Adversarial Scenario Generation Integrated Distributionally Robust Deep Reinforcement Learning for Survival of Critical Loads

Topology reconfiguration (TR) is an efficient resilient defense action to ensure the survival of critical loads (CLs) during sequential extreme events (SEEs). While deep reinforcement learning (DRL)-based TR approaches have been widely studied recently, most existing works ignored the interaction between scenario and training performance, which will risk policy robustness against environmental changes and potentially endangers CLs during events. This paper proposes a distributionally robust DRL (DR-DRL)-based TR method to address these challenges. Specifically, TR scheduling is formulated as a distributionally robust sequential decision-making problem to determine actions of remote control switch (RCS) for optimizing CL survival and microgrid operation under the worst-case scenario distribution. To solve the formulated problem, the DR-DRL implementation strategy is proposed including DRL for RCS switching policy learning and a novel adversarial scenario generative adversarial network (ASGAN) for dataset enhancement with worst-case environmental scenarios. A robust TR policy is obtained through a customized training algorithm that alternates between DRL and AS-GAN training. Finally, the effectiveness and efficiency of the proposed method are validated by a 7-bus system and the IEEE 123-bus system, and the results demonstrate superior robustness and adaptability for CL survival under dynamic SEEs with uncertainties.