10:00 〜 10:15
[ATT30-10] A MACHINE LEARNING APPROACH FOR RELIABLE NEAR-REAL-TIME PREDICTION OF SOLAR IRRADIANCE FROM GEOSTATIONARY SATELLITE IMAGERY
Accurate short-term forecasting of solar irradiance is crucial for improving our understanding of climate dynamics. This enhanced predictability supports informed decision-making in climate change mitigation, disaster risk reduction, and strategic energy planning, addressing critical environmental and societal challenges.
Obtaining accurate solar irradiance for a specific location requires measurements from a weather station equipped with a pyranometer and sun-tracking devices. This is necessary due to the diffusion of solar rays by clouds, air molecules, and dust particles. However, producing near real-time solar irradiance on a global scale continues to be a challenging task.
Most current research relies on historical weather data sources (CAMS and MERRA2) that provide global coverage for irradiance. However, these sources have significantly low spatial resolution (around 70 km) and are generally updated on a bi-weekly to monthly basis.
Therefore, this study aims to develop a novel Machine Learning (ML) based approach for near-real-time solar irradiance prediction with high spatiotemporal resolution (at a 2 km level with 5-min granularity) by combining geostationary satellite (GOES-16) imagery with geo-orbiting atmospheric satellite (AURA-OMI) data. Our experiments have been conducted on one location basis to ensure the validity of our approach, but have the potential to expand the capacity to forecast at any location on the globe.
The GOES-16 data captures frequent changes in various methodological parameters like aerosol and cloud optical depth, cloud particle size, surface albedo, and precipitable water vapor at a 2-km spatial resolution, allowing us to capture the unique aspects of local weather characteristics.
Highly localized weather data is used by physics-driven mathematical models (REST2V5, MAC2) to estimate clear sky irradiance. This irradiance data is then utilized by Cloud Modification Models (Simplified Solis) and Radiative Transfer Models (MODTRAN, SMARTS) to estimate the actual irradiance at that location, based on cloud characteristics such as cloud coverage ratio, cloud optical depth, and cloud particle size. Finally, Artificial Neural Networks (ANN) and Gradient-Boosting Models (GBM) are trained using these actual irradiance outputs as labels from the mathematical models.
The Machine Learning (ML) models predicted actual irradiance using the selected meteorological features based on their influence. This selection was analyzed by trend-seasonality analysis, principal component analysis, and feature importance from GBM. As a result, compared to conventional methods, this approach simplifies the prediction of irradiance. It eliminates the need for future mathematical models and reduces the dimensionality of the required data.
The proposed method underwent testing at the ground-based weather station (SURFRAD) in Bondville, Illinois, USA, from the 26th of June to the 10th of July in 2023. We tested the ML models during the sunshine hours (5 AM to 8 PM local time) on three specific days (6th, 9th, and 10th July 2023), while the remaining days were used for model training.
Our model achieved a coefficient of determination of 0.96 when compared to the ground truth (SURFRAD data). This was approximately 10% higher than the prediction from global reanalysis of weather data (MERRA2, CAMS).
Obtaining accurate solar irradiance for a specific location requires measurements from a weather station equipped with a pyranometer and sun-tracking devices. This is necessary due to the diffusion of solar rays by clouds, air molecules, and dust particles. However, producing near real-time solar irradiance on a global scale continues to be a challenging task.
Most current research relies on historical weather data sources (CAMS and MERRA2) that provide global coverage for irradiance. However, these sources have significantly low spatial resolution (around 70 km) and are generally updated on a bi-weekly to monthly basis.
Therefore, this study aims to develop a novel Machine Learning (ML) based approach for near-real-time solar irradiance prediction with high spatiotemporal resolution (at a 2 km level with 5-min granularity) by combining geostationary satellite (GOES-16) imagery with geo-orbiting atmospheric satellite (AURA-OMI) data. Our experiments have been conducted on one location basis to ensure the validity of our approach, but have the potential to expand the capacity to forecast at any location on the globe.
The GOES-16 data captures frequent changes in various methodological parameters like aerosol and cloud optical depth, cloud particle size, surface albedo, and precipitable water vapor at a 2-km spatial resolution, allowing us to capture the unique aspects of local weather characteristics.
Highly localized weather data is used by physics-driven mathematical models (REST2V5, MAC2) to estimate clear sky irradiance. This irradiance data is then utilized by Cloud Modification Models (Simplified Solis) and Radiative Transfer Models (MODTRAN, SMARTS) to estimate the actual irradiance at that location, based on cloud characteristics such as cloud coverage ratio, cloud optical depth, and cloud particle size. Finally, Artificial Neural Networks (ANN) and Gradient-Boosting Models (GBM) are trained using these actual irradiance outputs as labels from the mathematical models.
The Machine Learning (ML) models predicted actual irradiance using the selected meteorological features based on their influence. This selection was analyzed by trend-seasonality analysis, principal component analysis, and feature importance from GBM. As a result, compared to conventional methods, this approach simplifies the prediction of irradiance. It eliminates the need for future mathematical models and reduces the dimensionality of the required data.
The proposed method underwent testing at the ground-based weather station (SURFRAD) in Bondville, Illinois, USA, from the 26th of June to the 10th of July in 2023. We tested the ML models during the sunshine hours (5 AM to 8 PM local time) on three specific days (6th, 9th, and 10th July 2023), while the remaining days were used for model training.
Our model achieved a coefficient of determination of 0.96 when compared to the ground truth (SURFRAD data). This was approximately 10% higher than the prediction from global reanalysis of weather data (MERRA2, CAMS).