New FSU model for seasonal predictions

[yzerfan]

http://fsu.com/pages/2009/07/15/hurrica ... model.html

In addition, the model outperformed many statistical and other dynamical models, LaRow said. Statistical models use statistical relationships between oceanic and atmospheric variables to make a forecast, while dynamical models, such as the COAPS model, require major computing resources in order to make trillions of calculations using the equations of motion along with the best physical understanding of the atmosphere.

The COAPS model uses the university's high-performance computer to synthesize massive amounts of information including atmospheric, ocean and land data. A key component of the COAPS model is NOAA's forecast of sea surface temperatures. But COAPS researchers continue to study their own model in an effort to better understand the relationship between sea surface temperatures and climate predictability.

[wxman57]

No details in the article, but they forecast 8 named storms with 4 hurricanes for 2009. My 10/5/2 may be a little high.

[clfenwi]

See viewtopic.php?f=31&t=105785 for a couple of background links.

For those wondering why the forecast didn't include intense hurricanes as is the custom for other seasonal predictions, it is because the model cannot generate storms with surface winds greater than 50 m/s (~100 knots).

[Ed Mahmoud]

To actually generate and count storms, I assume this dynamic model actually runs like a dynamic model, ie, like the GFS or Euro, starting with an initial set of data at a given time, and actually develops tropical cyclones.


Shouldn't it be able to give time and location/landfall predictions, then? They may not be right, but for a dynamic model to develop 8 storms, four of which are hurricanes, it should be able to pick places and dates.

BTW, how often is this run? Once a week, once a month? How long does it take to run, and how does its grid scale/wave number compare to other global models.

[clfenwi]

Don't have time at the moment for a fuller response, but here's a couple tidbits drawn from the paper Atlantic Basin Season Hurricane Simulations (pdf).

Regarding landfall predictions, there's this concluding discussion of the 1986-2005 hindcasts done using the model (the body of the paper, of course, has more detail):

The tracks from the ensembles showed that the
model tends to move the storms away from the United
States too quickly, when observations show that storms
come much closer to the eastern seaboard before re-
curving. These results are an outcome of our atmo-
spheric model’s Atlantic large-scale steering flow (ap-
proximated by the 500-hPa streamlines). The 500-hPa
streamlines show that during the first half of the hurri-
cane season (JJA), the westward flow over the Atlantic
tends to turn northeast before it approaches the east
coast of the United States because of the break in the
high-pressure ridge. The use of a different convection
scheme resulted in better landfalling statistics, a result
of better large-scale steering flow and the fact that the
Zhang and McFarlane (1995) convection scheme gen-
erated about 30% more storms than the Hogan and
Rosmond (1991) convection scheme. However, the bet-
ter landfalling statistics did not translate into better in-
terannual storm variability

From the abstract, some details on the resolution/composition of the model:

An ensemble of seasonal Atlantic hurricane simulations is conducted using The Florida State University/Center for Ocean–Atmospheric Prediction Studies (FSU–COAPS) global spectral model (Cocke andLaRow) at a resolution of T126L27 (a Gaussian grid spacing of 0.94°). Four integrations comprising the ensembles were generated using the European Centre for Medium-Range Weather Forecasts (ECMWF) time-lagged initial atmospheric conditions centered on 1 June for the 20 yr from 1986 to 2005. The sea surface temperatures (SSTs) were updated weekly using the Reynolds et al. observed data. An objective-tracking algorithm obtained from the ECMWF and modified for this model’s resolution was used to detect
and track the storms.

[clfenwi]

This shows what's described in the first quote of my previous quote: The model's hindcast output of landfalling storms was well below the number of actual landfalls.

Comparing the resolution of this model to a few of the global models (TxxLyy, where x=number of waves and L=number of vertical layers; higher numbers=higher resolution):

FSU-COAPS T126L27 (.94°)
ECMWF T799L91 (.225°)
GFS T382L64 to 180h, T190L64 to 384h (.32°, .73°)
NOGAPS T239L30

The paper makes a brief assertion that higher resolution does not necessarily improve the forecasts storm-numbers wise and that the decisions made for model physics are more important.

As far as the time it takes to run the model goes, using the time it takes to run the coarse GFS as a rough guide, it would seem that it would take a day or two to run the model (assuming, of course, equal computing power). *If* the press release is right in stating that a forecast was issued on June 1, it couldn't take longer than that as the model was run using a May 30 forecast of SSTs.

As far as I can tell, it was a run once forecast, with the numbers being the mean of the forecasts generated by the four ensemble members.