Ensemble Machine Learning Methods for better Dynamic Assessment of Transformer Status

Abstract

Analyzing dissolved gases in the transformer's mineral oil helps to detect and classify the systemic faults in electric power transformers. Formerly, empirical methods such as Rogers ratio, Duval triangles 1–4–5, and pentagons 1–2 were used for transformer fault classification. Loose fit for every transformer type is one of the most prominent disadvantages of conventional methods. Formulating robust machine learning algorithms, such as the decision trees, can significantly overcome the loose fit issue. This paper focuses on implementing four different decision tree algorithms, including a regular decision tree classifier, a bagging classifier, a boosting classifier, and a stacking classifier to classify six different transformer fault types distinctly. Further, this study shows that the efficacy and accuracy of the four mentioned classifiers could be far exceeded when combined using a wisdom of the crowd approach. The wisdom of the crowd approach essentially merges the predicted classes from the four individual classifiers and decides on the final prediction via a hard-voting routine. The computational evaluation revealed that the given voting approach could significantly improve power transformers' online diagnostic accuracy up to 91%, thus aiding early forecast of power transformers' preventive maintenance.