Efficient power system corrective control with adaptive operation condition handling: A GATN-RL based method

Traditional corrective control algorithms often struggle to strike a balance between computational accuracy and speed. While deep reinforcement learning (DRL) has emerged as a promising alternative, conventional DRL methods cannot capture the complex topological structures and temporal dynamics inherent to power systems. This paper proposes a corrective control algorithm based on a Graph Attention Temporal Network with Reinforcement Learning (GATN-RL) to address these challenges. The method employs a topology embedding mechanism that encodes connectivity directly into node features, enhancing the model’s perception of the grid’s structure. Specifically, Graph Attention Networks (GATs) are employed to aggregate data from adjacent nodes and capture spatial dependencies. A Transformer architecture is then used to identify temporal dynamics from sequential power system observations. Case studies on the 36-node and 118-node test systems validate that the proposed method delivers fast and effective corrective control. Compared to conventional reinforcement learning methods, GATN-RL improves the success rate by 2.7 % to 25 % in the 36-node system and by 9.83 % to 69.5 % in the 118-node system, while requiring less than 0.2 % of the computation time of traditional corrective control algorithms. Furthermore, the algorithm demonstrates strong generalization capabilities across diverse network topologies and under varying levels of renewable energy integration.