Outstanding Operational Forecast Challenges Board

A case of week-2 forecast running to the opposite of the observation

     Date of initial condition : 11/21/06

     Forecast target period: 11/29-12/5/06

     Predictant:  Mean surface temperature anomaly

     Model:  NCEP GFS

     Problems:  (To get enlarged figure, click the figure link.)

i)    The model ensemble forecast was totally out of phase compared with the observation (Fig.1).

ii)   Further examined the near-range forecast and found that the 6-10 day forecast from 11/26 initial condition can correctly capture the 500 hPa height observed pattern, while that from 11/25 initial condition cannot.  (Fig.2)

Conjectures contributed by open discussions

  1. Tropical forcing interacting with high-latitude systems

  2. Deficiencies in model physics
  3. Predictability of weather regime change
  4. Land surface impact
  5. Data gaps and assimilation problems
  6. Ensemble formation

Interim summary of discussion  (2/28/2007)

This week-2 forecast problem from routine operation has got broad attention in both weather and climate communities at ESRL, ARL, CPC, EMC, NCAR, UA and IU. The implication of emerging ideas has gone far beyond the scope of this special case. Following is an interim summary, which stringed important pieces of idea together.

1. What do we need to focus on for improving the week-2 forecast?

a. Upstream regimes of weather system development

When we extending our forecast range from week-1 to week-2, our focused area should be moved from NA to its upstream areas. According to the ECMWF recent practice, improving precipitation modeling over the continental US had large benefit to their 1-2 week forecast over Europe and including NCEP GFS also improve their multi-model ensemble prediction skill. It would be beneficial to investigate critical regions in NA’s upstream, where weather system development influences the US w-2 forecast the most. In addition to those areas we’ve already paid a lot of attention to, e.g. tropical oceans, areas of AO/NAO, PNA, west of coasts ..., promising regions would also include the Tibetan Plateau, areas of prevailing Asian summer and winter monsoons, Eurasia snow cover, and the Arctic Ocean, where our model performance in each season is less assessed.

b. Day-1 forecast

As pointed out by Lorenz (1982), the best way to improve the weather forecast beyond day 1 is by improving the first day forecast. According to the critical information provided, the difference between the day-1 forecast started from 11/25 and that from 11/26 could be the key to understand the problem. If the difference looks like the random perturbation, the problem could be unpredictable. Otherwise we may conjecture that some important regional weather system development in the upstream could be missing in the model forecast. This would possibly be the case, since usually the global day-1 forecast skill is high, which implies the large scale features between the forecast and the verification are alike globally. By focusing on day-1, we can isolate/identify the problem more effectively and make clearer thinking on model deficiencies in physics and dynamics.

2. Physical processes and interactions

Tropical forcing, energy propagation and interaction with the strength of atmosphere-land/ocean coupling along the developing path have been discussed for the potential causes of the model forecast failure. Since the global model is far from faithfully representing regional characteristics, the regional model designed for regional applications are more capable to catch the regional system development than the global model. In such consideration, developing nested model with that the embedded regional model having two-way interactive lateral boundary, instead of one-way downscaling like most regional models currently doing, should be set as a R&D challenge to the research community.

(This case was presented by Mike Halpert at Review of the NCEP Products Suite, Camp Spring, 13 December 2006)