Virtual power plant enhanced load restoration in distribution networks: A distributed decision-making approach

Distribution networks may experience faults and blackouts due to various factors, resulting in the isolation of network sections from upstream substations. Conventional distribution network reconfiguration (DNR) relies solely on topological reconnection; even after establishing an optimal radial topology, DNR may still fail to restore loads due to insufficient local power supply capacity. Virtual power plants (VPPs), which aggregate distributed energy resources into a dispatchable entity, provide a complementary way to energize these topologically connected but power-deficient regions. However, integrating VPP coordination into scalable multi-agent reinforcement learning (MADRL) frameworks — while ensuring cross-region collaboration and avoiding action space explosion — remains a critical challenge. This paper proposes a multi-agent deep reinforcement learning approach for the joint optimization of network reconfiguration and VPP dispatch. A shared control strategy for critical tie-lines and VPP interconnection switches facilitates cross-community power delivery, while a dynamic action masking mechanism focuses agent learning on fault-affected zones. Evaluated on a real-world 80-node Chinese distribution network where conventional DNR yields limited restoration performance, comparative experiments on ablation variants demonstrate that the proposed framework significantly improves load recovery rate and system operational efficiency.