Biao Huang, Ramesh Kadali

Dynamic Modeling, Predictive

Control and Performance
Monitoring
A Data-driven Subspace Approach

4y Spri
nnger
g<
Contents

Notation XIX

1 Introduction 1
1.1 An Overview of This Book 1
1.2 Main Features of This Book 4
1.3 Organization of This Book 4

Part I Dynamic Modeling through Subspace Identification

2 System Identification: Conventional Approach 9

2.1 Introduction 9
2.2 Discrete-time Systems 9
2.2.1 Finite Difference Models 10
2.2.2 Exact Discretization for Linear Systems 10
2.2.3 Backshift Operator and Discrete-time Transfer Functions 11
2.3 An Example of System Identification: ARX Modeling 12
2.4 Persistent Excitation in Input Signal 13
2.5 Model Structures 15
2.5.1 Prediction Error Model (PEM) 15
2.5.2 AutoRegressive with Exogenous Input Model ( A R X ) . . . . 15
2.5.3 AutoRegressive Moving Average with Exogenous
Input Model (ARMAX) 16
2.5.4 Box-Jenkins Model (BJ) 17
2.5.5 Output Error Model (OE) 17
2.5.6 MISO (Multi-input and Single-output) Prediction
Error Model 18
2.5.7 State Space Model 18
2.6 Prediction Error Method 19
2.6.1 Motivation 19
2.6.2 Optimal Prediction 21
2.6.3 Prediction Error Method 24
XIV Contents

2.7 Closed-loop Identification 25

2.7.1 Identifiability without External Excitations 26
2.7.2 Direct Closed-loop Identification 27
2.7.3 Indirect Closed-loop Identification 28
2.7.4 Joint Input-output Closed-loop Identification 29
2.8 Summary 29

3 Open-loop Subspace Identification 31

3.1 Introduction 31
3.2 Subspace Matrices Description 31
3.2.1 State Space Models 31
3.2.2 Notations and Subspace Equations 33
3.3 Open-loop Subspace Identification Methods 40
3.4 Regression Analysis Approach 40
3.5 Projection Approach and N4SID 43
3.5.1 Projections 43
3.5.2 Non-steady-state Kaiman Filters 44
3.5.3 Projection Approach for Subspace Identification 46
3.6 QR Factorization and MOESP 48
3.7 Statistical Approach and CVA 49
3.7.1 CVA Approach 49
3.7.2 Determination of System Order 51
3.8 Instrument-variable Methods and EIV Subspace Identification.. 51
3.9 Summary 53

4 Closed-loop Subspace Identification 55

4.1 Introduction 55
4.2 Review of Closed-loop Subspace Identification Methods 57
4.2.1 N4SID Approach 57
4.2.2 Joint Input-Output Approach 59
4.2.3 ARX Prediction Approach 60
4.2.4 An Innovation Estimation Approach 61
4.3 An Orthogonal Projection Approach 63
4.3.1 A Solution through Orthogonal Projection 63
4.3.2 The Problem of Biased Estimation and the Solution . . . . 67
4.3.3 Model Extraction through Kaiman Filter State Sequence 69
4.3.4 Extension to Error-in-variable (EIV) Systems 71
4.3.5 Simulation 72
4.4 Summary 78

5 Identification of Dynamic Matrix and Noise Model Using

Closed-loop Data 79
5.1 Introduction 79
5.2 Estimation of Process Dynamic Matrix and Noise Model 81
5.2.1 Estimation of Dynamic Matrix of the Process 84
5.2.2 Estimation of the Noise Model 85
 Contents XV

5.3 Some Guidelines for the Practical Implementation of the

Algorithm 86
5.4 Extension to the Case of Measured Disturbance Variables 87
5.5 Closed-loop Simulations 89
5.5.1 Univariate System 89
5.5.2 Multivariate System 90
5.6 Identification of the Dynamic Matrix: Pilot-scale
Experimental Evaluation 94
5.7 Summary 96

Part II Predictive Control

6 Model Predictive Control: Conventional Approach 101

6.1 Introduction 101
6.2 Understanding MPC 102
6.3 Fundamentals of MPC 103
6.3.1 Process and Disturbance Models 103
6.3.2 Predictions 105
6.3.3 Free and Forced Response 106
6.3.4 Objective Function 107
6.3.5 Constraints 107
6.3.6 Control Law 108
6.4 Dynamic Matrix Control (DMC) 108
6.4.1 The Prediction Model 109
6.4.2 Unconstrained DMC Design 111
6.4.3 Penalizing the Control Action 111
6.4.4 Handling Disturbances in DMC 112
6.4.5 Multivariate Dynamic Matrix Control 113
6.4.6 Hard Constrained DMC 115
6.4.7 Economic Optimization 116
6.5 Generalized Predictive Control (GPC) 117
6.6 Summary 118

7 Data-driven Subspace Approach to Predictive Control 121

7.1 Introduction 121
7.2 Predictive Controller Design from Subspace Matrices 122
7.2.1 Inclusion of Integral Action 125
7.2.2 Inclusion of Feedforward Control 127
7.2.3 Constraint Handling 128
7.3 Tuning the Noise Model 130
7.4 Simulations 132
7.5 Experiment on a Pilot-scale Process 138
7.6 Summary 141
XVI Contents

Part III Control Performance Monitoring

8 Control Loop Performance Assessment: Conventional

Approach 145
8.1 Introduction 145
8.2 SISO Feedback Control Performance Assessment 146
8.3 MIMO Feedback Control Performance Assessment 150
8.4 Summary 155

9 State-of-the-art M P C Performance Monitoring 157

9.1 Introduction 157
9.2 MPC Performance Monitoring: Model-based Approach 158
9.2.1 Minimum-variance Control Benchmark 158
9.2.2 LQG/MPC Benchmark 159
9.2.3 Model-based Simulation Approach 160
9.2.4 Designed/Historical vs Achieved 161
9.2.5 Historical Covariance Benchmark 161
9.2.6 MPC Performance Monitoring through Model Validation 162
9.3 MPC Performance Monitoring: Model-free Approach 165
9.3.1 Impulse-Response Curvature 165
9.3.2 Prediction-error Approach 166
9.3.3 Markov Chain Approach 166
9.4 MPC Economic Performance Assessment and Tuning 167
9.5 Probabilistic Inference for Diagnosis of MPC Performance 171
9.5.1 Bayesian Network for Diagnosis 171
9.5.2 Decision Making in Performance Diagnosis 173
9.6 Summary 175

10 Subspace Approach to MIMO Feedback Control

Performance Assessment 177
10.1 Introduction 177
10.2 Subspace Matrices and Their Estimation 179
10.2.1 Revisit of Important Subspace Matrices 179
10.2.2 Estimation of Subspace Matrices from Open-loop D a t a . . 180
10.3 Estimation of MVC-benchmark from Input/Output Data 181
10.3.1 Closed-loop Subspace Expression of Process Response
under Feedback Control 181
10.3.2 Estimation of MVC-benchmark Directly from
Input/Output Data 183
10.4 Simulations and Application Example 190
10.5 Summary 193
 Contents XVII

11 Prediction Error Approach to Feedback Control

Performance Assessment 195
11.1 Introduction 195
11.2 Prediction Error Approach to Feedback Control Performance
Assessment 196
11.3 Subspace Algorithm for Multi-step Optimal Prediction Errors .. 201
11.3.1 Preliminary 201
11.3.2 Calculation of Multi-step Optimal Prediction Errors . . . . 202
11.3.3 Case Study 206
11.4 Summary 211

12 Performance Assessment with LQG-benchmark from

Closed-loop Data 213
12.1 Introduction 213
12.2 Obtaining LQG-benchmark from Feedback Closed-loop D a t a . . . 214
12.3 Obtaining LQG-benchmark with Measured Disturbances 217
12.4 Controller Performance Analysis 219
12.4.1 Case 1: Feedback Controller Acting on the Process
with Unmeasured Disturbances 219
12.4.2 Case 2: Feedforward Plus Feedback Controller Acting
on the Process 220
12.4.3 Case 3: Feedback Controller Acting on the Process
with Measured Disturbances 221
12.5 Summary of the Subspace Approach to the Calculation of
LQG-benchmark 222
12.6 Simulations 223
12.7 Application on a Pilot-scale Process 224
12.8 Summary 227

References 229

Index 237