A U.S. research team has developed a machine learning model that predicts variability in surface solar irradiance using cloud type and cloud cover as inputs. The model was originally developed and trained at a single site in Oklahoma, and the researchers have now tested its performance across 15 additional sites worldwide to evaluate how well it generalizes beyond its original training location.

“Riihimaki has developed in 2021 a machine learning model that predicts surface solar irradiance variability from cloud type and cloud cover from five years of cloud radar, lidars, and surface radiation observations at the Atmospheric Radiation Measurement Program (ARM) Southern Great Plains (SGP) site in Oklahoma,” the group said. “This study complements that study by evaluating the model’s performance and applicability in different climates at 15 additional sites.”

In the 2021 study, the group used data recorded between 2014 and 2018 at the Oklahoma site to train a random forest model. The model used cloud type and cloud cover as inputs to predict mean effective transmissivity (ET), the standard deviation of ET, and, in particular, the standard deviation of minute-to-minute changes in ET. This last metric captures rapid solar “ramp” events – sudden increases or drops in solar irradiance caused by moving clouds – which are important for grid operations.

The 2021 results showed that cloud type and cloud cover alone could explain 42% of rapid fluctuations in sunlight caused by moving clouds. This led the authors to hypothesize that the same relationship would hold across other climates. They therefore expanded the analysis to 15 additional sites, including other ARM locations in Alaska, Australia, Papua New Guinea, the Azores, Argentina, Texas, Colorado and California, as well as stations from the National Oceanic and Atmospheric Administration (NOAA) Surface Radiation Budget Network (SURFRAD) in Illinois, Nevada, Montana, Mississippi, Pennsylvania, South Dakota and Colorado.

However, as the scope of the prediction was expanded, modifications to the original model were required. In the 2021 study, cloud cover was derived from a Total Sky Imager (TSI), whereas in the more recent work it was obtained using RADFLUX, which estimates cloud cover from surface radiation measurements. The researchers also tested a second cloud-type method used at NOAA SURFRAD stations, based on radiation data and ceilometer cloud-base heights instead of cloud radar and lidar. This enabled them to assess whether the model remained robust beyond the original instrument configuration and could be applied more broadly.

“In terms of coefficient of determination (r2), half of the sites (53%) have the same r2 or better than in the original research. Of the remaining sites with a smaller r2, nearly half are within 0.1 of the original’s r2. This indicates that nearly three-quarters (73%) of sites have the same predictability or better than the original,” the academics explained. “In terms of mean squared error (MSE), all sites have small MSEs, and all are within 0.0015 of the original study (0.0035). The results here confirm the hypothesis that the relationship is largely applicable to locations with other cloud climatologies distinct from the central United States.”

However, the results also showed that some locations and cloud types exhibited lower predictability of solar variability. These were mainly sites in more extreme environments than the Oklahoma reference site, including mountainous, arid, tropical, and high-latitude regions. The site in Alaska showed the lowest r2 values for nearly all cloud types.

The new model was described in “Prediction of solar variability by cloud type and cloud cover,” published in Solar Energy. Researchers from the United States’ University of Colorado Boulder, NOAA Global Monitoring Laboratory, and NOAA Global Systems Laboratory have contributed to the study.

This content is protected by copyright and may not be reused. If you want to cooperate with us and would like to reuse some of our content, please contact: editors@pv-magazine.com.