RBF neural networks for modelling and predictive control: An application to a neutralisation process

Abstract

This paper describes a Model Predictive Control (MPC) algorithm in which a Radial Basis Function (RBF) neural network is used as a dynamic model of the controlled process and it reports training and selection of the RBF model of the benchmark system for MPC. In order to obtain a computationally uncomplicated control scheme, the RBF model is successively linearised on-line, which leads to an easy to solve quadratic optimisation problem, nonlinear optimisation is not necessary. Efficacy of the MPC algorithm is shown for a neutralisation system, which is a significantly nonlinear dynamic process. It is shown that the described MPC algorithm with on-line model linearisation gives trajectories very similar to those obtained in a truly nonlinear MPC scheme, in which the full nonlinear RBF model is used for prediction.