Hierarchical clustering-based deep reinforcement learning for adaptive frequency control of wind farm

As the inertia level of the power system declines, wind farms (WFs) are increasingly required to provide inertial support during frequency events. However, most existing adaptive frequency support methods do not simultaneously consider the frequency response requirements and the turbine rotor speed dynamics. This may cause the turbine to release excessive energy, affecting its operational recovery. This paper proposes an adaptive frequency regulation (FR) control method for WFs based on the deep reinforcement learning (DRL) Soft Actor-Critic (SAC) algorithm. Using the high exploration capability of the SAC algorithm, the DRL agent dynamically optimizes FR control gains by maximizing the reward function. A dynamic hierarchical clustering method based on real-time wind speed is established to group and control turbines, thereby improving the consistency of frequency responses and optimizing agent configuration. Simulation results show that the proposed method demonstrates superior exploration capability compared to traditional algorithms during training and provides just enough FR energy to keep the frequency within the allowable deviation range. Additionally, the rotor speed can be actively restored, ensuring system stability. The simulations further validate the generalization capability of the proposed method, demonstrating its effectiveness under various operating conditions.