Explainable AI–driven sustainable transformer health monitoring for resilient power distribution systems

Transformer failures in power distribution systems cause lengthy outages and operational complexities, resulting in substantial revenue losses, highlighting the need for a reliable and explainable transformer health assessment. This paper proposes an explainable machine learning (ML) framework for transformer health assessment using a publicly available time-series dataset containing measurements from 13 electrical and thermal sensor variables. The data preparation steps involve sorting the data by date and time; filling missing data with the median value; and applying rules based on predefined thresholds to classify the transformer condition as Good, Fair or Bad condition. In order to avoid temporal leakage and simulate the real world, a walk-forward validation approach and a future target generation approach are used. The models that are evaluated are as follows: Decision Tree (DT), Random Forest (RF), Multi-Layer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). A pipeline incorporating SMOTE for class balancing was used to train each model. Hyperparameter tuning is used to enhance the performance of the model, employing RandomizedSearchCV. The proposed system is able to achieve better classification rates than the baseline systems with respect to precision, recall and classification performance. The system is also tested with missing data scenarios to check the robustness of the system. To improve model interpretability, explainable artificial intelligence (XAI) techniques such as Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), were employed to identify the most important features for transformer degradation. It is noted that important features like voltage, load current, and oil temperature have a considerable effect on the classification of transformers’ health state.