Benchmarking satellite-based and numerical weather prediction models for national-scale intraday PV power forecasting

We introduce a novel framework that captures spatiotemporal dependencies for intraday photovoltaic (PV) power forecasting and employ it to systematically assess seven nowcasting models. The evaluated approaches cover a broad spectrum, ranging from satellite-driven deep learning models and optical-flow techniques to physics-based numerical weather prediction systems, and include both deterministic and probabilistic formulations. Their performance is examined with respect to accuracy, reliability, and forecast sharpness. The evaluation is conducted in two stages. First, forecast skill is assessed at the irradiance level using satellite-derived surface solar irradiance fields as reference. These irradiance predictions are then translated into PV power estimates through a station-specific machine learning model, which incorporates local irradiance together with solar azimuth and elevation angles as input features. This setup enables a consistent conversion from irradiance forecasts to power forecasts, which are subsequently validated against observations from 6434 PV systems distributed across Switzerland. To the best of our knowledge, this study constitutes the first nationwide assessment of spatiotemporal PV power forecasting. We further introduce new visualization techniques that reveal the impact of mesoscale cloud dynamics on PV generation at hourly and sub-hourly timescales. Our results show that satellite-based approaches outperform the Integrated Forecast System ensemble (IFS-ENS) at short lead times, although their skill decreases more rapidly with increasing forecast horizon. Among the evaluated methods, SolarSTEPS and SHADECast achieve the highest overall accuracy for both irradiance and PV power, with SHADECast also demonstrating the most reliable ensemble spread. While the deterministic IrradianceNet model yields the lowest root mean square error, probabilistic forecasts from SolarSTEPS and SHADECast provide better-calibrated uncertainty estimates. Forecast performance is found to deteriorate with increasing elevation, and conditions characterized by cloudiness and high variability remain particularly challenging. At the national scale, satellite-driven approaches accurately capture daily aggregate PV generation, achieving relative deviations under 10% on 82% of days over the 2019–2020 period. This demonstrates their strong robustness and potential for use in operational settings.