Research On Fault Line Selection Method Based On Convolutional Neural Network

The problem of fault line selection for small current grounding systems has long been a research hotspot.Because the fault current of each line of the distribution network is very small when a single-phase ground fault occurs,and it is easily affected by other factors,the current line selection method by manually extracting features is difficult to apply to various fault conditions.In order to improve the reliability and accuracy of single-phase grounding fault line selection for small current grounding systems,this paper transforms fault line selection into time series classification problems and proposes an end-to-end fault line selection algorithm based on deep learning.And based on the idea of integrated learning,a variety of line selection methods are fused to make full use of the fault characteristics of the system to achieve fault line selection.The work of this article is as follows:(1)By establishing a small current grounding system model and setting the fault,the data samples required for the training and testing of the line selection method are obtained.Using MATLAB / Simulink as the platform,10 k V distribution network simulation models with different topologies are built.The simulation parameters are set according to the outgoing line number of the fault,the initial phase angle,the distance from the fault point to the bus,and the size of the grounding resistance.Automated batch simulation to obtain the fault data set.For the N-outgoing model,the number of samples is 6000 N,and the training set and test set are divided according to the ratio of 5: 1.(2)Aiming at the fault line selection of small current grounding system,this paper proposes a neural network with two convolutional layers.The algorithm has excellent performance in the fault line selection task of distribution network with different number of outgoing lines.This model is the first convolutional neural network that directly acts on the current signal of the power system and performs single-phase ground fault line selection.(3)By further analyzing the characteristics of power system current,a new convolutional neural network model WDCNN is proposed based on a neural network with two convolutional layers.The first layer of convolutional neural network convolution kernel is a wide convolution kernel,and each subsequent convolution kernel is relatively small.In addition,the model uses batch normalization and convolutional layer dropout technology to improve the generalization ability.The experiment with the 4-outlet model as an object has a certain improvement in the accuracy of fault line selection compared with the two-layer convolutional neural network model.In addition,the difference between the anti-noise ability and the generalization ability of WDCNN and other algorithms that rely on manual feature extraction is compared,and it is concluded that WDCNN has better performance in fault line selection tasks.Finally,using data visualization technology,the process of fault line selection by the convolutional neural network model is shown.(4)A fault line selection algorithm is proposed which combines the results of WDCNN and other models.The model selects different kinds of neural networks,and uses three-phase current and zero-sequence current as the input of various models,and fuse the results obtained from each model.Experiments were conducted using different fusion methods.After testing and comparison,the optimal model fusion strategy was selected.After the model fusion,compared with the WDCNN model only,the accuracy and anti-noise ability are improved.The various experiments done in this paper are all dependent on the simulation data,which can provide a reference for the construction of the actual application of the line selection method.