Enhancing power transformer insulation assessment using gradient boosted decision trees and 2-FAL analysis

This study presents an empirical model for predicting the degree of polymerization (DP) and estimating the remaining useful life (RUL) of transformer insulation based on 2-furfuraldehyde (2-FAL) concentrations. Using 125 samples collected from multiple credible sources, a Gradient Boosted Decision Trees (GBDT) model was developed to improve prediction accuracy. Compared with existing models such as Chengdong, Burton, and Heisler, the proposed GBDT achieved superior performance, with lower Mean Absolute Error (MAE = 38.65) and Root Mean Squared Error (RMSE = 63.04). Graphical analyses confirmed a strong agreement between predicted and measured DP values, effectively capturing the nonlinear relationship between 2-FAL and DP. Sensitivity analysis showed that the model responds notably to small variations in 2-FAL at early degradation stages. The results enhance transformer diagnostics and enable proactive asset management through accurate, non-invasive, and data-driven monitoring of insulation aging.