Degradation in neural network prediction models of f0F2 with time

Ionospheric forecasting services fall short of the precision required for many practical applications. In-service models of ionospheric parameters often fail to offer much of an increase in performance over simple persistence or recurrence. In particular, predictions of storm events and disturbances are very poor. This performance shortfall can be attributed to several causes one of which is an inability to adequately model solar-magnetospheric-ionospheric physics. To circumvent this knowledge gap some authors have adopted knowledge independent (time-series) modelling techniques that can utilise typical ionospheric data sets. A number of experimental and theoretical studies have demonstrated the importance of non-linear behaviour within the solar-terrestrial environment. Attention has also turned to time series predictive methods derived from studies into artificial intelligence (AI) and the application of neural network (NN) based techniques to geophysical prediction problems. Francis et al. (2000) undertook a study of neural networks for the prediction of the ionospheric parameter f0F2 based on the use of historical data from the 1970s. The resultant model for the one hour ahead prediction of this parameter had an RMS error of only ~0.4 MHz. A version of this model, but re-optimised, was incorporated in the Ionospheric Forecasting Demonstrator (IFD). However, it soon became clear that the predictive capability of the IFD was degrading with time. This paper reports on our studies to understand the reasons for this degradation and to improve the model.