Fault detection of VSC based DC distribution network based on CNN and DCGAN

Coal mine DC power supply technology will greatly improve the reliability and safety of coal mine power supply due to its excellent power supply performance. However， because the power electronic equipment used in this technology has poor ability to withstand fault impulse current，its protection needs to utilize transient information within a few milliseconds identifying the fault area，which brings new challenges to the traditional AC grid relay protec⁃ tion. Aiming at the problems of low accuracy and weak robustness of DC grid fault detection，this paper proposes a new method for flexible DC distribution network fault detection based on the convolutional neural network (CNN) and deep convolutional confrontation generation network (DCGAN). First，the ensemble empirical mode decomposi⁃ tion method is used to decompose the transient current to obtain several intrinsic mode function (IMF) components，the correlation coefficient of each IMF is calculated，and then a new transient current signal is reconstruc⁃ ted. By taking the signal value through the sliding window， then combining it， a two⁃dimensional image is ob⁃ tained through signal⁃image conversion. The image is divided into a test set and a training set. Then，the training set is trained using DCGAN to train the discriminator and generator in turn，and the true value is approximated by multi⁃ ple trainings. The enhanced samples obtained in the generator are used as the expansion of the training set. The train⁃ ing set and the training set generated by DCGAN are used to train the network，and the CNN is further used for fault detection judgment. Finally，the impact of the sample enhancement on the recognition result and the effect of the size of the convolution kernel，the pooling method and the activation function on the network performance verifies the effectiveness of the algorithm proposed in this paper. The test results show that the four different working conditions analyzed in this paper can have high detection accuracy，and the average detection accuracy is 95.044%. For the im⁃ age with a resolution of 44×44，the detection speed can reach 20 frames / s.