基于图注意网络与深度确定性策略梯度的三相主动配电网供电恢复方法

The distribution service restoration as a fundamental resilient paradigm provides an optimally coordinated resilient solution to improve resilience through power restoration of the distribution system after extreme events. According to the property of the three-phase unbalanced distribution system, a graph attention network is utilized to reform the Actor-Critic-based deep reinforcement learning, and the agent learning ability in the unbalanced distribution system is improved by increasing the features of the network topology in this paper. A new method for power restoration of the three-phase unbalanced active distribution system based on graph deep reinforcement learning is proposed. In this method, the dynamic power restoration problem is designed as a novel Markov decision-making process, in which sample data is continuously generated and the agent is trained according to the proposed graph deep reinforcement learning algorithm. The agent realizes dynamic load restoration of the distribution system by optimally coordinating multiple microgrids. Its performance has been verified on IEEE 37-Bus and IEEE 123-Bus distribution systems.