Space Weather Physics, Prediction and classification of solar wind structures and geomagnetic activity using artificial neural networks

Sammanfattning: This thesis concerns the application of artificial neural network techniques to space weather physics. The networks applied include multi-layer error-backpropagation, radial basis function, and self-organized maps. Different parts in the solar-terrestrial chain are analysed with the emphasis on developing methods for real time predictions of geomagnetic activity. The neural networks are general models which utilize learning algorithms to adjust the free parameters of the models based on data samples. The models used here rely heavily on observations of solar magnetic fields, measurements of solar wind plasma and magnetic fields, and indices of geomagnetic activity. The thesis consists of an introductory part followed by 5 papers. The introduction describes part of the solar-terrestrial physics that is relevant to the papers and includes a summary of the applied neural networks used. Papers I and II describe the application of multi-layer error-backpropagation networks to the solar wind-magnetosphere coupling, where the geomagnetic activity is described by the Dst index. It is shown that real time predictions of the Dst index can be made one hour in advance. Papers III and IV examine the possibility to predict the daily average solar wind velocity from solar magnetic field observations. The model consists of a potential field model describing the solar coronal magnetic fields and a radial basis function neural network for the mapping from the corona to the solar wind. Paper V considers the analysis of hourly average solar wind structures at 1 AU using self-organizing maps. It is found that it is possible to identify specific solar wind events on the self-organized maps that are associated to geomagnetic storms occurring several hours later.