Real time forecasts

The first major component is the hydrological forecasts produced in real time. This gives the predicted conditions for the sub-seasonal and seasonal products to be compared to the climatologies to derive the forecast anomaly and uncertainty. The characteristics of the real time forecast simulations are described below. Where appropriate, the difference between EFAS/GloFAS and sub-seasonal/seasonal is specified. If there is no EFAS/GloFAS or sub-seasonal/seasonal mentioned, then the method is identical between the systems:

• Hydrological model used: LISFLOOD (LISFLOOD*)
• Meteorological forcing:
  • Sub-seasonal:
    • The combination of the 9km (horizontal resolution) ECMWF ensemble forecasts and the 36km ECMWF sub-seasonal forecasts from the previous run date (see EFAS Meteorological forecasts and GloFAS meteorological forecasts for further details)
    • The low-resolution sub-seasonal forcing is taken from the previous available forecast run, simply because of timing, as the sub-seasonal meteo forecasts for the same day are only available with several hours delay after the high-resolution forecasts. 
    • The high- and low-resolution forcing is combined by a simple mechanical blending by using the high-resolution meteorological forcing for days 1-15, while the low-resolution meteorological forcing for days 16-45, for each of the 51 ensemble members in the same order (i.e. member 1 of ECMWF ensemble forecasts and member 1 of ECMWF sub-seasonal forecasts). 
    • The blending can create inconsistencies locally over complex orographical areas, due to the resolution change at day15, and also there could be some smaller inconsistencies between the ensemble because of the forcing change from 15 to 16 day, due to the mechanical blending which will mix different weather conditions from the low- and high-resolution meteorological data. However, overall as an ensemble of 51 scenarios, the blending is not expected to lead to any larger discontinuities.
  • Seasonal:
    • 36 km seasonal forecasts (SEAS5, LINK)
  • Number of ensemble members: 51 ensemble members (one unperturbed and 50 perturbed) to represent the equally likely forecast scenarios of the future in both sub-seasonal and seasonal.
• River resolution:
  • 1 arcmin (~1.5 km) in EFAS
  • 3 arcmin (~5 km) in GloFAS
• Run frequency:
  • Sub-seasonal forecasts are generated daily at 00 UTC
  • Seasonal forecasts are generated monthly on the 1st of each month at 00 UTC
• Lead time:
  • Sub-seasonal has 45-day (1080 hours) lead time. Only 45 days, as although the low-resolution sub-seasonal meteo forecasts run out to 46 days, we can only rely on the 1-day-old (yesterday's) run, which means we can only run the simulations to 45 days ahead.
  • Seasonal has 7-month lead time (strictly speaking 215 days for each run date)
• Forecast step:
  
  • Sub-seasonal:
    • EFAS: 6-hourly
    • GloFAS: 24-hourly 
  • Seasonal: 24-hourly in both EFAS and GloFAS
• Hydrological forecast initialisation:
  • Sub-seasonal is initialised from a fillup simulation forced with the shortest-range ENS-Control (unperturbed member of the ECMWF ensemble forecast) meteorological conditions.
  • Seasonal is initialised from hydrological monitoring simulation, which is forced with gridded meteorological observations in EFAS and ERA5 meteorological reanalysis data in GloFAS.

Reforecasts

The second major component is the hydrological reforecasts, that are used to generate the model climatologies. These climatologies are range-dependent, i.e. they change with the lead time. The climatologies are produced from a large set of hydrological reforecasts, which allows for robust estimate of the climatological behaviour of the systems. The characteristics of the reforecast simulations are described below. Where appropriate, the difference between EFAS/GloFAS and sub-seasonal/seasonal is specified. If there is no EFAS/GloFAS or sub-seasonal/seasonal mentioned, then the method is identical between the systems:

• Hydrological model used: LISFLOOD (LISFLOOD*)
• Meteorological forcing:
  • Sub-seasonal:
    • The combination of the 9km (horizontal resolution) ECMWF ensemble forecasts and the 36km ECMWF sub-seasonal forecasts from the same run date (see EFAS Meteorological forecasts and GloFAS meteorological forecasts for further details)
    • The high-resolution medium-range and low-resolution sub-seasonal forcings are taken from the same forecast run date. The timing of the availability of the low-resolution forecasts is not an issue here, as these reforecasts are only produced retrospectively without any time constraint.
    • The high- and low-resolution forcing is combined by a simple mechanical blending by using the high resolution meteorological forcing for days 1-15, while the low-resolution meteorological forcing for days 16-46, for each of the 11 ensemble members in the same order (i.e. member 1 of ECMWF ensemble forecasts and member 1 of ECMWF sub-seasonal forecasts)
    • The blending can create inconsistencies locally over very complex orographical areas, due to the resolution change at day15, and also there could be some smaller inconsistencies between the ensemble because of the forcing change from 15 to 16 day, due to the mechanical blending which will mix different weather conditions from the low- and high-resolution meteorological data. However, overall as an ensemble of 11 scenarios, this is not be expected to lead to any larger discontinuities.
  • Seasonal:
    • 36 km seasonal forecasts (SEAS5, LINK)
  • Number of ensemble members:
    • Sub-seasonal has 11 ensemble members (one unperturbed and 10 perturbed) to represent the equally likely forecast scenarios of the future.
    • Seasonal has 25 ensemble members (one unperturbed and 10 perturbed) to represent the equally likely forecast scenarios of the future.
• River resolution:
  • 1 arcmin (~1.5 km) in EFAS
  • 3 arcmin (~5 km) in GloFAS.
• Run frequency:
  • Sub-seasonal reforecasts are generated for the past 20 years. Before 12 Nov 2024, they were generated twice-weekly on Mondays and Thursdays (at 00 UTC), while from 12 Nov 2024 they are generated on given days of the months, i.e. 1, 5, 9, 13, 17, 21, 25 and 29 (excluding 29 Feb).
  • Seasonal reforecasts are generated for each month of the past back to 1981, although we only use them from 2004-2023. The shorter period, in the same way as for the sub-seasonal, provides better better stationarity (see https://www.ecmwf.int/en/elibrary/81194-trends-glofas-era5-river-discharge-reanalysis), which can be very important for the reliability of the forecasts.
• Lead time:
  • Sub-seasonal reforecasts have 46-day (1104 hours) lead time. This is 1-day longer than the real time forecasts, as these for the reforecasts there is no time constraint and the blending can be done with the high-resolution and low-resolution meteorological forecasts of the same run date (no need to delay by 1 day the low-resolution).
  • Seasonal reforecasts have the same 7-month (or 215-day) lead time.  
• Forecast steps:
  • Sub-seasonal:
    • EFAS: 6-hourly
    • GloFAS: 24-hourly 
  • Seasonal: 24-hourly in both EFAS and GloFAS
• Forecast hydrological initialisation
  • All reforecasts are initialised from the hydrological monitoring simulation, which is forced with gridded meteorological observations in EFAS and ERA5 meteorological reanalysis data in GloFAS. There is no reason to use fillup, since these are produced for dates well in the past (as for the real time forecasts).

Climatologies

The third major component is the range-dependent climatologies, that are generated from the hydrological reforecasts. The climatologies give the reference point for the forecast anomaly computation. The reference points are the percentiles of the climate distribution from 1^st  to 99^th. The characteristics of the climatologies are described below. Where appropriate, the difference between EFAS/GloFAS and sub-seasonal/seasonal is specified. If there is no EFAS/GloFAS or sub-seasonal/seasonal mentioned, then the method is identical between the systems:

• Climate generation dates:
  • In the sub-seasonal, climate files are produced for each reforecast run date, so in total 8 (days per month)*12-1 = 95 dates in a calendar year with 1/5/9/13/17/21/25/29 of each month, excluding 29 February. The sub-seasonal climate files are generated continuously in a real time manner, right after the hydrological reforecasts are produced.
  • The seasonal climatology is produced only once for each month of the year (so 12 sets).
• Climate lead times:
  • In the sub-seasonal, for each of these climate dates, one climate file per daily lead time of the 46-day sub-seasonal range if produced. Here, all possible 7-day lead times, starting from days1-7, then days2-8, ..., out to days 40-46 (40 possible lead times) are considered. This way, the weekly mean climatology for all possible lead times will be available, and so for each real time forecast the right climatology can be used, with the correct lead time in days, depending on which day of the week the real time forecast run happens (i.e. which corresponding climate lead time to choose in order to get to the calendar weeks). For example, if the sub-seasonal forecast run is on a Wednesday, then the first lead time will be the following Monday-Sunday and the corresponding climatology will have the lead time of days6-12. However, if the forecast date is on Monday, then the first lead time will be days1-7 of the climatology, etc.
• Climate sample:
  • The sub-seasonal uses 20 years of hydrological reforecasts with run dates roughly twice a week. For the climate sample, always 3 run dates are combined, the actual climate date (we generate the climate for) and one reforecast date before and after. The 3 dates combined will guarantee more robust estimates of the percentiles (especially the most extreme ones), with no impact on the seasonal variability (i.e. combining too many dates could negatively impact on the behaviour in climate zones with rapid shift between seasons). For example, when generating the climate sample for 15 December 2024, all reforecasts produced for 11, 15 and 19 December will be used in the climate sample from years 2004-2023 (so, 11 Dec 2004, 11 Dec 2025, ... 11 Dec 2023, then 15 Dec 2004, 15 Dec 2025, ..., 15 Dec 2023, then 19 Dec 2004, 19 Dec 2005, ..., 19 Dec 2023). The total size of the climate sample is 3*20*11 = 660.
  • The seasonal also uses 20 years of hydrological reforecasts. For each month, this will mean 20 sets of reforecasts in the climate sample, using all 25 ensemble members. The total size of the climate sample is 20 * 25 = 500.
• Climate distribution: The climate will be generated from the climate sample by sorting the 660/500 values for each river pixel. The percentiles then are produced by dividing the climate range into 100 equally likely bins, separated by the percentiles from 1 to 99. The 1^st percentile is the value that is exceeded 99% of the time, while the 99^th percentile is the value that the reforecasts will exceed only 1% of the time, based on the 20-year climate period. These climate percentiles will be specific to the 40 possible lead times from days1-7 to days 40-46 and to all reforecast run dates in the sub-seasonal, and will be specific for all 12 months and for all the 7 monthly lead times.