Modeling temporal dynamics of solar irradiance using neural networks

Ambelu, Tebabal

dc.contributor.author	Ambelu, Tebabal
dc.date.accessioned	2019-03-04T04:33:34Z
dc.date.available	2019-03-04T04:33:34Z
dc.date.issued	2019-03-04
dc.identifier.uri	http://hdl.handle.net/123456789/9286
dc.description.abstract	Different observations of the Sun provide a vast array of measures of solar activity, including solar irradiance, which has been used in this thesis. The Sun’s irradiance, total solar irradiance (TSI) and spectral solar irradiance (SSI), incident at the top of the Earth’s atmosphere and normalized to one astronomical unit, have been measured with spaceborne instruments continuously since 1978. The irradiance varies on several timescales, ranging from minutes up to decades, and likely lasting even longer. This temporal variability of SSI significantly alters the Earth’s atmospheric density distribution and temperature, which drive variations in many upper atmospheric processes including satellite drag, ground-space communications, and GPS precision. Long-term TSI variability is essential in understanding past and future global climate changes. Direct measurements of the solar irradiance are available over the last four decades and are too short to derive conclusions about any possible long-term changes in solar irradiance and their possible influence on climate. It is, therefore, necessary to use models of solar output to deduce variations at earlier dates. One type of solar irradiance model is an empirical model, frequently called a proxy model, that is derived using linear relationships between a proxy of solar activity and direct observations of the solar irradiance. The main driver of the irradiance variations on time-scales of days to decades, and possibly longer, is believed to be associated with solar magnetic activity located in the active photospheric regions of sunspots and faculae. Empirical models incorporating the effects of magnetic features (both sunspots and faculae) are sufficient to account for most of the observed changes in TSI. However, proxy based empirical model outputs and observed TSI comparison do not support this conclusion; the model fails to explain the depletion that was observed in TSI since 2005. Conventional regression approach does not have the features that can address this kind of temporal variations. Therefore, a method aligning the empirical model with the observation in an adaptable manner is needed. This means that in order to accurately estimate the detailed characteristics of TSI, including TSI variability before 1947, an adaptive and more robust model is essential. In this thesis, a data-driven method of solar irradiance modeling from magnetic features is presented. The thesis employs a neural network (NN) modeling approach to reconstruct both total and spectral irradiance temporal dynamics using the solar proxy as the input drivers for the variations. The physical basis of this model is that all variations in solar irradiance are caused by changes in surface magnetic activity. To find the nonlinear mapping between solar magnetic features and solar irradiance, feed-forward neural networks were used due to their simplicity, flexibility, and ease of use. To determine the critical combination of magnetic features, we explored the influence of magnetic features on the solar irradiance variations by means of network-based empirical modeling. To train the neural network, the photometric sunspot index (PSI) and the magnesium II core-to-wing ratio (Mg II index) were used in the input space while the Physikalisch-Meteorologisches Observatorium Davos (PMOD) composite used as a target of the network. In order to facilitate the learning process, two training algorithms have been implemented: Levenberg- Marquardt and Bayesian approach. Using these approaches, we separately developed two TSI models. The performance of the network was estimated quantitatively by means of cross-validation and learning curve analysis on the data from the PMOD composite. The ability of the networks to model the TSI variations was also independently tested by comparing its performance with TSI data obtained from semi-empirical and linear regression models. The results indicate that the NN TSI model proposed has a good performance in representing TSI variations compared to the linear regression model. To deduce solar output variation prior to the satellite era, first, we estimated the Mg II index variation from F10.7 cm solar flux back to 1947 using NN modeling approach. By incorporating the PSI and the modeled Mg II index effects, we extend the TSI estimation back to 1947. The extrapolated TSI result indicates that the amplitudes of Solar Cycles 19 and 21 are closely comparable to each other, and Solar Cycle 20 appears to be of lower irradiance during its maximum.	en_US
dc.language.iso	en_US	en_US
dc.subject	Physics	en_US
dc.title	Modeling temporal dynamics of solar irradiance using neural networks	en_US
dc.type	Thesis	en_US