Search

The work of this thesis is concerned with investigating the application of artificial neural network techniques to the problem of predicting the pattern of behaviour of the solar activity cycle. This investigation measures the success with which these computational methods can predict various solar activity indices on different timescales and in each case compares the level of success with that obtained using a comparison benchmark prediction model. Taken individually, the two subjects involved in this thesis, solar activity prediction and artificial neural network methods, are enormous areas of active research. The overlap, however, has a very limited history and this thesis aims to redress some of these limitations. The most important aspect of my research has been to produce the most accurate predictions of future solar activity possible. In doing this, it is essential to compare the prediction accuracy achieved in my analysis with that currently provided by other methods. In carrying out this research though, I have investigated in detail the effect of varying the parameters which define a neural network on the prediction accuracy achievable. In this way I hope to go some way towards bridging the gap between the two different subject areas. The general outline of the thesis is as follows. The opening two chapters describe in turn the relevant history of solar activity prediction and basic neural network theory. Included in Chapter 1 is a historical perspective of the observations of solar activity phenomena concentrating on early measurements of sunspots. From these measurements the approximate 11-year solar cycle was discovered. The implications of high levels of solar activity for satellite drag, solar particle events and geomagnetic storms are detailed to demonstrate some of the observed interactions between the Sun, the interplanetary medium and the terrestrial environment. A brief review of the types of prediction models which are currently in use for estimating the future behaviour of the Sun is also carried out. Neural network computational techniques were born over fifty years ago but have experienced mixed fortunes in terms of popularity during these years. Since the mid-1980's interest has increased markedly in these methods, in the main due to the emergence of the 'back- propagation of errors' learning algorithm. This algorithm is described in Chapter 2 along with the historical perspective which gave rise to its formulation. Since this research is essentially numerical in nature, a link has to be provided between the theory and the practical implementation. This is done in Chapter 3 where the methodology required to analyse the time series solar activity data using the back-propagation algorithm is described. In particular, the software developed in the course of this research is described, showing how it relates to the back-propagation algorithm. The limited amount of previous work in this subject area is also described along with consideration of potential problems in neural network learning which can occur in a practical implementation of the theory. Chapters 4, 5 and 6 collectively discuss the results of the comprehensive investigations which have been carried out for this thesis. Chapter 4 is devoted solely to analysis of the smoothed monthly sunspot number. This is because this index has been most widely used as an indicator of solar activity levels and has been observed consistently for the longest time. A full description of the effect of changing the various parameters in a neural network is also given. The next chapter summarises the results obtained when a similar analysis is applied to unsmoothed monthly sunspot numbers, monthly and yearly solar flux data and also geomagnetic data in the form of the antipodal aa index. Thereafter, Chapter 5 discusses some alternative ways of presenting data as input to the networks, for example, a combination of neural network training with the McNish and Lincoln method is tested. Different styles of obtaining a prediction for n sample points into the future are also established. Finally, the last half of this chapter concentrates on the specific problem of predicting only the level of activity at the next maximum. This builds on some ideas previously suggested by Koons and Gorney (1990) and compares them with the procedures studied in the earlier parts of this thesis. Chapter 7 offers a brief overview of the results and conclusions of the previous chapters and completes this with a look at some of the prospective areas into which this research may be extended. The original work of this thesis is contained in Chapters 3 through 6 which detail the various investigations carried out into the application of artificial neural network methods to solar activity prediction. The initial results of these studies have been presented at various international conferences or workshops and appeared in subsequent proceedings. The more detailed and complete results are in preparation for submission for publication as full papers.