Spatial-Temporal Graph Convolutional Network for Fault Diagnosis in Weak Electrical Systems

Abstract

Weak electrical systems, such as low-voltage distribution grids and embedded sensor networks, are highly susceptible to faults due to their complex topology and limited fault tolerance. Accurate failure analysis and diagnosis in such systems are essential for maintaining operational reliability and safety. However, traditional diagnostic methods—such as rule-based systems or shallow machine learning—struggle with non-linear relationships, dynamic system behavior, and distributed component interactions. These limitations result in delayed or inaccurate fault detection, particularly in noisy or rapidly changing environments. To overcome these challenges, this paper proposes a failure diagnosis framework based on the Spatial-Temporal Graph Convolutional Network (ST-GCN), comprising a multi-channel CNN feature extractor for spatial pattern learning and an attention-guided temporal module for capturing temporal dependencies across system nodes. The architecture allows the model to learn complex spatiotemporal interactions and adapt to multi-modal sensor inputs effectively. The proposed ST-GCN achieves 96.7% accuracy, 0.95 F1-score, and 93.6% fault localization accuracy, significantly outperforming traditional methods. It also demonstrates sub-10 ms detection latency, 95.4% actual positive rate in the confusion matrix, and a Precision–Recall AUC of 0.96, while converging within 25 epochs and showing only 1.4% accuracy drop when scaled from a 33-bus to a 123-bus system. These results highlight the robustness, real-time applicability, and methodological effectiveness of ST-GCN for fault diagnosis in weakly meshed and low-voltage distribution grids.