Practical nonlinear predictive control algorithms for neural Wiener models

This paper describes three nonlinear Model Predictive Control (MPC) algorithms for neural Wiener models. In all algorithms the model or the output trajectory is linearised on-line and used for prediction. In the first case model linearisation is performed in a simplified manner for the current operating of the process. In the second algorithm the predicted output trajectory is linearised along an assumed future input trajectory once at each sampling instant whereas in the third approach trajectory linearisation is carried out in an iterative way. As a result of linearisation, the future control policy is easily calculated from a quadratic programming problem or from a series of such problems. Good control accuracy and computational efficiency of described algorithms are demonstrated for two nonlinear processes: a polymerisation reactor and a neutralisation reactor are considered. Unlike many control structures for Wiener models, discussed algorithms do not need an inverse of the steady-state part of the model.

► Model Predictive Control algorithms for neural Wiener models are described.
► The model or the output trajectory is linearised on-line and used for prediction.
► The algorithms do not need the inverse steady-state part of the model.
► The algorithms are computationally efficient: quadratic programming is used.
► Algorithms’ advantages are shown in the control systems of two chemical reactors.