HVDC GIS/GIL - Machine Learning Algorithms for Online PD Classification at DC Voltage

Abstract

The measurement of partial discharges (PD) is one of the most important diagnosis methods for investigating the condition of high-voltage (HV) equipment. The interpretation and classification of PD defects can be done either manually by human experts using visual interpretation tools or automatically by machine learning (ML) algorithms. The latter has many advantages, especially for online PD monitoring applications. Reliable PD monitoring systems with automatic PD classification already exists for AC applications, unlike for DC applications. The increasing use of HVDC systems makes it necessary to develop a significant and reliable automatic classification tool for PD monitoring at DC voltage as well. Therefore, ML algorithms are trained to automatically recognise hidden pattern in the data without visualisation, interpret the data and to initiate further steps. To achieve good and reliable classification results, feature extraction and selection are the first important steps in training an algorithm. Based on the fundamental parameters of PD measurements at DC voltage, the time of occurrence of each PD pulse ti and the PD magnitude qi, several features were extracted and already analysed in previous publications. This contribution deals with PD classification using ML algorithms and is the next crucial step towards an automatic PD classification at DC voltage. Several algorithms with different parameters, support-vector machines (SVM) and artificial neural networks (ANN), were trained and investigated. The feasibility and reliability of the algorithms were determined using online PD measurements of typical PD defects of gas-insulated systems in a laboratory test setup.