基于CNN与DCGAN的柔性直流配电网故障检测

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

煤矿直流供电技术凭借其优良的供电性能将大大提高煤矿供电的可靠性和安全性,但由于该技术所采用的电力电子设备耐受故障冲击电流能力差,保护需利用数毫秒内的暂态信息识别故障区域,给传统交流电网继电保护带来了新的挑战。针对直流电网故障检测正确率低、鲁棒性弱的问题,提出了一种基于卷积神经网络(CNN)与深度卷积对抗生成网络(DCGAN)的柔性直流配电网故障检测新方法。首先利用集合经验模态分解方法将暂态电流分解,得到若干个本征模态函数(IMF)分量,计算各个IMF的相关系数,并重构成新的暂态电流信号;通过滑动窗口取值、信号组合,将其经过信号-图像转换变为二维图像,该图像分为测试集和训练集;接着,将训练集利用DCGAN通过轮流训练判别器和生成器,经多次训练逼近真实值,从生成器中得到增强样本作为训练集的扩充;并将训练集与DCGAN生成的训练集来训练网络,进一步利用CNN进行故障检测判断;最后,从样本增强对识别结果的影响,以及卷积核大小、池化方式和激活函数等方面对网络性能的影响进行仿真测试,验证了所提算法的有效性。测试结果表明,该方法所分析的4种不同工况下都能有较高的检测精度,且平均检测精度为95.044%,对于分辨率为44像素×44像素的图像,检测速度可达20帧/s。

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 protection.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 decomposition 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 reconstructed.By taking the signal value through the sliding window,then combining it,a two-dimensional image is obtained 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 multiple trainings.The enhanced samples obtained in the generator are used as the expansion of the training set.The training 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 image with a resolution of 44×44,the detection speed can reach 20 frames/s.