Nonlinear dynamic matrix control using local models

This paper proposes the concept of using a local model network (LMN) to identify a highly nonlinear chemical process, and to implement a dynamic matrix controller (DMC) that uses the local model network as its internal model. The LMN is constructed of local linear autoregressive with external input (ARX) models, and is trained using a hybrid learning approach developed by McLoone et al. (1998). It is shown how this LMN structure is linked to a long range predictive controller, specifically dynamic matrix control. Originally, a linear step response model was used as the internal model of the controller, however, to extend to the control of a highly nonlinear process, step responses for different operating points are extracted from the LMN. Simulation results for the method, when applied to a pH neutralization process, indicate an improvement in control over a standard DMC controller.