Comparative Analysis of Partial Discharge Source Identification Using Self Organizing Map (SOM) And K-Nearest Neighbour Method (KNN)

Abstract

Partial discharge (PD) detection and source identification play a critical role in ensuring the reliability and longevity of high-voltage insulation systems. Accurate classification of PD sources enables timely maintenance decisions and reduces the risk of catastrophic equipment failures. This study presents a comparative analysis of two machine learning techniques—Self-Organizing Map (SOM) and K-Nearest Neighbour (KNN)—for effective PD source identification. PD patterns were extracted from experimental high-voltage test setups, and relevant statistical and waveform-based features were derived to train both models. The SOM, an unsupervised neural network, was employed to cluster PD signatures and visualize underlying data structures, while the KNN classifier, a supervised learning method, was used to categorize PD sources based on proximity in feature space. Performance evaluation was conducted using accuracy, clustering efficiency, computational complexity, and sensitivity to noise. Results indicate that KNN provides higher classification accuracy and faster convergence for well-labeled datasets, whereas SOM demonstrates superior capability in handling unlabeled data, revealing hidden patterns, and providing intuitive visualization of PD classes. This comparative study highlights the strengths and limitations of both algorithms and offers insights into selecting suitable methods for PD monitoring systems in power engineering applications.