Predicting geomagnetic storms from solar-wind data using time-delay neural networks

Abstract

We have used time-delay feed-forward neural networks to compute the geomagnetic-activity index D _st one hour ahead from a temporal sequence of solar-wind data. The input data include solar-wind density n, velocity V and the southward component B _z of the interplanetary magnetic field. D _st is not included in the input data. The networks implement an explicit functional relationship between the solar wind and the geomagnetic disturbance, including both direct and time-delayed non-linear relations. In this study we especially consider the influence of varying the temporal size of the input-data sequence. The networks are trained on data covering 6600 h, and tested on data covering 2100 h. It is found that the initial and main phases of geomagnetic storms are well predicted, almost independent of the length of the input-data sequence. However, to predict the recovery phase, we have to use up to 20 h of solar-wind input data. The recovery phase is mainly governed by the ring-current loss processes, and is very much dependent on the ring-current history, and thus also the solar-wind history. With due consideration of the time history when optimizing the networks, we can reproduce 84% of the D _st variance.