Structured neural networks for constrained model predictive control (Q5943842)

The structured neural networks (SNN) and corresponding training algorithms for an exact solution of a quadratic programming (QP) problem associated with constrained model predictive control (MPC) are developed. At first, the constrained MPC problem is exposed as a standard QP problem, which can be used to calculate time-varying control actions on-line every time instant. Then a projection neural network (PNN) and the special training algorithms that guarantee the convergence to an optimal solution are constructed. The PNN generates the projection of a state space point to the linear inequality constraint set. Finally, based on the PNN, the SNN is developed, which (by exact implementation of the known gradient projection algorithm) solves the main QP problem. It is stressed that the proposed SNN has a parallel structure well-suited for hardware implementation. The SSN is tested on the simplified model of a cross-direction basis weight and moisture control system of an industrial paper-making machine.