With each passing year, forecasters have ever-more-accurate numerical guidance on where tropical storms and hurricanes are most likely to track and how strong they’ll get. Several of the leading models have undergone noteworthy improvements over the past year. Track models have gotten steadily better over the last couple of decades, whereas improvements in forecasting intensity have been much more difficult to come by (see Figures 1 and 2 below), so a great deal of energy has been focused on the latter. Below is a summary of what’s new and cool, based on interviews and email exchanges with the following experts:
--Richard Pasch, Senior Hurricane Specialist, NOAA National Hurricane Center (NHC)
--David Richardson, Head of Evaluation, Forecast Department, European Center for Medium-Range Weather Forecasts (ECMWF)
--Julian Heming, Tropical Cyclone Specialist, UK Met Office (UKMET)
If you’re unfamiliar with the major models discussed below, see this overview by Jeff Masters
and a somewhat more technical summary
from NHC. Jeff’s post “Which Model Should You Trust?”
, from August 2014, offers another excellent set of guideposts. How are NHC’s hurricane forecasts doing?
Before diving into model improvements, let’s take a look at how NHC has fared over recent years in its own predictions, which rely heavily on the models below. Here are two graphics from NHC’s 2014 Forecast Verification Report
, released in March 2015.
Verification of official NHC hurricane track forecasts for the Atlantic, 1990 - 2014. Over the past 25 years, 1 - 3 day track forecast errors have been reduced by 60 - 75%. Track forecast error reductions of 30 - 40% have occurred over the past ten years for 4- and 5-day forecasts. Image credit: 2014 National Hurricane Center Forecast Verification Report.
Verification of official NHC hurricane intensity forecasts for the Atlantic, 1990 - 2015. After more than two decades with little improvement, intensity forecasts have become notably more accurate in the 2010s, due to model improvements as well as a relative lack of strong and/or rapidly developing hurricanes. Image credit: 2014 National Hurricane Center Forecast Verification Report.
The ECMWF’s Integrated Forecast System is widely considered to be the world's best at tropical cyclone (TC) track forecasting, although it is sometimes beaten out by the GFS model (below). The ECMWF got a big boost in attention after it performed admirably during 2012’s Hurricane/Superstorm Sandy
. The ECMWF was the first major model to call for Sandy to arc westward into New Jersey, a full week before the storm made landfall
and several days ahead of the GFS. The high-resolution version of the ECMWF (16 km between grid points) is run every 12 hours; it is accompanied by a 51-member ensemble that’s run at lower resolution (32 km between grid points). Each ensemble run starts out with slightly different initial conditions, generated randomly to simulate the actual uncertainty in starting-point observations of the atmosphere. Ensembles help flesh out when the future of a tropical cyclone is fairly straightforward to predict, or when it's highly uncertain. Another strength of the ECMWF ensemble runs is that they include interaction between the atmosphere and ocean, which can help improve intensity prediction. “Recent work shows the importance of this coupling, at least in some situations,” said ECMWF's David Richardson.
The latest version of the ECMWF model, introduced in May, has significant changes to model physics and the ways in which observations are brought into and used within the model. The overall improvements include better portrayal of clouds and precipitation, including a more accurate depiction of intense rainfall. The main effect of the model upgrade for tropical cyclones is slightly lower central pressure. During the first 3 days of a forecast, the ECMWF has tended to have a slight weak bias on tropical cyclones; the new version is closer to the mark. From Day 5 onward, however, the new version adds to the preexisting tendency in those time frames to make hurricanes and typhoons too deep.UKMET
Forecasters at the UK Met Office are already seeing benefits from two major upgrades to its Global Model. One is the adoption of a new dynamic core, improvements in model physics, and an increase in horizontal resolution (now 17 km between grid points) in July 2014, which affected many characteristics of the model. This provided a big improvement in TC track forecasts: errors were reduced by about 8% when the new version was tested offline in 2012, and by about 18% when it was run alongside the old one during actual events from April to July 2014.
“The Met Office Global Model has historically been far too weak for most TCs,” Julian Heming told me. “However, the new model configuration is far more energetic, and TC intensity errors are significantly reduced, particularly at longer lead times.” To produce even more improvement in TC intensity forecasting, a new assimilation scheme was introduced in February. Each model run now incorporates the observed locations and central pressures of tropical cyclones at the model initialization time, as provided by the various warning centers around the world. These data are also interpolated for the six hours before the model run and extrapolated for the two subsequent hours, again based on warning-center advisories. “This is a totally new method of initializing TCs for us,” Heming said. The change not only further reduced the UKMET’s weak bias on TC strength, but it had a somewhat unexpected benefit: track errors in testing went down by 6%. Together, the updates of July 2014 and February 2015 reduced track error by as much as 30% in one test period, according to Heming.
UKMET also runs a 24-member ensemble system, dubbed MOGREPS-G
, that includes the upgrades above. The ensemble is run every 6 hours, out to 7 days ahead, with a grid spacing of around 33 km.
. The last two years have seen marked improvement in the UKMET Global Model predictions of tropical cyclone intensity, especially at longer time frames. Shown here are Northern Hemisphere results through July 20, 2015, with the average error for each year (left axis) shown for various lead times (bottom axis). Image credit: UK Met Office
, courtesy Julian Heming.
A well-publicized upgrade to the GFS model at the start of 2015 was made possible by a large increase in available computing power. The upgrade significantly boosted the model’s horizontal resolution, which increased from 26 km to 13 km (a fourfold jump, since it includes both east-west and north-south directions). However, there may not be major improvements evident right away in TC track or intensity, because the representation of atmospheric physics in the new GFS has not yet been tweaked to maximize the value of the higher resolution. “The new GFS has been doing fairly well this year, although it lags other global models in predicting east Pacific tropical cyclone formation,” said NHC’s Richard Pasch. The ensemble version of the GFS (GEFS), which includes 20 members, is run at the coarser resolution of 55 km.
The highly regarded GFDL model has gotten a few tweaks this year, but changes are relatively minor compared to the other models above. According to Pasch, "we're going to see some improvement, but nothing earth-shattering." GFDL and HWRF (below) are the two leading models used by NHC in recent years for intensity prediction, along with statistics-based models. HWRF
NOAA’s version of the Weather Research and Forecasting model specifically tailored for hurricanes (the Hurricane WRF, or HWRF) has undergone a major improvement in resolution, implemented in June. HWRF features a triple nest of concentric model domains that narrow in resolution as they zero in on hurricanes. The previous resolutions of 27, 9, and 3 km are now 18, 6, and 2 km. "There’s no other regional operational model in the world at that resolution that I know of,” Pasch told me. The new version also features a number of physics upgrades, including an advanced land surface model and a new radiation scheme that allows for better depiction of nocturnal peaks in shower and thunderstorm activity within TCs. According to Pasch, “the new HWRF has been a good performer thus far for track and intensity.”
Blending the models
The most powerful approach to hurricane prediction is model consensus: averaging the results from a large number of model runs, so that the most consistent signals come to the forefront and the outliers fade into the background. One type of consensus is the average of all the lower-resolution ensemble runs from a single model (such as the GFS or ECMWF). While this has some value, its usefulness is limited by the coarse resolution of ensemble runs, and by the particular strengths and weakness of any particular model. Forecasters usually favor multi-model ensembles, where the higher-resolution runs from several different models are averaged. This approach reduces the negative impact of any one model's idiosyncrasies, and random errors are more likely to cancel each other out. In making its forecasts, NHC calls on a variety of model blends, which usually outperform any individual model. Official NHC track predictions are often very close to the output from a model blend called TVCA, which employs the five models above (ECMWF, UKMET, GFS, GFDL, and HWRF). Going even further in this direction are "superensembles," such as the one developed at Florida State University. A superensemble not only blends multiple models, but it also weighs each model based on its past performance and includes bias corrections for each. As NHC puts it
, "The [Florida State superensemble] is constantly learning from the past performance of the models that it comprises."
. Successive forecasts of wind speed (in knots) from the GFS operational model for Cyclone Pam in the Southwest Pacific, beginning at 0000 GMT on March 10, 2015. Each red line represents a single GFS model run. The blue line indicates the best estimate of Pam's maximum wind speed based on satellite imagery. Image credit: Kerry Emanuel, Massachusetts Institute of Technology.Impressive results from Cyclone Pam
As it tore through the Southwest Pacific in March with sustained winds that topped out at 165 mph
, fearsome Tropical Cyclone Pam
lived up to the ominous projections from long-range models. Several days ahead of time, operational runs of both the GFS and ECMWF models indicated that Pam had the potential to become a severe cyclone. Figure 4 shows the intensity forecasts (in maximum wind speed) from a number of successive GFS runs starting on March 10, by which point the model was already consistently and correctly predicting that Pam would become a Category 5 cyclone (if anything, the GFS was overdoing Pam's strength).
At the ECMWF, Pam offered forecasters an encouraging preview of the next version of their model. This upgrade, scheduled for early 2016, will bring the top resolution of the ECMWF ensemble members to 17 km, which is comparable to the current resolution of the ECMWF operational runs. Figure 5 shows that the average from the present-day ensemble (part a) brought Pam down to a minimal central pressure of 950 mb, whereas the average from the higher-resolution version to be implemented in 2016 (part b) produced a central pressure of 915 mb. The present-day operational model (gray line in part a) did even better, giving several days’ notice that Pam’s central pressure could dip below 900 mb. It bottomed out on March 14 at 896 mb.
The UKMET operational model predicted a minimum central pressure for Pam of 916 mb, the lowest in that model's history. Even lower values have shown up in UKMET runs for other recent tropical cyclones. “Prior to the two model changes made in the last year, it would not have gotten anywhere near these central pressure values,” said Julian Heming.
This year’s tepid Atlantic hurricane season won’t provide many case studies for the model improvements discussed above, but the active Pacific is giving the new incarnations a solid workout. Should the Atlantic perk up in 2016, even more upgrades will be in place by that point, giving forecasters and the public an even better sense of what to expect.
. Forecasts of Pam’s central pressure at mean sea level from 10 March 1200 UTC, showing (a) the operational high-resolution forecast (HRES) and the operational ensemble mean forecast (ENS mean) with vertical lines indicating the extreme members and blue bars representing the 25th to 75th percentile of the ensemble distribution, and (b) a higher-resolution (17 km) ensemble mean forecast. Image credit: ECMWF, reproduced with permission from the ECMWF Newsletter, Summer 2015, courtesy David Richardson.