Page History

...

It might seem attractive to identify the member which verifies best in the early part of the forecast (say at T+12) and assume this member will continue to provide the best forecast during the rest of the medium range period.  But this is not true; the performance of any member during the first 12hrs of the forecast has little relevance to its skill beyond T+48hrs in the same area.

In order to compress the amount of information being produced by the ensemble and highlight the most predictable parts, individual ensemble members that are "similar" according to some measure (or norm) can be grouped together.   This process is known as clustering.  

Fig8.1.1.1: A example of postage stamps showing PMSL forecasts from ensemble control, and all 50 ensemble members, DT 00UTC 19 May 2017, VT T+120hr at 00UTC 24 May 2017.  Some large differences in the pressure pattern can be seen on individual ensemble members.  Each member has been allocated a to a cluster, shown in a different colour above each frame for lead-times T+120 and above only.  The representative member of each cluster is here shown by arrows. The clusters are shown in Fig8.1.1.2.

In order to compress the amount of information being produced by the ensemble and highlight the most predictable parts, individual ensemble members that are "similar" according to some measure (or norm) can be grouped together.   This process is known as Clustering.  

Fig8.1.1.2: Clustering for  or the case shown in Fig8.1.1.1.  The three clusters for T+120hr are in the left column.  Clustering is based on 500hPa geopotential height pattern.  

...