This paper presents a LiDAR-informed adaptive-cost nonlinear model predictive control (NMPC) strategy for wind turbine pitch regulation. The proposed method uses a reinforcement learning (RL) agent as a supervisory cost-shaping module that adjusts the weights in the NMPC cost function. The pitch command is obtained from the constrained NMPC optimizer, which preserves the physical prediction model, actuator limits, and receding-horizon solution structure. LiDAR-derived preview wind-speed information is used as an estimate of the incoming disturbance and is introduced into both the prediction model and the agent state. This design helps the controller account for wind-field variation over the prediction horizon and adapt the relative emphasis on power regulation, load mitigation, and pitch-action smoothness. Compared with feedforward PID (FF-PID) and fixed-weight feedforward NMPC (FF-NMPC) controllers, the proposed controller shows stronger adaptability under abrupt and stochastic wind variations in OpenFAST-MATLAB/Simulink co-simulations.
Jin et al. (Fri,) studied this question.