Graph-Based Topological Embedding and Deep Reinforcement Learning for Autonomous Voltage Control in Power System

With increasing power system complexity and distributed energy penetration, traditional voltage control methods struggle with dynamic changes and complex conditions. While existing deep reinforcement learning (DRL) methods have advanced grid control, challenges persist in leveraging topological features and ensuring computational efficiency. To address these issues, this paper proposes a DRL method combining Graph Convolutional Networks (GCNs) and soft actor-critic (SAC) for voltage control through load shedding. The method uses GCNs to extract higher-order topological features of the power grid, enhancing the state representation capability, while the SAC optimizes the load shedding strategy in continuous action space, dynamically adjusting the control scheme to balance load shedding costs and voltage stability. Results from the simulation of the IEEE 39-bus system indicate that the proposed method significantly reduces the amount of load shedding, improves voltage recovery levels, and demonstrates strong control performance and robustness when dealing with complex disturbances and topological changes. This study provides an innovative solution to voltage control problems in smart grids.