Abstract: To solve the problem that the effectiveness and adaptability of heuristic algorithms in distribution network fault recovery decision become poor when the topology of distribution network changes,a fault recovery decision method based on graph reinforcement learning is proposed.First,the graph data is used to characterize the decision information in fault recovery,including distribution network topology and electrical characteristics.Then,the pre-graph neural network is set up in the graph reinforcement learning model to receive the graph data input and cope with the topological changes of the distribution network in the process of fault recovery.Finally,the reinforcement learning agent of the embedded graph neural network outputs the final fault recovery strategy to improve the decision speed.A modified PG＆E 69-bus distribution network case is used for validation.The results show that the proposed algorithm can reach millisecond level in solving speed,and improve the solving efficiency by 6%～7% compared with heuristic and genetic algorithms,Also,the load recovery rate of the fault recovery strategy is higher.