Abstract:

SF6 gas has traditionally been used in Gas- Insulated Switchgear, but due to its extremely high Global Warming Potential, there is growing interest in alternative insulating media, such as Green Gas for Grid and dry air. As a result, there is an increasing need to develop new partial discharge (PD) diagnostic methods tailored to these alternative media, while also addressing the challenge of limited fault data. In this paper, using high-pressure dry air as an example, we propose a methodology for adapting existing PD diagnostic models—originally developed for atmospheric conditions—to high-pressure dry air, leveraging transfer learning techniques. The proposed method first transforms the raw data of measured applied voltages and partial discharge voltages into six relevant features for each time window through feature engineering. These features are then fed into the PD pattern diagnosis model, which consists of a feature extractor, a PD fault type classifier, and a domain discriminator. Feature discrepancy loss, including maximum mean discrepancy, batch-based instance separation, and batch-based feature decorrelation, is added to the loss function to optimize the model’s parameters. We evaluate the prediction performance under varying levels of data scarcity for high-pressure dry air switchgear. Additionally, we compare the estimation performances of transfer learning versus deep learning and discuss the transition point between these two approaches as the dataset evolves.