Cost effective assessment of transformers using machine learning approach

Abstract

Furan content in transformer oil is highly correlated with the transformer insulation paper aging. In this paper, the ranges of furan content in power transformer is predicted using measurements of transformer oil tests like breakdown voltage, acidity and water content. Machine learning approach is adopted, and maintenance data collected from 90 transformers are used. A maximum of 67% recognition rate was achieved using Decision Tree classifier. The major challenge of the used data is the relatively low number of available samples in certain furan intervals. Two solutions have been proposed to overcome this imbalanced classification problem, namely, using an over-sampling technique and balancing data distributions by reducing the number of intervals to be predicted to three instead of five intervals. The recognition rate has improved to reach 80%.