Medium Range Forecasting - using ensemble forecasts

Ensemble forecasts (ENS) offer the most consistent method of achieving good and consistent forecasts.  Interpretation of the output gives the ability to assess most likely outcomes and the probability of extreme weather and even an assessment of how extreme that weather might be.  It is vital that users adopt ways of working successfully with the ensemble.

Note:

  • ECMWF strategy 2015-2025 is centred on ENS development
  • The horizontal and vertical resolutions of the medium range ensemble (ENS) became the same as for the HRES after the upgrade to IFS (Cy48r1) introduced in June 2023.  The HRES forecast and the unperturbed ENS control forecast are meteorologically equivalent and equally skilful on average.  They have the same physical and dynamical representation of the atmosphere and use the same parameterisation of sub-gridscale effects.  However, they can diverge on a day-to-day basis due to small technical differences and the chaotic nature of the atmosphere.  The HRES will continue for the time being for ease of use by customers and users.

Use of the ensemble control member (CTRL)

The control member of the medium range ensemble has the same horizontal and vertical resolutions as HRES.  This does not mean that it may be used simply as a substitute for HRES working as a deterministic model yielding deterministic results.  As a forecasting aid it is not a direct substitute for HRES.

There are 50 other members of the ensemble.  All are equally valid.  There is no reason to select the results from the unperturbed control member (CTRL) rather than any of the others.   And any member viewed in isolation cannot provide any estimate of forecast uncertainty or confidence.   

The higher resolution (currently 9km) brings advantages and disadvantages - smaller scale atmospheric features are modelled and forecast, and look beguilingly realistic.  Development of these atmospheric systems often is in response to inherent numerical instability (which affects all numerical models) and reliance on detail is inappropriate.  

The main strategy to adopt is to avoid over-interpreting non-predictable features.  Therefore the detail of the most recent ensemble members should not be used in isolation.   Run-to-run jumpiness can on the one hand be tackled as something negative that has to be dampened, but on the other hand as something positive which can enrich the forecast information by giving alternative scenarios.  Ensemble members can give an indication of the probability and the consistency of features of the forecast.

Ensemble control member and HRES

Since introduction of Cy48r1 in June 2023, HRES and the ensemble control member have the same horizontal and vertical resolution.  However, there are some minor differences in the model structure.

HRES should be considered as just another (unperturbed) member of the ensemble.  Sometimes quite large differences from other ensemble members occur.  But analysis of HRES and ensemble forecasts give some evidence that HRES and Control forecast are meteorologically equivalent.  See further comments on the similarity and differences between HRES and ENS control.  

Use of the ensemble mean (EM)

Generally, whether the ensemble spread is small or large, the EM (or median if applicable) will, beyond the short range, exhibit higher accuracy than the ensemble control (CTRL).  This is particularly true for parameters, such as mean sea level pressure (MSLP) and temperature.  However, with increasing spread amongst ensemble members, it becomes more appropriate to couch the forecast information in the form of probabilities rather than as predicted values  - this is particularly true for parameters such as precipitation and cloud amounts.  The ensemble mean also displays a higher degree of day-to-day consistency.  Jumpiness in the ensemble mean is also markedly less, on average, than seen in the ensemble control, particularly when examining forecasts beyond about Day3. 

Criticism of the ensemble mean (EM)

Ensemble mean forecasts provide more accurate and considerably less “jumpy” solutions.  Averages of forecasts from the same or different NWP models are similarly more accurate.  However, meteorologists are somewhat apprehensive about using them.  This reluctance derives mainly from three reasons:

  • Ensemble averages do not constitute genuine, dynamically consistent, three-dimensional representations of the atmosphere.
  • Ensemble averages are less able to represent extreme or anomalous weather events.  Event probabilities or the extreme forecast index (EFI) should be used instead.
  • Ensemble averages might lead to inconsistencies between different parameters.  For example, the ensemble cloud average (or median) might not be consistent with the average (or median) of the precipitation.
  • On average, gradients in ensemble mean fields (e.g. ensemble average mean sea level pressure) systematically reduce with lead time, which can give misleading guidance on other parameters (e.g. wind strength, which is commonly inferred from the isobaric gradient).

A synoptic example of combining ensemble mean and probabilities

It is important to avoid over-interpretation of the ensemble mean, in particular underestimation of the risk of extreme weather events.   To aid visual interpretation by the user, ensemble mean output should be presented together with a measure of the ensemble spread.  The ensemble mean and the probabilities relate naturally to each other and can be most effective when shown together.  So, for example, the ensemble mean of the MSLP (or 1000hPa) presented together with gale probabilities will put the latter into a synoptic context that will help interpretation (see Fig6.1-1).

Fig6.1-1: 1000hPa forecast from 12UTC 13 August 2010 T+156hr to 00UTC 16 August 00 UTC T+96 h, all valid at 00UTC 20 August 2010.   Full lines are the 1000 hPa geopotential EM overlaid by the probabilities of wind speeds >10m/s.  Probabilities are coloured in 20% intervals starting from 20%.