Machine Learning for Early Transformer Insulation Defect Risk Ranking

Early identification of transformer insulation defects during IEC 60270 induced-voltage acceptance testing could enable high-risk units to be prioritized before the guaranteed test voltage is reached. This study investigated whether early-ramp partial discharge characteristics could predict the partial discharge inception voltage (PDIV) and support defect-risk ranking. An HPWOA-optimized dual-output feedforward neural network (HPWOA-DO-FNN) was developed and compared with Linear Regression, Random Forest (RF) Regressor, Gradient Boosting Machine (GBM) Regressor, Logistic Regression Classifier, and RF Classifier. Ten features were extracted from 120 s sliding windows applied to the first 50% of IEC 60270 test records. The dataset comprised 1705 windows from 39 source files representing 18 power transformers tested in South Africa between 2019 and 2021. Five-fold file-level GroupKFold cross-validation prevented leakage between windows from the same test file. Uncertainty was estimated using 1000-replicate group-level bootstrap confidence intervals. Performance was evaluated using R2, RMSE, area under the receiver operating characteristic curve (AUC), F1 score, and top-quartile risk capture. All models demonstrated near-zero or negative coefficients of determination (R2) for absolute PDIV prediction (best result: GBM-Regressor PDIVRatio R2 = 0.097, 95% bootstrap CI [−0.237, 0.354]), confirming that PDIV cannot be reconstructed from 10 early-ramp features alone—an expected result given that PDIV is a transformer-level property, while windows sample correlated sub-sequences of a single test session. Despite this, relative defect-risk ranking remains achievable: the ranking task does not require accurate point prediction. GBM-Regressor achieved the highest PDIVRatio R2 (0.097) and lowest RMSE (0.323), although its confidence interval included zero. For defect discrimination, RF Classifier achieved an AUC of 0.714 (95% CI, 0.628–0.803) and the highest top-quartile capture (43.5%), while GBM-Regressor achieved an AUC of 0.712 (95% CI, 0.624–0.801) and an F1 score of 0.603. HPWOA achieved an AUC of 0.527 (95% CI, 0.389–0.662). Early-ramp features are insufficient for accurate transformer-level PDIV prediction but provide useful relative defect-risk ranking. GBM-Regressor and RF Classifier demonstrated the strongest discrimination and represent the most promising approaches for early-risk triage in this small tabular dataset.