Data-driven machine learning for enhancing predictive maintenance of distribution transformers

Predictive maintenance of distribution transformers remains a major challenge for electric utilities, as failure events represent a minority class and entail high operational and economic costs. This study presents a data-driven framework for transformer failure prediction using readily available operational and installation data, with explicit attention to class imbalance and risk-based maintenance prioritization. A comparative evaluation of classical machine learning models is performed, including LR, SVM, DT, RF, and XGBoost. Models are assessed using stratified 10-fold cross-validation and metrics suited to imbalanced classification and failure detection, with emphasis on Recall, F1-score, Precision–Recall AUC, and the Matthews Correlation Coefficient (MCC). Two experimental settings are analyzed: a baseline configuration and an imbalance-aware optimized configuration combining resampling and hyperparameter tuning. In the baseline setting, Random Forest demonstrates the most robust and stable behavior, achieving a strong balance between missed failures and false alarms (MCC ≈ 0.91). Under the optimized setting, XGBoost shows the largest improvement, reaching higher detection capability while maintaining controlled false-positive rates (Recall ≈ 0.92, F1-score ≈ 0.93, MCC ≈ 0.90), with consistent performance across folds. Model interpretability is addressed using SHAP applied to the optimized XGBoost model, highlighting coherent contributions from aggregated risk indicators, electrical stress variables, loading conditions, and installation context, thereby improving transparency without implying causality. The framework is validated on operational data from approximately 32,000 distribution transformers across 42 municipalities in the Cauca Department, Colombia (2019–2020). The results show that comparative machine learning combined with imbalance-aware evaluation and interpretability can effectively support practical, scalable, risk-based predictive maintenance strategies.