Investigation on transformer oil parameters using support vector machine

Abstract

Machine Learning has been used to predict the transformer oil parameters by using data obtained from Megger tests and transformer oil test. The relationship among the measured insulation resistance (among distribution transformer's low tension winding, high tension winding and ground) with breakdown strength, acidity, water content, and interfacial tension of transformer oil is modeled for the prediction. Support Vector Machine is the algorithm used for the prediction of the parameters. A cascaded network approach has been used where stage-wise division has been done to obtain different parameters depending on their correlation with each other. The cascaded network takes insulation resistances as input to predict breakdown and interfacial tension which are further used along with colour as input to predict water content which is further used to predict the acidity. Even though there was a lack of sufficient dataset for training the network the results seemed to be promising. Testing data was used to verify the network and the results were good as evident from the confusion matrices obtained. Therefore it is concluded that SVM is a good technique to predict transformer oil parameters with accuracy.