Ensemble WRF simulations of the 1938 New England Hurricane Using 20th Century Reanalysis
Such simulations can provide insight with a level of detail that was not previously possible although it must be restated that such simulations represent a range of possible evolutions against which the coarser historical record (e.g. the best-track of Jarvinen et al. 1984) of the TC intensity and track must be compared. Given the 56-member ensemble represented by the reanalysis, it also provides for the ability to begin to quantify the sensitivity of track, intensity, and structure, of historical TC events. Such simulations help us address, for the first time, the very relevant question, "how would our numerical models perform today if a similar [historic] event were to occur today?"
The case chosen for presentation here is the 1938 New England Hurricane. This case represents a rare event in terms of the intensity and track of the TC, the tilt and magnitude of the interacting trough, and the (un)fortunate confluent timing of both (Tannehill 1938; Pierce 1939). The simulations performed illustrate the sensitivity of the TC's evolution to initial conditions but also seek to examine the structural characteristics of the TC and trough during extratropical transition.
Version 3.2 of the NCAR ARW WRF was run using 36km, 12km and 4km resolution nests for a simulation length of 5 days. The 20CR Version 2 (20CR-V2) Ensemble Mean fields were used for WRF initializations from the 14th to the 22nd of September 1938, every 6h (totaling 36 simulations). Based upon the quality of the various simulations (comparing track and intensity evolutions to the best-track; Jarvinen et al. 1984), a single initialization (1200 UTC 18th) was then chosen as a focus. WRF model physics include the BMJ Convective and YSU PBL parameterizations, consistent with those suggested for TCs by prior studies. As sea surface temperature (SST) from 1938 is not available over the model domain, a range of SST for recent Septembers was tested, with 2008 shown here.
The resulting ensemble of simulations is surprisingly consistent with the best-track. Specifically, the consecutive five day simulations are, with a few exceptions, reasonably well confined to the best-track track - impressive given the length of the simulation (5 days) and the resolution of the source data (2 latitude/longitude 20th Century Reanalysis). The intensity evolutionsshow a hurricane that is too weak early in the evolution (expected given the coarse source data), but quickly ramps up the intensity toward a landfall intensity in New England that is quite reasonable (strong Category 2), but with notable variability present in the ensemble. As mentioned earlier, following these numerous simulations, one specific initialization time was chosen as a further control: 1200 UTC 18th September 1938. The 56 ensemble members initialized with this one time show a range of track, landfall timing, and landfall intensity, with the ensemble mean track and intensity evolution impressively in agreement in the 24hr leading up to and following New England landfall. The 12km simulation further improved on the track and intensity evolution given by the coarser 36km simulation. The estimated maximum surface winds (performed by approximating the same method as NHC in real-time: reducing 700mb winds by 75-90%, depending on distance from storm; Powell and Black (1990), Franklin et al. (2003)] shows a broad hurricane force wind field throughout all of New England, with sustained category two (and perhaps weak category 3) winds into central and eastern coastal New England.
One important caveat is that the timing of landfall in even the best WRF track was approximately 9 hours too late, a timing error that is very consistent with 3-4 day forecasts currently performed in real-time. Further understanding this timing error and quantifying it for other cases will be helpful for interpretation of future forecasts of impending extreme hurricane events. It is suspected that should the simulations shown thus far have happened in real-time, (with only a 9hr timing error on a New England landfall at 3-4 days length), such warning would have been considered a great success - although undoubtedly the time for preparation would have been cut short as the storm began to accelerate northward faster than forecast. Nonetheless, it must be remembered that these simulations are not forecasts, since the future grids used for boundary conditions to the model are "perfect" rather than (e.g. GFS or ECMWF) global forecast grids, as would be used in a real-time setting. As such, the output is labeled a simulation to distinguish it from a real-time forecast setting.