Intelligent Hyper parameter Optimization Framework for Explainable Machine Learning-Based Power Transformer Fault Classification

Abstract

Power transmission and distribution systems is intensive to be relatively very highly on Power transformer, a sudden failure in these systems can result in severe operational disruptions, including significant time loss, effluent equipment deficiency, financial. It is based on DGA features, machine learning classification, advanced hyper-parameter tuning and Explainable AI (XAI). This study is a quantitative predictive method based on a sample of 350 transformer condition observations including normal, partial discharge, low-energy discharge, high-energy discharge, thermal faults and combination electrical/thermal fault scenarios. Approaches includes pre- processing, feature normalisation, stratified sampling, Class-aware evaluation, model training hyper – parameter optimisation and Explainability analysis. Introductory descriptive and correlational statistics demonstrate key relationships between the severity of faults in transformers and the characteristics of dissolved gases. An illustrative test where these metrics topped 91.2% accuracy, 89.7% balanced accuracy and a macro-F1 of top 0.88 versus those from fine-tuning the model to discover that this optimisation framework brought them instead to: 95.4%, 94.1% and a macro-F1. The validation loss is lowered from 0.286 to 0.171, and novel model showed more consistent results during the whole training process. These results indicate that hyperparameter tuning could improve the classification performance of manually-constructed machine-learning classifiers, particularly in its multi-fault and/or class imbalance contexts. Finally, leveraging XAI helps in identifying the key features of DGA that contribute to each predicted outcome thereby increasing interpretability and allowing engineers to gain insight from the model. Future validation efforts should include field measurements, long-term monitoring data, and cohorts of automated transformers before actual deployment.