Lifetime Prediction and Interpretability Analysis of Power Transformers Based on Multi-Model Feature Selection and Stacking Ensemble

To address the challenge of accurately assessing transformer health under multi-source heterogeneous data conditions and to enable precise full-lifecycle lifetime prediction along with interpretable decision-making insights, we propose an ensemble learning framework that integrates multi-model feature selection with Stacking-Optuna optimization. First, we construct a full-lifecycle data framework that integrates a basic data layer with a derived feature layer, resulting in an 89-dimensional feature set. Second, feature importance is evaluated via equal-weighted integration of Random Forest, LightGBM, and XGBoost, and 13 key features are selected using an 80% cumulative importance threshold. On this basis, a Stacking model is constructed using CatBoost, XGBoost, and LightGBM as base learners, with Ridge regression as the meta-learner, while Optuna optimizes the model’s parameters and architecture. Finally, SHAP is employed to perform multi-level interpretability analysis. Experimental results show that the proposed model achieves MAE = 1.3411, RMSE = 1.6899, and R2 = 0.9502 on the test set, outperforming the best single model (CatBoost). SHAP analysis further reveals that commissioning year is the feature with the highest SHAP contribution (44.54%), and that feature contributions exhibit a nonlinear pattern, initially decreasing and subsequently increasing with service time. This method achieves high accuracy and interpretability, providing a reference for transformer condition assessment and operational decision-making.