A Hybrid Random Forest–XGBoost Framework for Power Transformer Health Index Prediction
In this study, a health index prediction framework for the power transformer is developed based on machine learning (ML) for evaluating the reliability and operational state of it. The feature engineering and Gaussian noise-based data augmentation are performed on the dataset to select the appropriate features and generate a large amount of data for ML algorithms. Several ML algorithms, including Linear Regression (LR), Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Lightweight Gradient Boosting Method (LightGBM), are employed. Among these algorithms, the two most superior are chosen based on their individual performance. The Optuna optimizer and residual learning strategy are applied to improve the performance of the proposed method and combine the predictions of superior ML algorithms for the final prediction. The simulation evaluation is done by collecting the dataset from the previous studies. The proposed method achieves an RMSE value of 2.49, an MAE value of 1.45, a MedAE value of 0.72, and an R² value of 0.98. Finally, the feature interpretation is done using the SHAP analysis, which shows that Dibenzyl Disulfide (DBDS), Interfacial V, Methane, and DBDS-Water exhibit the largest impact on transformer life expectancy.