AI-based cloud predictions sharpen short-term solar power forecasts

Combining AI-supported cloud predictions and infrared technology can improve overall solar radiation forecasting significantly, particularly in early mornings, according to researchers. By predicting cloud development with AI, researchers at the Fraunhofer Institute for Solar Energy Systems ISE managed to reduce the number of errors in short-term solar radiation forecasts by 11 percent on average, they said.

Solar radiation forecasts are important for grid operators and solar power suppliers, as the forecasts enable them to plan better. “PV forecasting systems play an important role in solar power trading, grid management and power plant deployment planning,” said Elke Lorenz, group leader for solar energy meteorology at Fraunhofer ISE, in a press release. “The more the expansion of fluctuating renewable energies progresses, the more helpful they are.”

While daily and annual radiation profiles can be predicted well based on the earth’s position in relation to the sun, cloud development is more unpredictable and has a major influence on solar radiation. The AI-supported forecasting method predicts cloud development for the next 15 minutes to four hours based on previous images, optimising short-term predictions. On average, the number of errors was 11 percent lower compared to the conventional reference model.

One challenge that remained in solar radiation forecasting with satellite images was making predictions for the early morning, as the images are less clear when there is no or limited sunlight. The researchers overcame this problem by adding two infrared channels to the images, which can work without sunlight. “Compared to a model that uses only images from the visible spectrum, we were able to significantly increase the availability of forecasts,” said Nils Straub, researcher at Fraunhofer ISE.