Fault Diagnosis Of High-voltage DC Transmission Lines Based On CNN-BiGRU

High Voltage Direct Current(HVDC)transmission technology based on Modular Multilevel Converter(MMC)plays an important role in long-distance and large-capacity efficient transmission projects in China.The stable operation of HVDC systems directly affects the stability of every connected system and even the entire power grid,while HVDC transmission lines are the components most susceptible to failures in the entire system.Therefore,it is of great significance to achieve rapid fault identification and accurate location on high-voltage DC transmission lines for maintaining system safety and reliable operation.The transient traveling wave of MMC-HVDC transmission line faults has the characteristics of time sequence and strong nonlinearity,which can lead to a decrease in fault recognition accuracy.To solve this problem,a fault recognition method for HVDC transmission lines based on Convolutional Neural Network(CNN)and Bidirectional Gate Recurrent Unit(Bi GRU)is proposed.First,the bipolar transient current wave on the rectifier side after the fault is used as the feature vector.The global features are extracted by CNN and noise and unstable components are removed from it to complete the dimensionality reduction processing of the data.Then,Bi GRU is used to capture the front and rear time information of the features extracted by CNN to further extract the time sequence features in the data to achieve fault recognition of HVDC transmission lines.Simulation experiments show that this method can accurately identify four types of faults: single-phase ground fault,bipolar short circuit fault,lightning fault,and lightning interference under different fault locations and different transition resistances.It has high reliability,strong tolerance to transition resistance,and certain anti-noise performance.Based on the Multi-Task Learning(MTL)idea,a fault location method is proposed that can locate multiple types of faults simultaneously.A Softmax classifier of CNN and Bi GRU are selected to construct specific task layers to complete the tasks of determining fault type and locating faults respectively.First,the rectifier side fault current wave is input into CNN for feature extraction.Then,according to the fault type output by Softmax classifier,CNN’s hidden layer is used as a shared layer input into Bi GRU corresponding to the corresponding fault type to complete the fault location task.Experimental results show that this method also has high positioning accuracy for high-resistance grounding faults with transition resistance in kilo-ohms.Compared with other traditional methods,it has better tolerance to transition resistance,anti-noise performance and generalization performance.