Artificial neural networks with stepwise regression for predicting transformer oil furan content

Abstract

In this paper a prediction model is proposed for estimation of furan content in transformer oil using oil quality parameters and dissolved gases as inputs. Multi-layer perceptron feed forward neural networks were used to model the relationships between various transformer oil parameters and furan content. Seven transformer oil parameters, which are breakdown voltage, water content, acidity, total combustible hydrocarbon gases and hydrogen, total combustible gases, carbon monoxide and carbon dioxide concentrations, are proposed to be predictors of furan content in transformer oil. The predictors were chosen based on the physical nature of oil/paper insulation degradation under transformer operating conditions. Moreover, stepwise regression was used to further tune the prediction model by selecting the most significant predictors. The proposed model has been tested on in-service power transformers and prediction accuracy of 90% for furan content in transformer oil has been achieved.