Abstract
This thesis presents studies of solar windmagnetosphere coupling using dynamic neural networks in combination with statistically correlative analysis. The primary contribution of the thesis is dynamic neural network models that can be implemented for near realtime predictions of geomagnetic storms from the solar wind alone. With acceptable accuracy, the prediction time has been extended up to 5 hours. This is of great socioeconomic significance in space weather forecasting. The secondary contribution of the thesis is the modeling of the magnetospheric dynamics, which optimizes combinations of solar wind parameters and coupling functions. The third contribution of the thesis includes the development of a Ccod of Elman recurrent network models, the development of an algorithm for pruning Elman networks and the algorithms for post network error analyses. The fourth contribution of the thesis is the exploitation of the hybrid of a timedelay network and a recurrent network by examining the role of a timedelay line in recurrent networks.
This thesis consists of five chapters. Chapter 1 provides an introduction to solarterrestrial physics. Chapter 2 is a general description of neural networks. Chapter 3 describes solar windmagnetosphere coupling and related studies. Chapter 4 is devoted to studies of a timedependent system, such as the solar winddriven magnetosphere, using neural networks. Chapter 5 is a summary of the papers included in this thesis.
Paper I pioneered exploitation of recurrent neural networks in prediction of geomagnetic activity and presents very accurate one hour ahead prediction of magnetic storms using only solar wind data.
Paper II is a study of solar windmagnetosphere coupling using a partially recurrent neural network to find the optimal coupling functions. The optimal coupling functions found are used to predict magnetic storms up to 5 hours with acceptable accuracy. This study was the first to present realtime one hour ahead prediction of magnetic storms using the satellite WIND realtime data.
Paper III made predictions of magnetic storms up to 8 hours. It presents the appropriate combinations of solar wind parameters for predicting magnetic storms, which reveals the relative importance of solar wind parameters. It is found that a magnetic storm is formed in the magnetosphere on a timescale of about 1 hour. In this study, we exploit a timedelay recurrent network which is a hybrid of a timedelay network and an Elman recurrent network, and prove it very helpful in improving predictions.
Paper IV studies solar windmagnetosphere interaction in detail, which finds the best coupling functions for accurate prediction of geomagnetic activity based on neural network modeling, in comparison with the results from crosscorrelation analyses. The algorithms for computation of confidence limits on the prediction accuracy are developed in this study.
This thesis consists of five chapters. Chapter 1 provides an introduction to solarterrestrial physics. Chapter 2 is a general description of neural networks. Chapter 3 describes solar windmagnetosphere coupling and related studies. Chapter 4 is devoted to studies of a timedependent system, such as the solar winddriven magnetosphere, using neural networks. Chapter 5 is a summary of the papers included in this thesis.
Paper I pioneered exploitation of recurrent neural networks in prediction of geomagnetic activity and presents very accurate one hour ahead prediction of magnetic storms using only solar wind data.
Paper II is a study of solar windmagnetosphere coupling using a partially recurrent neural network to find the optimal coupling functions. The optimal coupling functions found are used to predict magnetic storms up to 5 hours with acceptable accuracy. This study was the first to present realtime one hour ahead prediction of magnetic storms using the satellite WIND realtime data.
Paper III made predictions of magnetic storms up to 8 hours. It presents the appropriate combinations of solar wind parameters for predicting magnetic storms, which reveals the relative importance of solar wind parameters. It is found that a magnetic storm is formed in the magnetosphere on a timescale of about 1 hour. In this study, we exploit a timedelay recurrent network which is a hybrid of a timedelay network and an Elman recurrent network, and prove it very helpful in improving predictions.
Paper IV studies solar windmagnetosphere interaction in detail, which finds the best coupling functions for accurate prediction of geomagnetic activity based on neural network modeling, in comparison with the results from crosscorrelation analyses. The algorithms for computation of confidence limits on the prediction accuracy are developed in this study.
Original language  English 

Qualification  Doctor 
Awarding Institution  
Supervisors/Advisors 

Award date  1997 May 22 
Publisher  
Publication status  Published  1997 
Bibliographical note
Defence detailsDate: 19970522
Time: 14:00
Place: Physics Institute, sal B, Sölvegatan 14
External reviewer(s)
Name: Vassiliadis, Dimitris
Title: Dr
Affiliation: GSFC/NASA, USA

Subject classification (UKÄ)
 Astronomy, Astrophysics and Cosmology
Free keywords
 Geophysics
 kosmisk kemi
 rymdvetenskap
 Astronomi
 cosmic chemistry
 solar wind
 space weather
 magnetosphere
 solar windmagnetosphere coupling
 geomagnetic activity
 geomagnetic storms
 predictions
 modeling
 neural networks
 space research
 Astronomy
 physical oceanography
 meteorology
 Geofysik
 fysisk oceanografi
 meteorologi
 Fysicumarkivet A:1997:Wu