Abstract:In the context of high complexity, rapid response requirements, and signal fluctuations inherent in power grid dispatching, traditional dispatching models based on expert experience often fail to meet real-time control demands. This paper proposes a system dispatch optimization method that integrates a simulation operation mechanism, specifically designed for power system dispatch scenarios involving sudden disconnection faults. During the imitation learning phase, the simulation operation mechanism is combined with a multi-level action space reduction model to compress the action search space, thereby enhancing the learning efficiency of the baseline dispatching strategy using simulated expert experience data. In the reinforcement learning phase, the simulation mechanism is further employed to prioritize disconnection and reconnection actions, improving decision-making accuracy. Additionally, a differentiated reward function is designed to refine policy optimization. Simulation experiments involving disconnection faults and ablation studies demonstrate that, compared to the DDQN algorithm and the imitation learning model, the proposed method increases the comprehensive dispatch success rate by 50% and 100%, respectively, reduces dispatch action costs by 50% and 60%, respectively, and extends the stable operation time of the power grid by factors of 1.85 and 6.5, respectively.