Remarks on the sub-seasonal range

Sub-seasonal range structure

The sub-seasonal range ensemble is run daily based on 00UTC data.  The products cover the period up to Day46 and are derived from a 100 member ensemble with a control and has a resolution of 36km.   The sub-seasonal range ENS is independent of the medium range ENS and has it's own control and sub-seasonal range model climate (SUBS-M-Climate).

The sub-seasonal range ENS covers a time scale lying between:

  • medium range forecasts (ENS to day15).  These are mainly governed by atmospheric initial values (background plus new observed data) but less so on ocean temperature information.
  • seasonal forecasts. These are more reliant on predictability of the oceans and on the impact that tropical sea-surface temperatures have on the atmospheric circulation.

This is a particularly difficult time range as it is:

  • generally too long for the atmosphere to keep a memory of its initial conditions.
  • too short for the ocean variability to have an impact on the atmospheric circulation.

Sea-surface temperatures have an important influence upon the atmospheric evolution.  Ocean-atmosphere coupling is made at hourly intervals throughout the sub-seasonal range forecast period.  This high-frequency coupling:

  • can influence the development of synoptic-scale systems (e.g. tropical cyclones)
  • influences meteorological evolution and predictability in the extra-tropics
  • helps capture the propagation of Madden-Julian Oscillation (MJO) events, notably in the equatorial Indian Ocean and western Pacific ocean. 

The LIM2 subprogram (within NEMO) forecasts changes in the sea-surface temperature and sea-ice evolution.   Note: ECMWF uses LIM2 which is an earlier version of the Louvain-la-Neuve sea ice model currently available (Version 3.6).

The oceanic ensemble has been introduced to enable representation of the uncertainty of the sea-surface temperature and associated heat and moisture fluxes.  The initial conditions of the oceanic ensemble are perturbed using five ocean assimilations (1 control and 4 perturbed) produced by addition and subtraction of two randomly selected wind stress perturbations.  The same perturbations cannot be chosen for 2 consecutive months.  

Re-forecasts provide an sub-seasonal range climate (SUBS-M-Climate) and associated probability distribution functions (pdfs) for several variables.  The latest sub-seasonal range ensemble forecasts and the associated probability density functions can be compared with the SUBS-M-Climate.  The differences between the two are used as the basis of model products any model drift is effectively removed. 

Combating Model Drift

Drift of model calculations begins to be significant after 10 days of coupled integrations.  It displays similar patterns to seasonal forecasting after 6 months of integrations, but with less amplitude.

The strategy for dealing with model drift is straightforward: 

  • For all forecasts, the ocean, atmosphere and land surface are initialised to be as close to reality as possible.  Then the coupled models calculate the evolution of the atmospheric and oceanic systems.
  • No "artificial" terms are introduced to try to reduce the drift of the model.
  • No steps are taken to remove or reduce any imbalances in the coupled model initial state.

The effect of model drift can be estimated from re-forecasts and thus may be "removed" from the latest model solution during the post-processing.  This an expedient rather than a perfect solution.  Model drift characteristics may also depend somewhat on the prevailing synoptic pattern, and will not be accounted for fully.

The sub-seasonal range model climate (SUBS-M-Climate) drifts towards becoming rather too cold at longer lead-times in wintertime high latitudes.  Hence the anomaly in forecast temperatures against SUBS-M-climate temperatures may be too large.  The magnitude of the drift is not uniform.  At longer lead-times the trend in northern China is towards colder values but less so in Siberia and Canada.  The variation may be due to the analysed initial snowpack conditions and/or snowmelt in marginal snow cover areas in these areas.  Issues regarding this are being addressed. A multi-layer snow scheme is incorporated.  


After about 10 days of forecasts, the spread of the ensemble can become very large.  A significant shift can be detected by comparing probability distribution functions of the latest model and the SUBS-M-climate.

The re-forecasts are created twice a week (Mondays and Thursdays) and are ready a week before the real-time forecasting suite starts.  Real-time forecasts are calibrated using all the re-forecasts available in a one week window centred on the forecast start day and month. 



Fig5.2-1: Example of plumes for Dublin.  Extended Range forecast, DT00UTC 1 January 2018.  The plumes show increasing spread of forecasted values of 850hPa temperatures and 500hPa geopotential height within the extended range period.



(FUG Associated with Cy49r1)