Evaluation and improvement of surface shortwave downward radiation forecasts for solar energy production

In the current context of global change and energy transition, the share of solar energy in electricity production is increasing significantly. Forecasting solar energy production is both an issue in terms of facilitating its integration into the electric grid and ensuring its stability, and a challenge because of its high spatiotemporal variability and its strong dependence on meteorological conditions. In this context, forecasts of surface shortwave downward radiation (SWD) produced by numerical weather prediction (NWP) models are an essential element. However, the performance of NWP models in terms of solar radiation has rarely been evaluated, and very rarely over large areas and long periods. Quantifying the performance of NWP models is all the more important given that they remain limited. Furthermore, a detailed assessment of the various sources of error in SWD forecasts, which are multiple and complex, enables progress to be made in improving the models. This thesis responds to these challenges in two stages. Firstly by developing a methodology for evaluating the performance of high resolution NWP models for SWD, which is applied to Meteo-France's operational SWD model AROME, over the whole of mainland France and at a highly instrumented site, for the year 2020. Secondly, by suggesting possible improvements to the AROME model. More specifically, the AROME hourly forecasts for mainland France are initially based on SWD measurements from the network of 168 pyranometers operated by Meteo-France. Cloud products derived from geostationary satellite observations have also been used to classify high-frequency cloud situations across country. Analysis of these observations show that the situations contributing most to the SWD errors correspond to cloudy skies in the model and in the observations. These situations are very frequent situations and characterised by a significant positive annual bias. Missed cloudy situations and erroneously predicted clouds are relatively rare, so have a smaller impact, while the bias for well-predicted clear sky conditions is small. The positive bias in cloudy conditions seems to be mainly due to errors in the cloud optical thickness. Cloud fraction errors cannot be excluded, but are difficult to assess. In overcast conditions in the model, high clouds are associated with a positive SWD bias while low clouds are associated with a negative bias. A detailed analysis on the highly instrumented SIRTA site, where observations of cloud fraction and liquid water path (LWP) are also available, provides further details. It appears that the negative bias of SWD for boundary layer clouds is not only due to an overestimation of the modelled cloud fraction, but also to an overestimation of the LWP. The positive bias of SWD for high and geometrically thick clouds is associated with an underestimation of the LWP, possibly due to an underestimation of the total cloud water content or of the supercooled liquid water. Based on these findings, new AROME simulations were produced over two months, including modifications to the radiative and microphysical schemes. Improvements in SWD forecasts appear when the snow content is included in the radiative scheme. In addition, reducing droplet concentration improves the SWD for low clouds. Finally, SWD is very sensitive to the subgrid cloud heterogeneity factor. These various issues should be explored in future work with the aim of obtaining a version of AROME optimised for forecasting solar energy production. (author)