Analog‐based post‐processing of the ALADIN‐LAEF ensemble predictions in complex terrain

Abstract

AbstractThe main goal of this study is to assess the performance of the analog‐based post‐processing method applied to the Austrian ALADIN‐LAEF wind speed ensemble predictions through a set of sensitivity experiments. Evaluation of several analog‐based configurations using various meteorological variables as predictors is therefore conducted. The results of those experiments are compared to the ensemble model output statistics (EMOS) baseline model. The hypothesis further investigated is that using summarized measures, such as mean and standard deviation of an ensemble for several meteorological variables, is comparable to the analog post‐processing using all of the ensemble members. Results show that both analog‐based and EMOS experiments considerably improve the raw model forecast. Even though the improvement over raw model forecast is evident, large differences among nearby stations are noticed in the highly complex terrain. The processes at lower stations seem to be better represented by the raw model, which leads to a better input forecast to the post‐processing and a better overall result than for the mountain stations. The analog‐based method is overall comparable to or even outperforms the EMOS. Assessing the post‐processing performance for high wind speeds shows that the analog experiments can improve the raw forecast, exhibiting significantly higher skill than the EMOS. The difference among all analog experiments is less pronounced, especially the experiment using all of the raw model ensemble members and the one using summarized measures. Furthermore, it is demonstrated that the usage of summarized ensemble measures is an optimal way to improve the forecast skill, compared to the other analog‐based experiments and the EMOS model. Therefore, it is suggested that it is not necessary to increase the computational costs by using the full input spectrum of a raw probabilistic model, that is, all ALADIN‐LAEF members as predictors, as the summarized metric suffices.