Clustering

The clustering process groups individual ensemble members that are "similar" according to some measure (or norm).  This compresses the large amount of information produced by ensemble members and highlights the most predictable parts.  Clustering can be performed:

  • over different geographical areas
  • for different parameters
  • for each forecast time
  • over different forecast intervals.
  • or a combination of the above.

Currently clustering summarises the range of synoptic flow patterns from the ensemble over a restricted area in two forms: 

  • Weather scenario clustering (up to day10 of medium range forecast output). It is performed on medium range ensemble forecast output.  The area covers Europe and the northwest Atlantic.  It highlights the differences between medium range scenarios in terms of the large scale flow.  
  • Weather regime clustering (beyond day10 of medium range forecast output).  These summaries are based on one of four climatological regimes which have been derived from reanalysed data.  The area covers Europe, eastern Canada, and the north Atlantic.  It provides information about the possible transitions between regimes.
  • Weather regime clustering (seasonal forecast output).  These summaries are based on one of six climatological regimes which have been derived from reanalysed data over a slightly larger area.  The area covers Europe, eastern Canada, the North Pole region and the north Atlantic.  It provides information about the possible transitions between regimes.

Clustering is a compromise between:

  •  the advantage of condensing a large amount of ensemble output.
  • the disadvantage of losing possibly important information.

A convenient overview of ensemble members forecast information can be gained from using:

  • a cluster mean (the average over all members in a cluster).
  • a most representative member (MRM or cluster scenario).  This is the ensemble member closest to the centre of the cluster centre and typifies the members within that cluster.  Each most representative member is then attributed to one of the weather regime clusters.


(FUG Associated with Cy49r1)