Discussion and Conclusions

In this case, we have seen that the SREF ensembles did not properly capture the evolution of an inland Mid-Atlantic snowstorm, resulting in significant forecast errors in QPF and snowfall amount up to the day of the storm. We have attributed the SREF (and operational NWP model) forecast errors to:

Initial condition errors

The data assimilation systems for the models used in the SREF were having difficulty analyzing the short-wave trough that ultimately resulted in the snowstorm. This difficulty began before 12 UTC 6 January, but was most obvious at that time, with radiosonde observations significantly different from the analysis increments used to update the initial atmospheric state.

Data Assimilation "Catch-up"

If we consider that the operational model and SREF were gradually trending toward the final outcome, it is clear that the data assimilation systems of the operational Eta and the SREF ensembles were playing "catch-up" with the atmosphere. Because of the assumption in the data assimilations systems that the first guess (i.e. 3-hr forecast in the EDAS, 6-hr forecast in the GDAS) is the best place to start for a current analysis and should have a very heavy weight in constructing the analysis, it takes awhile to 'nudge' a first guess gone awry to a good rendition of the atmospheric state. Unfortunately for the forecasters in PA, the data assimilation errors had been reasonably corrected only by the time the snow was falling.

SREF failure to capture the range of uncertainty in the atmospheric initial state

This is most obvious in the 9 UTC 6 January 2002 SREF ensemble members, and can be seen both in comparisons to the raobs and to satellite observations. One or two of the ensemble members came close to capturing the initial evolution of the storm when we compared the 12 UTC 6 January raobs to the 3-hour ensemble forecasts. We showed an example of how to use that ensemble member to adjust the NWP forecast, based on subsequent observations.

Conclusions

• CHECK THE INITIAL CONDITIONS!
  • Determine whether the range of ensemble initial states incorporates all the uncertainty in the initial atmospheric state, using the same tools as for deterministic NWP forecasts.
  • In cases where the SREF (or any ensemble) forecast does not contain the initial atmospheric state, neither the ensemble mean nor any individual ensemble member will likely encompass the forecast evolution properly.
• The ensemble forecast can be adjusted to account for the expected resolution using physical and dynamical reasoning, using the ensemble member most resembling the initial condition and the initial forecast evolution, etc.