impulse response for the continuous and discrete-
time models is due to the pulse implementation on the discrete-time model.

Figure M4-2. Step responses of continuous and discrete van de Vusse

reactor model.

Figure M4-3. Impulse responses of continuous and discrete van de

Vusse reactor model. Notice the default labeling with a time unit of
seconds.
[ Team LiB ]
[ Team LiB ]

M4.6 Summary
The M ATLAB Control Toolbox can be used to create LTI objects. These objectives can then easily be
converted to other forms. Conversion from continuous to discrete is particularly useful when
implementing a digital control algorithm. The commands used include the following:

sys = ss(a,b,c,d) % state space form

sys = tf(num,den) % transfer function form
sys = zpk(z,p,k) % zero-pole-gain form
sysd = filt(num,den,Ts) % backward shift form
sysd = c2d(sys,Ts,'zoh' % continuous to discrete
sys = d2c(sysd,'zoh') % discrete to continuous
pi = pole(sys) % poles
zi = tzero(sys) % zeros
[y,t] = step(sys) % step response
[y,t] = impulse(sys) % impulse response

In addition, the following do not require the Control Toolbox:

[num,den] = ss2tf(a,b,c,d,iu) % state space to transfer function

[a,b,c,d] = tf2ss(a,b,c,d) % transfer function to state space

Examples for the most important commands can be found in Section M4.1.

[ Team LiB ]
[ Team LiB ]

Reference

More details on converting between model types can be found in the following: Bequette, B. W.,
Process Dynamics: Modeling, Analysis and Simulation, Prentice Hall, Upper Saddle River, NJ
(1998).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Consider the continuous state space model, where the time unit is minutes

a. Find the eigenvalues of the A matrix.

b. Find the continuous-time transfer function and calculate the poles and zeros. Is the
inverse of this transfer function stable?

c. Find the discrete-time transfer function for a sample time of 0.1 minutes. What do
you notice about the number of zeros? Is the inverse of this transfer function
stable for this sample time? You should increase the sample time to 1.8399
minutes to find that the transmission zero shifts to the right-half-plane.

d. Compare the step responses of the continuous model with the discrete models.

e. Find the "filter" form (backward shift notation) of the discrete transfer function for
sample times of 0.1 and 1.

2: Show that the following three state space models result in the same transfer function
model.

a. Model 1

b. Model 2

c. Model 3
 c.

[ Team LiB ]
[ Team LiB ]

Module 5. Isothermal Chemical Reactor

There are a number of important concepts that are presented in this module.

The nonlinear concept of input multiplicity

The challenge of inverse response behavior

The Ziegler-Nichols closed-loop oscillation tuning method

IMC

This module contains too many topics to cover interactively during a typical 2-hour session. It is
recommended that the reader (or instructor) select a topic of particular interest. The two different
controller design sections (M5.4 and M5.5) can be studied independently. The approximate time
needed and the focus (in parentheses) of each section of this module are as follows:

M5.1 Background

M5.2 Model (Chapter 2)

M5.3 Steady-State and Dynamic Behavior (Chapter 3)—1–1.5 Hours

M5.4 Classical Feedback Control (Chapters 5 and 6)—1.5 Hours

M5.5 Internal Model Control (Chapter 8)—3 Hours

[ Team LiB ]
[ Team LiB ]

M5.1 Background
The reactant conversion in a chemical reactor is a function of the residence time or its inverse,
the space velocity. For an isothermal CSTR, the product concentration can be controlled by
manipulating the feed flow rate, which changes the residence time (for a constant volume
reactor). The feedback strategy is shown in the schematic diagram shown in Figure M5-1.

Figure M5-1. Feedback strategy for the isothermal reactor.

Here, we consider a series-parallel reaction of the following form (known as the van de Vusse
reaction scheme):

The desired product is the component B, the intermediate component in the series reaction. In
this module, we find interesting steady-state and dynamic behavior that can occur with this
reaction scheme. Klatt and Engell (1998) note that the production of cyclopentenol from
cyclopentadiene is based on such a reaction scheme (where A = cyclopentadiene, B =
cyclopentenol, C = cyclopentanediol, and D = dicyclopentadiene).

[ Team LiB ]
[ Team LiB ]

M5.2 Model (Chapter 2)

The molar rate of formation (per unit volume) of each component is

Assuming a constant volume reactor, it is easy to derive the following equations (do this):

Notice that the first two equations do not depend on the concentration of components C or D.
Since we are only concerned about the concentration of component B, we only need to solve the
first two equations:

[ Team LiB ]
[ Team LiB ]

M5.3 Steady-State and Dynamic Behavior (Chapter 3)

At steady state, the component A balance yields a quadratic expression,

where the subscript s is used to indicate the steady state value. Solving this quadratic and using
the positive root (clearly, there cannot be negative concentrations), we find

and solving for the steady-state concentration of B,

Using these two expressions, we can find the steady-state input-output curve relating CBs and
Fs/V. Here we consider the dilution rate (F/V) to be the manipulated input. The main advantage is
that it is independent of scale (reactor volume). Whether the reactor is 1 or 10,000 liters, the
same dilution rate will yield the same concentrations (assuming the feed stream is the same
composition).

For the particular reaction under consideration, the rate constants are

and the steady-state feed concentration is CAfs = 10 gmol/liter.

Steady-State Input-Output Curve

Notice that this process has a nonlinear relationship between the steady-state dilution rate (F/V)
and the steady-state concentration of B (Figure M5-2). There exists a maximum concentration of
B that can be achieved. It is interesting to note that the reactor can not be controlled at this
maximum point because the process gain is zero. For a given desired value of the concentration
of B (as long as it is less than the maximum possible value), there are two dilution rates that can
achieve the concentration. This is known as input multiplicity.

Figure M5-2. Steady-state input-output curve. The dotted lines

indicate two possible flow rates to achieve a concentration of 1.117
gmol of B/liter.

In this example, a steady-state concentration of B of 1.117 gmol/liter can be obtained with either
Fs/V = 0.5714 min-1 (case 1) or F s/V = 2.8744 min-1 (case 2). Notice that the process gain (slope
of the steady-state input-output curve) is positive for case 1 and negative for case 2. For
controller design, it is clearly important to know whether you are operating on the left-hand side
(e.g., case 1) or right-hand side (e.g., case 2) of the "peak," since the sign of the controller gain
must be the same sign as the process gain. Also notice that a steady-state concentration greater
than 1.266 gmol/liter cannot be achieved regardless of the controller used.

Linear Analysis
The linear state space model is

where the states, inputs, and output are in deviation variable form. The first input (dilution rate)
is manipulated and the second (feed concentration of A) is a disturbance input.
You should linearize the two modeling equations at steady-state solution to find the following
state space matrices:

Case 1—Operation on the "Left-Hand Side" of the Peak Concentration

Based on the steady-state operating point of CAs = 3 gmol/liter, CBs = 1.117 gmol/liter, and Fs/V
= 0.5714 min-1, the state space model is

The manipulated input-output process transfer function for the reactor is [you can find this
analytically, using G(s) = C(sI-A)-1B, or numerically, using Module 4]

and the disturbance input-output transfer function is

The responses of the linear and nonlinear models to a step change in F/V of 0.1 min-1 are
compared in Figure M5-3.

Figure M5-3. Comparison of concentration of B response to step

change in F/V of 0.1 min-1. Nonlinear (solid) vs. linear (dashed).

Note: A motivated reader may wish to demonstrate that the linear and nonlinear models have the
same response for a step change of 0.01 min-1, while the differences in response to a step change
of 0.5 min-1 are substantial. Function files for integration of the nonlinear model are presented in
Module 3. The focus of the current module is on responses of the linear model.

[ Team LiB ]
[ Team LiB ]

M5.4 Classical Feedback Control (Chapters 5 and 6)

Closed-Loop Performance Requirements

It is desirable to be able to make setpoint changes of ± 0.1 gmol/liter. Also, the manipulated
dilution rate should not change by more than 0.4 min-1, and there should be less than 20%
overshoot. Here, we assume that the process is represented by the transfer function

Problem 1. Based on the transfer function given, answer the following questions:

a. For a desired concentration change of 0.1 gmol/liter, what is the expected steady-state
manipulated dilution rate change? If the volume is 1 liter, what is the expected steady-state
flow-rate change? If the volume is 1000 gallons, what is the expected steady-state flow-rate
change?

b. To keep the manipulated input dilution rate change less than 0.4 min-1, what is the
maximum concentration setpoint change that can be made?

c. What are the units of the controller proportional gain, kc?

Problem 2. Generate the SIMULINK block diagram shown in Figure M5-4. Recall that the default
SIMULINK controller block requires parameters in the decoupled (kc,kI,kD) form,

Figure M5-4. SIMULINK block diagram.

while we use the following form in this text:

If the default SIMULINK PID controller block is used, then the tuning parameters for the integral
and derivative modes are

Note that variables can be specified on the input window for the PID controller block, so that
numerical values can simply be entered in the MATLAB command window before each new
simulation.

Perform some initial simulations with arbitrary PID parameter values. Use a setpoint change of
0.1 at t = 1 minute. Adjust the tuning parameters to obtain an "acceptable" closed-loop response,
so that the concentration overshoot and maximum input flow-rate changes are satisfied.

Problem 3. Consider P-only control and a step setpoint change of 0.1 gmol/liter at t = 1 minute.

a. Find the offset when kc = 1 is used. Is this the amount expected? Show your calculation.

b. Discuss the behavior as kc is increased. At what value of kc does the system appear to go
unstable? Is this consistent with your analytical results from the Routh stability criterion?
Show those results here.

Note that the system solver (integration code), although accurate, may generate points that yield
a "coarse" plot. You can make the curves appear smoother by selecting the parameters pull-
down menu and changing refine from 1 to 2 or 4. This generates more points for plotting
purposes.

Problem 4. Perform a Ziegler-Nichols closed-loop design (based on Chapter 6).

a.
 a. Find the proportional gain that causes the response to continuously oscillate (actually, you
found the proportional gain for this in Problem 3b above), and the "ultimate period" of this
oscillation. Find the Ziegler-Nichols closed-loop controller parameters for P, PI, and PID
(ideal derivative) controllers for this process. Remember that the default S IMULINK PID
controller block uses a different form for the PID controller than we do.

b. Compare the responses of the Ziegler-Nichols P, PI, and PID controllers to a step setpoint
change of 0.1 gmol/liter at t = 1 minute.

c. Compare the responses for Tyreus-Luyben PI parameters (Chapter 6) with Ziegler-Nichols

PI.

d. Compare the responses of ideal and "real" (use a transfer function to generate a "real"
derivative) PID controllers, tuned using the Ziegler-Nichols approach.

Problem 5. Understanding the effect of tuning parameters—PI control.

a. Consider now the effect of integral action. Start with a nominal case of the proportional gain,
kc, set to 1 and the integral time, tI, set to 2 minutes. What happens to the output and
manipulated input responses as tI is decreased? Does there appear to be a value where the
closed-loop system goes unstable? Is this consistent with the Routh stability criterion? Show
your analytical work.

[ Team LiB ]
[ Team LiB ]

M5.5 Internal Model Control (Chapter 8)

Here, we assume that the concentration sensor has a measurement delay of 0.5 minutes, and we
use the following process transfer function:

Open-Loop Controller Design and Implementation

Consider the open-loop control block diagram shown in Figure M5-5, where r(s) represents a
desired setpoint change in the concentration of B (mol/liter), u(s) represents the manipulated
input dilution rate (min-1), and y(s) represents the measured concentration of B (mol/liter).

Figure M5-5. Open-loop control block diagram.

Problem 1. Design an open-loop controller that gives a first-order + time-delay response for a
setpoint change. Assume that the model perfectly captures the process behavior, so that
. For this part, do not factor out the RHP zero to form the controller. Perform
simulations to show that the manipulated variable grows unbounded when a step setpoint change
is made, although good performance of the output is achieved. Assume a setpoint change of
magnitude 0.1 mol/liter.

For the following parts, assume that the all-pass factorization technique is used.

Problem 2. Find q(s) using an all-pass factorization.

Problem 3. Assume the model is perfect.

a. Plot the transient response of y(t) to a setpoint change of r(t) = –0.1 mol/liter, for several
different values of l.

b. Plot the manipulated variable action as a function of l.

c. What is the smallest value of l you can use before the manipulated variable action becomes
unrealistic (the physical flow rate would be negative) for a setpoint change of –0.1 mol/liter?

d. Consider now an unmeasured load disturbance, where the disturbance is a change in the
 c.

d.
feed composition. The disturbance transfer function is

Show (simulate using S IMULINK) that a step load disturbance (feed concentration of A) of 1.0
mol/liter leads to offset in the output variable (see Figure M5-6).

Figure M5-6. Open loop diagram with a disturbance block.

Problem 4. Discuss how process gain uncertainty degrades the performance. For example, if the
real process gain varies between 0.3 and 0.8 (mol/liter)/min-1, what is the offset that results from
using the nominal model for control system design? Assume no disturbance and a setpoint change
of -0.1 mol/liter.

Internal Model Controller Implementation

A problem with open-loop "control" is that the model must be perfect to obtain offset-free
performance. Also, unmeasured disturbances can cause deviations from setpoint. The IMC
implementation, shown in Figure M5-7, handles these problems.

Figure M5-7. IMC implementation.

Use S IMULINK to construct this block diagram.

Problem 5. a–c. Using the same controller design (all-pass factorization) used in Problems 2 and
3, show that the IMC structure gives exactly the same results for a perfect model (process =
model) and no disturbance. That is, show the same results as for 2 and 3a–c.
 d. Consider now an unmeasured load disturbance, where the disturbance transfer function is

Show (simulate using S IMULINK) that a step load disturbance of 1 mol/liter does not lead to
offset in the output variable when the IMC structure is used.

Problem 6. Discuss how gain uncertainty degrades the performance. Keep the process model
transfer function the same but let the actual process gain vary between 0.3 and 0.8
(mol/liter)/min-1. Show the IMC performance for step setpoint changes of –0.1 mol/liter.

Question: Does the higher gain or lower gain process cause more of a control problem?

Problem 7. Discuss how dead-time uncertainty degrades the closed-loop performance. That is,
keep the model dead time the same, but vary the process dead time. Do process time delays that
are larger than the model time delays cause more of a problem than process time delays that are
smaller than the model time delays?

a. For an actual process time delay of 2 minutes (with the model delay of 0.5 minutes), what is
the smallest value of l that can be used before there are problems with the manipulated
variable action for a setpoint change of –0.1 mol/liter?

b. For an actual process time delay of 2 minutes (with the model delay of 0.5 minutes), what is
the smallest value of l that can be used before there are stability problems, when a setpoint
change of –0.01 mol/liter is made?

[ Team LiB ]
[ Team LiB ]

Reference

Klatt, K.-U., and S. Engell, "Gain-Scheduling Trajectory Control of a Continuous Stirred Tank
Reactor," Comp. Chem. Engng., 22 (4/5), 491–502 (1998).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Consider the case 2 operating point, F s/V = 2.8744 min-1. Show that the steady-state
concentrations are CAs = 6.0870 mol/liter and CBs = 1.117 mol/liter.

a. Find the resulting state space model and verify that the manipulated input-output
process transfer function for this operating point is

b. Find the gain-time constant form for this transfer function.

c. Show that a P controller (with the correct sign on the controller gain) cannot
destabilize this process unless there are additional measurement and valve lags.
Use the Routh stability criterion.

2: Consider the following parameters given by Klatt and Engell (1998) for a reactor
operating at 134.14°C.

a. Find the two steady-state flow rates that yield a concentration of B = 0.9
gmol/liter.

b. Find the process transfer function for the lower flow rate that yields CB = 0.9
gmol/liter.

Answer:

c. What are the poles and zeros of this transfer function?

[ Team LiB ]
[ Team LiB ]

Module 6. First-Order + Time-Delay

Processes
The most commonly used model to describe the dynamics of chemical processes is first order +
time delay. In this module, we relate the stability of feedback control of these processes to
frequency-response analysis. The important principles that we wish to illustrate are as follows:

Gain margin is a measure of gain uncertainty that can be tolerated

Phase margin is a measure of the time-delay uncertainty that can be tolerated

The "cross-over frequency" is related to the ultimate period from the closed-loop Ziegler-
Nichols approach

Ziegler-Nichols recommended tuning values do not provide adequate gain and phase
margins

IMC-based PI is easier to tune for robustness than are the other approaches

The sections of this module are as follows:

M6.1 Motivation

M6.2 Closed-Loop Time-Domain Simulation

M6.3 Bode Analysis

M6.4 Ziegler-Nichols Tuning

M6.5 IMC-Based PID Control

M6.6 Summary

It is assumed that the reader has the MATLAB Control Toolbox, which has the Bode and Nyquist
functions.

[ Team LiB ]
[ Team LiB ]

M6.1 Motivation
Since many chemical processes are modeled as first order + time delay (at least for controller
design), it is important to understand the effect of control structure (IMC, P, PI, PID) and tuning
parameters on the stability of the nominal (perfect model) and uncertain (actual) closed-loop
system. In this module, we use the following transfer function to represent the model

For illustrative purposes, we assume that the units of the gain are in degrees Celsius per percent
and that minutes are the time units.

[ Team LiB ]
[ Team LiB ]

M6.2 Closed-Loop Time-Domain Simulation

Here, we begin by using the SIMULINK diagram shown in Figure M6-1.

Figure M6-1. Simulink diagram. For best results, choose a stiff

integrator, such as ode15s (stiff/NDF), from the "parameters" pull-down
menu.

The system is on the verge of instability when kc = 3.8 %/°C, as shown in Figure M6-2; the
ultimate proportional gain is then kcu = 3.8 %/°C. The period of oscillation (ultimate period, Pu) is
approximately 17.2 minutes. Both of these values can be obtained from frequency-response
analysis, as shown in Section M6.3. Notice that the manipulated variable action is constant from t
= 1 to t = 6 minutes. This is due to the initial "proportional kick." The manipulated input does not
change until the process output is observed to change at t = 6 minutes (remember there is a 5-
minute time delay). If we used a controller with integral action, the manipulated input would have
kept increasing during this interval.

Figure M6-2. Closed-loop response for a setpoint change of 1°C at time

= 1 minute, with kc = 3.8 %/°C. The period of oscillation (Pu, peak-to-
peak time) is 17.2 minutes.
[ Team LiB ]
[ Team LiB ]

M6.3 Bode Analysis

Here, we enter the first-order + time-delay model in the MATLAB command window.

» fodt = tf([1],[10 1],'InputDelay',5)

Transfer function:
1
--------
10 s + 1

Input delay: 5

Now, let us perform a Bode analysis assuming a proportional controller gain of 1.

» kc = 1;
» [mag,phase,w] = bode(kc*fodt);
» subplot(2,1,1), loglog(w,squeeze(mag))
» subplot(2,1,2), semilogx(w,squeeze(phase),[0.01 1],
[-180 -180],'--')

The resulting Bode plot, with the –180° line drawn for convenience, is shown in Figure M6-3.
Notice that the crossover frequency is around 0.36 rad/min, where the amplitude ratio is
approximately 0.26. The gain margin is then 1/0.26 = 3.8. This means that the controller gain
could be increased to roughly 3.8 before the process goes unstable. This is consistent with the
finding of kc = 3.8 for instability, shown in the previous section (time-domain simulation).

Figure M6-3. Bode plot of gc(s)gp(s) for kc = 1.

The gain and phase margins can be calculated using the following MATLAB function:

» [Gm,Pm,Wco,Wpm] = imargin(squeeze(mag),squeeze(phase),w)

Gm = 3.80093442825351
Pm = Inf
Wco = 0.36655875121677
Wpm = NaN

where wco is the crossover frequency (where f = –180°) and wpm is the frequency where AR = 1.

Gain Margin
The gain margin of 3.8 indicates that the controller gain can be increased from 1 to 3.8 before the
process becomes unstable—again, this is consistent with the previous results. Also notice that the
phase margin is infinity. This is because the amplitude ratio is always less than 1 (it is exactly 1
only at 0 frequency, when the proportional gain is 1).

Crossover Frequency/Ultimate Period

The crossover frequency is 0.366 rad/minute. Converting this to cycles/minute, we find a
frequency of 0.0583 cycles/minute, which corresponds to a period of 1/0.0583 = 17.16 minutes.
This is exactly the value found in the time-domain simulations.
Nyquist Diagram
The Nyquist plot (Figure M6-4) can be generated by using the following commands:

» [re,im,w] = nyquist(kc*fodt);
» plot(squeeze(re),squeeze(im))

Figure M6-4. Nyquist plot of gc(s)gp(s) for kc = 1.

Problem 1. Use a controller proportional gain of 3.8 and perform a Bode analysis to show that the
amplitude ratio is 1 at the crossover frequency. What is the phase margin and frequency when AR
= 1 (for the controller proportional gain of 3.8)?

Solution

The Bode and Nyquist plots for gc(s)gp(s) with kc = 3.8 are shown in Figures M6-5 and M6-6.
Notice that the phase margin is 1 (the AR = 1 at the crossover frequency) and the phase margin
is 0 (the phase angle is -180° when the AR = 1).

Figure M6-5. Bode plot of gc(s)gp(s) for kc = 3.8. Notice that the gain
 margin is 1 and the phase margin is 0°, indicating that the system is
on the verge of instability.

Figure M6-6. Nyquist plot of gc(s)gp(s) for kc = 3.8. The Nyquist curve
passes through the critical point (–1,0), indicating the system is on
the verge of instability.
This is also shown by the gain and phase margin analysis from MATLAB:

» kc = 3.8;
» [mag,phase,w] = bode(kc*fodt);
» [Gm,Pm,Wco,Wpm] = imargin(squeeze(mag),squeeze(phase),w)

Gm = 1.00024590217198
Pm = 0.03106191271067
Wco = 0.36655875121677
Wpm = 0.36646150971664

Notice that the gain margin is effectively 1, the phase margin is effectively 0, and the crossover
and phase margin frequencies converge to 0.366 rad/min.

[ Team LiB ]
[ Team LiB ]

M6.4 Ziegler-Nichols Tuning

The tuning parameters recommended by Ziegler and Nichols (1942) were shown in Table 6-1.
Generally, these values are not very robust. The objective of this section is for you to calculate
the values of the gain and phase margins for each of these recommended values. Tyreus and
Luyben modified the suggested Ziegler-Nichols parameters for increased robustness (see Luyben
and Luyben, 1997); these recommended values were shown in Table 6-2.

Recall that the P-only controller tuning recommended by Ziegler and Nichols (1942) results in a
controller that has a gain margin of 2 (since the proportional gain is one half of the ultimate gain).
This is also shown by the following M ATLAB analysis.

» kc = 3.8/2;
» [mag,phase,w] = bode(kc*fodt);
» [Gm,Pm,Wco,Wpm] = imargin(squeeze(mag),squeeze(phase),w)

Gm = 2.00049180434395
Pm = 75.49091926803467
Wco = 0.36655875121677
Wpm = 0.16127245556240

The gain margin is 2 and the phase margin is 75°, indicating that a reasonable amount of
uncertainty can be tolerated before the closed-loop system goes unstable. The closed-loop
responses (measured output and manipulated input) for kc = 1.9 are shown in Figure M6-7.

Figure M6-7. Closed-loop response for a setpoint change of 1 at time =

1 minute, with kc = 1.9 (Ziegler-Nichols recommended P-only setting).
Notice the offset due to the P-only controller.
Problem 2. Find the gain and phase margins when Ziegler-Nichols PI settings are used. Compare
these with the gain and phase margins of the Tyreus-Luyben recommended PI settings. Compare
the closed-loop responses (time domain) for both of these controller settings; include plots of the
manipulated input action. Hint: Create a transfer function for the PI controller and multiply it with
the process transfer function. Recall that a PI controller transfer function is

The Ziegler-Nichols recommended parameters for a PI controller are then

The commands to M ATLAB are then

» gc = tf([24.45 1.71],[14.3 0])

Transfer function:
24.45 s + 1
-----------
14.3 s

» [mag,phase,w] = bode(gc*fodt);
» [Gm,Pm,Wco,Wpm] = imargin(squeeze(mag),squeeze(phase),w)

Thus far, we have used the gain margin to determine either how much the controller gain can be
increased before instability occurs, or how much the process gain can increase before instability
occurs. Next, we illustrate how the phase margin is related to time-delay uncertainty that can be
tolerated.

Problem 3. Use the Ziegler-Nichols recommended P-only value of kc = 0.5 kcu = 3.8/2 = 1.9.
What is the phase margin for this value of controller gain? What is the frequency associated with
this phase margin? Calculate the amount of additional time delay that can be tolerated from

Add this time delay to the process model and show that the new phase margin is now 0. Show
that the time-domain closed-loop simulations have an oscillation with a frequency of wpm. Include
plots of the manipulated input action.

[ Team LiB ]
[ Team LiB ]

M6.5 IMC-Based PID Control

Recall from Chapter 9 that Padé approximations for dead time could be used to develop a IMC-
based PI and PID controllers for first-order + time-delay systems. Here, we compare the gain and
phase margins for an IMC-based PI controller (using the "improved" settings) for various values
of l.

The improved IMC-based PI settings are as follows:

Problem 4. For the IMC-based PI (improved) settings, find the value of l that leads to a gain
margin of 2.5. Also, find the value of l that leads to a phase margin of 75°. Compare the time-
domain closed-loop responses for these two systems (always include plots of the manipulated
inputs). Hint: Recall that the improved IMC-based PI settings for first-order + time-delays
systems are

For l = 10, we find

» gc = tf([1.25*12.5 1.25],[12.5 0])

Transfer function:
15.62 s + 1.25
--------------
12.5 s

» [mag,phase,w] = bode(gc*fodt);
» [Gm,Pm,Wco,Wpm] = imargin(squeeze(mag),squeeze(phase),w)

Gm = 2.63261419901939
Pm = 63.19455846288847
Wco = 0.32379172211747
Wpm = 0.11488259758305

Here, we see that the gain margin requirement is satisfied. Increase l until the phase margin is
satisfied.
Problem 5. Compare the Ziegler-Nichols PI, Tyreus-Luyben PI and IMC-based improved PI (with l
= 10 minutes) control performance in time-domain simulations.

[ Team LiB ]
[ Team LiB ]

M6.6 Summary
Frequency-response techniques have been used to study the robustness of a simple first-order +
time-delay process model. The gain margin is a measure of gain uncertainty that can be
tolerated, while phase margin is a measure of dead time uncertainty that can be tolerated. An
uncertain time constant affects both the gain margin and phase margin.

[ Team LiB ]
[ Team LiB ]

References

The original Ziegler-Nichols parameters are developed in the following paper: Ziegler, J. G., and
N. B. Nichols, "Optimum Settings for Automatic Controllers," Trans. ASME, 64, 759–768 (1942).

The Tyreus-Luyben recommended values are presented in Chapter 3 of the following textbook:
Luyben, M. L. and W. L. Luyben, Essentials of Process Control, McGraw Hill, New York (1997).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: The transfer function for the van de Vusse reactor is given below; the input variable is
dilution rate (F/V, min-1) and the output is concentration of component B (mol/liter).
Find the gain and phase margins for the Ziegler-Nichols recommended P, PI, and PID
parameters. Show the resulting closed-loop time-domain responses to a unit setpoint
change.

2: The IMC-based PID design procedure for the van de Vusse reactor (Additional Exercise
1) results in a PID with first-order lag controller. Find how the gain and phase margins
vary as a function of the IMC filter factor, l. Vary l from 0.1 minutes to 1 minute. What
is the recommended value of l if the minimum gain and phase margins are 3 and 75°,
respectively?

[ Team LiB ]
[ Team LiB ]

Appendix M6.1
Answers to Problems 2, 3 and 4. The reader should show how these results are obtained and
generate the resulting closed-loop time domain plots.

Problem 2: Ziegler-Nichols PI: Gm = 1.97, Pm = 53, Wco = 0.33, Wpm = 0.16. Tyreus-Luyben:
Gm = 3.09, Pm =102, Wco = 0.351, Wpm = 0.0746.

Problem 3: Gm = 2.00, Pm = 75.49, Wco = 0.366, Wpm = 0.161. Additional time delay = 8.18
minutes.

Problem 4: For l = 9.5, kc = 1.315, Gm = 2.50. Increase l to obtain a Pm of at least 75°.

[ Team LiB ]
[ Team LiB ]

Module 7. Biochemical Reactors

The objective of this module is to illustrate a practical problem where the IMC-based PID
procedure can be used. The sections of the module are as follows:

M7.1 Background

M7.2 Steady-State and Dynamic Behavior

M7.3 Stable Steady-State Operating Point

M7.4 Unstable Steady-State Operating Point

M7.5 SIMULINK Model File

The procedure for a stable process will be used in Section M7.3, while the procedure for an
unstable process will be used in Section M7.4. In both cases, use the SIMULINK model file
presented in Section M7.5.

Although you will be using a linear controller, you will be implementing this controller on a
nonlinear process. The inputs and outputs to/from this process are in physical variables, while the
controller design procedure is based on deviation variables.

[ Team LiB ]
[ Team LiB ]

M7.1 Background
Biochemical reactors are used in a wide variety of processes, from waste treatment to alcohol
fermentation. Biomass (cells) consume substrate (sugar or waste chemicals) and produce more
cells. A typical control and instrumentation diagram, with biomass concentration as the measured
output, is shown in Figure M7-1.

Figure M7-1. Simplified control and instrumentation diagram for a

biochemical reactor.

Model
The modeling equations for a bioreactor are

Equation M7.1

where the state variables are x1 = biomass (cell) concentration = mass of cells/volume, and x2 =
substrate concentration = mass of substrate/volume.

The manipulated input is D = dilution rate = F/V = volumetric flow rate/reactor volume, and the
disturbance input is x2f = substrate feed concentration.

Two possible expressions for the specific growth rate are monod and substrate inhibition kinetics,
which include
Notice that the monod-specific growth rate model is a subset of the substrate inhibition model (k1
= 0).

Scale Up
One motivation for working with the dilution rate as the manipulated input is that the resulting
dynamic model is independent of scale. A reactor volume of 1 liter with a flow rate of 0.3
liters/hour has the same dynamic behavior as a reactor volume of 1000 liters with a flow rate of
300 liters/hour. Thus, a small-scale (laboratory or pilot plant) reactor can be used to predict the
behavior of a production-scale reactor.

[ Team LiB ]
[ Team LiB ]

M7.2 Steady-State and Dynamic Behavior

We use the following parameters for a substrate inhibition model for this control study:

The steady-state dilution rate is Ds = 0.3 hr-1 (and the residence time is 3.33 hours) and the feed
substrate concentration is x2fs = 4.0 g/liter.

Steady-State Conditions
The nonlinear process has the following three steady-state solutions (operating points) to
Equation M7.1 for a dilution rate of 0.3 hr-1:

Steady state Biomass concentration Substrate concentration stability

Equilibrium 1—wash out x1s = 0 x2s = 4.0 stable

Equilibrium 2—nontrivial x1s = 0.995103 x2s = 1.512243 unstable

Equilibrium 3—nontrivial x1s = 1.530163 x2s = 0.174593 stable

Notice that we definitely do not want to operate a equilibrium point 1. Here there is no reaction
occurring because the cells have been washed out of the reactor. The outlet substrate
concentration is the same as the inlet substrate concentration under these conditions.

Linear Model
The state-space model matrices are

where the partial derivative of the specific growth rate with respect to the substrate concentration
is
and where dilution rate is the manipulated input. Different control strategies have been used to
control continuous biochemical reactors. One is based on measuring the biomass concentration
and manipulating the dilution rate. Another is based on measuring the substrate concentration
and manipulating the dilution rate.

In the following two problems, the biomass concentration is the measured output and the dilution
rate is the manipulated input. You will be performing your control simulations using the nonlinear
process, so the inputs and outputs from the process are in physical variable form, while the linear
controller design is based on the use of deviation variables. Assume that the dilution rate is
physically constrained between 0 and 0.6 hr-1.

[ Team LiB ]
[ Team LiB ]

M7.3 Stable Steady-State Operating Point

Design an IMC-based PID controller to control the bioreactor at equilibrium point 3—the stable
nontrivial point. The steady state (also use this as the initial condition for your simulations) is

At this operating point, the state space model is

Use M ATLAB to find that the eigenvalues are –0.3 and –2.2640 hr -1, so the system is stable and
the IMC-based PID method for stable systems can be used (see Table 9-1). Find the transfer
function relating the dilution rate to the biomass concentration and use this for controller design.
You may wish to use the MATLAB function ss2tf to find the process transfer function

After placing the process model in gain and time constant form and recognizing pole-zero
cancellation, you should find

Notice that the time constant of 0.4417 hr is significantly shorter than the residence time (3.33
hr). Also notice from Table 9-1, that the IMC-based controller is a PI controller.

a. Show how the response to a small setpoint change varies with l. Suggested setpoint
changes are from 1.53016 to 1.52 and from 1.53016 to 1.54 g/liter.

b. For a particular value of l, show how the magnitude of the setpoint change affects your
response. This is where the nonlinearity comes into play. Try changing from the steady-state
value of 1.53016 to 1.0, 1.4, 1.6, and 2.5 g/liter. You will find problems with control
saturation and for larger setpoint changes.

c. Often measurements cannot be made instantaneously, and there will be a transport delay
associated with the measurement. Use the transport delay function in SIMULINK. Use q = 0.25
hour and discuss how the time-delay affects your choice of l.

d.
 c.

d. Consider now disturbance rejection. Do not make a setpoint change but do make a step
change in substrate feed concentration (x2f) from 4 to 3.5 at t = 1 hour. Compare
disturbance rejection results with the PI controller developed in row 2 of Table 9-1.

[ Team LiB ]
[ Team LiB ]

M7.4 Unstable Steady-State Operating Point

Design an IMC-based PID controller to control the bioreactor at equilibrium point 2—the unstable
nontrivial point. The steady state (also use this as the initial condition for your simulations) is

At this point, the state space model is

Use M ATLAB to find that the eigenvalues are –0.3 and 0.169836 hr -1, so the system is unstable and
the IMC-based PID method for unstable systems must be used (Table 9-3). Find the transfer
function relating the dilution rate to the biomass concentration and use this for controller design.
You may wish to use the MATLAB function ss2tf to find the process transfer function

After placing the process model in gain and time constant form (and cancelling common poles and
zeros), you should find

That is, the transfer function has a RHP pole at 0.1698 hr-1 which is consistent with the state
space model. Notice that we can use the first entry in Table 9-3, which is a PI controller.

a. Show how the response to a small setpoint change varies with l (show explicitly how the
PID tuning parameters vary with l). I suggest setpoint changes from 0.995103 to 0.985 and
1.005 g/liter.

b. For a particular value of l, show how the magnitude of the setpoint change affects your
response. This is where the nonlinearity comes into play. Try changing from the steady-state
value of 0.995013 to 0.5, 0.75, 1.5, and 2.0 g/liter.

c. Use the transport delay block in SIMULINK. Start with q = 0.25 hour and discuss how the time
delay affects your choice of l.

d.
 c.

d. Consider now disturbance rejection. Do not make a setpoint change but do make a step
change in substract feed concentration (x 2f) from 4 to 3.5 gmol/liter at t = 1 hour.

[ Team LiB ]
[ Team LiB ]

M7.5 SIMULINK Model File

In order to save time on constructing your block diagram, copy the block for the bioreactor from
the textbook web page (see Figure M7-2)

biofbc.mdl

Figure M7-2. SIMULINK block diagram for bioreactor.

I recommend that you make your proportional and integral terms in the PID controller a function
of lambda (l). Then, before each simulation run, you merely have to type in a new lambda in the
MATLAB window, rather than typing in new PI parameters. Always realize that the integral term in
the default SIMULINK PID block is our kc/tI.

Notice that the SIMULINK diagram contains two extra blocks. The transport delay block can be
inserted in the feedback loop to illustrate the effects of a measurement time delay. The PID
controller for the open-loop unstable point can be inserted when Additional Exercise 2 is being
solved.

The following must be entered into the MATLAB command window: initial conditions, x1i and x2i;
setpoint for biomass concentration, x1sp; final simulation time, tfinal; substrate feed
concentration, x2f; and the IMC filter factor, lambda.

The transfer function parameters are used by the PID controller block. The IMC-based PID
parameters are different depending on whether the process is open-loop stable (Additional
Exercise 1) or open-loop unstable (Additional Exercise 2). This is also seen by comparing
 Tables 9-1 and 9-3.

Notice that a saturation block is used so that the dilution rate is bounded between 0.0 and
0.6. This is consistent with "real-world" constraints on manipulated variable action.

The typical rule-of-thumb for l for open-loop stable systems is roughly one third to one half
of the dominant time constant. There are no good rules-of-thumb for open-loop unstable
systems.

When performing your simulations, plot the dilution rate to see if it is hitting a constraint.

[ Team LiB ]
[ Team LiB ]

Reference

The parameters used in this module were presented in the following paper: Agrawal, P., and H. C.
Lim, "Analyses of Various Control Schemes for Continuous Bioreactors," Adv. Biochem. Eng.
Biotechnol., 30, 61–90 (1984).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Data for a specific growth rate coefficient as a function of substrate concentration for a
biochemical reactor are shown below

x2, 0 0.1 0.25 0.5 0.75 1 1.5 3 5

g/liter
m, hr-1 0 0.38 0.54 0.63 0.66 0.68 0.70 0.73 0.74

a. Estimate the parameter values for a Monod-specific growth rate model.

b. The production rate of cells is Dx1. Find the steady-state value of the dilution rate
that maximizes the production rate of cells. The substrate feed concentration is 5
g/liter.

c. Find the steady-state concentration of biomass and substrate at this dilution rate.

d. Find the linear state space model at this dilution rate, with dilution rate and
substrate feed concentration as the input variables. Also, find the transfer function
relating dilution rate to biomass concentration.

e. Simulate the responses (using the nonlinear dynamic model) of the concentration
of biomass and substrate to step increases and decreases of 10% in the dilution
rate (changes are from the dilution rate found in question b.). Compare these
results with those of the linear system (remember to convert deviation variables
back to physical variables).

2: Literature search: Perform a literature review (using Current Contents, EI Compendex,

or some other professional search package, and NOT an Internet-based search engine)
on the control of fermentation processes. Select one or two articles that are most
interesting to you and discuss the importance of this particular process. Example: What
is a typical annual production rate?

[ Team LiB ]
[ Team LiB ]

Module 8. CSTR
Unlike many of the other modules, this module is too lengthy to be performed in a single 1- to 2-
hour session. Each instructor or reader is likely to have a different focus. Much of the initial
material is related to modeling and interesting nonlinear aspects. If the focus is on controller
design, the initial material can be skimmed before Problem M8.4 is completed. After studying this
module the reader should be able to understand:

The effect of scale-up on the steady-state and dynamic characteristics of a jacketed

chemical reactor. Further, understand how these characteristics affect the control system
design and operating performance.

These reactors can have input-output operating curves that are nonlinear; that is, the
"process gain" is not constant.

Larger reactors have a reduced capability of removing heat compared to smaller reactors.
Often, a larger reactor is operated at an unstable point, while the smaller reactor (with the
same residence time) is stable.

The major sections of this module are

M8.1 Background

M8.2 Simplified Modeling Equations

M8.3 Example Chemical Process—Propylene Glycol Production

M8.4 Effect of Reactor Scale

M8.5 Further Study: Detailed Model

M8.6 Other Considerations

M8.9 Summary

[ Team LiB ]
[ Team LiB ]

M8.1 Background
Chemical reactors often have significant heat effects, so it is important to be able to add or
remove heat from them. In a jacketed CSTR (continuous stirred-tank reactor) the heat is added
or removed by virtue of the temperature difference between a jacket fluid and the reactor fluid.
Often, the heat transfer fluid is pumped through agitation nozzles that circulate the fluid through
the jacket at a high velocity; an example recirculating jacket system is shown in Figure M8-1.
Notice that cascade control is shown, with a primary controller (reactor temperature) and a
secondary controller (jacket temperature). The output of the reactor temperature controller is the
jacket temperature setpoint. The output of the jacket temperature controller is the valve position
for the jacket make-up valve (usually it is assumed that this is directly related to the flowrate of
the jacket make-up fluid). The coolant can be chilled water, but is often a heat transfer fluid, such
as an ethylene glycol/water mixture or a proprietary mixture of hydrocarbons.

Figure M8-1. Cascade control, with jacket outlet temperature as the

secondary control variable.

A different cascade control strategy is shown in Figure M8-2. In this cascade control stategy, the
output of the reactor temperature controller is the setpoint for jacket inlet temperature. The
output of the jacket inlet temperature controller is the valve position of the coolant make-up
valve. The first cascade control strategy (Figure M8-1) is the most commonly used reactor
temperature control strategy.

Figure M8-2. Cascade control, with jacket inlet temperature as the

secondary control variable.
[ Team LiB ]
[ Team LiB ]

M8.2 Simplified Modeling Equations

Here we make the following assumptions

1. Perfect mixing in reactor and jacket

2. Constant volume reactor and jacket

3. Constant parameter values

In addition, to develop a simplified model, we assume that the jacket temperature can be directly
(and nearly instantaneously) manipulated. This assumption is very good if a boiling heat transfer
fluid is used, for example; changing the pressure on the jacket side would result in near
instantaneous change in jacket temperature (note that a recirculating jacket system would not be
used in this scenario). Even for the recirculating heat transfer system, the assumption of the
jacket temperature being directly manipulated can be good if the jacket dynamics are rapid
compared to the reactor dynamics.

Overall Reactor Material Balance

The reactor overall material balance equation is

where V is the constant liquid reactor volume, r is the density of reactor fluid and rin is the
density of the inlet stream. F in and Fout are the flow rates of the inlet and outlet streams.
Assuming a constant density (rin = r) and volume, it is easy to show that Fout = Fin = F; this will
be used in the component balance.

Balance on Component A
Here we consider, for simplicity, the simple reaction A B. The balance on component A is

Equation M8.1

where CA is the concentration of component A in the reactor and rA is the rate of reaction per unit
volume. The Arrhenius expression is normally used for the rate of reaction. A first-order reaction
results in the following

Equation M8.2
where k0 is the frequency factor, Ea is the activation energy, R is the ideal gas constant, and T is
the reactor temperature on an absolute scale (R, Rankine or K, Kelvin).

Reactor Energy Balance

The reactor energy balance, assuming constant volume, heat capacity (cp) and density (r), and
neglecting changes in the kinetic and potential energy is

Equation M8.3

where -DH is the heat of reaction, U is the heat transfer coefficient, A is the heat transfer area, Tf
is the feed temperature, and Tj is the jacket temperature.

State Variable Form of the Equations

We can write Equations (M8.1-M8.3) in the standard state variable form

Equation M8.4a

Equation M8.4b

Steady-State Solution
The steady-state solution is obtained when the two state derivatives are set equal to zero

Equation M8.5a
Equation M8.5b

Here, all of the inputs and parameters must be specified, allowing the two equations to be solved
for steady-state values of the two states. For these equations, a nonlinear technique, such as
Newton Raphson, would need to be used. It should be noted, however, that there is an easier
selection of the two unknowns to solve for. For a given steady-state temperature, solve Equation
(M8.5a) for the steady-state concentration. Then, solve Equation (M8.5b) for the value of the
jacket temperature (perform this in the following exercise).

Problem M8.1 Steady-State Concentration and Jacket Temperature for a Specified Reactor
Temperature

a. Solve for the steady-state concentration as a function of the specified steady-state

temperature, from Equation (M8.5a). Use a subscript s to indicate a steady-state value.

Result:

Equation M8.6a

b. Solve for the steady-state jacket temperature as a function of the steady-state reactor
temperature and concentration from Equation (M8.5b).

Result:

Equation M8.6b

Alternatively, one could specify the desired concentration, solve for the required reactor
temperature (from M8.5a), and then solve for the required jacket temperature (from M8.5b).
Again, this assumes that the residence time and all other parameters are fixed.

Linearization
The goal of the linearization procedure is to find a model with the form
Here we define the states, output, and inputs in deviation variable form

Notice that the first input (jacket temperature, Tj) is manipulated, while the last three inputs are
disturbances.

The Jacobian matrix is

Equation M8.7

where ks = k0 exp(–Ea/RTs) and .

The first column (related to the manipulated input, Tj ) of the B matrix is

Equation M8.8

The output matrix is, since we consider temperature (second state) to be measured,

Also, the input is not measured, so D = 0.

[ Team LiB ]
[ Team LiB ]

M8.3 Example Chemical Process—Propylene Glycol

Production
Roughly 1.3 billion pounds of propylene glycol are produced per year. It has a wide variety of
uses, including: anti-freeze applications, including aircraft deicing; a solvent for a number of
drugs; moisturizers; and as artificial smoke or fog, for fire-fighting training or theatrical
productions. Propylene glycol is produced by the hydrolysis of propylene oxide with sulfuric acid
as a catalyst

Water is supplied in excess, so the reaction is first-order in propylene oxide concentration. The
rate of reaction of propylene oxide (component A) is first-order

Parameter Values
This system has the following activation energy, frequency factor, and heat of reaction values

The other parameters are

Assume that the reactor is to be operated with the following residence time, feed concentration,
and feed temperature

Also assume the reactor is designed as a vertical cylinder with a height/diameter ratio of 2:1, that
complete heat transfer area coverage occurs when the reactor is 75% full, and that the reactor is
operated at 85% of the design volume.

[ Team LiB ]
[ Team LiB ]

M8.4 Effect of Reactor Scale

In this section we study the effect of reactor size on the steady-state and dynamic behavior of the
reactor. Each reactor is designed with a residence time of 15 minutes, with a desired conversion
of 50% (CA = 0.066 lbmol/ft 3), resulting in a reactor temperature of 101.1°F.

Reactor Scale
Consider now the following reactors, operated at 85% of the design volume.

Reactor Volume (nominal) 100 ft3 5000 gallon (668 ft3)

Heat transfer area 88 ft2 309 ft2

Diameter 4 ft 7.5 ft
Operating Volume, V 85 ft3 500 ft3

Operating Flowrate, F 340 ft3/hr 2000 ft3/hr

UA 6600 Btu/hr °F 23200 Btu/hr °F

Problem M8.2 Steady-State Jacket Temperatures

Find the steady-state jacket temperatures for the 100 ft3 and 5000 gallon reactors.

Solution: Substituting the parameter values and steady-state concentration and temperature into
Equations (M8.6a) and (M8.6b), the resulting jacket temperatures are 80°F (100 ft3) and 65°F
(5000 gallon).

Steady-State (Nonlinear) Results

Expanding on Problem M8.1 we can find the jacket temperature as a function of reactor
temperature, and solve for a number of reactor temperatures to construct a steady-state input-
output plot. This is shown in Figure M8-3 for different reactor sizes. The '+' symbol denotes the
design operating point for each reactor. The jacket and reactor temperatures at this design point
are consistent with the results in Problem M8.2.

Figure M8-3. Steady-state input-output plot of reactor temperature as

a function of jacket temperature. Notice that the larger reactors
exhibit multiple steady-state behavior (this is also known as
hysteresis). The desired steady-state operating point is denoted by
'+'.
There are several interesting things to note:

1. All of the input output curves are nonlinear. Recall that a process gain is change in
output/change in input, or the slope of the input-output curve. A linear system has a
constant gain and therefore a constant slope.

2. The larger (5000 gallon and 1000 ft3) reactors exhibit multiple steady-state behavior. That
is, for a certain jacket temperature there is the possibility of three reactor temperatures. For
these larger reactors the intermediate temperature (desired) operating point is unstable.
This will be shown in the linear stability analysis that follows.

3. For a given reactor temperature, a larger reactor requires a lower jacket temperature. This
is because a larger reactor has a reduced heat transfer area/volume (A/V) ratio and a
reduced ability to transfer heat.

Linear Open-Loop Results

The steady-state operating point is CAs = 0.066, Ts = 101.1°F.

a. For the 100-ft 3 reactor the state space model is (note that the time unit is hours)

and you should be able to show (analytically, or using MATLAB) that the process transfer
function is
 Equation M8.9

where the process gain is kp = 3.275°F/°F and the time unit is hours. You will want to factor
this into the gain-time constant form for controller design.

b. For the 5000-gallon reactor the state space model is

and you should be able to show (analytically, or using MATLAB) that the process transfer
function is

Equation M8.10

You will want to factor this into the gain-time constant form for controller design.

Problem M8.3 Open-Loop Stability

Calculate the poles and zeros of the process transfer function for both size reactors. You should
find that the 100-ft3 reactor is stable, while the 5000-gallon (668 ft 3) reactor is unstable.

Since the 5000-gallon reactor is open-loop unstable, it can only be operated at the design point of
101.1°F using a closed-loop control system.

Linear Closed-Loop Results

Recall that the closed-loop transfer function relating the setpoint to the output is

where the stability is determined by the roots of the denominator polynomial. The Routh stability
criterion (chapter 5) can be used to determine the ranges of the controller tuning parameters that
will assure a closed-loop stable system. See Additional Exercise 4 at the end of the module.

Problem M8.4 Closed-Loop Simulations

Develop a SIMULINK block diagram to perform closed-loop simulations for the two reactor sizes
(100 ft3 and 5000 gal). For simplicity, use the linear process transfer functions (M8.8 and M8.9)
to represent the respective reactor. Include two additional first-order transfer functions with time
constants of 0.1 hour to account for unmodeled mixing and jacket dynamics (although these are
included in your simulation studies, neglect them for any model-based controller design
procedure). Here any of your favorite control system design techniques can be used: (i) Ziegler-
Nichols, (ii) Tyreus-Luyben, (iii) IMC, (iv) IMC-based PID, (v) "ad-hoc" iterative tuning.

In the following we use the IMC-based PID procedure for both the 100-ft3 (stable, Table 9-1) and
the 5000-gallon (unstable, Table 9-3) reactors. The IMC filter factor (l) is adjusted to obtain a
satisfactory closed-loop response to a step setpoint change in reactor temperature (including a
consideration of the action of the manipulated jacket temperature). For the unstable reactor a
setpoint filter (see Figure 9-8) should be used for improved response.

Make plots of both the reactor temperature (output) response and the manipulated jacket
temperature (input). Since the linear simulations are in deviation variable form, add the steady-
state values to the deviation variables to plot the physical values. Your responses for a 5°F
setpoint change should be similar to Figure M8-4.

Figure M8-4. Closed-loop responses for reactor temperature setpoint

changes of +5°F. Linear simulations for 100 ft3 reactor. IMC-based PID
with l = 0.5 hr.

[ Team LiB ]
[ Team LiB ]

M8.5 For Further Study: Detailed Model

A more detailed model of a CSTR includes the effect of cooling jacket dynamics. The three
modeling equations are

Equation M8.11a

Equation M8.11b

Equation M8.11c

where Fj is the flow rate through the jacket, and Tjin is the jacket inlet temperature. The variable
definition should be clear from Figure M8-5. Notice that the manipulated input is the jacket make-
up flow rate (Fjf), which does not appear in the above modeling equations.

Figure M8-5. CSTR with recirculating jacket. Definition of variables.

A static energy balance can be used to find the jacket inlet temperature as a function of the
jacket make-up flow rate (Fjf)

Equation M8.11d

When (M8.11d) is substituted into (M8.11c), the following equations result

Equation M8.12a

Equation M8.12b

Equation M8.12c
which is the same set of equations that would be derived from a reactor with a "once-through"
jacket, shown in Figure M8-6.

Figure M8-6. CSTR with "once-through" jacket. Definition of variables.

This is an interesting result. The model for a CSTR with a "once-through" jacket is the same as
that for a recirculating jacket, assuming a static energy balance around the mixing point. This is
one reason that it is common to see Figure M8-6 used to represent a CSTR when Figure M8-5 is
the more common configuration in industry. It should be noted that the assumption of a constant
heat-transfer coefficient is much better for Figure M8-5 than M8-6, since the jacket velocity will
stay relatively constant for this system.

Linear Open-Loop Results

Here we again study the propylene glycol reactor presented in Section M8.3, with the following
additional jacket parameters

For the 100-ft 3 reactor the state space model is (see Appendix M8.1 for the linearization results)

Equation M8.13
where x = [CA – CAs T – T s Tj – Tjs]T, u = Fjf – F jfs, y = [T – Ts T j – Tjs]T

You should be able to show (using MATLAB) that the process transfer function relating the jacket
flow rate (u1) to the reactor temperature (y1) is

Equation M8.14

which has an unstable pole.

Nonlinear Open-Loop Results

There is also multiple steady-state behavior as shown in Figure M8-7. This is an interesting result,
because the 100-ft3 reactor was stable when the simplified model was used, and there was only
one operating point. The simplified model essentially assumes perfect control of the jacket
temperature. This means that the jacket temperature controller stabilizes the process.

Figure M8-7. Steady-state input-output plot of reactor temperature as

a function of jacket flow rate for the 100-ft3 reactor. The desired
steady-state operating point is denoted by '+'.
Linear Closed-Loop Results
Since the CSTR is open-loop unstable, a feedback controller must be designed to stabilize the
reactor operation. In the next problem, the Routh stability criterion is used to find the range of
controller gains that stabilize the CSTR.

Problem M8.5 A Stabilizing P-only Controller for the 100-ft3 Reactor

Use the Routh stability criterion to determine the range of controller gains, for a P-only controller,
that stabilize the reactor (based on the transfer function M8.14). Based on simulation results,
what value of gain do you recommend for the P-only controller?

Problem M8.6 Cascade Control of the 100-ft3 Reactor

Develop a SIMULINK diagram of a cascade controller CSTR, where a state space representation
[Equation (M8.12)] is used for the reactor. First, tune the secondary loop to stabilize the process.
Then, tune the primary loop for the desired setpoint response behavior.

[ Team LiB ]
[ Team LiB ]

M8.6 Other Considerations

Nonlinear Behavior
Although general nonlinear models were developed in this module, and some nonlinear steady-
state results were presented, dynamic simulations were based on linear models. Although the
responses on the physical (nonlinear) system will be similar to the linear results for small changes
around the designed operating point, the nonlinear behavior can be significantly different for large
changes.

Download the nonlinear SIMULINK/MATLAB files from the book web page and perform closed-loop
simulations. Show that responses of the nonlinear system are similar to the linear simulations, for
small setpoint changes. For a sequence of larger setpoint changes, show that the nonlinear
system begins to deviate significantly from the linear system.

Split-Range Control
Thus far we have only considered the case where coolant was circulated through the jacket at a
lower temperature than the reactor vessel, to remove energy due to the exothermic reaction.
During reactor start-up, however, it may be necessary to add energy to the reactor to bring the
temperature up to a point where the reaction can be initiated. In this case a warm fluid must be
circulated through the jacket. This requires a split-range controller as shown in Figure M8-8 and
M8-9. Notice that the cold fluid is fail-open, while the hot fluid is fail-closed.

Figure M8-8. Split-range strategy for secondary controller.

 Figure M8-9. Split-range controller action.

[ Team LiB ]
[ Team LiB ]

M8.7 Summary
Exothermic CSTRs can exhibit interesting, and challenging, nonlinear behavior. Even a simple
two-state model, with jacket temperature manipulated and the reactor temperature as the
(controlled) output variable, can have multiple steady-state behavior. For reactors with three
steady states, the intermediate steady-state is unstable, requiring a feedback controller for stable
operation.

CSTRs often have recirculating jacket fluid heat-transfer systems to improve the heat removal
capability. For tighter control, and particularly for start-up and transition control, it is important to
incorporate a split-range strategy; here, the jacket temperature controller output manipulates
either a hot fluid valve or a cold fluid valve, depending on the range of the controller output. For
most reactors it is desirable for the cold valve to fail open and the hot valve to fail closed.

[ Team LiB ]
[ Team LiB ]

References
The Fogler text uses the propylene glycol example for the analysis of an adiabatic CSTR. The
Seider et al. text uses it in the context of simulation-based process design. The original
experimental results for the kinetic rate expressions are provided in Furusawa et al. An analysis of
a two-state CSTR is presented in the Process Dynamics text by Bequette. The article by Lipták
provides an excellent overview of both batch and continuous reactor control.

Bequette, B. W., Process Dynamics: Modeling, Analysis and Simulation, Prentice Hall, New York
(1998).

Fogler, H. S., Elements of Chemical Reaction Engineering, 2nd ed., Prentice Hall, New York
(1992).

Furusawa, T., H. Nishimura, and T. Miyauchi, "Experimental Study of a Bistable Continuous

Stirred-Tank Reactor, J. Chem. Eng. Jpn., 2(1), 95–100 (1969).

Lipták, B. G., "Controlling and Optimizing Chemical Reactors," Chemical Engineering Mag., pp.
69–81 (May 26, 1986).

Seider, W. D., J. D. Seader, and D. R. Lewin, Process Design Principles, Wiley, New York (1999).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Perform a search on propylene glycol. What is the worldwide demand? What is the
current sales price? Enumerate some of the uses of this product.

2: Consider the simple CSTR model used in Section M8.4. Transfer functions for the 100-ft3
and 5000-gallon reactor were given.

a. Find the heat transfer area, diameter, and UA for a 1000-ft3 reactor (with an
operating volume of 850 ft3), designed with the same residence time and
height/diameter ratio as the other reactors. Also, find the required jacket
temperature to achieve the reactor temperature of 101.1°F.

b. Find the state space model and calculate the eigenvalues for the 1000-ft3 reactor.
Find the transfer function and calculate the poles. You should find that this reactor
is open-loop unstable. Now, find the range of proportional gains that will yield a
stable closed-loop system under proportional-only control.

3: Consider the propylene glycol example

a. For a desired propylene oxide conversion of 50% at a residence time of 0.25 hr,
show that the reactor operating temperature is 101.1°F.

b. For a desired propylene oxide conversion of 50% and a feed concentration of

0.132 lbmol/ft3, find the production rate of propylene glycol for two different feed
flow rates: (i) 340 ft3//hr, and (ii) 2000 ft3/hr. Assume that the plant operates for
8400 hours/year. If propylene glycol sells for $0.70/lb, what are the expected
annual revenues (before operating costs are considered) for each flowrate?

c. Assume complete liquid coverage of the heat transfer area when the reactor is
75% full. What is the heat transfer area for a 100-ft3 reactor? Remember to
include both the "bottom" and "side" of the reactor.

d. Again, assume complete liquid coverage of the heat transfer area when the reactor
is 75% full. What is the heat transfer area for a 5000-gallon (668 ft3) reactor?
Remember to include both the "bottom" and "side" of the reactor.

e. Comment on the A/V ratio as the volume of a reactor increases. For a specified
reactor conversion (and temperature) and residence time, how does the jacket
temperature change as a function of reactor volume?

4: Propylene glycol example, closed-loop behavior.

Perform the following for both the 100-ft3 and 5000-gallon (668 ft3) reactors.

a. For a PI controller, with an integral time of 1 hour, find the range of controller
gains for closed-loop stability using the Routh stability criterion.

b. Develop a SIMULINK block diagram to perform closed-loop simulations for the two
 a.

b.
reactor sizes. For simplicity, use the linear process transfer functions to represent
the reactor. Include two additional first-order transfer functions with time
constants of 0.1 hour (6 minutes) to account for unmodeled mixing and jacket
dynamics. Select tuning parameters within the range of those that you obtained in
part a. Adjust the parameters until you are satisfied with closed-loop response to a
step setpoint change.

Make plots of both the reactor temperature response and the manipulated jacket
temperature.

5: Consider the following CSTR parameters

Assume that the reactor is to be operated with the following steady-state inputs:

Solve for the steady-state concentration and temperature and find the corresponding
state space model.

[ Team LiB ]
[ Team LiB ]

Appendix M8.1
For Equations (M8.12) we develop a linear state space model, with the states, outputs, input, and
disturbances as the following deviation variables:

The Jacobian matrix elements are

where

The first column of the B matrix is (this column is related to input 1, the jacket feed flow rate)

If both the reactor temperature (state 2) and the jacket temperature (state 3) are measured for
a cascade control structure, the measured output matrix is
[ Team LiB ]
[ Team LiB ]

Module 9. Steam Drum Level

Steam utility systems are prevalent in almost all manufacturing complexes, as well as in many
other commercial systems (including hospitals, universities, etc.). Since these systems are so
important and have been in existence for many years, the control strategies are reasonably
advanced. In this learning module, the instructors and students can choose their preferred control
system design technique for an example steam drum system.

The sections of this module are as follows:

M9.1 Background

M9.2 Process Model

M9.3 Feedback Controller Design

M9.4 Feed-Forward Controller Design

M9.5 Three-Mode Level Control

[ Team LiB ]
[ Team LiB ]

M9.1 Background
Virtually every manufacturing plant has a steam utility system. The heart of the steam system is
the boiler (in general, there are a number of boilers in a plant). One of the important control loops
on a steam boiler is the steam drum level controller, which is shown schematically in Figure M9-1.
An actual boiler control problem consists of many other control loops, including steam pressure
control.

Figure M9-1. Control instrumentation diagram for boiler level.

The control of steam drum level is tougher than most other level control problems because there
is inverse response combined with a ramp in the level when a step increase in feed-water flow
rate is made. A typical step response is shown in Figure M9-2. The ramp nature (integrator) is
because if there an increase in the incoming water with no corresponding increase in outlet steam
flow, then the level must continually increase. The inverse response occurs because the incoming
cold water causes some of the bubbles in the steam drum to collapse, causing the level to drop
slightly. However, since the incoming flow rate is higher that the outgoing flow rate, the level
must eventually start rising.

Figure M9-2. Response of steam drum level (deviation from steady

state) to a step change in feed-water flow rate.
This is an example of a system where constraints on the process output (steam drum level) are
important. If the liquid level gets too high, liquid can enter the superheater section, expanding
rapidly and causing pipe rupture. If the liquid level gets too low, then water is no longer in the
pipes in the radiation section below the steam drum, causing the pipes to get too hot and fail.

[ Team LiB ]
[ Team LiB ]

M9.2 Process Model

The relationship between feed-water flow rate and drum level for the process studied in this
module is

Equation M9.1

Assume that the valve transfer function is

Equation M9.2

the unit of time is minutes, and the input and output have been scaled so that the full range is –1
to 1, with a nominal value of 0. The constraints on the inputs and outputs are then -1 and 1.

Now consider process disturbances. The primary disturbance (header pressure) affects the drum
level in the following fashion

Equation M9.3

[ Team LiB ]
[ Team LiB ]

M9 Feedback Controller Design

Instructors and students are asked to choose from a variety of controller design options. The
following are suggestions: IMC-based PID, Ziegler-Nichols (and Tyreus-Luyben), and frequency
response.

For the selected controller design technique, perform simulations for the following:

a. A level setpoint change of 0.1 (dimensionless) at t = 1 minute.

b. A steam demand change of 0.5 (dimensionless) at t = 1 minute.

In each case discuss the effect of controller tuning. In addition to the response of stream drum
level (process output), include plots of the manipulated input (valve position). You may wish to
revise the SIMULINK diagram shown in Appendix M9.1, for your simulations.

IMC-Based PID
Use the IMC-based PID procedure to find that the resulting feedback controller: assuming all-pass
and allowing the IMC controller to be semiproper. The resulting feedback controller will be lead-
lag [neglect the valve dynamics in controller design but include it in the simulation, that is, use
Equation (M9.1) for controller design].

a. What is the PID controller transfer function? (You should find that there is no integral term.)

b. Plot the closed-loop response for a step setpoint change of 0.1 at t = 1 minute. Compare the
closed-loop responses for several values of the IMC filter factor, l.

c. Plot the closed-loop response for a step load disturbance change of 0.5 at t = 1 minute.
Compare the closed-loop responses for several values of the IMC filter factor, l. You should
find that there is offset since there is no controller integral action.

You found that the feedback-only controller had offset when a load disturbance occurred. This is
because the controller does not have integral action. An IMC-based PID controller with integral
action can be designed for the integrating process with inverse response

by using the following IMC filter, and by not forming an all-pass factorization
and by letting g = 2l + b to obtain integral action. The resulting ideal PID controller parameters
are:

Ziegler-Nichols (or Tyreus-Luyben)

Using the Ziegler-Nichols continuous-oscillation procedure, find the critical gain and ultimate
period. Based on the critical gain and ultimate period, choose either the Ziegler-Nichols or Tyreus-
Luyben values for PI and PID controllers.

a. What are the values of the PI and PID tuning parameters?

b. Plot the closed-loop response for a step setpoint change of 0.1 at t = 1 minute. Compare the
closed-loop responses the PI and PID controllers.

c. Plot the closed-loop response for a step load disturbance change of 0.5 at t = 1 minute.
Compare the closed-loop responses for the PI and PID controllers.

Frequency Response
Design a PI controller with a gain margin of 2 and a phase margin of at least 60°.

a. What are the values of the PI controller? What are the gain and phase margins and the
relevant frequencies?

b. Plot the closed-loop response for a step setpoint change of 0.1 at t = 1 minute.

c. Plot the closed-loop response for a step load disturbance change of 0.5 at t = 1 minute.

[ Team LiB ]
[ Team LiB ]

M9.4 Feed-Forward Controller Design

Design a static feed-forward controller for this system.

Compare responses of the feed-forward control (combined with the feedback control designed in
the first problem) strategy with feedback control for a primary load step change of 0.5 at t = 1
minute.

[ Team LiB ]
[ Team LiB ]

M9.5 Three-Mode Level Control

In practice, a so-called three-mode level control strategy is used, as shown in Figure M9-3. The
purpose of the flow controller is to reject boiler feedwater flow disturbances (due to upstream
pressure changes or valve stiction, for example), while the feed-forward controller rejects steam
demand disturbances. The level controller provides additional long-term corrective action.

Figure M9-3. Control instrumentation diagram for three-mode steam

drum level control.

Develop a SIMULINK diagram for the 3-mode strategy, similar to that shown in Appendix M9.2.
Tune the flow controller for tight performance (closed-loop time constant of roughly 0.1 minutes)
and compare the 3-mode strategy with the standard feedback strategy for disturbances affecting
the feed water valve position.

[ Team LiB ]
[ Team LiB ]

Appendix M9.1: SIMULINK Diagram for Feed-

Forward/Feedback Control of Steam Drum Level

[ Team LiB ]
[ Team LiB ]

Appendix M9.2: SIMULINK Diagram for 3-Mode Control of

Steam Drum Level

[ Team LiB ]
[ Team LiB ]

Module 10. Surge Vessel Level Control

Surge vessels are used to help reduce the effect of flow rate variations between interconnected
process units. It is not necessary to maintain tight level control in a surge vessel, resulting in a
significantly different control objective compared to that of the steam drum level controller of
Module 9.

The sections of this module are as follows:

M10.1 Background

M10.2 Process Model

M10.3 Controller Design

M10.4 Numerical Example

M10.5 Summary

[ Team LiB ]
[ Team LiB ]

M10.1 Background
Much of this text has focused on tracking setpoint changes, where the goal has often been to
design a controller to yield a first-order type of response to a step setpoint change. There are
many processes, however, where disturbance rejection (regulatory control) is more important
than setpoint tracking, because setpoints may not be changed frequently. Consider the surge
drum level control system shown in Figure M10-1. The purpose of the surge vessel is to smooth
variations in the flow from process 1 and maintain a relatively constant flow rate to process 2.
The level can vary substantially from the setpoint, as long as the vessel does not overflow or go
dry. The main objective is to vary the manipulated flow rate (the outlet flow from the vessel) as
little as possible, while satisfying level constraints.

Figure M10-1. Surge vessel level control. (a) Vertical orientation. (b)
Horizontal orientation.

The goal of this module is to develop guidelines for tuning level controllers for surge vessels. The
performance specifications are significantly different than for the steam drum level systems
studied in the previous module.
[ Team LiB ]
[ Team LiB ]

M10.2 Process Model

Here, we develop a process model for the vessel in Figure M10-1a, which is a vertically oriented
cylinder that can be assumed to have a constant cross-sectional area. The development of a
model for the vessel in Figure M10-1b is left as an exercise for the reader. A material balance
around the vessel, assuming constant density and cross-sectional area (A), yields

Equation M10.1

The deviation variables for the manipulated input, load disturbance, and measured output are

Equation M10.2

The process and load disturbance transfer functions are then

Equation M10.3

Notice that this is an integrating process, since the pole = 0.

[ Team LiB ]
[ Team LiB ]

M10.3 Controller Design

Here we consider a P-only controller, since offset is less important than minimizing manipulated
variable movement. The question is:

What value of the proportional gain should be used?

It seems reasonable to select a value for the proportional gain that will just assure that
constraints will not be violated for the worst-case expected disturbance. Assume that level alarms
are set at 20% and 80% of the volume of the vessel. Let DH represent the magnitude of the
maximum allowable deviation of the tank height from setpoint, and let DL represent the maximum
magnitude of a step disturbance.

Proportional Gain
Assuming that the flow rate out is proportional to the deviation from setpoint,

Equation M10.4

where kc is a negative number since the flow rate must increase if the tank height is above the
setpoint value. The step change in the load disturbance is DL,

and the value at the new steady-state after a step disturbance is [from Equation (M10.1)]

which yields

Equation M10.5

Nonlinear Proportional Gain

Since small deviations from setpoint are not critical, it may be desirable to have a controller gain
that varies as a function of the magnitude of the error. A possible formulation is
Equation M10.6

Again, the value of the kc0 term can be determined from the proportional gain term in Equation
(M10.5),

Equation M10.7

[ Team LiB ]
[ Team LiB ]

M10.4 Numerical Example

Consider a surge vessel that has a diameter of 1 m, and a maximum height of 2 m, with low- and
high-level alarms set at 20% and 80%. Also, assume that the height setpoint is 1 meter, and the
steady-state flow rate is 0.5 m3/min. The constant cross-sectional area is then 0.785 m2. Since
the steady-state flow rate is 0.5 m3/min, assume that the inlet flow rate can vary between 0 and
1 m3/min.

You should show that the process and load disturbance transfer functions are

Since the physical variable minimum and maximum alarm heights are 0.4 and 1.6 m,
respectively, in terms of deviation variables, the minimum and maximum heights are

So, DH = 0.6 m. The minimum and maximum manipulated variable values are

So, DL = 0.5 m3/min. From Equations (M10.5) and (M10.7), we find kc = -0.833 m2/min and kc0
= -1.389 m/min.

Step Disturbances
Compare the step responses of the linear and nonlinear controllers to a small step disturbance of
0.05 m3/min and a large step disturbance of 0.5 m3/min. For a small step change in the load
disturbance (inlet flow rate), the linear and nonlinear controllers are shown in Figure M10-2.
Notice that the manipulated input changes more slowly for the nonlinear controller. The minor
disadvantage is that the offset is larger for the nonlinear controller, as designed. This is not a real
problem, as the offset is small compared with the capacity of the vessel.

Figure M10-2. Response to a small step in inlet flow rate (0.05

m3/min).
Responses to a large step in the inlet flow rate are compared in Figure M10-3. For this case, these
responses are similar and there is not much incentive to use the nonlinear strategy.

Figure M10-3. Response to a large step in inlet flow rate (0.5 m3/min).
Sinusoidal Disturbances
Consider a sinusoidal disturbance with an amplitude of 0.05 m 3/min and a frequency of 1/min.
The linear and nonlinear strategies are compared in Figure M10-4. Clearly, there is an incentive
for nonlinear control, since the manipulated variable does not change nearly as much for the
nonlinear case as for the linear case.

Figure M10-4. Response to a small amplitude (0.05 m3/min) sinusoidal

inlet flow rate disturbance.

The responses to large-magnitude sinusoidal disturbance with an amplitude of 0.5 m3/min and a
frequency of 1/min is shown in Figure M10-5. In this case, there is not much incentive for the
nonlinear control strategy.

Figure M10-5. Response to a large-amplitude (0.5 m3/min) sinusoidal

inlet flow rate disturbance.
You should show that a high-frequency disturbance is rejected very well by the nonlinear
strategy.

Additional Simulations
Consider the large-magnitude sinusoidal disturbance shown in Figure M10-5. Show that doubling
the amplitude of the disturbance leads to saturation of the manipulated variable, and to violation
of the high- and low-level alarms.

[ Team LiB ]
[ Team LiB ]

M10.5 Summary
For surge vessels it is generally more important to allow levels to "float" in order to minimize flow
rate variations. The approach shown in this module uses a gain that is proportional to the
absolute value of the error; this is effectively a nonlinear controller, where the control action is a
function of the square of the error. Nonlinear control was shown to lead to reduced control action
compared with simple P-only control.

[ Team LiB ]
[ Team LiB ]

Reference

Nonlinear level controllers with integral action are discussed in the following:

Marlin, T. E., Process Control: Designing Processes and Control Systems for Dynamic
Performance, 2nd ed. McGraw Hill, New York (2000).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Develop a nonlinear model for the rate of change of tank height for the horizontal vessel
orientation shown in Figure M10-1b. Consider a surge vessel that has a diameter of 1 m
and a length of 2 m, with low- and high-level alarms set at 20% and 80%. Also, assume
that the height setpoint is 0.5 m, and the steady-state flow rate is 0.5 m3/min. The
steady-state flow rate is 0.5 m3/min, so assume that the inlet flow rate can vary
between 0 and 1 m3/min. Design a proportional controller for this system. For the same
disturbances given in the example, compare the performance of the horizontal tank with
that of the vertical drum.

2: Implement a so-called gap controller with the following form

on the level control problems discussed in this chapter. Revise the SIMULINK diagram
shown in Figure M10-6 for your simulations. Discuss your selection of kcmin, kcmax, and
ebreak.

Figure M10-6. SIMULINK diagram for surge level control

simulations.

3: Consider a more general form of controller than Equation (M10.6). Perform simulations
 using a controller with the form

Discuss the effect of the two tuning parameters (kc0 and b) on the response to different
amplitude step and sign disturbances.

[ Team LiB ]
[ Team LiB ]

Appendix M10.1: The SIMULINK Block Diagram

[ Team LiB ]
[ Team LiB ]

Module 11. Batch Reactor

Batch chemical reactors are commonly used in the specialty chemical and pharmaceutical
industries. The focus of this textbook has been on continuous processes that operate at some
nominal steady-state value. Batch processes are dynamic by nature and usually have a time-
varying setpoint profile. Some of the challenges of batch reactor control are outlined in this
module.

After studying this module, the reader should be able to do the following:

Understand the effect of scale-up on the dynamic characteristics of a jacketed batch

chemical reactor. Further, understand how these characteristics affect the control-system
design and operating performance.

Understand some common control configurations and controller tuning techniques. Tuning
for ramp vs. step setpoint changes, and tuning for disturbance rejection.

M11.1 Background

M11.2 Batch Model 1: Jacket Temperature Manipulated

M11.3 Batch Model 2: Jacket Inlet Temperature Manipulated

M11.4 Batch Model 3: Cascade Control

M11.5 Summary

[ Team LiB ]
[ Team LiB ]

M11.1 Background
Jacketed batch chemical reactors are often used in the pharmaceutical and specialty chemicals
industries. Very often these reactors need to have both "heating" and "cooling" modes. That is,
the temperature of a reactor is often ramped up from the ambient reactor charge conditions (near
20°C) to a temperature where the reaction begins to "take off." If the reaction is exothermic, the
heat released through the reaction must be removed by circulating a cold fluid through the jacket.

A common strategy is shown in Figure M11-1, where both hot and cold jacket feed streams are
available. Here we have shown a split-range jacket temperature control strategy. When the jacket
temperature controller output is between 0 and 50%, the cold glycol valve is open. When the
jacket temperature controller output is between 50 and 100%, the hot glycol valve is open. The
split-range behavior is shown in Figure M11-2. Notice that the cold glycol valve is fail-closed, while
the hot glycol valve is fail-open.

Figure M11-1. Batch reactor temperature control. The jacket

temperature controller has a split-range output, where the cold glycol
valve is open during "cooling mode" and the hot glycol valve is open
during "heating mode."

Figure M11-2. Depiction of the split-range controller action.

An alternative approach is for the reactor temperature controller output to be the setpoint to a
jacket inlet temperature controller, as shown in Figure M11-3. In Section M11-2 we will focus on
the strategy shown in Figure M11-1. Simplifying assumptions will be used to ease the design of
the reactor temperature controller, and to better understand the efffect of reactor scale.

Figure M11-3. Batch reactor temperature control. The jacket inlet

temperature controller has a split-range output, where the cold glycol
valve is open during "cooling mode" and the hot glycol valve is open
during "heating mode."

[ Team LiB ]
[ Team LiB ]

M11.2 Batch Model 1: Jacket Temperature Manipulated

In this section we focus on the development of a mathematical model for the reactor temperature
dynamics; we will assume that the jacket temperature is directly manipulated. A number of other
assumptions will be made in the development of a mathematical model to describe the dynamic
batch reactor behavior. First of all, we will assume that the reactor is perfectly mixed. Also, for
simplicity we lump all reaction heat effects into a single heat generation term, qgen . Physical
parameters, such as density and the heat transfer coefficient, are assumed to be constant.

Equation M11.1

where U = heat transfer coefficient, A = area for heat transfer.

Notice that the accumulation term includes the effect of the "thermal mass" of the reactor wall

Equation M11.2

where (mcp)m is the "thermal mass" (mass * heat capacity, including the metal as well as the
glass lining) of the reactor wall and any other inert components (agitator, baffles, etc.). Also,
(mcp)f is the contribution of the reactor fluid (Vrcp). For small (laboratory) scale reactors the
contribution of the reactor wall and other inert components can be significant, while the relative
contribution is insignificant for large scale reactors. Equation (M11.1) can be rewritten

Equation M11.3

The fundamental model shown in (M11.3) is linear. The more familiar state-space form is

Equation M11.4a

Or, using deviation variable form

Equation M11.4b
where reactor temperature is the state (and output), jacket temperature is the manipulated
input, and heat generation is the load disturbance.

The resulting input-output transfer function models relating the manipulated and disturbance
inputs to the measured reactor output are first-order (see Additional Exercise 1)

Equation M11.5

Effect of Scale (Size)

Notice that the time constant is the "thermal mass" divided by the UA. If both of these terms
increased proportionally with the reactor volume, then studies conducted in a laboratory would be
directly applicable to large-scale production vessels. The problem is that the mass of the fluid
increases linearly with reactor volume, but the heat-transfer area does not. As a result, the time
constant of the reactor increases as the reactor gets larger. Essentially, on a per-volume basis,
the ability to transfer heat goes down as the volume increases.

In the following studies we will consider three different reactors: (i) laboratory scale (1 liter), (ii)
pilot-plant scale (0.5 m3), and (iii) manufacturing scale (10 m3). We will assume that the heat-
transfer coefficient is the same for each of these reactors, but that the heat transfer area
changes. The heat transfer coefficient used is

Assume that the reactor is a vertical cylinder with a height/diameter ratio of 2, and that a
heating/cooling jacket fully covers the bottom and sides of the reactor. This yields the results
shown in Table M11-1.

Quasi-Steady-State Behavior
If the heat-generation term is constant, then there exists a steady-state temperature difference
between the reactor and jacket such that the heat flows are balanced. From Equation (M11.1) we
see
 Table M11-1. Reactor Characteristics

V 0.001 m3 0.5 m3 10 m3

A 0.0523 m2 3.295 m2 24.28 m2

UA 0.0219 kW/°C 1.379 kW/°C 10.16 kW/°C

Problem M11.1 Effect of Reactor Scale on Heat Transfer Capability

For each scale of reactor, determine the heat flow, on a per-volume basis, if the temperature
difference between the reactor and jacket is 20°C. How much more "efficient" at transferring heat
is a 1 liter reactor compared to a 10-m 3 reactor?

Problem M11.2 Effect of Reactor Scale on Process Time Constant

For each scale of reactor, find the process time constant (tp) assuming that the process fluid is
water and neglecting the thermal effect of the reactor mass. For controller tuning, we suggest
that you use an initial l value of 1/4 tp for future simulations.

IMC-Based Design
In the Internal Model Control (IMC) design procedure used throughout most of this text we have
focused on the response to step setpoint changes. IMC can also be designed for ramp setpoint
changes or for the rejection of process input disturbances. What differs about the design is the
specification of the filter. For each case the filter is shown below.

Step setpoint change

Ramp setpoint change

Input load disturbance

where n is chosen to make the internal model controller physically realizable. For a first-order
process, you should verify the following results.

Design IMC, q(s) IMC-based PID, gc(s)

Step setpoint

(see Table 9-1A)

Ramp setpoint

Input disturbance (see Table 9-1B)

It is very important to use antireset windup (ARW) strategies when using PID controllers on batch
reactors, since the manipulated inputs are often constrained. For the following problems assume
that the jacket temperature is constrained to be between (–10°C and 150°C). For each of the
following problems use the corresponding controller design.

Problem M11.3 IMC-Based PI Control: Step Setpoint Tracking

Here we require that the reactor temperature setpoint be changed from 20°C to 50°C. Compare
responses to step setpoint changes, for each size reactor. Discuss the practical limitations to the
expected dynamic performance for each size reactor, especially considering the constraints on
jacket temperature.

Problem M11.4 IMC-Based PI Control: Ramp Setpoint Tracking

Here we require that the reactor temperature setpoint be changed from 20°C to 50°C. Compare
responses to ramp setpoint changes, for each size reactor. Discuss the practical limitations to the
expected dynamic performance for each size reactor, especially considering the constraints on
jacket temperature. Also, what are practical limitations on the ramp time as a function of scale?
For example, perhaps a 10-minute ramp time is fine for the laboratory scale reactor; is it feasible
for the larger scale reactors?

Problem M11.5 IMC-Based PI Control: Disturbance Rejection

After the reactor has been heated to 50°C, assume that a reaction heat flow of 0.05 kW/liter (50
kW/m3) is initiated. Compare the closed-loop performance as a function of reactor scale; compare
the maximum deviation from the desired value of 50°C for each scale reactor.

[ Team LiB ]
[ Team LiB ]

M11.3 Batch Model 2: Jacket Inlet Temperature

Manipulated
In this section we focus on the development of a mathematical model for the reactor and jacket;
we will assume that the jacket inlet temperature is manipulated, and the jacket flowrate is
constant. This model is appropriate for the strategy shown in Figure M11-3. A number of other
assumptions will be made in the development of a mathematical model to describe the dynamic
batch reactor behavior. First of all, we will assume that the reactor is perfectly mixed. Also, for
simplicity we lump all reaction heat effects into a single heat- generation term, qgen . Similarly, the
jacket is perfectly mixed; often, agitation nozzles are used to create a high "swirl velocity" in the
jacket, so this assumption will be fine on the time scale of interest. Physical parameters, such as
density and the heat transfer coefficient, are assumed to be constant. The dynamics of the
recirculation heat-transfer system are neglected in this model. Energy balances on the reactor
and jacket yield the following equations:

Equation M11.6

where the accumulation terms include the effect of any "thermal mass" of the reactor and jacket
walls. For example, the reactor energy balance includes the term

where (mcp)m is the "thermal mass" (mass * heat capacity, including the metal as well as the
glass lining) of the reactor wall and any other inert components (agitator, baffles, etc.). Also,
(mcp)f = Vrcpf is the contribution of the reactor fluid. The term represents the mass flow rate
of the jacket fluid.

Equation (M11.6) can be rewritten

Equation M11.7

The linear state space model is

Equation M11.8
The resulting input-output transfer function models relating the manipulated and disturbance
inputs to the measured reactor output are (see Additional Exercise 2)

Equation M11.9

Problem M11.6 Effect of Reactor Scale on Process Parameters

For each scale of reactor (Table M11-1), find the process parameters assuming that the process
and jacket fluid is water and neglecting the thermal effect of the non-fluid components of the
reactor and jacket. Assume that the jacket volume is 1/4 of the reactor volume.

IMC-Based PID Tuning Parameters

Using the IMC-based PID tuning rules, with a desired first-order closed-loop response with a time
constant of l, we find

Equation M11.10

We can also use the following order of magnitude analysis to reduce the expressions
so

Equation M11.11

Also, as in Section M11.2, let the closed-loop time constant be roughly 1/4 of the open-loop time
constant, so the resulting controller proportional gain is

and the proportional band is then 25%. For a PI controller, we can start with the same parameter
values, but set the derivative time constant to 0.

Problem M11.7 IMC-Based PI Control: Step Setpoint Tracking

Here we require that the reactor temperature setpoint be changed from 20°C to 50°C. Compare
responses to step setpoint changes, for each size reactor. Discuss the practical limitations to the
expected dynamic performance for each size reactor, especially considering the constraints on
jacket inlet temperature (–10 to 150°C).

[ Team LiB ]
[ Team LiB ]

M11.4 Batch Model 3: Cascade Control

The process model, if the jacket make-up flow rate is the manipulated input, is

Equation M11.12

As written, this assumes that the jacket is a "once through" flow and that there are cold and hot
jacket feed streams. It can also be shown that a jacket recirculating heat-transfer system has the
same model equations, where the jacket mass flow rate is simply the make-up flow to the
recirculating jacket heat-transfer system [compare Equations (M8.11) and (M8.12) for a CSTR].
Here a split-range control strategy would be used, with either the cold or hot fluid stream
admitted to the jacket, depending on the split-range controller output.

Linearizing Equation (M11.12), we find the following state space model matrices

where inputs 1 and 2 (corresponding to columns 1 and 2 of the B matrix) are the cold and hot
jacket feedstreams, respectively. The third input is the disturbance (reaction heat flow), qgen .
Since the measured outputs are the reactor and jacket temperature,

Problem M11.8 Simulation Model for Intermediate Scale Reactor

For the 0.5 m3 reactor, assuming that both the reactor and jacket fluids are water, and the jacket
volume is 25% of the reactor volume (and neglecting the thermal mass of the non-fluid
components), find the state space model. For the initial steady-state assume that there is no
reaction heat flow, so the cold and hot fluid flow rates are zero initially. What do you notice about
the eigenvalues of the A matrix? Find the transfer functions relating the manipulated inputs to the
outputs; what do you notice about one of the poles? Observe the responses to step changes in
the manipulated inputs. Are these realistic? (Hint: You should find that the responses are
integrating in nature; this is expected for the initial responses, but not for the long-term
behavior.)

Problem M11.9 Cascade Control with Split-Range Jacket Temperature Controller (this problem is
better for a longer-term project)

Develop a jacket temperature controller that has a split-range output (see Chapter 12 for an
example SIMULINK model for a stirred-tank heater). The controller output should be similar to
Figure M11-2. Design the jacket temperature (secondary) controller using any method that you
desire. Next, design the reactor temperature (primary) controller.

Perform simulations where the reactor temperature setpoint is changed from 20°C to 50°C.
Compare responses to step setpoint changes, for each size reactor. Discuss the practical
limitations to the expected dynamic performance for each size reactor, especially considering the
constraints on jacket inlet temperature (–10 to 150°C).

It is recommended that both the primary and secondary controllers include antireset windup
(Chapter 11) strategies.

[ Team LiB ]
[ Team LiB ]

M11.5 Summary
This module has assumed very simple models for batch reactors, where heat effects (due to an
exothermic reaction, for example) are captured by a simple heat generation term.

This resulted in low-order models that demonstrated clearly the importance of process scale. A
smaller reactor has a larger heat transfer area to volume ratio, making it much more efficient at
transferring heat compared to a larger reactor. That is, smaller temperature differences between
the reactor and jacket are needed to remove heat energy from the reactor. Smaller reactors then
are less prone to suffer from problems with jacket temperature constraints. In addition, the
smaller reactors have smaller process time constants, allowing better closed-loop temperature
control (faster closed-loop responses). Certainly, more complicated models can be developed,
which include reactor rate terms and semi-batch reactor operation (see Additional Exercise 4).
More comprehensive examples will be placed on the textbook web page; you are encouraged to
download these files for advanced study.

[ Team LiB ]
[ Team LiB ]

Reference
The following article provides an excellent overview of both batch and continuous reactor control:

Lipták, B. G., "Controlling and Optimizing Chemical Reactors," Chem. Eng. Mag., pp. 69–81 (May
26, 1986).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Find the transfer function relationships shown in Equation (M11.5).

2: Find the transfer function relationships shown in Equation (M11.9).

3: Pfaudler is a major manufacturer of reaction vessels. Search the Pfaudler web page to
find 100-gallon and 1500-gallon reactors. Do the heat transfer areas vary with volume
in the way that you would expect?

4: All of the examples in this module have included a heat generation term, for simplicity.
This resulted in simple, low-order models. Develop a more comprehensive model of a
semi-batch reactor (flow in, but not out), with a simple first-order irreversible reaction
(A B). Develop the modeling equations, including the following states: V (liquid
volume), NA (total amount of component A in the reactor), T (reactor temperature), and
Tj (jacket temperature). Allow the heat-transfer area to vary as a function of the reactor
volume. Include a steady-state energy balance on the recirculating jacket fluid heat-
transfer system, and assume that a split-range controller manipulates either the cold or
hot glycol flow rate.

[ Team LiB ]
[ Team LiB ]

Module 12. Biomedical Systems

There are a number of important applications of feedback control for biomedical systems. In this
module we present a number of examples and suggest that instructors or students pick one of the
topics for each interactive learning session. The goals are to learn more about modeling and
control by applying the basic concepts to important problems in biomedical systems.

The sections of this module are as follows:

M12.1 Overview

M12.2 Pharmacokinetic Models (Chapters 2 and 3)

M12.3 Intravenous Delivery of Anesthestic Drugs (Chapters 5–7)

M12.4 Blood Glucose Control in Diabetic Patients (Chapters 5–9)

M12.5 Blood Pressure Control in Post-Operative Patients (Chapters 6–9)

M12.6 Critical Care Patients (multivariable control—Chapters 13 and 14)

M12.7 Summary

[ Team LiB ]
[ Team LiB ]

M12.1 Overview
Physiological systems are composed of a large number of feedback control loops. Examples
include the baroreceptor reflex that regulates blood pressure, and the synthesis of insulin and
glucagon by the pancreas to regulate blood glucose. The focus of the examples presented in this
module is on exogenous feedback regulation using devices external to the human body.

There is a rich history of automated control applications in biomedicine. Initial studies on feedback
control of anesthesia were performed by Bickford at the Mayo Clinic in the 1950s. Several
different anesthetics were studied as the manipulated input, while the measured variable was the
integrated rectified amplitude of the EEG. The control strategy did not gain acceptance because it
was not clear what feature of the EEG signal should be used as an indicator of the depth of
anesthesia.

Modeling biomedical processes can be quite challenging. For physiological systems, drug material
balances can be written based on a compartmental model; that is the body or tissues are
"lumped" into "volumes" in the same fashion that stirred tanks are used to represent chemical
process behavior.

[ Team LiB ]
[ Team LiB ]

M12.2 Pharmacokinetic Models

Pharmacokinetics is the study of the dynamic behavior of drug concentrations in blood and
tissues. Compartmental models are widely used to conceptually describe this behavior. One
common 3-compartment model is shown in Figure M12-1.

Figure M12-1. Three-compartment model.

Compartment 1 is often used to represent blood plasma. The drug is assumed to be directly
infused to compartment 1. The drug can then diffuse between compartment 1 and compartments
2 and 3, based on the transfer parameters, k ij. For example, k12 represents the kinetics for
transfer of the drug from compartment 1 to compartment 2. Compartment 2 can be thought of as
well-perfused tissues, such as muscle and brain, while compartment 3 represents residual tissues
and bone.

Modeling Equations
If we let I represent the mass infusion rate (mass/unit time) of the drug, the overall material
balance can be written as

Equation M12.1
This has the form of a state space model

Equation M12.2

One is normally concerned more about drug concentrations rather than the total amount of drug.
The blood plasma concentration can be represented by C 1 = M1/V1. The state space model for the
measured output is then

Equation M12.3

Alternatively, the states can be formulated as concentrations, resulting in the following model:

Equation M12.4

[ Team LiB ]
[ Team LiB ]

M12.3 Intravenous Delivery of Anesthetic Drugs

Traditionally, most anesthetics have been vapor-phase, such as isoflurane. In recent years there
has been a move toward the intravenous (IV) delivery of drugs for anesthesia. One of the more
popular IV anesthetics is propofol. For a particular population of subjects, the following kinetic
values have been determined: k 10 = 0.152, k 12 = 0.207, k13 = 0.040, k 21 = 0.092, and k31 =
0.0048, all with units of min-1. Also, assume that the volume of the blood plasma compartment is
V1 = 12 liters.

1. What is the required steady-state propofol infusion rate to maintain a desired blood
concentration of 2 mg/ml?

2. Develop the state space model, based on Equations (M12.2) and (M12.3 or M12.4),
assuming that the plasma volume is 12 liters. What are the units of the manipulated input
(propofol infusion rate) and measured output (blood concentration)?

3. Develop the process transfer function model, if the manipulated drug infusion rate has units
of mg/hr, and the blood concentration has units of mg/ml.

4. Develop a feedback-control strategy to control blood concentration by manipulating the

propofol infusion rate. Make a setpoint change from 1 to 2 mg/ml. Are the closed-loop
response and drug infusion rate feasible?

[ Team LiB ]
[ Team LiB ]

M12.4 Blood Glucose Control in Diabetic Patients

The human body has many innate feedback-control loops. For example, blood glucose is regulated
by the production of insulin by the pancreas. When food is consumed and broken down by the
digestive system, the blood glucose level rises, stimulating insulin production. Cells use insulin to
break down the glucose. People with diabetes have a reduced capability of producing insulin. Type
I diabetes mellitus patients cannot produce any insulin and must administer insulin shots several
times a day to help regulate their blood glucose level. A typical patient is then serving as a control
system. Some actions are feedforward in nature, such as the administration of an insulin shot to
coincide with meal consumption. Other actions are feedback in nature, such as when the patient
changes dosages based on blood glucose measurements, obtained from "finger pricks" and
analysis of glucose strips. In the long-term, hyperglycemia (high blood sugar) can lead to
blindness and cardiovascular problems. In the short-term, hypoglycemia (low blood sugar) can
lead to fainting or a diabetic coma.

Clearly, there is a tremendous motivation for the development of closed-loop insulin delivery
systems. Current technology involves external pumps with insulin reservoirs that can deliver
insulin continuously, rather than having "pulses" due to shots administered several times a day.
With an external infusion pump, the patient must still perform tests to monitor blood glucose
levels.

There is a major research effort to develop implantable glucose sensors and insulin infusion
pumps. An implanted sensor will measure glucose and send a telemetry signal to a small,
wristwatch-sized controller that can send another telemetry signal to a small implanted insulin
pump. This closed-loop system essentially acts as an artificial pancreas, yielding a lifestyle that
differs little from a nondiabetic person on a day-to-day basis.

Nonlinear Model
One of the more widely used models of the effect of insulin infusion and glucose (meal) inputs on
the blood glucose concentration is known as the Bergman "minimal model" (Bergman et al.,
1981). This is described by three differential equations:

Equation M12.5

where G and I represent the deviation in blood glucose and insulin concentrations, respectively.
Also, X is proportional to the insulin concentration in a "remote" compartment. The inputs are
Gmeal, a meal disturbance input of glucose, and U, the manipulated insulin infusion rate. The
parameters include p1, p2, p3, n, and V1 (which represents the blood volume). Other parameters
are Gb and Ib, the "basal" (baseline or steady state) values of blood glucose and insulin
concentration. These values can be used to determine the basal infusion rate of insulin necessary
to maintain a steady state.

Linear Model
A linear state space model can be developed for use in control-system design. Here, define the
state, input, and output variables (in deviation form) as

Equation M12.6

where the first input is manipulated (insulin infusion) and the second input represents a meal
glucose disturbance. The state space model is

Equation M12.7

Example Set of Parameters

Consider a diabetic that is modeled using the following set of parameters (Lynch and Bequette,
2001):

It is necessary to be careful with units. Since the concentrations are in mmol/liter, and the
glucose disturbance has units of grams, we must apply a conversion factor of 5.5556 mmol/g to
the Gmeal term. Solving for the steady states, we find that the basal insulin infusion rate must be
Ub = 16.667 mU/min. For these parameters, the resulting state space model is

Equation M12.8
In the United States it is more common to work with glucose concentration units of mg/deciliter
rather than mmol/liter. Since the molecular weight of glucose is 180 g/mol, we must multiply the
glucose state (mmol/liter) by 18 to obtain the measured glucose output (mg/deciliter). This yields
the following state-output relationship

Equation M12.9

You should find that the process transfer function is

Equation M12.10

Due to pole/zero cancellation, the disturbance transfer function is

Equation M12.11

In practice, the glucose does not directly enter the blood stream. The processing in the gut before
entering the blood can be modeled as a first-order transfer function with a 20-minute time
constant. This means that the disturbance transfer function, including the lag in the gut, is

Equation M12.12

Also, unlike many chemical process disturbances, a meal disturbance is best modeled as a pulse.
Assume that a 50 g glucose meal is consumed over a 15-minute period; the pulse then has a
magnitude of 3.333 g/minute for a duration of 15 minutes.

Desired Control Performance

The steady-state glucose concentration of 4.5 mmol/liter corresponds to a measured glucose
concentration of 81 mg/deciliter. It is important that a diabetic maintain her/his blood glucose
concentration above 70 mg/deciliter (to avoid short-term problems with hypoglycemia, such as
fainting). Notice that the insulin infusion rate (manipulated input) cannot go below 0; it is critical
that you consider this constraint in your simulations.

A number of control strategies could be used. One suggestion is to simplify the process transfer
function (M12.10) to the following form

Equation M12.13

and design an IMC-based PD or PID controller. Also, since the diabetic knows when they are
consuming a meal, some type of feed-forward action can be used.

Perform disturbance rejection simulations and discuss your results. What feedback-only design
(and tuning) do you recommend? What are the minimum and maximum blood glucose values that
occur after a pulse meal consumption of 50 g glucose over a 15-minute period? Does your
strategy satisfy constraints on the insulin infusion rate? How much better is the performance if
feed-forward control is included? Compare the resulting blood glucose profiles with those found in
the literature.

[ Team LiB ]
[ Team LiB ]

M12.5 Blood Pressure Control in Post-Operative

Patients
Health care is an important area that can benefit from control systems technology. For example,
after undergoing open-chest surgery (such as cardiac bypass or lung surgery), many patients
have a higher than desired blood pressure; one reason that a lower blood pressure is needed is to
reduce bleeding from sutures. Typically, a critical care nurse will set up a continuous infusion of
the drug sodium nitroprusside to reduce the patient's blood pressure. The nurse must make
frequent checks of blood pressure and change the drug infusion rate to maintain the desired blood
pressure. In a sense, the nurse is serving as a feedback controller. There is a clear motivation to
develop an automated feedback control system, where a blood pressure sensor sends a signal to
a blood pressure controller, which adjusts the speed of a drug infusion pump. In addition to
providing tighter regulation of blood pressure, this strategy frees up the nurse to spend more time
monitoring patients for other problems.

A 70-kg patient exhibits the following input-output behavior for the effect of sodium nitroprusside
(SNP) on mean arterial pressure (MAP):

where the pressure unit is mm Hg, the drug flow units are ml/hr and the time unit is minutes. The
initial MAP (with no SNP infused) is 175 mm Hg. The desired MAP setpoint is 100 ± 15 mm Hg,
and the maximum allowable (short-term) SNP infusion rate is 150 ml/hr. The desired settling time
(time for the MAP to remain within limits) is 5 minutes.

1. Can a proportional-only controller be designed to yield an offset of ± 15 mm Hg while

meeting the infusion rate constraint?

2. Design an internal model controller for this system. What is the minimum l value that can
be used and still satisfy the drug infusion limit? What is the maximum l value that can be
used and still satisfy the settling time limit?

3. Design an IMC-based PI controller for this system, using the "improved" parameters (Table
9.2). Compare the closed-loop performance with a "pure" IMC strategy (developed in
problem 2).

[ Team LiB ]
[ Team LiB ]

M12.6 Critical Care Patients

Critical care patients have often suffered a "disturbance" to the normal operation of their
physiological system; this disturbance could have been generated by surgery or some sort of
trauma (e.g., a heart attack). A responsibility of the critical care physician is to main tain certain
patient outputs within an acceptable operating range. Two important outputs to be maintained
are mean arterial pressure (MAP) and cardiac output (CO). During surgery the anesthesiologist
will infuse several drugs into the patient in order to control these states close to the desired
values. A conceptual diagram is shown in Figure M12-2.

Figure M12-2. Drug infusion control.

The goal of this control system design is to manipulate the flow rate of two drugs, dopamine
(DPM) and sodium nitroprusside (SNP), to maintain the two outputs at their desired setpoints. A
successful implementation of such a strategy allows the anesthesiologist to spend more time
monitoring other patient states, such as "depth of anesthesia."

A simplified model representing the input-output behavior for a particular patient is

Equation M12.14

where inputs 1 and 2 are SNP and DPM (ml/hr), and outputs 1 and 2 are MAP (mmHg) and CO
(liters/min). The time constants have units of minutes.

1.
 1. Based on the RGA (Chapter 13), what are the natural pairings? Ordinarily, one would control
MAP by manipulating SNP and control CO by manipulating SNP. Does the RGA rule out this
strategy? Why or why not?

2. Consider the transfer function matrix. If the nominal steady-state drug infusion rates are 0,
can the cardiac output (CO) be decreased by infusing these drugs? Why or why not?

3. Perform an SVD analysis (Chapter 14), assuming that the transfer function matrix (M12.14)
is properly scaled. What are the strongest and weakest output directions for this process?

4. Assume that the initial steady states for MAP and CO are 120 mm Hg and 4 liters/min,
respectively. Design controllers and perform simulations for step setpoint changes to 100
mm Hg and 4.5 liters/min. Discuss controller performance and the different behavior if
setpoint changes are made individually versus simultaneously.

[ Team LiB ]
[ Team LiB ]

M12.7 Summary
Automation and control in biomedicine is an important area of research, with many potential long-
term health care benefits. Sensors, the difficulty of developing a reliable model, and "fuzzy"
performance specifications all create difficulties in feedback-control design. For each of the
examples presented in this module, it is recommended that you perform a literature review to
understand the current state of the art.

[ Team LiB ]
[ Team LiB ]

References
A nice textbook on this general topic is:

Northrop, R. B., Endogenous and Exogenous Regulation and Control of Physiological Systems,
Chapman and Hall/CRC, Boca Raton, FL (2000).

For examples of biomedical control applications, see the following special issue:

Bequette, B. W., and F. J. Doyle III (Eds.), IEEE Engineering in Medicine and Biology Magazine,
Special issue on automation in biomedicine, Jan/Feb (2001).

A comparison of sets of propofol pharmacokinetic parameters is presented in:

Vuyk, J., F. H.M. Engbers, A. G.O. Burm, A. A. Vletter and J. G. Bovill, "Performance of Computer-
Controlled Infusion of Propofol: An Evaluation of Five Parmacokinetic Parameter Sets," Anesth.
Analg., 81, 1275–1282 (1995).

Control of blood glucose in diabetics is discussed in the following papers:

Bellazzi, R., G. Nucci, and C. Cobelli, "The Subcutaneous Route to Insulin-Dependent Diabetes
Therapy," IEEE Engineering in Medicine and Biology Magazine, 20(1), 54–64 (2001).

Parker, R. S., F. J. Doyle III, and N. A. Peppas, "The Intravenous Route to Blood Glucose
Control," IEEE Engineering in Medicine and Biology Magazine, 20(1), 65–73 (2001).

The Bergman "minimal model" is discussed in the following paper:

Bergman, R. N., L. S. Philips, and C. Cobelli, "Physiological Evaluation of Factors Controlling

Glucose Tolerance in Man," J. Clin. Invest., 68, 1456–1467 (1981).

The parameters for the diabetes model are presented in the following paper:

Lynch, S. M., and B. W. Bequette, "Estimation-based Model Predictive Control of Blood Glucose in
Type I Diabetics: A Simulation Study," in Proceedings of the 27th Northeast Bioengineering
Conference, Storrs, CT, pp. 79–80 (2001).

A review of blood pressure control is presented by:

Isaka, S., and A. V. Sebald, "Control Strategies for Arterial Blood Pressure Regulation," IEEE
Trans. Biomed. Eng., 40(4), 353–363 (1993).

An experimental study of the simultaneous control of blood pressure and cardiac output is
presented in the paper:

Rao, R., C. C. Palerm, B. Aufderheide, and B. W. Bequette, "Experimental Studies on Automated

Regulation of Hemodynamic Variables," IEEE Engineering in Medicine and Biology Magazine,
20(1), 24–38 (Jan/Feb, 2001).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Much research has been conducted on the development of a measure of anesthetic

depth. A device that measured the bispectral index (BIS) is on the market and has been
used in some feedback-control studies. Perform a literature search and discuss at least
two papers that present feedback-control results based on a BIS measurement and the
manipulation of either an intravenous (such as propofol) or vapor-phase (such as
isoflurane) anesthetic.

2: Ventricular Assist Devices (VAD) help a weakened left ventricle to supply more
"pumping action" to the blood. Perform a literature search and discuss at least two
papers presenting VAD control studies.

[ Team LiB ]
[ Team LiB ]

Module 13. Distillation Control

Distillation remains the most common method of separating chemical components from a
mixture. In this module, we provide a simple example of a binary distillation column that has
behavior illustrative of more complex columns. The goal is to better understand the dynamic
behavior of multivariable, interacting systems. After studying this module the reader should be
able to do the following:

Understand the different control loops associated with distillation operation

Understand the importance of input "direction" when making step input changes

Understand how an input "direction" affects the magnitude and "direction" of the resulting
output change

Interpret SVD results in terms of these sensitive directions

Consider failure sensitivity when designing controllers

The sections of this module are as follows:

M13.1 Description of Distillation Control

M13.2 Open-Loop Behavior

M13.3 SISO Control

M13.4 RGA Analysis

M13.5 Multiple SISO Controllers

M13.6 Singular Value Analysis

M13.7 Nonlinear Effects

M13.8 Other Issues in Distillation Column Control

M13.9 Summary

[ Team LiB ]
[ Team LiB ]

M13.1 Description of Distillation Control

A schematic diagram of a typical binary distillation column control strategy is shown in Figure
M13-1. Notice that there are five measured variables (pressure, distillate receiver level, base
level, distillate composition, and bottoms composition) and five manipulated variables (condenser
heat duty, reflux flow rate, distillate flow rate, bottoms flow rate, and reboiler heat duty). In
addition, if the feed to the column is regulated, then there is an additional measurement (feed
flow rate) and manipulated input (also feed flow rate or valve position). Normally, it is assumed
that the pressure is controlled by manipulating the condenser heat duty (or the flow of cooling
water to the condenser), that the distillate receiver level is controlled by manipulating the
distillate product flow rate, and that the base level is controlled by manipulating the bottoms
product flow rate. It is also assumed that these loops are tightly tuned, and the process outputs
are "perfectly controlled" at their setpoint values.

Figure M13-1. Dual composition control of a distillation column.

For a number of reasons, composition sensors may not be installed or many not be in service. The
distillate and bottoms streams compositions can be inferred by measuring the temperatures of
trays near each end of the column, as shown in Figure M13-2. The tray temperature
measurement dynamics are typically more rapid than the product stream composition
measurements, so tighter disturbance rejection may be possible with this strategy. For
convenience, we assume throughout the rest of this module that composition measurements are
available. The analysis and design techniques presented in this module can easily be used for the
temperature-control system shown in Figure M13-2.

Figure M13-2. Dual temperature control of a distillation column.

Simplified Model
The column studied is a 41-stage binary separation. For ease of modeling, it is assumed that the
manipulated inputs are reflux flow rate and vapor boil-up rate. In practice, the steam flow rate to
the reboiler would be manipulated, but it is related to the vapor boil-up by the heats of
vaporization of the bottoms stream and the steam.

An input-output transfer function model for the dual composition control strategy is

Equation M13.1
The outputs are the distillate (xD) and bottoms (xB) composition, in mole fraction of light
component. The manipulated inputs are the reflux (L) and vapor boil-up (V) rates, in kilomoles
per minute. The disturbance inputs are feed flow rate (F, kmol/minute) and feed light component
mole fraction (zF). The unit of time is minutes. The steady-state compositions are 0.99 and 0.01
mole fraction light component for the distillate and bottoms product streams, respectively. The
steady-state manipulated inputs (reflux and vapor flow rates) are 2.706 and 3.206 kmol/minute,
respectively. The feed to the column is 1 kmol/minute, with a feed composition of 0.5 mole
fraction light component and a feed quality of 1.

[ Team LiB ]
[ Team LiB ]

M13.2 Open-Loop Behavior

Response to Reflux Change

Here, we consider open-loop step changes in the manipulated inputs, with additional time delays
of 2 minutes in the concentration measurements. The responses to step changes of ±0.01
kmol/minute in reflux at t = 10 minutes, while keeping vapor boil-up constant, are shown in
Figure M13-3. The response to a positive reflux flow change are shown as solid lines, while
responses to a negative reflux change are shown as dashed lines; also, we have chosen to plot
mole percent rather than mole fraction (the unit used in the transfer function matrix). The
responses are symmetric because we are using a linear model. Naturally, the bottoms
composition cannot decrease below 0, as obtained with a step decrease in reflux flow; this
illustrates a clear limitation to the use of a linear model.

Figure M13-3. Open-loop step response. Change of ±0.01 kmol/minute

in reflux flow rate.

Response to Vapor Boil-Up Change

The reader should generate responses to step changes of ±0.01 kmol/minute in the vapor boil-up
rate (V) at t = 10 minutes (see Additional Exercise 1). Compare and contrast these responses
with those of the reflux changes. Are these changes consistent with the transfer function matrix?
Response to Simultaneous Changes in Reflux and Vapor Boil-Up
It should be noted that the effect of the input "direction" is very important. Contrast the response
to reflux (+0.005 kmol/minute) and vapor (-0.005 kmol/minute) changes that are in the
"opposite" directions, with simultaneous positive (+0.005 kmol/minute) changes in reflux and
vapor. The responses shown in Figure M13-4 are two orders of magnitude larger than in Figure
M13-5, although the magnitude of the input changes are the same. The SVD analysis shown in
Section M13.6 is used to explain these results. Remember that this is a linear model, based on the
nominal steady-state conditions. These are not nonlinear effects; rather, they illustrate the
important of "direction" in multivariable processes, even when the processes are linear.

Figure M13-4. Open-loop response to a simultaneous +0.005

kmol/minute changes in reflux and –0.005 kmol/minute in vapor.

Figure M13-5. Open-loop response to simultaneous +0.005

kmol/minute changes in reflux and vapor.
[ Team LiB ]
[ Team LiB ]

M13.3 SISO Control

The IMC-based PID procedure was used to design the xD-L (loop 1) and xB-V (loop 2) controllers,
in each case assuming the "other loop" was open. Time delays were neglected in the control
system design, resulting in PI controllers

A first-order closed-loop response is achieved for each of these controllers operated independently
(only one loop closed at a time). The reader should verify, through simulation of the independent
controllers, that IMC filter factors of 25 minutes lead to closed-loop time constants of roughly 25
minutes (see Additional Exercise 2).

[ Team LiB ]
[ Team LiB ]

M13.4 RGA Analysis

The RGA (Chapter 13) for this system is

This RGA indicates that the xD-L (distillate composition-reflux flow) and xB-V (bottoms
composition-reboiler vapor flow) input-output pairings are the only possible pairings, since pairing
on a negative relative gain would lead to a failure-sensitive system. It also indicates that this
system is sensitive to model uncertainty and that significant tuning parameter changes must be
made (typically, the magnitude of the proportional gains must be increased) so that both loops
can be closed simultaneously with good performance. Very often, with high relative gain systems,
tuning parameters that work well with both loops closed will be unstable if one of the loops is
opened (either by an operator or by the failure or saturation of a valve). With high relative gain
systems, one must often decide to simply keep one loop under control and sacrifice the control of
the less important output variable.

[ Team LiB ]
[ Team LiB ]

M13.5 Multiple SISO Controllers

Setpoint Changes
In Section M13.3, the composition control loops were tuned independently using the IMC-based PI
technique, neglecting time delays. The control system was simulated assuming a 2-minute time
delay on all measurements. The IMC filter factors (l 1, l2) were both set to 25 minutes, one third
of the process time constant. Here, we show the responses for both loops closed simultaneously.
The response for a setpoint change of 1 mol% (0.01 mole fraction) in the distillate composition is
shown in Figure M13-6. Note that the controllers are highly interacting and the response time is
extremely slow (the outputs are still far from the setpoint after 1000 minutes).

Figure M13-6. Setpoint change of r = [1,0], mol% light comp. Note the
long timescale.

It is intriguing that simultaneous setpoint changes of 0.625 mol% in the distillate and 0.78 mol%
in bottoms composition lead to much more rapid closed-loop responses, as shown in Figure M13-
7. Notice here that the time scale is 100 minutes, and that the desired setpoint has been achieved
in 55 minutes.

Figure M13-7. Setpoint change of r = [0.625,0.78], mol% light comp.

Note the relatively short timescale compared to Figure M13-6.
Without any other information, one might assume that nonlinearities could be used to explain the
difference in the system behavior displayed in Figures M13-6 and M13-7. These simulations,
however, are based on the linear transfer function model. The differences observed are due to the
"directional sensitivity" of multivariable systems. Singular value analysis will be used in Section
M13.6 to better understand this behavior.

The reader should tune the controllers for tighter performance for the setpoint change in the
"slow" direction (Figure M13-8). Decrease l 1 and l2 (increase the magnitude of the proportional
gains) and observe the behavior (Additional Exercise 3). Also, once tighter control has been
achieved with both loops closed, consider the effect of opening one loop. You should observe that
the other loop goes unstable (Additional Exercise 4).

Figure M13-8. Responses to an increase in reflux (2.706–2.806) and

decrease in vapor (3.206–3.106) rates.
Disturbance Rejection
The focus of the previous sections has been on setpoint changes. It is also very important for any
control strategy to reject feed flowrate and feed composition disturbances. In Additional Exercise
9 you have the opportunity to perform simulations and discuss results for disturbance rejection.

[ Team LiB ]
[ Team LiB ]

M13.6 Singular Value Analysis

The SVD can be used to predict directional sensitivity of a process. The process gain matrix is
decomposed into three matrices,

where U is the left singular vector matrix, S the diagonal matrix of singular values, ordered, and V
the right singular vector matrix. The left and right singular vector matrices are both orthonormal
matrices; that is, each column of the matrix is orthogonal to all other columns and the columns
each are unit length. The diagonal singular value matrix is ordered so that the largest singular
value is in the (1,1) position. Note that the standard notation for SVD is to use U to represent the
left singular vector matrix. Please do not confuse this with the u vector commonly used to
represent the vector of manipulated inputs.

When performing a singular value analysis, it is important to scale the inputs and outputs to cover
the same range. For this system, we assume that we desire the output compositions to vary by
only ±0.01 mole fraction. Also, we assume that the inputs vary by ±0.5 kmol/minute. The scaled
process gain matrix is then

The M ATLAB SVD analysis of this system is shown below.

MATLAB SVD Analysis

» g = [0.878 -0.864;1.082 -1.096];

» gs = [100 0;0 100]*g*[0.5 0;0 0.5]

gs =
43.9000 -43.2000
54.1000 -54.8000

» [u,s,v] = svd(gs)

u =
0.6246 -0.7809
0.7809 0.6246

s =
98.6043 0
 0 0.6957

v =
0.7066 -0.7077
-0.7077 -0.7066

» cond(gs)

ans = 141.7320

The SVD of the scaled gain matrix is

The condition number is

which indicates that this is an ill-conditioned system. The first column of the left singular vector
matrix indicates that the most sensitive output direction is a simultaneous change in the distillate
and bottoms composition in the same direction. The first column of the right singular vector
matrix indicates that the strongest input direction is to change reflux and vapor boil-up by the
same magnitude, but in different directions (increase reflux and decrease vapor boil-up, or vice
versa). Physically, this is because these types of changes have a greater effect on the overall
material balance around the column.

We can view these SVD results to understand both open-loop and closed-loop effects.

Open-loop. An input in the most sensitive direction will have a large-magnitude effect on the
outputs. This means that an input in the most sensitive direction has a high "gain" effect on the
output; that is, a small input change causes a large output change. Similarly, an input in the least
sensitive direction (column 2 of the V matrix) has very little effect on the output. These effects
are illustrated by the open-loop responses in Figures M13-4 and M13-5. Figure M13-4 is a forcing
in the most sensitive input direction, while Figure M13-5 is a forcing in the least sensitive
direction.

Closed-loop. A desired output change in the most sensitive output direction will require an input in
the most sensitive input direction, and these are associated with the largest singular value. This
means that a setpoint change in the most sensitive output direction will not require as large a
magnitude input change as a setpoint change in the least sensitive output direction. It should be
noted that the response shown in Figure M13-7 is based a setpoint change in the most sensitive
output direction, which requires little manipulated variable action. The reader should show that a
setpoint change in the weakest direction will yield a slow response, similar to Figure M13-6.

[ Team LiB ]
[ Team LiB ]

M13.7 Nonlinear Effects

The results presented in previous sections are purely linear phenomena. In addition, distillation
columns have important nonlinear effects. For example, even an infinite reflux rate (requiring an
infinite vapor boil-up rate) will yield a mole fraction of the light component in the distillate that is
less than 1.0. If the steady-state value of the distillate composition is 0.99, it can increase by less
than 0.01, yet it can decrease to almost as low as the feed composition of 0.5. The range for
increasing purity is then much smaller than the range for decreasing purity.

Figures M13-8 and M13-9 show small open-loop changes in the manipulated inputs. The reader
should perform additional small open-loop changes to verify that the outputs do not change much
when the purity is increasing but can change tremendously when the purity is decreasing
(Additional Exercise 5).

Figure M13-9. Responses to a increase in reflux (2.706–2.806) and

vapor (3.206–3.306) rates.

Because of nonlinear effects, a controller designed at setpoint may not operate well when the
setpoint is significantly changed. The reader is encouraged to perform SISO control simulations
(Additional Exercise 6) and multiple loop simulations (Additional Exercise 7).

[ Team LiB ]
[ Team LiB ]

M13.8 Other Issues in Distillation Column Control

Distillation is probably the most widely studied multiloop control problem, with many papers and
books published on the topic. In this chapter, we have focused on "conventional control," where
reflux flow rate is manipulated to control distillate composition, and vapor boil-up (related to
reboiler duty) is manipulated to control bottoms composition. The pressure and level loops are
assumed to be perfectly controlled. An alternative control structure, often used on high-purity
columns, is known as material balance control. In material balance control, distillate flow rate is
manipulated to control distillate composition. (See Additional Exercise 8 for a study using material
balance control.) In practice, for this strategy to work, the distillate level must be very tightly
controlled by manipulation of the reflux flow rate.

In this chapter we have assumed that the pressure and level loops are perfectly controlled. In
Figures M13-1 and M13-2 we used a simplified representation for the pressure control loop, where
the pressure was controlled by manipulating the condenser cooling water flow rate. In practice,
there are many different ways to control column pressure. (See the references section for articles
discussing various pressure control methods.)

The multivariable interactions, particularly in high-purity distillation, make it very difficult to

control the distillate and bottoms compositions simultaneously. More often, the important product
is under feedback control, while the less important product is manually controlled. Alternatively,
the controller for the most important product could be tightly tuned, while the other composition
loop is loosely tuned.

[ Team LiB ]
[ Team LiB ]

M13.9 Summary
In this chapter, we have focused on "conventional control" of distillation columns, where reflux is
manipulated to control distillate composition and reboiler vapor boil-up rate is manipulated to
control bottoms composition. The RGA indicates that this is the only satisfactory pairing for this
strategy. We have found that this process is very sensitive to the multivariable direction of the
manipulated inputs (open loop), or the setpoints (closed loop). The SVD is a useful mathematical
tool for finding the strongest and weakest directions for a multivariable process.

[ Team LiB ]
[ Team LiB ]

References
The distillation model is presented in Bequette (1998), based on a column studied by Morari and
Zafiriou (1989) and Skogestad and Postlethwaite (1996).

Bequette, B. W., Process Dynamics. Modeling, Analysis and Simulation, Prentice Hall, Upper
Saddle River, NJ (1998).

Morari, M., and E. Zafiriou, Robust Process Control, Prentice Hall, Englewood Cliffs, NJ (1989).

Skogestad, S., and I. Postlethwaite, Multivariable Feedback Design, Wiley, New York (1996).

Sigurd Skogestad has published a nice series of articles that discuss the dynamics and control of
distillation columns. The following review paper can be found at his web site
(www.kjemi.unit.no/~skoge): Skogestad, S., "Dynamics and Control of Distillation Columns. A
Tutorial Introduction," paper presented at Distillation and Absorption '97, Maastricht, Netherlands
(8–10 September, 1997).

Other papers of interest include the followng:

Skogestad, S., and M. Morari, "Understanding the Dynamic Behavior of Distillation Columns," Ind.
Eng. Chem. Res., 27(10), 1848–1862 (1988).

Skogestad, S., and M. Morari, "Control Configuration Selection for Distillation Columns," AIChE J.,
33(10), 1620–1635 (1987).

For a discussion of the many practical issues in distillation control, see the following:

Shinskey, F. G. Distillation Control, 2nd ed., McGraw-Hill, New York (1984).

There are many ways to regulate the pressure in a distillation column. Two nice articles are as
follows:

Chin, T. G., "Guide to Distillation Pressure Control Methods," Hydrocarbon Processing, 86(10),
145–153 (1979).

Sloley, A. W., "Effectively Control Column Pressure," Chem. Eng. Prog., 97(1), 38–48 (2001).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Generate responses to step changes of ±0.01 kmol/minute in the vapor boil-up rate (V)
at t = 10 minutes. Compare and contrast these responses with those of the reflux
changes shown in Figure M13-3. Are these changes consistent with the transfer function
matrix?

2: Consider the SISO controllers. The reader should verify through simulation that the
independent controllers with IMC filter factors of 25 minutes lead to closed-loop time
constants of roughly 25 minutes.

a. First, close the xD-L loop and make a setpoint change of 0.005 mole fraction.

b. Next, close the xB-V loop (with xD-L open) and make a setpoint change of -0.005
mole fraction. In each case, show the change of both compositions.

3: Consider the case where both SISO controllers are closed simultaneously. Tune the
controllers tighter for the setpoint change in the "slow" direction (Figure M13-8).
Decrease l 1 and l2 (increase the magnitude of the proportional gains) and observe the
behavior.

4: Using the tuning parameters found in Additional Exercise 3, study the response to a
setpoint change if one of the loops is "opened." You should find that the other loop goes
unstable or has oscillatory performance.

5: Now perform open-loop step changes on the nonlinear model. Show that the purity
never changes much when it is increasing but can exhibit large changes when
decreasing.

6: Perform closed-loop SIS0 studies on the nonlinear system. For small setpoint changes
(say, from 0.990 to 0.991 on the distillate composition), the results are similar to the
linear case. For large setpoint changes (say, from 0.990 to 0.999), the differences can
become more substantial.

7: Perform closed-loop studies on the nonlinear system with both loops closed. For small
setpoint changes in the most sensitive directions (say, from 0.9900 to 0.9906 on
distillate and from 0.0100 to 0.0108 on bottoms), the results are similar to the linear
case.

8: Consider a so-called material balance control strategy, where distillate flow is

manipulated to control distillate composition. Assume the same scalings for the inputs
and outputs as for the conventional control strategy. Perform an SVD analysis on the
scaled gain matrix to determine the most and least sensitive setpoint change directions.
Compare and contrast closed-loop simulations with those for the conventional control
strategy. The transfer function model for a material balance control strategy (where D is
the distillate flow rate) is
 9: The focus of the previous problems has been setpoint tracking. In this problem consider
changes of ± 0.1 kmol/min in the feed flow rate and ± 0.05 mole fraction in the feed
composition. Base your simulations on the matrix transfer function model presented in
Section M13.1. Compare and contrast closed-loop performance when only single loops
are closed vs. both loops closed.

[ Team LiB ]
[ Team LiB ]

Appendix M13.1
The SIMULINK .mdl diagram for the open-loop transfer function model is shown in Figure M13-10.

Figure M13-10. SIMULINK diagram for linear model. dist_OLtf.mdl.

The nonlinear model of the open-loop system is shown in Figure M13-11.

 Figure M13-11. SIMULINK diagram for open-loop nonlinear system.
dist_blk.mdl.

[ Team LiB ]
[ Team LiB ]

Module 14. Case Study Problems

The purpose of this module is to present some realistic multivariable process control problems
that require the integration of a number of different concepts and techniques presented in the
text. SIMULINK files for each of these processes can be downloaded from the text web page. The
various case studies are as follows:

M14.1 Background

M14.2 Reactive Ion Etcher

M14.3 Rotary Lime Kiln Temperature Control

M14.4 Fluidized Catalytic Cracking Unit

M14.5 Anaerobic Sludge Digester

M14.6 Drug Infusion System

M14.7 Suggested Case Study Schedule

M14.8 Summary

A partial solution for the drug infusion case study is presented on the textbook web page.

[ Team LiB ]
[ Team LiB ]

M14.1 Background
The development and implementation of an industrial control strategy is complex and involves a
number of steps. A goal of this module is to present a number of interesting important process
control problems that require the synthesis of concepts presented in this textbook. One use of this
module is for a special project studied during the final few weeks of a typical undergraduate
process control course. Students, often working in small groups, select one of the challenge
problems discussed in this module. Here we discuss some of the critical steps of the control
system development.

It is important to consider the control objectives for a particular process; in addition, it one should
consider previously developed strategies and understand how successful they were/are.
Sometimes this information is available from a currently operating plant in the corporation. In
other cases, a literature search can provide important background material. The next step is to
develop a process model, either from first principles, or from testing an existing process. In this
module we assume that the step testing approach is being used; since we do not have direct
access to an operating plant, SIMULINK modules are used to mimic plant behavior; these can be
downloaded from the textbook web page. Models can then be used to develop SISO control loops,
and to understand/predict possible multivariable interaction problems when all loops are closed. If
the closed-loop performance of multiple SISO loops is not satisfactory, then multivariable
techniques can be developed.

When working on these problems, use the following steps and answer the following questions:

What types of control strategies and objectives have been used on similar processes?

What types of input tests should be performed to develop models for control system design?
Do the responses of the models adequately match the actual plant responses?

Do the orders of magnitudes of the model gains, time constants, and time delays make
physical sense?

If using independent SISO control loops (MVSISO), what pairings are suggested by the RGA
analysis?

If an SISO controller is tuned based on the other loops being open, does the tuning need to
change when all the other loops are closed?

Should a decoupling (steady-state or dynamic) type of control strategy be used for real
(MIMO) multivariable control? What about multivariable IMC?

Can you think of any other measured or manipulated variables that could be used?

What are the important process disturbances?

When tuning the independent SISO loops (assuming the other loops are open), it is probably best
to use a procedure that you are familiar with, such as IMC-based PI. You will generally use a
closed-loop time constant (l, the IMC filter factor) of roughly one half of the open-loop time
constant (if this satisfies the minimum requirements for l based on the dead time).

When testing controller performance for these multivariable processes, it is probably best to
initially make a step change in one setpoint while keeping all other setpoints constant (for
example, change r 1 without changing r 2). Do this for each loop. Compare how a setpoint change
in one output variable affects the other output variables. Also consider the manipulated variable
responses.

[ Team LiB ]
[ Team LiB ]

M14.2 Reactive Ion Etcher

An important process in semiconductor device manufacturing is plasma etching (or reactive ion
etching). In this unit operation, reactive ions are used to selectively remove (etch) layers of solid
films on a wafer. A major objective of a control system is to achieve relative etch rates
(angstroms/second) by manipulating the process inputs. Currently, etch rate measurements are
not readily available, so it is assumed that by measuring and controlling other variables, good
control of the etch rate can be obtained.

Two variables important to control are the voltage bias of the plasma and the fluorine
concentration, since they ultimately determine the etch rate. Two important manipulated
variables are the RF power and the outlet valve position (throttle). A schematic of a reactive ion
etcher is shown in Figure M14-1.

Figure M14-1. Reactive ion etcher.

A reactive ion etcher has power and throttle position inputs (P and T) and voltage (Vbias) flourine
concentration (F) as outputs.

Questions to Answer

Are fundamental (first principles) models of plasma etchers easy to develop?

Will fundamental models or input/output tests (i.e., step tests) normally be used for control
system design?

The nominal and range of operating conditions for this system are as follows:

Inputs: power 1000 W, range from 0 to 2000 W; throttle, 50%, range from 0 to 100%.
 Outputs: voltage, 250 V; fluorine, 50% of range.

Disturbances of up to ±25 W in power are possible.

A SIMULINK block diagram for a reactive ion etcher is shown in Figure M14-2. The inputs and
outputs of the etcher are in physical (not deviation) variables.

Figure M14-2. SIMULINK diagram for reactive ion etcher.

[ Team LiB ]
[ Team LiB ]

M14.3 Rotary Lime Kiln Temperature Control

Lime kilns are used in the paper industry to convert lime mud to lime. A schematic control
instrumentation diagram for a lime kiln is shown in Figure M14-3. Lime mud enters the "back" or
the "cold" end, while the lime product exits the "front" or "hot" end. The kiln rotates at
approximately one revolution per minute and is inclined so that the mud flows from back to front.
The quality of the lime depends on the temperature profile along the kiln. Typical temperatures
are 2250°F at the front end and 425°F at the rear end, with nominal damper positions of 50% of
range. The inputs are fuel gas flow rate and the damper position, and the outputs are the front
and rear temperatures.

Figure M14-3. Rotary lime kiln temperature control. The manipulated

variables are fuel gas flow rate (u 1) and damper position (u 2). The
measured variables are front temperature (y1) and rear temperature
(y2).

[ Team LiB ]
[ Team LiB ]

M14.4 Fluidized Catalytic Cracking Unit

A fluidized catalytic cracking unit (FCCU) is used to produce the majority of gasoline in an oil
refinery. A simplified process schematic and instrumentation diagram is shown in Figure M14-4.
Heavy oil feedstock from the crude unit is mixed with recirculating catalyst and reacted in a riser
tube. The catalyst cracks the oil to form lighter hydrocarbons. The lighter hydrocarbon products
are separated from the catalyst in the reactor (which is actually a separator; it is called a reactor
for historical purposes). The spent catalysis (which contains a great deal of carbon) is sent to the
regenerator, where partial combustion is used to remove the carbon from the catalyst.
Regenerated catalyst (manipulated variable) is recirculated back to mix with the inlet feed oil from
the crude unit. The important measured variables are reactor (separator) temperature (T1) and
the regenerator gas temperature (Tcy). The manipulated variables are the catalyst recirculation
rate (Fs) and regenerator air rate (F a).

Figure M14-4. FCCU control strategy. Fs and Fa are manipulated. T1 and

Tcy are measured.

The steady-state input values are 294 kg/s (Fs), 25.35 kg/s (Fa) and the output values are
776.9°F (T1) and 998.1 (Tcy), the regenerator temperature (Trg) is 965.4°F. Assume that the
manipulated inputs can vary by ± 50% of the nominal steady-state values.

[ Team LiB ]
[ Team LiB ]

M14.5 Anaerobic Sludge Digester

Anaerobic sludge digesters are used to degrade compounds (in waste water that is high in
suspended solids) to carbon dioxide and water. A schematic of an anaerobic digester is shown in
Figure M14-5.

Figure M14-5. Anaerobic digester. The level and pressure controllers

are assumed to be perfect.

The time unit is days, and the inputs and outputs have been scaled to cover the same ranges.
The outputs and inputs are y1 = substrate concentration, y2 = temperature, u1 = inlet flow rate,
and u 2 = heat addition rate (°C/day). The steady-state inlet flow rate is 300 m3/day, the heat
addition rate is 10°C /day. The steady-state temperature is 55°C and substrate concentration is
300 mg/liter.

Besides considering setpoint responses, assume that disturbances in the inlet flow rate occur.

[ Team LiB ]
[ Team LiB ]

M14.6 Drug Infusion System

Critical care patients have often suffered a "disturbance" to the normal operation of their
physiological system; this disturbance could have been generated by surgery or some sort of
trauma (e.g., a heart attack). A responsibility of the critical care physician is to maintain certain
patient outputs within an acceptable operating range. Two important outputs to be maintained
are mean arterial pressure (MAP) and cardiac output (CO). Often the anesthesiologist will infuse
several drugs into the patient in order to control these states close to the desired values. A
conceptual diagram is shown in Figure M14-6.

Figure M14-6. Drug infusion control.

The goal of this control system design is to manipulate the flow rate of two drugs, dopamine
(DPM) and sodium nitroprusside (SNP), to maintain the two outputs (MAP and CO) at their
desired setpoints. A successful implementation of such a strategy allows the anesthesiologist to
spend more time monitoring other patient states, such as "depth of anesthesia."

The steady-state (initial) drug infusion rates are both 0, while the steady-state (initial) mean
arterial pressure is 120 mmHg and cardiac output is 130 ml/(kg min). A partial solution to this
problem is presented on the textbook web page. The maximum flow rates for both drugs is 10
mg/(kg min).

[ Team LiB ]
[ Team LiB ]

M14.7 Suggested Case Study Schedule

It is suggested that these case studies be assigned during the last 4–5 weeks of a typical one-
semester course. The suggested time frame is realistic enough to allow for the study of other
modules and for short homework assignments during this period.

1. Literature review and motivation for problem 1 week

2. Model development (step tests) 2 weeks

3. SISO controller design 3 weeks

4. MVSISO controller design 4 weeks
5. Oral presentation 4–5 weeks

6. Written report 4–5 weeks

Each problem (unit operation) will have an "advisor" who will answer questions that you have
regarding the problem and possible solutions techniques. After you turn in part 1, you will be
assigned an advisor. You should probably plan to meet with the advisor once per week, during
office hours, during the project period.

An example partial solution to a biomedical engineering control problem (drug infusion control) is
presented on the web page.

1. Literature review and motivation for problem (due after 1 week)

Write a short memo (two pages or fewer), giving the following information:

The example that you have decided to study.

The importance of the unit operation to the particular industry. What are typical sizes
or production rates?

A literature search with references related to control of the unit operation. A web-
based search (using Google or some other search engine) is unacceptable. We are
interested in sources published in books, journals, magazines, or conference
proceedings. Appropriate databases include Current Contents and the EI Compendex.

The SIMULINK files for each problem are available on the textbook web page. You will apply
step inputs to identify the process transfer functions for the next phase of the project.

2. Model development (due after 2 weeks)

Write a short memo (one page or under) summarizing your modeling results. Attach the
following information to the memo:

The transfer function matrix relating each input to each output

Comparison of open-loop step tests with model step responses (superimpose the model

3.
 and plant responses)

3. SISO controller design (due after 3 weeks)

Write a short report summarizing the control system design for each independent SISO
controller. Please include the following:

A description of the design procedure (IMC-based PID, Ziegler-Nichols, direct synthesis,

etc.) and the tuning parameters used.

Closed-loop responses for each independent loop. It is appropriate to demonstrate the

effect of different values for the tuning parameters.

How does a setpoint change in a controlled loop affect the uncontrolled loop?

4. MVSISO controller design (due after 4 weeks)

The RGA analysis: Based on the RGA, suggest how the variables should be paired in a
MVSISO control structure (i.e., y1 paired with u1 and y2 paired with u2).

The suggested PI or PID tuning parameters that you would use if the loops did not
interact: For example, what tuning parameters would you recommend for loop 1 if loop
2 was open; you may use IMC-based PI or PID, Cohen-Coon, Ziegler-Nichols, Direct
Synthesis, or frequency-response-based tuning (satisfying certain gain and phase
margins).

A couple of plots: One for the setpoint response of loop 1, with loop 2 open; another
for the setpoint response of loop 2, with loop 1 open.

Do you suspect that the control loops must be detuned if all loops are closed? Use the
RGA to assist you with this.

Are there dynamic reasons that you might want not pair outputs and inputs based on
the RGA?

What are the practical limits to the magnitude of setpoint changes?

5. Oral presentation (performed after 4–5 weeks)

Prepare a short (fewer than 15 minutes) oral presentation with overhead transparencies.
Each group member should make an equal contribution. The problem advisor and at least
one other consultant will observe the presentation and ask questions.

6. Final Report (due after 4–5 weeks)

Write a short final report including the following:

a. A title page, giving a list of the group members. Include a one paragraph abstract that
summarizes the results presented in the report.

b. A short background section describing the problem. Include literature references.

c. Summarize the model developed in the preliminary reports.

d. A short section with the preliminary pairing selection based on RGA analysis.

e.
 c.

d.

e. A single-loop results section giving the setpoint responses of each control loop
separately (i.e., setpoint response for loop 1 assuming loop 2 is open).

f. A MVSISO results section for the setpoint responses of each control loop. In this, you
will show the results of a setpoint change in loop 1 with loop 2 closed. Did you need to
detune either loop?

g. Disturbance rejection results. If there are no disturbance transfer functions, assume

disturbances to the inputs.

h. A recommendations section. Do you feel that MVSISO control is sufficient? Do you feel
that a multivariable technique such as decoupling should be used? Do you feel that the
system needs to be studied further using more advanced techniques?

i. The previous reports (with the advisor comments) placed in the appendix of the report.

[ Team LiB ]
[ Team LiB ]

M14.8 Summary
We have presented module of case study example problems. The simulation files can be
downloaded from the textbook web page, allowing a complete, realistic control system study. This
includes model development, SISO controller design, MVSISO design (including retuning of
controllers and consideration of failure tolerance). Additional case study problems will be added to
the web page.

[ Team LiB ]
[ Team LiB ]

Additional Exercises
For each of the case studies, the following additional studies can be performed, depending on the
interests of the students and instructor:

Implement antireset windup techniques (Chapter 11)

Implement digital control (Module 16)

Provide SVD and operability analysis (Chapter 14)

Implement decoupling (Chapter 14)

Implement multivariable IMC (Chapter 14)

Implement multivariable MPC using the MPC Toolbox

[ Team LiB ]
[ Team LiB ]

Module 15. Flow Control

The objective of this module is to discuss some of the practical issues in flow control. After
studying this module, the reader should be able to do the following:

Determine the process gains for typical pieces of instrumentation

Understand how to determine the proportional band of a controller

Discuss the flow characteristics of equal percentage, quick-opening, and linear control valves
and understand when each should be used

Realize that control valves should be specified as fail-open or fail-closed

The sections of this module are as follows:

M15.1 Motivating Example

M15.2 Flowmeters

M15.3 Control Valves

M15.4 Pumping and Piping Systems

M15.5 Summary

[ Team LiB ]
[ Team LiB ]

M15.1 Motivating Example

Consider the control schematic diagram for the flow control loop shown in Figure M15-1. In
practice, this loop contains a number of components, as detailed in Figure M15-2.

Figure M15-1. Schematic instrumentation diagram for a flow

controller.

Figure M15-2. More detailed instrumentation diagram for a flow

controller.

We can see that this loop consists of the following components.

Flowmeter— often an orifice plate meter. The volumetric flow rate is proportional to the
square root of the pressure drop across the orifice plate

DP (differential pressure) cell— measures the pressure drop across the orifice plate
 P/I (pressure to current) transducer— converts the pressure signal to a 4- to 20-mA current
signal

Controller— the input to the controller is a 4- to 20-mA current signal. The output of the
controller is a 4- to 20-mA current signal

I/P (current to pressure) transducer— converts the 4- to 20-mA signal to a 3- to 15-psig

pressure signal

Control valve actuator— translates the 3- to 15-psig pressure signal to a valve position

Notice that this control loop is based on having an electronic analog controller. In practice, there
are still a number of pneumatic analog controllers, where the input and output signals are
pressure signals (typically 3–15 psig). Also, more of the "field equipment" (flow transmitters, etc.)
are microprocessor-based, so often the signals to and from the controller are digital in nature.

Now we further analyze each component of instrumentation in terms of its transfer function.

Flowmeter
We can think of the input to the flowmeter as the actual flow rate of the process fluid. The output
of the flowmeter is the pressure drop across the orifice plate. Here, we determine the gain of the
flowmeter by simply considering the minimum and maximum flows and how they relate to the
pressure drop

Consider a flowmeter where the flow rate varies from 0 to 5 gpm, while the pressure drop varies
over a range of 0–2 psig. The flowmeter gain is then

The dynamics of flowmeters are so fast compared with the process dynamics that we can
normally neglect the time constant and assume that the transfer function is represented as a
static gain. Also, in practice the gain is not constant because the pressure drop is a nonlinear
function of the flow rate. This is discussed in depth in Section M15.2.

DP Cell
The DP cell measures the pressure drop across the orifice plate. We can think of the gain as being
part of the flowmeter gain. The dynamics of differential pressure cells are so fast compared with
the process dynamics that we can normally neglect the time constant and assume that the
transfer function is represented as a static gain.

Pressure to Current Transducer

The input to the P/I transducer is the pressure drop across the orifice plate. Assume that input to
the P/I transducer is scaled so that 0–2 psig is 0–100% of the input range. The range of the
output is 4–20 mA, so the gain for the P/I transducer is

Again, the dynamics are so fast that they are normally neglected.

Controller—Concept of Proportional Band

The input to the controller ranges from 4 to 20 mA and the output of the controller ranges from 4
to 20 mA. This might lead us to believe that the controller gain is one. We must recognize,
however, that the controller gain, kc, is a tuning parameter. Remember that the proportional gain
does not act on the actual input to the controller; it acts on the error [the difference between
setpoint and the process measured variable (input to the controller)].

Often a controller uses proportional band (PB) as a tuning parameter, rather than proportional
gain. The PB is defined as the range of error that causes the controller output to change by the
full range (100%). We see that relationship can be written

Equation M15.1

For example, if a change in error (De) of 10% causes a 50% change in controller output (Du),
then the PB is 20.

We recognize that the controller gain we are familiar with can be written

Equation M15.2

Notice also that Equation (M15.1) can be written

Equation M15.3
Comparing Equations (M15.3) and (M15.2), we see that

which leads to the following definition of PB:

Equation M15.4

We have always used a simplified representation for a controller for analysis purposes. In
practice, a controller performs the computations shown in Figure M15-3.

Figure M15-3. Schematic of computations and scaling performed by a

controller.

We see that the controller transfer function operates on signals that are scaled from 0 to 100% of
range. The following scaling factors are used:

Sei converts electronic signal to physical units (e.g., gpm/mA)

Si converts engineering units to percentage of range of measured variables

So converts percentage of range of controller output to control signal

Current-to-Pressure Transducer
The input to the I/P transducer is a 4- to 20-mA signal and the output is a 3- to 15-psig signal.
The I/P gain is then
Control Valve

Control valves will be covered in more depth in Section M15.3. Consider here the two cases: a
fail-closed valve and a fail-open valve.

Fail-Closed Valve

If the pressure signal to the valve changed, this valve would close, hence the term fail-closed.
These valves are also called air-to-open. Consider a case where flow rate through the valve is 5
gpm when it is fully open (15-psig signal). Assuming that the valve is fully closed at 3 psig (or
less), the valve gain is

Fail-Open Valve

If the pressure signal to the valve changed, this valve would open, hence the term fail-open.
These valves are also called air-to-close. Consider a case where the flow rate through the valve is
5 gpm when it is fully open (3-psig signal). Assuming that the valve is fully closed at 15 psig (or
more), the valve gain is

One of the most important things that a process engineer can do is to specify whether a valve is
air-to-open or air-to-close. For example, if the control valve was manipulating the cooling water
flow to a nuclear reactor, then a fail-open valve should be specified.

Valve dynamics can become significant for larger valves. A laboratory valve may have a time
constant of perhaps 1–2 seconds, a valve on a 3-inch process pipe may have a time constant of
3–10 seconds, and a valve on the Alaska pipeline may have a time constant of minutes.

[ Team LiB ]
[ Team LiB ]

M15.2 Flowmeters
An orifice plate meter is the most common flow measurement device. A constricting orifice is
placed between two flanges, as shown in Figure M15-4. As the fluid "speeds up" through the
orifice, the pressure drops. The volumetric flow rate is related to the pressure drop across the
orifice. As shown in the motivating example, a DP cell is typically used to measure the pressure
drop; here, we use a manometer for illustrative purposes.

Figure M15-4. Orifice meter.

The equation for flow through an orifice is (McCabe and Smith, 1976)

Equation M15.5

where vo is the average velocity through the orifice;

b is the ratio of orifice to pipe diameters, Do/Dp;

gc is Newton's law gravitational constant, 32.174 ft lb m/lb fs2; r is the fluid density;

Co is the orifice coefficient (typically 0.61); and

P1,P2 are the upstream and downstream pressures.

The volumetric flow rate, F, is

Equation M15.6
so the volumetric flow rate [combining Equations (M15.5) and (M15.6)] is

Equation M15.7

which can be written

Equation M15.8

where

Equation M15.9

Notice that C1 will be a constant for a given system. Notice also that an orifice plate has a
nonlinear input-output relationship. Consider the input to be the volumetric flow rate. The output
is the differential pressure. After rearranging Equation (M15.8) we find Equation (M15.10) and
notice that the pressure drop is a function of the square of the flow rate,

Equation M15.10

This means that the orifice plate gain (pressure drop change/flow rate change) increases with
increasing flow. The orifice plate gain is

Equation M15.11

and we see that if the flow rate doubles, then the orifice plate gain doubles. If a controller is tuned
to obtain good control when the flow rate is low, there is a chance that the control loop will
become unstable when the flow rate is high, because of the increase in overall control loop gain,
which is due to the increasing orifice plate gain. A common way of handling this varying gain
problem is to use a square root extractor.

Square Root Extractor

Assume that the input to a square root extractor is the pressure drop signal. The output is a 4- to
20-mA signal that is proportional to the square root of the pressure drop. Let Pop represent the
signal entering the square root extractor and I represent the signal out of the square root
extractor. The input-output relationship is then

Equation M15.12

where C2 is a constant. The gain of the square root extractor is

Equation M15.13

Notice that the square root extractor gain is inversely proportional to the square root of the
pressure drop. The purpose of a square root extractor can be seen by combining Equations
(M15.12) and (M15.10) to yield

Equation M15.14

The gain between flow rate and the signal out of the square root extractor is

where . That is, we have a signal that is now directly proportional to the flow rate.

[ Team LiB ]
[ Team LiB ]

M15.3 Control Valves

The relationship for flow through a valve is

Equation M15.15

where F is the volumetric flow rate, Cv is the valve coefficient, x is the fraction of valve opening,
DPv is the pressure drop across the valve, s.g. is the specific gravity of the fluid, and f(x) is the
flow characteristic.

Three common valve characteristics are linear, equal-percentage, and quick-opening. For a linear
valve,

Equation M15.16

For an equal-percentage valve,

Equation M15.17

For a quick-opening valve

Equation M15.18

The three characteristics are compared in Figure M15-5.

Figure M15-5. Flow characteristics of control valves. a = 50 for equal-

percentage valve.
At this point it is reasonable to ask why we would ever want a valve characteristic that is not
linear. After all, it seems reasonable that we would want the increase in flow to be linearly related
to the increase in valve opening.

Think about the orifice plate nonlinearity. If a square root extractor is not used, then the orifice
plate gain is low at low flow rates and high at high flow rates [see Equation (M15.11)]. Notice that
the quick-opening valve has a high gain at low flow rates and a low gain at high flow rates,
exactly the opposite of the orifice plate. Since the flowmeter and control valve are both elements
in a control loop, the overall gain of the loop may be relatively unchanged by combining the
nonlinearity of the direct-acting valve with the nonlinearity of the orifice plate flowmeter. This is
probably less important in this day and age because square root extractors or computer
computation of the actual flow rate can compensate for the orifice plate nonlinearity. A quick-
opening valve is then no longer necessary. We discuss this because a number of older plants still
have old analog equipment, which may not have nonlinear compensation (e.g., square root
extraction) for the orifice plate. Perhaps a process was originally designed with the direct-acting
valve and an orifice plate flowmeter with square root extraction. If a recent "retrofit" (installation
of new equipment) to change the control instrumentation involved adding nonlinear compensation
(square root extraction, etc.) but did involve the replacement of the control valve, then the
overall loop gain could change tremendously as a function of the flow rate. These thoughts may
help you when you are troubleshooting a control problem.

Notice that the curves in Figure M15-5 are based on a constant pressure drop across the valve.
This may be reasonable, for example, if the pressure drop through a piping system was negligible
and all the pressure drop was due to the valve. In this case, a linear valve is desirable because
there is a linear relationship between the valve position and the actual flow rate.

There are many systems, however, where the pressure drop due to piping is a significant portion
of the pressure drop through the system and the pressure drop through the valve is not constant,
and thus it may be desirable to use a nonlinear valve. This is shown more clearly in the next
section.

[ Team LiB ]
[ Team LiB ]

M15.4 Pumping and Piping Systems

Consider the pumping and piping system shown in Figure M15-6. The pump is taking suction on
the surge tank and pumping a liquid through a chemical process system (perhaps a heat
exchanger and a reactor).

Figure M15-6. Process flowsheet for a pumping and piping system.

A typical pump head curve is shown in Figure M15-7. The pump head is a direct measure of the
pressure increase between the suction and discharge of the pump that can be obtained as a
function of the flow rate through the pump.

Figure M15-7. Pump head curve for a typical centrifugal pump [see
Figure 6-7 from Perry and Green (1984)].
where hp is the pump head in feet. The process will operate at the intersection of the pump head
and system head curves, as shown in the following discussion.

Assume that the pressures Pa and P4 are known. Pa is the ambient (or atmospheric) pressure and
P4 is the pressure at some known point downstream. If the height of liquid in the tank is known,
then P0 can be calculated for any given flow rate as

Equation M15.19

where h is the height of fluid between the tank level and the suction side of the pump and DPs p is
the pressure loss due to friction in the suction piping. The pressure drop is generally proportional
to the square of the volumetric flow rate.

The pump discharge pressure, P1, can be determined either using P0 and the pump head curve as
a function of flow rate,

Equation M15.20

or using the pressure drop through the discharge piping system and the final pressure, P4.

Using the second method, we can basically determine the system head as a combination of the
pressure drop through the chemical process (DPcp) and across the orifice plate (DPop).

Equation M15.21
The pump discharge pressure can be found as

Equation M15.22

Obviously, the problem solution is where P1 from Equation (M15.20) is equal to P1 from Equation
(M15.22).

Numerical Example

Consider the pump head curve shown in Figure M15-7. Assume that the maximum flow rate is
240 gpm. The pump head at this flow rate is 76 feet of water. Now, let's make some assumptions
about the other pressures in the system. First, we know that P1 calculated from Equation
(M15.20) is equal to P1 from Equation (M15.22), or

Equation M15.23

For this particular example, assume that P4 = P0, so

Equation M15.24

At the maximum flow rate of 240 gpm, assume that the system head is

so the valve pressure drop must be

We can easily determine the system head at any other flow rate (F 2) as

Equation M15.25

That is, .

The system head is compared with the pump head curve in Figure M15-8. Notice that in practice,
the pressure drop can be predicted by knowing the pipe diameter, the number of elbows and
fitting, the pipe roughness, and so forth, and by calculating an overall friction factor. The pressure
drop is a function of the square of the flow rate, so we are assuming that the pressure drop has
been calculated or measured at a particular flow rate and we simply use the ratio of the square of
the flow rates to determine the pressure drop at any other flow rate. The difference between the
pump head and the system head is the pressure drop across the valve.

Figure M15-8. Pump head and system head curves.

The relationship for flow through a valve is

Equation M15.26

Now assuming that the valve is wide open at the maximum flow rate of 240 gpm (x = f(x) = 1)
and the fluid is water, we can calculate the valve coefficient as

Notice that the unit of pressure drop that we are using is feet of water.

Then, for a given flow rate and valve pressure drop, we can calculate the fraction that the valve
must be open. That is,

And for a linear valve, the fraction that the valve is open is
and for an equal-percentage valve, the fraction that the value is open is

In each case, DPv = DPp - DPsys. These relationships are shown in Figure M15-9.

Figure M15-9. Comparison of installed characteristics of control

valves.

Notice that the linear valve has a more linear characteristic at low flow rates, while the equal-
percentage valve has a more linear characteristic over a wider range of flow rates, and in the
higher flow rates in particular. Notice also that the gain of the linear valve is high at low valve
opening and low at high valve openings. The equal-percentage valve has a gain that is roughly
the same at high and low valve openings and has a maximum gain that is roughly twice the
minimum gain. The linear valve has a maximum gain that is roughly 10 times the minimum gain.
If a control system using the linear valve was tuned at a high flow rate (valve gain low, controller
gain high), the control system might go unstable when operated at low flow rates. This problem is
less likely to occur with the equal-percentage valve.

[ Team LiB ]
[ Team LiB ]

M15.5 Summary
We have seen that a direct-acting valve might be used to compensate for the orifice plate
nonlinearity if no square root extractor is used. We also noted that a linear valve should be used if
the pressure drop across the valve is constant (say, the upstream and downstream headers are
maintained at constant pressure, and there is negligible pressure drop through the piping). We
notice that an equal-percentage valve should be used if a substantial amount of the pressure drop
at high flow rates is due to piping, while most of the pressure drop at low flow rates is across the
valve.

Many industrial controllers use proportional band rather than proportional gain. You can still use
any of the techniques presented in this text to find the proportional gain, and simply convert to
PB using PB = 100/kc.

[ Team LiB ]
[ Team LiB ]

References

McCabe, W. L., and J. C. Smith, Unit Operations of Chemical Engineering, 3rd ed., McGraw-Hill,
New York (1976).

Perry, R. H., and D. Green (ed.), Perry's Chemical Engineering Handbook, 6th ed., pp. 6–8, New
York: McGraw Hill (1984).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Consider the control valve diagram presented in Figure M15-5 and Equations
(M15.16)–(M15.18). Find the gain between the fraction of valve opening, x, and the
output of the orifice plate, DP, as a function of flow rate for each of the valves shown in
Figure M15-5. Assume that the sum of the pressure drop across the valve and the
orifice plate is constant. Assume that the constant pressure drop is 30 psig, and that at
maximum flow, three quarters of the pressure drop is across the valve and one quarter
is across the orifice plate. The gain calculation is shown conceptually below.

2: Instrumentation Search. Select one of the following measurement devices and use
Internet resources to learn more about it. Determine what types of signals are input to
or output from the device. For flow meters, what range of flow rates can be handled by
a particular flowmeter model?

a. Vortex-shedding flowmeters

b. Orifice-plate flowmeters

c. Mass flowmeters

d. Thermocouple-based temperature measurements

e. Differential pressure (delta P) measurements

f. Control valves

g. pH

[ Team LiB ]
[ Team LiB ]

Module 16. Digital Control

This module provides a tutorial overview of digital control system analysis and design. After
studying this module, the reader should understand the following:

Discrete and continuous controller performance is similar for short sample times

The discrete equivalent to a continuous-system RHP zero is a zero outside the unit circle

A discrete control system is stable if the poles of the discrete closed-loop characteristic
equation are inside the unit circle

The discrete IMC procedure factors zeros outside the unit circle and negative zeros inside
the unit circle into the "bad stuff"

A continuous process transfer function with no RHP zeros may have an equivalent discrete
transfer function with a zero outside the unit circle, if the relative order (difference in
denominator and numerator orders) is 3 or higher.

The major sections of this module are as follows:

M16.1 Background

M16.2 PID Controllers

M16.3 Stability Analysis for Digital Control Systems

M16.4 Performance of Digital Control Systems

M16.5 Discrete IMC

M16.6 Summary

[ Team LiB ]
[ Team LiB ]

M16.1 Background
The focus of this textbook has been on continuous systems although, in practice, most controllers
are implemented digitally. Fortunately, the sample time is usually small compared with the
process dynamics, so continuous systems theory is usually sufficient.

Processes are best described as continuous systems (with differential equation models), with
inputs that are applied continuously and outputs that are available continuously. With digital
controllers, however, measurements are made at discrete intervals of time and control moves are
made at discrete intervals of time. Discrete-time models were developed in Chapter 4 and the
conversion of continuous time to discrete time models (using LTI objects in MATLAB) were
presented in Module 4.

Recall that continuous systems are stable if their poles are negative [in the left-half plane (LHP)];
also, process models can only be inverted to form stable controllers if the models have zeros in
the LHP. Models with RHP zeros must be factored into "good" and "bad" stuff; only the portion
with LHP (good stuff) zeros can be inverted to form a controller. For continuous systems, then,
the focus is on the LHP vs. the RHP. Discrete-time systems are stable if their poles are "inside the
unit circle," that is, with a magnitude <1. If a discrete-time process model has a zero outside the
unit circle (magnitude >1), it cannot be inverted to form a stable controller; the zero must be
factored into the bad stuff before the good stuff is inverted to form a stable controller.

Stable continuous process models will always have stable discrete process transfer functions. That
is, if all poles of a continuous process model are negative, then the discrete process model will
have all discrete poles inside the unit circle. It is possible, however, for a continuous process with
no RHP zeros to have a discrete zero that is outside the unit circle. This is possible for continuous
transfer functions that have a relative order (difference between the denominator and numerator
polynomial orders) of 3 or higher. See Additional Exercise 4 for an example of this problem.

The purpose of this interactive module is to obtain a better understanding of the effect of
discretization on control system performance. We focus initially on PID control, then study IMC.

[ Team LiB ]
[ Team LiB ]

M16.2 PID Controllers

Consider the continuous (analog) form of an ideal PID controller,

Equation M16.1

We use a constant sample time, Dt, and the index k to represent the value of the continuous time
signal at discrete step k. For example, the error at discrete sample time k is equal to the
continuous system value at time t = tk.

Equation M16.2

We approximate the integral term

Equation M16.3

and the derivative term is approximated using backward finite differences:

Equation M16.4

Substituting Equations (M16.2)–(M16.4) into Equation (M16.1), we find

Equation M16.5
Notice that the manipulated variable action is calculated at discrete intervals. The most common
way of implementing the discrete control actions on a continuous process is to assume a zero-
order hold. A zero-order hold means that the control action is held constant between the sample
times, and that the manipulated variable changes instantly at each sample time.

Equation (M16.5) is known as the position form of the discrete PID controller equation. The
"velocity form" can be found by subtracting the position form at step k - 1 from that at step k:

Equation M16.6

One of the major advantages to the velocity form of Equation (M16.6) is that it is naturally
"antireset windup" (see Chapter 11 for a discussion of reset windup and procedures to prevent it).
If the manipulated input saturates, the position form of Equation (M16.5) has an integral term
that continues to accumulate, causing a potential windup problem.

Most of the PID control system design techniques covered earlier in this text (direct synthesis,
IMC-based PID, etc.) should yield successful tuning parameters for the discrete PID version, if the
sample time is sufficiently small. Generally, the sample time should be 0.01 to 0.1 times the
dominant time constant. In practice, most digital control loops are sampled several times per
second, which is far more frequently than is actually needed for most process systems.

Z-Transform Representation of PID Control

In the same way that Laplace transforms are used to characterize continuous-time systems, Z-
transforms are used to represent the input-output relationships for discrete-time systems.
Consider the discrete PID algorithm in Equation (M16.6). This can be written as

Equation M16.7

where

In an analogy to continuous-time systems, where we let u(s) represent the Laplace transform of
u(t), we let u(z) represent the Z-transform of u(k). It is common to refer to z-1 as the backward
shift operator:

handling e(k), etc., similarly, we can write Equation (M16.7) as

which can be written

Equation M16.8

We can also multiply the numerator and denominator polynomials by z2/z2, to write

Equation M16.9

Note that we can view this as an input-output block

A discrete closed-loop block diagram is similar to that of a continuous block diagram. The closed-
loop relationship for a setpoint change is

Equation M16.10

[ Team LiB ]
[ Team LiB ]

M16.3 Stability Analysis for Digital Control Systems

For a standard discrete feedback control system, the closed-loop system is stable if the poles of
the closed-loop transfer function are all inside the unit circle. The closed-loop transfer function is

so the roots of

must be less than 1 in magnitude.

Example M16.1: Stability of a Discrete Control System

Consider a first-order process, where the output is temperature (°C), the manipulated input is
valve position (%), and the time unit is minutes.

If discretized with a sample time of 1 minute, the discrete-time process (the M ATLAB Control
Toolbox can be used to convert continuous to discrete models—see Module 4, Section M4.3) is

With a P-only controller, we must check the roots of

and solving for the root, we find

where the magnitude of the root is less than 1 for –1<kc<20.01. Since the process gain is
positive, only positive values of the proportional gain make sense. This is result is verified in
Figure M16-1, where the discrete P-only controller is unstable, while the continuous P-controller is
stable with only a small amount of offset; kc = 20.01 for both the continuous and discrete P
controllers.

Figure M16-1. Comparison of setpoint responses for continuous and

discrete P-only controllers on the first-order process. kc = 20.01
%/°C.

Problem M16.1 (see the SIMULINK .mdl file in Appendix M16.1)

Consider the first-order process given in Example M16.1. For a discrete PI controller, with tI = 10
minutes, find the proportional gain that just makes the closed-loop system unstable. Verify your
analytical result via simulation; show plots similar to Figure M16-1.

[ Team LiB ]
[ Team LiB ]

M16.4 Performance of Digital Control Systems

The performance of a digital PID controller is rarely significantly different from that of a
continuous PID controller, as long as the sample time is relatively short compared with the
dominant process time constant and the desired closed-loop response time. A rule of thumb for
digital control systems in general is that the sample time should be less than one tenth the
dominant time constant. In practice, digital PID controllers have a sample time much shorter than
that. Many distributed control systems have a sample rate that is several times per second, so
the sample time is less than 1 second. This makes sense for flow control loops but is much shorter
than necessary for other loops. The main disadvantages to sampling too fast are that (i) more
frequent computations than necessary are being performed, (ii) numerical round-off could be a
problem if only small control moves are being made at each sample time, and (iii) there could be
increased measurement noise sensitivity.

Example M16.2: Effect of Sample Time on PI Control Performance

Consider the first-order + time-delay process

where the output is temperature (°C), the manipulated input is valve position (%), and the time
unit is minutes. In Module 6 we used this example to compare the robustness of different tuning
procedures for PID-type controllers. The IMC-based "improved PI"parameters (see Chapter 9),
with l = 7.5 minutes, yields kc = 1.67%/°C and tI = 12.5 minutes for an analog PI controller. In
Figure M16-2 we compare analog PI control with discrete PI control for two sample times (1 and 4
minutes). Clearly, the control performance degrades as the sample time increases.

Figure M16-2. Comparison of setpoint responses for different sample

times. IMC-based PI controller tuning used (l = 7.5 minutes).
Problem M16.2 (see the SIMULINK .mdl file in Appendix M16.1)

Perform simulations with different sample times and compare the performance of the digital
controller with the continuous controller while keeping the PID tuning parameters the same as
shown in the example.

a. Is there much difference in performance when a sample time of 0.25 minutes is used?

b. Does the discrete control system go unstable for any sample times greater than 4 minutes?
You may wish to increase the integration final time for this simulation.

[ Team LiB ]
[ Team LiB ]

M16.5 Discrete IMC

The design of discrete IMCs is similar to that of the continuous IMC design procedure detailed in
Chapter 8. The primary difference is that the factorization is performed on a discrete-time (Z-
domain) model.

Factorize into good stuff and bad stuff the discrete model

Equation M16.11

where contains zeros outside the unit circle and zeros inside the unit circle that negative.
The reason for this is that negative zeros inside the unit circle cause oscillatory poles in the
controller if not removed. Make certain that the numerator of is one order less than the
denominator polynomial. Also, form an "all-pass" in by including a pole at 1/zi for each zi
outside the unit circle. Make certain that "gain" (found by setting z = 1) of the "bad stuff" is equal
to one, while the gain of the "good stuff" is equal to the gain of the unfactored transfer function.

The design procedure is then analogous to that developed for continuous IMC in Chapter 8. The
controller is the inverse of the "good stuff" times a discrete filter to make the controller physically
realizable (numerator and denominator polynomials are the same order).

Equation M16.12

where f(z) has the following form

Equation M16.13

where a is a discrete tuning parameter. It is related to the continuous IMC filter factor by

Equation M16.14

where Dt is the sample time. The discrete IMC procedure is slightly more complex than the
continuous-time case for a couple of reasons. First of all, the continuous system parameters such
as time constants are much easier to interpret than the discrete-time parameters; discrete
parameters are a function of the sample time of the digital system. Also, it is easier to keep
mental track of gains in continuous-time, since setting s = 0 normally eliminates most terms in a
transfer function. In discrete-time, setting z = 1 results in the sum of a number of coefficients,
and is simply not as natural to handle.

The following example presents a step-by-step application of the discrete IMC procedure.

Example M16.3
Consider the van de Vusse reactor model studied in Module 5; this was also used as an example
for the conversion between model types in Module 4. The continuous transfer function is

The discrete-time model (for a sample time of 0.1 minutes) is

This can be written in discrete pole-zero form as

and it is easy to verify that the gain (found by setting z = 1) is 0.5851, which is consistent with
the continuous time model (found by setting s = 0).

Now, the first step of the factorization procedure is to remove the zeros that are outside the unit
circle (there is a zero at 1.334), as well as the negative zeros inside the unit circle (there are none
in this particular problem)

The next step is to form the "all-pass"

Next, make certain that the gain of the "bad stuff" is 1 and that the gain of the "good stuff" is the
same as the gain of the original transfer function

The IMC is obtained by inverting the good stuff

and using a first-order filter to make it realizable:

where the discrete tuning parameter, a, is related to the continuous tuning parameter, l, by

Here, we compare the responses of the discrete control system with the continuous system. For
this sample time and choice of l = 0.2 min, there is virtually no difference, as shown in Figure
M16-3. The block diagram and m-file used for this simulation is presented in Appendix M16.2.

Figure M16-3. Comparison of continuous and discrete IMC for l = 0.2

minutes. Discrete sample time, Dt = 0.1 minutes.
Problem M16.3

Choose Exercise 2, 3, or 4 from the Additional Exercises at the end of this module for further
study of the IMC design procedure. Exercise 2 is a straightforward first-order example where
there are no particular performance limitations. Exercise 3 illustrates the importance of factoring
out negative zeros inside the unit circle. Exercise 4 is an example where the continuous-time
transfer function is third-order, but minimum-phase (no RHP zeros or time delays) so there is no
limitation to achievable IMC performance. The discrete-time model, however, has a zero outside
the unit circle. This poses substantial performance limitations for the discrete-time system.

[ Team LiB ]
[ Team LiB ]

M16.6 Summary
The unit circle plays the same role in discrete-time systems that the real line does in continuous-
time systems. A zero outside the unit circle causes the same performance limitations that an RHP
zero causes for continuous systems. It is possible, however, for a continuous-time system with no
RHP zero to have a discrete zero that is outside the unit circle, if the continuous-time system has
a relative order of 3 or higher.

Discrete-time models with zeros outside the unit circle must be factored into the bad stuff for
discrete IMC controller design. This procedure is similar to continuous-time models, where RHP
zeros must be factored into the bad stuff. The main difference between continuous and discrete
systems is that even zeros than are inside the unit circle, but negative, must be factored out; if
they are not, the controller will often be too oscillatory for acceptable performance.

[ Team LiB ]
[ Team LiB ]

References
In this textbook we have used the Control Toolbox to convert continuous-time models to discrete
time models (assuming a zero-order hold). For more details on how to perform this conversion
analytically, see the following textbooks:

Stephanopoulos, G., Chemical Process Control, Prentice Hall, Upper Saddle River, NJ (1984).

Seborg, D. E., T. F. Edgar, and D. A. Mellichamp, Process Dynamics and Control, Wiley, New York
(1989).

The basic discrete IMC design procedure is covered (at an advanced level using different notation
than in our textbook) in the following:

Zafiriou, E., and M. Morari, "Digital Controllers for SISO Systems: A Review and a New
Algorithm," Int. J. Cont., 42(4), 855–876 (1985).

Morari, M., and E. Zafiriou, Robust Process Control, Prentice Hall, Englewood Cliffs, NJ (1989).

[ Team LiB ]
[ Team LiB ]

Additional Exercises

1: Consider the van de Vusse reactor, which has the following state space model at the
nonminimum-phase operating point:

where

Find the sample time that will bring the zero (discrete time) that is outside the unit
circle to the unit circle. Compare the discrete-step response for this sample time with
the continuous system step response.

2: Consider the first-order process (where the time unit is minutes and the gain is °C/%)

If discretized with a sample time of 1 minute, the discrete-time model is

Design continuous and discrete IMC controllers and compare performances on a unit
step setpoint change. Study various values of the IMC tuning parameter, l.

3: Consider the second-order process (where the time unit is minutes and the gain is
%/%)

If discretized with a sample time of 3 minutes, the discrete-time model is

 Consider factorizations for the discrete IMC design when (a) the negative zero inside is
not factored out, and (b) the negative zero inside the unit circle is factored out.
Compare the two digital controllers and continuous IMC performances on a unit step
setpoint change. Study various values of the IMC tuning parameter, l.

4: Consider the third-order process (where the time unit is minutes, and the gain is %/%)

which has no zeros (that is, it is minimum phase). If the model is perfect, we know that
a continuous-time controller can be tuned arbitrarily tightly.

If the continuous model is discretized with a sample time of 1 minute, the discrete-time
model (in factored form) is

Notice that one of the process zeros is at –1.7790 (outside the unit circle); if the inverse
of the model is used for controller design, an unstable controller results. It is interesting
that the continuous-time model has no RHP zero, yet the discretized model has a zero
outside the unit circle.

Consider factorizations for the discrete IMC design when (a) the zero at –1.7990 is not
factored out, and (b) the zero at –1.7990 is factored out. Compare the two digital
controllers and continuous IMC performances on a unit step setpoint change. Study
various values of the IMC tuning parameter, l.

[ Team LiB ]
[ Team LiB ]

Appendix M16.1: SIMULINK .mdl File for Example M16.2

The .mdl figure Figure M16-4 can be used to simulate Example M16.2. To use this for Example
M16.1, simply remove the "transport delay" blocks.

Figure M16-4. .mdl figure.

[ Team LiB ]
[ Team LiB ]

Appendix M16.2: SIMULINK .m and .mdl Files for Example

M16.3
Use the following m-file to run simulations based on the cd_IMC.mdl SIMULINK model, and make
plots. See Figure M16-5.

% Run continuous and discrete IMC simulations

% For use with the simulink diagram: cd_IMC.mdl
%
% B.W. Bequette
% 14 Sept 00
%
% plant values are (notice the other models are embedded in the
% simulink diagram

a = [-2.4048 0;0.8333 -2.2381];

b = [7; -1.117];
c = [0 1];
d = 0;
%
% first enter the value of lambda desired in the command window
%
alpha = exp(-0.1/lambda);
%
tfinal = 5; % final simulation time
sim('cd_IMC')
%
figure(3)
title('comparison of discrete and continuous IMC')
subplot(2,1,1),plot(t,yimc,'--',t,yd,t,r,'-.')
legend('continuous','discrete')
xlabel('time, min')
ylabel('y')
subplot(2,1,2),plot(t,uimc,'--',t,ud)
legend('continuous','discrete')
ylabel('u')
xlabel('time, min')

Figure M16-5. Comparison of discrete and continuous IMC.

[ Team LiB ]
[ Team LiB ]

About the Author

B. Wayne Bequette is a professor of chemical engineering at Rensselaer Polytechnic Institute.

His teaching and research interests are in process systems and control engineering. Applications
of interest include biomedical systems (anesthesia and diabetes), pharmaceuticals (batch reactor
scale-up and control), chromatography (protein separations), and fairly classical chemical
processes, particularly those with interesting nonlinear behavior. Dr. Bequette is an associate
editor of Automatica and the General Chair for the 2003 American Control Conference, and has
served as the AIChE Director on the Board of the American Automatic Control Council. He is the
author of Process Dynamics: Modeling, Analysis, and Simulation, also published by Prentice Hall
(1998), and has published over 90 technical papers.

Dr. Bequette's career path can be viewed as either a series of random moves or a well-planned
deterministic process. While completing a B.S. in chemical engineering at the University of
Arkansas, he worked at Arkansas Eastman (handling utility and waste treatment problems) and
Cosden Oil and Chemical (where he did process trouble shooting during the day, and played third
base on a championship softball team at night). After his undergraduate studies, he was a
process engineer at American Petrofina, where he had the chance to serve as a process operator
during two work stoppages. This sparked his interest in process automation and control, enticing
him to the University of Texas at Austin to earn a Ph.D. with a focus on multivariable control
system analysis and design. He spent a year as a visiting lecturer at the University of California at
Davis (developing a taste for extra dry cappuccinos) before becoming a professor at Rensselaer in
1988. While at Rensselaer, he has had the good fortune to serve as the advisor for ten Ph.D. and
four M.S. students, in addition to teaching chemical process dynamics and control to at least 900
undergraduate students. In recent years, he has had little time for softball and pole-vaulting,
devoting more time to teaching applied trajectory analysis to his son's T-ball team and helping his
daughter apply corrective feed-back action to an open-loop unstable transportation device
(bicycle).

[ Team LiB ]