Transformer Partial Discharge Identification Based on Optimised Convolutional Neural Networks

With advancements in machine learning, automated pattern recognition has significantly improved various fields, especially in power systems. Accurate and rapid identification of partial discharge (PD) types in transformers is vital for system stability and safety. This paper presents a novel method for transformer PD pattern recognition based on optimized convolutional neural networks (CNN) and phase-resolved partial discharge (PRPD) pattern. The approach involves creating a PRPD dataset from ultra-high-frequency (UHF) sensor signals, followed by data preprocessing and labeling. By optimizing the CNN’s network structure and hyperparameters, the method achieves a remarkable recognition accuracy of $99.8 \%$, outperforming traditional algorithms like Support Vector Machines (SVM) ($\mathbf{9 3. 7 \%)}$ and Random Forests (RF) (89.3%). This study demonstrates the superior engineering applicability of the proposed method, highlighting the importance of network layer depth, convolutional kernel size, and activation function selection in improving model performance and generalization.