Stable neural network based model predictive control

The paper presents nonlinear model predictive control designed using recurrent neural network. A recurrent neural network is trained to act as the one-step ahead predictor, which is then used succesively to obtain k-step ahead prediction of the plant output. Based on the neural predictor, the control law is derived solving a constrained optimization problem. The stability of the considered predictive scheme is also investigated showing that a cost function is monotonically decreasing with respect to time. Derived stability conditions are used to redefine a constrained optimization in order to calculate a control, which guarantees the stable work of the control system. The quality of the proposed stable predictive scheme is tested using a tank unit simulator realized in MATLAB/Simulink environment.