The Research On Recognition Algorithm Of Power Quality Disturbances Based On Recurrent Neural Network

With the development of industrialization,the proportion of non-linear loads in the power grid is increasing,and the problem of power quality is getting more and more attention.Monitoring the type and timing of power quality disturbances in the power grid has important theoretical and practical significance.Studying the disturbance problems can reduce the problem of power outages caused by the disturbance,allowing enterprises to have better economic benefits,and can protect the grid equipment for the power grid itself At the same time,it can reduce the occurrence of accidents and protect people’s lives and property safety by analyzing disturbances.For the problems of the existing network power quality disturbance database with fewer disturbance types and shorter disturbance time,the Monte Carlo idea in statistics is used to solve the data problem.According to the international and domestic standards on power quality,mathematically model the existing 14 types of disturbance definitions,and generate the probability of disturbances through the rich libraries and functions in Python.Finally,a disturbance database of 14 types of transient steady-state disturbances is obtained as a sample.For the problems that the existing disturbance feature extraction algorithm is not accurate enough to extract the disturbance feature and the identification of transient disturbances is poor,the empirical mode decomposition algorithm of Hilbert-Huang transform is used as the feature extraction algorithm,which uses the empirical mode decomposition.The modal components are input to the recognition algorithm as feature vectors,That is,the eigenmode component is used as the input of the dynamic recurrent neural network,and the simulation results verify the effectiveness of feature extraction.In terms of disturbance recognition algorithms,select the recurrent neural network as the basic algorithm,and add the variational automatic encoder to the network with the GRU gating unit as the basic structure,and the network structure is composed by four layers instead of three layers,the latent layer makes the network more adaptable to data.Finally,according to the output probability curve,the disturbance type and the start and end time of each disturbance are determined.Feature extraction uses empirical modal decomposition.After 150 generations of dynamic recurrent neural network,the average recognition accuracy of 14 types of disturbances exceeds 85.8%,which is higher than the average recognition accuracy of the other five algorithms,so the final recognition result is verified and identify the effectiveness of the algorithm.