Neural-net based multi-steps nonlinear adaptive model predictive controller design

Concerns nonlinear model predictive control, and particularly the nonlinear optimization problem. Usually the control sequence can be determined by using some effective numerical iteration approaches, especially for multistep predictive control. This work focuses on the multistep adaptive NMPC controller design using neural-net. The main ideas are (A) initialisation of the multistep control laws by using one-step ahead predictive control law; (B) linearization of the neural-net predictor at every operating point; and (C) tuning of the neural-net predictor through online learning using teacher signals generated by closed-loop system input-output data. As an illustrative example of our approach, an explicit control laws are derived for the control horizon Nu=2 case.