Structured neural networks for constrained model predictive control

High computational demand in solving the optimization problems associated with the model predictive control (MPC) schemes is a major obstacle when applying the methods to large-scale or fast-sampling systems. In this paper, we propose a new structured neural network approach to solving the quadratic programming problem in the constrained MPC. This new approach has the advantage of solving large-scale quadratic programming problems in a massively parallel fashion. The structured neural network consists of a projection network and a network for implementing the gradient projection algorithm. where the projection network is constructed from specially structured linear neurons with a special training algorithm. We prove that the training algorithm converges to the optimal solution. Finally, we test the method on a simplified paper machine model. (C) 2001 Elsevier Science Ltd. All rights reserved.