HVDC GIS/GIL – Classification of PD Defects using NoDi* Pattern and Machine Learning

Partial discharge measurement is one of the most important diagnosis methods for high voltage equipment. For equipment under AC voltage exists reliable interpretation tools such as the PRPD pattern. Due to the lack of phase information at DC voltage, many of the established methods for interpreting PD cannot be used. At DC voltage, the classification of PD defects can be realised with NoDi* pattern or automatically with machine learning algorithms. For this purpose, features are extracted from the measured PD data and provided to the trained algorithm. Five typical PD defects of HVDC GIS/GIL were investigated
– bouncing particles, particles in firefly effect, floating electrodes, and chamber and busbar corona. New aspects of floating electrode with respect to the dynamic range of the input amplifier of the PD measurement system were investigated. For this purpose, two measurement systems in parallel with different gain settings were used. Furthermore, the possibility to distinguish between busbar and chamber corona at DC voltage was examined. The extracted features for the machine learning algorithms were analysed and both manual and automatic feature selection were performed and compared to verify the manual feature selection. The results were in good agreement. In a next step, the necessary number of PD pulses was investigated to obtain the same seperability between the PD defects, whereas a number of 500 PD pulses seems sufficient for a PDM system.