An explainable AI-based predictive maintenance framework for transformer lifespan assessment and load forecasting in high-voltage power systems

Abstract

Abstract Accurate load forecasting and reliable transformer lifespan prediction are essential for ensuring the stability and efficiency of modern high-voltage power systems. However, existing approaches often rely on single-model architectures that inadequately capture temporal dependencies and nonlinear interactions, while providing limited model interpretability and maintenance decision support. This study proposes an explainable stacking-based hybrid machine learning framework for simultaneous load forecasting and transformer remaining lifespan prediction in 150 kV power systems. The framework combines Artificial Neural Networks (ANN), Random Forest (RF), XGBoost (XGB), and Long Short-Term Memory (LSTM) with an attention mechanism to model nonlinear relationships and temporal dynamics. The proposed framework is trained and evaluated using a physics-informed dataset comprising 26,255 hourly samples with multi-domain operational and thermal features. Experimental results demonstrate that the proposed stacking model achieves competitive performance among the evaluated machine-learning models for load forecasting, with an RMSE of 12.78 MW and an R² of 0.809. For transformer remaining lifespan prediction, the proposed model achieves an MAE of 1.470 years, an RMSE of 2.057 years, an MAPE of 6.39%, and an R² of 0.962, indicating its capability for reliable long-term asset health assessment. To enhance model transparency, SHapley Additive exPlanations (SHAP) identify hotspot temperature, aging acceleration, and load ratio as the dominant factors influencing transformer health. Furthermore, the predicted load and remaining lifespan are translated into maintenance-oriented risk levels and maintenance recommendations, supporting maintenance prioritization and asset management decision-making. Overall, the proposed framework provides an accurate, interpretable, and practically applicable solution for intelligent transformer condition assessment and predictive maintenance in modern high-voltage power systems.