A Temperature-Compensated CNN-Based Method for Transformer Partial Discharge Localization

Abstract

An acoustic time reversal-convolutional neural network (ATR-CNN) approach is proposed for localizing partial discharge (PD) in power transformers with temperature compensation. A digital twin model is developed through multiphysics coupling analysis to accurately describe temperature distributions in oil-immersed natural air-cooling (ONAN) transformers. The temperature-compensated dual-sensor configuration demonstrates a root-mean-square error (RMSE) of 4.48 mm in PD localization, exhibiting a minimal accuracy degradation of 1.4 mm in unseen datasets while maintaining consistent performance across noise levels (0%–10%). Comparative analyses reveal the ATR-CNN methodology’s superior localization accuracy over traditional machine learning algorithms and enhanced performance in non-line-of-sight regions compared to the time difference of arrival (TDoA) approaches. A significant 264 000-fold reduction in computation time is achieved relative to ATR implementations. Integrating deep learning with ATR techniques offers an enhanced approach to PD localization in complex transformer environments.