Research On Fault Location Of Flexible DC Distri-Bution Network

Flexible DC distribution is the key technology to support the clean and low-carbon transition of power grid because of its strong power supply capacity and good controllability,especially its excellent compatibility with distributed generations and DC loads.However,due to the changes of distribution system and the high-permeability of power electronic devices,the temporal and spatial evolution mechanism of flexible DC distribution network’s faults based on MMC converters is more complex,leading to the inability of existing methods to achieve accurate fault location and limiting the development of DC distribution technology.To solve the problem that the fault information is limited and difficult to extract,a high-precision fault location algorithm based on fault feature extraction and convolutional neural network is proposed by combining signal processing with deep learning.MMC inverter’s control strategy and the key equipment are theoretical analyzed,and the simulation model of DC distribution network with power supply at two-ends is constructed.Then the distribution line between two converter stations is taken as the research object,and the characteristics of the line fault are analyzed.The problem is that the pole-to-ground fault circuit of sub-module’s capacitance and transient resistances are unknown and the fault location cannot be directly calculated by fault data.Instead,the mapping relationship between fault frequency-domain features and fault location is sought.The theory and process of fault feature extraction are presented.According to the feature distribution law,abundant fault features are selected independently by signal processing based on wavelet transform,and the influence of noise is eliminated while the input is increased to obtain the dual features in time-domain and-requency domain.The structural characteristics of convolutional neural network and the basic principle of key units are briefly analyzed,and the updating formula of network parameters during training is given.The final process of the fault location algorithm is as follows:the fault voltage signal is decomposed by discrete wavelet transform with multi-resolution,then the transient energy function is constructed,and the fault characteristic with rich frequency band is selected for signal reconstruction.Reconstructed signal through continuous wavelet-transform forms 2D time-frequency gray image,the image after image enhancement is as a convolution of the neural network input.Relying on convolutional neural network generalization ability and powerful,the mapping relationship between fault features and fault locations independent of device parameters and control strategies is established.Finally,this paper achievs 30m of fault location precision in distribution line with 1.5 km,only a single end measurement point needed,effectively reducing equipment investment and operation and maintenance costs.A large number of simulation results show that the fault features extracted by this method can accurately reflect the fault location,and overcomes the influence of the converter control strategy and the number of sub-module’s switching capacitor.It still has a high accuracy in the distributed photovoltaic output fluctuation.In addition,the comparative test of different algorithms further highlights the superiority and anti-noise ability of the algorithm in paper.