Research On Methods Of Feeder Fault Type Identification In Distribution Network

As the last link to users in power system,distribution network must operate safely and stably to guarantee electrical safety,reliability,benefit of users.There are more than 90%power faults emerging in distribution network.After faults happening,it is very dependent on fault type identification timely and accurately to deal with accident including distance measuring,location,exclude on fault zone,reconstructing fault net,accident analysis,etc.Therefore,it is of significance to research on methods of feeder fault type identification in distribution network.First of all,the research directions in field of power system fault identification home and abroad are summarized in this paper.Developments on methods of fault identification in distribution network are outlined around topics both of extraction methods of fault feature and choice of intelligent classifier.After that,both superiority of Hilbert-Huang Transform(HHT)in dealing with nonlinear and unstable signals and potential of deep learning applying in fault identification is emphasized highly.Based on transient and steady analysis on fault feature in distribution network,the idea that means extraction combining transient and steady feature is determined.And then,noise reduction and flexibility of HHT band-pass filter is detailed in an example of signal decomposition.Theory of singular Value Decomposition(SVD)is introduced to explain that singular value is suitable as fault feature because it can express the modal characteristics of signals.After expounding theories of support vector machines(SVM)and convolutional neural network(CNN),a nine-level SVM and a seven-layer CNN are designed to classify fault type in distribution network.Based on above analysis results,two methods of feeder fault type identification in distribution network are proposed in this paper.The one is a method based on SVD and multi-level SVM.First,7 fault waves including three phase and zero sequence voltage of bus bar,and three phase current on main transformer second side are pretreated by HHT band-pass filter algorithm.And 7 time-frequency matrixes are reconstructed.After that,the time-frequency matrixes are decomposed by SVD and parts of normalized effective singular values are extracted to be characteristics for training and testing SVM.The other one is a method based on block time-frequency spectrum and CNN.Reconstructed time-frequency matrixes are processed to get block time-frequency spectrums by calculating energy block.The normalized block time-frequency spectrums are input images for training and testing 7-layer CNN.Two models of lOkV distribution network are constructed to acquire samples for training and testing.The one is constructed by PSCAD/EMTDC.And another one is a model of physics experiment based on distribution network dynamic simulation system.The test results show that identification accuracies of two methods are excellent for identifying 10 fault types including AG?BG,CQ ABG,ACG,BCG,AB,AC,BC,ABC in distribution network.What's more,two methods behave good adaptability in circumstances of noise,asynchronous sampling,distributed power access,changes of network structure,load current,system equivalent impedance,and arc suppression coil grounded system.And two methods behave excellent sample compatibility in mixed sample test.It is no need to construct and extract characteristics artificially for the method of feeder fault type identification in distribution network based on CNN.And the method based on CNN is more excellent than another method in terms of robustness and adaptability.