Post-processing of global horizontal irradiance forecasts using machine learning

The accuracy of solar forecasting is key for the electricity grid. The increase in installed solar capacity is a major challenge for the grid and the installation of solar power plants is becoming increasingly popular. Numerical weather prediction (NWP) models provide the best forecasts in the day-ahead timeframe, but these forecasts can also be improved by post-processing.In this research, data from two different numerical weather prediction models were analysed: the AROME (Applications of Research to Operations at Mesoscale) and ECMWF (European Centre for Medium-Range Weather Forecasts) models for global horizontal irradiance forecasts for 6 different locations in Hungary. Five machine learning models were used, within which 12 predictor combinations were created. Initially, the two numerical weather prediction models were treated separately, but the impact of combining the two on the accuracy of the results was also investigated. At the end of the work, the goodness of the forecasts was analysed on the basis of the uniform verification criteria.The results showed that an increase in altitude leads to larger prediction errors. Kékestető, which is the highest point in Hungary, had the highest RMSE (root mean square error) value, while Debrecen, which is located on the plain, had the lowest RMSE value. On Kékestető, the best models lowered the initial 137.3 W/m
²
AROME RMSE and 128.0 W/m
²
ECMWF RMSE values to 116.5 W/m
²
, in Debrecen 124.0 W/m
²
AROME RMSE and 108.3 W/m
²
ECMWF RMSE values to 97.6 W/m
²
.