To compress the amount of information being
produced by the EPS and highlight the predictable and thus relevant
parts, individual EPS forecast members that are "similar" according
to some norm are traditionally grouped together and averaged to
constitute new forecast fields, known as clusters. There is
no objective measure to determine which type of clustering is
“best” and the norm for judging what is
“similar” can be defined in different ways. The
clusters are not supposed to be used as forecasts but, rather, give
a comprehensive overview of the ensemble forecast information.
Clustering can be
performed over different geographical areas and on different
parameters; it can be done for each forecast time or for different
forecast trajectories. Every possible clustering is a
compromise: the advantage of condensing information is balanced by
the disadvantage of losing information that, on some occasions, in
hindsight, might have been important.
At
ECMWF two types of clustering are currently applied, one is based
on “weather scenarios” and the other on “weather
regimes”; another variant, “tubing”, is a
combination of a
“refined” ensemble mean and ensemble
outliers.
|
