nage c “Bees he ___with Constr’Predictive Control with Constraints J.M. Maciejowski iuKonn a PEN Aninprnt of Pearson Education Harlow, England - London » New York » Reading, Massachusetts - San Francisco - Toronto - Den Mills, Ontario - Sydney Tokyo - Singapore - Hong Kong - Seoul -Taipet - Cape Town - Madrid - Mesico City - Amsterdam » Munich - Paris - MilanContents Preface xi List of Mini-Tutorials xiv List of Figures xv 1 Introduction 1 1.1. Motivation u 1.2. The ‘receding horizon’ idea 7 1.3. Computing the optimal inputs 9 1.4 Asimple MATLAB program 14 1.5 Offset-free tracking 17, 1.6 — Unstable plant 22 1.7 Early history and terminology 25 1.8 Predictive control in the control hierarchy 26 1.9 General optimal control formulation 28 1.10 What is in this book 31 Exercises 32 2 Abasic formulation of predictive control 36 2.1 State-space models 36 2.1.1 Form of the model 36 2.1.2 Linear model, nonlinear plant 37 2.1.3. First-principles models and system identification 40vi Contents a on 2.4 25 2.6 rie A basic formulation General features of constrained predictive control Alternative state variable choices Allowing for computational delay Prediction 2.6.1 No disturbances, full state measurement 2.6.2 Constant output disturbance 2.6.3. Using an observer 2.6.4. Independent and realigned models Example: Citation aircraft model Exercises 3. Solving predictive control problems 3.1 3.2 a 3.4 Unconstrained problems 3.1.1 Measured state, no disturbances 3.1.2 Formulation as a least-squares problem 3.1.3. Structure of the unconstrained controller 3.1.4 Estimated state Constrained problems 3.2.1 Formulation as a QP problem 3.2.2 Controller structure Solving QP problems 3.3.1 Active Set methods 3.3.2. Interior point methods Softening the constraints Exercises 4 Step response and transfer function formulations 41 4.2 Step and pulse response models 4.1.1. Step and pulse responses 4.1.2. Relations to state-space models 4.1.3 Prediction using step response model 4.1.4 State-space models from step responses Transfer function models 4.2.1 The basics 4.2.2 Prediction using transfer functions 4.2.3 Prediction with a disturbance model 4.2.4 The GPC model 4.2.5 State-space interpretations 4.2.6 Multivariable systems Exercises 5 Other formulations of predictive control 51 Measured disturbances and feedforward 41 47 49 50 53 54 56 57 62 64 70 74 74 74 77 79 80 81 81 84 88 89 94 97 104 108 108 108 111 113 115 121 121 124 127 133 134 142 143 145 145Contents vii 5.2. Stabilized predictions 148 5.3 Decomposition of unstable model 150 5.4 Non-quadratic costs 151 5.4.1 Characteristics of LP and QP problems 151 5.4.2 Absolute value formulations 154 5.4.3, Min-max formulations 155 5.5 Zones, funnels and coincidence points 156 5.6 Predictive Functional Control (PFC) 159 5.7 Continuous-time predictive control 163 Exercises 165 Stability 167 6.1 Terminal constraints ensure stability 169 6.2 Infinite horizons 172 6.2.1 Infinite horizons give stability 172 6.2.2 Constraints and infinite horizons — stable plant 176 6.2.3 Constraints and infinite horizons — unstable plant 178 6.3 Fake algebraic Riccati equations 180 6.4 Using the Youla parametrization 183 Exercises 185 Tuning 188 7.1 What are we trying to do? 188 7.2 Some special cases 191 7.2.1 Effect of control weighting 191 Mean-level control 192 Deadbeat control 192 ‘Perfect’ control 193 7.3. Frequency-response analysis 199 7.4 Disturbance models and observer dynamics 201 7.4.1 Disturbance models 201 7.4.2. Observer dynamics 203 7.5 Reference trajectory and pre-filter 211 Exercises 214 Robust predictive control 217 8.1 Formulations of robust control oe 8.1.1 Norm-bounded uncertainty 218 8.1.2. Polytope uncertainty 220 8.2 The tuning procedure of Lee and Yu 221 8.2.1 Simplified disturbance and noise model 221 8.2.2 Tuning procedure 224 8.3 The LQG/LTR tuning procedure 225 8.4 LMl approach 233viii Contents 8.5 8.4.1 Overview 8.4.2 Robustness without constraints 8.4.3 Robustness with constraints Robust feasibility 8.5.1 Maximal output admissible sets 8.5.2 Robust admissible and invariant sets Exercises 9 Two case studies a 9.2 Shell oil fractionator 9.1.1. Process description 9.1.2 Control specifications 9.1.3 Initial controller design 9.1.4 Controller performance and refinement 9.1.5 Constraint softening 9.1.6 Robustness to model errors Newell and Lee evaporator Exercises 10 Perspectives 10.1 Spare degrees of freedom 10.1.1 Ideal resting values 10.1.2 Multiobjective formulations 10.1.3 Fault tolerance 10.2 Constraint management 10.3 Nonlinear internal models 10.3.1 Motivation and approaches 10.3.2 Sequential quadratic programming 10.3.3 Neural net models 10.3.4 Sub-optimal nonlinear MPC 10.4 Moving-horizon estimation 10.5 Concluding remarks References A Some commercial MPC products Al A2 A3 A4 AS A6 Aspentech: DMCPlus Honeywell: RMPCT Simulation Sciences: Connoisseur Adersa: PFC and HIECON ABB: 3dMPC Pavilion Technologies Inc.: Process Perfecter 233 234 237 239 240 241 246 248 248 248 251 253 258 261 263 269 274 292 305 306 308 309 311 311 311B_ MATLAB program basicmpe C The MPC toolbox C1 General remarks C2 Functions scmpe2 and scmpen12 C.3 Functions scmpe3 and scmpc4 Author index Subject index Contents ix 313 318 318 321 cree 323 326Predictive Control, or Model-Based Predictive Control (‘MPC” or ‘MBPC”) as it is sometimes known, is the only advanced control technique — that is, more advanced than standard PID control — to have had a significant and widespread impact on industrial process control. The main reason for this is that it is © The only generic control technology which can deal routinely with equipment and safety constraints. Operation at or near such constraints is necessary for the most profitable or most efficient operation in many cases. The penetration of predictive control into industrial practice has also been helped by the facts that ® Its underlying idea is easy to understand, Its basic formulation extends to multivariable plants with almost no modification, ® It is more powerful than PID control, even for single loops without constraints, without being much more difficult to tune, even on ‘difficult’ loops such as those containing long time delays. Predictive control was developed and used in industry for nearly 20 years before attracting much serious attention from the academic control community. This commu- nity tended to ignore its potential for dealing with constraints, thus missing its main advantage. In addition, it tended to point to the fact that, when constraints are ignored, predictive control is equivalent to conventional, though generally ‘advanced’, linear control, and hence apparently nothing new. This is true, but again misses the important point that issues such as tunability and understandability are crucial for the acceptance of a control technology, at least in the process control environment. Fortunately, the academic community has for some years now appreciated that predictive control really does offer something new for control in the presence of constraints, and has provided much analysis, and new ideas, to such an extent that it has gone beyond current industrial xixii Preface practice, and is preparing the ground for much wider application of predictive control — potentially to almost all control engineering problems. The constant increase in computing speed and power certainly makes that a real prospect. ~ In this book I have attempted to bring everything together. I have tried to convey the simplicity of the underlying concepts of predictive control, but also to show how it relates to existing control theory, and indeed how much more can be done with it when its use is informed by a knowledge of standard control techniques, such as state estimation, disturbance modelling, and frequency response analysis. Predictive control involves optimization, and I have included details of how to solve at least some of the optimization problems encountered. I have also tried to display the main directions of current research, and give some indication of likely future directions, both of research and of applications. This book assumes that the reader has some previous exposure to control theory, such as a first course and some contact with state-space models. It will be suitable for graduate students taking systems and control courses, but I hope that it will also prove useful to practising industrial engineers. In order to increase its usefulness to non-students, and also to alleviate the problem of prerequisites for students, I have included a number of Mini-Tutorials. These are one-page summaries of topics, such as observers or Lyapunov equations, that are needed in order to understand the material at certain points. I believe that the Mini-Tutorials contain enough explanation to allow the reader to follow the developments in the book, but of course they are not meant to replace in-depth study of these important topics. It is essential to have access to suitable software in order to solve any non-trivial predictive control problems, and to gain experience of how it actually works. This book assumes that the reader has access to MATLAB, together with its Control System Toolbox and Model Predictive Control Toolbox. Some simple MATLAB files have also been written, and some files which augment those available in the Model Predictive Control Toolbox — some of these also require the Optimization Toolbox. All such additional software is available on this book's Companion Web Site: http://www. booksites.net/maciejowski/ The versions of software used for the preparation of this book were: MATLAB 53.1 Control System Toolbox 42.1 Model Predictive Control Toolbox 1.0.4 Optimization Toolbox 2.0 My students Eric Kerrigan and Simon Redhead were responsible for most of the modifications to Model Predictive Control Toolbox functions which are available on the web site. This book originated in a course of lectures given at the Faculty of Aerospace Engineering in Delft, during November and December 1997. I would like to thank Bob Mulder, Hans van der Vaart and Samir Bennani for inviting me to spend a sabbatical term at Delft, for making all the arrangements, for making my stay at Delft both pleasant and interesting, and above all for having enough vision to believe that a course on predictivePreface xiii control was not out of place in an aerospace department. I would also like to thank Ton van den Boom for giving me feedback on the first few chapters, Rob de Vries (both of TU Delft) for some valuable help with Chapter 4, and Hans van der Vaart for providing me with the linearized model of the Citation aircraft which is used in a number of illustrative examples in the book. The use of an aircraft example may seem quirky, since predictive control has been used almost exclusively in the process industries. There are two reasons for this. Firstly, it reflects my conviction that predictive control has great potential in all application sectors, as a result of continuing increases in real-time computational possibilities. Secondly, most process control examples require considerable explanation of the context to those who are not chemical engineers, whereas most readers will be able to understand the (very simplified) aircraft example from their own experience. Of course there are also examples based on process control in the book, and the major case studies in Chapter 9 are both taken from process control. T gave a graduate course based on the material in this book to the Centre for Process Control Design at Lund University, and received several very valuable suggestions from both faculty and students who attended that course. Andrew Ogden-Swift of Honeywell Hi-Spec Solutions, Sean Goodhart of Aspentech, David Sandoz of SimSci-Predictive Control, and Jacques Richalet of Adersa have ali been very generous of their time, have provided me with details of their companies’ products, and have discussed the practicalities of predictive control most valuably and stimulatingly. David Sandoz and David Clarke both provided extremely useful, but also encouraging, feedback at various stages of the development of the manuscript, and Fred Loquasto of UC Santa Barbara read the almost-final version in record time, and picked up several errors and infelicities. Any errors remain my own responsibility, of course. JM. Maciejowski 14 October 2000 A Companion Web Site accompanies Predictive Control by Jan Maciejowski. Visit the Predictive Control Companion Web Site at -www.booksites.net/maciejowski Here you will find valuable teaching and learning material including: For Students and Lecturers: ® MATLAB files allowing you to model and simulate many of the systems presented in the book Links to software needed to run the simulations (commercial Matlab-based fiware from The Mathworks, and free extensions and other software from the author) © Links to websites related to predictive control For Lecturers: Downloadable solution's manualxiv List of Mini-Tutorials 1 2 3 4 5 6 7 8 9 i a Quadratic forms. State observers. Convex optimization, QP and LP problems. The singular value decomposition. Diophantine equations. Stability and Lyapunov functions. Linear quadratic optimal control. Interpreting frequency responses. The Kalman filter. Linear matrix inequalities. 42 38 84 117 126 110 181 200 226 235List of Figures 1,10 noah aes 29 2.10 aa Semi-batch reactor. Set-point improvement available from the use of predictive control. Surge level control. Predictive control: the basic idea. Results from basicmpc, exact model. Results from basicmpe with incorrect model gain. Results from trackmpe with incorrect model gain. Input and output disturbances and measurement noise. Simulation of system with input and output disturbances and measurement noise using noisympc. Control of unstable helicopter using unstampe. Typical use of predictive control — current practice. Use of predictive control — future trend? Paper machine with headbox. The boundary between controller and plant — a matter of convenience. Assumed timing of measuring and applying signals. Output disturbance. Response of Citation aircraft to 40 m step change in altitude set-point. Response of Citation aircraft to 400 m step change in altitude set-point. Response of Citation aircraft to 400 m step change in altitude set-point, with altitude rate constraint. Response to disturbance with altitude rate constraint. Response to disturbance with altitude rate set-point. Response of Citation aircraft to 400 m step change in altitude set-point, with altitude rate constraint and plant-model error. 22 24 27 28 46 49 oan i) 65 66 67 68 69 70 xvxvi List of Figures 3. ae 33 3.4 | S4 3.5 5.6 Structure of controller with no constraints and full state measurement. Structure of controller with no constraints and state observer. Structure of controller with constraints and state observer. A quadratic cost surface and a linear inequality constraint: constraint inactive. A quadratic cost surface and a linear inequality constraint: constraint active. Basic interior point method for Example 3.1. Plant-model error and hard constraints result in infeasibility, Soft output constraints restore feasibility. I-norm penalty with p Response to disturbance with softened altitude rate constraint, p = 5 x 10° Response to disturbance with softened altitude rate constraint, p = 104. Step response of 3-input, 2-output distillation column, Solid curves: original responses. Dotted curves: responses of 8-state model. The first 15 singular values of 11259 for the example. Spectral density of crew disturbance, modelled by (4.109) with p = 0.98, Generation of 6(k|k) — one interpretation. GPC applied to unstable helicopter. Feedforward from measured disturbances. Decomposition of unstable plant as a feedback interconnection of two stable systems, Decomposition of unstable plant as a stable model and ‘measured quasi-disturbance’. Coprime factorization of an unstable system, and its approximation. Cost contours and constraints for LP (left) and QP (right) problems, The black dot shows the optimal solution in each case, Using a QP formulation to place the desired operating point just inside the feasible region, Left: x, deviations cheaper than x2 deviations. Right: x2 deviations cheaper than x; deviations. A zone objective. A funnel objective. A funnel objective with a straight line boundary. A reference trajectory with only a few coincidence points. ‘A future input built up from a polynomial basis Finite and infinite horizons (no disturbances, perfect model). Finite horizon and deadbeat response. ‘Youla parametrization of all stable feedback systems (if the plant is stable). ‘Two degree of freedom feedback system. Mean-level control: set-point change on r). Mean-level control: set-point change on r3. Default (DMC) observer: response to step disturbance. Deadbeat observer: response to step disturbance. Deadbeat set-point response. Step demand on ry. 100 102 103. 120 121 129 130 142 146 150, 151 Isl 152 154 156 157 158 158 ed 173 174 184 190, 195 196 196 197 1987.10 aaa ve ie 7.14 ae 7.16 aa 7.18 8.1 9.10 Hel 9.12 9.13 List of Figures ‘Perfect’ controller: set-point response. Step demand on r}. ‘Perfect’ controller: set-point response. Step demand on r3. Singular values of $(z) (solid line) and T(z) (broken line) for mean-level control of the headbox. ‘A ‘proportional + integral’ (PI) controller. Predictive controller, unconstrained case. Predictive controller, simplified. Predictive controller, simplified again. Control of water temperature with perfect plant model, and model of air temperature disturbance. W = 1, V = 10°. The broken line shows the air temperature. Control of water temperature with model of air temperature disturbance. Perfect (solid line) and imperfect (broken line) models. W = 1, V = 1. Control of water temperature with model of air temperature disturbance, heater power constraints, and imperfect model. W = 1, V = 1. Controller structure with reference trajectories: non-anticipated set-point. Controller structure with reference trajectories: anticipated set-point. Feedback combination of controller and plant, with additive uncertainty model. The disturbance and noise model used by Lee and Yu [LY94]. The LQ state feedback regulator. ‘The Kalman filter as a feedback system. The closed loop with LQG controller. Disturbance and noise model used for LOG/LTR. The ‘Shell’ heavy oil fractionator. ‘Open-loop step responses of the ‘Shell’ heavy oil fractionator. Hankel singular values of the ‘Shell’ heavy oil fractionator, with T, = 4 minutes. Response with initial tuning parameters. Response with increased weight on z; and z» tracking errors, and penalized control moves. Response to measured disturbance. Response to unmeasured disturbance. Response to measured and unmeasured disturbances, and soft constraints. Response to measured and unmeasured disturbances, and hard constraints on 2(k + ilk) for é > 3. Largest singular values of sensitivity S(z) (broken line) and complementary sensitivity T(z) (solid line). Largest singular value (solid line), and estimates of complex j« (broken line) and real jz (dotted line) of K(z)S(z)W(z). Response with plant-model mismatch, unmeasured disturbance, and soft constraints. Response with plant-model mismatch, unmeasured disturbance, soft constraints, and increased control move penalty. xvii 198 199 201 203 204 205 205 208 209 210 212 214 219 222 ae 228 228 a 249 ae 256 259 260 260 261 262 264 266 266 267