A comparison of nonlinear predictive control techniques using neural network models

In this paper a comparison between two approximations of the general constrained nonlinear optimization problem is made. The proposed techniques are tested on a highly nonlinear process modeled with an affine combination of multilayer feedforward neural networks. Such network structures are suitable to be further integrated into feedback linearization schemes providing, under some mild assumptions, an easy way to feedback linearize a nonlinear process. Simulation results have revealed a superior closed-loop performance when comparing this technique with the classical linearization through Taylor's expansion of the expanded nonlinear prediction model.