Fault Detection Of Flexible DC Distribution Network Based On Convolutional Neural Network And Deep Convolutional Confrontation Generation Network

Through the close integration of electronic power technology and flexible DC distribution network,the flexible DC distribution system can gradually transition from partial DC to full DC power supply.Due to its excellent power supply reliability,flexible DC power supply technology greatly improves the reliability and safety of flexible DC power supply,and is one of the important development directions of future power distribution systems.Aiming at the problems of low accuracy and poor robustness of fault detection in DC power grids,related research on fault detection and identification of flexible DC distribution networks is carried out.(1)First,the current situation of fault detection and identification of flexible DC distribution network is expounded,then the structure,operation principle and control method of flexible DC distribution network are studied,and finally different faults of flexible DC distribution network are studied.(2)Aiming at the problem that the fault information of the flexible DC distribution network is not obvious,and the fault detection accuracy of traditional methods is low,this paper proposes a new method based on Convolutional Neural Networks(CNN)and deep convolutional adversarial generation network(Deep Convolutional Neural Networks,CNN).A new method for fault detection in Convolutional Generative Adversarial Networks(DCGAN).Firstly,the fault transient current is decomposed by Ensemble Empirical Mode Decomposition(EEMD).After EEMD decomposition,the fault transient current can obtain several intrinsic mode function(IMF)components,and then set The threshold is set to 0.25.By calculating the Pearson correlation coefficient and the Spearman correlation coefficient,the IMF components greater than 0.25 are selected.At this time,the selected IMF components contain the most fault information,and then the selected IMF components are used for waveform reconstruction,reconstruct a new one-dimensional transient waveform;convert the one-dimensional transient waveform into a two-dimensional image through sliding window and signal combination,and then divide the transformed two-dimensional image into training set and The test set is tested by setting network structures such as different convolution kernel sizes,different pooling layers,and different activation functions.However,the recognition accuracy of CNN may decrease under small samples and lead to DCGAN network to solve this problem.(3)Aiming at the defect that the CNN model needs a large number of training sets to improve the accuracy,this paper proposes a fault detection algorithm for flexible DC distribution networks based on CNN and DCGAN.The role of DCGAN is to generate a generated image with the same distribution as the original image after passing the limited training sample set through the discriminator and generation of DCGAN,and then increase the number of training sets.At the same time,aiming at the synchronization requirements of the generator and the discriminator,a control threshold method is proposed to adjust the synchronization relationship between the generator and the discriminator according to the requirements during training,so as to improve the stability of the network and speed up the convergence speed of the model.And through the experimental part,it is verified that the recognition accuracy reaches 96.044% after adding the threshold control.This verifies the reliability of DCGAN in the case of small samples.