The Study On Statistical Feature Optimization And Patter Recognition Of Partial Discharge In GIS

Gas Insulated Substation (GIS) is widely used in urban power substation becauseof its advantages of small area coverage, operational reliability and low electromagneticpollution. However, due to some reasons in structure and transportation, there areinevitable insulation defects which may cause partial discharge in regular operation. Thesafe operation of grid is threatened seriously by insulation damage. Therefore, theon-line monitoring and fault type recognition in GIS has been the focus of research inthis field.In this paper, statistical characteristics of different insulation defects PD signals arestudied based on analyzing researches about PD type recognition home and abroad.Support Vector Data Description method is presented for PD type recognition in GIS.The main work and achievements are as follows.①Large data of discharge experiment in different intensity is acquired by PDUHF monitoring system which has been developed for the detection of four kinds oftypical insulation defects model in the laboratory. The φ-u-n3D PD image and thecorresponding φ-u, φ-n two-dimensional image has been constructed. The results showthat: the differences of four PD image shape are rather obvious, the same type of PDimage shape remains unchanged in different voltage levels. Obtaining the13statisticalfeatures based on the two dimensional image can lay a foundation for the research ofPD type recognition.②The KPCA method is presented for the extraction of feature subset of fourkinds of insulation defects discharge comprehensive characteristic. The MaximalRelevance Minimal Redundancy method is put forward to obtain different characteristicparameters for dimension reduction of the original statistical characteristics.Fromdifferent angles of partial discharge pattern, two algorithms are combined to constructoptimal feature subset TBESTto identify four kinds of defects, which effectively retainedthe characteristics of each class of the PD signal and reduced the redundancy ofstatistical characteristics.③SVDD is introduced into PD type recognition, based on the principle ofMaximum interval of support vector machine and one to multiple of multipleclassification method, an optimal radius support vector data description algorithm(OR-SVDD) is proposed to solve the disadvantages of missing and wrong classification in SVM and the lack of classification margin small in SVDD. The principle of“one-to-many” is adopted to solve the difficult problem of multi-class defectclassification, and to improve the identification ability and application value.④The result that analysis the effects of different features obtaining method forfour kinds of PD type recognition shows that TBESTgets the optimal recognition ratebecause of considering the characteristics of each defect discharge type based ondimension reduction of original statistical features in the KPCA method. It is provedthat OR-SVDD algorithm is better than SVM or SVDD, the average recognition ratecan reach87.9%. In addition, the correctness and effectiveness of the OR-SVDDmethod is verified through analysis of the PD image of four kinds of discharge model.