Research On GIS Partial Discharge Pattern Recognition Technology Based On UHF Method

Gas Insulation Switchgear(GIS)is widely-used in 60kV and above level's power grid for its compactness,reliability and durability.However,insulation defects will develop inevitably during manufacturing,delivering,installment and running process of GIS,which can bring about Partial Discharge(PD).PD is not only a sign of insulation defects but also an incentive of further insulation deterioration.A series of physical and chemical changes accompanied with PD are principal references for detection of PD.Main detection methods can be classified into electrical,vibrational and spectral detection.Ultra-High-Frequency(UHF)method is recognized as a mainstream detection method due to its high sensitivity and anti-interference ability.Considering that distinctive PD patterns represent different insulation defects as well as different effect on insulation damage,PD pattern recognition is regarded as a premise of GIS insulation evaluation and conjunct maintenance plan.A PD online detection system was designed to conduct PD experiments before 100 sets of UHF PD signal data was collected respectively corresponding to four typical partial discharge pattern.A comparison was made between EMD(Empirical Mode Decomposition)and Pulse Clustering Analysis based on the collected data to denoise the signal.Effective feature extraction of PD UHF signal is crucial in guarantying the recognition accuracy rate of a classifier.By referring to relevant documents,the author selected 7 dimensions of feature parameters as input of the classifier.Classifiers based on SVM(Support Vector Machine)and K-means Cluster was also designed in this article.By inputting proposed 7-dimentional feature parameters,the effectiveness of these parameters on the pattern extraction of different PD UHF signal as well as the accuracy of proposed classifiers was verified with accuracies of over 90%.A performance analysis was made on the proposed method,then improved K-median Cluster and SVM with optimal kernel parameter and L2 regularization was applied with accuracies of over 93%.At last,Principal Component Analysis(PCA)was used to reduce the dimensions of the data on the premise of a 95%variance.