Autonomous Formation for Post-disaster Multi-Microgrid Via Multi-Agent Reinforcement Learning

Abstract

Post-disaster power system restoration often requires the rapid formation of multiple islanded microgrids under severe infrastructure damage, limited communication, and high combinatorial complexity. Traditional centralized optimization approaches struggle to provide timely and scalable solutions in such scenarios. To address these challenges, this paper proposes a Multi-Agent Microgrid Formation framework (MA-MMGF) for autonomous post-disaster power system restoration. To improve learning efficiency and scalability, a connection-only mechanism (COM) and a self-organizing network mechanism (SONM) are introduced to significantly reduce the action-space dimensionality while preserving feasible energization paths. A graph convolutional network (GCN)-based Q-network is employed to capture topological dependencies among neighboring microgrids, enabling topology-aware decision making. In addition, an experience screening strategy is adopted to enhance training stability under sparse and imbalanced rewards. Furthermore, to comprehensively quantify restoration quality, we propose an Integrated Resilience Indicator (IRI), which synthesizes recovery rate, overload rate, and time-efficiency into a unified metric. The proposed approach is evaluated on modified IEEE 30-bus test systems and IEEE 300-bus systems using Monte Carlo line-damage simulations.Extensive experimental results demonstrate that MA-MMGF significantly outperforms comparison method in terms of distributed generation utilization, load restoration ratio, and system stability.The results verify that the proposed framework enables fast, scalable, and robust autonomous microgrid formation, offering a promising solution for resilient power system operation in post-disaster environments.

Keywords: Deep reinforcement learning, Multi-microgrid, Power system resilience, Load recovery