Energy and Buildings 42 (2010) 10841092

Contents lists available at ScienceDirect

Energy and Buildings

A software framework for model predictive control with GenOpt

University of California, Berkeley, USA Concordia University, Montreal, Canada c Public Works and Government Services, Ottawa, Canada
b

A R T I C L E I N F O

A B S T R A C T

Article history: Received 15 December 2009 Accepted 29 January 2010 Keywords: Building simulation Optimization Model predictive control Demand response

There is a growing interest in integrated control strategies for building systems with numerous responsive elements, such as solar shading devices, thermal storage and hybrid ventilation systems, both for energy efciency and for demand response. Model predictive control is a promising way of approaching this challenge. This paper presents a exible software framework for model predictive control using GenOpt, along with a modied genetic algorithm developed for use within it, and applies it to a case study of demand response by zone temperature ramping in an ofce space. Various areas for further research and development using this framework are discussed. 2010 Elsevier B.V. All rights reserved.

1. Introduction There is currently a strong demand for very low energy commercial buildings. Passive technologies, such as increased insulation and climate-appropriate orientation and window placements, are likely to account for part of the demanded efciency improvements. But active technologies, such as low energy HVAC and variable-transmissivity window systems, are also expected to play a signicant role in most low energy commercial buildings. Such low energy active systems can also play a role in demand response by shedding or shifting electrical load when necessary, which will likely become even more important with more intermittent renewable electricity production being added to the grid. Better integrated supervisory control strategies are needed to fully address the potentials and challenges presented by these very low energy systems and by demand response control objectives. One such advanced control strategy is model predictive control (MPC), where a building model is used within the supervisory control system, automatically testing at each control time-step various possible set-point congurations with simulations before choosing the best one for use in the building. Optimization algorithms determine set-points, rather than relying on rule-based approaches that specify set-points based on schedules or in response to particular conditions. There has been a modest amount of research on MPC for buildings applications, but this body of

research is growing quickly and is expected to grow further in the next few years. The underlying motive of the research presented in this paper is to facilitate future research in this area by developing a standard approach and software for model predictive control (and for non-predictive model-based control) that can be easily applied by buildings researchers to many different cases. A Java software framework based around GenOpt [1] is presented, along with an example MPC algorithm for use within it, and an example case study. A beta version of the software is freely available at [2]. It allows the use of any text-based simulation tool and simplies the setup process for simulationbased studies or real-time implementations. It also facilitates the use of optimization starting points and dynamic search-space constraints based on the results of previous time-steps and on heuristic rules, which is a promising and often overlooked technique in MPC for buildings. 2. Background 2.1. Model predictive control in buildings research Although still relatively uncommon in building engineering, MPC has seen extensive research and application in other elds, particularly in chemical, electrical and mechanical engineering (see [3,4] for seminal papers, and [5] for an example application). Within the buildings research eld, the idea of MPC for supervisory control was noted at least as early as 1988 [6], but because of its computational requirements it has not received much research attention until the past decade. It has been applied to a number of different types of systems by buildings researchers, often using standard simulation tools for the online building

* Corresponding author. E-mail addresses: [email protected], [email protected] (B. Coffey). 0378-7788/$ see front matter 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.enbuild.2010.01.022

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

1085

models. Research interest seems to be growing, and building controls problems have recently attracted MPC experts from outside of the buildings research eld. The use of MPC for energy minimization through ice storage and active building mass thermal storage has been studied by Henze et al.: a real-time implementation experiment was carried out for the control of a building with these systems, using a 24-h future horizon and a 1-h controller time-step, with TRNSYS used for the building model and Matlab used to control the optimization [7,8]. They noted some of the expected technical glitches (such as sensor malfunctions and problems with information passing), but aside from this the approach was generally successful, except for a few points in time where the optimization algorithm was stuck in local minima and actually performed worse than the default strategy would have. They have also analyzed various aspects of MPC that are peripheral to the central concept but essential in its application: they have studied the importance of forecast and model accuracy [9]; worked on the development of a method for automated model calibration to ensure continued accuracy over time [10]; and worked on the development of a hybrid control system that attempts to incorporate the aspect of continual updating (found in the learning-algorithm approach to control) with the MPC approach [11,12]. The problem of determining the optimal start time for heating has been considered by a number of researchers. One such study brought together researchers from Honeywell Controls Systems Ltd. and the University of Strathclyde, and worked out some of the interfacing concerns between a standard control system and a simulation tools (ESP-r in particular) and optimizer [13]. A modication of the optimal start problem was also considered by Kummert et al. [14,15], who looked at the optimal control of passive solar buildings with night setback, in an attempt to minimize the energy consumption while also minimizing occupant discomfort due to morning undercooling and afternoon overheating. A simplied model (a linear state-space representation) was used and the optimization was done by quadratic programming. Non-predictive model-based control has also been used by a number of buildings researchers, for example in daylighting research (controlling blinds and lights) by Mahdahvi et al. Their experimental work has focused on the control of automated shading devices [1618], using lighting simulation tools such as LUMINA, but they have also considered control for natural ventilation [19]. The purpose of using simulation in these cases is to capture the complexities of a snapshot of the system, rather than to capture the systems dynamics, so no prediction horizon was used in these studies.1 Other examples of non-predictive model-based supervisory control include studies of VAV control (using highly simplied and fast-running system models) [20,21], and chilled water plant control (using detailed TRNSYS models, but conceived not for real-time implementation but for derivation of heuristic control rules, so it did not have to contend with the computation time challenge associated with global HVAC control using a detailed simulation model) [22]. It is also worth noting the extensive existing research and applications of neural network control in building systems. An
1 Note that the use of predicted values over a future horizon is necessary for any system involving energy storage, but it may also be necessary for any system in which there is a penalty incurred for changing control congurations. For example, if the system requires x kWh to change the position of a solar shading device, and such an action would save less than x kWh over any one particular controller timestep, this action would not be selected without the consideration of further points in the future, even though one change in position might save substantial energy over the course of a day. So a prediction horizon may still be desired in solar shading applications if an energy or occupant-distraction penalty is considered for position changes.

illustrative example of how this approach can be conceived in a way that is similar to the MPC approach (with a neural-network empirical online model instead of a physics-based online model) is the EDIFICIO project carried out by a number of European research agencies and headed by a group from EPFL [23]. Clarke [24] outlined the differences between these learning-algorithm approaches and the model-based approach, and discusses the advantages and disadvantages of each. In general, learningalgorithm based controllers are easier to install and operate, and do not have the same concerns about model accuracy or computation time that MPC has. But unlike model-based controllers, learning-algorithm controllers usually require an in situ learning period before use, they do not deal well with physical or occupant-use changes in the building, and they tend not to perform well under anomalous conditions. And since they lack any building physics in their internal calculations, they cannot be used to explain why particular conditions or congurations produce particular effects, so they are of less use in building diagnostics or retrot than are the models used in MPC. But the ability to improve performance over time, as embodied by these learning-algorithm approaches, is a useful feature that could be incorporated into software and methodologies for MPC. Recent and ongoing research in MPC for buildings is being carried out by research teams involving both buildings experts and MPC experts. Recent collaborative work between UC Berkeley, Lawrence Berkeley National Lab and United Technologies Corporation has implemented MPC for a campus chilled water storage system [25], and there is an ongoing research project on MPC in buildings involving ETH Zurich, Siemens, EMPA Dubendorf and MeteoSwiss [26]. 2.2. Simulation, optimization and controls software There is a broad collection of commonly used building energy simulation software available, but most of this was created with design applications in mind and tends not to be ideal for use for the online model in MPC because of slow run times, lack of optimizer access to the underlying equations, and difculties in reinitializing the model states for a particular time-step (most require a warm-up period). Some existing programs (e.g. TRNSYS, SPARK) present fewer difculties for use in MPC than others, but they are still not ideal. The development of a Modelica library of buildings components [27] is currently underway, in part to try to address these problems. Some MPC researchers have turned to the creation of models from scratch in Matlab (which is the tool of choice for controls researchers, and the Matlab MPC Toolbox [28] is helpful) or the creation of Matlab lookup tables off-line by simulating the modeled components over their range of expected input conditions and then using these lookup tables as the model in the online optimization instead of using the rst-principles directly. This development of more controls-friendly building simulation tools is helpful. But it will require some time, and buildings researchers and practitioners who are more familiar with the standard building simulation tools may wish to use these for MPC studies in the near-term. And those for whom the software cost is prohibitive will always prefer to use free or inexpensive tools. As such, the software framework described herein was constructed to make use of any text-based simulation software (so either existing or new-generation tools can be used), and uses GenOpt [1] for the optimization. GenOpt was developed in an effort to simplify and standardize the process of optimization with building simulation. It is a Java program and is freely downloadable. It essentially acts as an interface between any text-based building simulation program and any optimization algorithm. It has a standard library of

1086

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

optimization algorithms, which includes generalized pattern search algorithms, particle-swarm algorithms and algorithms for parametric studies, and the library can be extended by the user by adding new optimization algorithms. 3. Framework The development of the software framework aims to provide a free MPC tool that researchers and practitioners can use alongside their modeling software of choice. It essentially acts as an extension of GenOpt, providing the organization necessary to use it as the optimizer in a model-based controller that is connected to a central building control system, and providing some functionality specically appropriate to MPC problems. We begin by looking at the MPC problem denition, then describe the GenOpt-based framework developed to address it. 3.1. Generic problem denition Fig. 1 provides a simple illustration of model predictive control, where the projected control sequence over the prediction horizon is determined by an online optimization with a system model. It can be used for multi-input multi-output control of nonlinear dynamic systems, and is most often applied to systems with slow dynamics, application-critical and possibly nonlinear state constraints (as in many chemical process applications), and/or complex interactions between inputs and outputs. Buildings applications rarely have critical state constraints, but many supervisory problems have complex interactions between components (e.g. between solar shading, space cooling and lighting), combined with relatively slow dynamics. The supervisory controller time-step in buildings applications could range from minutes to days. Since the objective function is evaluated by running a simulation of the online model, there is an important tradeoff that must be made between the controller time-step, the computation time required by the simulation, and the precision (and global-vslocal characteristics) of the optimizer. Each time that the controller is asked to determine new values for the control variables, it is faced with the optimization problem dened by the set of Eqs. (1)(4) (in an energy minimization case) or the set of Eqs. (5)(8) (in a demand minimization case), where g 1 is the objective function (the output of the online model), g 2 are the constraints on the state and output (which may be nonlinear),

f maps the evolution of the system states (the evolution of the model states), p is a terminal cost, and U are the input constraints. Note that for non-predictive control, one can use same ideas but set t 1 as a constant, eliminate the terminal cost p and ignore Eqs. (2) and (6). min
u N X g 1 ut ; xt ; wt pxN t1

1 t 1; . . . ; N t 1; . . . ; N t 1; . . . ; N 2 3 4

s:t:

xt1 f xt ; ut ; wt g 2 ut ; xt ; wt  0 ut 2 U

min
u

t 2 1...N

max g 1 ut ; xt ; wt pxN t 1; . . . ; N t 1; . . . ; N t 1; . . . ; N

5 6 7 8

s:t:

xt1 f xt ; ut ; wt g 2 ut ; xt ; wt  0 ut 2 U

An important aspect of the MPC optimization problem is that each problem faced is quite similar to the problem at the previous time-step. And other earlier time-steps may have had similar conditions to the current one. This feature of the MPC optimization problem turns out to be quite useful in generating good optimization starting points, although it is often overlooked, and to the authors knowledge it has not yet been taken advantage of in previous buildings research. One must be careful to avoid getting stuck in local minima when using this idea, but this concern can be overcome in a number of ways, one of which is to use a multi-start optimization algorithm, as used in the example below. 3.2. Framework overview Within the software framework, the problem is divided into three layers. The rst layer consists of the simulation software that is being used for the on-line model. The second layer is the optimization layer, which runs one of many possible algorithms to try to solve the optimization problem for a given time-step. The third layer is an organization layer, which interfaces with the building central control system, sets up the optimization problem at each time-step, and performs a number of other organizational and learning tasks. Text les are used to pass information between layers. The basic structure is shown in Fig. 2, along with how it interfaces with the buildings energy management system and with a prediction module (which is considered as a separate problem from the optimization problem itself, and is not dealt with in detail here). The simulation layer of the software framework can use any energy simulation program, as long as it reads and writes to text les and can be called from the command line. The optimization layer uses GenOpt as the interface between the optimization algorithms and the simulation program. The organization layer uses a java program named SimCon, which was written to meet the requirements of this layer. For a given time-step, SimCon begins the computation process by reading the text les for the current conditions and the predicted future conditions, and it ends by writing the decisions text le. So from the point of view of a central control system, its use requires just that it write some of its sensor information to text les, call SimCon, wait a pre-determined amount of time, and then read a decision text le which tells it what set-points to use. Details of the text le congurations, syntax, source code and example les are included with the beta version download [2].

Fig. 1. Receding horizon control.

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

1087

of various model inaccuracies (by changing parameter values so that they are different in the two models), improper initialization of the online model, or prediction inaccuracies. 4. Example algorithm An example optimization instructions algorithm and optimization algorithm pair was developed for this framework, based on a genetic algorithm. Although genetic algorithms tend to have slow convergence rates, it was used here because of interesting possibilities for coupling with SimCon. In particular, a genetic algorithm can be initialized with numerous starting points, so they can use some rule-based or learning-based suggestions for starting points and still have other randomly generated starting points that may allow it to discover unexpected points that may be better than those around the given points. 4.1. Basic genetic algorithm description
Fig. 2. Structure overview.

3.3. The organizational layer Aside from its roles of interfacing and overall coordination, the organizational layer does two important things: (1) it stores the results from previous time-steps; and (2) it writes an optimization instructions le which gets read by the optimization algorithm within the optimization layer. Similar to how GenOpt has an extendable library of optimization algorithms, SimCon has been developed with an extendable library of optimization instruction algorithms, from which the user can select (or develop) the algorithm that will write the optimization instructions le. Since this le must be read by the optimization algorithm in GenOpt, these two algorithms must be coordinated. In the case study considered below, the optimization instructions le is used to specify starting points and constraints for the optimization algorithm. (Other uses are possible, as discussed in Section 6.) In particular, the optimization instruction algorithms used here specify optimization starting points based on the results from the previous time-step, and starting points and constraints based on knowledge-based rules. 3.4. Virtual testing environment A single run of SimCon is just for one time-step. To test its performance, it needs to be used within a real-time environment, with a central controller sending it conditions and predictions at each time-step, and then acting on its decisions. This can be an actual buildings control system, or it can be connected to the control system of a virtual building. This link between SimCon and another building model can be done a number of different ways. The Building Controls Virtual Test Bed, recently developed at LBNL [29] provides a good framework for this. In the case study considered below, the virtual building is a copy of the online model in the controller, but slowed down to real-time (through the addition of a TRNSYS object that causes the model to sleep for a specied time period at each time-step), and with the addition of a supervisory control object in TRNSYS that writes conditions les, calls SimCon and reads the decisions le. This two-model virtual testing arrangement allows for easier testing of SimCon than would be possible through an actual supervisory controller implementation in a real building. But as importantly, it also allows one to test various MPC congurations under identical conditions, and allows one to systematically consider the impacts

Genetic algorithms (GAs) are inspired by biological evolution, using concepts of mating, mutation and natural selection to impel successive populations of individuals (with each individual representing a point in the search space) towards an optimization of the objective function. An initial population of points is chosen, usually at random from throughout the search space, and as the population evolves from one generation to the next new points emerge, with the evolutionary processes favouring the emergence of points that are ever closer to optimal. Their path through the search space is stochastic (there is some randomness in the mutation and mating processes through which new points are chosen), so they can sometimes be less efcient as optimizers than other possible algorithms, but their stochastic search path also makes them more able to avoid being trapped in local minima. Genetic algorithms have been used extensively in previous buildings research for design optimization problems [3035], but their use in control optimization has been limited. A genetic algorithm for GenOpt was written for this example, and is included in the SimCon beta download [2]. It uses one-point crossover and uniform mutation, and allows the user to specify the population size, the percentage elite to maintain, and the rates for crossover and mutation. To avoid problems that may result in later generations due to a depletion of diversity, the algorithm is congured to test for stagnation and if necessary it increases the mutation rate to 100% for a generation. It also allows the user to stipulate the maximum number of simulations to run (some sort of time-limiting factor is necessary for use within control optimization), and the user can choose to keep the values considered for each variable to a desired level of precision by having the GA search over points on a user-dened grid. 4.2. Using the optimization instructions le The genetic algorithm in GenOpt is also congured to read and use starting points and constraints from the optimization instructions le. A companion optimization instructions algorithm was written for the SimCon library which writes starting points and constraints to the instructions le, based on the following three considerations: 1. The values obtained in the previous optimization solution (which are stored by SimCon) are transferred over to an initial point in the current optimization by moving them all back one time-step, and nding an arbitrary way of lling in the last timestep values for the current point (in the example case below, the last time-step value is set equal to the second-last one).

1088

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

2. Using rules for suggesting starting points: When Henze et al. performed their case study on thermal storage control using model predictive control [7,8], they found a number of points where the controller actually performed worse than the standard heuristic controller they were comparing against, as it was stuck in a local minimum. This situation of performing worse than a heuristic controller can be avoided if the knowledge embedded in the heuristic controller is also embedded in the model predictive controller in the form of optimization starting points. (Note that, however, with an imperfect model or poor predictions, this knowledge is not necessarily enough to keep MPC from performing worse than the heuristic controller.) 3. Using rules to set search-space constraints: In order to make the time-constrained optimization problem more tractable, it is helpful to eliminate parts of the search space that are unlikely to produce good control points. (In the case study below, for example, it may be assumed that the optimal point does not include set-points near the end of the trim that are less than 25  C, so the constraints for these variables can be modied to 25 < T < 26  C, rather than their original 22 < T < 26  C.) 5. Example application study 5.1. Demand response with zone temperature ramping A number of previous studies [36,37] have shown that demand response with zone temperature reset can be an effective demand trimming strategy, helping to avoid blackouts or the need for new electrical production capacity, by producing substantial decreases in cooling demand over a critical period (usually in the range of 1 3 h) without causing too much occupant discomfort. Most of these previous studies have considered very simple control strategies, focusing on the demand response potential with strategies that could easily be mass-deployed. However, research by Lee and Braun [3841] has looked more closely at the optimal control of this set-point modication to minimize the peak demand over the trimming period. Their work has been along two related tracks: (1) the development of an inverse model of a small commercial building, and the use of this model to determine optimal set-point trajectories [38,41]; and (2) the development and evaluation of simple control rules that can approximate optimal control [39,40]. In the rst track, they used the fact that the optimal set-point trajectory produces a at demand over the trimming period. They used the inverse model to calculate, at any given point in time, what the set-point must be to produce a particular demand level. They then iterated over the demand level to nd the minimum level possible within the problem constraints, and the collection of set-points associated with that level is thus the optimal set-point trajectory. This is a clever approach to the problem. The implementation is only slightly different than the model predictive control approach considered in this paper, in that it does not march along making control decisions each time-step in real-time, but rather determines the full trajectory before the trim. But it could easily be made to march along and re-do the optimization at each timestep as new information became available (as noted in [41]). The use of the demand-target approach by backing out the set-point values for a specic demand level was possible because they were using a simplied inverse model, but it would be very difcult with a more detailed model and a standard simulation tool, as used in the example study below. The second track of Lee and Brauns research in this area, that of developing and evaluating near-optimal rule-based strategies for demand response with zone temperature ramping, is what is used most in this paper. In particular, they have determined that a very
Fig. 3. No trim.

effective strategy is to use a logarithmic set-point trajectory over the trim period. In general, the problem of optimizing demand response with temperature ramping has been well covered by Lee and Braun, and their simplied rule-based approaches are very effective. Having such well-charted territory is useful for the case study, and their insights and heuristic rules are used both as a starting point and as a basis for comparison. 5.2. Case study description The case study considered is for optimal demand response using zone temperature set-point ramping for an ofce space in Ottawa. The ofce space considered is area 8B1 of Place-du-Portage, which has a fairly standard open plan conguration. It is the home of the Innovations & Solutions Directorate of Public Works and Government Services Canada, and it has been outtted with a highly instrumented experimental personal environmental control (PEC) system that gives occupants control over their lighting and over a jet-diffuser that provides conditioned air to their workspace. The substantial data availability with this experimental system makes it amenable to the creation and calibration of a detailed thermal model,2 which was done in TRNSYS. This model was used both as the online model within the controller, and also to act as the actual building for the virtual tests. In the case study, the personal airconditioning controls are over-ridden in favour of a demand response action, for one particularly hot afternoon in July. The goal of the demand response action is to minimize the maximum 15-min cooling demand between 1:00 and 4:00 p.m., with the control action allowed to begin at 12:30 p.m., and the zone temperature set-point constrained to remain between 22 and 26  C. Fig. 3 shows the base case with no action. 5.3. Results Some heuristic controls were considered for comparison with MPC. Fig. 4 (jump trim) shows the effect of simply increasing the set-point to 26  C immediately and leaving it at that level throughout the demand response period (this is the most common simple strategy used in applications). Fig. 5 (linear trim) shows
2

Model details are available in [42].

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

1089

Fig. 4. Jump trim.

Fig. 6. Log trim.

the effect of increasing the set-point linearly over the demand response period. Fig. 6 (log trim) shows the effect of increasing the set-point logarithmically over the demand response period, as suggested by Lee and Braun. (The log trim requires the setting of a parameter to determine the steepness of the curve. For this study, that parameter was selected by an iterative search, simulating possible values before using the best one.) It should be noted here that these heuristic controls were also implemented with a 15-min control time-step in the virtual tests (as were the MPC cases), but in reality such simple control strategies could be implemented on a much shorter time-step and thus may perform slightly better than shown here, but the improvement would be minor. The MPC cases were split into two types: ideal cases, where the online model exactly matched the virtual test simulation model, the model was initialized perfectly and the prediction was perfect; and non-ideal cases, considering the effects of model mismatch, imperfect initialization and imperfect prediction. The case shown

in Fig. 7 is an ideal case with the computation time constraint removed so that the model predictive controller can be used to more closely approximate optimal control (the computation was allowed to run over three nights rather than being constrained to the 3-h window it would have in a real-time application). Note that, intuitively and supported by [38,41], the optimum should produce a horizontal line in the load during the trim period. This case comes closer to that optimum than do any of the heuristic cases, but it is still not quite optimal. The results of the heuristic control cases and some of the MPC cases considered are shown in Table 1. It was found that the logarithmic trim was nearly ideal in this case, and should be recommended for use in other similar cases. The optimal set-point conguration, however, is actually quite complex and likely highly case-dependent, and it would be difcult to imagine a rule-based approach that would capture it. The logarithmic approach produced a 28.0% trim, while the optimal was estimated to be

Fig. 5. Linear trim.

Fig. 7. MPC trim with loosened time constraints.

1090 Table 1 Summary of results: trim percentages. Heuristic Base case Jump trim Linear trim Log trim

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

23.0% 13.7% 28.0% 30.1% 29.5% 28.0% 5.9% 27.1% 10.1%

Model predictive

Ideal, no time constraints Ideal, with time constraints Ideal, no learning Ideal, no rule suggestions Imperfect calibration Imperfect initialization

slightly greater than the 30.1% produced by the ideal non-timeconstrained MPC. The application of MPC in the ideal case and with the real-time computation constraints produced a 29.5% trim. Without rule-based initial suggestions or previous-time-step learning, MPC (as congured with the genetic algorithm) does not perform nearly as well, because it does not converge quickly enough towards the optimum. And the performance of the controller drops off as circumstances move away from the ideal: with imperfect model calibration but perfect initialization and prediction, MPC performed slightly worse than the logarithmic approach, producing a 27.1% trim; and with perfect model calibration but imperfect initialization and prediction, the MPC results became more sporadic and much further from the optimal. (It should be noted that since the genetic algorithm is stochastic, the results can vary from test to test. But in the ideal cases with logarithmic starting points, MPC always performs somewhere between the log trim and the optimal.) 6. Discussion The software framework presented here opens up a variety of possibilities for further research. 6.1. Appropriate complexity level for application studies In the case study presented above, heuristic control was shown to perform nearly as well as MPC (or possibly better, if one considers model mismatch and imperfections in initialization and prediction). But if the scenario were more complex, for example if light dimming is being applied alongside the zone temperature ramping or if more detailed consideration is to be given of occupant comfort under these transient thermal conditions [43], then MPC may be worth considering. And one must consider other benets of MPC as, such as ability to easily adapt to changing electricity rate structures, and its links with other model-based building operations improvements as discussed below. But in general, the relative performance of heuristic control and MPC in this case study suggests that there might be a certain level of problem complexity below which MPC is simply not appropriate. It is difcult to assess what this level might be, but future research should be considering more complex problems (with more control variables and/or more complex control objectives), if we are to test practical applications of MPC in buildings. 6.2. Improving optimization results Genetic algorithms have the benet of allowing multiple starting-point suggestions alongside other random starting points and thus can more easily avoid local minima, but they are perhaps not ideal for this control-optimization application because they do not converge to solutions very quickly. Other optimization algorithms should be developed for use within this supervisory

control framework. (For example, the generalized pattern search algorithms available in the standard GenOpt library should be explored for use in this context.) Hybrid algorithms are also likely worth developing for this application, with different approaches used for a global search phase and for a local search phase, and the optimization algorithm could be made to switch from one to the other after a certain number of simulations. And there could be more advanced ways of using the organizational layer to inform the optimization algorithm. The organizational layer could search for more complex patterns in the previous results and pass these on to the optimization algorithms: it could nd correlation strengths between variables in optimization outputs, noting conditions where variables should take on similar values (e.g. if in previous time-step optimizations the variable values tend to change very little over the time horizon, then the optimization layer could restrict its search); and the organizational layer could perform sensitivity analyses on the previous results to determine which variables have the greatest effect on the objective function, then tell the optimization algorithm to focus on these variables rst before tweaking the other variable values. It is worth noting that with an increased use of such techniques, SimCon algorithms could begin to take on some of the characteristics and benets of learning-algorithm (e.g. neural network) approaches. Other possible improvements to the use of organization-layer information include nding ways of treating future time-steps differently the further into the future they are (e.g. to decrease their relative precision or constrain their search ranges as a way of placing more focus on the moreimmediately upcoming time-steps), and developing or rening heuristic starting point rules based on results from previous optimizations. All of these changes could signicantly improve the efcacy of the optimization. However, they all still walk around the simulation program, treating it essentially as a black box, generating and testing possible points. If simulation tools were to develop in such a way as to provide the optimizer more access to the problem equations and/or to ensure differentiability, the optimizer could use this to parse the problem into sub-regions or to analytically derive gradients, both of which could dramatically improve computational efciency. Such development could also allow for the backing-out of set-point values given a desired objective, as was used in the inverse-model studies of Lee and Braun [38,41]. 6.3. Imperfect models, initializations and predictions The case study results have highlighted the need for more consideration of model accuracy, initial conditions and predictions. If not dealt with effectively, these aspects could undermine the performance of any practical or theoretical application of MPC. More research is required in these areas, but the following are some good starting points. 1. Continuous model calibration: The automated error minimization approach to model calibration [10,44,45] allows for the model to be continuously tuned to become more accurate over time. The conditions and optimization results database in SimCon could be used for this. 2. Initialization: As noted earlier, this is an inherent problem with the use of most building simulation tools, which do not allow for explicit state initialization, but instead require a warm-up period. Development of more controls-friendly building simulation tools will help get around this. But in the meantime, it would be helpful to devise ways of rewinding or copying the simulation process such that the warm-up period only needs to be run once per optimization.

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092

1091

6.4. Other potential research areas with SimCon One of the most alluring aspects of simulation-based control is that it overlaps well with a number of other possible applications of simulation in building operations. A controllers well-tuned physics-based model of the building could be of great use in evaluating possible retrot strategies. And the overlap between the idea of a model-based controller with automated calibration and a model-based diagnostics tool [46,47] is interesting. In the rst case you want an accurate model of how the building is actually performing, but in the second case you want an accurate model of how the building should be performing. But the models must share the same structure, and could be treated as two sides of the same model within an integrated controls-and-diagnostics system, differing only in terms of some parameter values. Although this paper has focused on the study of real-time MPC, it should be acknowledged that another possible use of SimCon (and perhaps a much more practical use in the short and mid term, and possibly even in the long term) is for the development and evaluation of standard heuristic control sequences for complex systems. By relaxing the computational time constraints and providing ideal circumstances (perfect model calibration, initialization and prediction), SimCon could be used to approximate optimal control for any given (simulated) set of conditions. Iterating over various sets of conditions and analyzing the SimCon outputs could allow researchers to devise rules for near-optimal control. There is an R&D policy interest in knowing the technical energy savings potential with near-optimal control of highly dynamic integrated building systems, but there is currently only a modest amount of literature available on the energy savings potential through integrated systems control (e.g. [48,49]), and none of this extends to full integration of HVAC, lighting and dynamic facades. SimCon could be used to provide an approximation of optimal control for any type of complex idealized building system, and so could be useful in estimating an upper bound on the energy savings potential. Personal control tends to lead to better occupant satisfaction and can also lead to decreased energy use in some cases [50,51], but not all. Perhaps the best conguration would involve some combination of automated control and occupant control. SimCon could be used to determine appropriate control for the automated parts of a building given stochastic occupant control of other parts of it. Or it could be used to determine optimal congurations and suggest these to the occupant to accept or ignore. Or it could be used to determine when to allow occupant control with a redgreen-light system. 7. Conclusions The example application study and example algorithm have provided an illustration and test of the software framework. And the study has supported the ndings of Lee and Braun that a logarithmic trim is near optimal for demand response by zone temperature reset. The framework should be used for more complex application studies in the future, and more optimization and organization-layer algorithms should be developed. Model predictive control holds promise for energy efciency and demand response in integrated supervisory building control systems. It is hoped that the framework presented herein will help to facilitate further research and development in this eld. Acknowledgement The authors gratefully acknowledge the support of Public Works and Government Services Canada for this research.

References
[1] M. Wetter, GenOpt: generic optimization program, http://gundog.lbl.gov/GO/. [2] B. Coffey, Simcon beta, http://ftp.tech-env.com/./pub/Coffey/SimCon. [3] M. Morari, J. Lee, Model predictive control: past, present and future, Computers and Chemical Engineering 23 (1999) 667682. [4] D. Mayne, J. Rawlings, C. Rao, P. Scokaert, Constrained model predictive control: stability and optimality, Automatica 36 (6) (2000) 789814. [5] F. Borrelli, A. Bemporad, M. Fodor, D. Hrovat, An MPC/hybrid system approach to traction control, IEEE Transactions of Control Systems and Technology 14 (3) (2006) 541552. [6] G. Kelly, Control system simulation in North America, Energy and Buildings 10 (1988) 193202. [7] G. Henze, D. Kalz, S. Liu, C. Felsmann, Experimental analysis of model-based predictive optimal control for active and passive thermal storage inventory, HVAC&R Research 11 (2) (2005) 189213. [8] G. Henze, M. Krarti, Predictive optimal control of active and passive building thermal storage inventory: nal report, DOE Award Number: DE-FC-2601NT41255, 2005. [9] G. Henze, D. Kalz, C. Felsmann, G. Knabe, Impact of forecasting accuracy on predictive optimal control of active and passive building thermal storage inventory, HVAC&R Research 10 (2) (2004) 153177. [10] G. Henze, S. Liu, Calibration of building models for supervisory control of commercial buildings, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [11] S. Liu, G. Henze, Experimental analysis of simulated reinforcement learning control for active and passive building thermal storage inventory, Part 1, Theoretical foundation, Energy and Buildings 38 (2006) 142147. [12] S. Liu, G. Henze, Experimental analysis of simulated reinforcement learning control for active and passive building thermal storage inventory, Part 2, Results and analysis, Energy and Buildings 38 (2006) 148161. [13] J. Clarke, J. Crockroft, S. Conner, J. Hand, N. Kelly, R. Moore, T. OBrien, P. Strachan, Simulation-assisted control in building energy management systems, Energy and Buildings 34 (2002) 933940. [14] M. Kummert, P. Andre, A. Argigiou, Performance comparison of heating control strategies combining simulation and experimental results, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [15] M. Kummert, P. Andre, Simulation of a model-based optimal controller for heating systems under realistic hypotheses, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [16] A. Mahdavi, Simulation-based control of building systems operation, Building and Environment 36 (2001) 789796. [17] A. Mahdavi, B. Spasojevic, K. Brunner, Elements of a simulation-assisted daylightresponsive illumination systems control in buildings, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [18] G. Suter, O. Icoglu, A. Mahdavi, B. Spasojevic, Position uncertainty in space scene reconstruction for simulation-based lighting control, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [19] A. Mahdavi, C. Proglhof, A model-based method for the integration of natural ventilation in indoor climate systems operation, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [20] S. Wang, X. Jin, Model-based optimal control of VAV air-conditioning system using genetic algorithm, Building and Environment 35 (2000) 471487. [21] N. Nassif, S. Kajl, R. Sabourin, Simplied model-based optimal control of VAV air-conditioning system, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [22] B. Flake, Parameter estimation and optimal supervisory control of chilled water plants, PhD thesis, University of Wisconsin-Madison, 1998. [23] C. Priolo, S. Sciuto, F. Sperduto, Efcient design incorporating fundamentals improvements for control and integrated optimisation: nal report, http:// lesowww.ep.ch/anglais/techint/edicio.pdf, 2001. [24] J. Clarke, Domain integration in building simulation, Energy and Buildings 33 (2001) 303308. [25] Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, A. Packard, M. Wetter, P. Haves, Model predictive control for the operation of building cooling systems, in: American Control Conference, 2010. [26] G. Ddfd, et al., Opticontrol project, http://www.opticontrol.ethz.ch/. [27] M. Wetter, Modelica-based modeling and simulation to support research and development in building energy and controls systems, Journal of Building Performance Simulation, 2009. [28] http://www.mathworks.com/products/mpc/. [29] M. Wetter, P. Haves, A modular building controls virtual test bed for the integration of heterogeneous systems, SimBuild 2008 (2008) 6976. [30] M. Wetter, J. Wright, A comparison of deterministic and probabilistic optimization algorithms for nonsmooth simulation-based optimization, Building and Environment 39 (2004) 989999. [31] W. Wang, R. Zmeureanu, H. Rivard, Two-phase application of multi-objective genetic algorithms in green building design, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005.

1092

B. Coffey et al. / Energy and Buildings 42 (2010) 10841092 response using a modied genetic algorithm, M.A.Sc. thesis, Concordia University, 2008. G. Newsham, C. Donnelly, S. Mancini, R. Marchand, W. Lei, K. Charles, J. Veitch, The effect of ramps in temperature and electric light levels on ofce occupants: a literature review and a laboratory experiment, ACEEE Summer Study on Energy Efciency in Buildings, 2006. J. Sun, T. Reddy, Calibration of building energy system simulation programs using the analytic optimization approach (rp-1051), International Journal of HVAC&R Research 12 (1) (2006) 177196. T. Reddy, Literature review on calibration of building energy simulation programs: uses, problems, procedures, uncertainty, and tools, ASHRAE Transactions 112 (2) (2006) 226240. P. Xu, P. Haves, M. Kim, Model-based automated functional testingmethodology and application to air-handling units, ASHRAE Transactions (2005) OR-05-13-4. P. Xu, P. Haves, D. Curtil, A library of HVAC component models for use in automated diagnostics, in: Proceedings of Simbuild, 2006. E. Lee, D. DiBartolomeo, F. Rubinstein, S. Selkowitz, Low-cost networking for dynamic window systems, Energy and Buildings 36 (2004) 503513. P. Haves, R. Hitchcock, F. Rubinstein, P. Xu, Assessment of building control systems, Final report to BT, OERE, DOE, Contract No. DE-AC02-05CH11231, 2007. R. de Dear, G. Brager, The adaptive model of thermal comfort and energy conservation in the built environment, International Journal of Biometeorology 45 (2001) 100108. G. Newsham, J. Veitch, C. Arsenault, C. Duval, Lighting for vdt workstations 1: effect of control on energy consumption and occupant mood, satisfaction and discomfort, Research report, Institute for Research in Construction, National Research Council Canada, B3208.3 RR-165, 2004.

[32] J. Wright, A. Alajmi, The robustness of genetic algorithms in solving unconstrained building optimization problems, in: Proceedings of the 9th International Building Performance Simulation Association (IBPSA) Conference, Montreal, Canada, 2005. [33] L. Zhou, F. Haghighat, Optimization of ventilation systems in ofce environment, Part I: methodology, Building and Environment 44 (2008) 651656. [34] L. Zhou, F. Haghighat, Optimization of ventilation systems in ofce environment, Part II: results and discussion, Building and Environment (2008) 44. [35] L. Magnier, F. Haghighat, Multiobjective optimization of building design using TRNSYS simulations, genetic algorithm, and articial neural network, Building and Environment 45 (2010) 739746. [36] P. Xu, P. Haves, J. Braun, L. Hope, Peak demand reduction from pre-cooling with zone temperature reset in an ofce building, in: ACEEE Summer Study on Energy Efciency in Buildings, 2004. [37] P. Xu, Evaluation of demand shifting strategies with thermal mass in two large commercial buildings, in: Proceedings of SimBuild, 2006. [38] K. Lee, J. Braun, Development and application of an inverse building model for demand response in small commercial buildings, in: Proceedings of SimBuild, 2006. [39] K. Lee, J. Braun, Development of methods for determining demand-limiting setpoint trajectories in commercial buildings using short-term data analysis, in: Proceedings of SimBuild, 2006. [40] K. Lee, J. Braun, Evaluation of methods for determining demand-limiting setpoint trajectories in commercial buildings using short-term data analysis, in: Proceedings of SimBuild, 2006. [41] K. Lee, J. Braun, Reducing peak cooling loads through model-based control of zone temperature setpoints, in: Proceedings of the American Control Conference, 2007. [42] B. Coffey, A development and testing framework for simulation-based supervisory control with application to optimal zone temperature ramping demand

[43]

[44]

[45]

[46] [47] [48] [49] [50]

[51]