电力系统调度决策：一种示教学习辅助加速的安全强化学习方法

With the growing penetration of renewable energy generation and the increasing complexity of the power system environment, it is necessary to formulate a hybrid knowledge-data-driven dispatch strategy for modern power systems. In light of this, an imitation learning approach is proposed to exploit the expert knowledge, which provides demonstrations for power system economic dispatch strategy using neural networks. Furthermore, an accelerated safe reinforcement learning approach is proposed by incorporating learning from demonstration, which can make fast decisions in real-time operation. By incorporating the learning from demonstration approach, the algorithm convergence speed is significantly accelerated, the dispatch strategy is optimized, the operation cost is reduced, and the power flow violation risk is alleviated. Numerical simulation results verify the advantages of the proposed approach in improving the convergence efficiency of the reinforcement learning algorithm and promoting the security and economy of the power systems.