1. Introduction
Ensemble predictions of tropical cyclone (TC) tracks have become available from a number of numerical centers for TCs that are already present in the initial conditions. For example, the European Centre for Medium-Range Weather Forecasts (ECMWF) integrates the Variable Ensemble Prediction System (VarEPS) in addition to its medium-range deterministic model (van der Grijin et al. 2005). The new generation of ECMWF products includes an extension of the tracking and strike probability maps from 5 to 10 days and probabilistic information on the storm intensity (Vitart et al. 2012). Although verifications of the ensemble mean tracks might be expected to be less skillful than the tracks of the “mother” deterministic high-resolution model because the ensemble model is integrated at a coarser horizontal resolution, some studies indicate the ensemble mean may have marginally more skill at longer forecast intervals. Elsberry (2010) notes that the ensemble products are not available to the forecasters until 6 h later than the consensus of deterministic models that many forecasters use as the primary track guidance, and this time delay must be taken into account when comparing ensemble track errors with the official track forecast errors.
Rather than just generating a mean track that might be marginally better, a primary objective of the track ensemble is to provide information on the uncertainty in the track forecast. The ECMWF constructs strike probabilities that the center of the storm will pass within 120 km of a point within the next 10 days. Superposing the ensemble mean track on these strike probabilities then provides a visualization of the ensemble spread around the mean, and this swath is an indication of the track uncertainty. Dupont et al. (2011) have evaluated calibrated probabilistic track forecasts in the South Indian Ocean based on the ECMWF VarEPS and found skill relative to climatology in predicting track uncertainty for lead times up to 3 days. Whereas Majumdar and Finocchio (2010) had found that the ECMWF was the most skillful ensemble they tested, the VarEPS forecasts in the Atlantic were underdispersive (i.e., ensemble spread underestimated the track uncertainty). Yamaguchi et al. (2012) also found the ECMWF ensemble was the best individual ensemble in their multicenter grand ensemble.
A new “genesis” prediction from ECMWF that was provided in 2012 is the strike probability for the occurrence of a TC-like vortex with a maximum wind speed of at least 8 m s−1 passing within 300 km of a given location during some period in the medium-range forecast (Vitart et al. 2012). A similar display of the TC activity within 300 km is available from the ECMWF 32-day ensemble integrated twice a week from initial conditions at 0000 UTC on Mondays and Thursdays.
Vitart (2009) had demonstrated that an ECMWF 15-member ensemble hindcast could reproduce the main characteristics of the observed distribution of TCs, although the TC activity in the model was higher than observed. Vitart (2009) had attributed the skill of these activity forecasts to the ability of the ECMWF ensemble model to predict the Madden–Julian oscillation (MJO). Vitart et al. (2010) also demonstrated that after calibration the ECMWF ensemble could predict the TC activity in the Southern Hemisphere up to 3 weeks with skill comparable to the statistical models. Belanger et al. (2010) examined the predictability of North Atlantic tropical cyclones using the 32-day ECMWF VarEPS during the 2008 and 2009 seasons. Predictability to 15–21 days was indicated in the main development region for Atlantic TCs. Belanger et al. (2012) have also studied the predictability of TC genesis and regional outlooks of TC activity for the Arabian Sea and the Bay of Bengal. They found similar levels of predictability as in the North Atlantic.
Elsberry et al. (2010) evaluated the predictability of western North Pacific TC formations and tracks out to 4 weeks during the 2008 season using the ECMWF 32-day ensemble forecasts. Rather than a strike probability approach, they combined similar ensemble member vortex tracks and used a weighted mean vector motion technique (Elsberry et al. 2008) to form ensemble storm tracks. Elsberry et al. (2011) examined the more typical and active 2009 season using the same approach and found improvements in predicting the formations and tracks of typhoons and the improvement also applied to most of the tropical depressions on time scales of 5–30 days.
Elsberry et al. (2010, 2011) used a two-step objective plus subjective approach for verifying the ensemble storm tracks relative to the Joint Typhoon Warning Center (JTWC) best tracks. Their approach was to first treat the JTWC track as another storm and compare all of the applicable forecast ensemble storms at each 12-h forecast time to determine if at least one point on the ensemble storm track matched within a separation distance ε(t) the JTWC position at exactly that 12-h time. Second, a subjective assessment was made to assign a quality metric (excellent, above average, good, below average, and poor) of the match of the ensemble storm to the JTWC track.
Tsai et al. (2013) have developed an objective track analog verification technique to select all ensemble storm tracks predicted by the ECMWF 32-day ensemble that match the overall JTWC tracks. For ensemble storms within specified time and space differences of the JTWC track that are potential analogs, four metrics of shortest distance, average distance of the matched points, distance at formation time, and distance at ending time are calculated. An objective quality measure that assesses the overall track similarity between the potential analogs and the JTWC storm is calculated in terms of membership functions for the four track metrics. Weighting factors multiplying these membership functions were adjusted to approximately match the subjectively determined quality measures for the ECMWF ensemble storm forecasts during the 2009 season (Elsberry et al. 2011).
Objective verifications for the 2009 and 2010 seasons were then summarized by Tsai et al. (2013) in terms of hits, misses, false alarms, and correct negatives that no TC would be present in the western North Pacific. The most important result was that the ECMWF ensemble was able to predict out to 4 weeks nearly all of the TCs in both seasons and with only a small number of misses that generally were short-lived tropical depressions. Good performance in terms of correct negatives was achieved, especially during the 2010 season, which had the least TC activity in recent times.
False alarms were defined by Tsai et al. (2013) to be all ensemble storms that could not be matched with JTWC storms within the specified thresholds. Evaluations of the characteristics of the false alarms indicate seasonal and geographic biases and that about 50% of the false alarms during the first week originate from the initial conditions in the model. After a minimum of false alarms originating in the week-2 forecasts, which was attributed to the decrease in horizontal resolution in the model that occurs at day 10, a steady and nearly uniform increase in false alarms originating in the week-3 and -4 forecasts was documented.
The purpose of this study is to evaluate the ECMWF 32-day ensemble forecasts, which were kindly supplied by ECMWF in support of the National Aeronautics and Space Administration (NASA) Hurricane and Severe Storm Sentinel (HS3) field experiment, for TC events in the Atlantic during May–December 2012. Since essentially the same techniques for generating ensemble storms as in Elsberry et al. (2010, 2011), and for objectively verifying the forecasts as in Tsai et al. (2013), were utilized, these will only be briefly summarized in section 2. An overview of the ECMWF ensemble performance during the 2012 Atlantic season will be given in section 3. Some case studies of individual storms will be presented to illustrate some successes and some failures in extended-range forecasts. A description of the false alarms in the Atlantic will be given in section 4. Finally, a summary and future tasks will be provided in section 5.
2. Methodology
While the reader will find the details in Elsberry et al. (2010, 2011) and Tsai et al. (2013), a number of features of the technique and methodology will be summarized in the appendix, including (i) characteristics of the ECMWF ensemble compared to the earlier studies and (ii) a detailed summary of the modified track analog technique for verifying the ensemble storm tracks that is essential for understanding the following verifications during the 2012 Atlantic season.
Following Elsberry et al. (2010, 2011), the evaluations of the ECMWF ensemble predictions will begin for the Monday or Thursday forecasts before the National Hurricane Center (NHC) begins to follow the system, which will be defined as week 1. An example of the 0000 UTC 30 July 2012 ECMWF ensemble forecast of pre–Hurricane Ernesto is used to illustrate several aspects of the ensemble storm generation and validation as applied in the Atlantic (Fig. 1). The first entry in the NHC working best track for Ernesto (AL05) was at 1200 UTC 1 August at 11.6°N, 46.7°W (Table 1). Notice that the ECMWF ensemble forecast initiated 2.5 days earlier already had a well-defined track that began about 14° longitude to the east. The ensemble vortex tracks that make up ensemble storm 1 then had a small spread about the Ernesto track until 70°W, when a number of vortex tracks began to turn poleward. However, the majority of the vortices continued to the west-northwest, so the weighted-mean vector motion ensemble storm track was almost perfect until it passed to the west of the Yucatan Peninsula. Whereas Ernesto actually moved inland over Mexico and decayed, ensemble storm 1 was predicted to remain and moved over the western Gulf of Mexico until it sharply turned to the northeast and moved over the southeast United States. As Elsberry et al. (2011) documented in the western North Pacific, the ECMWF ensemble vortices are often predicted to continue large distances over land, as in Fig. 1.
Ensemble storm 1 (dark gray) generated from the weighted-mean vector motion of ECMWF 32-day ensemble vortices (light gray) in the week-1 forecast initiated at 0000 UTC 30 Jul 2012, which was then matched by the analog track verification technique with the NHC WBT of Hurricane Ernesto (AL05, dashed black) with an LHV = 0.55. The first position in the WBT is at 1200 UTC 1 Aug, which is 2.5 days after the first position of ensemble storm 1.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
Storm number and name, as well as time, intensity category, and space characteristics from the 2012 Atlantic WBTs grouped according to their forecastability by the ECMWF 32-day ensemble (see text and Fig. 2 for the corresponding LHVs during weeks 1–4).
The ensemble storm 1 prediction in Fig. 1 is referred to as a week-1 forecast because it was initiated on Monday, 30 June, in the range of 1–7 days before pre-Ernesto existed according to the NHC WBT. The next forecast initiated on Thursday, 2 August, is not evaluated because Ernesto (AL05) began on 1 August and the interest here is on the formation and subsequent track. The forecast on Thursday, 27 July, would also be a week-1 forecast since it is also in the range of 1–7 days before pre-Ernesto existed. The Monday, 23 July, and Thursday, 20 July, forecasts would be designated as week-2 forecasts, etc. Fewer ensemble vortex tracks exist to form an ensemble storm at longer forecast intervals than in week 1 (Fig. 1), but a minimum of three vortices and a track of at least 3 days was imposed to include a maximum number of potential ensemble storms without considering short-lived systems that are not likely to be representative of tropical storms and hurricanes.
When the ensemble storm 1 track is validated with the objective track analog technique of Tsai et al. (2013; see summary in the appendix), the quality measure called the likelihood value (LHV) is only 0.55, even though the track overlies the Ernesto track from 40° to 90°W. This reduction from LHV = 1.0 is due to the deduction for the forecast ending point being far to the northeast, and partly for the forecast starting point being to the east. However, the NHC forecasters may suggest the weighting factors for the ending point and the formation point may need to be reduced to better reflect the ensemble storm 1 track forecast overlapping most of the Ernesto track and be more accepting of the early start of the forecast track. Nevertheless, the same weighting factors of the four metrics in the quality measure LHV will be specified here as they were in Tsai et al. (2013) for the western North Pacific.
Occasionally, two or more ensemble storms with similar tracks existed in a forecast and were both matched with a NHC storm by the track analog verification technique. Such multiple similar tracks are more common in the western North Pacific (Elsberry et al. 2010) and code was developed to merge two ensemble storms that were subjectively determined to have similar tracks. As this capability to combine two similar tracks had only been applied during the later part of the 2012 Atlantic season, in this evaluation the ensemble storm track with the higher LHV will be selected if multiple ensemble storms from the same forecast were matched to an NHC storm.
3. Evaluation of EMCWF 32-day ensemble performance during the 2012 Atlantic season
In addition to the period of the NASA HS3 field campaign of August–November, ECMWF also provided the forecasts starting from 31 May and extending to 17 December. Specifically, forecasts of Tropical Storms (TSs) Alberto (AL01) and Beryl (AL02) were not included in the study period.
a. Overview of the 2012 season
The performance of the ECMWF 32-day ensemble forecasts of Atlantic TC events during the 2012 season is summarized in Fig. 2 in terms of the quality measure LHV for the ensemble storm that matched an NHC storm (see the LHV definition in the appendix). For example, the summary for Ernesto (AL05) in Fig. 1 is LHV = 0.58 (0, 0.94, 0.29) for week 1 (2, 3, 4), which would be designated as good (miss, excellent, below average) in Tsai et al. (2013).
Quality metric LHV (ordinate) for weeks 1–4 (symbols in inset) for the AL03–AL19 TSs and H1–H3 hurricanes during the 2012 Atlantic hurricane season. An LHV ≥ 0.2 is considered a hit and an LHV < 0.2 is considered a miss. Note that LHV = 0 for all four weeks for AL03, AL07, AL13, AL16, and AL19.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
An immediate conclusion from comparing the 2012 Atlantic summary of LHVs in Fig. 2 with the western North Pacific summary for the 2009 and 2010 seasons (Tsai et al. 2013) was that the ECMWF ensemble performance on extended-range time scales was more limited in the Atlantic. Since a small sample of 17 Atlantic storms during 2012 cannot be used to assess predictability, this evaluation of the ECMWF ensemble performance will be characterized in terms of a “forecastable” metric for certain types of TC events. That is, the objective is to inform the NHC forecaster that some Atlantic TC events may be forecastable in the extended range, but other types of Atlantic TCs are not. Indeed, the variation of forecastability is from 4 weeks to no forecastability even 1 week before the storm has appeared in the NHC WBT.
b. Forecastability of individual storms
Four groupings of the storms (Table 1) will be described. These groupings are characterized as 1) highly forecastable (EF4), which is defined as cases where all forecasts for weeks 1–4 had a match; 2) somewhat forecastable (EF3), which will be for ensemble storm forecasts that had LHV ≥ 0.2 in only three of the four weeks; 3) limited forecastable (EF2), which were ensemble storm forecasts that had LHV ≥ 0.2 in only the first two weeks; and 4) not forecastable (EF0), storms for which the ECMWF 32-day ensemble forecasts failed to have a match in all four weeks.
1) Highly forecastable (EF4) storms
Only two hurricanes (AL11 and AL12) and two tropical storms (AL10 and AL15) were successfully forecast in all four weeks, which is in strong contrast with Tsai et al. (2013), who found that nearly all of the typhoons and even many of the tropical depressions in the western North Pacific during 2009 and 2010 were successfully forecast out to 4 weeks. Some characteristics of these EF4 storms are given in Table 1a.
The ensemble storm forecasts that could be matched with Hurricane Leslie (AL12) during all four weeks are given in Fig. 3. Ensemble storm 8 in the Thursday, 23 August, forecast that is designated as week 1 (Fig. 3a) had a track that overlapped the Leslie track for much of the forecast period. As in Fig. 1, the week-1 forecast for Leslie had an LHV of only 0.54 (Fig. 2) due to deductions for a large forecast track ending difference and a premature starting location more than 25° longitude to the east. While the large-scale environmental conditions do not accurately predict the timing of formation according to NHC, they do determine the track type and two factors in the LHV measure the track characteristics, which are arguably more important than the precise formation time.
Ensemble storm, member vortices, verifying WBT, and LHV as in Fig. 1, but for Hurricane Leslie (AL12, black dashed) based on ECMWF 32-day ensemble forecasts initiated during (a) week 1 on 23 Aug, (b) week 2 on 20 Aug, (c) week 3 on 13 Aug, and (d) week 4 on 6 Aug.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
For the week-2 forecast (Fig. 3b), the ensemble storm 19 track forecast starting position was quite close to the actual position. In addition, the subsequent track forecast was quite similar to the Leslie track, so both the closest position and the average position difference metrics were highly scored. However, the forecast track ending position was quite poor, so the LHV was fully reduced by 0.22 (weighting factor) to 0.56 (Fig. 2). For the week-3 forecast (Fig. 3c), the ensemble storm 39 starting position and the early portion of the forecast track agreed quite well with the Leslie track. However, Leslie turned sharply northward and almost stalled for several days. While one ensemble vortex track had a close similarity to this track change, other vortex tracks had a more westward path, so that a deduction was made for a reduced average position metric and the large forecast ending position difference. Thus, the LHV = 0.70 is in the above average category. Finally, the week-4 forecast (Fig. 3d) that was initiated on 6 August, or 23.25 days before pre-Leslie started, had a forecast starting position and an early track that agreed well with the Leslie track, but an excessive ending distance metric at the time corresponding to day 32 in the forecast led to the quality measure LHV = 0.78.
Hurricane Kirk (AL11) was the second hurricane that was classified as a highly forecastable (EF4) storm with a quality measure LHV > 0.4 in all four weeks (Fig. 2). A large fraction of the ECMWF ensemble members were predicting pre-Kirk in the 0000 UTC 23 August forecast (Fig. 4a), which is the week-1 forecast, as Kirk began 30 h later at 0600 UTC 24 August (Table 1a). The good agreement in the initial positions and in the early track contributed to an LHV = 0.73 (Fig. 2). However, Kirk had a relatively sharp recurvature that only a few ensemble vortices predicted. Rather, the majority of the ensemble vortices recurved the storm about 10° longitude farther west than Kirk. A significant number of vortex tracks had the recurvature much farther to the west, but the weighted-mean vector motion calculation would give little weight to these tracks compared to the larger number of vortices with tracks closer to the prior ensemble storm 3 track.
As in Fig. 3, but for Hurricane Kirk (AL11) for forecasts initiated during (a) week 1 on 23 Aug, (b) week 2 on 13 Aug, (c) week 3 on 6 Aug, and (d) week 4 on 30 Jul.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
Ensemble storm 33 in the week-2 forecast (Fig. 4b), which initiated at 0000 UTC 13 August, again had a starting point on the coast of Africa and then had a more westward track that resulted in a recurvature about 17° longitude west of Kirk. Although the shortest distance metric is satisfied because the forecast track crossed Kirk’s track, the deductions for the forecast starting point, average distance, and forecast ending point resulted in an LHV = 0.51 (Fig. 2). The week-3 forecast (Fig. 4c) is rated at LHV = 0.81 even though the starting point is again on the coast of Africa and the initial track is more westward because two crossings of the Kirk track lead to a zero shortest distance metric and a smaller average distance as the forecast recurvature point is only 7° longitude west of Kirk. Finally, the week-4 forecast (Fig. 4d) is also highly rated (LHV = 0.92) due to an excellent forecast starting point, a forecast track crossing, a small average distance, and that the distance ending time is only considered to the end of the 32-day ensemble forecast (i.e., the later part of the actual track is not verified).
TSs Joyce (AL10) and Oscar (AL15), which are also classified as EF4 storms (Fig. 2), began off the west coast of Africa, as did Leslie and Kirk, but both had short tracks. In both the week-1 forecast—see Fig. S1a in the supplemental material—and week-3 forecast of Joyce (Fig. S1c), the only LHV deduction of 0.22 was for a much longer forecast track. For the week-2 (Fig. S1b) and week-4 (Fig. S1d) forecasts, the track was only slightly longer than for TS Joyce, and the primary deduction was for the average distance and again a forecast starting on the coast of Africa rather than downstream. Because the week-1 forecast for TS Oscar (Fig. S2a in the supplemental material) was initiated at 0000 UTC 10 October (Table 1a) that was also the starting time for pre-Oscar, the forecast starting position was good and the large number of ensemble vortices contributed to an excellent forecast of a reversed-S track. Again the LHV deduction of 0.22 was for the longer forecast track. The week-2 forecast (Fig. S2b) had a similar shape as the track of Oscar. However, it started on forecast day 17 and was 3 days late (rather than the usual premature start), so the forecast track was shifted to the northwest and the LHV was only 0.55. Both the week-3 (Fig. S2c) and week-4 (Fig. S2d) ensemble storm-track forecasts started on the coast of Africa rather than downstream, but the track lengths were short, similar to TS Oscar, and thus the LHVs of 0.77 and 0.68 were in the above average category.
In summary, these four EF4 storms were forecast by the ECMWF on extended-range time scales (5–30 days). It should be noted that these four highly forecastable storms generally occurred in the midseason and all developed from African systems. Although the forecast tracks were frequently too long (and thus resulted in an LHV deduction of 0.22), they are considered to be highly forecastable and potentially useful outlooks of specific TC events in the Atlantic.
2) Somewhat forecastable EF3 storms
Four hurricanes and one tropical storm were classified as EF3 storms (Table 1b), as the EMCWF 32-day ensemble was able to successfully forecast these storms in three of the four weeks (Fig. 2). In Hurricanes Isaac (AL09), Nadine (AL14), and Sandy (AL18), the unsuccessful forecast was in week 4 (as might be expected), but for Hurricane Ernesto (AL05) and TS Florence (AL06), the unsuccessful forecast was in week 2.
The successful week-1 forecast of Ernesto (AL05) was illustrated in Fig. 1. The week-3 forecast (Fig. 5a) initiated on 12 July was even more successful with an LHV = 0.94 (Fig. 2). Even though ensemble storm 16 started 48 h prior to the pre-Ernesto starting time of 1200 UTC 1 August in the NHC WBT (Table 1b), the forecast location was good. In addition, the forecast track overlaid the Ernesto track and the forecast ending point was relatively close to the location for Ernesto, which accounts for the very high LHV for the week-3 forecast. By contrast, ensemble storm 24 in the week-4 forecast (Fig. 5b) had a starting position almost 15° longitude east of Ernesto, and although the first half of the forecast track was parallel to Ernesto’s track, the latter half was northwestward rather than westward. Thus, the LHV = 0.29 (Fig. 2) is marginal and would be in the below average category of Tsai et al. (2013).
As in Fig. 3, but for Hurricane Ernesto (AL05) for only forecasts initiated during (a) week 3 on 12 Jul and (b) week 4 on 9 Jul.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
Tropical Storm Florence (AL06) began just 24 h after Ernesto, but Florence formed 30° longitude to the east at 11.5°N, 17.8°W (Table 1b), which is just off the coast of Africa. The week-1 forecast for Florence (Fig. S3a in the supplemental material) initiated at 0000 UTC 30 July is noteworthy, as ensemble storm 3 had a perfect quality measure LHV = 1.0 (Fig. 2), which indicates that all four metrics that make up LHV were within tolerance. Ensemble storm 21 in the week-3 forecast (Fig. S3c), which initiated at 0000 UTC 16 July, had a parallel track relative to the Florence track, but the forecast starting time of 0000 UTC 5 August was 2.5 days late. Thus, the LHV for this storm forecast was 0.69. The first half of the ensemble storm 27 track in the week-4 forecast (Fig. S3d) that initiated at 0000 UTC 9 July was displaced to the north of the Florence track, but the second half overlaid the Florence track and ended at almost the correct time. Thus, the LHV of 0.82 was quite satisfactory.
It is somewhat unclear why the week-2 forecasts of Ernesto and Florence were unsuccessful. Elsberry et al. (2011) documented a similar case of a failure in week-1 forecasting a pretyphoon, but their proposed explanation of a large-scale circulation interaction with the multiple coexisting typhoons would not apply for Ernesto, which developed 1.5 months after TS Debby. As indicated above, Ernesto and Florence formed within 24 h, and TS Helene formed in a similar region and just 4.5 days after Florence (Table 1b). Thus, some degradation in the large-scale circulation in the week-2 forecast for Florence may have occurred.
Due to its significance, Superstorm Sandy (AL18) is selected for extended discussion as an EF3 storm in which the ECMWF 32-day ensemble week-4 forecast failed to predict its formation. As has been well publicized, the ECMWF deterministic (and 32-day ensemble; not shown) model provided excellent forecasts of the track after Sandy became well organized in the northern Caribbean Sea. In the week-1 forecast (Fig. 6a) initiated at 0000 UTC 18 October (3.5 days before the start of Sandy; see Table 1b), a large fraction of the ECMWF ensemble member vortices were indicating a storm would exist, but a large spread in tracks existed. Because of this track diversity, the weighted-mean vector motion was first westward and then slowly northward with almost no indication from the member vortices of the continued poleward track over the western Atlantic. In addition to the deduction in the LHV for the poor forecast ending point, the average position difference was also poor, which resulted in an LHV = 0.45 (Fig. 2).
As in Fig. 3, but for Hurricane Sandy (AL18) for forecasts initiated during (a) week 1 on 18 Oct, (b) week 2 on 8 Oct, (c) week 3 on 4 Oct, and (d) week 4 on 27 Sep.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
Elsberry et al. (2011) also found reduced forecastability by the ECMWF ensemble for late-season storms in the western North Pacific that had poleward tracks from the beginning. Uncertainty in the week-1 forecast starting position was due in part to the spread among ensemble member vortices over northern South America as well as over the Caribbean Sea. This uncertainty in initial positions also was a factor in the week-2 forecast (Fig. 6b) initiated at 0000 UTC 8 October, in which ensemble storm 19 actually begins over land.1 Again, the forecast of a slow translation to the north-northwest and a very poor forecast ending point led to an LHV = 0.54 (Fig. 2).
The week-3 forecast (Fig. 6c) initiated on 4 October also had an initial position near the coast and a poor forecast ending point. However, the two forecast track crossings and a small average position difference over the Caribbean Sea still resulted in an LHV = 0.70 (Fig. 2). The week-4 forecast (Fig. 6d) initiated at 0000 UTC 27 September did have a number of ensemble vortices in the region where Sandy formed on 21 October. However, these vortices occurred at the end of the ECMWF 32-day ensemble forecast and did not last the 3 days required to be designated as an ensemble storm. In summary, the ECMWF 32-day ensemble provided some advance notice of the formation and early poleward track across the Caribbean Sea, but the later (and more critical) track was not predicted in those forecasts initiated prior to the formation of Sandy, which justifies its classification as just somewhat forecastable.
Hurricane Isaac (AL09) was an EF3 storm that also made landfall in the United States. As indicated in Table 1b, pre-Isaac started just prior to TS Joyce at 0600 UTC 18 August near 12°N, 27°W. In the week-1 forecast (Fig. S4a available in the supplemental material) initiated at 0000 UTC 13 August, ensemble storm 3 starts at 1200 UTC 17 August, but is about 10° longitude to the west of pre-Isaac. The forecast westward translation of ensemble storm 3 is parallel to the track of Isaac until about 70°W, but its translation speed is much faster. When those ensemble member vortices leading to this rapid westward motion terminate over Central America, the weighted-mean vector motion then continues with the remaining vortex tracks that were now moving eastward after recurving, which leads to the strange eastward track prediction across southern Florida and the Bahamas (Fig. S4a). With the deductions for a poor forecast ending point, large average distance, and somewhat deficient forecast starting point, the quality measure LHV is only 0.25 (Fig. 2). Nevertheless, the early track to about 70°W might still be useful to a forecaster who would recognize the doubtful reality for the later portion of the track.
Although almost all of the path of ensemble storm 11 in the week-2 forecast (Fig. S4b) closely matches that of Isaac, the early (4.75 day) start and displaced starting position plus the poor forecast ending position leads to an LHV = 0.48. In the week-3 forecast (Fig. S4c), the starting time and position were excellent, the forecast track was quite good, but the forecast ending position was again not good. The resulting quality measure LHV = 0.75 does indicate useful information. In conjunction with the similar week-2 forecast track (Fig. S4b), the NHC forecaster might have an increased awareness of a potential landfall in the United States that will start from near the African coast in 2 weeks’ time.
Hurricane Nadine (AL14) was one of the longest-lasting Atlantic storms in recent times, and was a primary target of the NASA HS3 field experiment. Because the initial time of 0000 UTC 10 September for the week-1 forecast (Fig. S5a available in the supplemental material) coincided with the starting time of pre-Nadine, many of the ECMWF ensemble member vortices had an excellent initial position and early track to the recurvature point near 30°N, 50°W. However, the track of Nadine then became extremely complex and the ensemble forecast track spread became very large. Because of the later portion of the track, and that the forecast ending point was very different from that of Nadine, the week-1 LHV was only 0.64 (Fig. 2). In both the week-2 forecast (Fig. S5b) initiated at 0000 UTC 30 August and the week-3 forecast (Fig. S5c) initiated at 0000 UTC 23 August, the forecast starting position and early track were somewhat satisfactory, but the more westward track led to a forecast recurvature point that was about 15° longitude too far to the west. Consequently, the LHV was only 0.48 (Fig. 2) for the week-2 and -3 forecasts.
The inability of the ECMWF ensemble to forecast the formations and tracks in week 4 for Hurricanes Sandy, Isaac, and Nadine indicates that these are only somewhat forecastable EF3 storms. The late-season Sandy was strongly impacted by interactions with deep midlatitude troughs. Similarly, the extremely sharp recurvature and subsequent complex track indicate that Nadine was strongly impacted by midlatitude circulations, and these interactions are not forecastable on 4-week time scales. In the case of Isaac, it is only the LHV deductions for the forecast starting and ending points that make the week-2 and -3 forecasts appear less desirable. No explanation for the absence of a week-4 forecast of Isaac is available at this time.
3) Limited forecastable EF2 storms
Two hurricanes and one tropical storm were classified as EF2 storms (Table 1c), as the ECMWF 32-day ensemble was able to successfully forecast these storms in only the first two weeks (Fig. 2). Hurricane Rafael (AL17) was a late season storm that seemed to have its origins at low latitudes, but moved poleward after reaching tropical storm stage at 14.7°N, 62.7°W at 1800 UTC 12 October (Table 1c). Both TS Debby (AL4) and Hurricane Gordon (AL08) had their origins near 20°N and had short poleward tracks before recurving to the northeast.
The week-1 forecast (Fig. 7a) initiated at 0000 UTC 8 October had a number of ensemble member vortices that started near the first position (8.3°N, 37.5°W; Table 1c) in the NHC WBT for Raphael. Whereas the WBT has gaps [note that the North Atlantic hurricane database (HURDAT) only starts Raphael at 35 kt (1 kt = 0.51 m s−1) at 14.7°N, 62.7°W at 1800 UTC 12 October], the ensemble storm 2 track is continuous and parallel to the WBT until 60°W. Whereas ensemble storm 2 is forecast to sharply turn northward, this turn occurs about 16° longitude to the west of Raphael’s turn. Rather than continuing to move poleward, ensemble storm 2 is predicted to stall at ~25°N, and so the full deduction (0.22) for a poor forecast ending point is applied. In conjunction with a deduction for the average distance, an LHV = 0.63 (Fig. 2) is in the lower ranges of an above average quality. While the track of ensemble storm 7 in the week-2 forecast (Fig. 7b) initiated at 0000 UTC 10 October has a similar shape to Raphael’s track, the forecast starting point is far to the east at the coast of Africa. While one member vortex track extends to 40°N, the ensemble storm track 7 ends at 23°N, which contributes to the low LHV of 0.25 (Fig. 2).
As in Fig. 5, but for Hurricane Rafael (AL17) for forecasts initiated during (a) week 1 on 8 Oct and (b) week 2 on 1 Oct.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
The week-1 forecast track of ensemble storm 4 (Fig. 8a) was present in the initial conditions at 0000 UTC 13 August; this forecast did not have the correct location for the formation of pre-Gordon at 26.1°N, 54°W (Table 1c) 2.25 days later. Similarly, the sharp recurvature was forecast well. Because the forecast starting and ending points were somewhat off, what many forecasters might consider to be an excellent forecast is only rated at LHV = 0.55. In the week-2 forecast (Fig. 8b) initiated at 0000 UTC 6 August, ensemble storm 11 is predicted to start at 0000 UTC 11 August, which is 4 days before pre-Gordon actually started. Although ensemble storm 11 is made up of vortices that begin over a range of initial positions and have a variety of track types, the shape of the forecast track is quite similar to that of Gordon. Indeed, a shift of the forecast track about 6° latitude and longitude to the south and east would have resulted in a near-perfect track. Consequently, the LHV of 0.55 indicates a useful forecast of Gordon 2 weeks in advance. However, a storm such as Gordon beginning at 25°N involves both tropical and midlatitude influences, which are not forecastable beyond 2 weeks (EF2) in this case.
As in Fig. 5, but for Hurricane Gordon (AL08) for forecasts initiated during (a) week 1 on 13 Aug and (b) week 2 on 6 Aug.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
Pre–TS Debby (AL04) began at 1200 UTC 22 June at 22.6°N, 88.1°W (Table 1c) just north of the Yucatan Peninsula (Fig. S6 available in the supplemental material) and moved slowly northward before turning eastward and moving eastward across northern Florida. Because the week-1 forecast was initiated just 36 h prior to the start of pre-Debby, a large number of member vortices were predicting the system, but with at least two track types: one group with a track quite similar to Debby’s track and another group with a larger number of vortices predicting a slow northwestward or westward motion. Thus, the weighted-mean vector motion was a slow northwestward track toward the Gulf of Mexico coast near the border of Texas and Louisiana before turning northeastward. Given the slow motion of Debby and the early forecast track, the average distance plus the excellent forecast starting position still led to an LHV = 0.48.
Ensemble storm 4 in the week-2 forecast (Fig. S6b) initiated at 0000 UTC 14 June is made up of two types of vortex tracks. The first set of tracks again begins over land in Colombia and moves northwestward along the Central America to the Yucatan Peninsula and the east coast of Mexico. This first set of tracks is then joined by five tracks that are more representative of the formation and track of Debby. Indeed, one vortex track forecast has almost the exact path of Debby. Had the second set of tracks been considered separately, that ensemble storm would have been rated highly. However, the evaluation of ensemble storm 4 is only rated at LHV = 0.25 due to its unrepresentative forecast starting and ending points and poor average distance.
These three EF2 storms (Table 1c) represent the limited forecastability of the ECMWF ensemble for somewhat different reasons. The low-latitude origin and poleward track of the late season Rafael has some of the same difficulties as Sandy due to interactions with a deep midlatitude trough. The importance of interaction with the midlatitudes for the formation track and intensification of Hurricane Gordon can account for its limited forecastability. Finally, Debby’s formation adjacent to the Yucatan Peninsula and its short poleward track before recurving may indicate limited predictability due to land and ocean interaction as well as midlatitude interaction.
4) Not forecastable EF0 storms
The most surprising result of this evaluation of the 2012 Atlantic season was that the ECMWF 32-day ensemble forecasts on Mondays or Thursdays did not predict even in week 1 the formations and tracks of five TCs (Table 1d). That is, the LHVs of three tropical storms [Helene (AL07), Patty (AL16), and Tony (AL19)] and two hurricanes [Chris (AL03) and Michael (AL13)] were equal to zero for all four weeks (Fig. 2). The tracks of these five storms from the NHC WBT are shown in Fig. 9.
WBTs for five Atlantic TSs and hurricanes in the not-forecastable EF0 category (see names and characteristics in Table 1d).
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
In the case of TS Helene (AL07), the first entry in the WBT was at 0000 UTC 7 August at 11.7°N, 26.6°W (Table 1d). This was a marginal tropical depression or even a 20-kt tropical disturbance along its long path across the tropical Atlantic, Caribbean Sea, Central America, and the Yucatan Peninsula. It was only for a 12-h period over the Bay of Campeche that Helene was classified as a tropical storm. It was during this period that the Monday or Thursday ECMWF 32-day ensemble forecasts failed to predict Helene. Pre–TS Patty (AL16) started at 0000 UTC 11 October near 25.8°N, 72.7°W at the southern end of a long, north-oriented cloud band from the midlatitudes. Patty remained almost stationary (Fig. 9) and reached maximum intensity 24 h after formation. Patty had a compact structure that appeared to evolve from mesoscale processes, which (in addition to its short life) may explain the failure of a Monday or a Thursday ECMWF ensemble to predict its formation and track. Pre–TS Tony was a late season storm (21 October; Table 1d) that began relatively far north (19.8°N, 50.0°W; Fig. 9) and initially moved poleward and then recurved quickly. Tony (AL19) achieved tropical storm intensity at 26.4°N, 49.8°W, as it was moving into the midlatitude westerlies. As the intensification of Tony was likely influenced by mixed tropical and baroclinic processes, and given its short period as a tropical storm, it is perhaps not that surprising that Tony was not predicted as a week-1 storm in either a Monday or Thursday ECMWF 32-day ensemble forecast.
The more serious failures were the two hurricanes that the ECMWF ensemble failed to predict in advance of their formation time. Hurricane Chris (AL03) was an early season storm that started at 0000 UTC 17 June at 28.8°N, 68.8°W (Table 1d) and immediately moved poleward (Fig. 9). Since Chris only attained TS status at 1800 UTC 19 June when it was at 39.5°N, 58.0°W and moving eastward, baroclinic processes must have been a contributing factor to its intensification. While Hurricane Michael (AL13) was a midseason storm starting 1800 UTC 2 September (Table 1d), it was not a typical African wave–type development. Rather, its starting position was at 26.4°N, 40.1°W and it had a meandering poleward track (Fig. 9). Given this starting location and the implied weak steering, the formation and intensification of this storm was likely strongly influenced by baroclinic processes, which may account for Michael not being forecast in advance.
The tentative conclusion from four of these five EF0 storms that began in the central Atlantic is that strong midlatitude influences make their formation locations and timing not forecastable by the twice-weekly ECMWF 32-day ensemble. Diagnostic studies performed to demonstrate the types and magnitudes of the interactions of the midlatitude circulations with the pre-TCs in the ECMWF ensemble fields are beyond the scope of this study.
4. Characteristic false alarm tracks
As in Tsai et al. (2013), the false alarm tracks are defined as those ensemble storm tracks that could not be matched with any of the NHC WBTs within the specified tolerances. A substantial fraction (35%) of all false alarms occurs during the first week and often are in the initial conditions, which was also the case in the western North Pacific (Tsai et al. 2013). Of course, the NHC forecaster would be aware that no potential tropical cyclone seedlings existed in the area by T + 15 h when the ECMWF ensemble storms tracks are received and processed. Since the ECMWF ensemble tends to begin the storms 2–3 days (or more) before NHC begins its WBT, this may indicate the need for a modified vortex tracker definition. Once the ECMWF ensemble begins to predict a false alarm storm, it tends to sustain that storm in the next 32-day forecast. So if an ensemble storm in the Monday forecast that may be associated with the same ensemble storm in last Thursday’s forecast has still not shown signs of developing after 4.5 days, it is likely a false storm.
A more straightforward detection of Atlantic false alarm storms is simply from the geographical distribution of their formation locations and tracks (Fig. 10). Type 2 false alarms that predominantly begin inland over the United States or near the Gulf of Mexico coast are easily recognized. The most frequent (43 of 107) type 3 false alarm tracks begin near the northern coast of South America and along the coast of Central America, or move inland over these regions. The least frequent (10 of 107) type 4 false alarms begin at or north of 20°N and mainly have northward tracks. While these may be representative of baroclinically forced systems, the result in section 3b(4) that the ECMWF ensemble did not accurately predict the location and subsequent track of any such storms even 1 week in advance would lead to a presumption that such a track forecast is highly likely to be a false alarm.
Four types and numbers (see inset) of false alarm ensemble storm tracks starting during the first week of the twice-weekly ECMWF 32-day forecasts between 31 May and 17 Dec 2012 (total of 58 forecasts). The starting location is indicated by a circle.
Citation: Weather and Forecasting 29, 2; 10.1175/WAF-D-13-00104.1
The type 1 (22 of 107) false alarm tracks in Fig. 10 that begin near the west coast of Africa will not be easily recognizable—especially those that have long, recurving tracks that resemble climatological tracks in the Atlantic. The shorter forecast tracks might be suspect because limited time would be available to form a hurricane, but some real tropical storms do have short tracks. Similarly, the tracks that begin inland over Africa might be suspect, but the ECMWF ensemble tends to begin storms 2–3 days early and may be detecting vigorous African waves that do have a chance to develop. While the African wave–type tracks were shown in section 3 to be the more forecastable, those ensemble storms that verified well did not include all 51 ensemble members. Indeed, the fraction of the 51 ensemble members may be thought of as a raw probability of occurrence. Especially for the type 1 false alarms, their numbers compared to the hits reflect some measure of the reliability of the ECMWF ensemble for African wave–type systems. Of course, a multiseason sample of Atlantic storms would be required to statistically establish the reliability.
The four types of false alarm tracks in week 1 continue to be present in weeks 2–4 with totals of 70, 72, and 54, respectively (Fig. S7 available in the supplemental material). The numbers of type 2 (over the United States) and type 3 (South America and Central America) false alarms are considerably smaller in week 2 (Fig. S7b) than in week 1, but the tracks are similar. Note that the number of type 1 (African wave) false alarms in week 2 is almost identical to that of week 1, with 21 such false alarms over the 58 forecasts in this sample. The type 2 and 3 false alarm frequencies in week 3 (Fig. S7c) are the same as in week 2, albeit with shorter tracks. No decrease in the number of type 1 false alarms during week 3 is noted. While the overall number of false alarms is smaller during week 4 (Fig. S7d), only a small fraction of the reduction is attributable to the type 1 (African wave) false alarms. In summary, this consistent pattern of false alarm tracks that may be easily recognized and removed (except for type 1) will make the ECMWF ensemble storm forecasts more acceptable to forecasters.
5. Summary and future work
This evaluation of the performance of the ECMWF 32-day ensemble predictions of Atlantic TC events (formation and tracks) was made possible by the ECMWF kindly providing their twice-weekly forecasts in near–real time in support of extended-range mission planning for the NASA HS3 field experiment. Although the ECMWF also provided daily 15-day ensemble forecasts that could be used for medium-range forecasting of existing TCs, the focus in this evaluation was on the number of weeks in advance of the beginning of NHC storms AL 03–AL 19 that the ECMWF 32-day ensemble predictions were successful in forecasting their existence. A single season is not adequate to make statements about the predictability of Atlantic TCs. Rather, this evaluation is simply a description of the forecastability of the AL 03–AL 19 tropical storms and hurricanes with the ECMWF 32-day ensemble using essentially the same technique as in Elsberry et al. (2010, 2011) to form ensemble storm tracks that could be verified with the objective track analog technique of Tsai et al. (2013).
An important distinction is made here between forecasts of TC formation (or genesis) alone versus both the formation and the subsequent track, which are also represented in the LHV quality measure by the average track difference, shortest track difference, and the forecast ending distance. Given a multiyear sample of the formation locations from these twice-weekly 32-day ensemble forecasts, a comparison could be made with the climatological distribution of formations. Similarly, a multiyear sample of tracks could be compared with the climatological track distribution. At this time, only the 2012 Atlantic season is available and such comparisons with climatology are not possible.
Whereas Elsberry et al. (2010, 2011) and Tsai et al. (2013) had found that the ECMWF 32-day ensemble could predict out to 4 weeks most of the typhoons, tropical storms, and even some tropical depressions in the western North Pacific, this was not the case during the 2012 Atlantic season. Four categories of Atlantic storms have been tentatively classified as to whether they were forecastable on the extended-range time scale: 1) EF4 storms in all four weeks in advance of the starting date in the NHC WBT, 2) EF3 storms in three of the four weeks, 3) EF2 storms limited to the first two weeks, and 4) EF0 storms that were not forecast 1–7 days in advance of any Monday or Thursday ECMWF 32-day ensemble forecast. Two hurricanes [Kirk (AL11) and Leslie (AL12)] and two tropical storms [Joyce (AL10) and Oscar (AL15)] were successfully forecast in all four weeks. The common characteristics of these EF4 storms were that they generally occurred in midseason from African wave type formations and the hurricanes had long recurving tracks that may have been more controlled by the subtropical high than by an interaction with an active midlatitude trough. In many of the African wave–type formations, the ECMWF ensemble has a starting point at the coast rather than downstream, which resulted in an LHV deduction of 0.22. Especially for the TSs with shorter tracks, the ECMWF ensemble tracks were too long, which also resulted in an LHV deduction.
The somewhat successful EF3 events of Hurricane Ernesto (AL05) and TS Florence (AL06) were not forecast in week 2, which was surprising since they had African wave origins with westward tracks. The explanation for the week-2 failure may be related to these two storms and TS Helene (AL07) having started in the same area within 7 days. That is, multiple interactions among the three storms in the ECMWF ensemble prediction during that week may have led to degradations in the forecasts of the large-scale circulation over the Atlantic during the second week such that formations of AL05 and AL06 were not forecast in that week. Strong interactions with midlatitude circulations affecting the formations and tracks of the late-season Sandy (AL18) and the long-lasting Hurricane Nadine (AL14) are considered to be the reason limiting their forecastability to weeks 1–3. The tendency for the midseason TCs to recurve over the Atlantic in nature and in the ECMWF ensemble may be the explanation for the lack of a match in week 4 for the long westward track of Hurricane Isaac.
Strong interactions with the midlatitude circulation during and following the formation at low latitudes of the late-season Hurricane Rafael (AL17) are considered to be the factor limiting forecastability beyond 2 weeks (EF2). Similarly, the strong midlatitude influence on the formation and track in the subtropics of Hurricane Gordon (AL08) was a factor in its limited forecastability. TS Debby (AL04) is a special case as it formed early in the season in the Gulf of Mexico near the Yucatan Peninsula coast, and did not have a precursor African wave that could be forecast for weeks in advance. Thus, the timing and location of the formation of Debby may have involved mesoscale processes and land–ocean processes that are not forecastable beyond a few days.
Among the five not forecastable EF0 storms, TS Helene (AL07) near the east coast of Mexico and TS Tony (AL19) in the eastern Atlantic were weak tropical storms for a short period of their lifetime. Similarly, TS Patty developed for a short time in a limited region of favorable conditions. Elsberry et al. (2010, 2011) reasoned that the success in extended-range prediction for typhoons could be attributed to favorable environmental conditions that had to exist not only at the time of formation, but also over a large area to allow time for intensification to a typhoon, and that such large-scale circulation conditions were forecastable in the western North Pacific. By contrast, favorable environmental conditions for formation plus intensification to hurricanes did not exist for these three tropical storms. For the early season Hurricane Chris (AL03) that formed so far north, the ECMWF vortex tracker definition that distinguishes between tropical cyclones versus extratropical cyclones may have been a contributing factor. However, it seems more likely that the mixed-phase physics of baroclinic and convective processes may not be well predicted in the ECMWF ensemble model, at least beyond a few days. The failure in forecasting the formation in the subtropics of the long-lasting Hurricane Michael (AL13) may well be due to the treatment of baroclinic and convective processes in the ECMWF ensemble. While the week-1 forecast of Hurricane Gordon, which has some similarities to Hurricane Michael, was successful, this success may have been due to an African wave precursor (Fig. 8a).
Although based on a small sample, this evaluation has demonstrated a more limited extended-range forecastability of TC events in the Atlantic than in the western North Pacific (Elsberry et al. 2010, 2011; Tsai et al. 2013). It seems reasonable that the most forecastable TC events are those long-lasting that occur in midseason from well-defined African waves. The limited forecastability beyond weeks 3 (EF3) or 2 (EF2) may be due to less strong tropical forcing (e.g., Madden–Julian oscillation), the absence of a large-scale monsoonal influence as in the western North Pacific, or the larger midlatitude impacts on TC formation, intensification, and tracks in the Atlantic. A larger sample of Atlantic TC forecasts on the extended-range time scale will be required to attribute these causes of limited forecastability. It might be said that the four EF0 TC events that occurred in the mid-Atlantic are not that important because they posed no threat to the United States or Caribbean islands. However, they included two hurricanes that were a maritime threat. Furthermore, these systems interact with and modify the large-scale environment (consider the long-lasting Hurricane Nadine), which may affect the forecastability of a subsequent TC. A larger sample of Atlantic TC events is required to explore the source(s) of the EF0 failures.
This study has demonstrated that the ensemble storm generation techniques of Elsberry et al. (2010, 2011) and the track analog verification technique of Tsai et al. (2013) can be applied for the ECMWF 32-day ensemble forecasts in the Atlantic. One of the advantages of this technique relative to the more common studies of TC activity is the capability to associate the successes (hits) and failures (misses) to specific TC events. A second advantage is the capability to study the false alarms [defined as in Tsai et al. (2013) as all ensemble storms that could not be matched with an official track]. Understanding these false alarms will be an important second step in developing an operational version of the technique for the extended-range prediction of TC events in the Atlantic.
As suggested by a reviewer, it would be helpful to the forecasters to know how the number of ensemble members that are forecasting a formation is related to an actual formation. That is, if 40 of the members are predicting a formation versus only 20 members, can the forecaster be twice as confident that a formation will occur? While it is generally true that the larger the number of ensemble members the more reliable are the ECMWF ensemble forecasts, there have been notable exceptions. Additional seasons need to be examined to document the reliability. A systematic study is also needed of the contributions of the twice-daily ECMWF 15-ensemble forecasts in adding confidence in week 2 to those events that were predicted in weeks 3 and 4 of the prior ECMWF 32-day forecasts. That is, does week-to-week continuity in events lead to higher and higher confidence?
Acknowledgments
Dr. Hsiao-Chung Tsai is a National Research Council postdoctoral research associate. He and coauthors Russell Elsberry and Mary Jordan are supported with funding from the Marine Meteorology section of the Office of Naval Research. Dr. Frederic Vitart of the ECMWF has been instrumental in the provision of real-time access to the ensemble predictions. Dr. Jack Beven of NHC provided some insights as to physical processes involved in Hurricane Rafael and Sandy. Mrs. Penny Jones provided excellent manuscript preparation support.
APPENDIX
A notable consideration for this evaluation of the 2012 ECMWF ensemble forecasts relative to the earlier studies has been an improvement in the treatment of convection in the tropics. In addition, higher horizontal resolution of T639 (32 km) for the first 10 days of the forecast and T319 (65 km) from day 10 onward was implemented in 2010 and an ensemble-based data assimilation technique was introduced in 2010. Lang et al. (2012) had demonstrated the impacts of various perturbation methods on TC track predictions. Thus, it might be expected that the 2012 track forecasts in the Atlantic have been improved compared to earlier years.
A major advancement prior to the 2012 season was that the ECMWF 32-day ensemble was available twice a week (Mondays and Thursdays from 0000 UTC initial conditions) rather than only once a week (Thursday). As in Elsberry et al. (2010, 2011), the focus is on both the formation and the track of the Atlantic tropical storms and hurricanes. It is well known that the “mother” ECMWF deterministic model is quite skillful in predicting the TC track after the TC warning center has named the storm, and this is also true of the ECMWF ensemble when as many as 51 members may be representing the storm. Similar to the experience in the western North Pacific, the ECMWF ensemble predicts vortices in the Atlantic that begin before the NHC begins tracking the corresponding system in its “working best track” (WBT). Thus, the combination of similar vortices into an ensemble storm, which by definition must contain at least three vortices and last at least 3 days, also begins before the first position in the NHC WBT.
The basic concept of the track-matching approach developed by Tsai et al. (2013) is to select an ensemble storm as a potential analog to an observed track. The main focus in this procedure is to match the overall track rather than the point-to-point comparisons in Elsberry et al. (2010, 2011). Given a NHC storm track, the ensemble storm tracks for the season are searched to find all ensemble storms within the allowable time difference at any time along the NHC track: ±3 days for week 1, ±4 days for week 2, and ±5 days for week 3 and 4 forecasts. For ensemble storms within specified time differences of the NHC track that are potential analogs, an objective quality measure that assesses the overall track similarity between the potential analogs and each observed track is calculated in terms of membership functions for the four metrics: shortest distance, average distance of the matched points, distance at formation time, and distance at ending time. Membership functions and weighting factors multiplying these membership functions are used and adjusted to match the subjectively determined quality measures for the ECMWF ensemble storm forecasts during the 2009 season (Elsberry et al. 2011). The largest weighting factor of 0.3 is assigned to the average distance to emphasize the similarities with the overall observed track. The next largest weighting factor (0.25) is given to the shortest distance, which in combination with the average distance will favor the ensemble storm that most nearly overlaps with the NHC storm track. Finally, weighting factors of 0.23 and 0.22 are used to represent the differences in the formation and ending positions.
Five quality measures of excellent, above average, good, below average, and poor are then assigned as a linear function with an LHV interval of 0.2, with an LHV < 0.2 not considered to be a real match. The same membership functions and weighting factors as in Tsai et al. (2013) are used with the following exceptions: (i) The metric of the distance at the ending time is only applied to the end of the 32-day ensemble forecast. If an ensemble storm track ends at 32 days but the NHC storm track extends beyond that date, the ending distance is calculated at day 32. (ii) An additional screening step for the distance at formation time is applied, which requires the difference be less than 2220 km.
REFERENCES
Belanger, J. I., Curry J. A. , and Webster P. J. , 2010: Predictability of North Atlantic tropical cyclones on intraseasonal time scales. Mon. Wea. Rev., 138, 4362–4374.
Belanger, J. I., Webster P. J. , Curry J. A. , and Jelinek M. T. , 2012: Extended prediction of north Indian Ocean tropical cyclones. Wea. Forecasting, 27, 757–769.
Dupont, T., Plu M. , Caroff P. , and Faure G. , 2011: Verification of ensemble-based uncertainty circles around tropical cyclone track forecasts. Wea. Forecasting, 26, 664–676.
Elsberry, R. L., 2010: Critical time requirements for operational use of deterministic and ensemble tropical cyclone track forecasts. Asia-Pac. J. Atmos. Sci., 46, 217–220.
Elsberry, R. L., Hughes J. R. , and Boothe M. A. , 2008: Weighted position and motion vector consensus of tropical cyclone track prediction in the western North Pacific. Mon. Wea. Rev., 136, 2478–2487.
Elsberry, R. L., Jordan M. S. , and Vitart F. , 2010: Predictability of tropical cyclone events on intraseasonal timescales with the ECMWF monthly forecast model. Asia-Pac. J. Atmos. Sci., 46, 135–153.
Elsberry, R. L., Jordan M. S. , and Vitart F. , 2011: Evaluation of the ECMWF 32-day ensemble predictions during 2009 season of western North Pacific tropical cyclone events on intraseasonal timescales. Asia-Pac. J. Atmos. Sci., 47, 305–318.
Lang, S. T. K., Leutbecher M. , and Jones S. C. , 2012: Impact of perturbation methods in the ECMWF Ensemble Prediction System on tropical cyclone forecasts. Quart. J. Roy. Meteor. Soc., 138, 2030–2046.
Majumdar, S. J., and Finocchio P. M. , 2010: On the ability of global ensemble prediction systems to predict tropical cyclone track probabilities. Wea. Forecasting, 25, 659–680.
Tsai, H.-C., Elsberry R. L. , Jordan M. S. , and Vitart F. , 2013: Objective verifications and false alarm analyses of western North Pacific tropical cyclone event forecasts by the ECMWF 32-day ensemble. Asia-Pac. J. Atmos. Sci.,49, 409–420, doi:10.1007/s13143-013-0038-6.
van der Grijin, G., Paulsen J.-E. , Lalaurette F. , and Leutbecher M. , 2005: Early medium-range forecasts of tropical cyclones. ECMWF Newsletter, No. 102, ECMWF, Reading, United Kingdom, 7–14.
Vitart, F., 2009: Impact of the Madden-Julian oscillation on tropical storms and risk of landfall in the ECMWF forecast system. Geophys. Res. Lett., 36, L15802, doi:10.1029/2009GL035089.
Vitart, F., Leroy A. , and Wheeler M. C. , 2010: A comparison of dynamical and statistical predictions of weekly tropical cyclone activity in the Southern Hemisphere. Mon. Wea. Rev., 138, 3671–3682.
Vitart, F., Prates F. , Bonet A. , and Sahin C. , 2012: New tropical cyclone products on the web. ECMWF Newsletter, No. 130, ECMWF, Reading, United Kingdom, 17–23.
Yamaguchi, M., Nakazawa T. , and Hoshino S. , 2012: On the relative benefits of a multi-centre grand ensemble for tropical cyclone track prediction in the western North Pacific. Quart. J. Roy. Meteor. Soc., 138, 2019–2029, doi:10.1002/qj.1937.
Dr. J. Beven of the NHC (2013, personal communication) indicates that the global models often develop vorticity centers downstream of the Guajira Peninsula of Colombia.
