After a Distribution Fault, Speed Is Not the Only Goal: Graph Attention Cuts Self-Healing Decisions to 2.3 ms Near the Optimum

Online self-healing decisions for a faulted distribution network can be made in milliseconds while staying within a few percent of the mathematical-programming optimum. Earlier routes forced a choice: mixed-integer linear programming returns near-optimal restoration plans but its solve time grows sharply with system size, while deep reinforcement learning decides within milliseconds yet treats the grid as flat data and ignores topology, with an opaque decision process. On the IEEE 33-bus system the method decides in 2.3 ms against about 1.2 s for mixed-integer linear programming, a 2.8% optimality gap; on the larger IEEE 123-bus system the optimizer exceeds 10 s while the method stays at millisecond level.

Post-fault restoration is cast as a sequential decision problem on a graph, and the agent learns switching operations and distributed-generation dispatch with resilience as the training objective. Resilience assessment and real-time control have largely been separate, with resilience indices seldom entering the control objective, and policy networks have mostly been fully connected or convolutional, unable to exploit the topology of feeders and tie switches. Buses are nodes and feeders and tie lines are edges; node features carry voltage magnitude, injections and load level, edge features carry impedance, flow and switch state. A graph attention network serves as the shared backbone of actor and critic, and its attention weights adapt to input, so the agent follows the graph as it changes after a fault.

Decision evidence arrives through two complementary channels: attention weights inside the model show which buses the agent attends to, and post-hoc attribution ranks each state feature's marginal contribution. Conventional reinforcement-learning policies are black boxes, so dispatchers cannot check the restoration plan, which limits trust in safety-critical settings. In the single-branch fault scenario the attention heatmap concentrates on faulted branch 12–13 and adjacent buses, on tie switch 18–33 that offers a transfer path, and on distributed-generation node 25; the top five SHAP features include bus-18 voltage, branch 7–8 flow, bus-33 voltage, tie-switch 18–33 state and DG-25 output, with bus-18 voltage highest at 0.32.

The demonstrated scope is two standard test systems, IEEE 33-bus and 123-bus, under predefined fault scenarios: a single-branch fault opening branch 12–13 and a double-branch fault opening branches 2–19 and 6–26, with distributed-generation output discretized in the action space. The method targets distribution dispatch that needs online restoration decisions, trained offline and executed online. The authors propose continuous–discrete hybrid action spaces, hierarchical reinforcement learning, and domain randomization with meta-learning for extreme scenarios as next steps.

The authors note three qualifications: the action space is discretized, leaving room for joint optimization of continuous distributed-generation output and switching; training data come from predefined fault scenarios, so generalization to unseen extreme scenarios needs validation; and global SHAP attribution does not yet capture the real-time causal chain of a single decision. The 2.8% optimality gap arises mainly from action-space discretization, which the authors say could shrink with finer DG dispatch levels.