Transmission Line Galloping Curve Reconstruction Based On Conditionalgenerative Adversarial Networks

As the main means of long-distance power transmission,the operation status of overhead transmission lines is closely related to the stability of power system.By the end of 2018,more than 1000 conductor galloping accidents occurred in China that voltage levels affected ranged from 35 kV to 1000 kV.Those accidents caused heavy economic losses of tens billions RMB.Therefore,the research of real-time transmission lines on-line monitoring is imminent.Based on the research background of transmission lines galloping curve visual reconstruction,in order to solve the shortcomings such as high operation and maintenance cost and large errors of reconstructed signal in the previous monitoring schemes,a new galloping curve reconstruction scheme based on conditional generative adversarial network(CGAN)is proposed.Start with the network structure design and loss function optimization to achieve low sampling rate reconstruction of the galloping curve.The research content specifically include the following two directions.One scheme is the transmission line galloping curve reconstruction based on deep convolution conditional generative adversarial network.Taking the spatial distribution of each point on the transmission line during galloping and the whole galloping state change between different moments as the joint feature information.With GAN’s excellent fitting ability for unknown data distribution,and the transmission line galloping reconstruction can be combined with the signal reconstruction at low sampling rate.The galloping data model is the Yao double line three-split conductor in the Zhongshan crossing of Hubei province.The basic model architecture is deep convolution network(DCGAN),which composed of "Convolution Layer-Batch Normalization-ReLU Activation" and the other similar convolution units,it’s used to extract data features.Sampling signal is used as input of generator in the model to obtain corresponding reconstruction data.Sampling signal and original signal are used as conditional mapping relation and discriminator to optimize the generator’s performance.Finally,after the training,the complete state of transmission line galloping model with spacing of 1055 meters could reconstructed by only 9 sampling points.The other galloping curve reconstruction scheme is based on an improved W-CGAN model with arc length and curve smoothing joint constraints.In order to solve the problem of large reconstruction errors caused by the gradient disappearance of the cross entropy loss function in the training process,we proposed a conditional generative adversarial network model based on Wasserstein distance.The Wasserstein distance has excellent gradient performance and can improve the model reconstruction performance.In addition,with the single-time curve galloping reconstruction fluctuation feature and the time-domain curve feature at multiple moments,a joint constraint loss function that based on arc length and curve smoothing is proposed.At the same time,the generator network structure removes the batch normalization operation and selects the maximum pooling layer,so that the convolution unit becomes the form of "Convolution layer-Convolution layer-Maximum Pooling layer-ReLU activation".The simulation results show that compared with the simple W-CGAN algorithm and the classical cubic B-spline interpolation reconstruction algorithm,the reconstruction performance and average error are improved and optimized.