2005 American Control Conference FrA08.

2
June 8-10, 2005. Portland, OR, USA

Optimization-Based PI/PID Control for a

Binary Distillation Column
Jiann-Shiou Yang

 multicomponent feeds, many can be approximated by

Abstract — This paper presents the PI/PID control of a binary (or pseudo binary) mixtures. However, due to the
binary distillation column via a genetic searching algorithm strong cross coupling and significant time delays inherent
(GSA). The time-domain design criterion, expressed as an in the distillation column, the simultaneous control of
integral of the squared error, is reformulated in the overhead and bottoms composition using reflux and steam
frequency-domain using the Parseval's relation and Padé flow as the control variables is still difficult. Attempting to
approximation. A genetic algorithm is then used to search
use tuning rules such as the well-known Ziegler-Nichols
over the stability region in the controller parameter space for
the best settings to minimize the design criterion. Our results rule to adjust the PI/PID controller for each individual loop
indicate that the genetic algorithm can provide better often leads to deteriorated control performance for the
solutions for the PI/PID control schemes as compared to those overall system. This is because these tuning rules do not
using the single-loop and multi-loop Ziegler-Nichols tuning take the interaction between the control loops into
methods. We found that the GSA is easier to use than consideration ([3], [9]).
traditional optimization techniques. In addition, no knowledge
of complex mathematics is required to use the GSA In this paper, instead of using a tuning rule we focus on
effectively. using a genetic algorithm to determine the optimal PI/PID
control settings for a pilot-scale binary distillation column
model. A genetic searching algorithm (GSA) is chosen to
I. INTRODUCTION
test whether such algorithms offer any practical advantages

T HE distillation column is probably one of the most

popular and important processes studied in chemical
engineering literature. Distillation is used in many
over traditional optimization algorithms. For the test
example, we select the Wood and Berry binary distillation
model [4] because it was derived from an operational
chemical processes for separating feed streams and for distillation column and is representative of many chemical
purification of final and intermediate products. It is known control processes. Determination of the optimal gain
that high-purity distillation columns are highly nonlinear settings for the PI/PID controllers with the Wood and Berry
and the composition interaction between the stages due to model serves as our optimization problem. A parameter
the counter flow of vapor and liquid is also large (e.g., [1], space method [10], [11] is used to determine the stability
[2]). Thus, the control of columns to give multiple products region in the controller's parameter plane (space), i.e., the
of constant composition is very difficult. It is perhaps one search space for the GSA. Limiting the search of the
of the most challenging problems in process control. controller parameters to such a region guarantees the
Various methods of controlling distillation columns have system stability. A standard integral of the squared error
been reported in the literature (e.g., internal model control (ISE) is used as our control criterion, which can be
method [1], ratio control [3], [4], non-interacting control analytically expressed in terms of the controller gains via
[3]-[5], µ-synthesis method [6], linear-quadratic Gaussian the Parseval’s Theorem. A second-order Padé
with loop transfer recovery (LQG/LTR) method [7], fuzzy approximation is also used to approximate the time delays
control [8]). Despite many advanced/modern control design involved in the model. Based on the criterion, a GSA is
techniques developed in recent years, PI and PID implemented and used to search over the stability region
controllers still represent the majority of the controllers for the controller gains that achieved the best performance.
used in industry. It is, therefore, interesting to develop an The simulation results are given, and they are also
appropriate approach to adjust the parameters of the PI and compared with those using the single/multi-loop Ziegler-
PID controllers for the control of distillation columns. Nichols tuning methods. We found that the GSA is easier
to use than traditional optimization techniques.
In this paper, the PI/PID control of a binary distillation
column is studied. Although most columns handle
II. THE BINARY DISTILLATION COLUMN
Final manuscript submitted February 22, 2005.
The author is with the Department of Electrical and Computer
Engineering, University of Minnesota, Duluth, MN 55812 USA (email: The binary distillation column we studied basically
[email protected]). separates a mixture of two components having different

0-7803-9098-9/05/$25.00 ©2005 AACC 3650

boiling points. Moreover, the process is greatly enhanced Note that R. K. Wood and M. W. Berry developed the
by forcing separation to occur in stages within the column. above mathematical model using a pilot scale replica of an
One way to accomplish this is by directing a liquid stream actual binary distillation column [4]. Obviously, the control
of high purity distillate, or reflux, back into the column of the distillation column transfer matrix P(s) = [Pij(s)] is
through an arrangement of sieve trays. For example, complicated by the dependent relationship that exists
consider a feed mixture consisting of component A with a between the steam flow and reflux flow. The interaction
low boiling point and component B with a higher boiling between the two loops makes controller tuning more
point. As feed enters the column, some of the mixture will difficult. To compensate for undesirable control loop
immediately vaporize. The remainder will concentrate at interactions, consider the block diagram of the decoupling
the bottom of the column where further boiling is induced scheme [pp. 463-467, 3] shown in Fig. 1, where D12(s) and
by the re-boiler. Consequently, a vapor stream rich in D21(s) are the decouplers used for the decoupling purpose,
component A is created and rises in opposition to the liquid while the feedback controllers C1(s) and C2(s) are to be
stream. The sieve tray arrangement creates “pockets” in designed to provide a satisfactory set point tracking for
which the vapor stream’s contact with the liquid stream is both channels. From Fig. 1, it is easy to see that in order to
maximized. This increases the likelihood that the liquid eliminate the effect of overhead and bottoms composition
stream will entrain the residual B component from the control actions on the bottoms and overhead composition,
vapor stream and direct it back toward the bottom of the respectively; the two decouplers should be chosen as
column. Most importantly, the sieve tray arrangement P21 ( s ) P12 ( s )
provides the mechanism needed for multi-staged boiling. D21 ( s )  ; D12 ( s )  (2)
Each sieve tray can be viewed as a separate boiler of P22 ( s ) P11 ( s )
increased efficiency as the vapor stream rises toward the Note that alternative decoupling techniques are possible
top of the column. (e.g., [2], [3]). In this paper, we use the decoupling scheme
shown in Fig. 1 before attempting to design feedback
While the benefits of reflux to the column’s efficiency are controllers. Using the two decouplers given in Eq. (2), the
certainly to be desired, the resulting dynamics greatly system equation becomes
complicates the simultaneous control of the bottoms and
overhead product compositions. Consider the case in which
it is desired to increase the overhead product purity by
ª xD ( s ) º ªT11 ( s ) 0 º ª u1 ( s ) º
« x ( s) » « 0 (3)
increasing reflux flow with steam flow held constant. ¬ B ¼ ¬ T22 ( s ) »¼ «¬u2 ( s ) »¼
Increasing reflux will certainly remove more of the residual
component B from the vapor stream, but it will also whereT11(s)=P11(s)+P12(s)D21(s),
increase the amount of the component A in the bottoms T22(s)=P22(s)+P21(s)D12(s). Note that in reality, the exact
product. It can be seen that there is a dependent cancellations of coupling can seldom be realized because of
relationship between steam flow and reflux flow that the modeling error and various uncertainties in the process.
affects both product compositions. This dependence is also However, Seborg et al. claimed that even approximate
apparent in the mathematical model of the distillation cancellations could be very beneficial in reducing control
column, which can be expressed as loop interactions while simplifying the controllers [3].
Besides that, the approximated “ideal” decouplers should
ª xD ( s ) º ª P11 ( s ) P12 ( s ) º ª R ( s ) º also work fine except for high-purity distillation columns
« x (s) » « P ( s) P ( s)» « S ( s) » (1) with large relative gains [12]. Therefore, we will use the
¬ B ¼ ¬ 21 22 ¼¬ ¼ decoupling scheme described above. Since the purpose of
this work is to study the feasibility and advantages of using
where the outputs xD(s) and xB(s) are the overhead and a genetic algorithm to tune the controller parameters, a
bottom product mole fraction of methanol, respectively, perfect decoupling is assumed so that we can compare the
while R(s) is the reflux flow rate and S(s) is the re-boiler available simulation results with some other methods
steam flow rate. The transfer functions Pij(s) = Ke – Ȝ s /IJ s+1 discussed later. After decoupling, we then close the loops
(i, j =1, 2) have the parameter values shown below: and try to design the controllers C1(s) and C2(s) to control
xD(s) and xB(s) independently. Our objective is to design
Parameters K Ȝ IJ the PI/PID controllers to minimize the integral of the
P11(s) 12.8 1.0 16.7 squared error (ISE) in both channels. That is, our tracking
P12(s) -18.9 3.0 21.0 performance is formulated as
P21(s) 6.60 7.0 10.9 f

P22(s) -19.4 3.0 14.4 min ³

{C1 ( s ), C2 ( s )}
0
eT (t )e(t ) dt (4)

3651
where the vector e(t) is defined as e(t)=[e1(t) e2(t)]T (the ª a0 0 0 0 0 0 0 0º
superscript T means the transpose) and ei(t) represents the «a
tracking error in channel i (i =1,2). « 2 a1 a0 0 0 0 0 0 »»
« a4 a3 a2 a1 a0 0 0 0»
« »
III. DESIGN CRITERION REFORMULATION
: « a6 a5 a4 a3 a2 a1 a0 0»
« a8 a7 a6 a5 a4 a3 a2 a1 »
After the system is decoupled, the ISE criterion in each « »
loop can be rewritten as «0 0 a8 a7 a6 a5 a4 a3 »
f
«0 0 0 0 a8 a7 a6 a5 »
Ji min ³ ei2 (t )dt (5) « »
{Ci ( s )}
0 ¬« 0 0 0 0 0 0 a8 a7 ¼»

From Fig. 1, we also have ei(s) = ri(s)/(1+Tii(s)Ci(s)) where

Tii(s) represents the decoupled plant transfer function in
channel i (i =1, 2). By using the Parseval’s Theorem [13], ª a0 0 0 0 0 0 0 d0 º
Eq. (5) can be expressed in the frequency-domain as «a a1 a0 0 0 0 0 d1 »»
« 2
« a4 a3 a2 a1 a0 0 0 d2 »
f
2 1 jf
« »
ISE ³ e (t )dt ei ( s )ei ( s)ds
2S j ³ jf
(6)
0
i
:1 « a6 a5 a4 a3 a2 a1 a0 d3 »
« a8 a7 a6 a5 a4 a3 a2 d4 »
and ei(-s)ei(s) can be further expressed as
« »
«0 0 a8 a7 a6 a5 a4 d5 »
ri ( s )ri ( s ) B( s) B( s)
«0 0 0 0 a8 a7 a6 d6 »
ei ( s )ei ( s )  « »
(1  Tii ( s )Ci ( s ))(1  Tii ( s )Ci ( s )) A( s ) A( s )
¬« 0 0 0 0 0 0 a8 d 7 ¼»
(7) (9)
where
where B(s) and A(s) are Hurwitz polynomials in s. It is
clear that the presence of time delays in Tii(s) prevents us d0 b02 , d1 b12  2b0b2 , d2 b22 2bb
1 3  2bb
0 4,
from obtaining the last expression in Eq. (7). That is, we d6 b62 2bb
5 7
cannot find the polynomials A(s) and B(s). To overcome
this problem, we can approximate the time delay using a d3 b32 2bb
2 4 2bb
1 5  2b0b6 , d4 b42  2b3b5  2b2b6  2bb
1 7
2nd-order Padé approximation [3], [14] given below
d7 b72 d5 b52  2b4b6  2b3b7
Ds D 2s2 By substituting the above two into -| ȍ1| / (2a8)| ȍ, we
1  obtain the ISE, which is a function of ai, bj and these
e D s | 2 12 (8)
Ds D 2s2 coefficients are again functions of the PI gains Kp and Ki,
1  (or the PID gains Kp, Ki, and Kd). Therefore, our objective
2 12
is to use a genetic algorithm to search over the set of
PI/PID gains in the controller’s parameter plane (space)
to replace all the time delays involved in Tii(s). After such
that maintains the system stability and, at the same time,
an approximation, we can then easily obtain B(s) =
7 8
minimizes the ISE criterion.
¦ bk s k and A(s) = ¦ ak s k , where the coefficients ai, bj
k 0 k 0
III. THE STABILITY REGION
are functions of the PI/PID gains. Note that the order of the Due to the popularity and wide acceptance of PID control
polynomial A(s) is 6 if a first-order Padé approximation is among process industries, we use this control for the
used, and its order becomes 10 for a third-order distillation column. Consider the system shown in Fig. 1
approximation. with the controllers C1(s)= kP1+kI1/s+kD1s and
C2(s)=kP2+kI2/s+kD2s where kPi, kIi, and kDi (i=1,2) are the
Using [15] with some derivations, the criterion Eq. (6) can proportional gain, integral gain, and derivative gain,
be further expressed as ISE = (-1/2 a8) (|ȍ1|/|ȍ|), where the respectively. To achieve the optimal tracking performance
notation |·| means the determinant and the matrices ȍ1, ȍ (i.e., Eq. (4)), we will find the optimal (kPi, kIi , kIi) (i=1,2)
are defined, respectively, as by using a genetic searching algorithm. Before performing
the search, we use the Siljak's parameter plane method to

3652
find the set of all PID gains that maintain the system that perhaps a better solution exists close by. Rather than
stability. That is, for each loop of the decoupled system relying solely on the crossover and mutation operators to
(i.e., Eq. (3)), we determine the stability region in the kP kIi reach this conclusion randomly, it is much more efficient to
kDi-space after inserting the PID controller in that loop. The strategically manipulate the lower order bit positions of the
search via our genetic algorithm will then be limited to the best elite population member chromosome. This
stability region for optimality of the error criterion. For manipulation produces a small number of children whose
details about the method to get the stability region in the search space coordinates are very close to the best solution
controller parameter space, please refer to [10,11]. obtained for the previous generation.

IV. THE GENETIC SEARCHING ALGORITHM V. RESULTS ANALYSIS

A genetic algorithm is a function optimization technique
based on the ideas of evolutionary genetics and the natural 1. Tuning Based on the GSA
selection process [16], [17]. Ideally, the algorithm creates Based on the implemented GSA described in Section IV,
new population members (children) who, with each the algorithm searched for the optimal (kP, ki, kd) in the
successive generation, are better equipped to succeed in stability region to maximize the fitness function. This is
their present environment (or, have a higher degree of applied to each loop separately. Results were obtained after
fitness). In terms of function optimization, this process performing five consecutive searches for each control loop
equates to incorporating various weighted operators to with the number of generations set at 50 and population
randomly select values for the independent variable(s) that size set at 100. For the PI control scheme (i.e., kd = 0), we
have a high probability of producing successively higher found that the GSA has no difficulty converging to the
values of the dependent variable until some global optimum same solution during separate runs of the algorithm.
is reached. The implementation of a basic genetic algorithm However, the GSA showed less repeatability for the PID
(GA) can be found in the literature. Our study focuses on control scheme, giving three possible solutions for T11 and
using a genetic approach to determine the optimal settings two possible solutions for T22. Despite the loss of
for the classical PID controller. A genetic searching repeatability, each of these PID settings yields better
algorithm is chosen to test whether such algorithms offer performance than optimally tuned PI controllers as
any practical advantages over traditional optimization expected. The results we tested can be summarized as
algorithms. Desirable properties of the GSA include: (1) follows:
No reliance on integral or derivative operators to steer its
operation. This allows the GSA to search for a function’s x PI Controller Results
global minimum or maximum without regard to the
continuity of the function, (2) Many points within the Parameters Kp Ki ISE
search space are evaluated each generation. This reduces T11 - loop 0.5524 0.07478 2.0284
the probability that the algorithm will converge to some T22 - loop -0.1651 -0.02118 4.5179
local minimum or maximum within the search space, (3)
No dependence on the selection of a starting point within x PID Controller Results
the search space to initialize the search process, and (4)
Ease of use. The GSA does not require detailed knowledge Example 1
of complex optimization techniques. Parameters Kp Ki Kd ISE
T11 - loop 0.6212 0.1569 0.4647 1.4348
In addition to the basic genetic operators, we choose to T22 - loop -0.1825 -0.04167 -0.3139 3.3318
incorporate two intuitive operators of our own. The first is Example 2
designed to eliminate an adverse side effect of elitism, that
Parameters Kp Ki Kd ISE
is, the general population members are excluded from
T11 - loop 0.5994 0.1474 0.4216 1.4389
mating with the elite population set. After some
T22 - loop -0.1826 -0.03533 -0.2728 3.3479
experimentation, we found that elitism leads to the average
Example 3
fitness levels of both populations reaching plateau fitness
values for many generations. This is minimized by Parameters Kp Ki Kd ISE
inserting a copy of the fitness elite population member in T11 - loop 0.6306 0.1373 0.3736 1.4535
place of the least fit general population member prior to T22 - loop N/A N/A N/A N/A
selection of offspring producing parents. The second
operator is designed to force the GSA to focus a small Error response plots for the best gain setting obtained for
amount of its attention on the search space region each control scheme are shown in Fig. 2 and Fig. 3.
immediately surrounding the coordinates of the best elite
population member for a particular generation. It makes 2. Comparison Study
sense that if a good solution is obtained at a particular point

3653
An accurate comparison of the GSA solutions to results performance using PID controllers tuned with the GSA, we
obtained by others is difficult for a number of reasons. are not able to guarantee that these are the best possible
There are a number of criteria that can be used to evaluate a gain settings.
controller’s performance. Gain settings that provide REFERENCES
optimal performance for the ISE criterion may not provide
the same performance for some other criterion. Differences [1] M. Morari and E. Zafiriou, Robust Process Control, Prentice Hall,
Englewood Cliffs, NJ, 1989.
in test methods also make comparison difficult. For
[2] W. L. Luyben, Process Modeling, Simulation, and Control for
example, Wood and Berry evaluated the performance of Chemical Engineers, 2nd ed., McGraw-Hill, New York, NY, 1990.
their PI controllers empirically using a different [3] D. Seborg, T. F. Edgar, and D. A. Mellichamp, Process Dynamics and
performance criterion [4]. A “rough” comparison of the Control, John Wiley & Sons, New York, NY, 1989.
GSA PI controller results to the single-loop Ziegler- [4] R. K. Wood and M. W. Berry, “Terminal composition control of a
binary distillation column,” Chemical Engineering Science, vol. 28,
Nichols method (single-loop/Z-N) and multi-loop Ziegler- pp. 1707-1717, 1973.
Nichols method (multi-loop/Z-N) [3], using the same ISE [5] W. H. Ray, Advanced Process Control, McGraw-Hill, New York, NY,
criterion, is shown in the following table. The results are 1981.
also shown in Figs. 4 and 5. These comparisons show that [6] H. E. Musch and M. Steiner, “Robust PID control for an industrial
distillation column,” IEEE Control Systems, pp. 46-55, August 1995.
our results seem better. [7] C. Zhou, J. R. Whiteley, E. A. Misawa, and K. A. M. Gasem,
“Application of enhanced LQG/LTR for distillation control,” IEEE
Tuning Method ISE (T11 – loop) ISE (T22 – loop) Control Systems, pp. 56-63, August, 1995.
[8] R. Stenz and U. Kuhn, “Automation of a batch distillation column
GSA 2.028 4.518
using fuzzy and conventional control,” IEEE Trans. on Control
Single-loop/Z-N 2.069 5.009 Systems Technology, vol. 3, no. 2, pp. 171-176, 1995.
Multi-loop/Z-N 10.839 4.653 [9] K. J. Astrom and B. Wittenmark, Adaptive Control, Addison Wesley,
Reading, MA, 1989.
[10] D. D. Siljak, “Analysis and synthesis of feedback control systems in
VI. CONCLUSION the parameter plane, part I- linear continuous systems,” IEEE Trans.
Applications Industry, vol. 83, pp. 449-458, 1964.
PI and PID control of the Wood and Berry distillation [11] D. D. Siljak, “Generalization of the parameter plane method,” IEEE
column via a genetic searching algorithm has been Trans. Automatic Control, vol. 11, pp. 63-70, 1966.
[12] T. McAvoy, Interaction Analysis Theory and Application, Instrum.
presented in this study. The 2x2 multivariable plant is first Soc. of America, Research Triangle Park, NC, 1983.
decoupled by two decouplers so that the system can be [13] A. V. Oppenheim and A. S. Willsky, Signals & Systems, 2nd ed.,
converted into two independent SISO subsystems. We use Prentice Hall, Englewood Cliffs, NJ, 1997.
a PI/PID controller to control each loop separately. The [14] G. F. Franklin, J. D. Powell, and A. Emami-Naeini, Feedback Control
of Dynamic Systems, 4th ed., Prentice Hall, NJ, 2002.
time-domain tracking error criterion (ISE) is reformulated [15] E. I. Jury and A. G. Dewey, “A general formulation of the total square
by using Parseval’s theorem, and the set of the controller integrals for continuous systems,” IEEE Trans. Automatic Control,
gains to maintain each subsystem's stability is also vol. 10, pp. 119-120, 1965.
identified in the parameter plane (space). We then search [16] D. E. Goldberg, Genetic Algorithms in Search, Optimization and
Machine Learning, Addison Wesley, Reading, MA, 1989.
for the optimal PI/PID gains in the stability region to
minimize the ISE criterion. A genetic algorithm was [17] L. Chambers, Practical Handbook of Genetic Algorithms - New
Frontiers, vol. 2, CRC Press, Boca Raton, FL, 1995.
implemented and used in the searching process. The
simulations show small tracking errors for both channels.
Our results also give a better performance when compared
- XD
with the single-loop/Z-N and multiloop/Z-N approaches.
+ U1 R
The proposed method can be easily used to control systems C1 P11
where the controller structure is fixed with several + +
r1 +
adjustable parameters to be determined. +

We are able to show that the GSA is certainly capable of D21 P21
tuning PI and PID controllers for this particular application.
Our GSA gives good results and is very easy to use. Once
stability regions and a performance criterion are D12
P12
established, the operator needs only to initialize the GSA
parameters and perform the search. No knowledge of
complex mathematics is required to use the GSA r2 + +
+ +
effectively. However, we did not find enough evidence to C2 P22
prove or disprove the hypothesis that GSAs provide + U2 S
-
superior search results when compared to other XB
optimization techniques due to the limited search space in
this example. While we are able to obtain better Fig. 1 Block diagram of the model showing the
controllers and decouplers.
3654
Fig 2 PI versus PID control for overhead loop (PI: curve
Fig. 4 Performance comparison of PI tuning for
labeled 1; PID: curve labeled 2).
overhead (GSA-1; Single-loop/Z-N: 2; Multi-loop/Z-N:
3).

Fig. 3 PI versus PID control for bottom loop (PI: curve

labeled 1; PID: curve labeled 2). Fig. 5 Performance comparison of PI tuning for bottoms
loop (GSA-1; Single-loop/Z-N: 2; Multi-loop/Z-N: 3).

3655