A Machine Learning Approach for FDS Based Power Transformer Moisture Estimation

This paper presents a machine learning approach for estimation of moisture content (MC) in power transformer insulation based on its frequency domain dielectric spectroscopy (FDS). The proposed method is based on principal component analysis and support vector machine approaches. It classifies transformer insulation into three classes based on the estimated MC levels i.e. dry (MC<1%), intermediate (1%3%). The proposed algorithm was trained and tested using FDS data sets consisting of 112 laboratory modeled samples and 20 field transformer samples. The laboratory modeled data were estimated using XY model from FDS measurements on pressboard samples having different moisture levels. A range of transformer geometries, aging conditions of pressboard and oil conductivities were considered when deriving the data set. The results showed that the moisture classification was accurate up to 96% thus indicating as a simple and promising method for transformer moisture estimation.