A Rainbow DQN-Based Strategy for Post-Disaster Load Restoration in Distribution Networks

Low-probability, high-impact extreme events may result in distribution system failures, making the rapid and effective restoration of affected loads a critical challenge for enhancing the resilience of the distribution system. Conventional model-based load restoration methods are constrained by significant limitations, owing to dependence on precise model parameters and exponentially increasing computational complexity as system scale increases. To address these challenges, this paper proposes a novel load restoration strategy based on the rainbow deep Q-network (rainbow DQN) with an adaptive action-masking mechanism. Rainbow DQN improves training performance through prioritized experience replay and n-step returns, enabling efficient exploration in a high-dimensional action space. Furthermore, an adaptive action-masking mechanism dynamically excludes unsafe or repeated actions, ensuring operational feasibility and adherence to safety constraints. Comparative experiments on the modified IEEE-123 bus system demonstrate that rainbow DQN achieves faster convergence, higher load restoration rates, and stronger generalization compared to traditional DQN and double DQN. It also maintains better compliance with operational constraints. The proposed method offers a practical and efficient solution for rapid, intelligent restoration of large-scale distribution networks under extreme conditions.