Spatiotemporal Graph Networks for Predicting Transformer Failures in Regional Power Grids

Abstract

Power transformers are critical assets in regional electricity infrastructure, and their unexpected failures frequently trigger cascading outages with severe economic consequences. Traditional fault diagnosis approaches, including dissolved gas analysis (DGA) and periodic physical inspections, remain inadequate for proactive real-time monitoring across large and topologically complex grid networks. This paper proposes a spatiotemporal graph network (STGN) framework that jointly models the topological structure of a regional power grid and the temporal dynamics of transformer operational data to achieve early failure prediction. The architecture integrates spatio-temporal convolutional blocks, each comprising temporal gated convolution layers flanking a spatial graph convolution layer with gated linear unit activations, to capture both spatial propagation patterns between interconnected devices and time-varying deterioration signals at individual nodes. Panoramic state information, including dissolved gas concentrations, load data, and qualitative maintenance records, is encoded through sequential LSTM processing to support multi-horizon failure probability estimation. Experiments on a real-world regional transmission dataset demonstrate a precision of 91.4%, recall of 89.7%, and F1 score of 90.5%, surpassing support vector machine, LSTM, and standard convolutional neural network baselines by margins of 8 to 17 percentage points. Performance comparison across Correlation, CSI, FAR, and POD metrics at multiple prediction horizons confirms the sustained accuracy advantage of the spatiotemporal formulation over non-graph baselines.