Clouds can change the amount of sunlight reaching solar panels in minutes, complicating electricity planning. A new forecasting model that combines sky photographs with historical weather measurements reduced prediction errors in tests using data from two US locations.
What researchers studied
Researchers developed a machine-learning system called TSGAN to forecast global horizontal irradiance, a measure of incoming solar energy. The model combines recent ground measurements with sequences of sky images, adjusting how much attention it gives each source.
It was tested against numerical-only, image-only and other combined-data models using public datasets from Albuquerque, New Mexico, in 2021 and Eugene, Oregon, in 2024. Observations were aligned at ten-minute intervals, and the system predicted irradiance ten, twenty and thirty minutes ahead.
The researchers separated training, validation and testing data chronologically to limit information leakage and analysed results across four seasons.
What they found
Relative to image-only baselines, the model reduced mean absolute error by 32.38% in the Albuquerque dataset and 25.98% in the Eugene dataset. Normalised root mean squared error fell by 23.32% and 17.02%, respectively.
The researchers also compared the system with models that use historical measurements alone and with other multimodal approaches. Their approach generally performed better across the two sites and maintained an advantage over different forecast horizons, although accuracy deteriorated as predictions extended further ahead.
The findings suggest that cloud images add timely information about changes that historical numerical readings can struggle to capture.
What it means
For solar-power operators, better very-short-term sunlight forecasts could eventually support battery scheduling, grid balancing and responses to sudden changes in renewable electricity production.
This is a model comparison, however, not a demonstration that a power grid saved money or avoided outages. Its value lies in showing a promising method for using complementary measurements, with operational benefits still to be established.
Important limitation
Evaluation relied on two locations and selected years, with forecasts no more than thirty minutes ahead. The model produced single-value predictions rather than uncertainty intervals, and the paper calls for cross-site testing under changing conditions before broader deployment.
Source Information
Study: State-aware adaptive multimodal fusion for ultra-short-term solar irradiance forecasting using sky cloud images
Authors: Houhang Zhao, Aiyi Song, Yu Jin and Haoxiang Yu
Journal / Source: PLOS ONE
Published: 3 November 2026
