Dear Jonathan,
Thanks again for the response.
We already thought of using realization. However, current operational
and research systems are more complex than what the "realization"
standard name envisages.
For instance, imagine a multi-model forecast (for the medium-range,
monthly, seasonal or interannual time scales) built up from forecasts
issued by several GCMs or by a number of versions of a single GCM. Each
GCM or model version is supposed to have the possibility of issuing an
ensemble of forecasts by perturbing the initial conditions. The members
of such an multi-forecast system ensemble forecast can not all be dealt
with in a similar way. For a given GCM or single-model version, the
ensemble members might have similar statistical properties (e.g., in an
initial-condition perturbation ensemble setting), but those ensemble
members will likely have different statistical properties when compared
to those from another GCM or single-model version, which are also part
of the same multi-forecast system ensemble forecast. In my opinion, it
is impossible to identify to which forecast system each member belongs
to when only using the realization variable.
The example above is only one among a set of multi-forecast system
predictions currently generated operationally (eg, the EUROSIP
multi-model seasonal forecasts) or under development (TIGGE medium-range
forecasts or the ENSEMBLES seasonal-to-decadal predictions). The
closeness to the operational activity prevents some of these communities
from using NetCDF because of the lack of adequate standards (and use
GRIB1 or GRIB2 instead). We believe that the variables
experiment_identifier (along with its auxiliary variables) and
ensemble_member_number can handle this problem.
Best regards,
Paco
Jonathan Gregory wrote:
> Dear Paco
>
> ----- Forwarded message from Francisco Doblas-Reyes <Francisco.Doblas-Reyes at ecmwf.int> -----
>
>> A NetCDF file with multi-forecast system ensemble data requires, at
>> least, the experiment id to be a dimension of the file. For instance,
>> ensembles generated with different models, which are part of the same
>> multi-model ensemble forecast, need to be identified in case the user
>> wishes to give different weights to the models or needs to bias correct
>> them separately. That is the reason why the information can not be
>> encoded as global attributes and we require the new standard names.
>
> For that purpose, perhaps you could use the new standard name of realization,
> which was recently added for a similar purpose. To quote from Alison's posting:
>
> realization; 1; realization is used to label a dimension that can be thought of
> as a statistical sample, e.g., labelling members of a model ensemble.
>
> This is expected to be a dimensionless number. If you need to use strings,
> we probably need a different standard name.
>
>> Concerning the use of time coordinates, we can try to use two variables
>> (reftime and leadtime) with standard names forecast_reference_time and
>> forecast_period but using a single time dimension, as in your example of
>> structure (b). I prefer the use of forecast_period because it makes more
>> obvious that we are dealing with forecasts.
>
> That makes sense to me. If structure (b) is acceptable to you, I am happy.
>
> Best wishes
>
> Jonathan
> _______________________________________________
> CF-metadata mailing list
> CF-metadata at cgd.ucar.edu
> http://www.cgd.ucar.edu/mailman/listinfo/cf-metadata
--
________________________________________
Francisco J. Doblas-Reyes
European Centre for Medium-Range
Weather Forecasting (ECMWF)
Shinfield Park, RG2 9AX
Reading, UK
Tel: +44 (0)118 9499 655
Fax: +44 (0)118 9869 450
f.doblas-reyes at ecmwf.int
_______________________________________
Received on Wed Oct 18 2006 - 06:49:23 BST
