Predictive Anomaly Detection and Self-Healing in Active Distribution Networks Using a Graph Neural Network Digital Twin

Abstract

Active distribution networks (ADNs) are exposed to two anomaly classes that resist unified detection: physical line outages and cyber false-data-injection (FDI) attacks. The standard defense, weighted- least-squares (WLS) bad-data detection, catches FDI but is structurally blind to topology changes, because a reconfigured network still yields internally consistent measurements. This paper identifies a specific property of a predictive graph neural network (GNN): when trained to forecast the one-step-ahead voltage profile, its prediction residual exposes both classes, and its graph structure is what converts that residual into a line-level fault localization, not a more accurate detector. On the IEEE 33-bus feeder across six baselines, the GNN residual records 0 false positives, localizes the faulted line, and reconnects the islanded subtree in one switching action, whereas a CUSUM change-point detector and a graph-agnostic MLP match its detection F1 (0.99 and 1.00) yet provide no localization. The contribution is therefore a specific discovery,the GNN’s graph structure buys fault localization, not detection accuracy, together with the closed-loop self-healing it enables. A reproducible Python pipeline produces every reported number.