Predicting the Accuracy of Transformer Oil Classification by Goodness-of-fit Statistics

Abstract

Classification of transformer oil as per IEC C.57.106-2006 is performed using Partial Discharge (PD) measurements. Fresh and used oil samples procured from the State Electricity Board are used for the classification. Histogram Similarity Measures (HSM) like Kolmogorov Smirnov (KS) test, Chi-square test and Cross-correlation are used to find the similarity between the histograms of the PD data and classify them. Subsequently, to predict the accuracy of the decision made while classifying, a probability is associated with each classification. The test statistics of HSM are fitted using Beta, KS and Chi-square distributions and their class likelihood probabilities are evaluated. Eventually, the class assignment and related probabilities from different PD measurements are added up to get a final class assignment and probability value for the test oil sample.