Partial discharge severity assessment in gas-insulated switchgear using unsupervised feature learning algorithm

Summary

Gas insulated switchgear (GIS) is used widely in power system for its excellent performance, and it’s meaningful to monitor its operation Status. Evaluating the severity of partial discharge (PD) in GIS is an important part to learn about the insulation condition of GIS. In this study, the step voltages method is applied to simulate the development of partial discharge under typical protrusion defect in experimental platform, and the ultra-high frequency (UHF) PD signals in various PD stages are collected and converted to f–u (discharge phase and discharge amplitude) and f–n (discharge phase and discharge times) spectrograms. The test results suggest that discharge times, amplitudes, and phases showed regular changing trends with the development of PD. But it is not convincing to describe PD severity only by discharge times, discharge amplitude and discharge phase in currently study. Hence, an unsupervised feature learning mode – sparse auto-encoder (SSAE) is proposed to realize the feature extracting for PD severity assessment. The soft-max classifier and SVM classifier are employed to identify the PD severity under protrusion defect based on the SSAE model features and traditional statistical features. The comparison test of PD severity assessment accuracy based on the features extracted by SSAE model and the statistical features indicate that the features extracted by SSAE model can effectively characterize the severity of PD. The recognition performance of SSAE algorithm was found to be better than SVM algorithm based on statistical features with average assessment accuracy up to 86.6%. In above, the proposed unsupervised feature learning algorithm is proved to be effectively to learn the features from raw PD data, and it is a good practical application value when we are lack of prior knowledge about the PD development process.

Additional informations

Authors

MANTILLA, PAWAR, RABIE, SPENCER, LIA

Keywords

Gas-insulated switchgear; partial discharge development; feature learning; assessment