Residual-aware health prediction of power transformers via spatiotemporal graph neural networks

Accurate health state prediction and timely fault detection of power transformers are critical for ensuring the reliability and resilience of modern power systems. This paper proposes a residual-aware spatiotemporal graph neural network (STGNN) framework that jointly models dynamic topological dependencies among multivariate Supervisory Control and Data Acquisition (SCADA) signals and their temporal evolution. The proposed approach constructs time-varying sensor graphs via attention mechanisms to capture evolving inter-variable relationships and applies Chebyshev spectral graph convolution for localized spatial feature aggregation. Temporal dependencies are modeled using gated recurrent units (GRUs) enhanced with residual connections, enabling robust forecasting under nonstationary operating conditions. A health indicator (HI) is derived from node-level prediction residuals, and anomalies are detected using a quantile-based thresholding strategy. Extensive experiments are conducted on a synthetic SCADA dataset simulating five major transformer subsystems—winding, core, cooling, insulation, and tap changer—under both nominal and faulty conditions. Results demonstrate that the proposed STGNN achieves superior forecasting accuracy and significantly outperforms baseline methods in anomaly detection, particularly under noisy and dynamic scenarios. The framework offers a scalable, interpretable, and deployment-ready solution for intelligent condition monitoring in substation automation systems.