Enhancement of Distribution Network Resilience: A Multi-Buffer Invalid-Action-Mask Double Q-Network Approach for Distribution Network Restoration

The increasing occurrence of extreme weather events presents significant challenges to the resilience of distribution networks. Distribution network restoration (DSR), typically solved by using optimization-based methods, suffers from drawbacks such as reliance on accurate models and slow solution convergence, leading to low scalability. To address these issues, this paper proposes the multi-buffer invalid-action-mask double Q-network (MI-DDQN) approach for DSR. Firstly, the DSR problem is formulated as a Markov decision process. Next, the efficiency of experience replay is enhanced through the utilization of multi-buffer and prioritized experience replay. Additionally, an invalid-action mask is employed to mask out invalid actions, thereby improving model convergence speed. The effectiveness of the proposed method is validated by using the IEEE 37-node system.