Estimating cloud coverage with deep learning methods: Key to solar power production

July 13, 2021

In a study published in a special issue of the journal Atmosphere, a team of researchers from Argonne National Laboratory presented a novel ground-based approach using machine learning for estimating cloud cover.

Many of us might wonder, “Don’t we have satellites to detect clouds?” Indeed we do, but these have been used principally for estimating coverage over large areas. For hyperlocal analysis – analysis at the neighborhood scale – using satellite images is costly. Researchers therefore prefer ground-based sky-facing instruments and have used machine learning methods with such instruments for segmenting cloud regions based on colors or shapes and features in the images obtained. But here, too, a problem arises: segmentation is difficult under overcast conditions or thin clouds.

To address this limitation, the Argonne team examined three deep neural networks and two regression models. They used over 1,500 images from the Singapore Whole sky Imaging SEGmentation (SWIMSEG) and HYTA datasets, as well as a new dataset they created using the Argonne-developed Waggle platform. The models were trained on the Chameleon computing cluster, with each model calculating the probability of each pixel representing a cloud and classifying the cloud pixels. After being validated, the six trained models were used to estimate cloud cover and solar irradiance. Four cases were analyzed: clear, haze cloudy, partially cloudy, and overcast with a mixture and thick and thin clouds. See Fig. 1.

Fig. 1: Results from six trained models used in the study

“Our tests showed that cloud cover was correlated with solar irradiance and that solar irradiance was highly correlated with solar power production. These results are critical for understanding solar irradiance and estimating production from photovoltaic solar facilities,” said Seongha Park, a postdoctoral appointee in Argonne’s Mathematics and Computer Science Division and lead author of the paper describing the team’s research.

Arguably, several challenges still remain. For example, new methods are needed to better distinguish cloud thickness, perhaps using infrared camera images to provide thermal information or using video to incorporate cloud motion and formation. The researchers believe, however, that by deploying the new machine learning prediction approach on an edge computing platform such as Waggle or SAGE, real-time in situ estimates of solar irradiance and solar power production will be possible.

For the full paper, see S. Park, Y. Kim, N. Ferrier, S. Collis, R. Sankaran, and P. Beckman, Prediction of Solar Irradiance and Photovoltaic Solar Energy, Atmosphere 12(3)395, 2021. https://www.mdpi.com/2073-4433/12/3/395, in the special issue “Machine Learning Applications in Earth System Science”