Deep graph neural network for fault detection and identification in distribution systems

Effective fault diagnosis is critical for ensuring power system reliability, preventing failures, and enhancing resilience. However, existing methods encounter challenges in fault detection accuracy, limited fault type coverage, and scalability, issues exacerbated by increasing renewable integration, and practical data noise. This paper proposes a deep graph attention network (GAT) for detecting and managing fault events on distribution systems, addressing above the first two limitations. Our proposed one-dimensional GAT leverages spatial and temporal data more effectively than traditional data-driven methods. High-accuracy mu PMU measurements are exploited as multi-dimensional attributes to enhance feature representation. The validation results on two test systems show that the proposed method improves fault event detection accuracy by 1%-2%, fault type classification by 4%, and fault localization by 5%. Additionally, the method demonstrates improved robustness under noise conditions, achieving a 1.3% to 3% improvement over traditional methods as noise levels increase from 3% to 10%.