Research On Spatial Domain Image Fusion And Fault Line Selection Method Of Small Current Grounding System

Single-phase grounding fault line selection for small current grounding systems has always been a difficult problem in the power sector,especially for resonant grounding systems.Traditional line selection methods still need to be improved in the aspects of line selection accuracy and model adaptability.At present,artificial intelligence technology has been implemented in many fields such as medical treatment,security,machinery and so on through image recognition,and the application scenarios are becoming more and more abundant.At the same time,the digital informatization process of my country’s power grid has continued to deepen,and power companies have successively proposed the construction goals of "Smart Grid" and "Ubiquitous Electric Internet of Things".This shows that the use of a new generation of artificial intelligence technology represented by deep learning to solve the problem of fault line selection has the realistic conditions and actual needs.In order to make full use of the advantages of deep learning in the image field,this paper explores a new way to turn fault line selection into an image processing problem.At the same time,it combines deep learning algorithms and transfer learning strategies to achieve fault line selection.The main contents of this article are as follows:(1)The distribution network model of resonant grounding system is built,and the fault data is generated in batches.First,the main factors affecting fault line selection are determined by theoretical analysis.Then,distribution network models with six different topological structures are built through MATLAB/Simulink.Finally,the fault data of different distribution network structures are divided into training/verification/test data sets according to the ratio of 0.6/0.2/0.2.(2)According to the characteristics of three-phase current signals in threedimensional space,a method for generating single-phase ground fault images in space domain is proposed.In this method,the images of the three-phase current in the threedimensional space are respectively projected to the AOB,BOC and AOC planes,and three projection images can be obtained for each feeder.This method not only preserves all the fault information of the fault process completely,but also realizes the transformation from 3D time domain current signal to 2D image.(3)According to the mechanism and advantages of deep learning algorithm for image recognition,a two-dimensional image fusion method for single-phase ground faults is proposed,and the feasibility and superiority of the line selection method are verified.The core of this method is the weighted average fusion of pixels.In order to reduce the complexity of the deep learning network,two pixel-level image fusion processing strategies are adopted to achieve the goal of generating only one RGB color image when a fault occurs.Taking the four-outlet distribution network as the line selection object,the influence of different factors on the fault characteristics of fusion image is compared,it proves the superiority of the method.Taking the fault fusion image as input,the convolutional neural network with four convolutional layers as the main body,through the visualization of the line selection algorithm,the line selection process is intuitively displayed,and high line selection accuracy has been achieved under different test conditions,Verified the feasibility of the image fusion line selection method.(4)In view of the structural diversity of the actual power grid,a depthwise separable convolution algorithm is used to realize the lightweight of the line selection network,and the transfer learning strategy is used to enhance the adaptability of the line selection network.Firstly,the depthwise separable convolutional neural network is pre-trained with the fault data set of six-outlet distribution networks as the source domain data set.Then the transfer learning strategy is used to transfer the pre-trained line selection network to the other five different distribution network structures.A small amount of training data can be used to realize the training of another five line selection networks with different distribution network structures,and the accuracy of the transfer line selection network meets the line selection requirements.It shows that the method in this paper can effectively enhance the adaptability of the line selection network and can provide a solution for the smaller amount of data in the training process of the line selection network.