Fast algorithm for adaptive generalized predictive control based on BP neural networks

Abstract

The traditional way of reducing the on-line computing time required for the standard adaptive generalized predictive control (GPC), is by using a short predictive horizon or a short control horizon. However, it breaches the intrinsic principle of this long range receding horizon strategy, and sometimes it could lead to poor control performance in some processes. A kind of fast algorithm for self-tuning GPC (FGPC) is presented by using back propagation (BP) neural networks. The algorithm involves the derivation of the GPC in a new way and the training of BP neural networks to represent nonlinear relations between the GPC controller coefficients and system open-loop parameters, control horizon, smoothing factor and control weighting factor. The method developed does not involve the Diophantine equation and avoids the matrix operation, therefore substantially alleviates the mathematical computation problems associated with the standard adaptive GPC, especially for a long control horizon. The efficacy of this algorithm is demonstrated with a comparative simulation study.