Transformer Incipient Fault Diagnosis using Machine Learning Classifiers

Power transformers are important equipment in the electric grid. The availability loss of the transformer is a high impact event for various stakeholders in the power system network. Hence transformer health monitoring has garnered much significance. Dissolved Gas Analysis is by far the most prevalent condition monitoring technique among utilities. However, identifying the condition of the transformer from the DGA observations is a difficult task. Machine learning algorithms based on the principles of probability and decision making theories are now gaining momentum in this area. In this study, the performance of SVM, k-NN and ensemble method algorithms are compared in interpreting DGA data. Data transformation is performed on raw data to improve its quality. The DGA test datasets are generally skewed which hamper the performance of the classifiers. Data random sampling is employed to balance the data. The effects of the data transformations and data imbalance have been studied. The results obtained show that the ensemble method algorithm performed the best in most cases. Also, the performance of the classifier algorithms has been found to increase through data pre-processing and data balancing.