Machine learning-based prediction of moisture and oxygen in a large power transformer with online monitoring validation

Abstrak

This study presents a predictive modeling approach for monitoring moisture and dissolved oxygen dynamics in a newly commissioned high-capacity power transformer. Using over 48,000 real-time observations collected across three years via an advanced online monitoring device installed on a 326 MVA generator step-up transformer (GSUT), machine learning models were developed to estimate moisture and oxygen concentrations based on correlated operational parameters. Multiple regression-based algorithms were trained and evaluated using performance metrics including root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Linear regression achieved superior performance with an RMSE values as low as 0.05888 ppm for oxygen and 0.0153 ppm for moisture. The models were further validated using data from a sister transformer, demonstrating generalizability and reliability across similar transformer units. This work contributes a scalable and accurate solution for real-time transformer health assessment, with practical implications for predictive maintenance strategies in power utilities.