A Dual-Graph Physics-Informed Graph Attention Network for Fault Location in Active Distribution Networks

INTRODUCTION: Fault localization in active distribution networks (ADNs) is challenging because feeder topology and electrical coupling may become inconsistent under operating conditions with high penetration of distributed energy resources. Moreover, purely data-driven models often provide limited physical interpretability.

Objectives

This study aims to develop a dual-graph learning framework for fault localization in ADNs, with particular attention to scenarios where topology-based message passing alone may be insufficient due to topology–electrical mismatch.

Methods

A dual-graph physics-informed graph attention network (DG-PIGAT) is proposed. The method uses a topology graph to describe feeder connectivity and an electrical-distance graph to characterize impedance-related coupling among buses. Electrical-distance-guided propagation is introduced to improve feature aggregation under electrical coupling variations, while Kirchhoff’s Current Law (KCL)-related residual information is incorporated as a weak physical regularization term to support physically plausible learning.

Results

Experiments on the IEEE 33-bus benchmark under four consolidated scenarios show that DG-PIGAT achieves the highest mean accuracy across the tested scenarios. Its clearest advantage appears under topology–electrical mismatch, where it achieves 86.92% accuracy and outperforms GraphSAGE by 8.29 percentage points (p = 0.0025). Sensitivity and robustness tests also indicate better tolerance to measurement noise, missing data, and topology–electrical inconsistency.

Conclusion

The results suggest that combining feeder topology with impedance-related electrical coupling is beneficial for robust fault localization in ADNs. The KCL-related component should be interpreted conservatively as a physical regularization mechanism rather than the dominant source of accuracy improvement.