Guide to Tropical Cyclone Forecasting: Tropical Cyclone Motion
Download 197.74 Kb.
|
- Navigate this page:
- Climatology and Persistence
- Other Considerations and Synoptic Methods
- Consensus-Based Approach
- Corrected Consensus Scheme
- Guidance on Guidance
- Selective Consensus
Persistence Since the flow in which a TC is embedded often evolves gradually, persistence can be important in forecasting motion within the first 12 to 24 hours. Since, at any instant in time, TC motion tends to be governed by the horizontal advection of relative vorticity, TC motion is mostly a function of the large-scale steering flow. Beyond that time, steering mechanisms and their interactions become complicated and TC track is best predicted using dynamical model guidance instead of persistence. In order to use persistence, a forecaster must determine a representative estimate of TC motion. The initial motion is typically computed with an averaging period over the previous 6, 12, 18, or possibly even 24 hours. An interval of 6 to 12 hours is best if known changes in the track are occurring (e.g., like recurvature), but a longer interval may be chosen if the center location is uncertain. Forecasters should use great care and deliberation when choosing a short-time interval for averaging in the case of TCs near landfall. In general, the time interval for averaging can be thought of as a smoothing factor, with more smoothing and a longer averaging time.
As forecasts from dynamical models have improved, the operational use of statistical models such as CLIPER for track forecasting has declined sharply. There are still several reasons, however, to consult CLIPER guidance in real time. Climatology and persistence, as a very basic forecasting technique, can be used as an early step in the forecast process and may serve as a reasonable “first guess” for the forecast, especially in areas where deviations from climatology are fairly small (i.e, near the equator). In cases where the climatological signal is strong and model spread is small, the forecast rationale could lean more heavily on climatology. Barrett et al. (2005) show that Hurricane Ivan was incorrectly and repeatedly forecast by a majority of the dynamical models to move significantly poleward of the verifying track. Additionally, the right-of-track-biased dynamical guidance was routinely at odds with a climatology-based prediction technique similar to CLIPER, which favored a more westward component of motion (Barrett et al. 2005). A forecaster should examine the strength of the climatological signal for a given forecast scenario and determine to what degree climatology should play a part in the forecast process. Over-reliance on the wealth of numerical guidance without an adequate examination of the forecast scenario cannot guarantee success. Apart from its limited use in an operational setting, CLIPER or similar methods serve as a benchmark against which the performance of dynamical models can be evaluated (Aberson 1998, Bessafi et al. 2002). This evaluation is accomplished by comparing the forecast errors from numerical guidance aginast those from CLIPER. The error output from CLIPER represents a “no-skill” level of accuracy since it contains no information about the current state of the atmosphere. If CLIPER errors for a TC or a season are low (high), the interpretation is that a particular TC or season’s storms were inherently easier (more difficult) to forecast.
Maintaining continuity with the previous forecast should be a priority in the forecast process. Typically, small, incremental changes are made to the forecast from cycle to cycle in an attempt to avoid the sometimes abrupt changes in numerical model guidance (e.g., Figs. 27-29). A forecaster who does not constrain forecast changes with continuity may make a large track change in one direction and then have to adjust the track significantly back in the other direction shortly thereafter. This type of large swing in the track forecast is known as the “windshield wiper” effect. Avoiding these types of radical changes is important since they could cause the public and other users to lose confidence in the official forecast. There are occasionally situations, however, when breaks in continuity and large changes to the forecast track are required, such as the reformation of the center or a significant, unexpected deviation from the previous forecast motion. 
Figure 27. Image of available numerical model guidance at 1200 UTC 6 July 2005 for Hurricane Dennis, then centered over the central Caribbean Sea. The white line represents the official NHC forecast track, and the dashed, multi-colored lines represent other forecast models. The case of Hurricane Dennis in the Atlantic in 2005 is a good example of why maintaining continuity benefits the end user of the forecast. Figure 27 shows numerical model output for 1200 UTC 6 July 2005 for Dennis, then centered over the central Caribbean Sea. Notice how the official forecast (in white) is on the eastern side of the guidance envelope, very close to the GUNA model consensus, the GFDL and UKMET (where GUNA is a simple averaging of GFDL, UKMET, GFS, and NOGAPS). During the next forecast cycle (Figure 28) nearly all of the individual and consensus models shift their track significantly westward. While maintaining continuity, the forecaster only nudged the official forecast track slightly toward the west. One advantage of maintaining continuity is that model guidance tends to vary from cycle to cycle at long lead times, and by making relatively small changes the forecaster can gradually trend toward a solution with time. In this case, if the model guidance shifted back to the right in the next cycle or two, major changes to the official forecast would be required if it closely followed the guidance in every forecast package. To avoid this “windshield wiper” effect the forecaster should make incremental adjustments over several forecast cycles until a clear trend emerges in the guidance and a general scenario is consistently predicted from one cycle to the next, increasing confidence in the forecast. In this Hurricane Dennis example little overall change is noted in the track models (Fig. 28) in the 0000 UTC 7 July cycle. However, the model guidance has shifted back to the east by 0600 UTC 7 July (Figure 29), closer to the original set of guidance from the day before. 
Figure 28. Numerical model guidance through 120 hours for 0000 UTC 7 July 2005 for Hurricane Dennis, then located over the central Caribbean Sea. The white line represents the official NHC forecast. The multi-colored tracks represent other forecast models. 
Figure 29. Numerical model guidance out through 120 hours for 0600 UTC 7 July 2005 for Hurricane Dennis, then nearing eastern Jamaica. The white line represents the official NHC forecast, and the dashed, multi-colored lines represent other forecast models. Other Considerations and Synoptic Methods Interactions between a TC and the surrounding environment are also important to TC motion (Ross and Kurihara 1995). Large, stronger TCs are more likely to have a more significant impact on their environment than smaller, weaker ones, and this environmental modification could, in turn, influence the TC motion. One example is the transport of heat by upper-level outflow poleward of a TC, which could enhance a mid- to upper-level anticyclone and help keep the TC on a general westward heading. TC motion can also be affected by terrain interaction, which varies with the strength of the TC and characteristics of the terrain. Changes in TC motion due to land interaction are sometimes related to convective asymmetries within the storm that are caused by asymmetric surface fluxes and frictional effects (Wong and Chan 2006). The conservation of potential vorticity also plays an important role. Observational and numerical studies have demonstrated that an acceleration and a significant poleward track deflections can occur upstream of a rugged land mass like Taiwan for westward-propagating TCs (Brand and Blelloch 1974; Chang 1982; Bender et al. 1987; Yeh and Elsberry 1993). These same studies show that stronger TCs tend to have more continuous tracks while weaker TCs have track discontinuities as they tend to “jump” or reform on the lee side of elevated terrain. Similar track deflections have also been observed in TCs passing over other islands such as Cuba, Hispañiola, Puerto Rico, and Luzon in the Philippines (Lin et al. 2006). Past need for critical forecast skills such as pattern recognition gave rise to a number of notable case studies on understanding and predicting TC motion. One of the best examples of this is the assessment of the TC in relation to larger-scale subtropical ridges and upper-level troughs in the surrounding environment. For example, the presence of a strong and broad subtropical ridge poleward of a TC will result in a general westerly motion. Conversely, the development of a broad mid- to upper-level trough west of a TC is an excellent indicator of future recurvature (Dunn and Miller 1964, George and Gray 1976). While these more rudimentary tools are not primary forecasting techniques, understanding them is very useful for evaluating model output. Smaller-scale-TC interactions can also be important. The interaction of two or more closely-spaced TCs can lead to what is known as the “Fujiwhara effect” (Fujiwhara 1921, 1923, 1931; Brand 1970) or binary interaction (Dong and Neuman 1983). Fujiwhara (1921, 1923, 1931) explains how same-sized cyclones are mutually attracted to one another and orbit cyclonically about a mid-point. Brand (1970) showed that TCs begin to interact significantly when their centers come within 700-800 nm (1125-1290 km) of each other. Holland and Dietachmayer (1993) found that variations in the maturity of beta gyres surrounding storms can extend the distance at which one vortex is able to influence the other. The degree of interaction increases rapidly as the distance between cyclones decreases. Storm-storm interaction also depends on storm size, intensity, and variations in the ambient steering flow, with storms of unequal sizes likely to have a greater interaction than two of the same size (Prieto et al. 2003). Once two storms come into proximity there are several possible outcomes: one could be captured by the other; there could be a stable, cyclonic orbiting about their common centroid; the two cyclones could merge; one cyclone could escape the interaction; or one of the cyclones could dissipate. The absorption of one TC by another is a rare event and is most likely to occur when one cyclone is much larger and stronger than the other (Lander and Holland 1993). For example, Tropical Storms Gil and Henriette in the eastern North Pacific in 2001 merged once they came within 85 km of each other, after having already rotated about 540 degrees around their centroid (Prieto et al. 2003). Ritchie and Holland (1993) found that vortices must come within 150-300 km of each other before a merger or a “shearing-out” process can occur. Binary interactions are most prevalent in the western North Pacific, where more TCs and the existence of a monsoon trough result in multiple opportunities for TCs to encounter one another (Dong and Neumann 1983). They show that the expected annual number of binary interactions in the western North Pacific is 1.5, considerably higher than the 0.33 expected annually in the Atlantic. When normalized with respect to total storm numbers, however, the western North Pacific experiences the same amount as most other global basins do, except the Atlantic basin (reference??). Observational limitations can sometimes make it difficult to forecast the track for a single TC, and that difficulty can be compounded when considering the interactions of two or more TCs (Dong and Neumann 1983). Carr and Elsberry (1998) presented a detailed conceptual model that classifies binary interaction events by the degree of interaction between storms (direct, semidirect, and indirect), the dominance of one system over another, and a cyclone’s interaction with neighboring steering features or the mean steering flow. Unfortunately, the scheme is quite qualitative in practice, and it can sometimes fail to detect binary interaction or delay the detection. Initially, from the forecaster’s perspective, binary interaction may be masked by the influence of the environmental flow, making it difficult to identify the onset of binary interaction. Contemporary numerical models can better resolve binary interaction than they did a decade or two ago. Although no systematic study has assessed model forecast skill in cases of binary interaction, a forecaster should compare model and actual data near TCs undergoing binary interaction in order to anticipate model error. Models often fail to correctly analyze TC intensity, size, and structure, and this can result in a degradation of the forecast, even in the short-term. To the extent that models do not accurately initialize the convective structure and vorticity, there can be additional forecast error, especially in the case of binary interaction. Oscillations in TC track are also possible through other means. Holland and Lander (1993) discuss how TCs tend to oscillate about a mean path, even if they are in a particularly well-defined steering current. This can lead to erratic motion such as short-period, trochoidal motion. Willoughby (1988) demonstrated that a mass source and sink couplet exhibited trochoidal oscillations along the cyclone path, and these were documented observationally in Atlantic Hurricane Belle in 1976 ( Lawrence and Mayfield 1977). Willoughby (1984, 1990) and Holland and Lander (1993) further discuss the role that convective asymmetries, such as mesoscale convective systems rotating within the core of the TC, can have on short-term TC motion. Ritchie and Elsberry (1993) acknowledge the possibility that significant track changes on the order of tens to even hundreds of kilometers over several days could result between the interaction of a TC and a persistent MCS rotating within its circulation. Ritchie and Elsberry (1993) show the effects of a long-lived MCS within the core of Typhoon Robyn and then simulate these effects using a high-resolution model. They found that a persistent burst of convection, especially in an asymmetric pattern, can cause a localized pressure fall through latent heat release which can cause a “jump” or reformation of the center. Models and forecasters have not routinely predicted convectively-related meanders observed in TCs. This was the case on 12 July 2003 when the center of Tropical Storm Claudette, then located over the central Gulf of Mexico, unexpectedly jogged north of the forecast track. Claudette was embedded in a persistent southwesterly shear environment, which was causing the bulk of the convective activity to be distributed asymmetrically from north and northeast of the center. The asymmetry appeared to tug the center northward throughout the day, with the center eventually reforming closest to the deepest convective cells several hours later. Neither forecasters nor the track guidance anticipated this motion, and the likelihood of being able to forecast short-term track motion with a time scale of this type of convective asymmetry is not very high. If there is reason to think that an asymmetric convective pattern will persist, this may allow a forecaster the opportunity to deviate from the track guidance (Ritchie and Elsberry 1993). However, extrapolating a short-term motion too far into the future, especially after the initial forcing agent has dissipated, is not advisable. Consensus-Based Approach Although a wide variety of models is available for TC track forecasting, no single model is consistently the best. The reliability of even the best-performing models fluctuates from season to season due to changes in model formulation or data assimilation schemes. These fluctuations can also occur from one run to the next due to errors in model initial conditions and how individual models handle critical features, including the TC itself. In the past decade, it has been found that grouping several independent models together to form a multi-model consensus can yield more skill than can be found in any individual model. Goerss (2000) illustrated the value of the consensus-based approach by forming a simple global-model consensus of three models and a regional-model consensus constructed with two. In each case, Goerss (2000) demonstrated that the consensus track errors were significantly less than that of the best individual models. The addition of new models to the consensus often leads to minor but steady gains in skill, especially if the number of models in the consensus is small. In fact, the inclusion of a model that is generally less skillful than the other consensus members can still yield a positive contribution. This improvement stems from the fact that an ever-larger consensus produces smaller mean errors by reducing the influence of large errors associated with any individual model. In other words, when a single model performs poorly, the other models mitigate the impact of any poor performer within the consensus. The “smoothing out” of random errors from individual, independent models is demonstrated in a comparison of various individual and consensus models from a series of hurricane seasons, as presented in Figure 30. Notice how the GUNA model consensus, an average of the GFDL, UKMET, NOGAPS, and GFS models, outperforms not only all of the deterministic model runs but also the GFS ensemble mean. In addition, Rappaport et al. (2009) report that the GUNA consensus outperformed the GFDL by 18% between 2004 and 2006; the GFDL was the most skillful individual model of the GUNA constituents. Several other possible combinations of models are noted in Figure 30, which shows the current of makeup of consensus models used at the National Hurricane Center. The building of a consensus must take into account model availability, which can differ considerably from model to model. The ECMWF global model, for example, is the latest-arriving of all the guidance at the National Hurricane Center, making its inclusion into any model consensus a critical factor. Some consensus-based solutions require that all members of the consensus be available for computing the model consensus (e.g., GUNA or TCON, an average of GFS, UKMET, NOGAPS, GFDL, and HWRF), while others require a minimum of two members available (e.g. TVCN, an average of at least two of the GFS, ECMWF, NOGAPS, UKMET, GFDL, GFDN, and HWRF). These are referred to “fixed” or “variable” consensus models, respectively. Increasing the number of consensus members increases the chance of having the minimum number of models available to form an accurate forecast consensus. A variable consensus, however, may introduce consistency issues which a forecaster must remember, as the 120-hour forecast may be based on a different set of models than the 96-hour forecast position. The effectiveness of a model consensus relative to each individual member is a function of the independence or effective degrees of freedom of the models forming the consensus. Sampson et al. (2006) showed that the inclusion of a relatively simple barotropic model in the consensus produced a large positive impact in spite of that model’s generally larger errors. Therefore, even average- or below average-performing models can contribute to the skill of the consensus. Corrected Consensus Scheme One technique that attempts to further increase forecast skill forms a “corrected” consensus after analyzing and correcting for the error of the individual models that comprise the consensus. The corrected consensus is developed by first statistically regressing model tracks against the ‘best track’ and then assigning weights to each member model solution according to that model’s past performance (Kumar et al. 2003). In essence, the corrected consensus scheme is updated as it learns from past events. The weighting factors can be computed in real time and applied during the season or they can be calculated from season to season (Weber et al. 2003). The FSU Superensemble (FSSE) is one example of a corrected consensus scheme, where different weights and corrections are applied to individual models depending on their biases (Williford et al. 2003). A linear regression model is run to determine model weights during a training period, which is normally taken to be at least a few seasons. Another corrected consensus (TVCC) is formed by applying east-west and north-south predictors to TVCN (simple model consensus), which are either statistical or bias correctors (Goerss 2007 ??). The statistical correctors are found using the same techniques used for the Goerss Predicted Consensus Error (GPCE: discussed in the section Guidance on Guidance), except that the predictand is east-west or north-south error rather than consensus position error (Goerss, personal communication). The same predictors used for GPCE are used in this regression, although the final predictors and weights are different for TVCC. Recent verification of corrected consensus methods illustrate the difficulty in using past performance as a guide for \attempting to correct future model errors. For example, the past performance of member models may not be representative of their current performance, typically due to major changes made to a model between seasons. If changes to an individual model are made before new weighting factors can be computed, the performance of the consensus can be degraded. As a result, in recent years the operational performance of the corrected consensus has been inferior to that of the simple consensus methods. This may be due to the increasingly frequent model changes and also the small sample sizes of TCs from which conclusions about storm behavior cannot be made due to meteorological conditions which are too varied across the available sample (Franklin 2009).
Additional skill in track forecasting is possible through the use of a selective consensus technique, where the forecaster subjectively chooses which models will be part of the consensus. The rejection of a particular model from a standard consensus requires an evaluation of its performance, which could come from its error statistics. In the western North Pacific an expert system named the Systematic Approach Forecast Aid (SAFA) was introduced in 2000 to assist forecasters in determining error characteristics for the available track models (Carr et al. 2001). Elsberry and Carr (2000) demonstrated that eliminating a model suspected of underperforming improved the consensus in nearly all cases they examined where forecast errors were critically large. However, operational use of this method has yielded mixed results at best, with many more cases seen where little to no appreciable improvement to the forecast has been noted. Consequently, the use of selective consensus remains a difficult challenge. A selective consensus can occasionally produce positive results, but the forecaster must recognize when it is appropriate to utilize this method. Sampson et al. (2006) argue that, aside from adding more models to the consensus, model improvements in recent years have made it difficult to outperform a common consensus. This is particularly true in cases of small model spread, when beating the consensus becomes increasingly unlikely. When model spread is relatively large, the opportunity for the forecaster to increase model skill by selectively choosing a consensus is rather limited and in only rare cases results in an improvement. The removal of “bad” guidance should be attempted with caution, since the mistaken elimination of “good” guidance can degrade the consensus, especially when model spread is small. Nonetheless, a selective consensus may be useful in cases of larger-than-average model spread, if available information leads a forecaster to favor one forecast scenario over another or completely reject a scenario. NHC forecasters were able to register marginal skill at nearly all forecast lead times when forming selective consensus solutions in the Atlantic basin in 2007 (Figure 31). However, no details of model spread characteristics were recorded. With only small sample sizes for each forecast time and only one year of data, the results cannot be considered statistically significant. Figure 31. Comparison of a selective consensus approach (RYOC) and the CONU, a simple consensus method.
Download 197.74 Kb. Share with your friends:
|
