Advancements in Partial Discharge Risk Estimation through AI

Partial Discharge (PD) monitoring is increasingly utilized in critical equipment within our high voltage network. For Gas Insulated Switchgear (GIS) and Transformers, the UHF method is frequently preferred due to its superior signal-to-noise ratio at high frequencies compared to the standard IE60270 method. Additionally, the heightened signal damping at high frequencies assists in locating the PD source. PD monitoring systems commonly incorporate Artificial Intelligence (AI) to automatically identify signal root causes and prevent false alarms triggered by external PD or noise. The next logical advancement in PD evaluation is PD risk assessment, and this study aims to contribute to this area. Machine learning algorithms can be trained to predict whether an actual PD measurement is detected at PD inception voltage (presumably less risky) or at a significantly higher voltage level (potentially indicating higher risk). This paper employs state-of-the-art AI toolboxes to identify regression algorithms capable of automating the prediction of the relative voltage level above PD inception for five types of GIS PD faults: Moving Particles (on AL and Paint), Particles on Insulation, Floating Electrode, Voids in Insulators, and Protrusions. It is demonstrated that regression models can effectively be implemented for all fault types with a good degree of accuracy, with the key to success lying, as with all AI approaches, in having a robust and extensive dataset to learn from.