Enhancing the Retrieval Accuracy of Surface Solar Radiation from Meteosat with AI-Based Emulation and Fine-Tuning

Short-term solar radiation forecasts are becoming increasingly important for electricity grid operators as solar forecasts allow for predictions of expected photovoltaic (PV) power generation of the upcoming minutes to hours. Satellites in geostationary orbit, such as Meteosat with its Spinning Enhanced Visible and Infrared Imager (SEVIRI), deliver spectral observations that are essential for many surface solar irradiance (SSI) estimation and forecasting techniques. Yet, existing solar forecast methods still depend on derived level-2 SSI products, causing considerable delays and inaccuracies in the forecasts (Carpentieri et al., 2023; Paletta et al., 2023). We propose a deep learning based SSI retrieval that estimates SSI from Meteosat level-1 observations of Earth without relying on traditional radiative transfer based retrieval algorithms (Verbois et al., 2023).

We show that convolutional residual networks can estimate SSI at higher accuracy and lower latency than traditional radiative-transfer-based retrievals such as Heliosat-based SARAH-3 (Cano et al., 1986; Pfeifroth et al., 2023). We demonstrate that the convolutional residual networks can accurately emulate instantaneous SSI from SARAH-3 using the SEVIRI visible and infrared channels across Europe. The SSI estimates of our emulator exhibit similar accuracies on pyranometer ground stations across Europe and North Africa. Fine-tuning the SSI emulator on pyranometer ground stations induces a further leap in accuracy while maintaining geographically unbiased SSI estimates. The fine-tuned SSI emulator significantly surpasses the performance of the Heliosat SARAH-3 algorithm, especially in conditions dominated by clouds. We characterise the SSI emulator and discuss the relevance of the various visible and infrared channels for accurate SSI estimates based on feature importance.

REFERENCES

D. Cano, J. M. Monget, M. Albuisson, H. Guillard, N. Regas, and L. Wald, 1986,

A method for the determination of the global solar radiation from meteorological

satellite data, Solar Energy 37, 31–39. doi:10.1016/0038-

092X(86)90104-0.

A. Carpentieri, D. Folini, D. Nerini, S. Pulkkinen, M. Wild, and A. Meyer, 2023, Intraday probabilistic forecasts of surface solar radiation with cloud scale-dependent autoregressive advection, Applied Energy, vol. 351, doi:10.

1016/j.apenergy.2023.121775.

Q. Paletta, G. Terren-Serrano, Y. Nie, et al., 2023, Advances in solar forecasting: Computer vision with deep learning, Advances in Applied Energy, vol. 11, p. 100 150, doi:10.1016/j.adapen.2023.100150.

U. Pfeifroth, S. Kothe, J. Drücke, J. Trentmann, M. Schröder, N. Selbach,

and R. Hollmann, Surface Radiation Data Set - Heliosat (SARAH) - Edition 3,

2023, doi:10.5676/EUM_SAF_CM/SARAH/V003.

H. Verbois, Y.-M. Saint-Drenan, V. Becquet, B. Gschwind, and P. Blanc, 2023,

Retrieval of surface solar irradiance from satellite imagery using machine

learning: Pitfalls and perspectives, Atmospheric Measurement Techniques, vol. 16, no. 18, pp. 4165–4181, doi:10.5194/amt-16-4165-2023.

- Indicates paper has been withdrawn from meeting

- Indicates an Award Winner

See more of: AI Applications in the Energy Sector
See more of: 16th Conference on Weather, Water, Climate, and the New Energy Economy

Previous Abstract | Next Abstract >>