Physics–Causality Coupled Residual Graph Attention Networks (PCCResGAT) for Anomaly Detection and Root-Cause Localization in Distribution Grids

Diagnosing abnormal events and localizing their sources in distribution grids are complicated by DER-driven operating variability, limited and heterogeneous sensing, and long-tail anomaly distributions. This paper presents PCCResGAT, a physics–causality coupled residual graph-attention framework for unified, node-level anomaly detection, anomaly-type classification, followed by event-wise, propagation-aware root-cause localization. From multivariate electrical measurements, we construct (i) an admittance-derived physics graph that encodes network constraints and (ii) a data-driven temporal causal graph that captures directed influence and time-lag relations among nodes. These two graphs are fused by an adaptive gated residual GAT backbone to obtain topology- and causality-aware representations for diagnosis. For root-cause localization, we perform Top-Kranking using a propagation representation together with a multi-factor scoring rule that integrates causal strength, temporal precedence, and physical-impact cues. An imbalance-aware training objective and calibrated decision rules further improve robustness under long-tail distribution conditions. Experiments on a physics-consistent SimBench-based benchmark with injected anomalies (overload, non-technical-loss/theft, and reactive interference) show that PCCResGAT consistently improves detection and classification performance while yielding more actionable root-cause rankings across diverse operating conditions. Notably, despite inherent measurement noise and stochastic load fluctuations in the real-world data, the proposed framework maintains high robustness.