Deep reinforcement learning-driven wind farm flow control considering dynamic wind

Mitigating power losses caused by the wake effect is crucial for improving the efficiency of operational wind farms. Wind farm flow control represents a key approach to achieving this objective. However, dynamic wind conditions, including variations in wind speed and direction, along with environmental uncertainties, present significant challenges to effective flow control. To address these challenges, this paper proposes a wind farm flow control method via deep reinforcement learning that considers dynamic wind. Initially, dynamic wind fluctuation characteristics are extracted from LiDAR-measured data, which provides a comprehensive dataset. Subsequently, a flow control method is developed, using dynamic wind as input to maximize wind farm power output through yaw angle adjustments. Finally, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is introduced to drive a control model for real-time optimization and online learning. The model is capable of addressing uncertainties through experience replay and exploration mechanisms. Simulation results demonstrate that optimization considering mean wind is effective only when the wind direction standard deviation is below 4, whereas optimization considering dynamic wind is effective across all wind conditions. Considering dynamic wind results in a 3.3% improvement in power generation compared to optimization considering mean wind.