Research On Intelligent Identification Method Of Internal Overvoltage In Distribution Network

Distribution network is the terminal of transmission network in power system,whose stability and reliability has influence on the safety of power users,the productivity of the industrial sector and the stable development of the society directly.The distribution network has a large grid structure and contains a large number of equipments.Besides,it's insulation level is low.According to statistics,70%-80%of overvoltage faults in power system have emerged in distribution network.It is significant to research on the methods of internal overvoltage identification in distribution network to improve the insulation level and reduce the probability of overvoltage.The research directions in the field of extracting features and overvoltage identification at home and abroad are summarized in this paper.By compareing with other algorithms,it is proposed that Hilbert-Huang Transform(HHT)has superiority in dealing with nonlinear and unstable signals,and deep learning has potential when applyed in overvoltage identification.The process and principle of 9 kinds of internal overvoltage are analyzed in detail.The transient and steady state process exist together when internal overvoltage happens.In order to select the proper characteristic quantity,the idea that means extraction combining transient and steady feature is determined.The implementation steps of HHT band-pass filte algorithm are listed,the noise reduction characteristics of the method are pointed out in an example.Singular Value Decomposition(SVD)theory is used to extract the singular values of time-frequency matrix as the features of overvoltage signals.After setting forth in support vector machines(SVM)and convolutional neural network(CNN),an intelligent 8-layer SVM and a 9-layer CNN are designed to recognize internal overvoltage faults in distribution network.Two methods are proposed in this paper to identify internal overvoltage in distribution network.The first method uses HHT band-pass filter algorithm to decompose the three-phase voltage waveforms with different bandwidths to construct time-frequency matrices.Then SVD is used to analyze the matrix to get singular values as features.Finally the parameters of multistage SVM are optimized using the training samples and the test samples are used to verify the accuracy of the SVM method.The second method process the time-frequency matrices of frequency bands to calculate the energy blocks in time domain and the time-frequency energy spectrum is formed.The time frequency energy spectrum is input into the 9 layer CNN model for training and testing.A 10kV neutral ungrounded software system of distribution network is built and 9 kinds of internal overvoltage are simulated using PSCAD/EMTDC software.MATLAB software is used to analyze the overvoltage data to compare the accuracy and speed of overvoltage identification in distribution network between Method 1 and Method 2.It is showed that the method based on time-frequency energy matrix and CNN are better than those based on SVD and multilevel SVM in all aspects.Single phase overvoltage data in physical simulation system and the sample data in different software system are used to validate Method 2.It is found that the total recognition rate of the sample data is over 96%.