On Generalized Additive Models for Representation of Solar EUV Irradiance

In the context of space weather forecasting, solar EUV irradiance specification is needed on multiple time scales, with associated uncertainty quantification for determining the accuracy of downstream parameters. Empirical models of irradiance often rely on parametric fits between irradiance in several bands and various solar indices. We build upon these empirical models by using Generalized Additive Models (GAMs) to represent solar irradiance. We apply the GAM approach in two steps: (a) A GAM is fitted between FISM2 irradiance and solar indices F10.7, Revised Sunspot Number, and the Lyman-alpha solar index. (b) A second GAM is fit to model the residuals of the first GAM with respect to FISM2 irradiance. We evaluate the performance of this approach during Solar Cycle 24 using GAMs driven by known solar indices as well as those forecasted 3 days ahead with an autoregressive modeling approach. We demonstrate negligible dependence of performance on solar cycle and season, and we assess the efficacy of the GAM approach across different wavelengths. Modeling solar irradiance at extreme ultraviolet wavelengths is very important for describing the behavior of the upper atmosphere. Many empirical models represent solar irradiance by describing it as linearly-dependent on measurements of other quantities that are very strongly correlated with it. These methods have shown great promise, but require building their models from many sources of data. We build upon these methods by showing that using only four sources of data (three solar proxies and irradiance from the FISM2 model), solar extreme ultraviolet irradiance can be modeled in different wavelengths efficiently. We use Generalized Additive Models (GAMs) for our approach, which are used to describe irradiance in terms of a sum of smooth functions of solar proxies. We show how this approach can be used to forecast solar EUV irradiance. Solar EUV irradiance is modeled using Generalized Additive Models parameterized using F10.7, revised Sunspot Number, and Lyman-alpha Relative irradiance error is evaluated for solar cycle and seasonal dependence and compared to that of TIMED/SEE and SDO/EVE Solar EUV irradiance hindcasts during Solar Cycle are generated to demonstrate the suitability of the GAM approach for forecasting