HVDC GIS/GIL – Behaviour and Machine Learning Classification of Partial Discharges in SF6 and Dry Air

For many decades, SF6 has been used as an insulating gas and for arc extinguishing in gas-insulated switchgears (GIS) and gas-insulated lines (GIL) due to its excellent electrical properties. However, it has the highest known global warming potential. For this reason, intensive research into environmentally friendly alternative gases has been initiated in the last 10 years. In addition, Regulation (EU) 2024/573 of the European Union severely restricts the use of SF6 from 2028 or 2032 depending on system ratings. Dry air under pressure has emerged as a potential alternative gas for HVDC GIS/GIL.
To date, the partial discharge (PD) behaviour of alternative gases has been insufficiently researched, and although the standardisation of PD measurements at direct voltage has been steadily expanded, very little progress has been made in the last 50 years. This highlights the need to study typical GIS/GIL defects in dry-air insulated DC systems, particularly in comparison with the established SF6 technology. Prior studies focused mainly on classifying defects under DC voltage using SF6. Similar to PRPD patterns in AC systems, the NoDi* pattern enables human experts to identify PD defects under DC stress. For automated monitoring, machine learning (ML) algorithms have achieved high classification accuracies of around 90% under SF6. The present contribution investigates the PD behaviour of defects in dry air, the classification by NoDi* patterns and the feasibility of ML-based classification of typical PD defects, including floating electrode, particles on insulator, free-moving particles, busbar corona, chamber corona and surface discharge. PD measurements were conducted under DC voltage in test cells filled with SF₆ and dry air at varying pressures.
The PDIV in dry air decreased significantly compared to SF6 for most of the PD defects, whereas the PD magnitudes in dry air increased or were equal to those in SF6. The NoDi* patterns showed clear differences in the case of busbar corona and a wire as free-moving particle, while for floating electrode and surface discharge, similar patterns were achieved. Despite these differences, the ML classification achieved an accuracy of around 90% for both gases, with only 27 features extracted from the PD pulse sequences.