Research On Fault Diagnosis Method Of Multi-terminal DC Transmission Line Based On Deep Learning

High-voltage direct current(HVDC)transmission technology is rapidly advancing and gradually demonstrating a trend towards the development of multi-terminal HVDC transmission systems from two-terminal systems.Multi-terminal HVDC transmission lines exhibit characteristics such as fast rise in fault currents,high peak values,and difficulties in fault localization.Additionally,multi-terminal HVDC grids have disadvantages such as a large number of transmission lines,long spans,and complex terrains,which increase the difficulty of line inspection and maintenance.Accurate and rapid identification of faulty lines and fault types is of significant importance for improving line inspection efficiency,quickly restoring faults,reducing power outage losses,and ensuring reliable power supply.This study focuses on the research of deep learning-based fault diagnosis methods for multi-terminal HVDC transmission lines.The operating mode and characteristics of multi-terminal HVDC grids are analyzed,and a fourterminal HVDC transmission grid simulation model is constructed in PSCAD/EMTDC.The Random module is utilized to randomly alter fault attributes and simulate various fault conditions.To enhance the efficiency of obtaining fault sample data,the model is simulated in batches.The acquired fault sample data is preprocessed and used as a dataset for fault diagnosis features in deep learning models,which are then applied in the study of fault diagnosis algorithms for multi-terminal HVDC transmission lines.Three models with strong feature extraction capabilities,namely Deep Belief Nets(DBN),Long Short-Term Memory(LSTM)networks,and Convolutional Neural Networks(CNN),are established.The three fault diagnosis models are compared in terms of diagnosis accuracy for multiterminal HVDC transmission lines and differences in the training process.Simulation results indicate that DBN has a risk of overfitting and a more challenging training process,LSTM requires a longer training time,while CNN exhibits the highest recognition accuracy and the shortest training time.Fault samples with Gaussian white noise at signal-to-noise ratios of 30dB-70dB are input into the three pre-trained network models(DBN,LSTM,and CNN)for testing,and the simulation results demonstrate CNN’s outstanding anti-interference capability.The influence of different configuration parameters of CNN on fault diagnosis is studied,and optimization is performed to obtain an optimal CNN network parameter model.The t-SNE algorithm is used to visualize the output of the CNN model,showing a clustering phenomenon for the same type of fault data and no overlap with distinct boundaries between different fault types.This demonstrates CNN’s excellent feature extraction capability.When applied to fault diagnosis in multi-terminal HVDC transmission lines,CNN achieves a high fault diagnosis accuracy.The diagnosis accuracy for fault types and fault lines reaches 99.92%and 99.97%respectively,making it suitable for fault diagnosis in multi-terminal HVDC transmission lines and almost unaffected by fault location,fault type,fault line,and transition resistance.