A neural predictive control scheme for nonlinear plants

Since the last three decades predictive control has shown to be successful in control industry, but its ability to deal with nonlinear plants is still under research. Generalized Predictive Control (GPC) was one of the most famous linear predictive algorithms. The control law of GPC contains two parameters that describe the system dynamics: system free response ( f ) and system impulse response matrix (G ). Often these parameters are calculated from the discrete linear model. For nonlinear systems, either a nonlinear system model is instantaneously linearized or a nonlinear optimization is used. The validity of the linear model is the shortcoming of the first one and the possibility of non-uniqueness of local minimum is that for the second. In this paper, a neural network (NN) model is used as a predictor to calculate these parameters for GPC. The nonlinear system free response is obtained instantaneously while dynamic response is linearized every batch of time. The method is tested on a benchmark nonlinear model. Results are compared with that of others neural predictive techniques found in previous literature. Also, the method is applied and validated on a realistic multivariable aircraft model. The simulation results show that this method has some good advantages over others neural predictive techniques. In one hand, the system dynamics parameters are calculated more accurately directly from the nonlinear NN model. And in the other hand, the used linear GPC has a cost function with only one global minimum. The method, as a trade-off between nonlinear neural predictive control (NPC)and instantaneous linearization approximate neural linear predictive control (APC), is promising for control of nonlinear systems