Investigation of electrical tree growth characteristics and partial discharge pattern analysis using deep neural network

• Published in: Electric power systems research Vol. 220; p. 109287
• Main Authors: Sahoo, Rakesh, Karmakar, Subrata
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2023
• Subjects: Convolutional neural network; Cross-linked polyethylene; Electrical tree; Growth characteristics; Partial discharge; Pre-trained neural network; XLPE; NN; CNN; Conv2D; Partial discharge; Electrical tree; DNN; Convolutional neural network; FM; DL; MCC; HV; FEM; Cross-linked polyethylene; PRPD; PD; Pre-trained neural network; VGG; Growth characteristics; FC; ML
• ISSN: 0378-7796
• DOI: 10.1016/j.epsr.2023.109287

•Formation of the electrical tree structure due to PD in XLPE insulation under sinusoidal high voltage.•Analyses of the tree growth characteristics with presence of a void at different applied voltage.•Identify the degradation of XLPE insulation and corresponding images of PRPD pattern during electrical tree growth instances.•PRPD image feature extraction, PD detection, and classify the level of insulation degradation using CNN and pre-trained NN.•Further, analyze the electric field strength in XLPE insulation in the presence of a void using COMSOL multiphysics. Partial discharge (PD) activities due to void and crack in high-voltage cross-linked polyethylene (XLPE) cables deteriorate insulation properties. The existence of void amplifies the electric field stress caused by variations in the dielectric constant between air and the insulation that is all around it. PD activities on such XLPE insulation under high voltage stress are responsible for the growth of electrical- trees. Deep learning techniques have given opportunities for automated feature extraction and classification of the severity level of discharge activity during tree growth. In this work, the XLPE sample specimens are aged at different applied AC voltages, and the respective PD signals are continuously saved for every 500 cycles. PD signals are used for three different stages of pattern classification including fast-growth, slow-growth, and breakdown phases during the growth of an electrical tree with the help of a convolutional neural network (CNN) and pre-trained neural network (VGG, Inception, ResNet, Xeception, DenseNet, NasNet) architectures. For each stage, 500 phase-resolved PD (PRPD) pattern images are obtained for training the neural network (NN) model. The performance of CNN with Adam optimizer obtains 98.44% total accuracy, whereas Densenet201 with Nadam achieves 99.78%.