Second SRNWP Workshop on Statistical and Dynamical
Adaptation 5-7 May 2003, Vienna, Austria
Or as it has been said to the
participants: What would you answer to your Director if he would ask you to
work now to the improvement of the model and no longer for the adaptation? Some answers: - Models are improving but are
not yet perfect: it is thus perfectly justified to continue to work to the
improvement of the model results. - We are now asked to make very
local weather forecasts. Today's model resolutions cannot take into account
very local effects. Thus we must relay on statistical adaptation. - If you want to correct a weakness
in the results of a numerical model, you can do it much more rapidly with
a statistical correction than by changing the model formulation. - Some needed parameters are
not computed in the model. Example: visibility. This has to be done in the
post-processing.
Is rmse a good
score? Maybe, but one
thing is sure: it must only be used with other scores. The reason is well
known: it rewards smooth, undetailed
results (as 2D-fields) and punish the scientist who does work at high resolution
or develops schemes prone to handle extreme situations
(cf. presentation of Anders Persson). If rmse is used
for the verification of 2D-fields, it must be accompanied by the variance.
Thus very smooth fields, which will have low rmse, will be punished with the
variance. It has been said
that for point verification (normally an observing station) we use to many
scores. After definition of thresholds, "hit rate" and "false alarm rate"
should be sufficient. Probability
forecasts
The works presented
were based on the ECMWF ensembles of 51 members. Participants
have been asked whether they have remarks about the ECMWF products. Two remarks have been made: - 51 members with a good spread
are enough. No need is seen for more members (the Centre plans to increase
the ensemble to 100 members) - a colleague thinks that the
deterministic high resolution 12 UTC integration should be suppressed and
the effort put on an increase of the resolution of the ensemble.
Next meeting A colleague proposed to hold
this workshop annually instead of every two years. But the majority of the
participants who aired their opinion on that
point clearly favored a 2-year cycle. The next meeting will take place
in May 2005 in Vienna.
Jean Quiby SRNWP Network Coordinator
Is statistical adaptation still necessary?
Assessment of the model results: are we using
the right scores?