Study On The Prediction Model Of High-energy Electron Integral Flux At GEO Based On Deep Learning

Generally speaking,the flux of high-energy electrons with energies greater than 2 Me V at GEO will increase dramatically during the recovery phase of geomagnetic storms.This phenomenon is known as high-energy electron flux enhancement event.Since high-energy electrons have very strong penetrability,the occurrence of such events will have a fatal threat to satellites running at GEO.In this study,the deep learning method of Long Short-Term Memory(LSTM)was applied to develop a model to predict the daily >2Me V electron integral flux at GEO one day ahead.The inputs to the model include geomagnetic and solar wind parameters,and the value of >2Me V electron integral flux itself over the last five consecutive days.The magnetopause subsolar distance was also utilized in modeling.There are two categories of forecasting models,one used the daily integral fluxes of >2Me V electron as input and output to the model,and the other used the hourly integral electron fluxes.In the first category of model,we utilized the observations of high-energy electron flux obtained from GOES satellites(GOES-8,10,11 and 13).The model was trained on the data from 1999-2007 and 2011-2016,and the efficiency of the model was tested on 2008-2010 years.We have tried different inputs and found that when the model takes daily >2Me V electron integral flux,daily averaged magnetopause subsolar distance,and daily summed Kp index as inputs,the prediction efficiency for 2008,2009 and 2010 are 0.833,0.896 and 0.911,respectively.In addition,the corresponding linear correlation coefficients are 0.921,0.948 and 0.955,respectively.Because of the differences between observations of high-energy electron flux obtained from different GOES satellites due to longitude location,only the data obtained from GOES-11 was utilized in the second category of model.The prediction efficiency and linear correlation coefficient can reach up to 0.900 and 0.950 for 2008 in the second category of model,when it takes hourly integral flux of >2Me V electron,hourly magnetopause subsolar distance,and daily summed Kp index as inputs,trained on the datasets during June 19,2003 to April 13,2010 except 2008.The prediction efficiency of the persistence model and 27-order autoregressive model tested on 2008 year are 0.679,0.743,respectively.Compared to the persistence model,autoregressive model and some previous models,the prediction results of LSTM model has been improved significantly during the corresponding period.At the same time,comparing with the performance of the models based on daily integral electron flux,the forecasts of the model based on hourly integral electron flux also shown that high resolution data used in modeling can further improve the performance in prediction.