THRAN: Temporal Hierarchical Residual Attention Network — an Intelligent Multi-Domain Feature Fusion Framework for Transformer Health Condition Monitoring and Prediction

Abstract

Power transformers are the lifeline of electrical power transmission and distribution systems and their sudden failure can result in major outages, significant economic losses, and significant safety risks. Thus, the key operational challenges for asset managers in the utility and industrial sectors are reliable health condition monitoring and remaining useful life (RUL) prediction. In this paper, a novel three-stage hybrid ensemble framework, termed THRAN (Temporal Hierarchical Residual Attention Network), is proposed to combine Multi-Domain Feature Fusion (MDFF), which is constructed using thermal polynomial interactions, chemical log-exponential gas signatures and spectral sinusoidal projections, with an Adaptive Fault-Aware Attention layer based on entropy and kurtosis contextual anomaly features, and a LightGBM residual stochastic gradient boosting meta-learner over the enhanced representation. The dataset used for testing THR is a transformer health dataset consisting of 10 features: oil temperature, winding temperature, dissolved gas analysis (H2, CO, C2H2), moisture content and operating hours, against nine reference algorithms: SVR (RBF), Ridge Regression, Bayesian Ridge, Lasso Regression, SVR (Linear), KNN (k=15), KNN (k=7), MLP (Medium), MLP (Small). THRAN achieved the highest performance across all six evaluation metrics: RMSE=1.9279, MAE=1.4940, R2 =0.9223, MAPE=1.9465%, Pearson r=0.9634, and CV R2 =0.9206, representing a 5.3% improvement in R2 over the next best model (SVR RBF, R2 = 0.8763). The proposed framework promotes the monitoring of smart/ resilient power infrastructure and promotes reliable clean energy transmission systems.