Multiagent collaborative decision-making method for self-healing control of distribution networks

Abstract

Fast self-healing of active distribution networks requires coordinated fault isolation, load pickup, distributed-energy resource dispatch, and switching-safety verification. This paper proposes a multiagent collaborative decision-making (MAS-CDM) method in which feeder, zone, distributed energy resource (DER), and load agents exchange compact graph-attention messages and generate local restoration actions under centralized training and decentralized execution. Compared with a single reinforcement-learning controller, the proposed method explicitly separates state sensing, neighbor communication, candidate-action generation, and physics-based safety projection. To make the method more suitable for automation-control scenarios, camera, infrared, and LiDAR inspection flags from charging piles, traffic hubs, and logistics parks are encoded as auxiliary state inputs. The resulting control pipeline integrates message aggregation, actor-critic training, safety projection, reward shaping, and online command screening, and it is evaluated through baseline comparison, ablation, robustness, and delay/noise sensitivity tests. Experiments on modified IEEE 33-bus and 69-bus feeders show that MAS-CDM restores 98.0% weighted load with a 0.46 s online response and reduces operational violations from 0.31 to 0.04 per episode compared with graph reinforcement learning. The results further show that graph-attention messaging improves load pickup, while the safety shield is the decisive component for secure field deployment.