Prediction of geosynchronous electron fluxes using an artificial neural network driven by solar wind parameters

There are hundreds of satellites operating at the geosynchronous (GEO) orbit where relativistic electrons can cause severe damage. Thus, predicting relativistic electron fluxes is significant for spacecraft safety. In this study, using GOES satellite data during 2011–2020, we propose two neural network models with two hidden layers to predict geosynchronous relativistic electron fluxes at two energy channels (>0.8 MeV and > 2 MeV). The number of input neurons of the two channels (>0.8 MeV and > 2 MeV) are determined to be 36 and 44, respectively. The > 0.8 MeV model has 22 and 9 neurons in the hidden layers, while the > 2 MeV model has 25 and 15 neurons in the hidden layers. The input parameters include the north–south component of the interplanetary magnetic field, solar wind speed, solar wind dynamic pressure and solar wind proton density. Through the analysis of different time delays, we determine that the optimal time delays of two energy channels (>0.8 MeV and > 2 MeV) are 8 days and 10 days, respectively. The training set and validation set (Jan 2011-Dec 2018) are divided by the 10-fold cross-validation method, and the remaining data (Jan 2019-Feb 2020) is used to analyze the model performance as a test set. The prediction results of both energy channels show good agreement with satellite observations indicated by low RMSE (～0.3 cm^-2sr^-1s^?1), high PE (～0.8) and CC (～0.9). These results suggest that only using solar wind parameters is capable of obtaining reasonable predictions of geosynchronous relativistic electron fluxes.