TRANTECH Hurricane Model Description
by Gary Gray
A. General Overview:
A new scheme has been developed in the forecasting of tropical cyclones.
It is known as TRANTECH (TRopical ANalog TECHnique). For certain reasons
TRANTECH is a method that is quickly dismissed by some. This is because the
scheme has been developed solely on the basis of climatology. However, as
will be discussed in the following paper, it uses, indirectly, current
atmospheric dynamical fields. With no direct dynamic fields input into the
model it cannot be called "dynamical". As such, we will coin the term
"pseudo-dynamical" for this model. With the existence of the CLIPER
(CLImatology and PERsistence) model (Neumann) one automatically questions the
authenticity (and usefulness) of a "new" climatology based forecasting
technique. This will be the first topic to be covered in this paper.
The statistical CLIPER (Neumann) model developed in the 1970's differs
greatly from the TRANTECH method being introduced herein. In developing the
CLIPER model a storm's movement was determined to be a function of certain
parameters. CLIPER contains coefficients generated from historical data.
Given these coefficients and current parameters a forecast is then generated.
This scheme has only minor similarities to the TRANTECH scheme. The same can
be said in regard to the statistical SHIFOR model versus TRANTECH, in regard
to tropical cyclone intensity forecasting.
The new method of statistical forecasting uses another well known scheme,
the analog method. Using analogs has many advantages over the CLIPER model.
One advantage is the direct usage of actual past storm tracks rather than
usage via an indirect coefficient. Another advantage is the constant
upgrading of a climatological database. As each year passes storm tracks are
added to the database thus increasing the sample size for certain storm
scenarios. Conceivably, one could update the CLIPER coefficients each year,
but this is certainly more cumbersome and tedious than simply updating a storm
track database. Updating the CLIPER coefficients would also mean updating the
actual software for the CLIPER code, because the way it is currently coded has
the coefficients hard-coded within it. Of course, one could modify this to
soft-code the coefficients and automate the coefficient update. However, this
is work that must be done merely to bring CLIPER up to speed with something
such as TRANTECH; it does not make CLIPER a superior option. Also, one must
keep in mind that some past tracks may be adjusted slightly based on a
post-analysis of the data. Integrating this change into the climatological
database is much simpler than regenerating CLIPER coefficients then changing
code. In fact, it wouldn't even be worth it to update the coefficients based
on such subtle changes unless, again, this process were automated.
Two other advantages with the new TRANTECH forecasting scheme revolve
around the time intervals of the forecast. Official past-track information is
stored in six hour intervals. Thus a forecasting technique directly using
past-track information has a natural six hour interval between forecast
periods. Also, while there is no attempt to claim superb skill in extended
forecast periods, there is no real limit to the time extent which can be
predicted (other than degeneration of the storm). Therefore, while official
NHC forecasts and CLIPER/SHIFOR generated forecasts issue only 12, 24, 36, 48,
and 72 hour forecasts (SHIFOR performs a 60 hour forecast as well), TRANTECH
produces 6, 12, 18, 24, 30, 36, 42, 48, 54, 60, 66, 72, 78, 84, 90, 96, 102,
108, 114, and 120 hour forecasts (six hour intervals out to five days).
Another advantage of the analog method over CLIPER is that CLIPER makes no
attempt at forecasting tropical cyclone intensity and SHIFOR makes no attempt
to forecast storm track. The analog system forecasts intensity in the same
intervals as it forecasts position. Testing from TRANTECH indicates a
significant correlation between track and intensity. Therefore, for the best
possible forecast it is imperative that any statistical model utilize both
track and intensity. Finally, TRANTECH has no influence from persistence.
This is another critical difference as persistence has no basis in either
an analog or climatology method. It is merely an empirical attempt to keep
the model forecast from running amok. Persistence is, of course, used in
CLIPER.
So, how is TRANTECH pseudo-dynamical? Parameters such as latitude,
longitude, direction, speed, and storm strength are used as input into the
analog program. It then analogs this storm to each storm in the
climatological database. Certain storms which are too dissimilar to truly
represent an analog are discarded. Storms remaining as analogs are then
weighted based on how well they analog to the current system. Many of these
input parameters are a reflection of the current dynamical environment that
the storm is in. For example, a tropical cyclone moving between 320 and 40
degrees is likely encountering a trough. Another example, if the storm is at
a low latitude and moving due west it will likely continue on that course, at
least in the immediate future. Therefore, these input parameters all contain
some representation of the current dynamical influences on the storm. In
turn, when the analogs are done the most heavily weighted storms will be
storms most closely matching the current storm's parameters. It follows then
that these analog storm tracks will be a dynamical representation of the
current storm. It is in this way that a purely climatologically based model
can be considered pseudo-dynamical. Moreover, TRANTECH's coupling of track
and intensity provides even more detailed information on the nature of the
surrounding environment. For example, a rapidly weakening, southern latitude,
westward moving tropical cyclone is likely encountering a low level high
pressure beneath upper level troughing. The CLIPER and SHIFOR lack of this
coupling does not enable those models to "see" the environmental influences.
In this manner, TRANTECH is pseudo-dynamical.
B. Specific Forecasting Technique:
In past versions of TRANTECH there were actually seven different methods
which were used in the TRANTECH forecasting technique. They were as follows:
1) Linear Climatology Model [LCM], 2) Integrated Climatology Model [ICM],
3) Integrated Climatology & Persistence Model [ICPM], 4) Subjectively Adjusted
ICPM [SAI], 5) Discreetly Weighted ICPM [DWI], 6) Compound SAI & DWI [CSD],
7) Ensemble Forecast. These methods have been abandoned in favor of a single-
solution model most closely resembling the DWI technique. It also incorporates
some of the desired aspects of the other various techniques as well. It also
purposefully abandons any use of persistence, which is a slight change to the
DWI. Persistence was found to hurt, rather than help the forecast and, as
discussed earlier, is inappropriate for a pseudo-dynamical model. Thus, for
informational purposes only, each forecasting method is discussed below.
**Note that since this writing major changes and upgrades have been incorporated
into TRANTECH. It has been found that the re-introduction of noise into the
system yields significant errors. As such, although the Integrated techniques,
in theory, are clearly superior, they provide an inferior result to the
Linear technique. On the other hand, a detriment of the Linear method is that
storms that veer abnormally off course are not dropped from the analog set. As
such one aspect of the Integrated method has been held over; the weights for
each analog storm are reculculated at each time step. Another modification is
that the initial offset correction (discussed below) is applied on each analog,
so that each analog storm provides a unique, actual forecast. Also, the actual
method for applying the analog storm track has been modified somewhat. Thus,
the individual scheme discussions below are outdated. Still, they remain in
this document to serve as an example of the TRANTECH methodologies and to show
the complexity of what would otherwise seem to be a simplistic forecast tool.
The LCM is the simplest of the forecasting techniques, but it is also the
most continuous. As will be discussed below, with the other schemes it is
possible for storms to analog during part of the forecast then be removed from
the set of analogs later in the forecast period. The reverse is true as well.
The LCM is more stable. It's method of forecasting is as follows: First, the
climatological storms are compared with the current storm to determine whether
or not each one qualifies as an analog. If it does it is weighted based on
how well it analogs. Certain input parameters are weighted differently in the
total storm weight based on empirical studies. The weighted average of the
initial parameters of the analog storms is calculated to determine the initial
offset. The future positions (in six hour increments) are then averaged from
the analog storms. These positions are adjusted by the initial offsets to
come up with a final forecast.
The ICM is the second technique in the TRANTECH scheme. It is considered
"integrated" because it re-analogs (integrates) the storm after each six hour
forecast period. For example, a six hour forecast is generated for storm X in
an identical method as described in the LCM. The parameters for the six hour
forecast period are then treated as initial parameters. These "initial"
parameters are re-introduced into the system, analogued, and a six hour
forecast is again produced. This six hour forecast, being generated from the
original six hour forecast parameters is, therefore the 12 hour forecast.
This method of generating a six hour forecast and then integrating those
forecast parameters back into the system to obtain a forecast six hours
further out is repeated until the forecast reaches 120 hours (five days).
The third method, the ICPM, is nearly identical to the ICM. Both use the
integrated, six hour time-stepping, technique. The significant difference in
the ICPM as compared to the ICM is the inclusion of persistence in the ICPM.
Persistence means a continuation of the current direction, speed, and
intensification rate. Persistence is given a weight along with the normal
analog storm weights to produce an alternative weighted average and thus, an
alternative forecast. It should be noted that in this forecasting technique
the persistence applied to the forecast is a "running" persistence. That is
to say that, for example, the 84 hour forecast employs the 72 to 78 hour
direction, speed, and intensification rate as its persistence; it does not use
the 00 hour (initial) direction, speed, and intensification rate for
persistence. Due to the nature of the running persistence, the entire purpose
of including persistence in the forecast becomes less and less important the
further along the forecast gets. For this reason the weight given to
persistence slowly diminishes throughout the forecast period. **Note that
under the new processing system persistence has been abandoned.
The next method in the TRANTECH scheme is the SAI method. It is very
similar to the above ICPM method. It employs an integrated technique along
with an inclusion of persistence in the forecast. The primary addition in the
SAI method is to apply subjective adjustments on the numerical forecast
output. It should be made clear that while these adjustment are subjective in
that they are applied as best guess adjustments to the forecast, they are not
done subjectively in the sense that the adjustments are applied on a case-by-
case basis. This would remove the entire purpose of utilizing a model.
Instead, these subjective adjustments are very general and can therefore be
included as part of the model code. These subjective adjustments are somewhat
simplistic and are intended to be protection against "veering" storms. Analog
storms which obtain a climatologically uncharacteristic motion or intensity
fluctuation may adversely affect the current storm forecast. The subjective
adjustments are intended to keep these adverse effects under control. One may
argue that, as part of climatology, these abnormal storms should be included
like any other storms. This is a legitimate argument and, at times, the SAI
method does appear to worsen the forecast with its subjective adjustments.
This is why TRANTECH employs several different methods of forecasting.
**Note that in the new processing system subjective adjustments have been
abandoned.
TRANTECH - DWI is the next forecast method on the list. This "discreetly
weighted" model employs a technique somewhat varied from the previous ones.
As described in the LCM the way weighting has been handled in all previous
models is to obtain a weight for several parameters of a given storm. A
weighting function is then used to combine all the parametric weights into one
generic weight for each analog storm. The DWI method carries through with the
first part of this, obtaining a weight for each individual parameter. From
this, however, three different weighting functions are used. One weighting
function generates a latitude weight for each analog storm; the next function
generates a longitude weight; the third generates an intensity weight. This
is where the term "discreetly weighted" came about. In the forecast, each
forecast parameter then has its own weighted average. The advantage of this
method is more exact parametric analoguing. The main disadvantage is that, by
splitting up storm parameters the weighted averaging may be considered
somewhat disjointed.
The last actual model (the second to last forecast scheme) is the CSD
model. It employs a combined SAI and DWI method. As such, it is, of course,
an integrated method with an inclusion of persistence. As described in the
DWI paragraph above, the CSD method discreetly weighs each parameter and uses
three different weighting functions to obtain its forecast. After each six
hour forecast the subjective adjustments, as described in the SAI paragraph,
are applied to the numerical forecast. This results in the most complex
(although not necessarily the most accurate) of the forecasting schemes. This
is also the final TRANTECH forecasting scheme model.
The Ensemble Forecast is not a forecasting model, but it does generate
its own unique forecast. A confidence factor is determined for each forecast
period for each of the above six forecasting techniques. The confidence
factor is determined using a few different factors. Through extensive tests
certain forecasting techniques have been determined to be superior or inferior
to the others. This plays a role in determining the confidence factor. The
confidence factor is also determined by the number of analog storms beings
used (because of the integrated forecasting method used in most of the models,
each time period for each model need not have a consistent number of analog
storms). Obviously, a higher number of analog storms leads to a higher
confidence factor. A third factor in determining confidence is how well the
analog storms actually analog with the current storm. Finally, in some cases,
confidence drops as a forecast approaches the 120 hour forecast period. These
confidence factors act as weights and a weighted average is then performed on
the six forecast models. The result is a confidence weighted ensemble model.
**Note that a confidence level is still applied to the forecasts under the
new scheme, but it is not calculated in the same way as with the Ensemble
run of the old scheme.
C. Strike Probabilities:
After generating the TRANTECH forecast for a particular situation
TRANTECH then generates a chart containing several other key pieces of
information. There are several different parameters which are deemed to be
critical for hurricane preparedness. These include the chances of landfall at
any particular location, how far the storm will pass, and when the storm will
pass. TRANTECH uses the previously generated forecast to statistically
generate results for these parameters. It should be noted that the table it
produces is completely automated as part of the TRANTECH system. There is no
human intervention. So, in the unfortunate cases when TRANTECH does not
generate a particularly accurate forecast, the charted parameters will not be
very accurate either. However, when the forecast is quite accurate you are
guaranteed an objective and at least fairly accurate set of strike statistics.
The parameters are output in a chart format with each parameter making up
one column and each row designated to a particular location. The parameters
(columns) contained in the chart are as follows:
1. Location - There are many cities (up to 800, depending on the
list in use) available for output and the list of cities can
be easily updated. Only cities determined to have at least a
minimal chance of being hit by the storm are output to the
table.
2. Strike Probability - This is the chance (in percent) of the storm
passing within 60nm of the location through 120 hours.
3. Distance - This is the distance (in nm) of the closest pass of
the storm to the given location.
4. Time - This is the time (in hours from the initial forecast time)
that the storm makes its closest pass to the given location.
5. Intensity - This is the storm intensity (in kts) at the time when
the storm is making its closest pass to the given location.
6. Location Intensity - This is the actual estimated forecast wind
speed at the given location. This is an extremely rough
estimate attempting to account for distance, friction, etc.
7. Storm Surge - This is the predicted maximum storm surge, not
including natural tidal fluctuations. This is a general
estimate and, as of yet, does not incorporate any ocean floor
topography or coastline shapes, which can drastically alter
the storm surge in some localized areas.
8. Watches - This column contains output from an automated
watch/warning system. TRANTECH determines what, if any,
watches or warnings need be posted for the given location.
D. Output
The format of the output should be reasonably self-explanatory. The
track output file is very simple. It contains the TRANTECH forecast along
with the forecast from CLIPER and SHIFOR. The title of each forecasting
scheme is given followed by the actual forecast; at present, of course, there
is only one TRANTECH scheme in use. Under TRANTECH there are 21 entries; the
first is the initial conditions followed by the 20 forecast periods (6 hours
apart out to 120 hours). Each row contains the time period, then the forecast
latitude for that time, then the forecast longitude for that time, and lastly
the forecast maximum sustained winds (in kts) at that time. The CLIPER
forecast follows with six rows of data (initial, 12, 24, 36, 48 and 72 hours
forecasts). CLIPER outputs only track (latitude and longitude). SHIFOR follows
with the same time intervals as CLIPER, with 60 hours added, and outputs only
the intensity forecast.
The strike probability file might be slightly more difficult to
understand, simply because the header for each column may not be terribly
descriptive. The column headers and their descriptions are as follows (note,
each of these parameters was explained in the previous section):
LOCATION = Location being analyzed
STR = Strike probability
DIST = Distance from location
TIM = Time of closest pass to location
INTNS = Intensity at pass
LOCST = Storm strength realized at the location
SS - Storm Surge
Remarks = Watches, warnings, or advisories
Also note that all output is in ASCII and graphical formats. The
graphical interface is a new upgrade.
E. Statistics
Statistics on the accuracy of the TRANTECH model are, as of this writing,
under development. Details of the statistics can be found via a link from the
main TRANTECH page. Thus, they will not be given in detail herein. However,
the following findings should be noted:
1) The current version of TRANTECH is showing skill in the track forecast at
all time periods.
2) The track forecast skill is greater in the short time periods than in the
longer time periods (e.g. 72 hours).
3) The intensity forecasts generally exhibit considerably more skill than the
track forecasts.
4) The use of a topography dataset may erroneously indicate the intensity skill
as being to high; however, an examination of non-landfalling tropical cyclones
yielded little difference in statistics. Thus, TRANTECH is no more superior to
SHIFOR over land than over water, meaning that the solid intensity skill is not
a mere figment of the topography database.
Some important points need to be made about these statistics. First and
foremost these statistics are likely biased in favor of CLIPER and the NHC
official forecasts. The reason for that is simply that the TRANTECH software
is not yet finalized and optimized. I would expect to see improvements before
final statistics are generated. Also, note that any statistics shown on the
statistics page for CLIPER may not reflect precisely what is available from
other sources. These other sources make the serious error of initializing
their parallel forecasts for CLIPER and the model being tested with the actual
initialization data that was used in real-time. This may seem reasonable since
that was the data available at the time. However, it does not properly reflect
the model skill. A statistical model such as CLIPER is far more sensitive to
the initial conditions (LAT, LON, SPD, DIR, INTENSITY) than is a dynamic model.
The dynamic model is keyed in more on the surrounding environment. As such,
any initialization errors will propogate and expand in the statistical model;
they will damp out in the dynamic model. As such, initialization must be
performed from the "best track" final analysis data. This is typically not
done and, as such, model skill is overstated by most. However, in the case of
the TRANTECH versus CLIPER analysis, both models were initialized from the
best track data. In doing so, this also provided for a vastly greater data
set from which to generate forecasts and statistics. This, too, is another
problem with most analyses. Many organizations will perform an analysis on a
season or a few seasons. Because patterns can be persistent throughout a
season, and because some models perform better under certain atmospheric
conditions, a sample size of only a few seasons is unacceptable. The TRANTECH
statistics go back to 1940, and can be run on earlier data (1940 was selected
as the limit as that is around the time when data quality began to improve).
With the above stated, it should be made clear that the preliminary
statistics as they stand are neither exceedingly good, nor poor. The skill
at all time periods is a new development and did not exist in the original
document. So, clearly, TRANTECH is now serving as a useful forecast tool,
showing skill at all times, in both intensity and track. On the other hand,
the skill from the purely dynamical models continues to far exceed TRANTECH.
Though the skill from these dynamical models may be exaggerated in most
reports (see the above comments on this), they still far exceed the skill of
the TRANTECH track forecasts in the longer time periods. In the earlier time
periods TRANTECH is competitive; also, in intensity TRANTECH is far and away
the most superior model currently available. Thus, the current status of
TRANTECH is reasonably positive. However, to have a greater impact on the
user community, TRANTECH must exhibit greater skill in the track forecast
for longer time ranges. Clearly, TRANTECH will never match the skill of
most purely dynamical models. However, since run-to-run error variation can
be significant in even the best of models, if TRANTECH can exhibit "significant"
skill in the longer ranges it could easily rival the dynamical models on any
given forecast. But, further progress must be made to bring TRANTECH up to
this level, and some may doubt that it is even possible. However, given the
current state of the TRANTECH software system, and the number of planned
test upgrades, it seems highly unlikely that TRANTECH has hit its limit.
Thus, though the amount of skill to eventually be reached remains an open
issue, there is little question that TRANTECH's skill will grow further with
pending upgrades. Therefore, with the short-range track forecast and the
short and long-range intensity forecasts already showing significant skill,
TRANTECH is on the verge of becoming a legitimate team member in the suite
of tropical cyclone numerical prediction models.
