Research On Fault Diagnosis Method For Hybrid HVDC System Based On Deep Learning

Hybrid HVDC transmission technology combines the advantages of the line commutated converter(LCC)and the voltage source converter(VSC),which can improve the voltage support capacity of the system while effectively avoiding commutation failure.Hybrid HVDC transmission technology carries the important mission of long-distance and large-scale transmission,which is of great significance to respond to the national dual-carbon strategy and promote the development of new power systems.However,as a new technology,hybrid HVDC transmission has relatively little theoretical research and operating experience,and there is an urgent need to study fault diagnosis techniques suitable for hybrid HVDC transmission to improve the safety of system operation.Taking the hybrid HVDC transmission system as the main research object,this thesis takes the unique advantages of deep learning in terms of fault data self-learning and adaptation to deeply study the fault diagnosis method of hybrid HVDC transmission line in the case of short circuit fault,mainly including the fault nature identification method and fault line location method,and considering the limitation of obtaining fault data in actual engineering,research on fault data enhancement method.The main tasks of this article are as follows:Firstly,the operating principle of the hybrid HVDC transmission system is introduced from the perspectives of the topology of the model and the converter control strategy;the simulation model of the three-terminal hybrid HVDC transmission system under the actual engineering operating parameters is constructed on the PSCAD/EMTDC platform;and the batch acquisition of a double-layer module is constructed to quickly obtain the electrical quantity data under multiple fault conditions,laying a solid data foundation for the subsequent data-driven fault diagnosis method.Secondly,aiming at the shortcomings of traditional fault identification methods in obtaining fault information,a fault identification method for hybrid HVDC transmission system based on the combination of wavelet packet energy spectrum and convolutional neural network(CNN)is proposed.The rich frequency domain information in the original fault data is extracted by wavelet packet decomposition,and then the key fault information is identified by entering the CNN network in the form of energy characteristics to achieve accurate diagnosis of fault lines and fault types.The proposed model shows that the fault identification accuracy is high,and it has strong anti-interference and transition resistance resistance capabilities.Then,aiming at the problem of complex topology in hybrid HVDC transmission system and the difficulty of detecting the location of fault points caused by long-distance transmission,a fault location method based on sparrow-optimised CNN-LSTM is proposed.Combining the advantages of self-extraction of CNN fault features and the memory characteristics of long short-term memory(LSTM)for timing information,a CNN-LSTM fault location network is constructed,and improved sparrow search algorithm with Logistic chaos mapping(LSSA)is used to optimize the hyperparameters of network structure,so as to form an optimized fault location network model,which can meet the accuracy requirements of actual fault location.The proposed network model and method are less affected by the transition resistance,and the fault location accuracy and diagnosis time can meet the needs of actual engineering,which has certain advantages compared with traditional back propagation(BP)neural network and deep learning algorithms such as stacked autoencoder(SAE)and LSTM.Finally,based on the above fault identification method and fault location method,considering the difficulty of obtaining fault data and insufficient data set samples in actual engineering,a fault sample data enhancement method based on Wasserstein generative adversarial network-gradient penalty(WGAN-GP)is proposed,and the Wasserstein distance and gradient penalty are introduced into the GAN network model to improve the stability of network model training.Through the fault identification network and fault location network in the first two chapters,the generated samples are verified,and the proposed method can effectively obtain the generated samples that are highly similar to the real fault samples,so as to expand the sample set and improve the comprehensive diagnostic performance of the network.