Research On Fault Detection Of Distribution Network Based On Improved Graph Convolution Network

Fault detection of distribution network aims to locate and diagnose faults through methods such as fault data analysis.Fast and accurate fault detection methods are very important for fault repair and power supply restoration.With the continuous expansion of distribution network scale,traditional fault detection methods cannot efficiently process massive distribution network data,which has problems such as low accuracy of fault location and slow speed of diagnostic.As a new artificial intelligence algorithm,graph convolution neural network has the ability to process big data and learn graph structure data.Because the topology of distribution networks is a natural graph structure,how to use graph convolution neural networks for fault detection has gradually received attention.This thesis takes single-phase ground fault and interphase short circuit fault in line faults of distribution network as the research object.According to the characteristics of fault data,a graph convolution neural network model ADGCN was proposed in this thesis.The main work and contributions are as follows:(1)Due to the large scale of distribution networks and the variety of fault conditions,traditional detection methods are difficult to extract fault features and cannot reasonably use the structure information of distribution network.To solve the above problems,graph convolution neural network is used to extract fault features independently,without the need to manually extract features based on experience.Graph convolution neural networks can fully consider the topology information of distribution networks through the feature learning method of convolution aggregation of adjacent node information.(2)In order to solve the insufficient processing ability of graph convolutional neural networks on disassortative and sample imbalanced graph data,as well as the unique over-smooth problem of graph convolutional networks,this thesis designed a graph convolutional network model ADGCN based on adaptive frequency and dynamic node embedding.The ADGCN model uses adaptive frequency aggregation functions to rationally utilize high-frequency and lowfrequency information in the graph signal,improving the classification performance of the model on disassortative graph data.Using a dynamic node embedding mechanism to alleviate the oversmooth problem allows the model to deepen the network layers to generate more discriminative node embedding.By adjusting the loss function,the model can pay more attention to categories with fewer samples and difficult to classify samples to address sample imbalance issues.Experimental results on a benchmark dataset demonstrate the effectiveness of the ADGCN model.(3)Using simulation software PSCAD,an IEEE33 node distribution network system is built and fault experimental analysis is conducted.According to the topological structure of the distribution network,the distribution network is abstracted as graph structure data by taking the lines as nodes,the connections between lines as edges,and measuring the three-sequence current and voltage of the lines as node characteristics.The ADGCN model is used for fault detection on the graph generated from simulated distribution network fault data.Multiple experiments such as model comparison experiments and adaptability experiments were designed.The experimental results show that the ADGCN model has better fault detection performance and the ability to alleviate over smoothing,and verify the effectiveness of the adaptive frequency function and sample balance strategy.In addition,in the adaptability experiment,the interference situation of distribution network data collection was simulated,and the interference experiment results showed that the model has strong anti-interference ability.