Evaluation of surface solar irradiance forecasts by the NWP model AROME

In the context of global warming, the share of renewable energy sources (RES) is drastically increasing. A promising source to transit towards a more sustainable energy system is solar energy. However, like many others RES, solar energy is intermittent with high spatio-temporal variability and high dependence on weather conditions, so that the control over production is limited. As a consequence surface solar irradiance (SWD) forecasts based on Numerical Weather Prediction (NWP) models are essential to help incorporating solar energy into the electrical grid and ensure the network stability.

The performances of NWP models in terms of solar radiation have rarely been studied, though, especially over large domains and long periods. However, the growing interest from various end users including the photovoltaic community is now shedding light on this critical question. Errors in NWP models can be consequent. For instance the annual mean bias and RMSE for 2020 amount 18Wm^-2 and 97Wm^-2, respectively, for AROME, the operational NWP model of Météo-France at 1.3km horizontal resolution. Our objectives is to characterize the performance of AROME for solar radiation, which will allows in a second step to improve these performances by refining the physical parameterizations impacting solar radiation, primarily the microphysical scheme and the radiative code.

To this end, a full year of hourly AROME forecasts is compared to corresponding in-situ SWD measurements from the network of pyranometers operated by Météo-France. This network gathers about 180 high-quality pyranometers over metropolitan France. A thorough analysis of cloud satellite products is also carried on to identify situations which contribute most to radiation errors. Then physical processes are prioritized to be improved in AROME to reduce errors.

Results show that the first source of errors occurs when the sky is cloudy in both model and observations with an annual bias of 24Wm^-2 (contributing to 89% of the total bias) against 3Wm^-2 in clear skies (contributing to 3% of the total bias). The missed cloudy situations and the false alarms contribute respectively to 14% and -6% of the total bias. While part of the bias in clear sky condition is due to an unrealistic representation of aerosols in AROME, the bias in cloudy conditions seems to be mostly related to incorrect cloud optical properties. Further investigations allows to split these errors into liquid water path and effective radius errors, and to focus on the most problematic cloud types. In particular, not accounting for snow in the radiation scheme contributes to the positive bias, and misrepresenting cirrus clouds seem to contribute mostly to solar radiation errors.