Integrated remote sensing and machine learning techniques for solar forecasting and resource assessment

Abstract

Solar energy is set to be one of the major power sources to enable deep decarbonization owing to its sustainability. However, the integration of solar energy into power systems is still facing challenges due to the variability and intermittency of available solar irradiance. Ground measurements are the most reli­able data source for designing solar energy projects, but complete and long-term historical data are scarce. Therefore, it is important to retrieve ground-level solar irradiance from modern geostationary satellites with much improved spatio-temporal resolutions to fill the data gap. Meanwhile, solar forecasting is also a cost-effective method to reduce the negative impacts of solar variability and help with the system integration and management. Both solar forecasting and resource assessment rely on the clear-sky model, which estimates the clear-sky irradiance under cloudless conditions. To address the difficulty in applying physical clear-sky models and the compromised performance of empirical clear-sky models, an improved turbidity estimation method is proposed for estimating clear-sky irradiance based on common meteorological measurements. It is shown that the clear-sky irradiance from improved turbidity estimation exhibits lower divergences compared with the monthly climatological means. The Ineichen-Perez model based on the improved turbidity estimation is then applied for global horizontal irradiance (GHI) estimation using semi-empirical satellite methods. The results show the GHI estimates have comparable performance to the referenced physical solar model in the national solar radiation database (NSRDB), but with less complexity. To further expand the applicability of the turbidity estimation method, a transferable model is proposed for estimating turbidity and clear-sky irradiance. The results show that clear-sky GHI estimates and day-ahead persistent forecasts are compara­ble with the physical clear-sky models. Given that the forecasts of meteorological information are much more accurate than solar irradiance forecasts, the transferable turbidity estimation therefore shows valuable potential for solar energy applications. Since clouds are the major factor attenuating available ground-level solar irradiance, the irradiance es­timation under cloudy conditions exhibits huge uncertainties for both semi-empirical and physical models. To better account for the cloud effect, deep-learning models based on multispectral satellite images are proposed for GHI and direct normal irradiance (DNI) estimations. When compared with NSRDB, deep-learning methods show better overall results for both GHI and DNI estimations. The combination of deep learning and remote sensing shows potential in better extracting cloud information, which can better sup­port solar resource assessment, especially for cloudy conditions. Meanwhile, due to the spatio-temporal nature of solar irradiance, satellite data and satellite-derived products are extensively used in solar forecast­ing. To investigate the potential benefits of satellite-derived irradiance products and their improvements for solar forecasting, a comparative study for deterministic solar forecasts is performed. The results show that satellite-derived irradiance products generally outperform raw spectral satellite images, and the improved accuracy in satellite-derived irradiance products is likely to produce better forecasts. Similarly, solar fore­casting under frequent cloudy conditions is also associated with larger uncertainties. Nevertheless, satellite-derived irradiance with improved accuracy might lead to better forecasts, which is beneficial to a wide range of stakeholders in solar energy.