Corrosive Dibenzyl Disulfide Concentration Prediction in Transformer Oil Using Deep Neural Network

Dibenzyl disulfide (DBDS) is the most prevalent corrosive sulfur in transformer oil. It reacts with the transformer windings to produce copper sulfide (Cu2S) and gets deposited on the insulating paper’s surface, leading to interturn faults within the transformer windings. Hence, this article proposes a deep neural network (DNN) to predict the DBDS content in transformer oil. The parameters like interfacial tension (IFT), breakdown voltage (BDV), water content (WC), oxygen, neutralization number (NN), color, furan content, and specific gravity (SG) were used as features to train and test the DNN model. The performance of the regression model was evaluated using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (
${R}^{{2}}{)}$
. Moreover, extensive analysis is carried out by varying feature combinations and test-train ratios to obtain the best prediction model. The values of DBDS predicted by DNN were further used to determine the corrosive sulfur concentration in transformer oil. The proposed method is validated on real-life transformer data obtained from the online dataset and on data obtained from the local power utilities. A comparative study showed better efficacy of the proposed DNN model than other prediction models for accurate DBDS prediction.