- Machine learning model successfully predicted solar variability across 15 additional global research sites.
- Nearly 75% of locations delivered predictability equal to or better than the original benchmark study.
- Findings could strengthen next day solar forecasting for utilities, grid operators and energy traders.
Researchers have confirmed that a machine learning model designed to predict solar variability using cloud type and cloud cover can perform reliably across a wide range of climates and geographies, marking an important development for renewable energy forecasting and electricity market operations.
The study titled ‘Prediction of solar variability by cloud type and cloud cover’, expanded on earlier work conducted at the Atmospheric Radiation Measurement Program Southern Great Plains site in Oklahoma, where scientists first developed a model capable of forecasting short term fluctuations in surface solar irradiance based on cloud conditions.
To test whether the model could be applied more broadly, researchers evaluated its performance across 15 additional sites, including Atmospheric Radiation Measurement Program locations around the world and National Oceanic and Atmospheric Administration Surface Radiation Network sites spread across the continental United States.
The results showed that the relationship between cloud conditions and solar variability remained consistent regardless of location, climate or observation network. Researchers found that 73% of the sites achieved the same or better predictability than the original study, with r2 values ranging from 0.37 to 0.54 compared with the original benchmark of 0.42.
Importantly, all sites recorded low mean squared error values, remaining within 0.0015 of the original study’s result of 0.0035. According to the researchers, this demonstrates that the model is largely location independent and capable of operating effectively across different cloud climatologies.
The study also confirmed that the model was not overly dependent on specific instruments or cloud observation products. Previous research relied on Total Sky Imager data for cloud cover measurements, while the latest study used RADFLUX observations, which require only surface radiative flux measurements and less specialised instrumentation.
Researchers said this is a significant advantage because RADFLUX data is more widely accessible, potentially allowing broader adoption of the forecasting method across energy markets and solar forecasting applications.
The findings further showed that solar variability predictability remained robust regardless of the instruments used to classify cloud types. While Atmospheric Radiation Measurement Program sites use advanced cloud radar and lidar systems, NOAA SURFRAD sites rely on ceilometers and surface radiation measurements.
However, the study identified some limitations in more extreme environments. Lower predictability was observed in mountainous, tropical, arid and high latitude regions, particularly at the Alaska North Slope site, where the model demonstrated little forecasting skill for several cloud types.
Researchers believe this may be linked to the higher occurrence of super cooled liquid and mixed phase clouds in such regions, which can produce greater optical depth than expected and complicate irradiance forecasting.
Despite these challenges, the study concluded that the forecasting approach remains broadly applicable and represents another step toward operational next day solar variability forecasting based on numerical weather prediction model output.
The researchers noted that future work will focus on improving forecasting performance in difficult conditions, simplifying cloud classifications and incorporating cloud optical depth information to further improve prediction accuracy.
The long-term goal is to create an operational solar variability forecasting product capable of supporting utilities, grid operators and renewable energy market participants as solar generation continues to scale globally.
Link to the full paper HERE
Author: Bryan Groenendaal
