The main objectives of the investigations in this thesis are:
1. (a) To develop a predictive controller based on FIR models with 'ℓ2' regres-sion norm along with input and input-rate constraints with a simple estimator and to evaluate the performance of the controller related to the uncertainty of impulse response co-ecients.
(b) To develop a regularized l2 moving horizon estimator based on nite impulse response (FIR)models with input and input-rate constraints and to evaluate the closed loop performance of the above estimator with a predictive regulator.
2. (a) To develop a robust soft constraints based predictive controller with simple estimator for linear systems.
(b) To compare the performance of the constrained controller with nominal predictive controller by simulation.
3. (a) To implement the soft MPC in a real time cement mill circuit and compare the performance with that of the other controllers
(b) To evaluate by simulation the performance of soft MPC handling the large sample delay measurements
The organization of the thesis is as follows:
Chapter 2 provides the basic motivation of the work with a detailed literature survey on model predictive controllers with hard and soft constraints and control strategies for cement mill circuit.
Chapter 3 provides the discussion on Model Predictive Control Based on Finite Impulse Response Models with simple estimator. This chapter also presents the
5
details on deriving a Moving Horizon Estimation. Also this chapter presents the details on Model Predictive Control with Soft Output Constraints is provided.
Chapter 4 gives Comparison of Soft MPC with conventional MPC using simula-tion. First the controllers are compared with simple SISO system. Then a model of cement mill is considered for comparing the closed loop performance of the controllers using Matlab.
Chapter 5 gives the basics on Cement Manufacturing Process and Cement Milling circuit. Also the basic control strategy of cement mill circuit is discussed.
In Chapter 6 applications of Soft MPC to Cement Mill Circuit are discussed. A transfer function model of cement mill obtained and the controller is implemented in the simulator. The performance of the controller is then compared with con-ventional MPC in simulator. The soft MPC is then implemented in real plant and the performance of the controller is compared with already existing Fuzzy Logic controller.
Chapter 7 provides the detailed study on Implementation of MPC to a Large Sam-ple Delay System. Here the controller performance is investigated using the cement mill simulator rst with every minute sample and then with model including the sample delay.
Summary and Conclusions are given in Chapter 8.
Appendix A gives the formulation of Quadratic program for FIR based MPC, MHE and Soft Constraints based MPC.
Appendix B gives the ow chart for the MPC execution in MATLAB.
Appendix C gives the generalized form of deriving Quadratic program Appendix D provides the basic algorithm of Interior point methods.
Appendix E provides the Matlab codes for design of soft MPC and simulation in closed loop.
Appendix F gives a brief description of ECS/CEMulator system where the per-formance of the controllers are compared.
6
CHAPTER 2 Literature Survey
In this chapter, the published literature is reviewed on the model predictive con-trollers generally used in industries and the MPC with hard constraints. A brief review of soft constraint based MPC and controller for cement industries is also presented.
Excellent review on model predictive control is available. Reviews on stability of model predictive control are given by Mayne et al. (2000), Zheng (1998), Zheng and Morari (1995) and Limon et al. (2006) and a review on tuning methods have been provided by Garriga and Soroush (2010). Detailed survey reports on industrial applications of model predictive control are given by Bemporad and Morari (1999) and Morari and Lee (1999) and Qin and Badgwell (2003) and Bemporad and Morari (1999). Garcia et al. (1989) have discussed the basic theory on model predictive control .
2.1 Model Predictive Control in Industries
There are many control strategies in use today like intelligent control, adaptive control, stochastic control, optimal control etc. Optimal control is such a control technique in which we minimize certain cost index to achieve desired performance.
The two types of optimal control techniques are
• Linear Quadratic Gaussian (LQG)
• Model Predictive Control (MPC)
Model Predictive Control technique is the most widely used technique in industry as opposed to LQG based controllers. The LQG controllers were termed as failure and the reasons for this failure are given by Garcia et al. (1989) and Richalet
7
et al. (1976). They have provided the reasons that the LQG controllers are not successful because they cannot handle the following:
• constraints
• process nonlinearities
• model uncertainty (robustness)
• unique performance criteria
Further, MPC is classied into Linear MPC and Non-Linear MPC depending on the specic problem statement. Both linear and nonlinear systems have specic problem statements and utilize dierent optimization methods. Non- Linear MPC uses non-linear models for prediction and it requires iterative solution of optimal control problems on a nite prediction horizon. But non-linear MPC cannot be solved as convex optimization problem. Some of the work on non-linear MPCs are given by Miller et al. (2000) and Santos et al. (2008). They have provided a tool to analyze the stability of constrained non-linear model predictive control.
Linear MPCs are most commonly used techniques in industry because of compu-tational simplicity and faster solutions in solving real time optimization problems.
Further, linear MPC used in real time applications can be classied into following types
• Dynamic Matrix Control (DMC)
• IDCOM (Identication- Command)
• General Predictive Control (GPC)
• Moving Horizon Control (MHC)
These major classication of MPC is based on the type of algorithm used for solving optimization problem. While the MPC paradigm encompasses several dierent variants, each one with its own special features, all MPC systems rely on the idea of generating values for process inputs as solutions of an on-line (real-time) optimization problem.
8