GLOBAL OPTIMIZATION IN MULTIVARIABLE ROBUST PREDICTIVE CONTROLLER OF NONLINEAR SYSTEMS BASED ON NEURAL NETWORKS

摘要

This paper proposes a constrained robust model predictive control for discrete-time multivariable nonlinear systems. The neural networks approach is used to model the unknown dynamic behavior of the system, where the outputs layer weights are affected by bounded uncertainties. When structured uncertainties are explicitly used, the control problem is nonconvex and it can be formulated as a constrained minimax optimization one subject to the uncertain parameters of the neural model and the control signal constraints. Classic optimization method used to resolve this kind of problem can only lead to a sub-optimal solution. In this work, the Generalized Geometric Programming technique is iteratively used to reduce the constrained nonconvex optimization problem into a convex one in order to compute the optimal control actions. Simulation results are provided to demonstrate the effectiveness of the proposed neural predictive controller.