An Empirical Analysis of Machine Learning Models for the Prediction of Incipient Fault in Power Transformer

Abstract

The urgent need to monitor oil-filled power transformers’ health on daily bases is due to the incipient faults that lead to economic loss. However, the most used traditional technique which is dissolved gas analysis (DGA) for incipient fault detection is characterized by their inability to categorize the state of the faults. This is because the DGA datasets can be imbalanced, insufficient and overlapping; imposing limitation in obtaining accurate diagnosis. This study investigated an ensemble of classifiers used to build fault detection diagnostic model for power transformers using DGA. The proposed methods include using data transformation techniques, machine learning algorithms: Support Vector Machine, Logistic Regression, Multilayer Perceptron, and their ensembles: voting, stacking, boosting, bagging, and random forest classifiers. The prediction model was applied on 298 data samples with seven independent attributes. The research results showed that the AdaBoost Radom Forest ensemble model with an accuracy of 100% performed better than other methods for the prediction of incipient faults in power transformers. The findings, therefore, suggest that the performance of the use of ensemble of classifiers could be influenced by the type and size of the datasets, and models’ parameters.