A Physics-Informed Machine Learning Approach utilizing Multiband Satellite Data for Solar Irradiance Estimation

TL;DR

This paper presents a physics-informed machine learning model that leverages multiband satellite data and prior physical knowledge to accurately estimate solar irradiance across different locations, addressing overfitting issues.

Contribution

The study introduces a novel estimation approach combining machine learning with physical models and satellite data for improved solar irradiance prediction.

Findings

Significant accuracy improvement at validation sites.

Effective handling of clear-sky and thin cloud conditions.

Enhanced applicability across diverse locations.

Abstract

Solar irradiance is fundamental data crucial for analyses related to weather and climate. High-precision estimation models are necessary to create areal data for solar irradiance. In this study, we developed a novel estimation model by utilizing machine learning and multiband data from meteorological satellite observations. Particularly under clear-sky and thin clouds, satellite observations can be influenced by surface reflections, which may lead to overfitting to ground observations. To make the model applicable at any location, we constructed the model incorporating prior information such as radiative transfer models and clear-sky probability, based on physical and meteorological knowledge. As a result, the estimation accuracy significantly improved at validation sites.