Improved satellite-based estimation of direct normal solar irradiance using clearness index with deep learning methods

As the capacity addition of solar energy systems continues to increase, solar resource assessment is necessary in supporting the feasibility study. This can reduce associated risks of solar energy projects and improve their reliability. Considering the scarcity of on-site measurements, satellite-based irradiance retrievals with high temporal resolutions (i.e., 5-min) have been extensively used as an alternative in solar resource assessment. However, satellite-to-irradiance algorithms, either physical or statistical, focus more on the global horizontal irradiance. The direct normal irradiance (DNI) in most satellite-derived irradiance products is associated with more uncertainties because of its high sensitivity to the atmosphere. To further improve the accuracy in end-to-end satellite-based 5-min DNI estimations, the clearness index based on extraterrestrial solar irradiance is proposed as the target in deep learning satellite-to-DNI models with images of eight selected spectral bands. The results show that clearness index can better account for attenuation effects of the atmosphere, and thus the DNI estimations are associated with lower uncertainties. The use of clearness index offers additional advantages on computing extraterrestrial solar irradiance and pre-processing 5-min spectral satellite data, which is beneficial for large-scale applications. Although the satellite-to-DNI estimation shows high errors under clear-sky condition at some stations, and more efforts are still required in better extracting the atmospheric information (e.g., clouds and aerosols) from satellite images, especially at low solar elevations, the clearness index provides a new perspective in 5-min satellite-to-DNI retrievals with reduced uncertainties. This is beneficial to the reliability in designing solar energy projects.