Adaptive Estimation Method Of Voltage Sag State Based On Deep Transfer Learning

With the rapid economic development;the sensitive load in the power grid is increasing constantly,and the voltage sag is easy to cause the abnormal operation or shutdown of the sensitive load,resulting in serious economic losses.Therefore,this paper takes voltage sag in distribution network as the research object to study the voltage sag state estimation problem:(1)Accurate voltage sag state estimation based on data-driven method.(2)In view of the lack of voltage sag data in changing scenarios,adaptive estimation of voltage sag state is studied.This paper introduces the definition,cause,harm,characteristic quantity and monitoring quantity of the monitoring system of voltage sag,and uses the method of Monitor Reach Area(MRA)to optimize the selection of voltage sag monitoring points in view of the problem of voltage sag state estimation.That is,on the basis of using MRA to determine the observable matrix of voltage sag,the minimum number of monitoring points is taken as the objective function,and the observability of voltage sag is taken as the constraint to determine the selection of monitoring points.One Dimensional Convolutional Neural Networks-Gated Recurrent Unit(1D CNN-GRU)deep learning model is adopted to achieve accurate estimation of voltage drop state.One Dimensional Convolutional Neural Networks(ID CNN)are used to extract the high-dimensional features from the voltage sag monitoring information.After that,the timing features of highdimensional features are extracted through the Gated Recurrent Unit(GRU),and the accurate estimation of voltage drop status is finally realized.The range of voltage drop values of all the other nodes without monitoring devices is estimated.In view of the lack of voltage sag monitoring data in changing scenarios,adaptive voltage sag state estimation is implemented based on the feature decoupling transfer learning method.1D CNN-GRU was used to extract the features of the input information of monitoring points under the original scene and changing scene,and the features important to the estimation of voltage sag state in the original scene and changing scene were obtained through feature decoupling.After that,the distribution difference of important features between the original scene and changing scene was reduced by feature migration,and the migration of features from the original scene to the changing scene was completed.Realize voltage sag state adaptive estimation.In this paper,the IEEE-33 node network is taken as an example to determine the optimal selection of monitoring points,and the simulation verifies the accurate estimation of voltage sag state and the effectiveness of the adaptive estimation of voltage sag state under changing scenarios.Finally,the experimental verification is carried out based on the experimental platform of "RT-LAB+Raspberry PI Embedded Computer".The results show that the selection of monitoring points can realize the observability of voltage sag in the whole network.The voltage sag state estimation method based on 1D CNN-GRU can achieve accurate estimation of voltage sag state under the premise of sparse monitoring points.The transfer learning method based on feature decoupling can realize the adaptive estimation of voltage sag state in the case of lack of voltage sag data in changing scenes.