Sub-synchronous oscillation suppression strategy in wind power systems with DFIGs based on improved deep reinforcement learning algorithms

The characteristics of sub-synchronous oscillations (SSOs) in Doubly-Fed Induction Generator (DFIG)-based wind farms connected to series-compensated grids are influenced by various factors such as operating conditions, control parameters, and external environments. The complex and variable interaction between the wind farm and the grid leads to diverse oscillation patterns in interconnected systems. This study aims to develop an adaptive sub-synchronous damping controller (SSDC) that enhances system damping and operational stability under dynamic real-world conditions. First, a joint distribution adaptive transfer learning method (JDA-TFL) is employed to locate oscillation sources under multiple operating conditions. The optimal location of the SSDC at the oscillation source significantly reduces control costs. Second, the SSDC design problem is redefined as a Markov Decision Process (MDP), with the sensitivity of measured electrical parameters to system stability being embedded in the reward function construction. Third, the twin delayed deep deterministic policy gradients algorithm incorporating an exploration network (EMTD3) is utilized to generate optimal control signals. The encouragement mechanism improves the model's ability to explore the optimal strategy. Additionally, a multi-experience probability replay strategy is employed to accelerate training convergence. Finally, the effectiveness and robustness of the proposed method are validated through simulations across multiple scenarios. The proposed method offers a novel solution for mitigating SSOs in wind power systems, with significant theoretical and practical implications for enhancing the reliability of these systems.