Linear and nonlinear model predictive control

A linear model predictive controller and two nonlinear model predictive controllers are investigated in this work. These control strategies differ in the modeling techniques used. The first, NMPC-FF, uses a feedforward neural network to model the system whereas the second, NMPC-R, uses a recurrent neural network. In NMPC-FF, the predictions are made using different neural networks for each prediction. In NMPC-R, a single neural network is recursively used to get all future predictions. Both controllers are developed based on the principles of linear model predictive control. Maximum move size limitation is used in both nonlinear controllers instead of move suppression. This provides the necessary dampening of the manipulated variables.; The feedforward neural network is trained using classical Error Back Propagation technique. NeuralWare®, a commercial software for training neural networks is used for training the feedforward network. The recurrent neural network is trained using Random Optimization Method. Since commercial software is not available, the training routine for recurrent neural network has been developed as part of this study. The error function that is minimized during training has also been modified to account for the prediction horizon.; The linear model predictive controller, ONLINE®, is implemented on a binary distillation column. The linear controller is implemented without the inclusion of ambient temperature as a measured disturbance variable. The need for the addition of ambient temperature as a measured disturbance is demonstrated and the controller is implemented again with the addition of ambient temperature. These results are a baseline for the results with nonlinear controllers.; The nonlinear controllers are tested on the Dual Process Simulator, which approximates a nonlinear pH process, and on a binary distillation column. For the Dual Process Simulator, NMPC-FF is found to be better for regulatory control and NMPC-R is found to be very good for servo control. NMPC-R works well in all ranges of servo control and fails for one situation in regulatory control.; The results for nonlinear controller implementation, in simulation mode, on the distillation column indicate there are discrepancies in the models. Even though the testing set from the neural network training gives excellent results, when the controller is implemented, the results are not as expected.; Only the NMPC-FF controller is implemented on the actual distillation column. The model weaknesses do not allow for good controller performance on the distillation column. Model inadequacy is expected to hamper the performance of NMPC-R controller and therefore this controller is not implemented on the distillation column. A way to fine-tune the existing models needs to be developed before either nonlinear controller can give good results on the distillation column.