Research On Identification Of Single-phase Grounding Fault In Distribution Network Based On Deep Neural Network

In China,the distribution network can keep operating for 1~2h even if single-phase grounding fault occurs.However,with the development of the city,the scale of modern distribution network has expanded yearly.The length and number of feeder line increases,which significantly contributes to the increase of capacitive current of the system.Under this situation,the single-phase grounding fault that occurs in the distribution network,should be removed in time.Otherwise,the fault current amplitude is too large to be extremely likely to cause further development of the fault and pose huge threats to safe operation of distribution network and threatens life safety of citizens.Therefore,it is necessary to ascertain the cause of the fault timely and take corresponding measures to cut out the fault.Distribution network fault identification is beneficial to identifying the cause of the network breakdown,and is conducive to taking appropriate action to renovate the broken distribution network.Meanwhile,fault identification is also a prerequisite for fault line selection.Therefore,fault identification is of great significance for ensuring the safe operation of the distribution network,and thus,it is of great research value.This thesis studies the theoretical model of distribution network fault in China,and then exerts MATLAB/Simulink to build a single-phase grounding fault simulation model of small current grounding system with 10 k V voltage level.On this basis,the thesis focuses on the simulation and fault models of three kinds of arc grounding faults and high-resistance grounding faults.Considering the complexity and dynamic characteristics of the arc grounding fault in the actual operation of the distribution network,this thesis establishes two kinds of arc fault models,one with even arcing interval and another with uneven arcing interval,in order to simulate the actual arc grounding fault occurring in the real world to the full extent.The simulation model traverse different fault condition to simulate various single-phase grounding faults occurring in the distribution network,and then gathers the zero-sequence voltage of the neutral point.Combined with wavelet transform theory,wavelet packet transform theory and entropy theory,this thesis analyzes the zero-sequence voltage signals to obtain each layer's modulus maxima of the wavelet decomposition coefficients,to extract wavelet band energy,wavelet energy entropy,wavelet frequency entropy,wavelet entropy weight,wavelet packet band energy,wavelet packet energy entropy,wavelet frequency entropy and wavelet packet entropy weight.Based on the fault characteristics extracted before,this thesis proposes a method of grounding fault identification based on modulus maxima and wavelet entropy,and then tests its validity.The test results show that this method has excellent fault identification precision,which proves the correctness and effectiveness of this method.This thesis studies the theory of deep neural network and applies it to fault identification of distribution network.The combination of various fault characteristics extracted before are inputted into deep neural network to train the deep network and identification effect is examined by test samples.The identification effect of the deep network based on different fault characteristics is compared with each other.The identification effect of deep neural network is compared with the recognition effect of the method based on the modulus maxima and wavelet entropy characteristics,and is also compared with the identification effect of traditional neural network.Adding noise to the zero-sequence voltage signal and putting its fault characteristics into the network for identification testing,this thesis compares the identification result of the deep network with the traditional neural network's.The comparison shows that the fault identification method based on the deep neural network algorithm is more accurate.At the same time,the generalization ability of deep network is much stronger.