Intelligent Wind Farm Control via Grouping-Based Reinforcement Learning

This paper aims to maximize the total power generation for wind farms subject to strong wake effects and stochastic inflow wind speeds. A data-driven control method that only requires the accessible measurements of every turbine in the farm is proposed via deep reinforcement learning (DRL). We employ a grouping strategy to mitigate the high computational complexity induced by DRL and enhance our method's applicability to large-scale wind farms. Based on the levels of aerodynamic interactions among turbines, this grouping strategy divides the whole farm into small sub-groups. Therefore, one can execute DRL on these sub-groups instead of carrying on a complicated learning process for the entire farm. Simulations verify the advantages of the proposed DRL-based wind farm control method over the commonly employed greedy strategy. Results also show that the proposed method can significantly reduce the overall computing cost compared with the direct execution of DRL on the whole wind farm.