Non-linear model based predictive control through dynamic non-linear partial least squares

The extension of model predictive control (MPC) to non-linear systems is proposed through dynamic non-linear Partial Least Squares (PLS) models. PLS has been shown to be an appropriate multivariate regression methodology for modelling noisy, correlated and/or collinear data. It has been applied extensively, within a 'static' framework, for the modelling and analysis of industrial process data. The contribution of this paper is the development of a non-linear dynamic PLS framework for applications in MPC. The non-linear dynamic PLS models make use of an error based non-linear partial least squares algorithm where the non-linear inner models are built within an AutoRegressive with eXogeneous inputs (ARX) framework. In particular, quadratic and feedforward neural network inner models are considered. The application of a dynamic PLS model within a MPC framework opens up the potential of using multivariate statistical projection based methods not only for process modelling, inferential estimation and performance monitoring, but also for model predictive control. A benchmark simulation of a pH neutralization system is used to demonstrate the application of a non-linear dynamic PLS framework for model predictive control.