Xie Yaoheng

Severity diagnosis and assessment on defects in GIS based on partial discharge detection

In complex high voltage components local insulation defects can cause partial discharges (PD), especially in highly stressed gas insulated switchgear (GIS) these PD affected defects can lead to major blackouts. PD detection is recognized as one of the most effective and most important methods of insulation condition assessment in GIS. In this paper, we designed a scheme to diagnose and assess the severity levels of the PD provoked by defects in GIS. With the application of gradually increased voltage, long-term tests were conducted on a well-established 252kV GIS experiment platform to observe the entire evolution process of PD from its very initiation till the eventual flashover. Ream-time measurement was undertaken during the tests to capture the trend curve of as a result of test time, including the scatter plot, histogram, grey-scale map, etc. The results indicate that PD initiated by defects in GIS can be classified into three severity levels, namely, petty discharge, medium discharge, and threatening discharge. Moreover, on the basis of the features of phase distribution and the corresponding spectra, a procedure based on k-means cluster analysis are proposed to diagnosis and assess severity of PD automatically.

A robust discrimination method for inrush current based on neural network

Discrimination between internal fault and magnetizing inrush current in transformers is widely studied in recent years. A robust method that can resist strong noise and disturbance is proposed in this paper, which can identify inrush current waveform fast and efficiently. Since discrimination of internal fault current and magnetizing inrush current is essentially a pattern classification problem, a feedforward neural network is applied as the classification model in proposed method. As the most effective feature vector, the harmonic information is firstly extracted from the recorded fault data, then the feature vector is sent into the well-trained neural network and the output of the neural network is a Boolean value that indicates the identification result of the waveform. Experimental data proved that the proposed algorithm is robust and flexible. By comparing with existed algorithms, our method shows better robustness with respect to strong noise and disturbance and less dependency on prior knowledge.