Wind farm control under time-varying wind using deep reinforcement learning

A control method to maximize the power generation of a wind farm in time-varying wind is developed using deep reinforcement learning (DRL). Although DRL-based control methods have shown promising performance in enhancing the power generation of wind farms, conventional DRL-based methods assume statistically steady wind conditions (Dong et al, 2021, Appl. Energy), limiting their application to practical wind farms. The present method utilizes time-histories of the past and predicted future wind to consider complex wake interactions in an unsteady manner. A multi-fan wind tunnel has been employed to produce statistically unsteady wind for a miniaturized wind farm. The training is then performed to maximize power generation by determining the optimal pitch and yaw angles of each wind turbine in the wind farm. The present method demonstrates the feasibility of a DRL-based control method in stochastic wind conditions, including real wind scenarios, and shows a significant improvement in power generation compared to the conventional greedy control method. The wake velocity induced by upstream turbines is measured using a particle image velocimetry for analyzing the improvement in power generation by the present method.