Suboptimal Non-linear Predictive Control Based on MLP and RBF Neural Models with Measured Disturbance Compensation

This paper is concerned with a computationally efficient (suboptimal) non-linear Model Predictive Control (MPC) algorithm based on two types of neural models: Multilayer Perceptron (MLP) and Radial Basis Function (RBF) structures. The model takes into account not only controlled but also the uncontrolled input of the process, i.e. the measured disturbance. The algorithm is computationally efficient, because it results in a quadratic programming problem, which can be effectively solved on-line by means of a numerically reliable software subroutine. Moreover, the algorithm gives good closed-loop control performance, comparable to that obtained in the fully-fledged non-linear MPC technique, which hinges on non-linear, usually non-convex optimisation.