Neural network approach for optimization of batch polymerization processes

This dissertation focuses on the development and implementation of on-line strategies for end-point optimization of batch polymerization reactors. The traditional approach to end-point optimization requires the numerical solution of a two-point-boundary-value problem which is computationally expensive and hence difficult to implement on-line.; In this dissertation a novel on-line strategy based on neural networks is developed which calculates the optimal operating policy based on initial conditions of the batch. This novel strategy accounts batch to batch variations and computes the optimal state trajectories very quickly (in milliseconds). For this reason, this strategy is suitable for on-line implementation. To track the optimal state trajectory calculated by the neural network, a sliding-mode robust controller is developed. This nonlinear controller keeps the system on the desired optimal trajectory despite uncertainty in model parameters.; These on-line optimization and control strategies are tested via numerical simulations on two different batch polymerization processes. In order to test the on-line strategy experimentally an experimental batch polymerization system involving a three liter batch reactor is constructed. This fully instrumented experimental facility equipped with a plug-in data acquisition board has the capability to do both data acquisition and control in real time. Experiments are conducted to validate the on-line optimization strategy using au industrially important batch polymerization process--Methyl Methacrylate Polymerization.