Research On Fault Diagnosis Of AC-DC Hybrid Power Grid Based On Deep Learning

China has built the largest AC-DC hybrid power grid with the most complicated operating conditions in the world.The AC-DC hybrid power grid has strong power transmission capacity,which greatly alleviates the uneven distribution of energy and load in China.It has become a new form of power grid development.Due to the complex structure of the AC-DC hybrid power grid,when a fault occurs somewhere in the line,it is very easy to develop into a large-scale accident.The safe and stable operation of the AC-DC hybrid power grid has always been a key concern in the power system field.This thesis focuses on the fault diagnosis of the AC-DC hybrid power grid.On the basis of existing fault diagnosis methods,the thesis focuses on the fault diagnosis method of AC-DC hybrid power grid based on deep learning.Firstly,the thesis analyzes the current research status of power grid fault diagnosis at home and abroad,introduces the fault characteristics of AC-DC hybrid power grid,summarizes the smart algorithms commonly used in power grid fault diagnosis,and analyzes the advantages and disadvantages of the existing fault diagnosis methods,and mainly puts forward the research on fault diagnosis of AC-DC hybrid power grid based on deep learning methods.Secondly,the thesis introduces the basic composition and operation mode of the AC-DC hybrid power grid,and analyzes the main fault conditions of the AC-DC hybrid power grid.Then the thesis builds an AC-DC hybrid power grid model in MATLAB/Simulink software.By simulating different fault conditions,a batch simulation method is used to obtain a large number of fault samples containing multiple fault conditions,and the fault data samples are preprocessed.These fault data samples are used to verify the fault diagnosis capability of the algorithms proposed below.Subsequently,a power grid fault diagnosis method based on the deep learning convergence network(DLCN)is proposed.The deep learning convergence network forms by stacking the low-level autoencoders and the high-level convolutional neural network.It effectively realizes the fault characteristics extraction of the network.DLCN forms the corresponding relationship between sample characteristics and fault conditions,it reduces the influence of human factors on fault diagnosis result.The test result shows that the proposed method has strong feature extraction ability for power grid fault samples,it can accurately diagnose the fault lines and fault types corresponding to the power grid fault samples,and realizes the fault diagnosis of the power grid.Then,the thesis proposes a power grid fault diagnosis method based on gated recurrent unit(GRU)network.It analyzes the advantages of GRU applied to power grid fault diagnosis from the perspective of time sequence,and debugs the network parameters of GRU by network training.The network classification result of the test set samples shows that GRU has strong fault diagnosis ability for power grid fault samples,its fault diagnosis accuracy is better than other recurrent neural networks,and is hardly affected by different fault conditions and the size of the transition resistance.Finally,the thesis studies the fault diagnosis of AC-DC hybrid power grid based on transfer learning method.The mentioned deep learning method above forms the optimal network parameters after training with a large number of input samples,but there are few fault samples in the actual power system,and it is difficult to train the deep learning network to the optimum.In this regard,a fault diagnosis method using deep learning networks for power grid with fewer samples is proposed: the trained DLCN is regarded as the source domain network,and the transfer learning method is adopted to transfer the DLCN parameters in the source domain to the target domain,which is used to carry out fault diagnosis research on another power grid model.The test result shows that the transfer learning method based on DLCN parameters also has excellent fault diagnosis capability for other power grid models.The proposed method provides a new idea for applying deep learning to power grid fault diagnosis with a small number of samples.