Grid- Versus Station-Based Postprocessing of Ensemble Temperature Forecasts

Statistical postprocessing aims to improve ensemble model output by delivering calibrated predictive distributions. To train and assess these methods, it is crucial to choose appropriate verification data. Reanalyses cover the entire globe on the same spatiotemporal scale as the forecasting model, while observation stations are scattered across planet Earth. Here we compare the benefits of postprocessing with gridded analyses against postprocessing at observation sites. In a case study, we apply local Ensemble Model Output Statistics to 2-m temperature forecasts by the European Centre for Medium-Range-Weather Forecasts ensemble system. Our evaluation period ranges from November 2016 to December 2017. Postprocessing yields improvements over the raw ensemble at all lead times. The relative improvement achieved by postprocessing is greater when trained and verified against station observations. Plain Language Summary To this day, weather forecasts are uncertain and subject to error. Statistical postprocessing aims to remove systematic deficiencies from the output of numerical weather prediction models. To apply these statistical methods, training and reference data are required. Weather observation sites are scattered across planet Earth. An alternative source of training and reference data is provided by so-called analyses, which combine weather observations with past forecasts to provide gridded pseudo-data with full global coverage. In this study we consider forecasts of surface temperature from the European Centre for Medium-Range Weather Forecasts. We find that the benefits of postprocessing are greater when it is performed directly on observational data, as opposed to using gridded analyses. In both cases, statistical postprocessing yields improved temperature forecasts at lead times from a single day to more than 2 weeks ahead.