Training of neural models for predictive control

Abstract

This paper emphasises the link between neural model training and its role in model predictive control (MPC) algorithms. This role is of fundamental importance since in MPC at each sampling instant a model is used on-line to calculate predictions of future behaviour of the process and an optimal future control policy. Taking into account this particular function of models in MPC, a training algorithm of neural dynamic models is derived. An example identification problem of a methanol–water distillation process is discussed. The prediction accuracy of models obtained using the described algorithm and the classical backpropagation scheme is compared, which yields one-step ahead predictors.