Spatio-Temporal Adaptive Graph Convolutional Networks for Real-Time Power Equipment Monitoring and Fault Detection

Abstract

Traditional methods for monitoring the operational status of power equipment often suffer from limited feature representation and inaccurate dynamic modeling, resulting in low early-fault detection accuracy and elevated false-alarm rates. To address these limitations, this paper proposes a comprehensive, big-data-driven monitoring and fault warning framework, optimized end-to-end from data preprocessing to online alert generation. The system incorporates multi-scale sliding-window feature extraction, weighted trend-deviation quantification, and redundant-feature compression to enhance the representation of degradation signals. A spatio-temporal adaptive graph convolutional network (STAGCN) is employed to jointly capture equipment topology and temporal dependencies. In the warning module, an anomaly discrimination strategy based on confidence-score reconstruction is deployed on a hierarchical, parallel big-data platform, enabling sub-second alerts for critical faults (e.g., line short circuits, insulation degradation). Experimental evaluations reveal that, following feature compression, overall detection accuracy increases from 89.0% to 92.8%, while the false-alarm rate decreases from 4.8% to 3.5%. For typical line short-circuit faults, the proposed system achieves a warning accuracy of 96.5%, with response times under 3.1 seconds under full-load conditions. These results demonstrate that the proposed framework delivers high real-time performance, accuracy, and reliability in complex grid environments.