Partial discharge defect classification in cast-resin transformers using machine learning-based algorithms

Abstract Cast-resin transformers have been widely installed in a power system due to excellent arc-extinguishing characteristics, easy installation, and low expense of maintenance. From the perspective of condition monitoring, identification and classification techniques of internal defects based on partial discharge (PD) measurements are getting more important. This paper studies PD defect classification using two kinds of machine learning (ML) algorithms, random forest (RF) and artificial neural network (ANN) models. Four typical PD defect models were designed: metal protrusion, a particle on insulator, delamination, and a void. PD single pulses and phase-resolved partial discharge (PRPD) patterns at each partial discharge inception voltage (PDIV) were measured by the printed circuit board (PCB) based Rogowski-type PD sensor. Various kinds of PD features were extracted from each PD single pulse and PRPD pattern. From the experimental results, the two different ML algorithms, used in this paper, could classify the PD defects with over 90%, and the PD classification rate using RF model was slightly higher than that of the ANN model.