Bi-directional Long Short Term Memory Based Classification of Sulphur Corrosion in Power Transformers Using Mineral Oil Data

In high voltage power transformers, mineral oil is the most popular type of oil insulation as well as coolant. Mineral oil became corrosive in nature if there is the presence of various sulphur compounds (i.e., mercaptans, sulphides, free sulphur etc.) inside the oil tank. Sulphur compounds react with copper, aluminium and other metals present inside the transformer tank to form copper sulphide (Cu2S), aluminium sulphide (Al2S3) and other inorganic sulphides. The presence of these compounds is the reason for various kinds of faults inside the tank. In this study, a new deep learning (DL) algorithm had been used to classify the corrosive nature of the mineral oil as per ASTM D130 standard. If mineral oil became corrosive in nature, then there is a change in interfacial tension (IFT), furan (2FAL), oxygen (O2), dibenzyl di-sulfide (DBDS) concentration and colour code of the oil. That’s why these properties were taken as input features in the DL algorithm. An DL algorithm i.e., bi-directional long short-term memory (Bi-LSTM) was used to train and test the model, where the grid search method was used to select the best hyperparameters. Finally, the corrosiveness of the mineral oil is classified into four classes i.e., slight tarnish (ST), moderate tarnish (MT), dark tarnish (DT) and corrosion (CC). The performance of the DL model was compared with other DL models based on accuracy, recall and precision. In comparison with various other models, the proposed DL algorithm established the highest efficiency and real-time application effectiveness.