Factorized f-step radial basis function model for model predictive control

This paper proposes a new factorized f-step radial basis function network (FS-RBF) model for model predictive control (MPC). The strategy is to develop a f-step predictor for nonlinear dynamic systems and implement it with a RBF network. In contrast to the popular NARX-RBF model, the developed FS-RBF model is capable of making a designated sequence of future output prediction without requiring the unknown future process measurements. Furthermore, the developed FS-RBF model is factorized into two parts, with one part including past plant input/output and the other part including the future input/output. When this model is used as the internal model in the MPC, the factorization enables an explicit objective function for the on-line optimization in the MPC. Thus, the computing load in solving the optimization problem is greatly reduced. The developed model is used in MPC and applied to a continuous-stirred tank reactor (CSTR). The simulation results are compared with that of MPCs with other two models. The comparison confirms that the developed model make more accurate prediction so that the MPC performance is better, it also uses much less computing time than the other two models based MPC.