| Geosynchronous orbit(GEO)is located in the region of extraterrestrial radiation belt,which is about 36,000 kilometers above the Earth’s surface.About 400 communication,meteorological and navigation satellites operate there.High-energy electron flux is the main component of geostationary orbit,and its rapid enhancement will affect the safe operation of these geostationary orbit satellites.So it is significant to accurately predict high-energy electron flux.This paper,based on a variety of machine-learning methods,uses geomagnetic fluctuation parameters,solar wind parameters,geomagnetic parameters to study the short-medium term prediction of electron flux.The main research contents about electron flux are as follows:(1)The time series parameters used in prediction are cleaned.The fusion data is preprocessed including interpolation supplement of missing values,integrated unified resolution,delayed analysis,normalization,which provides the data meeting the requirements for machine-learning models.(2)The Empirical Mode Decomposition algorithm has been integrated into the Long-short Term Memory Network.The EMD-LSTM prediction model of >2 Me V is constructed by using geomagnetic pulsation Pc5 and various geomagnetic indices from 2001 to 2011.Firstly,the correlation between input parameters and electron flux is analyzed,and the best correlation between them appears in less than 5 days.We select 30 parameters of forecasting factors,including Pc5,AP,Kp,AE,>0.6 Me V and historical values of electron flux during previous five days,as the input of predicted model.Then,EMD method is used to decompose the highenergy electron flux sequence to obtain 9 subcomponents and a residual component.After decomposition,adam optimizer is used to construct LSTM model with full-connected layer and multi-node LSTM layer stacking as hidden layer for reconstruction prediction.Data from 2001 to 2008 is used as training set,and data from 2009 to 2011 is used as testing set.Finally,the result is compared with LSTM model and EMD-KLM model.The prediction efficiency of EMD-LSTM in 2010 and 2011 are 0.93 and 0.82,respectively.The prediction efficiency of LSTM is 0.89 and 0.75,and that of EMD-KLM is 0.88 and 0.77,respectively.The prediction efficiency of EMD-LSTM model is better than that of LSTM and EMD-KLM,and its correlation coefficient and root mean square error are also better.This indicates that the EMDLSTM model,which uses geomagnetic pulsation Pc5 and other geomagnetic indices(compared with solar wind parameters,geomagnetic data acquisition is more stable and less costly)as input,is an effective way to predict high-energy electron flux.At the same time,combined with the EMD algorithm,the prediction method of "decomposition before reconstruction" can effectively deal with the non-stationary problem of electron flux,so as to reduce the prediction error of the model.(3)The Gaussian Process Regression(GPR)model based on prior data evaluation is introduced into the prediction of >2 Me V electron flux.By analyzing the delays of input parameters like >2 Me V electron flux,solar wind parameters and geomagnetic parameters,it is found that the delayed analysis is a good method to select the prediction factor.We used the RBF kernel function and the fmin＿l＿bfgs＿b optimizer to improve the model performance.The prediction efficiency one day in advance is better than that of four widely used models including RDF model,FLUXPRED model,Low-E model,and REFM model(NASA SWPC).It is also better than the results of other intelligent algorithms such as LSTM,SVR,BPNN,DT.At the same time,it is found that GPR model can predict the extreme disturbance events with sudden increase or decrease of several orders of magnitude,which further verifies the validity of model.In addition,a rolling prediction method is proposed to solve the problem that PE of high-energy electron flux is relatively low 2 and 3 days in advance.Rolling method refers to that the predicted value of one day in advance is used as the input parameter of the prediction model two and three days in advance(similar to a dynamic correction method),so as to improve the performance of multi-step prediction.It is combined with GPR model as Rolling-GPR,and the prediction efficiency of Rolling-GPR is much better than that of FLUXPRED model(Ling et al.,2010),which proves that the Rolling-GPR combined prediction model is effective in predicting high-energy electron flux.And its predicted advantages become more obvious with the increase of predicted step size. |
