An interpretable stacking-optuna-SHAP framework with causal inference for remaining useful life prediction of power transformers

Power transformer remaining useful life (RUL) prediction is critical for grid reliability, yet existing methods suffer from insufficient feature integration, incomplete prediction mechanisms, and poor interpretability. This paper proposes a Stacking-Optuna-ML-SHAP framework to address these issues. Using data from 38 transformers in a southern China provincial grid, a three-dimensional feature system (main body, oil quality, dissolved gases) is constructed. A Stacking ensemble model with Optuna-tuned hyperparameters and a base-learner selection metric achieves superior accuracy: on the test set, R 2 = 0.9747 and MAE = 1.1724 years; validation on 18 decommissioned transformers yields MAE = 1.2274 years and RMSE = 1.7561 years, reducing MAE by 46.64% and RMSE by 28.50% compared to the traditional health index method. SHAP analysis reveals service time as the most influential feature (54.59% importance) and provides quantitative decision thresholds. LiNGAM causal inference identifies service time as the key causal driver (effect −1.14 years per service year) with high reliability. The proposed framework delivers accurate, interpretable, and causally grounded RUL predictions, eliminating dependence on hard-to-measure parameters and supporting condition-based maintenance.