Detecting Transformer Fault Types from Dissolved Gas Analysis Data Using Machine Learning Techniques

Asset management in power systems is a critical for maintaining reliability. Current methods rely on preventive maintenance to upkeep asset health. However, preventive maintenance can be time- and cost-intensive if not needed. Instead, predictive maintenance can be used to act as needed, depending on real-time data analytics, thereby, allowing corrective actions to be quicker and more targeted. The goal of this project is to identify power transformer fault types using dissolved gas analysis data. The data consists of the ppm values of dissolved gases in the New York Power Authority's transformers' insulating oil over a 2-year period. The gases include Acetylene, Ethylene, Ethane, Methane, Carbon Monoxide, Carbon Dioxide, and Hydrogen. There are three sub-methods involved in detecting the fault types. Firstly, is the Key Gas Method. An Isolation Forest model is used to detect outliers for each gas with 100% accuracy. Secondly, the Basic Gas Ratio is computed and using a supervised learning model, k-Nearest Neighbors, the fault type is predicted with 88% accuracy. Finally, the prediction is validated using a graphical Duval Triangle method. Overall, detecting transformer fault types using machine learning techniques allows the operators and technicians to prepare for faults and predictive maintenance to be done.