A neural network-based geosynchronous relativistic electron flux forecasting model

Abstract

[1] A multilayer feed-forward neural network model has been developed to forecast >2 MeV electron flux at geosynchronous orbit. The model uses as input 10 consecutive days of historical electron flux values and 7 consecutive days of daily summed values of the planetary Kp index with two neurons in a single hidden layer. Development of the model is discussed in which the size of the training set interval and the retraining period are investigated. Problems associated with neuron saturation which limit the ability of the network to generalize are shown to be circumvented through a daily retraining regimen. The model performance is evaluated for the period 1998–2008 and compared with the results produced by the REFM model. The neural network model is demonstrated to perform quite well relative to the REFM model for this time period, producing mean prediction efficiencies for 6 month test intervals of 0.71, 0.49, and 0.31 for 1 day, 2 day, and 3 day forecasts, respectively.