Solar Wind and Geomagnetic Activity - Predictions Using Neural Networks

Abstract: This thesis shows how artificial neural networks (ANNs) can be applied to predict geomagnetic activity from solar-wind data. It introduces and summarizes five papers where development of ANN models are reported, and where predictions of geomagnetic activity are discussed. The studies cover geomagnetic disturbances characterized by global-scale indices and geomagnetic variations that are locally observed. Several types of ANN are utilized: time-delay networks, radial-basis function networks, and partially recurrent networks. Methods and procedures that can be applied to forecasting based on real-time data are emphasized. The first, introductory part of the thesis begins with an outline of the solar-terrestrial space environment, focusing on the processes and circumstances that play a role in the generation of geomagnetic disturbances. The ANN methods that have been used in the present studies are then briefly described, and put into a wider context of other modeling and prediction techniques. The introductory part of the thesis ends with short summaries of the five papers, which are reprinted in the second part of the thesis. Paper I describes predictions of the ring-current index Dst from solar-wind data. It is shown that magnetic storms can be predicted with time-delay networks. The influence of the solar-wind input sequence length on the different storm phases is discussed. Papers II, III, and V present studies of the solar wind-auroral electrojet relations using time-delay networks II,III, and recurrent networks V. The relative importance of different solar-wind variables and coupling functions is studied in paper II. Paper III describes the influence of the Dst level on the modeled solar wind-auroral electrojet relations. In paper V, the capabilities of recurrent networks are evaluated, and compared to time-delay networks. In paper IV , predictions of locally observed geomagnetic variations are studied. The daily, quiet-time variations are modeled with radial-basis function networks that account for annual and solar-cycle modulations. The horizontal magnetic disturbance field is modeled with gated, time-delay networks taking local time and solar-wind data as input.