Comparison of irradiation data from different numerical weather models and their combination in multi-model forecasts

The forecast of solar radiation is increasingly gaining economic importance for various applications such as building control in smart home systems and grid integration or direct marketing of solar power. Day-ahead forecasts of PV production are based on irradiation forecasts from weather models. These weather models have been continually improved in recent years, and some of them are compared in this research paper with the usual error measures. More and more often probabilistic methods based on ensemble predictions are used. A weakness of these ensembles is often that the individual ensemble members behave in parallel, especially in critical ramp events, so that the ramp occurs in all ensemble members at the same time, even if in reality the weather front often one hour earlier or later arrives. An alternative probabilistic multimodal approach, which is already applied in the market, is averaging from different weather models. The different weather models use different algorithms and rely on different data assimilations, which is the reason for greater variation than with ensemble forecasts of single-models.