An Identification Approach for Industrial Nonlinear Model Predictive Control

Abstract

This paper presents an empirical modeling approach for nonlinear model predictive control. Industrial applications require feasible, reliable, accurate and efficient identification methods to obtain nonlinear process models for MPC control. This paper discusses several key issues for industrial MPC application. These issues include neural networks' modeling capability and shortcoming, model structure selection and long-term prediction for MPC control. An identification scheme and algorithm are introduced. The identification starts with a linear state-space model and uses PLS and internally balanced realization algorithm to determine model order and identify a set of initial state variables, which is similar and related to the subspace identification methods. Then the algorithm identifies a hybrid linear-neural network model using PLS. This approach addresses the robustness of the identification and resultants in relatively simple, but sufficiently accurate model for industrial nonlinear MPC. Two typical nonlinear examples including a pH neutralization process and a polymer reactor are presented to demonstrate the features of the approach.