Generalised predictive control for non-linear process systems based on lazy learning

Abstract

This paper presents an approach to model and control a discrete-time non-linear dynamical system on the basis of a finite amount of input/output observations. In order to improve the accuracy of local modelling based on lazy learning, a similarity criterion together with distance measure and angle measure is advanced. When the system works in the steady state or varies slowly, an adaptive tuning algorithm is used to alleviate the modelling computational burden. Also, this model is cast into a Generalised Predictive Control (GPC) framework. As a result of the integrated methodology, the cost, time and effort required to develop a non-linear generalised predictive control strategy is greatly reduced. The advantages of the methodology are demonstrated on non-linear function (a non-minimum phase system) and Continuous Stirred Tank Reactor (CSTR) simulations.