A neural network forecasting of relativistic electron flux at geostationary orbit: solar activity phase dependence

Abstract

Geostationary-satellite anomalies are mainly caused by extreme space environment such as flux enhancements of relativistic electrons. A prediction of the relativistic electrons at geostationary orbit (GEO) is crucial to alert of the flux enhancements before it reaches a harmful condition. This work investigates contribution of solar activities at various phases in the daily electron flux (E > 2 MeV) at GEO measured by GOES satellites during the solar cycles 23-24. An alternative method based on the solar activity and feed forward neural network (NN) scheme with back-propagation learning in Waikato Environment for Knowledge Analysis is adopted to predict the electron fluxes in GEO by sorting the solar activity into maximum, descending, and ascending phases. As the model inputs, the historical daily sum K p and the > 2 MeV electron fluxes are used. Results indicate that the prediction capability of the NN model is dependent on the solar activity in relation to the sum K p. Crossing the test over the similar solar phases gives a better prediction result than crossing over different phases. The NN model based on the lagged sum K p and log-flux input are more suitable for the forecasting during the descending and ascending phases, while the lagged log-flux input is more suitable during the maximum phase.