This figure provides climatological averages of precipitation, which are the means of weekly values of cumulative precipitation of the given calendar week computed over the 1990-2019 historical period based on DWD data. For the same calendar week, cumulative precipitation forecast from ECMWF Ensemble Extended is bias corrected and shown on the right. One can then compare climatological long-term averages with precipitation forecasts the upcoming weeks. DWD station data are interpolated using the External Drift Kriging (EDK) developed by Samaniego et al. (2011).

This figure provides climatological averages of maximum air temperature, which are the means of the maximum values of the variable computed over the 1990-2019 historical period based on DWD data. For the same calendar week, maximum air temperature forecast from ECMWF Ensemble Extended product is bias corrected and shown on the right. One can then compare climatological long-term averages with forecasts of the upcoming weeks. DWD station data are interpolated using the External Drift Kriging (EDK) developed by Samaniego et al. (2011).

Sub-seasonal to Seasonal (S2S) Forecasts

Sub-seasonal to seasonal (S2S) forecasts aims to bridge the gap between medium-range weather forecasts (up to 10 days) and seasonal climate predictions (above a month). Compared to medium-range, sub-seasonal predictions (sometimes referred as extended) are issued on a daily basis, or once/twice per week but with increased lead times up to about 45 days. A real-time S2S forecasting system could provide tailored information for early warning of high impact extreme events. For HS2S application, 6-hourly real-time and re-forecast products from European Centre for Medium-Range Weather Forecasts (ECMWF) are retrieved. For more information, see ECMWF Ensemble Extended forecasts.  

Forecast Terminology

• Forecast start date: The day on which a forecast is issued (e.g. July 6, 2020, Monday).

• Forecast (target) period: Time period analyzed in a forecast. In other words, the aggregated period for which a forecast is valid. It will be provided in weekly basis based on calendar weeks in this web-page.

• Forecast lead time: The time period between the forecast start date and the beginning of the forecast target period. For example, the forecasts issued on Monday has 0-week lead-time for the same calendar week (target period).
  

• Initial condition uncertainty: The output of numerical weather predictions are uncertain. On source of the uncertainty is the initial conditions of the atmosphere. By accepting this uncertainty, centers issuing weather and climate forecasts produce probabilistic forecasts based on Ensemble Prediction System (ENS). Different realisations are generated by perturbation of an unperturbed (control) forecast.
  

Pre-processing Data

Data generated from Earth-system models usually requires pre-processing. All information need to be converted into file formats which are conventional for the description of Earth sciences data. Big data consists of many records from in-situ observation and numerical weather prediction are processed in this stage to be compatible with other software tools in modeling chain. This step includes quality control of data and temporal aggregation of the original atmospheric model outputs (6 hourly) to the target period in which the forecast is disseminated or to the resolution required for next modeling chain.

Bias Correction

The raw outputs of Numerical Weather Prediction systems have model biases raised from the many sources of systematic errors. A comparison of ECMWF real-time ensemble forecasts and DWD grided data between April 2020 to March 2021 has been done for precipitation. Mean bias ranges between -1 and 1 mm/day for all forecast lead-times. Systematic error needs to be removed from atmospheric model outputs to avoid propagation to model chain (e.g. hydrological modeling).    

This figure illustrates ECMWF ensemble extended deterministic skill (mean bias) for all real-time forecasts initialized between April 1, 2020 and March 31, 2021 (114 start day in total). All forecasts issued in each month are averaged first, then the spatial average is calculated for all grid cells located in Germany in native model resolution (0.2 degree for 0 and 1-week lead times and 0.4 degree for 2 and 3-week lead times). DWD data are interpolated based on near real-time in-situ observations using External Kriging Drift (EDK) to generate grided observation fields.

To remove the bias from atmospheric forecast ensembles, a trend-preserving approach will be applied before hydrological simulation. It follows a modified version of the method propsed by Lange (2019) for S2S time scale. More information on the method can be found in Ulysses project.