| The performance of Model Predictive Control (MPC) is strongly influenced by the quality of the reference model which is used in the MPC algorithm to predict the system outputs into the future. To date, most applications of MPC are based on linear models which cannot be expected to describe very nonlinear processes. Nonlinear MPC can readily be formulated, but nonlinear process models are often more complex and require considerable effort to develop. The development of neural networks in recent years offers the possibility of powerful and flexible modeling of a wide range of nonlinear systems with fast computation speed. By using the strength of neural network modeling techniques, an advanced and general model predictive control can be formed.; The Radial Basis Function Neural Network (RBFN), a type of feedforward neural network, is a particularly attractive form of neural network for use in MPC applications because it offers the possibility of rapid retraining, facilitating adaptation to changing process behavior. The usual RBFN model is a one-step ahead predictor. In MPC, multi-step ahead predictions are needed. Although the RBFN model can be iterated to get multi-step ahead prediction, errors may accumulate during the successive iterations. Therefore, a multi-step Time-Lag Recurrent Radial Basis Function Neural Network (TLRRBFN) was used. A new and efficient training algorithm was developed to train the TLRRBFN which facilitates on-line adaptation of the weights of the TLRRBFN model.; To implement the Neural Network Based Model Predictive Control (NNMPC), the TLRRBFN model is used as a reference model. In this work, the NNMPC was evaluated on three simulated test systems, including a multi-input-multi-output (MIMO) Continuous Stirred Tank Reactor (CSTR) (Li and Biegler, 1988). In each case, for set point change problem, the NNMPC algorithm shows good tracking performance without offset. The NNMPC algorithm also shows good disturbance rejection ability. It was also found that including a filter in the feedback path was useful in improving the stability of the NNMPC algorithm. For all three nonlinear processes studied, the NNMPC outperformed conventional PID controller. |
