A direct mapping between shallow cumulus cloud field properties and three-dimensional surface solar irradiance

Abstract:

Shallow cumulus clouds are encountered frequently across the globe and modulate surface solar irradiance (SSI). The three-dimensional (3D) spatial structure of these cloud fields leads to complex variability in SSI beneath the clouds, which is captured by the SSI probability density function (PDF). Our interest is in the relationship between the cloud field properties and the shape of the SSI PDF. Observed SSI PDFs are typically bi-modal, representing separately the cloud shadows and the gaps between. To simulate the observed SSI PDF shape correctly requires computationally expensive 3D radiative transfer calculations that are not appropriate for operational use. As an alternative, we investigate the feasibility of machine learning algorithms to provide a direct mapping between shallow cumulus cloud field properties (inputs) and their associated SSI PDF shape (outputs). Cloud field properties are derived from large eddy simulation and the SSI PDF shape is quantified by fitting distributions to Monte-Carlo 3D radiative transfer output. Both random forest and artificial neural network algorithms are applied, and in this presentation we will demonstrate the extent to which they can predict variability in the SSI PDF shape using just a handful of key cloud field properties. We will also present differences in how the two algorithms arrived at these predictions using state-of-the-art machine learning inference techniques. This novel approach provides an opportunity to bypass the computational expense of 3D radiative transfer while maintaining realism in simulated SSI. The combination of efficiency and accuracy has immediate applications for assessments of solar renewable energy potential and paves the way for several other applications in the atmospheric sciences.