Neural network model predictive control of a ultra high temperature milk treatment plant : a thesis presented in partial fulfilment of the requirements for the degree of Master of Technology in Engineering and Automation at Massey University
This thesis reports the development of a Model Predictive Control system for a Ultra High Temperature milk treatment pilot plant. This control system utilises an Artificial Neural Network model of the plant dynamics. The entire process was divided into two parts for modelling purposes. Separate models were trained; one for simulating the dynamics of the hot water heating loop and the second the dynamics of the heat exchanger circuit. The two sub-models, when concatenated, form a complete model of the plant referred to as a composite neural network model. The results of training and testing of the sub-models with various sets of plant data were presented. Of all the possible combination of sub-models, the best trained and tested sub-models were concatenated to form the best composite network model, and the combination of worst sub-models to form worst composite network model. Two model predictive control (MPC) systems for the process were developed, one using the best composite network model for prediction purposes and to act as the plant, and the other using the worst composite model for prediction and best composite model as the plant. Both the developed MPC systems were evaluated in terms of setpoint tracking and disturbance rejection. As a part of these performance tests, a PI (Proportional-Integral) control system of the UHT plant was developed in a simulated environment using the best composite neural network model to act as the plant. The responses of both the MPC control systems were studied and compared with the responses of the PI control system.