Retrieving global radiation from Meteosat Level-1 with deep learning and pyranometer fine-tuning

Short-term forecasts of solar radiation are gaining importance for power grid operators, providing estimates of surface solar irradiance (SSI) and anticipated photovoltaic (PV) power generation for minutes to hours ahead. Geostationary satellites such as the Meteosat Second Generation equipped with the Spinning Enhanced Visible and Infrared Imager (SEVIRI) provide valuable spectral imagery for deriving SSI across wide geographical areas. While these spectral measurements are foundational for many SSI estimation and forecasting methods, current approaches still use Level-2 products of SSI, which leads to significant delays in the forecasting process. We demonstrate that a convolutional residual network can accurately emulate SSI produced by the SARAH-3 algorithm.

A generalized deep-learning model is trained to estimate the instantaneous SARAH-3 SSI from all channels of the SEVIRI imager over a large region in Europe. The SSI emulator shows a similar bias and root mean square error (RMSE) for a large validation pyranometer set across Europe and Northern Africa. The emulator directly infers SSI from the Level-1.5 spectral imagery of SEVIRI within just 15 seconds, strongly reducing the 10+ minutes runtime of non-machine-learning radiation retrievals such as SARAH-3. We present a characterisation of the SSI emulator's performance depending on location and season and quantify the channel importance for different regions and surface albedos. We also present the SSI emulator's performance over snowy surfaces where retrieval algorithms such as SARAH-3 have been struggling to distinguish between snow and clouds. To further improve the emulator, advanced fine-tuning methods are applied based on ground observations without degradation in regional bias. The fine-tuning improves the RMSE on average from 80 to 65 W/m^2 and outperforms the SARAH-3 algorithm.