Combining ensembles using a new synoptic-scale curvature diagnostic

ECMWF is now running its extended-range 101-member ensemble, with an approximate grid spacing of 36 km, every day from 00UTC, out to 46 days ahead. This is in conjunction with the 9 km medium-range 51-member ensemble that is run 4 times a day (out to 15 days ahead from 00 and 12 UTC, and 6 days ahead from 06 and 18 UTC). It means that there are effectively 152 ensemble members available from 00UTC in the overlap period, and even more than that if time-lagging is used to include earlier forecasts.

Now that the two ensembles are run every day, work is underway to investigate the potential benefit from combining them and to get a better understanding of how they differ. Many previous studies have already shown that an increase in ensemble size or a combination of ensembles improves skill, especially for longer lead times as members diverge more.

One of the difficulties of blending ensembles at two different resolutions is that each has different biases and representativeness of point observations. Hence the assumption that all members are equally likely is not true since they form two different distributions. That problem could be alleviated by calibrating one to the other or both to observations, which would involve applying statistical methods to a very large ensemble which is not a trivial undertaking.

The work presented here introduces a new, very simple, diagnostic that identifies cyclonic and anticyclonic regions using surface pressure or geopotential height, computed over scales ranging between 500 and 6000km. This has many benefits for assessing ensembles as well as more generally. The prediction of cyclonic and anticyclonic regions is one of the most fundamental aspects of a weather forecast to get correct and hence diagnose. The approach removes the need for any calibration because the scales are much larger than the grid spacing of each ensemble, enabling comparison with analyses processed in the same way. It is also possible to apply standard or spatial ensemble verification metrics, detect any systematic differences in the synoptic pattern between ensembles, and measure predictability over different scales. 

Once the diagnostic has been introduced, highlights of comparative behaviour and verification scores will be presented for the individual and combined ensembles using one year of forecasts, along with any selected cases that are helpful for understanding the overall scores.