：The secondary control problem of post-disaster island microgrids based on reinforcement learning was studied. Firstly, for a typical microgrid topology, two types of challenging disaster scenarios were characterized: the first type was the local paralysis of lines, with severe imbalance of power flow and voltage in the circuit. The controller must complete power redistribution and topological elastic recovery within milliseconds; otherwise, local overload or voltage collapse will cause cascading disconnection. The second type was the partial failure of distributed power sources, with frequency-voltage coupling oscillations amplifying within seconds. The controller needs to instantly coordinate the power margin of the remaining power sources; any delay will lead to the entire network"s loss of synchronization and paralysis. Clearly, in the above situations, traditional control strategies based on precise models are no longer applicable. A post-disaster island microgrid intelligent recovery controller using double-delay deep deterministic policy gradient (TD3) was proposed. Through offline training of disaster scenarios, the controller measures the electrical parameters of the system in real time during the online stage and controls the output power of storage elements to ensure the stability of voltage and frequency. Finally, the effectiveness and advancement of the proposed controller were verified by comparing it with traditional control strategies such as PID, fuzzy control, and the reinforcement learning deep deterministic policy gradient (DDPG) algorithm.