Fault Prediction in Power Transformer Using Ensemble Models

Abstract

One of the highly important elements of electrical system networks is the power transformer. There is an increasing amount of research being done on early warning systems and faults detection because the failure of these elements can ground economic activities. More so, using dissolved gas analysis (DGA) as one of the mostly used conventional techniques is deficient in locating these incipient faults as this may be caused by a variety of factors which includes but not limited to imbalance problem, inadequate and overlap in the DGA datasets, thereby restricts its capacity to obtain precise diagnosis. Therefore, this paper proposed an ensemble machine learning methods for incipient faults prediction using DGA datasets comprising 166 samples and eight variables. This research compares the accuracies of four ensemble machine learning methods: Bagging, Adaboost, Stacking, and Voting methods using multilayer perceptron and support vector machines respectively. The results obtained ranges from 90.50% to 100% with the Adaboost (MLP) achieving the highest accuracy, whilst the misclassification percentage ranges from 1.62% - 18.06% with Stacking method as the least performing. In the end, our findings highlighted the importance of the use of ensemble methods and has future prospects for further advancement