A spatiotemporal framework for intraday PV power forecasting using satellite-based and numerical weather prediction models

We introduce a novel spatiotemporal framework for intraday photovoltaic (PV) power forecasting and apply it to a systematic comparison of seven PV nowcasting approaches, assessing their accuracy, reliability and sharpness. The benchmarked methods range from satellite-based deep learning and optical-flow techniques to physics-based numerical weather prediction models, and include both deterministic and probabilistic configurations. Model performance is first evaluated at the irradiance level using satellite-derived surface solar irradiance fields as reference data. The irradiance forecasts are subsequently converted into PV power estimates using a station-specific machine-learning-based irradiance-to-power model, which takes local solar irradiance and local solar azimuth and elevation angles as predictors. This approach enables the transformation of solar irradiance forecasts into PV power forecasts. The latter are validated against measured production from 6434 PV installations across Switzerland. To our knowledge, this work represents the first national-scale analysis of spatiotemporal PV power forecasting. In addition, we present novel visualizations illustrating the influence of mesoscale cloud dynamics on national PV generation at hourly and sub-hourly temporal resolutions. The results indicate that satellite-based models consistently outperform the Integrated Forecast System ensemble (IFS-ENS) at short forecast horizons, although their performance degrades more rapidly than that of IFS-ENS as lead time increases. SolarSTEPS and SHADECast yield the highest accuracy in both irradiance and power predictions, with SHADECast exhibiting the most reliable ensemble dispersion. While the deterministic IrradianceNet model achieves the lowest root mean square error, probabilistic forecasts from SolarSTEPS and SHADECast provide superior uncertainty calibration. Forecast skill is found to decline with increasing elevation. Moreover, cloudy and high-variability weather conditions remain the most challenging for PV power forecasting. At the national level, satellite-based models reproduce daily total PV production with relative errors below 10% for 82% of days during 2019–2020, highlighting their robustness and suitability for operational deployment.