SUMMARY

The configuration of the NCEP global ensemble forecasting system has not been upgraded since March 1994. The new IBM SP supercomputer now makes it possible to improve the ensemble system. It is proposed that (1) the horizontal resolution for all perturbed ensemble members is increased from T62 to T126; (2) the number of ensemble members at the 1200 UTC cycle is increased from 4 to 10; and (3) the postprocessing of the ensemble forecasts is changed to reflect these upgrades. Results from both operational forecast statistics and preimplementation tests indicate that the changes, as expected, have a positive impact, making the ensemble forecasts more useful overall.
 

MOTIVATION

The current configuration of the NCEP global ensemble forecasting system was implemented more than 6 years ago, in March 1994, after the aquisition of the Cray90 supercomputer. Though there have been minor changes implemented, the basic configuration of the system has not been changed since. The aquisition of a new IBM supercomputer in 1999 makes it possible to consider an upgrade to the existing ensmeble forecast system.

CURRENT CONFIGURATION

The current operational ensemble configuration consists of running:

1. At 0000 UTC each day:
• T170 high resolution control out to 7 days, after which this run gets truncated and is run out to 16 days at a T62 resolution
• T62 control that is started with a truncated T170 analysis
• 10 perturbed forecasts each run at T62 horizontal resolution. The perturbations are from five independent breeding cycles (see Toth and Kalnay, 1993, 1997). Each perturbation is both added to, and subtracted from the control analysis.
2. At 1200 UTC:
• T170 control out to 3 days ("Aviation forecast"), that gets truncated and run at T62 resolution out to 16 days
• Two pairs of perturbed forecasts based on two independent breeding cycles (four perturbed integrations out to 16 days.)

 

BACKGROUND

The value of the current ensemble forecast system is limited in the following ways.
1)    The perturbed forecasts are run at the low T62 horizontal resolution, therefore the forecasts have a suboptimal performance both in terms of their systematic and random errors. This is a well known problem noted by the users.
2)    The system's ability is further limited by the relatively small number of perturbed forecasts in:
a)    Identifying extreme/rare events
b)    Serving customers with high or low cost-loss ratios
c)    Providing flow dependent covariance information used in targeted observations and possibly in the analysis
d)    Provinding twice daily perturbed boundary conditions for the NCEP limited area ensemble. Without more global ensemble forecasts introduced at the 12 UTC cycle, the planned regional ensemble system cannot be run.

Research conducted during the past few years, primarily at ECMWF, demonstrated that problem (1) can be well addressed by the use of a higher resolution model version. The introduction of a T159 resolution model in their ensemble, in place of the T63 model, imporved the overall performance of their ensemble significantly (Buizza et al., 1998). These studies also showed that the increase of ensemble membership (in their case, from 33 to 50 members) has substantial benefits with respect to problems listed under points 2a and 2b above. Regarding problem 2c above, a study by Houtekamer and Mitchell (1998), and Toth and Szunyogh (1997) demonstrated that analysis and targeted observational work, respectively, can greatly benefit from increased ensemble membership.
 

PROPOSED CONFIGURATION

To solve or ameliorate the problems listed above, we propose the implementation of the follwoing changes to the existing system:
1)    Increase the horizontal resolution of all perturbed ensemble members from T62 to T126 for the first 84 hrs of integration
2)    Increase the number of perturbed forecasts at the 1200 UTC cyle from 4 to 10
3)    Implement corresponding changes in the postprocessing of the ensemble data

With the above changes, the proposed configuration of the global ensemble system is:
0000 UTC:
a)    T170 (MRF) control forecast out to 7 days, truncated to T62 afterwards, integrated out to 16 days
b)    T126 ensemble control out to 84 hrs, T62 afterwards, out to 16 days
c)    10 perturbed forecasts with T126 resolution out to 84 hrs, T62 afterwards, out to 16 days

1200 UTC:
a)    T170 (MRF) control forecast out to 3 (or 5) days, truncated to T62 afterwards, integrated out to 16 days
b)    T126 ensemble control out to 84 hrs, T62 afterwards, out to 16 days
c)    10 perturbed forecasts with T126 resolution out to 84 hrs, T62 afterwards, out to 16 days
 

With the new configuration, the ensemble forecasts will be symmetrically arranged at 00 UTC and 12 UTC, which will have positive consequences with respect to their general use. Note that the forecasts will have a uniform resolution for the short range forecast period, providing high quality forecast guidance, and  boundary conditions for the limited area ensembles.

The proposed configuration amounts to an increase in cpu resources of a factor of 3.7, well below the actual cpu increase of 4.5-5, due to the implementation of the new IBM supercomputer. Note that the resolution of the MRF control forecast was recently changed from T126L28 to T170L42, which amounts to a cpu increase of the same order (factor of 3.7).
 

EXPERIMENTAL RESULTS

It is widely known that the T126 version of the MRF model has a superior performance to that of the T62 version of the same model. We present two aspects of this superiority in terms of operational forecast verification statistics. First, the T126 control model forecasts exhibit less rms error (see figure), and higher pattern anomaly correlation scores (not shown):

Second, when the two control forecasts are evaluated in the framework of the full ensemble, the T126 forecasts are found to be the closest member to the verifying analysis field more often than the T62 control (or the T62 pertubed forecasts):
Note the large difference in the average frequency values of a forecast being the best in favor of the T126 resolution  model: around 4 days lead time, the high resolution model is almost twice as likely to be the best than the low resolution version. These results clearly indicate that the higher resolution of the T126 model version helps the individual forecasts perform better.

To evaluate the impact of the proposed increase in resolution for the ensemble, two 10-member, otherwise identical sets of ensembles were generated  for a 30-day period between January 13 - February 12, 1999. The first set was a replica of the currently operational system at 0000 UTC, while the other used T126 resolution for the first 3 days, and T62 thereafter. A comparison of the performance of the ensemble mean forecasts from the two 10-member ensembles for the Northern Hemisphere (NH) extratropics revealed that the ensemble that had  the higher T126 resolution for the first 3 days had considerably higher Pattern Anomaly Correlation (PAC) scores out to seven days and beyond (see figure). Predictability is extended by the proposed system as demonstrated by  its PAC values dropping to that of the level of the operational system 8 (16) hours later around the 3 (6) days lead time.

The RMS errors for the two systems confirm the above results. On the next figure shown is the rms error for the MRF control forecast (T126 out to 7 days), and for the proposed ensemble configuration (first 3 days T126, then T62), for the NH extratropics. The error for the ensemble mean, as it should be for a well performing system, is at or below the level of the control forecast error. Note that this is not the case for the operational T62 resolution ensemble, whose mean forecast performs poorer than its control during the first 3 days of the integration (see figure). An analysis of the results indicates that the the difference in the performance of the two ensemble means is most pronounced near orographic areas. This is not surprising since the only difference between the two ensembles compared is their horizontal resolution.  The impact of increased resolution is in fact larger for the ensemble mean forecast than it is for the control forecast  (see figure). Increasing the resolution for the first three days of integration results in an rms error reduction of 11-5% for the ensemble, and only 7-2% for the control, evaluated over the first week of forecasts.  We speculate that the lower resolution model may exhibit some unrealistic instabilities related to erroneously resolved orography and that these instabilities may contaminate the ensemble forecasts more than the control forecasts. Therefore the resolution seems to benefit more the ensemble system than the control forecasts.

The performance of the proposed system was also evaluated over the Southern Hemispheric extratropics and over the tropics. In these areas the increased horizontal resolution still has a positive, though less pronounced impact.
 

POSTPROCESSING CHANGES

Currently 2.5x2.5 latitude/longitude grid output files are generated for the global ensemble forecasts, for selected variables.
1)    The increased resolution of the ensemble forecasts necessitates the introduction of 1x1 lat/lon files for the first 84 hrs, for maximizing the value of these forecasts to the users. These files will be generated in addition to the 2.5x2.5 resolution files.
2)    At the same time, we plan to introduce a new set of files called "ensstat" that will contain different statistics based on the ensemble that can be later augmented by additional information. At this time, we plan to introduce ensemble mean and standard deviation, and normalized ensemble spread, for each lead time separately. For spread normalization, we will compute the average ensemble spread over the most recent past 30-day period. For the same period, the average error in the ensemble mean forecast will also be computed. To the degree the average error is larger than the average spread, the ensemble's spread is insufficient. We will evaluate whether a simple bias correction in the spread of the ensemble can be applied to the ensemble forecast data to ameliorate this problem, and to provide a zero order calibration for probabilistic forecasts.
3)    The average of the ensemble mean forecasts for each lead time for the most recent 30 days, and the corresponding average for the verifying analyses will also be computed. This will allow for the implementation of a simple bias correction to reduce the systematic error in the first moment of the forecasts, based on the mean error over a 30-day preceding period.
 

OUTLOOK

The current changes represent the first step in the upgrade of the global ensemble forecast system. Once the second phase of the IBM SP machine is operational, we expect to propose a second upgrade, where the T126 resolution would be extended to 7.5 days, and additional ensemble members would be introduced at the 0600 and 1800 cycles (10 at both cycles). The two upgrades arranged this way are to maximize the utility of the ensemble forecasts, given the increased computational resources, while keeping changes to the system at a minimum.
 

CREDITS

Yannick Tremolet provided invaluable help in the process of setting up the ensemble forecast system in a parallel computational environment. Mark Iredell, Hua-Lu Pan, and Steve Lord gave advice and encouragement during the course of the research.
 

REFERENCES

Buizza, R., Petroliagis, T., Palmer, T. N., Barkmeijer, J., Hamrud, M., Hollingsworth, A., Simmons, A., and Wedi, N., 1998: Impact of model resolution and ensemble size on the performance of an ensemble prediction system.  Q. J. R. Meteorol. Soc., 124, 1935-1960.
Houtekamer, P. L., and H. L. Mitchell, 1998: Data assimilation using an ensemble Kalman filter technique. Mon. Wea. Rev., 126, 796-811.
Toth, Z., and E. Kalnay, 1993: Ensemble Forecasting at the N MC: The generation of perturbations. Bull.  Amer.  Meteorol.  Soc., 74, 2317-2330.
Toth, Z., and E. Kalnay, 1997: Ensemble forecasting at NCEP and the breeding method.  Mon.  Wea. Rev, 125, 3297-3319.
Toth, Z., and I. Szunyogh, 1997: Review of the use of ECMWF ensemble data for targeting upstream observations during the FASTEX field experiments. In: Proceedings of the FASTEX Upstream Observations Workshop, NCEP, Camp Springs, Md, April 10-11, 1997. NOAA/NWS/NCEP Office Note 420, p. 131. [Available from: Environmental Modeling Center, 5200 Auth Rd., WWB, Rm. 207, Camp Springs, MD 20746.]