| Distribution network plays a vital role in the whole network.In the grid,DC transmission is the most important one,and also the most important one.However,the interference of the outside world and improper operation of personnel will bring great negative impact to the distribution network.The structure of single-phase high resistance distribution network is complex and its transient process is often unstable,which makes it difficult for traditional devices to monitor and operate effectively.In order to reduce the probability of fault,the fault location,fault type and grounding protection are studied in detail.Aiming at this problem,this project intends to carry out the research on feature extraction and identification of high impedance single phase fault.Firstly,according to the characteristics of high resistance fracture,the sensitive zero-sequence current of high resistance fracture is selected and analyzed.Secondly,Emanuel model is introduced into distribution system to model high resistivity grounding fault in order to solve the problems of nonlinear deformation,high harmonic content and short circuit current.On this basis,for 10 k V distribution system,the variation of zero-sequence current under the initial phase angle,fault location and instantaneous resistance is studied.Through the statistics of the fault data of 10 k V distribution network in a city,the correlation analysis is carried out.A new method based on CNN-SVM is proposed for zero-sequence high resistivity fault,which is trained by CNN-SVM without extracting the feature vector manually.On this basis,we will introduce the entropy loss function of mutual information and train the model by selecting the optimal deep learning parameters.In order to solve the problem that convolution neural network only relies on one dimension image and can’t solve any dimension image and subgraph efficiently,this project intends to design and construct multi-scale perception network.In order to preserve the semantic and features of the final convolution neural network,the higher-order Runge-Kutta method is applied to the convolution neural network.Experiments show that the proposed algorithm can effectively reduce the iterative number of the model,improve the accuracy of the model identification,and have a better noise suppression.Aiming at the problems of difficult samples,large amount of data and high level in deep learning,this thesis proposes an improved residual network and unbalanced data enhancement technology to classify unbalanced data.Firstly,the DCGAN algorithm based on deep learning is studied,and the deep residual network is constructed to extract the image features of distribution system. |
