Solar PV installation is growing fast in recent decades across the world but high variability of solar power hinders its further penetration to the energy market. This variability mainly comes from cloud coverage, water vapor content and aerosol loadings, and has the greatest effect in short-term solar power prediction. This high volatility nature of solar insolation makes it difficult to integrate PV output to electricity grid. A more accurate short-term solar power prediction helps to develop bidding strategies for real-time markets or to determine the need for operating reserves. This work aims to tackle this problem by employing comprehensive spectral radiative models to calculate longwave and shortwave radiation through the atmosphere, estimating cloud properties from remote sensing data with the atmospheric model and building convolutional neural network model to model and forecast solar radiation.

First, a Line-by-Line (LBL) spectral radiative model is built to capture details of the highly wavenumber-dependent nature of the irradiance fluxes. Then the broadband empirical model serves as a benchmark to validate the LBL model. For longwave spectrum that is emitted and absorbed by gases, aerosols, clouds and the ground, a high-resolution two-flux model with a recursive scattering method is developed. For the shortwave (solar) part of the spectrum, which includes scattering from atmospheric constituents and the ground, 3D comprehensive Monte-Carlo simulations are used. Beyond the basic model, some corrections or calibrations are made. Comprehensive Monte Carlo simulations are used for correcting deviations on the atmospheric downwelling longwave (DLW) flux caused by isotropic scattering assumptions in high aerosol loading regimes.The $\delta$-M approximation input-based scaling rule is validated for a wide range of aerosol loading values except for very high aerosol loading conditions. This proposed scaling rules minimize substantially the computational effort of calculating anisotropic downwelling radiation from diverse types of aerosols under these extreme conditions. Earth curvature effect(air mass correction) is also tested. Although for solar zenith angles larger than 75°, the attenuation of the direct solar beam is overestimated in a plane-parallel atmosphere comparing to in a real spherical atmosphere, for most solar rays, a plane-parallel atmosphere approximation is accurate enough for modeling.

A Spectral Cloud Optical Property Estimation (SCOPE) method that integrates the high-resolution imagery from GOES-R satellite and a two-stream, spectrally-resolved longwave radiative model was proposed, for the estimation of cloud optical depth and cloud bottom height. An improved model SCOPE 2.0 is also proposed which considers multi-layer clouds, clouds with ice crystals and aerosol corrections. A shortwave Monte Carlo simulation is developed and used to validate the derived cloud optical properties. With this comprehensive cloud cover estimate model, a convolutional neural networks (CNN) model is developed to correlate global horizontal irradiance (GHI) to the satellite-derived cloud cover (a "now-cast"). The performance of SCOPE method as well as CNN+SCOPE model is evaluated using one year (2018) of downwelling longwave (DLW) radiation and GHI measurements from the Surface Radiation Budget Network, which consists of seven sites spread across climatically diverse regions of the contiguous United States. CNN+SCOPE model achieves test-set root-mean-square error (RMSE) of 30.5 - 62.6 $W m^{-2}$ with an average of 47.2 $W m^{-2}$, which is better then the National Solar Radiation Database (NSRDB) model (average RMSE is 66.9 $W m^{-2}$). A reference CNN model is also tested which directly use satellite ABI data that the SCOPE model uses with an average error equal to 69.4 $W m^{-2}$. This success at CNN+SCOPE ”now-cast” model points to possible future uses for short-term forecast.