This installment of the periodic learning series focuses on hurricane track forecasting and is divided into four parts:
1) Factors influencing storm motion
2) Analysis methods of features influencing storm motion
3) Historical methods of predicting storm motion
4) Current state of track prediction
As always, if you have any questions about anything presented here, please don't hesitate to ask!

Factors influencing storm motion

Tropical cyclones are complex beings with tracks influenced by a wide variety of factors, ranging from synoptic-scale (thousands of kilometers/miles) to internal features on the mesoscale (tens of kilometers). The former are more important toward determining the long-term track of a cyclone, while the latter are more important toward refining landfall locations as well as for sheer curiosity. Let's take a look at the major factors in both categories.

On the large scale, tropical cyclones are predominantly steered by the deep layer steering flow in the atmosphere, such as that ahead of a trough of low pressure or along the periphery of a ridge of high pressure. To a first degree, the stronger a storm is, the higher up in the atmosphere that the storm "feels" and is subsequently steered by. This means that storm intensity prediction is inherently tied to storm track prediction! Previous research has shown that a deep layer steering flow is a good approximation to storm motion (see http://cimss.ssec.wisc.edu/tropic/other/dlm_faq.html for specific references) and forms the basis of the steering layer products from the Univ. of Wisconsin at http://cimss.ssec.wisc.edu/tropic/real-time/atlantic/winds/winds-dlm.html. Precise details within the larger scale steering flow become key when predicting specific paths of motion, as does a correct prediction of how the factors influencing a given storm's track will evolve over time. Minor errors in predicting the timing or intensity of a particular feature can lead to huge forecast errors, such as those in the timing of Hurricane Wilma impacting Florida and in the ultimate landfall location of Hurricane Katrina.

On even smaller scales, factors such as topography and internal oscillations can play a role in storm motion. The latter are most often seen in the form of trochoidal oscillations, typically found only with the most intense of storms, and are characterized by wobbles along a mean storm motion vector. The former have been shown to impact storms in the Taiwan region (see for an example) and potentially have a similar effect in Jamaica as well (e.g. Hurricanes Charley and Ivan).

Moving to slightly larger scales once again, Coriolis plays a role in influencing storm motion primarily in terms of something called the beta effect. The change in Coriolis -- or planetary vorticity/spin -- across the storm's circulation induces a north-south vorticity gradient across the storm, serving to want to send the storm poleward. Concordant with this is this Beta effect, where the change in Coriolis parameter to the north and to the south of the storm induces what are called Beta gyres, tending to send the storm to the NW (in the northern hemisphere) at a significant (about 2kt) rate. Thus, the natural tendency of a storm is to escape the tropics! Note that this influence is taken into account naturally within all model track forecasting systems. Also, the vertical tilt of the tropical cyclone and the horizontal and vertical changes of the steering flow can also play a role in influencing storm motion, although these are of lesser significance than the other factors noted here.

Analysis methods of factors influencing storm motion

The simplest manner of analyzing short-term factors influencing storm motion on the large scale is the water vapor satellite channel. This channel captures the steering flow associated with mid-level (~400-600mb) cyclonic and anticyclonic systems and is, to a first approximation, a good proxy for the deep layer steering flow across the storm. However, it is not always possible to depict the evolution of steering features out beyond a day or so just by looking at water vapor imagery (or even by looking at upper air data from observing stations across the hemisphere), as their movement and intensity changes are controlled by dynamical processes not easily captured by the human eye. The layer mean steering product from the Univ. of Wisconsin referenced above uses satellite data to construct layer mean steering flow and is a good tool to use in conjunction with the water vapor imagery, but it too suffers from the same limitations of not being able to directly see into the future.

In terms of internal factors influencing motion, the best means to go off of to analyze the magnitude and significance of such features is an educated guess. Typically, storms located further to the north and of larger size are going to be more influenced by Coriolis-induced steering anomalies than those located further south and of smaller size. Trochoidal oscillations can be expected for intense storms, but no matter the intensity of the storm it is tough to predict when/if they will occur and their significance to overall (long-term) storm motion is dubious at best. Finally, topographical influences almost certainly require either mesoscale modeling (at high resolution, about 10-15km) or simply going off of climatological impacts of a particular topographic feature upon past storms. This is not always ideal, of course, but is about all we have.

Historical methods of predicting storm motion

While storm motion forecasts are generally of high quality these days, that has not always been the case. Just last decade, the data available to forecasters to help forecast storm motion was crude and primitive at best. Our best global models had large forecast errors and did not feature the complexity, fine-scale dynamical and physical representations, or data assimilation capabilities that today's models do. Satellite data was sparse at best other than standard visible/infrared/water vapor satellite imagery; satellite-derived products were virtually non-existant.

As a result, the primary methods of forecasting storm track as late as the early- to mid-1990s came from climatology, persistence, mean steering flow models, and the crude representations of global models. These all form the basis for such models as the NHC90, LBAR, BAM-series, A98E, and other models that are still run but given little to no weight these days. As you may know, the baseline for determining how skillful a storm track forecast was is given by the CLIPER (CLImatology and PERsistence) model; what you may not know is that this was recently one of the best methods to predict where any given storm would track in real-time. While the theories about deep-layer storm motion have existed for some time, it is only recently that we have had the data over the oceans necessary to even get a good handle on what that steering layer flow actually is and be able to use it to forecast storm track.

Similar to the CLIPER model, analogs have been used in the past for storm motion forecasting, where similar storms (intensity and location and time of year) have been identified for any given storm and composited to get an idea of where the current storm may go. While of dubious value at best, it too was once one of the better techniques used for forecasting storm motion, whether in a forecaster's head or from a storm database. You can still see this in use to some degree today with the Weather Underground's "Historical Storm Tracks" page given with any active named system.

Current state of track prediction

Dynamical forecast models have improved by leaps and bounds over the past 15 or so years in many areas, largely driven by advances in understanding convective and microphysical parameterizations (since we aren't quite at the point to be able to resolve those interactions in the atmosphere directly) as well as massive increases in computational power and data availability. Errors have been halved in just the past 10 years from these models, leading to better track forecasts from the NHC and those in charge; such improvements formed at least part of the basis for extending the storm forecast period out to 5 days several years ago. Mesoscale models (such as the MM5 and WRF) have been introduced and refined, offering hope in storm motion (and especially intensity) prediction in terms of a better representation of the tropical cyclones themselves.

Beyond this, however, great forecast improvements have come from the use of ensemble predictions schemes and model consensus forecasts; these tools represent the state-of-the-art in storm track forecasting (yet are still often beat by forecaster experience at the NHC). Examples of standard ensembles are provided by the GFS (our global model) and ECMWF (the European's global model) model ensembles. These systems are based off of the idea that we may not always be able to accurately capture the initial state of the atmosphere correctly, so if we introduce perturbations to the initial state (e.g. in temperature, winds, and so on) and let those forecasts run in parallel with the main forecast, we can get a measure for storm track sensitivity and uncertainty.

Model consensus forecasts are yet another example of an ensemble, this time termed as the "poor man's ensemble." These models, such as the GUNA and GUNS, are based off of the idea that each model has its strengths and weaknesses and that averaging all of their forecasts together will tend to accentuate the positive aspects and eliminating the negative ones, leading to a better forecast. This isn't always true in reality, but works quite well in practice as these models often have the lowest track errors of any forecast system. Perhaps better known to the general public, however, is the idea of a "superensemble," such as that offered by the FSU Superensemble. The basic premise behind the Superensemble is the same as for the model consensus forecasts, except that instead of averaging all of the forecasts together, we can use past performances with many storms of the models to understand where their biases lie and to correct for those individually. In theory, this should lead to a better track forecast than with any other model, though this is not always the case as features in the atmosphere are not always repeatable. However, in 2003 and 2004, the Superensemble did provide the best track forecasts in the Atlantic basin. As an aside, note that the superensemble/model consensus forecasts are all statistical prediction schemes rather than actual forecasts of the atmosphere; this shows us that we still have some room for improvement in actually predicting the weather rather than using the biases of the models in our favor.

In the future, understanding the mesoscale features influencing storm track is likely where we will see some of the greatest storm track improvements. Beyond this, obtaining more data in the vicinity of storms -- such as those by the dropsonde missions ahead of storms threatening the US or by satellite-based observing systems -- that can go into the model to get a better handle on the synoptic-scale features influencing storm motion will be another source of large model improvements. To a large degree, the current suite of models do a good job capturing and forecasting the storms influencing storm motion, but minute details in timing and intensity still lead to forecast error; we can improve upon this with more data. We will never be able to have a perfect forecast, but as the past 15 years have shown -- with 3-day NHC forecast errors halved from ~300 n. miles to ~150 n. miles -- we haved improved and can continue to improve.