Holistic Capture Graph Convolution Network With Ambient Signals for Line Short-Circuit Fault Diagnosis

Accurate diagnosis of fault factors in power system lines is crucial for restoring power supply and distribution. However, the diagnostic accuracy of current fault diagnosis methods is limited due to the single source of training data and the imbalanced fault categories. This article proposes a fault diagnosis algorithm based on ambient signals and graph convolutional networks (GCNs), approaching from two perspectives, broadening the input information sources and enhancing feature distinguishability to address this issue. First, this algorithm constructs a GCN adjacency matrix from system topology and enhances fault information by incorporating ambient signals derived from small disturbances during normal conditions. The fault and ambient signals are then decomposed into various intrinsic modes, serving as inputs of the diagnostic model to extract system information embedded in nonfault signals, thus achieving information increment. Subsequently, building upon the GCN architecture, the holistic capture GCN (HCGCN) is developed by integrating spatiotemporal attention mechanisms and integrative pyramid networks across temporal, spatial, and feature dimensions. This enables iterative updating of adjacency matrix weights and full-dimensional feature extraction. Finally, an imbalanced fault set is constructed based on practical operating conditions for comparative and ablation experiments to validate the proposed method. Various communication fault scenarios are also considered to test the method’s accuracy and robustness. The diagnostic accuracy up to 99.17% under no communication fault condition and 98.22% under communication fault demonstrate that the proposed algorithm significantly improves line short-circuit fault diagnosis performance in power systems.