The Prediction Of High-Energy Electron Fluence At GEO Based On EMD Algorithm And LSTM Network

During the recovery of magnetic storms,the relativistic electrons with MeV energy diffuse from the outer radiation belt to geosynchronous orbit.The electrons with energy are larger than 2 MeV could penetrate the surface of satellites and accumulate inside them.Such an electron flux effect could cause satellites to be unable to function properly or to fail completely.Due to external forcing by the solar wind,relativistic electrons change very rapidly during magnetic storms and are very non-stationary,it brings very big difficulties to accurate forecasts.It is found through the experiment that the Empirical Mode Decomposition(EMD)algorithm can deal with the non-stationary problem of high-energy electron flux data series.High-energy electron flux is significantly enhanced during the recovery phase of the magnetic storm,sometimes suddenly increasing by 3~4 orders of magnitude.Most of the existing prediction models are difficult to accurately describe the events in which the highenergy electron flux suddenly increases.We introduce the Long Short Term Memory(LSTM)to predict >2MeV electron flux.The algorithm is easier to capture the nonlinear relationship in the data set and can make more accurate predictions of electron flux based on historical usage data information.We used EMD algorithm and LSTM network to deal with the non-stationary and nonlinear problems of >2MeV electron flux data series respectively and used these two methods to predict >2MeV electron flux of GEO.The >2MeV electron flux data of 2001-2009 was used as the training set,the data of 2010-2013 was used as the testing set.The prediction results are ideal,and we compared the prediction results with the Persistence model,Kalman filter model,LSTM network model and a combined model using EMD algorithm and Kalman filter algorithm.The results showed that the model which used the EMD algorithm and LSTM network is the most ideal.The prediction efficiency in 2010-2013 is above 0.80,and the highest is 0.92.In addition,we compared the prediction results in 2003-2006 with existing prediction models,the results showed that the combined model which used the EMD algorithm and LSTM network has the best prediction results.The effectiveness of this model was further verified.