Global verification of numerical solar radiation forecast

Renewable energies are becoming increasingly important in energy production due to proceeding climate change. This makes the energy sector one of the fastest growing and changing sectors worldwide. Solar energy is expected to account for large parts of the world’s future energy supply.

Solar radiation forecast will become more important with the future expansion of existing energy supply structures with solar energy resources. Due to the natural fluctuation of solar radiation, a reliable forecast is required for an efficient use of solar energy. This study deals with a solar radiation verification of the global numerical weather prediction models NEMSGLOBAL, GFS, ICON, ARPEGE-World, as well as the reanalysis model ERA5. The aim is to find out which models performs the best and how multi-model approaches can lower the forecast error.

For the verification 81 global solar radiation measurements from the Basic Surface Radiation Network (BSRN) and the World Radiation Data Center (WRDC) were used. Hourly forecast data are compared to quality controlled and aggregated measurement data for the years 2018 until 2020. Statistical analysis is conducted for each measurement location separately to evaluate and compare the performance of each raw or multi-model by using the error metrics like mean absolute error (MAE) and mean bias error (MBE).

Among all models, the reanalysis model ERA5 performed the best with a MAE of 43 Wm^-2 and a MBE of 8 Wm^-2. For the weather forecast models, ICON showed the lowest error with a MAE of 48 Wm^-2 followed by GFS, NEMSGLOBAL and ARPEGE-World with MAEs of 51 Wm^-2, 61 Wm^-2 and 67 Wm^-2. Unlike the other forecast models, NEMSGLOBAL and ARPEGE-World tend to underestimate solar radiation. Within the best performing multi-models, ICON is usually weighted the highest with up to 50 %. Implementing a multi-model approach for a solar radiation forecast, the MAE can be reduced as the number of models increases. Including two models in the multi-model the MAE can already be reduced by 4.6 Wm^-2, while three models reducing the MAE by 7 Wm^-2, and four models by 8.8 Wm^-2.

Results indicate that the multi-model approach can improve solar radiation forecast. That shows the potential of further investigation of forecast models and their combination.