In Germany, grid operators mainly use solar power forecasts based on numerical weather forecasts to estimate the share of photovoltaic power in their grids. The forecast is usually made one day in advance. Short-term refinements of the forecast – 15 minutes to a few hours in advance – for marketing solar power on the intraday market or for grid load management are calculated on the basis of real-time measurements of PV feed-in and solar radiation forecasts from satellite images.

To further optimize the prediction of solar radiation for these short-term periods, scientists at Fraunhofer ISE developed an AI-based method. This method predicts cloud development in satellite images and evaluates it against a conventional method based on cloud movement vectors.

"With the help of the new method, we were able to significantly reduce prediction errors compared to the reference model across all prediction horizons examined – from 0 to 4 hours in 15-minute resolution intervals," summarizes Nils Straub, doctoral candidate at Fraunhofer ISE and lead author of the method. "On average, the prediction errors were 11 percent lower."

One weak point of predictions based on satellite images from the visible spectral range is the early morning hours. The quality of the images is impaired when the sun is low, and before sunrise they are completely black, so that no predictions can be calculated during these hours. 

The research team addressed this problem by adding two additional infrared channels to the visible images. These do not rely on direct sunlight and therefore also work in the dark. "Compared to a model that uses only images from the visible spectrum, we were able to significantly increase the availability of forecasts," adds Straub. "In Germany, over the course of a year, from before 8 a.m., from around 22 to up to 100 percent. A significant improvement in daytime forecasts was also a welcome side effect." The task of AI in the method is to predict cloud development and, in sunrise situations, to make infrared channels usable for radiation prediction.

"PV forecasting systems play an important role in solar power trading, grid management, and power plant deployment planning," explains Dr. Elke Lorenz, group manager for solar energy meteorology at Fraunhofer ISE. "The more the expansion of fluctuating renewable energies progresses, the more helpful they are. Although storage systems are increasingly taking on a stabilizing role for the power grid, PV feed-in forecasts can also ensure cost-efficient use here."

In addition to the daily and annual radiation profile, which can be calculated relatively easily and with a high degree of accuracy from the sun/earth constellation, clouds have the greatest influence on radiation at ground level and are much more difficult to predict. Radiation forecasting is therefore closely linked to the prediction of future cloud conditions. Solar radiation maps for the next few hours are then calculated from the predicted cloud conditions.