Fault Classification in Power Transformers via Dissolved Gas Analysis and Machine Learning Algorithms: A Systematic Literature Review

In electrical power systems, from generation power stations down to distribution substations, power transformers play a key role in ensuring reliable electricity transfer in the correct range from the generating source to the end-users. Over time, due to their operational demands and other various factors, transformers become susceptible to failures which threaten their reliability and life span. To address this issue, various transformer fault diagnosis methods are employed to detect and monitor the state of transformers, such as the dissolved gas analysis (DGA) method. In this paper, a systematic literature review (SLR) is conducted using the Preferred Reporting Items for Systematic Reviews (PRISMA) framework to record and screen current research work pertaining to the application of machine learning algorithms for DGA-based transformer fault classification. This study intends to assess and identify potential literature and methodology gaps that must be explored in this research field. In the assessment of the literature, a total of 124 screened papers published between 2014 and 2024 were surveyed using the developed PRISMA framework. The survey results show that the majority of the research conducted for transformer fault classification using DGA employs the support vector machine (32%), artificial neural network (17%), and k-Nearest Neighbor (12%) algorithms. The survey also reveals the countries at the forefront of transformer fault diagnosis and a classification based on DGA using machine learning algorithms. Furthermore, the survey shows that the majority of research conducted revolves around fault diagnosis with an emphasis on improving the accuracy of techniques such as SVM and ANN. At the same time, limited effort is put into other key metrics such as precision, Mean Squared Error, and R-Squared, and also, current works surveyed do not explore regularization techniques for preventing overfitting and underfitting of the proposed models.