ABSTRACT
Model predictive control techniques such as generalized predictive control (GPC) (Clarke et al 1987) and dynamic matrix control (DMC) (Rovlak and Corlis 1990) have proven successful when applied to the control of industrial processes. It has been demonstrated that such linear predictive control techniques can be improved by including nonlinear system models (Morningred et al 1990). In particular, both GPC (Montague et al 1991) and DMC (Hernandez and Arkun 1990) have been extended by utilizing a nonlinear neural predictive model of the process. This industrial case study focuses on the application of neural modeling to improve the control of a polymerization reactor. The industrial system is introduced, highlighting the problems of accurate polymer viscosity control, based on a delayed measurement. This work presents a nonlinear predictor developed around the multilayer perceptron that can be used to remove this measurement delay. Finally, a platform is proposed to allow for the on-line implementation of neural-network-based predictive controllers.
