Research On Fault Diagnosis Method Of Distribution Network Equipment Based On Deep Learning

The rapid advancement of China’s industrialization process has led to the continuous expansion of the power system and the increasingly complex structure,which puts forward higher requirements for the safety,reliability and stability of the power system.As the end of the entire power system,the role of the distribution network is crucial.It is of great significance for the safe and reliable operation of power system to find and locate the faults of distribution network in time and accurately.Combined with new technologies such as data acquisition,Internet of Things and deep learning,the fault of distribution network can be judged efficiently,real-time and accurately,so as to provide scientific maintenance scheme and decision-making,and ensure the stable operation of power system.Therefore,based on the deep learning method,this thesis studies the fault diagnosis of box-type substations in key equipment of distribution network in the following aspects:(1)To improve the real-time performance and reduce the waste of transmission resources in cloud computing models for fault diagnosis,we introduce edge computing and establish a cloudedge collaborative fault diagnosis framework.An "End-Edge-Cloud" fault diagnosis mode is proposed,with the edge serving as the data processing and real-time application layer,and the cloud as the resource output and decision-making layer.This mode achieves real-time processing of data at the edge,efficient model training using cloud computing resources,and enables realtime diagnostic transfer to the edge.This mode reduces network latency and bandwidth consumption,computing load on the cloud,and emphasizes real-time response for edge fault diagnosis.(2)In practical application scenarios,data loss and collection anomalies often occur due to network packet loss or unstable collection devices.To ensure high-quality training data can be obtained at the cloud,research is conducted on data preprocessing at the edge.The missing value imputation and anomaly detection in the collected data are respectively carried out by using the missing forest algorithm and the isolation forest algorithm.The missing forest algorithm is used for completing anomaly data interpolation.Then,the Z-score method is utilized to normalize the data.Through edge data preprocessing,the adverse effects of missing and abnormal data on fault diagnosis data mining are avoided,such as incorrect gradient direction and loss of fault knowledge.It also unifies the data dimension,improves the efficiency of data mining,and improves the performance of fault diagnosis to a certain extent.(3)To address the problem of insufficient training and low diagnostic accuracy of fault diagnosis models caused by the sparsity of box-type substation fault data,an improved conditional table generative adversarial network(CTGAN)data derivative model with a selfattention mechanism is proposed.The self-attention mechanism is introduced in the generator of the CTGAN to maintain the correlation between the input data indicators.The differential evolution algorithm is used to optimize the hyperparameters of the data derivative model,such as the learning rate,number of iterations,and training batch size.By balancing the game between the generator and discriminator,as well as the conditional constraints of the classifier in the CTGAN,high-quality synthetic data can be generated from a small number of data samples,which provides sufficient data support for the training task of the fault diagnosis model.(4)When using deep learning network model to process one-dimensional vector input such as box-type substation monitoring data,the depth of the network is usually difficult to control.To address the issues and improve the feature extraction capabilities of fault diagnosis models,a Residual-CapsNet fault diagnosis model has been proposed.This model uses skip connections to facilitate residual learning of convolutional layers and improve feature extraction without impacting performance.Additional techniques like local capsules to maintain model view invariance and dynamic routing mechanisms for capsule layer updates help model strong correlations and achieve high-accuracy fault diagnosis.Finally,a box-type substation fault diagnosis system based on cloud-edge collaboration is designed and developed.The feasibility and effectiveness of the data preprocessing scheme,SACTGAN data derivation model and Residual-CapsNet fault diagnosis model proposed in this thesis are verified.It provides an important basis for fault diagnosis of distribution network equipment and fault maintenance and scientific decision-making of distribution network.