Error tolerant predictive control based on recurrent neural models

Abstract

This chapter is focused on developing a feasible model predictive control (MPC) based on time dependent recurrent neural network (NN) models. A modification of the classical regression neural models is proposed suitable for prediction purposes. In order to reduce the computational complexity and to improve the prediction ability of the neural model, optimization of the NN structure (lag space selection, number of hidden nodes), pruning techniques and identification strategies are discussed. Furthermore a computationally efficient modification of the general nonlinear MPC is proposed termed Error Tolerant MPC (ETMPC). The NN model is imbedded into the structure of the ETMPC and extensively tested on a dynamic simulator of an industrial crystalizer. The results demonstrate that the NN-based ETMPC relaxes the computational burden without losing closed loop performance and can complement other solutions for feasible industrial real time control.