Challenges for improving solar nowcasting using Meteosat Third-Generation satellite imagery

Accurate short-term solar energy forecasting is essential for the safe and stable integration of a growing share of photovoltaic electricity generation into power grids. Intra-day forecasts up to a 6h prediction horizon are essential for grid management, trading on the electricity spot market and energy storage exploitation. Estimate irradiance using images from geostationary meteorological satellites is particularly appropriate for intraday projections, giving better performances than numerical weather prediction models.

The deployment of third-generation satellites provides unprecedented spatial and temporal resolution by delivering images every 10 minutes on a 500 meter grid at nadir. This high-frequency data offers a significant opportunity to improve short-term irradiance nowcasting via precise cloud displacement analyses. However, standard methods based on cloud motion vectors (CMV) computation do not inherently benefit from this finer resolution. Instead, they face increased sensitivity to parallax and cloud shadow effects alongside difficulties in extracting global motion tendencies. Recently, deep-learning has been extensively applied to satellite based forecasting tasks. ConvLSTM architectures, combining temporal recurrence with spatial convolutions, proved effective in reproducing the evolution of complex cloud structures.

This study explores novel forecasting approaches using Meteosat Third Generation (MTG) imager data. Firstly, a CMV based model has been improved by applying a parallax and cloud shadow correction to the input satellite images. Secondly, a ConvLSTM model has been set-up on MTG data to take advantage of the finer spatio-temporal resolution.

High-resolution images from the year 2025 captured every 10 minutes by the visible narrow channel centered on a 0.6μm wavelength of the MTG sensor have been used to produce CMV forecasts as well as train and test the deep learning model. Results accuracy is assessed against ground based measurements from several Baseline Station Radiation Network (BSRN) pyranometers located across Europe. Cloud top height data obtained from an operational version of the SAF-NWC/GEO software are used to correct parallax and shadow effects in addition to training the deep learning model.

Our results demonstrate that the improved CMV based model yields competitive accuracy scores while unlocking the benefits of high-frequency data, particularly for predicting ramp events and rapid fluctuations. The ConvLSTM approach captured sophisticated spatio-temporal patterns inherent to high-resolution imagery on numerous case studies. These findings suggest that, while finer data introduces more complexity, dedicated correction and non-linear modeling are key to improving solar forecasts.