Guide to Tropical Cyclone Forecasting: Tropical Cyclone Motion

While the development of consensus models has helped to improve track forecasting in recent years, further exploration of model consensus errors can yield additional tools. In an effort to predict the error of the consensus prior to the forecast advisory deadline, Goerss (2007) tested several possible predictors, including consensus model spread, initial and forecast intensity, initial TC position, TC speed and motion, and the number of members available in the consensus model. The most significant predictor of that group was consensus model spread, followed by TC intensity (either the initial or the forecast intensity). Goerss (2007) then used a stepwise linear regression model and the most skillful predictors to forecast the consensus model error at all forecast times. It was discovered that the model was able to predict 15-20% of the forecast variance at shorter lead times and up to 45-50% of the variance at the longest lead times (Goerss 2007). In order to make these results useable to forecasters, Goerss (2007) then used the predicted errors to draw radii around each of the consensus model forecast positions so that they would contain the verifying observed forecast positions 72-74% of the time. Since the predicted radii are positively correlated with model consensus spread, the forecaster can use the size of the radii to determine how much confidence to assign to the consensus model forecast. Making a forecast outside the GPCE circle would require a strong signal that the model consensus forecast is clearly in error.

Figure 32. Predicted 72% confidence radius (solid circle) surrounding the 120-h CONU forecast for (a) Hurricane Isabel, 0000 UTC 13 Sep 2003 and (b) Hurricane Kate, 0000 UTC 30 Sep 2003. The

individual model tracks used to create the CONU track are shown along with the 120-h radius (dotted circle) used by the NHC potential day 1–5 track area graphic. Reproduced from Goerss (2007).

Initialization

An inspection of the model analysis is necessary to determine how well the model captures the initial state of the atmosphere and the intensity, size, and structure of the TC, in particular. In some cases, model initialization are obvious (Fig. 33). Here the GFS analysis of sea-level pressure and 10-m winds during the genesis stage of Hurricane Gordon at 1200 UTC 11 September 2006 is much too weak, showing only an inverted trough and little cyclonic vorticity in the wind field when the cyclone is already a 35-kt tropical storm.

Figure 33. An example of a poor GFS model initialization fields of 10-m wind (in kt) and surface pressure (mb) for 1200 UTC 11 September 2006. Newly upgraded Tropical Storm Gordon is evident a few hundred miles northeast of the northern Leeward Islands.

In addition, a region of anticyclonic flow and low-level divergence is noted immediately west of the developing circulation, which appears unrealistic. Due to the poor analysis of the TC itself, it is unlikely that the GFS forecast track for this system will be realistic. Figure 34 shows another example from 0000 UTC 14 September 2004 where three TCs are analyzed by the GFS: Hurricane Javier and Tropical Storm Isis in the eastern North Pacific and Hurricane Ivan in the Yucatan Channel. While Javier is nearly as intense as Ivan, the model only analyzes Javier as a minimal tropical storm at best.

All available data sources, including satellite imagery, surface observations, and cloud-tracked winds, should be examined when assessing the quality of a model analysis. Using software packages that enable the analyst to overlay these data with the model analysis in real-time allow for a more robust comparison. Models can then be ranked in terms of how far out of tolerance they are with respect to observations. In the example of Javier, the GFS analysis is so poor that the model forecast swas poor beginning almost immediately. Models with poor analyses should be viewed with suspicion, however, some useful information might still be obtained from them.

A significant challenge to TC modeling has been developing a method by which the relatively small-scale TC vortex can be properly analyzed by relatively coarse models in regions with little data. One method involves constructing a TC vortex by using “bogus” or “synthetic” observations, which is done in the NOGAPS and UKMET models, respectively. In the GFDL, the vortex is “spun up” in a separate model and then introduced at model initialization. Other schemes such as “vortex relocation” are also utilized. The “vortex relocation” scheme in the GFS model does not modify the analyzed TC vortex but repositions it to the official NHC position estimate and has resulted in more accurate forecasts by that model. The GFS model 72-hour average track error decreased considerably, to 193 nm, after the inclusion of this procedure from 2000-2003 (Elsberry 2005), an astonishing increase in forecast skill of 42% with respect to CLIPER. However, the ECMWF global model, which does not use any TC vortex relocation or bogus scheme, has emerged as one of the more skillful track models in the Atlantic in recent years. For additional information on the vortex relocation procedure, refer to either Chapter 11 of this document or Liu et al. (2000).

Figure 34. GFS model initialization fields of 10-m winds (in knots) and surface pressure (in isobars) at 0000 UTC 14 September 2004. Tropical Storm Isis and Hurricanes Javier and Ivan are featured from left to right in the image, with intensity indicated in white text next to each cyclone.

Forecasters should evaluate model initializations using a step-by-step approach. First, the forecaster could attempt a manual analysis of observations of several atmospheric fields. The analysis process can familiarize the forecaster with the current data and synoptic pattern. Next, one should survey areas of importance such as the upstream storm environment to look at the density of data. On occasion data may be modified or deleted from the analysis, and forecasters reviewing model initialization data should realize when either action is taken and why. It is worth noting, however, that no analysis scheme in use by operational centers is likely to exactly match observations.

The forecaster should then compare the model analysis and observations, identifying the critical differences and assessing significance in the context of the present forecast problem. To do this the forecaster can qualitatively assess a model initialization. For example, Model Diagnostics forecasters at the Hydrometeorological Prediction Center (HPC) in Camp Springs, Maryland, attempt to compare upper tropospheric model fields with water vapor imagery upstream over the North Pacific to determine whether model initializations have accurately depicted key weather systems in the short- and medium-range forecasts over North America. Over the often data-void North Pacific the quality of upper tropospheric model fields is sometimes degraded because the initial fields do not match observations well. This can result in errors that propagate through the forecast, which subsequently can affect the evolution of significant weather systems several days into the future. Although identification of model initialization errors does not necessarily reveal the impact on the model forecast, an erroneous initialization often has an effect on forecast skill. However, with the tremendous amount of satellite observations and other observations that are part of ever more sophisticated data assimilation schemes, locating errors of substantial importance has become a harder task.

Models are characterized as either early or late, depending on their availability to the forecaster during the regular 6-hourly forecast cycle. Consider the forecast cycle at the NHC that begins at 1200 UTC and ends with the issuance of the forecast at 1500 UTC. After assimilating 1200 UTC observational data, the GFS analysis and model integration take several hours. As a result, GFS model forecasts through 120 hours are not accessible until 1600 UTC, after NHC’s forecast deadline. The unavailability of GFS model fields makes it a late model. On the other hand, the BAMS models are relatively simple and are available within minutes of their initialization; the rapid availability of the BAMS models makes them an example of an early model. Model timeliness is given in Table 3.

To address the issue of late models, the latest available run of a model can be adjusted to the current synoptic time and initial conditions. For example, the forecast from the 0600 UTC run of the GFS can be smoothed and then adjusted so that the 6-hour forecast position matches the observed 1200 UTC position. This adjustment process creates an early version of the GFS model that can be viewed in the context of all other model data for the 1200 UTC cycle. For historical reasons, the adjusted models have become known as “interpolated” models. Although the interpolation process generally offsets the problem of a poor initial analysis and six-hour forecast, errors in both of these can still lead to large errors later in the forecast.

Short-Term Forecasts

Short-term forecasts (i.e., those out to 48 hours) are critical to the issuance of watches and warnings and can serve as the basis for making evacuation decisions. Forecasts at the beginning of this range can take advantage of persistence, and changes to the forecast track are generally small for the first 12 to 24 hours, since dynamical model errors and model spread have become quite small at these ranges (Franklin 2009). Even though differences in model runs out to 24 hours are generally negligible, there are situations when model guidance can be discounted because of erroneous initial motion. Changes to the forecast track later in the short-range period are made incrementally and are based upon new but consistent trends in the guidance. Fine-tuning the forecast for TCs approaching land can be done using more frequent and precise fixes from radar and aircraft. Observations of short-term oscillatory movement or trochoidal motion are not uncommon but should not be confused with a longer-term trend. Wobbles in the track should be smoothed out when determining an initial motion estimate (Lawrence and Mayfield 1977). In cases where a TC approaches the coast at an oblique angle, differences between short-term, oscillatory track behavior and longer-term trends can lead to large variations in the locations that experience tropical storm and hurricanes conditions and the location of landfall. As a TC moves to within a few hours of landfall, short-term trends ultimately determine when and where hazardous conditions will occur.

Extended-range forecasts (e.g., beyond 48 hours) rely most heavily on dynamical models and consensus model approaches. Sensitivity to initial position analysis is still an issue, since errors in the model analysis can grow at extended time ranges. While global dynamical models are not designed to predict TC track explicitly they have demonstrated increasing skill due to improved forecasts of the large-scale steering flows that control TC motion. Regional dynamical models, such as GFDL, GFDN, and HWRF, tend to be better performers in short-range track forecasts due to their high resolution and efforts to properly initialize the TC vortex. However, forecasts begin to lose skill at extended ranges (Franklin 2009) due in part to the limited domain of these regional models, which means that critical, upstream weather features are sometimes initially located outside the model domain and either under-represented or not represented at all in the forecast cycle (M. Bender, personal communication). Since global models do not suffer from these boundary limitations they are better able to represent upstream steering features important later in the forecast period.

Of the global dynamical models, those of the highest resolution like the ECMWF, are generally consistent, better performers at longer ranges (reference??). One measure of global model forecast skill is the correlation between forecast and actual 500 hPa geopotential height anomalies over the Northern Hemisphere. It can be seen that the skill of the ECMWF degrades more slowly in comparison to other global models (Fig. 35). This measure shows how well the model predicts the large-scale circulation features that can affect TC motion, especially at longer time ranges. While there is some tendency for some models to regularly outperform others, no global model consistently performs best with TC track forecasting year to year. This may explain the limited success of the corrected consensus techniques for tracking forecasting.

When assessing model forecasts, it is important to understand why a model is showing particular forecast scenarios and to decide which scenario seems most realistic. This requires detailed knowledge of the strengths and weaknesses of the models and how these might vary in different circumstances. For example, even though the UKMET is a generally a good performer for TC track, it sometimes produce more westerly tracks of TCs (left of the remaining guidance). Over time, qualitative comparisons of UKMET fields with other models have shown a tendency for stronger subtropical ridging and weaker mid-latitude troughs in this model relative to others. Awareness of a model bias such as this one can influence decision-making practices when composing a TC track forecast.

Figure 35. Plot of correlation of forecast and anomalous 500 hPa northern hemispheric geopotential height.

Some of the more significant forecast challenges in the extended-range period occur in situations when TCs encounter ill-defined steering flows and move erratically. TCs can also interact with complex mid-latitude, sub-synoptic-scale flow, resulting in sharp turns or loops in the track. Numerical models struggle to predict abrupt halts or loops in track (Fig. 36). For example, no model correctly forecast the counter-clockwise loop made by Tropical Storm Hanna in the vicinity of the southeastern Bahamas (Fig. 36b). However, several models did suggest the development of a rather strong, deep-layer ridge over the eastern United States that could block Hanna’s northward progress and result in a southward or southwestward motion with some slowing of forward speed.

Comparing several consecutive runs of a particular dynamical model over a period of time can reveal trends in the forecasts from an individual model, while examining trends in the entire suite of guidance over time shows trends in the predictability of a forecast scenario. For example, comparing a particular model’s over a day might provide clues about the range of probable forecast scenarios and possibly identify outlier solutions. An example is shown from Hurricane Wilma in 2005 where three GFDL solutions generally show Wilma recurving off the U.S. East coast in a few days, although they suggest significant uncertainty with the forward speed of Wilma during recurvature (Fig. 37). However, the 1200 UTC GFDL solution appears to be an outlier compared to the other runs that day, as it forecasts the hurricane to remain in the Caribbean the next five days. The lack of consistency in these model solutions suggest a low predictability scenario with the interaction of Wilma and the trough that may cause recurvature.

In the Hanna example from earlier, an examination of several days’ guidance reveals that each model handled the development of the deep-layer ridge near the U.S. east coast differently, resulting in variations in Hanna’s predicted track. Despite these differences, there were two common threads in the guidance: was 1) the slowing of Hanna’s forward speed and 2) a possible motion toward the west-southwest or southwest. By comparatively evaluating a full suite of models over a period of time, additional information could be discerned about Hanna’s motion over the next few days that might not be apparent when looking at only one cycle of track guidance.

TCs that interact with the mid-latitude flow can also exhibit unusual track behavior. Two examples of this are Hurricane Kate (2003) around the time the TC stalled a few hundred miles southwest of the Azores (Fig. 36a) and Tropical Storm Jeanne (2004) located just north of Hispañiola (Fig. 38). In both cases the track guidance is quite varied, suggesting several possible motions in all directions. This is a situation where the model consensus may not perform well due to the disparate spread of model guidance. A selective consensus may be useful in this situation if some of the model solutions can be dismissed as unreasonable. However, this is often very difficult to determine in real time.

Figure 37. (a) Numerical model guidance at 1200 UTC 29 September 2003 Hurricane Kate over the central north Atlantic. The verifying track is given in solid white, while the individual models are color-coded. (b) Numerical model guidance at 1800 UTC 29 August 2008 for Tropical Storm Hanna. Solid while line indicates the best track, while color-coded lines depict individual models.

Figure 36. GFDL model output for Hurricane Wilma on 19 October 2005, starting with (a) 0000 UTC, (b) 0600 UTC, (c) 1200 UTC, and (d) 1800 UTC. The 5-day forecast track is given as a solid black line.

Figure 38. Track model guidance for Hurricane Jeanne at 1200 UTC September 19 2004. A lightly-shaded white line depicts the track of Jeanne until advisory time. A solid white line depicts the best track of the storm after the forecast time. Colored lines correspond to different track models.