FIFTH INTERNATIONAL WORKSHOP on TROPICAL CYCLONES

3.3.1 Background:

Ensemble forecast techniques have been available for use in weather and climate prediction for several decades. However, it is only in the past decade or so that ensemble forecasting methods have developed to the extent that they are now an integral part of both research and operations in numerical weather prediction (NWP) systems at major weather centers. This is particularly true for centers that have responsibility for tropical cyclone (TC) forecasting and issuing of warnings. The techniques are variously purely statistical, purely numerical, or a hybrid of both approaches. The success of the approach continues to grow as the refinement of ensemble prediction methods proceeds and the extension expands.

Apart from some earlier simple attempts at ensemble prediction, the specialist application of ensemble forecast methods to TC prediction has occurred only over approximately the past five years. The application to TCs, like much research in the tropics, followed the extension of ensemble forecasting to severe weather events in the mid-latitudes. In particular, it followed successful research applications to short-range ensemble forecasting (SREF) of explosive mid-latitude cyclogenesis.

From the mid-1990s to the present, the development of ensemble perturbation methodologies for TCs has developed rapidly, especially in the past few years. Ensemble predictions of TC tracks are now becoming available on various websites including some of the major weather centers and other institutions that have responsibility for NWP over the TC basins. In recent years, ensemble techniques have become invaluable and increasingly reliable. The example of TC Lili (below) that affected the Gulf of Mexico in early October, 2002 is a case in point.

An example of an ensemble forecast for TC Lili is shown in Figure 1 below. The ensemble approach was a multiple model ensemble with multiple initial conditions. In this case, the members of the ensemble are forecast tracks from different weather centers, as shown in the legend above the figure. There was very little spread among the ensemble members, so the mean forecast track was little different from any ensemble members. This close correspondence between ensemble members continued over most of the lifecycle of Lili after it reached the southeastern Gulf, near the southwest tip of Cuba. The ensemble consistently predicted landfall in Louisiana out to four days in advance. The fact that the ensemble members are all close together is often, but not always, a sign that the mean error of the ensemble prediction is small. It certainly was true in the case of TC Lili. In a TC context, Aberson et al. (1995) examined the relationship between ensemble spread and ensemble mean error. More recently, Elsberry and Carr (2000) have quantified consensus error as a function of consensus spread, extending earlier work on consensus dynamical TC track forecasts by Goerss (2000). However, as we shall discuss later, it has been established that a small ensemble spread is not always an indicator of a small ensemble mean error and the correlation between ensemble spread and error is an ongoing research area. TC Lili also provided an excellent example of a marked deterioration in the forecast accuracy as Lili approached and made landfall. Both the timing and intensity predictions became significantly less accurate than was the case when Lili was in the Gulf of Mexico, well away from land. This deterioration in forecasts as TCs approach landfall has been documented by various researchers. For example, the recent article by Aberson (2001) discusses in considerable detail the problem of lack of improvement of TC ensemble track forecasts at landfall for the North Atlantic basin.

3.3.2 Methodologies:

It is generally accepted that ensemble prediction methods were developed with at least two main aims in mind. The first goal was to build on established theory for midlatitude baroclinic systems that a probability density function (PDF) could be generated by perturbing the initial state of a forecast model. PDFs could be generated at various times in the future and then be used to generate additional information that a single forecast could not provide. Most notable of the additional information provided is knowledge of the various statistical moments of the forecasts. In particular, it had been shown that the ensemble forecast mean theoretically is more accurate than any individual ensemble forecast member. Additional benefits also were readily available, including valuable estimates of error growth rates and associated predictability time scales. Such information was almost entirely lacking at the time and has since become a key area of research and applications. A second force driving the development of ensemble prediction techniques was the need for much higher resolution models than computational platforms allowed at the time. During the mid 1990s, the modeling of explosive cyclogenesis, whether tropical or extratropical, was severely limited by an inability to achieve model horizontal resolutions commensurate with the spatial scales of the cyclones being modeled. As a consequence, ensemble prediction techniques were viewed as a possible means of achieving the effective skill of a higher resolution forecast without needing the punitive resources required when increasing the model resolution. Remember that for most models an increase in horizontal and vertical resolution of a (modest) factor of two requires a time step reduction of a similar factor and an overall computational demand that is sixteen times as great as the original resolution! When the extra information provided by an ensemble prediction is taken into account, such an alternate approach had obvious appeal. However, it is now widely appreciated that adequate model resolution is a necessity and that rather than replace high resolution models, ensemble methods complement rather than replace the single model approach.

As stated in the Background, the ensemble prediction approach generates a set of forecasts by perturbing the NWP system in various ways and producing forecasts from the perturbations. From the early perturbation generation based on Monte Carlo procedures, or simple weighted averaging of forecasts, a plethora of approaches now exists. At least ten methods have been developed over the years and most are still in use. No attempt will be made here to describe each method as they all are well-documented in the literature. They include:

• Monte Carlo methods in their original formulation in which random values within observational errors are added to analyzed model variables at model grid points. The fields are then usually are truncated at an “appropriate wave number” to restrict the generation of gravity waves.
  

• Modified Monte Carlo methods in which the random perturbations are applied to the original data and re-analyses are carried out to generate the final perturbed initial states.
  

• The Perturbed Observation approach, which can be based on a Monte Carlo procedure but avoids the gravity wave problems introduced by the randomness of the conventional Monte Carlo schemes.
  

• The Lagged-Average Forecast method, originally introduced at the National Centers for Environmental Prediction (NCEP) as an alternative to the original Monte Carlo method. In this method, a sequence of forecasts with different initial times is averaged at the same final time.
  

• The Singular Vector Decomposition method developed at the European Centre for Medium-range Weather Forecasts (ECMWF) to identify the fastest growing modes.
  

• The Breeding of Growing Modes method developed at NCEP as an advance on the Lagged-Average Forecast method and is also, like the ECMWF method, an attempt to capture and treat the fastest growing modes.
  

• The use of analyses from multiple NWP centers as initial conditions for a use in a one model to generate an ensemble forecast based on the particular model at each of the participating NWP centers.
  

• The use of model forecasts from different NWP centers. In this approach, a given center need only to carry out one forecast and the remaining ensemble members are provided by other NWP centers. The final ensemble is derived from these members by a range of averaging or weighting procedures.
  

• The EOF Eigenmode perturbation method developed at Florida State University. This approach is one of the very few that have been designed specifically for TCs.
  

• The Super-Ensemble method (also developed at Florida State University), which is an extension of the approach using models from different weather centers and then applies a statistical adjustment to reduce bias.
  

• The Self-adapting Analog Ensemble technique, which is a purely statistical method, developed at the University of Hamburg.
  

• The trade-off between using computational resources for ensemble methods rather than for increasing resolution and sophistication of a single deterministic model.
  

• An over-emphasis on the initial state that overshadows the importance and need for increased knowledge of model errors.
  

• For regional models, the need for and impact of perturbing boundary conditions.
  

• The need for a continued effort in improving existing perturbation methods and in searching for new ensemble techniques
  

• A requirement for procedures that make better use of the perturbation method output. There are many examples here, and many problems with how best to make use of the evolving PDF. In their raw form, it has been found in many applications that ensemble spread and skill are not sufficiently highly correlated to provide a reliable estimate of the skill of the individual forecasts. Attempts to counter this problem include various cluster analysis methods, which are showing great potential as research efforts and applications increase in this topic.
  

• How well are rare events captured by procedures in which the statistical methodologies are inherently conservative and may fail to forecast outliers? These outliers often are the very systems that cause the most destruction.
  

• How good are current verification procedures as measures of skill? It is generally agreed that there is a need for further work in this most important research and development area.
  

• What is the relationship between skill and value in a TC ensemble forecast? This is an emerging research area that is of great economic importance.
  

• Sensitivity studies are required to assess the importance of and need for proxy and synthetic data, including outgoing longwave radiation (OLR) and TC bogus data. Many studies have shown that forecasts can be very sensitive to these kinds of data. The study by Leslie and Holland (1995) demonstrated clearly that forecasts are highly sensitive not only to the kind of bogus vortex but also to the location of the bogus vortex in the environmental flow.
  

• Ensembles should not neglect the far-environment as TCs often move into that environment and many forecast failures occur as a result of the lack of care in specifying the far environment and hence its subsequent evolution. Returning to the above TC Lili example during early October 2002, the ensembles approach provided excellent guidance concerning track and timing when TC Lili was well away from land. Upon nearing landfall, the skill of timing and intensity predictions fell away as hurricane-induced cooler water, unfavorable vertical wind shear and intrusion of drier air from a neighboring system may have all contributed (in an as-yet-unknown order of importance) to dramatic structural change and weakening of TC Lili.
  

• A comparison among the many methods listed above.
  

• Establish a theoretical basis for TC initial conditions, as has been carried out for baroclinic systems.
  

• A major push in transition to operations. The research effort on TC ensemble prediction appears to be weighted far more heavily than implementation in operations. This emphasis varies between TC basins.
  

• Assessment of the TC tracks currently produced on a routine basis by the major weather centers. The TC tracks often are simply a by-product of the emphasis that still exists on mid-latitude ensemble prediction systems.
  

• As a consequence of the previous point, there is a need for the establishment of groups within the operational centers or collaboration with groups outside the centers whose members should evaluate the TC tracks, and develop appropriate techniques for the utilization and training of forecasters on the use of the predictions in an ensemble setting.
  

• There is a need to work more closely with mathematicians/statisticians on developing encouraging areas of research such as cluster analysis of the ensemble predictions.