Ensemble model-based transformer insulation life assessment using insulating paper degradation-linked parameters

• The study demonstrates ensemble learning models' ability to provide accurate, non-invasive DP estimation. • Gradient boosting achieved the highest accuracy for DP prediction, while random forest excelled in predicting elapsed life. • The model uses Spearman and CAM for effective feature selection, identifying 2FAL, paper damage, and tensile index as the most influential parameters. • Bootstrapping shows the best generalization performance among other resampling techniques. The degree of polymerization (DP) is regarded as a key parameter for assessing the health and remaining lifespan of power transformers. However, measuring DP is challenging due to the difficulty of extracting paper samples from operational transformers. While many studies have relied on markers dissolved in insulating liquid to estimate DP, this approach has limitations. Different types of paper produce varying markers, and the commonly used marker, 2FAL, only becomes detectable when DP is approximately 600. In addition, retro-filling can influence the exact concentration of these markers as the insulating liquid might have been replaced several times. To address these challenges, this study employs parameters directly linked to paper degradation to accurately predict DP and the elapsed life of insulating paper. As a method, the first ensemble models were used to predict the DP and the predicted DP values were subsequently employed to estimate the elapsed life of the insulating paper, utilizing the two best-performing models for the second ensemble model. The gradient-boosting regression model emerged as the best for DP prediction, achieving a correlation coefficient of 0.9921, while the random forest regression model excelled in predicting elapsed life, with a correlation coefficient of 0.9940.