An ongoing and important task of the Environmental Modeling Center (EMC), as an integral part of the National Centers for Environmental Prediction (NCEP), is to provide increasingly skillful numerical guidance to meet the various national and international demands for improved weather forecasts. Toward this end, a concerted effort at the EMC is directed at developing, testing, and evaluating potential upgrades to the global analysis–forecast system as a part of the worldwide effort of all major forecast centers to ensure the steady advancement in operational numerical weather prediction.
This effort involves an ongoing development of more sophisticated methodologies for data assimilation techniques to best maximize the use of observations that are now routinely being made available by both regional and global state-of-the-art observing platforms, as well as to continue a high degree of concentration on improving model representation of physics. Commensurate with this effort is a continual critical assessment of theimpact of all potential upgrades on model performance for applications over the extratropics and Tropics ranging from the larger-scale forecasts to forecasts involving smaller-scale phenomena over more specific areas of interest.
A number of papers documenting the Medium-Range Forecast Model (MRF) analysis–forecast system upgrades, assessing the overall model performance, and evaluating model behavior in terms of systematic error studies has been carried out by Kanamitsu (1989), Kanamitsu et al. (1991), Derber et al. (1991), Caplan and White (1989), White and Caplan (1991), and Caplan et al. (19931997).
During the 1994 hurricane season several systematic deficiencies in the Aviation (AVN) tropical forecasts were identified by Surgi (1994). Some of the more prominent errors in the hurricane forecasts included a left or westward track bias that was particularly conspicuous for recurving storms such as Hurricanes Chris and Florence. It was also noted that the forecasts consistently displayed a vertical decoupling of the storm vortex after 24 h that led to a serious degradation of the forecasts with time. The vortex decoupling was found to be related to the failure of the model to maintain a vigorous and deep circulation. In the larger-scale flow, a mid- to upper-level easterly bias, which was previouslyidentified by White and Caplan (1991) for the global Tropics, was shown to be large over the Caribbean. This proved to be problematic for intensity forecasts since the prediction of easterly flow is favorable to storm genesis and intensification. Fitzpatrick et al. (1995) explored this bias in relation to a deficiency in the AVN to properly maintain the tropical upper-level midoceanic trough and associated westerly flow.
During October 1995 an improved version of the NCEP global model was implemented operationally. This occurred after extensive testing and evaluation of the various proposed changes to the global analysis–forecast system and after assessing their impact on the overall predictive skill of the model. The National Hurricane Center (NHC) in collaboration with EMC actively participated in this assessment. The performance of the model was evaluated with relation to the model forecasts in the Tropics, and more specifically, the hurricane forecasts, to determine if the proposed upgrades made a positive impact to reduce any of the above cited model deficiencies. The NHC relies on NCEP for numerical guidance support as an integral part of the overall hurricane forecasting process at NHC. This guidance is provided by the AVN 72-h hurricane track forecasts as described by Lord (1991). Additionally, the overall quality of the global model system is extremely important since it provides the initial fields for other hurricane forecast models used by NHC—that is, the GFDL Hurricane Prediction System, NHC90, VICBAR, and the BAM models. A list of acronyms defining these models in given in Table 1 and further details on model characteristics are provided in Aberson and DeMaria (1994). Thus, this study was quite relevant to the NHC’s forecast interests.
The 1995 MRF upgrades are described in section 2, as is the methodology used for this study. The results showing the impact of the upgrades on the large-scale tropical flow and for the hurricane forecasts are described in section 3. The conclusions are presented in section 4.
2. Design of model evaluation
a. Changes to the MRF
The changes that were implemented in the model that are most relevant to this study were made to the modelphysics in the boundary layer and in the deep convection schemes. In the parameterization of deep convection, the simplified Arakawa–Schubert scheme of Pan and Wu (1994) has been further modified by a change to the closure assumption, which has led to an enhancement in the convective heating in disturbed regions and a decrease in convective heating in suppressed regions. The original closure of the scheme is an adjustment of the cloud work function to a set of climatological values [cloud depth dependent, calculated from a variety of observations as described by Lord and Arakawa (1980)] over a timescale of 1 h. The cloud work function is a measure of conditional instability of a column of air. Conditional instablitiy characterizes the atmospheric stability over the tropical oceans so that a lack of favorable forcing such as large-scale subsidence can inhibit convection and prevent the model atmosphere from returning to neutral conditions. Alternatively, in disturbed regions, where strong convergence and rising motion often exist, when the convection scheme adjusts the soundings to the climatological cloud work functions, the atmosphere will remain conditionally unstable even when the parameterized heating stops. This can lead to large-scale supersaturation in the forecast and the heating will primarily be in the lower troposphere. The new closure change allows the atmosphere to adjust to more neutral conditions when the cloud base rising motion gets stronger. This change has generally led to enhanced gradients of convective precipitation that results from larger maxima due to the adjustment toward more neutral conditions.
The changes in the planetary boundary layer (PBL) diffusion scheme has been reported in Hong and Pan (1996) and will be only briefly described here. The major change is the use of a bulk Richardson number of the entire boundary layer to specify the coefficients of diffusivity in place of a local Richardson number. While the scheme was designed to improve the diurnally varying PBL over land, it also improves the maintenance of the oceanic PBL. One important result of the new scheme is the stronger mixing of the moisture in the PBL as described in Hong and Pan (1996) resulting in weaker vertical moisture gradients in the PBL. The effect of this change is the replenishment of the PBL moisture after a convective event or the restoration of the conditional instability. This has had a significant influence on the convection parameterization in the model by increasing the moisture supply to the free atmosphere. In fact, the two changes to the deep convection and the boundary layer have contributed nearly equally in magnitude in the enhancement of the surface latent heat fluxes and the deep convective heating.
Also, in addition to the above-mentioned changes, the statistical interpolation system (SSI) analysis scheme was modified to directly incorporate the use of satellite radiance data. This change was shown to have a very positive impact on the forecasts in the extratropics (most notably in the Southern Hemisphere) as described byDerber and Wu (1997). This study does not include forecast sensitivity to this change since all the forecasts presented here were rerun from the new global data assimilation system (GDAS) with the incorporation of the radiance data.
b. Design of experiments
During the summer of 1995, for approximately two months (6 June–28 July), two versions of the global model were run in parallel. For the purpose of this study, the operational model—that is, without the 1995 changes—will be referred to as the MRF, while the model version including the upgrades will be referred to as the MRX.
As a preliminary study, these parallel runs were evaluated daily during this period with a general emphasis on examining the flow over the tropical Atlantic and eastern Pacific basins. The preliminary findings suggested that the MRX was a “moister” model than the operational MRF with a more realistic distribution of precipitation and moisture near the intertropical convergence zone (ITCZ) as qualitatively determined from satellite imagery. Also, the tropical convective systems appeared to be more sharply defined in the MRX with more concentrated regions of vorticity and convective precipitation.
Based on those encouraging findings, the MRF and the MRX were rerun for the months of August and September 1995 at 0000 UTC to determine what impact the new physics might have on the hurricane forecasts and on improving some model deficiencies mentioned above as noted for the 1994 hurricane season. The forecast fields from these runs were archived at a 1° horizontal resolution. Out of the six storms that formed during this period, several case studies were chosen to closely examine the details of the forecasts. The cases selected were those where the MRX had clearly shown an improvement over the MRF forecast. Moreover, the forecast scenarios that we selected were ones that proved difficult for other models to forecast, for example, NHC90, BAMS, VICBAR, and the GFDL model.
3. Impact of model upgrades—MRF versus MRX
a. Large-scale tropical forecasts
To provide an objective basis to evaluate the overall performance of the MRF and the MRX, root-mean-square (rms) errors for an ensemble of 72-h forecasts were calculated over a two-week period in August and for another two-week period in September. These calculations were carried out for the tropical and subtropical Atlantic and eastern Pacific Oceans (10°–35°N, 20°–140°W, Fig. 1a), for a smaller area over the tropical Atlantic including the Caribbean and eastern Gulf of Mexico (Fig. 1b) between 90° and 50°W (where the upper-level wind errors have been systemically large)and for the entire Northern Hemisphere. Predictive skill is measured relative to the analyses. Results are presented for August only since there was quantitively little difference between the two periods. Figures 1a and 1b show the results of these calculations. The solid lines are the results from the MRF; the dashed lines are from the MRX.
Figure 1a shows the 200-mb rms vector wind errors (top plots) and zonal wind errors (bottom plots) between 140° and 20°W. The vector wind errors show an overall modest improvement in the MRX 72-h forecasts as indicated by the area and time mean value (area averaged over this domain over the 14-day ensembles) of 8.86 ms−1 errors compared to 9.25 m s−1 from the MRF averaged. Although the errors from both runs show the same trend over certain days in the first half of the forecast period, the variability of the 72-h error is somewhat reduced in the MRX over the second half of the forecast period. The variability of the 72-h rms error for the 500-mb temperature is also reduced in the MRX but over the entire forecast period (not shown). The contribution of the 72-h zonal wind error at 200 mb (bottom plots) is similar to the total vector wind error. There is a modest improvement in the MRX average rms error values; however, a decrease in the variability of the error is shown throughout the forecast period. Figure 1c shows the result of this calculation for the vector wind over the Caribbean including the Gulf of Mexico. Although the magnitudes of the errors are larger in this region than we would like to see in both the MRF and the MRX, the mean rms error of the MRX is reduced to 9.30 m s−1 from 10.09 m s−1 from the MRF.
To assess the broader impact of the upgrades on the predictive skill of the MRX, from the same ensemble of 72-h forecasts, the rms error of the 500-mb heights were calculated for the Northern Hemisphere (not shown). In general, there is a moderate overall improvement in the MRX 72-h error and the area and time average errors are reduced in the MRX to 22.14 gpm from 22.82 gpm in the MRF, which are consistent with the above results. The details of the error statistics for the global and extratropical forecasts are described in Pan et al. (1997).
Although the objective errors were encouraging for the upper-level winds, disappointing results were obtained for the 850-mb tropical winds. At low levels, the MRX performed worse at 72 h than the MRF over the entire Tropics. The low-level tropical wind error is one that has long plagued the NCEP global model, and other global models (Surgi 1989). One possible source for this model bias on larger scales of motion could be related to an improper treatment of the shallow nonprecipitating clouds that form along the trade wind inversion that are important transporters of heat, moisture, and momentum into the deep Tropics. The low-level wind error is an area of active investigation.
b. Tropical storm/hurricane forecasts
The “hurricane forecasting problem” to date remains one of the most difficult and challenging forecast problems in numerical weather prediction. This is true for track forecasting in spite of a significant increase in numerical forecast guidance skill over the past two decades and most conspicuously for intensity forecasts that have shown little or no skill in our operational model forecasts (DeMaria and Kaplan 1997). One of the reasons for the limited progress in numerical prediction of tropical cyclones is due to a fundamental lack of understanding of the details of the scale interaction problem governing these systems in the real atmosphere.Largely, this is due to a historical lack of consistent data over the tropical oceans that has limited our progress in providing an adequate description of the details of the physical and dynamical interactions within the tropical systems and with the larger-scale environment for a variety of storm situations that impact storm motion and intensity. While a detailed treatment of the numerical prediction problem is well beyond the scope of this paper, Krishnamurti et al. (1993) provide insight into the recent advances of hurricane forecasting with very high resolution models; and Elsberry (1995) provides a comprehensive description on the more general role of the dynamical modeling effort in the advancement of tropical storm prediction.
The NHC makes use of a hierarchy of models ranging from purely statistical models—that is, CLIPER (climatology and persistence) to mixed statistical-dynamical models, NHC90 to barotropic models, VICBAR to multilevel fully dynamical models with state-of-the-art physics, and the AVN and the GFDL model. The reason for maintaining a number of models operationally is purely a pragmatic one: some models perform better under some situations than others. Hurricanes interacting with midlatitude systems are usually handled better by the baroclinic models as are storms near land where real-time data can be optimally used. Aberson and DeMaria (1994) provide an assessment of VICBAR performance and describe scenarios in which barotropic dynamics are adequate for accurate track predictions. Most dynamical models perform better for well-developed rather than weaker storms and most models perform better with well-developed storms and strong rather than weak steering currents. The reasons why one model outperforms another for a given storm are, however, not always that clear. And less clear are the inconsistencies that seem to randomly occur in the predictions between one model forecast period and the next, perhaps partly reflective of the various states of predictability of the larger-scale atmosphere. (Certainly the hurricane forecasters who rely on model forecasts for guidance every 12 h are keenly sensitive to this problem.)
Trying to isolate and determine exact causes for particular “failed” or “successful” forecasts in a global dynamical model is a very difficult and daunting task. It is difficult due to the strong nonlinear interactions between the boundary layer and deep convection parameterization schemes and between these physics and dynamics of the model. This is true in trying to isolate causes for model errors in general. Determining the best method in which to identify the source of the errors in another important aspect of this problem. A recent contribution to the study of both mean and transient systematic model errors is described in Kanamitsu and Saha (1995). Moreover, this difficulty escalates tremendously for the hurricane forecast problem due to all the complexities discussed above. In this study, we make use of our knowledge of the systematic model biases andmodel behavior in particular forecast scenarios identified during the 1994 hurricane season (described in section one), to evaluate the 1995 model upgrades in the convective and boundary layer parameterizations within the context of those deficiencies.
The results are presented for three forecast scenarios in which the upgrades dramatically improved upon the hurricane forecasts from the operational model. Also, from those cases, we selected scenarios deemed to be more “difficult” forecasts—that is, recurving storms and/or weak steering currents. These scenarios were typically where the other NHC model guidance showed decreased forecast skill as well.
1) Hurricane Felix
An illustration of a “difficult” forecast period is shown in Fig. 2. These were the various track model forecasts received at NHC for Hurricane Felix initialized on 0000 UTC 16 August 1995 before recurvature. The observed track of Felix is denoted by the tropical storm/hurricane symbols. What is notable here is the disparity in the forecasts between the models that predict landfall and the models that predict recurvature. This caused serious concern among hurricane forecasters who were issuing warnings along the coastline during this time. In the next forecast period, 12 h later, all the models forecasted recurvature.
Figure 3 shows a comparison of the MRF and MRX track forecasts for Hurricane Felix, initialized for the same forecast period as above, at 0000 UTC 16 August. The MRX shows a clear recurvature of the storm (although about 220 km too far to the west) followed by an acceleration toward the east. The MRF, however, essentially stalls the storm off Cape Hatteras. That is a scenario usually associated with weakening storms and/or weak steering currents. To help determine the factors that contributedto a much improved track forecast in the MRX, the MRF and MRX forecasts of the hurricane and of the larger-scale steering flow were examined.
Figures 4a–d show the 500-mb flow for the 36- and 72-h forecasts for the MRF and MRX forecasts. At 36 h, just prior to recurvature, Felix is located at 36°N, 75°W and is a stronger vortex in the MRX (Fig. 4b) than in the MRF (Fig. 4a) as denoted by the 15 m s−1 isotach over the eastern semicircle of the storm in the MRX. Also, from the difference field of the 500-mb flow for this forecast period—that is, the MRF subtracted from the MRX (not shown)—although the magnitude of the difference of the vortex is not that large at this level (5–10 m s−1), it is significant that the isotach difference pattern is essentially wrapped around the vortex showing a more circular hurricane wind pattern in the MRX with the maximum difference over the eastern semicircle of the storm.
With respect to the synoptic-scale flow, a high pressure area was building over the Ohio Valley with associated northerly flow strengthening along the Atlantic coast. This feature is shown in the 36-h forecasts of both the MRF and the MRX; however, the forecast of the MRX suggests a greater buildup of the high eastward, with the northerly part of the circulation reaching the coastline. In fact, it was the very strong northerlies that developed along the coastline in association with the high building northward into Canada that became an important steering influence that kept Felix well offshore. A major difference is seen in the simulations of this feature in the MRF and MRX 72-h forecasts as shown in Figs. 4c and 4d, respectively. A much stronger northerly component of the flow parallels the coastline in the MRX with a stronger northward building of the ridge. This is not the case in the MRF simulation, whichshows a markedly weaker synoptic flow where the synoptic- and storm-scale waves are hardly distinguishable. With respect to the simulation of Felix, there is also a stronger midlevel storm circulation in the MRX as indicated by the 20 m s−1 isotach, and although the streamline pattern does not display a well-developed closed circulation, it is nevertheless a marked improvement over the elongated trough impinging onto the coastline that represents the storm in the MRF. To further examine the simulations of the vortex in greater detail, vertical cross sections through the storm were made at the storm locations shown in Figs. 4a–d to assess the impact of the new physics on maintaining the vertical structure of Felix.
A vertical cross section through the storm of the vorticity and vertical velocity at 36 h and 72 h is shown in Figs. 5a–d. At 36 h, the MRX maintains a deep and strong vortex with a vigorous vertical circulation (as depicted by the contours of vertical velocity) extendingthroughout the depth of the troposphere. The MRF within 36 h shows a weakened and shallow circulation with the strongest vorticity becoming concentrated in the lower levels. By 72 h, a marked vertical decoupling of the storm in the MRF is shown at mid- to upper levels as the vorticity becomes diffuse and the area of maximum rising motion (0.5 Pa s−1) is seen to be well displaced from the center of the vortex (Fig. 5c). This is to be compared to the MRX 72-h forecast (Fig. 5d), which develops a very strong vertically coherent circulation with maximum rising motion of 2.4 Pa s−1.
To try to gain some insight on the role that the changes to the physics had on the apparent structural disparities shown in the above figures, the latent heating from the deep convection was examined. Figure 6 shows a comparison of the vertical profiles of deep convective heating through the center of Felix for the 48-h forecasts for the MRF and the MRX. Not only is the magnitude of maximum heating increased in the MRX by about 5°day−1 in the upper troposphere, but the increased heating is spread over a greater depth of the column. The maintenance of the strong profile of condensational heating is essential in maintaining the vertical structural integrity of the storm and thus the storm circulation throughout the forecast. Analogously, a storm weakens in the real atmosphere when it encounters strong vertical shear by causing a disruption of the latent heating associated with the deep convection. It is crucial for a developing storm that the upper-level latent heat release remain above the low-level circulation center. By 72 h (not shown), the MRF convective heating weakens substantially at the mid- and upper levels (about 50% of the heating at 36 h), which is consistent with the decoupling of the vortex shown in Fig. 5a, while the strong convective heating is maintained throughout the forecast period in the MRX. The increased heating is a result of the new closure assumption described in section 2a.
The change to the PBL was also examined via the energy supplied to the storm from the surface fluxes,which were found to be significantly enhanced in the MRX. At 36 h, the maximum surface latent heat flux over the eastern hemisphere of Felix produced by the MRX was 325 W m−2, compared to approximately 210 W m−2 for the MRF (not shown). Also, the horizontal gradients of the fluxes were much larger in the MRX. At 72 h, as shown in Figs. 7a and 7b, the very large fluxes in the MRX (∼700 W m−2) are quite realistically indicative of a strongly disturbed surface layer in association with the passage of a hurricane, whereas the MRF fluxes do not at all indicate the presence of a storm.
The results presented above strongly suggest that a positive feedback between the boundary layer and the deep convection serve to strengthen the vertical structure of the storm by enhancing the vertical distribution of the heating. In the simplified Arakawa–Schubert scheme, the condition at the cloud base depends strongly on the moisture structure in the PBL. The rising parcel can lose buoyancy when it encounters dry subcloud layers. When convection is initiated, the drying associatedwith the subsidence reduces the buoyancy and stabilizes the atmosphere. A PBL scheme that vigorously mixes the moisture from the surface to the top of the PBL can restore the buoyancy and maintain the convection. In fact, a set of sensitivity experiments were made to investigate the individual contributions of the deep convection and PBL to the surface fluxes and convective heating by only including one of the changes at a time. These results (not shown) indicated that each contributed nearly equally in magnitude to the enhancement of the surface fluxes and to the convective heating, which is a good indicator of an improved interaction between the convection and the PBL parameterization schemes.
The general result of the improved physics is to simulate a stronger and improved storm structure that contributes to an improved track forecast. Additionally, in this simulation, we found an improvement in the forecast of the amplitude of the larger-scale flow that also influences the recurvature of Felix. Here, it is suggested that the improved forecast of recurvature is a result of an improved simulation of both these features. Moreover, this also suggests the subtle but important role of diabatic heating in providing energy for the maintenance of the larger-scale circulations.
Finally, with respect to the westward track bias, although an improved track simulation was shown for this forecast scenario in the MRX, a general tendency forwestward motion still exists in the earlier forecast periods of Felix. It is not, however, any worse than the MRF forecasts. The westward bias appears early in the forecast period and is closely related to the bogussing problem as described by Lord et al. (1996, manuscript submitted to Mon. Wea. Rev.).
2) Hurricane Iris
Another interesting forecast improvement was for Hurricane Iris initialized at 0000 UTC 24 August, 6 h after being upgraded to a hurricane. This was a difficult forecast period due to the complex interaction between Iris and Hurricane Humberto that was strengthening about 1390 km to the east. The relatively close position of Humberto was influential in causing Iris to make a rather abrupt turn toward the west southwest from a previous northwest heading. In fact, all of the other numerical guidance showed difficulty in changing from the northward heading. Further difficulty arose in the AVN forecasts for this storm since the bogussing system was inoperative during this period that also negatively affected the other models initialized by the NCEP GDAS.
Figure 8 shows the MRF and MRX tracks corresponding to this period. Although the MRX did not correctly forecast the more southerly track component,it did correctly forecast the turn toward the west in the first 12 h. Figures 9a and 9b show the 12-h forecasts of Iris initialized at 0000 UTC 24 August for the MRF and MRX, respectively. At this time Iris was located at 14.8°N, 55.1°W and was a 64-kt hurricane. The main difference between the MRF and the MRX is a stronger and more distinct circulation of not only Iris but also for Humberto, which was a 90-kt hurricane at this time (located at 15.4°N, 42.6°W) in the MRX forecast. Also, the position of Humberto is too far north in the MRF. At the midlevels the MRF begins to diffuse the vorticity and develops an elongated trough. This is readily apparent by 36 h and is shown in Figs. 9c and 9d in a comparison of the 500-mb flows. The MRX, on the other hand maintains the circulations of Iris and Humberto and evidence is also seen in the eastern Atlantic of the spinup of what was to become Tropical Storm Karen (located at 11°N, 27°W). Figures 10a and 10b show a vertical cross section of the vorticity and vertical motion along the latitudes of the storms for the 36-h forecast from the MRF (Fig. 10a) and from the MRX (Fig. 10b).A clear signature of Iris is shown in both forecasts of the vorticity but the vertical circulation in the MRX forecast is somewhat stronger. What is strikingly different in the MRX is the clear spinup of Humberto at 50°W. Again, this spinup is without the benefit of an artificially induced initial vortex and arises solely from the ability of the model to generate storm-scale vorticity via the strong interactions of the model physical and dynamical processes. Finally at 72 h, the MRF (Fig. 11a) has lost the signatures of all the storms while the MRX (Fig. 11b) correctly forecasts not only Iris (at 15.1°N, 61.4°W) but Humberto (at 25°N, 50°W, slightly north and east of observed position), Karen (15.6°N, 35.4°W), and to some extent what was to become Hurricane Luis, although the position of Luis is too far west.
3) Tropical Storm Pablo
Last, we present results that addressed the deficiency in the model to properly simulate the mid-oceanic tropical upper-level trough over the tropical Atlantic that affected the intensity forecasts for Tropical Storm Pablo. Due the inability of the model to forecast the very strong westerly shear associated with the southward extension of the trough into the deep Tropics, no hint was provided in the global model forecast of the rapid weakening of Pablo upon encountering the westerlies within three days after being upgraded to storm status. At 0000 UTC 5 October, Pablo was a depression located at 8.4°N, 32.8°W or about 3100 km east of the Lesser Antilles. At this time, Pablo was located in easterly flow, which was associated with an upper-level anticyclone located over the eastern Atlantic. This scenario supported further strengthening from a depression to tropical storm intensity. Within 48 h, the anticyclone was pushed to the south as the upper-level trough deepened over the tropical mid–Atlantic Ocean pushing southward and extending westward with time. Figures 12a and 12b show the 48-h forecasts of this feature for the MRF and MRX, respectively. The position of Pablo is indicated in the figures by the dots. The verifying analysis for 7 October is presented in Fig. 12c. The important difference between the forecasts is the more southward extension of the trough in the MRX, essentially bringing the westerlies over the circulation of Pablo. In the MRF, Pablo remains under northeasterlies associated with the poor forecast of the deepening of the trough and the building of a ridge over the northeastern Caribbean toward the east. Comparing the MRX 48-h forecast (Fig. 12b) with the verifying analysis (Fig. 12c), the entire configuration of the trough is extremely well simulated with the axis extending northwestward.
Figures 12d and 12e show the 72-h 200-mb flow for MRF and MRX forecasts with the verifying analysis on 8 October shown in Fig. 12f. Again, the MRX provides an excellent forecast of a well-developed trough with associated northwesterly flow over Pablo that corresponds well with the analysis of this feature. To illustrate the magnitude of the difference between the forecasts, Fig. 12g shows the MRF subtracted from the MRX forthe 72-h forecasts. The cyclonic circulation, with the center located at approximately 18°N, 52°W, is indicative of the deeper trough in the MRX forecast. Also note that Pablo is located between the 10 and 15 m s−1 isotachs of the westerly wind difference between the two models, highlighting the area of the stronger MRX westerlies. It was 18 h later that Pablo dissipated at 57.5°W due to the very strong westerly vertical shear associated with the trough. Although we did not run Pablo in parallel systems to explicitly examine the differences in the forecasts of the storm in detail, we are confident based on the above result for the large-scale flow that an improvement in the intensity forecast would have been obtained.
c. Impact of MRX upgrades on the NHC model verification system
Since it became operational two years ago, the GFDL hurricane forecast model (Kurihara et al. 1995) has provided the best track verification scores over any other model. The model makes use of the NCEP global model analysis during its initialization process. Therefore, todetermine the impact of all of the upgrades on the GFDL track forecasts, the storms of August and September were rerun for the GFDL model from the new global model system. Figure 13 shows the skill of the model relative to CLIPER as a function of forecast time. The GFDL is the version of the model run from the operational system without the upgrades. The GFDX is the forecast run from the global system with the upgrades. The number of cases indicates the number of runs at 12, 24, 36, 48, and 72 h for the hurricanes that occurred between 1 August and 30 September 1995. As indicated from Fig. 13, the GFDX shows an improvement in the track forecasts compared to the GFDL at all forecast times. From these statistics, roughly a 20% reduction in the track error was found for these cases. These results were statistically significant at all time levels (Bender 1996, personal communication). The track error at 36 h has been reduced from 189 to 152 km; at 48 h from 236 to 199 km; and at 72 h from 368 to 326 km.
Finally, to provide an overall context to assess the global model upgrades by comparing its performance for the entire 1995 hurricane season (without the upgrades) with the 1996 hurricane season (with the upgrades), and to further assess their impact on all theother operational models that rely on the global model for initialization—for example, BAM, NHC90, VICBAR, as well as the GFDL model—the verification skill of all the models is shown for the 1995 and 1996 hurricane seasons in Figs. 14a and 14b, respectively. All model skill is relative to CLIPER.
It is shown that the AVN at 48 h has increased skill by approximately 10% relative to CLIPER from the 1995 season to the 1996 season (cf. 0.12 in Fig. 14a to 0.22 in Fig. 14b) For the 72-h forecasts, the skill has increased by about 25% between the 1995 season and the 1996 season (cf. 0.05 in Fig. 14a to 0.30 in Fig. 14b). Most noticable here is not only is there a substantial gain in skill of the global model after 36 h, but that there is a complete reversal in the trend of the skill after 48 h. In the 1995 hurricane season a marked loss of skill occurred for the AVN after 48 h (this was also true for the simpler models). This deterioration in skill has been eliminated in the global model. We are confident that this improvement is a direct consequence of the improved physics that dramatically impacted the storm circulation and environment after 36 h as shown in the previous sections.
The verification scores for the eastern Pacfic for theglobal model were also calculated (not shown). From these results the AVN also showed substantial improvement for the 1996 season and showed comparable skill to the BAMS models, whereas in the 1995 season the AVN showed no skill even relative to CLIPER.
Additionally, the GFDL model shows a 15% increase in the 48-h forecast skill between the seasons (cf. 0.35 in 1995 to 0.5 in 1996) and an approximately 18% improvement in skill at 72 h (cf. 0.35 in 1995 to 0.53 in 1996), which is comparable to the preliminary results for the test cases of the 1995 season mentioned above. Furthermore, all of the other models show an increase in skill during the 1996 season. Although the barotropic model still loses skill after 48 h, the skill at 48 h has increased 5% relative to CLIPER between the seasons;NHC90 has 5% increased skill at 48 h and reverses a negative trend from the previous season out to 72 h and the BAMS models show about a 5% increase in skill.
This study has evaluated the impact of recent changes to the deep convective and boundary layer parameterizatons in the NCEP global model. We have examined the impact on the large-scale tropical forecasts and on hurricane forecasts during the latter half of the 1995 hurricane season. Our diagnostics were based on known deficiencies in forecasting tropical systems during the 1994 hurricane season.
A modest overall improvement in the 72-h 200-mb rms vector wind error was found in the MRX across the tropical/subtropical Atlantic and eastern Pacific Oceans. This improvement was consistent with the lower rms errors for the 500-mb height fields over the Northern Hemisphere. In particular, the upper-level wind errors over the Caribbean, which had been identified as a problem area for hurricane intensity forecasts, were reduced. Forecasts of the low-level tropical flow, however, were degraded with the new model upgrades. Trying to identify the source of this error will be the focus of future investigations.
Changes in the deep convective and boundary layer parameterizations have increased convective activity and diabating heating for tropical disturbances, which has led to stronger vertical coupling and an improved vertical structure of the forecast model vortex. Improvedvertical vortex structure contributed to an improved track forecast for Hurricane Felix and ameliorated a problem noted in earlier 1994 cases. Additionally, a stronger amplitude in the forecast of the synoptic wavepattern over the eastern part of the United States also contributed toward improving the recurvature in the track forecast.
The MRX generated and maintained stronger circulations for Hurricanes Iris and Humberto, and indicated the regions of cyclogenesis for Karen and Luis. Generally, the improved model appears to develop realistic storm vorticity via strong interactions of the model physics and dynamics. Although the model’s ability to generate andmaintain a stronger vortex has led to improved track forecasts, the leftward track bias has not been totally eliminated. This bias may be related to the storm initialization problem and is being investigated (Lord et al. 1996, manuscript submitted to Mon. Wea. Rev.).
Additionally, it was shown that the MRX provided a much improved simulation of the midoceanic tropical upper-level trough. The maintenance of the trough was important in the intensity forecasts for Tropical Storm Pablo because of the associated westerly shear. This model deficiency is one that had been documented over the Caribbean in the previous year. It is however, adeficiency, which we suggest, is reflective of the more global upper-level easterly bias problem that the NCEP global model and other global models have been investigating over a period of years. As a possible contributor to the planetary-scale problem, we plan to address the maintenance of the upper-level troughs within the framework of the transfer of energy between the scales of motion via the divergent and rotational circulations.
The positive impact of the new model physics was also shown for the August and September 1995 hurricane forecasts from the GFDL hurricane model. This resulted in an overall 20% increase in skill in the GFDL hurricane track model forecasts. Furthermore from the verification scores for the NHC numerical guidance models comparing their performance for the 1995 hurricane season with the 1996 hurricane season, improved forecast skill was shown for all of the model guidance used at NHC. These results showed an increase in skill between 1995 to 1996 from 10% to 25% for the global model for the 36–72-h forecasts. There was also substantial improvement in the performance of the global model over the eastern Pacific wherein the previous season the model showed no skill at all in this basin. Furthermore, a 15%–18% improvement was shown in the GFDL model with the other models increasing between5% and 15% in performance skill as well. The improvement in the forecast performance of the global model and in the other models that are impacted by changes to the global model, we believe, is due to the improvements to the deep convection and boundary layer model physics.
The authors would like to thank Ron McPherson, Eugenia Kalnay, and Robert C. Sheets for encouraging and supporting this collaboration. Thanks are also due to John Manobiano for a very thorough and helpful review of the manuscript and to the other reviewers who helped in improving this work. Much appreciation is expressed to Joan David for her expert drafting of some of the diagrams.
Aberson, S. D., and M. DeMaria, 1994: Verification of a nested barotropic hurricane track forecast model (VICBAR). Mon. Wea. Rev.,122, 2804–2815.
Caplan, P. M., and G. H. White, 1989: Performance of the National Meteorological Center’s medium-range model. Wea. Forecasting,4, 392–400.
——, ——, and J. G. Jiing, 1993: Skill of the NMC global model in the Tropics. Preprints, 13th Conf. on Weather Analysis and Forecasting, Vienna, VA, Amer. Meteor. Soc., 225–228.
——, J. Derber, W. Gemmill, S. Y. Hong, H. L. Pan, and D. Parrish, 1997: Changes to the 1995 NCEP operational medium-range forecast model analysis/forecast system. Wea. Forecasting,12, 581–594.
DeMaria, M., and J. Kaplan, 1997: An operational evaluation of a statistical hurricane intensity prediction scheme (SHIPS). Preprints, 22d Conf. on Hurricanes and Tropical Meteorology. Fort Collins, CO, Amer. Meteor. Soc., 280–281.
Derber, J. C., and W.-S. Wu, 1998: The use of TOVS cloud-cleared radiances in the NCEP SSI analysis system. Mon. Wea. Rev., in press.
——, D. F. Parrish, and S. J. Lord, 1991: The new global operational analysis system at the National Meteorological Center. Wea. Forecasting,6, 538–547.
Elsberry, R. L., 1995: Recent advancements in dynamical tropical cyclone track predictions. J. Meteor. Atmos. Phys.,56, 81–99.
Fitzpatrick, P. J., J. A. Knaff, C. W. Landsea, and S. V. Finley, 1995:Documentation of a systematic bias in the aviation model’s forecast of the Atlantic tropical upper-tropospheric trough: Implications for tropical cyclone forecasting. Wea. Forecasting,10, 433–446.
Hong, S.-Y., and H.-L. Pan, 1996: Nonlocal boundary layer vertical diffusion in a medium-range forecast model. Mon. Wea. Rev.,124, 2322–2339.
Kanamitsu, M., 1989: Description of the NMC Global Data Assimilation and Forecast System. Wea. Forecasting,4, 335–342.
——, and S. Saha, 1995: Spectral budget analysis of the short-range forecast error of the NMC Medium-Range Forecast Model. Mon. Wea. Rev.,123, 1834–1850.
——, and Coauthors, 1991: Recent changes implemented into the global forecast system at NMC. Wea. Forecasting,6, 425–435.
Krishnamurti, T. N., and H. S. Bedi, K. S. Yap, and D. Oosterhof, 1993: Hurricane forecasts in the FSU Models. Adv. Atmos. Sci.,10, 121–131.
Kurihara, Y., M. A. Bender, R. E. Tuleya, and R. J. Ross, 1995: Improvements in the GFDL hurricane prediction system. Mon. Wea. Rev.,123, 2791–2801.
Lord, S. J., 1991: A bogussing system for vortex circulations in the National Meteorological Center Global Forecast Model. Proc. 19th Conf. on Hurricanes and Tropical Meteorology, Miami, FL, Amer. Meteor. Soc., 328–330.
——, and A. Arakawa, 1980: Interaction of a cumulus cloud ensemble with large-scale environment, Part II. J. Atmos. Sci.,37, 2677–2692.
Pan, H.-L., and W.-S. Wu, 1994: Implementing a mass flux convection parameterization package for the NMC Medium-Range Forecast Model. Preprints, 10th Conf. on Numerical Weather Prediction, Portland, OR, Amer. Meteor Soc., 96–98.
Surgi, N., 1989: Systematic errors of the FSU Global Spectral Model. Mon. Wea. Rev.,117, 1751–1766.
——, 1994: Performance of the AVN model for the 1994 hurricane season. Proc. 49th Interdepartmental Hurricane Conf., Washington, DC, Office of the Federal Coordinator, A-13.
White, G. H., and P. M. Caplan, 1991: Systematic performance of the NMC Medium-Range Model. Preprints, Ninth Conf. on Numerical Weather Predition, Denver, CO, Amer. Meteor. Soc., 806–809.
Hurricane guidance models.