MPC relevant identification method for Hammerstein and Wiener models

Abstract

This work focuses on obtaining models that may produce a better performance of Model Predictive Controllers - MPC. Several papers published in the last 25 years have proposed methods based on the minimization of multi-step ahead prediction functions. These methods have been called MPC Relevant Identification (MRI). Most of the papers focused on obtaining linear models. In the last 5 years, some methods have been proposed to obtain nonlinear models based on the minimization of the same cost function. These papers were based on the direct minimization of the nonlinear cost function to produce models with NARMAX structure. However, simplified MPC schemes may be obtained using models with Wiener and Hammerstein structures. This paper presents some new theoretical results which allow the development of MRI identification algorithms for models with Wiener and Hammerstein structures, without the need to perform the minimization of the nonlinear cost function. The statistical properties of the models identified by the algorithm proposed are evaluated using Monte Carlo simulations and the prediction capability of the algorithm is evaluated using didactic plants. Results reveal a good multi-step ahead prediction capability.