| With the advantages of high modularity,good expandability and high output quality,modular multilevel converters are widely used in high-voltage direct current transmission,power system power quality management,high-power power transmission systems and other fields.However,due to the large number of submodules integrated in the MMC,each submodule is a p otential point of failure,therefore,the probability of MMC failure is high.How to diagnose sub-module faults through efficient fault diagnosis strategies to improve the safety and reliability of system operation is an urgent problem to be solved.In thi s thesis,the open circuit fault diagnosis method of MMC submodule is studied in depth based on data-driven methods to improve the fault diagnosis accuracy of MMC submodules and reduce design complexity.The main research work of this thesis is as follows:(1)The basic working principle and topology of MM C are analyzed,the common modulation strategies of MMC and the control strategies of MMC are introduced,and a three-phase MMC model is established according to the aforementioned MMC-related theories,and the MMC sub-module faults in rectifier mode and i nverter mode are simulated and analyzed respectively as the basis for sub-module fault diagnosis.(2)In order to solve the problem of complex MMC structur e and high similarity of fault signals,an optimized Kernel Extreme Learning Machine(KELM)based on composite feature extraction combined with Sparrow Search Algorithm(SSA)is proposed.The MMC submodule fault diagnosis method is presented.This method uses the wavelet energy entropy of the three-phase phase voltage signal and its positive half-period scale factor to construct the composite features,introduces SSA to optimize the KELM parameters,and uses the optimized KELM model to diagnose the fault of the MMC submodule operating in the inverted state,and proves the advantages of the proposed method by comparing and analyzing with other methods.(3)In order to solve the problem that feature selection is too manual and not adaptive in traditional machine learning methods for fault diagnosis,based on deep learning theory,a combination of convolutiona l neural network(CNN)and gated recurrent unit(GRU)has its own advantages,combined with Attention Mechanism(AM)to highlight important features,and p roposes a CNN-GRU-AM MMC sub-module fault diagnosis method.Through the training and testing of the constructed network and comparative analysis with other deep learning methods,the advantages of the method proposed in the fault diagnosis of MMC sub-modules are proved,and the end-to-end fault diagnosis of MMC sub-modules is realized.(4)To further improve the fault diagnosis accuracy of the MMC submodule,the CNN-GRU-AM model is optimized using an optimization strategy to construct a hybrid depth network model.First,Depthwise Separable Convolution(DS C)is used to replace the normal CNN in the CNN-GRU-AM network in order to reduce the network training parameters,while Bidirectional Gated Recurrent Unit(Bi GRU)is used to tap the bidirectional timing information.Bi GRU is used instead of GRU in order to tap the bidirectional timing information,and fin ally,SSA-KELM is used to replace the Softmax layer of the conventional CNN for classification by combining the advantages of SSA-KELM.The proposed hybrid depth network is trained and tested and compared with other methods to verify its excellent performa nce and its advantages in MMC submodule fault diagnosis. |
