Research On Large-scale Power Grid Neural Network Fault Diagnosis Model Based On Multi-GP

With the development of "energy interconnection",the scale of power grids and their network topologies are becoming larger and more complex,and the fault diagnosis methods applied to large-scale power grids will face great challenges.When faults occur,fault diagnosis methods based on artificial intelligence technology are essential to assist dispatchers in quickly and accurately identifying faulty components and restoring safe and stable operation for the subsequent power grid.Therefore,this paper will be devoted to research on large-scale grid diagnosis methods based on neural network.Firstly,to address the problems of low diagnostic efficiency and easy dimensional disaster during large-scale grid fault diagnosis based on neural networks,a grid partitioning diagnosis method based on the fusion of improved convolutional neural networks(CNNs)and D-S evidence theory is studied.The partitioning strategy is used to divide the large-scale grid into smaller-scale grids to reduce the input dimensionality of the model,and using improved CNN models to achieve intra-regional fault diagnosis;D-S evidence theory is used to make secondary diagnosis of inter-regional tie lines.Eventually,the fault diagnosis of the whole large-scale power grid is realized to improve the fault diagnosis efficiency.Experimental results show that the method effectively diagnoses faults within the region as well as faults in the tie lines between overlapping regions and has good diagnostic accuracy and model generalization capability.Secondly,to further improve the diagnostic efficiency of large-scale power grids and the ability to analyze uncertainties such as rejection or mis-operation in fault information,we study a fault diagnosis model based on T-S fuzzy neural network(T-S FNN)power grid component.Establish corresponding T-S FNN models for grid components and use the idea of parallel diagnosis to selectively trigger the diagnosis models corresponding to each component,to achieve parallel diagnosis among components,so as to improve fault diagnosis efficiency.Using an improved gainingsharing knowledge algorithm to simultaneously determine the structure and consequent parameters of T-S FNNs to improve the diagnostic accuracy.The experimental simulation proves that the model is more capable to analyze the presence with uncertainty situation in the fault information and can diagnose correctly even in the face with more serious fault situations.Finally,in order to further improve the neural network model more efficiently and accurately for grid fault diagnosis,a large-scale grid fault diagnosis model based on a multi-GPU change architecture-convolutional bidirectional long and short-term memory neural network(CA-CNN-Bi LSTM)is studied.Based on the methods of grid partition diagnosis and element-oriented diagnosis,we build corresponding fault diagnosis models in multi-GPU computing environment respectively to improve the model building efficiency,and thus improve diagnosis efficiency.An intelligent optimization algorithm is used to build CA-CNN-Bi LSTM diagnostic model to improve the model diagnostic accuracy.simulation results show that the method can significantly improve the diagnosis model efficiency and it has a high correct diagnosis rate.