A Feature-Enhanced Approach to Dissolved Gas Analysis for Power Transformer Health Prediction Through Interpretable Ensemble Learning and Multi-Model Evaluation

Abstract

Dissolved Gas Analysis (DGA) is a diagnostic strategy that monitors oil-immersed transformers by correlating their health status with various insulation degradation by-products, where the Health Index (HI) offers a unified metric for asset evaluation. Existing studies frequently emphasize classification accuracy or single-model regression, overlooking interpretability, feature reduction, and systematic benchmarking. This paper introduces a feature-enhanced multi-experimental methodology for HI prediction incorporating SHapley Additive exPlanations (SHAP) in a dual role—as both an interpretability and a feature selection tool. Models from four algorithmic families (linear, kernel/tree-based, boosting, and hybrid ensembles) were systematically benchmarked using a publicly available dataset. Results demonstrate that the proposed LightGBM–CatBoost hybrid ensemble, enhanced by SHAP-guided feature pruning, achieves superior predictive accuracy while reducing model complexity and improving transparency. Unlike prior works carried out using the same dataset, the proposed framework not only provides a balanced approach that combines interpretability and reduced complexity, but also surpasses previous regression-based approaches, reducing MAE and RMSE by 4.93% and 2.31%, respectively, and enhancing HI predictive accuracy by 1.45%.