Solar energy forms an important pillar of climate change mitigation. Short-term

forecasts of surface solar irradiance (SSI) are gaining more importance for power grid

operators seeking to balance supply and demand in a secure and economical way.

Regional-scale SSI forecasts are essential since most solar power is provided by

decentralized PV plants. Regional-scale SSI estimates can be obtained from infrared

and visible Earth imagery of geostationary satellites such as Meteosat. Solar nowcast

models oEer SSI predictions at forecast lead times of minutes to hours. Current

regional-scale solar nowcast models often require satellite-derived SSI products based

on SSI satellite retrievals like Heliosat SARAH-3. Moreover, current solar nowcast

models are typically deterministic, lacking the ability to quantify forecast uncertainty.

I provide an overview of recent developments in spatiotemporal forecasting of solar

energy across Europe. For example, I will introduce the first two regional-scale solar

nowcast models, SolarSTEPS and SHADECast, which provide probabilistic satellitebased

solar forecasts for minutes to hours ahead (Carpentieri et al., 2023, 2024). Our

models simulate multiple forecast realisations for any given time and location which

enables accurate forecasts and uncertainty quantification for regions up to thousands

of kilometers in size. Our solar nowcast models perform autoregressive and generative

AI simulations. We demonstrate that they enhance forecast accuracy under all

cloudiness conditions. Our generative SHADECast model enhances the forecast

horizon of the state-of-the-art SolarSTEPS model by 26 minutes for lead times of up to

two hours. Additionally, we present a deep learning emulator of the Heliosat SARAH-3

SSI product (Pfeifroth et al., 2021), which predicts kilometer-scale SSI estimates at 15-

minute intervals based on visible and infrared Meteosat imagery. The emulator, a

convolutional residual network, estimates instantaneous SSI across Europe with

accuracy comparable to SARAH-3. It even surpasses SARAH-3 in SSI accuracy when

retrained on pyranometer stations (BSRN, IEA-PVPS, national weather services),

allowing for more accurate SSI initialization of solar nowcast models.