Statistical postprocessing aims to reduce systematic biases and dispersion errors of forecast ensembles
provided by state-of-the-art numerical weather prediction (NWP) models. Often, this also entails an implicit mapping from NWP forecasts valid for grid cell means to point forecasts valid at specific points of interest within a given grid cell. In general, statistical postprocessing increases forecast skill considerably. However, due to the strongly non-Gaussian distribution and low predictability of precipitation at points, statistical postprocessing of precipitation forecasts is particularly difficult. ecPoint, a postprocessing approach developed at the European Centre for Medium-Range Weather Forecasts (ECMWF) is tailored to point forecasting of precipitation and proved to increase forecast skill considerably. Moreover, ecPoint is data efficient in that it needs only one year of training data from the global ECMWF-IFS NWP model. The combination of data efficiency and forecast skill of ecPoint calls for a comparison with (postprocessed) probabilistic forecasts of precipitation provided by high-resolution limited area NWP models like COSMO-E, which is a MeteoSwiss configuration of the Consortium for Small-scale Modeling (COSMO) model.
In this study, we compare ecPoint forecasts for 12 h accumulated precipitation with ensemble model output statistics (EMOS) as a reference postprocessing method. We assess the performance of raw ecPoint and raw COSMO-E alongside EMOS applied to pooled ensembles constructed using either COSMO-E and ECMWFIFS or COSMO-E and ecPoint with varying weights. Verifying the different forecasts on a set of about 850 gauge stations in Switzerland and neighboring areas confirms the good performance of ecPoint. For long lead times and heavy precipitation, ecPoint tends to be more skillful than EMOS. However, further research is needed to assess the impact of the lengths of the respective training periods on the relative skill of ecPoint compared with EMOS. Moreover, it will be beneficial to identify in which regions and in which meteorological regimes ecPoint ordinarily outperforms forecasts based on a high-resolution limited area model, and vice versa.
Statistical postprocessing aims to reduce systematic biases and dispersion errors of forecast ensembles
provided by state-of-the-art numerical weather prediction (NWP) models. Often, this also entails an implicit mapping from NWP forecasts valid for grid cell means to point forecasts valid at specific points of interest within a given grid cell. In general, statistical postprocessing increases forecast skill considerably. However, due to the strongly non-Gaussian distribution and low predictability of precipitation at points, statistical postprocessing of precipitation forecasts is particularly difficult. ecPoint, a postprocessing approach developed at the European Centre for Medium-Range Weather Forecasts (ECMWF) is tailored to point forecasting of precipitation and proved to increase forecast skill considerably. Moreover, ecPoint is data efficient in that it needs only one year of training data from the global ECMWF-IFS NWP model. The combination of data efficiency and forecast skill of ecPoint calls for a comparison with (postprocessed) probabilistic forecasts of precipitation provided by high-resolution limited area NWP models like COSMO-E, which is a MeteoSwiss configuration of the Consortium for Small-scale Modeling (COSMO) model.
In this study, we compare ecPoint forecasts for 12 h accumulated precipitation with ensemble model output statistics (EMOS) as a reference postprocessing method. We assess the performance of raw ecPoint and raw COSMO-E alongside EMOS applied to pooled ensembles constructed using either COSMO-E and ECMWFIFS or COSMO-E and ecPoint with varying weights. Verifying the different forecasts on a set of about 850 gauge stations in Switzerland and neighboring areas confirms the good performance of ecPoint. For long lead times and heavy precipitation, ecPoint tends to be more skillful than EMOS. However, further research is needed to assess the impact of the lengths of the respective training periods on the relative skill of ecPoint compared with EMOS. Moreover, it will be beneficial to identify in which regions and in which meteorological regimes ecPoint ordinarily outperforms forecasts based on a high-resolution limited area model, and vice versa.