Nonlinear adaptive model predictive control based on self-correcting neural network models

Two major issues in process control. are the nonlinearities and variations with time that are observed in the dynamics of the processes. In most cases these problems are confronted by robust linear controllers, which are frequently retuned to take into account changes in the operating region or the system dynamics. Obviously, the performance of these controllers is limited by the degree of nonlinearities and the frequency of process variations. In this paper we present a new model predictive control (MPC)framework that can deal with both these issues. The proposed methodology is based on a nonlinear dynamic radial basis function (RBF) model of the process that is able to correct itself as new information about the process dynamics becomes available. The adaptive training algorithm that is used is able to update both the structure and the parameters of the RBF model. The typical formulation of the on-line optimization problem is augmented by a persistent excitation condition that guarantees that enough perturbation is introduced to the system by the control moves. The proposed MPC framework is applied on the digester control problem and proves to be superior to other MPC configurations. (c) 2005 American Institute of Chemical Engineers.