Changes to the 1998 NCEP Operational MRF Model
John Derber, Hua-Lu Pan, Jordan Alpert, Peter Caplan, Glenn White, Mark Iredell, Yu-Tai Hou, Ken Campana and Shrinivas Moorthi
National Centers for Environmental Prediction
W/NP23, World Weather Building,
Washington DC 20233, USA
An extensive package of changes will be made to the NCEP global analysis and forecast system on 15 June 1998 at 1200 UTC. The changes include (1) An increase in horizontal and vertical resolution from T126l28 to T170L42 for the analysis and first 84 forecast hours; (2) physics updates, affecting radiation and clouds, land surface parameterization, cumulus convection and gravity wave drag; (3) analysis and data assimilation updates, including changes in time interpolation, nonlinear interpolation, limits on supersaturation and negative moisture, background error covariance, 3-D ozone, level 1-b polar orbiter data, GOES radiances, and Y2K compliance. During the first two months in which the new package was run in parallel with the operational, the following features were observed: (1) reduction in the wet bias in precipitation over land (Fig. 5.6b); (2) reduction of the cold bias over much of the atmosphere, including the stratosphere and near-surface layers, together with the appearance of a warm bias over land in the lower atmosphere, all reflecting changes in the surface physics and in the radiation, and (3) improvement in the model's ability to maintain transient eddy kinetic energy throughout the forecast (Fig. 5.2a) and (Fig. 5.2b). The objective performance statistics for 1-2 months of testing show clear improvement in precipitation bias, reduction in jet-level rms vector error versus analysis - especially in the short range - (Fig. 5.3), increased accuracy of tropical winds versus analysis (Fig. 5.4), and no significant change in 5-day 500-hPa anomaly correlations for geopotential (Fig. 5.5).
==>[POST-IMPLEMENTATION CHANGES: See Section 9]<==
On 15 June 1998 changes to the following areas in the MRF analysis/forecast system will be implemented:
In this Technical Procedures Bulletin, for the sake of clarity the model that is to be replaced will be referred to as the "old" or "operational" model, while the new model will be referred to in the present tense, even though the TPB was posted before the implementation.
2.0 RESOLUTION INCREASE (M. Iredell)
Because small-scale features in the atmosphere tend to be more transient than the
larger-scale features, it was expected that increased resolution in the global model
would most benefit the short-term forecasts and in particular the data assimilation
system. Therefore, both horizontal and vertical resolution have been increased in
the global model to T170 L42 for the Final global data assimilation system, the full
4-per-day 78-hour Aviation forecasts and the first 84 hours of the Medium Range
Forecast (MRF). (An exception is in the rare case when a forecast must be run on
the backup computer, in which case it would be run in degraded resolution.) The
MRF (beyond forecast hour 84) as well as the other ensemble members remain
unchanged in resolution from the previous configuration. All the T170 L42
forecasts are run with the new prognostic ozone; all the T126 L28 and T62 L28
forecasts are run with climatological ozone. A summary of the different resolutions
for all the global model forecasts can be found in
The horizontal resolution has been increased to global spectral triangular truncation
T170 from T126. Thus the wavelength of the shortest resolvable wave on the globe
has been reduced to just over 2 degrees from almost 3 degrees, which is roughly
equivalent to increasing the grid resolution to 80 km from 105 km.
The model nonlinear dynamics is computed on an alias-free 512 by 256 Gaussian dynamics grid. However, the interpolation to observations in the analysis and the physics parameterizations in the model are computed on the same 384 by 190 Gaussian physics grid as previously. This physics grid will resolve any feature in the T170 spectral model, since it still has significantly higher resolution than its so-called linear grid resolution of 340 by 170. Both the orography and the diabatic physics terms in the spectral model are lightly filtered near the end of the spectrum (up to a 40% reduction at wavenumber 170) in order to reduce the direct forcing of the shortest waves. This filter follows Lander and Hoskins (1997), which suggested that the shortest waves in a spectral model be reserved purely for nonlinear dynamical forcing. The model post-processing grid remains an equidistant 360 by 181 cylindrical grid. The isobaric fields on this grid are filtered at around wavenumber 120 compared to wavenumber 90 in the previous model.
shows model orography over the eastern U.S. for
the T170 resolution and the T126 resolution. Note that the higher
resolution now allows a finer characterization of the White Mountains in New
Hampshire and the Great Smokies on the border between Tennessee and North
Carolina. Also note that the 10-meter Gibbs rolls in the Atlantic Ocean have been
greatly reduced. A general reduction of Gibbs phenomena in the orographic forcing
and in the physics forcing is due to both higher resolution and the light filter of
The vertical resolution has been increased to 42 sigma levels from 28 sigma levels.
The extra resolution is fairly evenly distributed in the atmosphere, with about 3 new
levels in the stratosphere, about 8 new levels in the middle to upper troposphere and
about 3 new levels in the boundary layer. In the jet region around 250 hPa, the
vertical resolution is now less than 1.0 km where in the previous implementation it
was greater than 1.5 km. It is anticipated that this increase in resolution should help
to improve both the accuracy and the sharpness of jet wind forecasts for aviation
purposes. The top sigma level is now around 2.0 hPa, up from the 2.7 hPa top level
in the previous 28-layer model. The vertical distribution of levels is now minimally
adequate to simulate prognostic ozone in the stratosphere.
The new vertical distribution of 42 sigma levels is shown in
Note that the distribution of levels is smoothly varying in the vertical. This is necessary
in order to minimize the truncation error in vertical differentiation and integration.
Furthermore, there is an emphasis on resolution in the boundary layer, in order to
provide greater characterization of inversions and fronts as well as to minimize
vertical truncation error. Although the vertical resolution is fine in the stratosphere
when viewed in the pressure coordinate, it resolution is quite coarse in the height
coordinate, the appropriate coordinate for stratospheric dynamics.
The change of resolution for the MRF at hour 84 (to T126 L28) and hour 168 (to
T62 L28) is accomplished with minimal impact on the forecast. A digital filter
initialization is run in the new lower resolution with the added cost of 3 forecast
integration hours in order to reduce the transient noise caused by any imbalances
resulting from the change of resolution.
Several steps have been taken to fit the higher-resolution model into the available
time windows on the computer. First, several data processing steps have been
better optimized. The SSM/I and ERS-2 surface winds processing has been moved
earlier in the cycle in order to minimize the delays caused by these steps. The SSI
analysis has been greatly optimized to multiprocess its work across the whole
computer. The spectral model also is better spread over the computer. The cost of
a quadratic model physics grid had been found to be much greater than its benefit,
so its relative grid size has been reduced as mentioned above. The T170 L42 SSI
analysis now takes about 15 minutes on the Cray C916 and the T170 L42 global
spectral model takes about 15 minutes per forecast day.
References:Lander, J. and B.J.Hoskins, 1997: Believable Scales and Parameterizations in a Spectral Transform Model, Mon. Wea. Rev., 125, 292-303.
3.0 CHANGES TO THE OPERATIONAL FORECAST MODEL
3.1 Land-surface parameterization upgrades (H.-L. Pan)
3.2 Convection changes (H.-L.Pan)
The model uses a Simplified Arakawa-Schubert
scheme with the quasi-equilibrium closure:
Mb = ( A - alpha(w) Ac ) / tau(w)
where A is the cloud work function, Ac is a critical cloud work function that is cloud-depth dependent, w is the vertical motion, alpha and tau are tuned parameters based on cloud bottom large-scale vertical motion, and Mb is the mass flux that will provide the balance between grid scale and the cloud scale environment. We have now modified alpha to reduce the critical cloud work function to allow for earlier initiation of convection when the cloud base vertical motion is upward. We also modified the upper limit of alpha to change the scheme to a CAPE elimination scheme with a lower value of omega. This is done only over the ocean to reduce the precipitation bulls-eye problem encountered during the summer of 1997.
In addition to the above changes, a modification of the evaporation formulation for convective rain has also been made. The current MRF model has a strong tendency to dry out the regions of convection, leading to a negative bias in the precipitable water field. This was suspected in 1995 when the model was run for a time with SSM/I precipitable water data and a marked spin-down in relative humidity resulted. Since observations show that significant evaporation occurs when the environmental shear is strong, we have modified the evaporation efficiency factor from a constant value of .07 to a factor that depends on wind shear, following Fritsch and Chappell (1980). This change has led to significant improvement in the global humidity climatology of the model. This change was made only over ocean grid points.
References:Fritsch, J. M. and C. F. Chappell, 1980: Numerical Prediction of convectively-driven mesoscale pressure systems. Part I: Convective parameterizaion, J. Atmos. Sci., 37, 1722-1733.
3.3 Enhanced gravity wave drag (and improved modeled leeside mountain cyclogensis(Jordan Alpert, Song-You Hong [General Sciences Corporation] and Young-Joon Kim [UCLA])
Gravity wave drag arises from the interaction of sub-grid scale gravity waves generated by the wind and orography and the subsequent vertical propagation of these waves and their interaction with the atmospheric momentum in the lower and upper troposphere. The model response to GWD is to warm the "too-cold" polar regions, improve the too-low height bias, and improve jet stream and cyclone track positions in large spatial scales and climate of the model - a general improvement to NWP model systematic errors. However, the modeled treatment of GWD parameterization in the lower troposphere is lacking in a number of respects. Improvements to the operational GWD are from two new orographic files in addition to the orography variance, (h')2:
There is evidence that NWP models often fail in predicting the correct intensity of leeside
mountain cyclogenesis and the often accompanying movement of cold air
outbreaks. Indeed, the model treatment of small-scale boundary layer friction drag and
the interaction with the variance in orography is still known incompletely in NWP
models. The GWD parameterization, while weakly-acting, influences the cyclone
track speed and development intensity by removing small-scale biases as well as
continuing to reduce large-scale biases, benefiting the model, by reducing
mid-latitude westerlies and warming the polar atmosphere.
References:Kim, Y.-J., 1995: Improvement of orographic gravity wave parameterization using a mesoscale gravity wave model. J. Atmos. Sci,52, 1875-1902.
The NCEP operational shortwave (SW) radiation parameterization
was developed at the Geophysical Fluid Dynamics Laboratory, using
portions of the scheme from Lacis and Hansen (JAS, 1974, p118).
While the longwave (LW) parameterization has been upgraded
several times in the past ten years, the SW method has not
changed. A newer, more accurate, method has been developed at
NCEP, based on the work of NASA's Chou and Lee
(JAS, 1996, p1204, as well as Chou, JAS, 1992, p762). The new SW
scheme consists of a multi-spectral band technique, an improved
calculation in cloudy atmospheres, the addition of climatological
aerosol effects, and a new surface albedo, which is a function of
14 distinct surface types.
Monthly forecast experiments with a T62 model, using the changes
to the SW scheme, showed significant improvement to the model-computed
radiation budget (Table 1), at both the top of the atmosphere (TOA) and
earth's surface (SFC). Improvements in the
surface albedo can be seen for the July 1985 test in
where the control (CNTL) contains the currently operational SW
algorithm, the experimental (EXP) contains the new SW scheme, and
the observed (ERBE) is derived from TOA flux data. The operational
model overestimates global planetary albedo, while it is
significantly reduced to more realistic values in the new scheme.
Atmospheric SW radiative heating is enhanced in the new model,
resulting in a warmer troposphere, especially in the tropics
(generally 0.5 - 1.0 degree K in the zonal averages).
Computational overhead with the new SW scheme is less than 5%.
TABLE 1 Global SW Radiation Budget (W/m2)
(data at earth's surface estimated from satellite measurements)
|MONTH||OBSERVED||NEW SW||OPNL SW|
|TOA-up clear||JUL 85||51.5||49.6||50.3|
|Net at SFC||JUL 85||151.6||158.5||160.7|
Concurrent with the change to the SW parameterization, the LW
scheme is slightly modified in regions of model multi-layer
cloud. In the operational model, radiative heating rates within these multi-layer
clouds are adjusted to the same value for all cloud layers,
that is, there is no large LW cooling at the cloud top, as is
observed in the atmosphere. This 'LW adjustment' is a leftover
from an earlier NCEP era, when zonal mean cloud was persisted
during a global forecast. Large values of continuous cloud top
cooling within the persisted clouds was found to be detrimental
to the forecast, and the heating rate adjustment was found to
ameliorate the problem. Additionally, the adjustment was
consistent with the procedure used in the old SW scheme. Since the new
SW scheme has removed this feature, the LW algorithm has thus been
made consistent with it.
The diagnostic cloud formulation is used only in the radiation
calculations (Campana etal, 1994, 10th AMS Conference on NWP).
The operational model diagnoses cloud coverage from model
temperature and moisture, in the form of relative humidity,
throughout the model troposphere. A set of cloud/humidity
relationships was developed from daily mean US Air Force RTNEPH
cloud analyses during August 1993-February 1994. Clouds in the
lowest 10% of the atmosphere are allowed only in suspected
marine stratus regions (i.e. oceanic moist inversion, capped by
very dry air). Consequently, especially over land, low stratus
is not well modeled, with occasional detrimental effects on
near-surface (sensible) weather. The scheme has now been retuned to
more recent synoptic (not daily mean) RTNEPH data (March-December
1995) and modified to include cloud/humidity relationships for
lower atmospheric clouds everywhere over the globe.
shows an example of the improvement to modeled low stratus cloud
over the eastern US in the new scheme. The top portion of the
figure contains the average lower atmosphere cloud fractional
coverage during hours 18-21 of a T126L28 global forecast, where
the old (left) and new (right) schemes are shown for a case in
January 1995. The bottom of the figure shows the US AirForce
observations for that time (the missing values in the
observations result from data that are too old for the 18-21 hour
forecast or where low cloud is obscured by higher cloud
(satellite observations over the oceans). The cloudiness being
advected down the northeast and middle Atlantic coast appears
better modeled in the new scheme.
When all the new components of the forecast model are included in the T170 model, the improvements to radiation and cloud parameters are quite striking. During the 10-day period, 14-23 March 1998, global planetary albedo values dropped from 35.6% to 32.6%, downward SW radiation at the earth's surface dropped by 15 W/m2, and total cloud cover increased from 47% to 53% for the new model. All changes are in the correct sense. Fig. 3.6, shows the mean of the daily 12-36 hour total cloud cover forecast for the 14-23 March 1998 period, for the new model, operational model, and RTNEPH analyses. The new clouds appear superior to the operational T126L28 data when compared with the mean of the RTNEPH analyses over this period.
3.5 Prognostic Ozone (S. Moorthi and M. Iredell)
Ozone is generated through photochemical processes above ~ 25 km from the earth's surface and carried to the stratosphere and upper troposphere through advection and mixing. The vertical profile of ozone generally has significant values above an altitude of 8 km with a peak value around an 10 to 30 km. Its strong absorption of solar ultraviolet radiation, is the dominant heat source in the stratosphere and is very important to the stratospheric general circulation. Ozone also plays an important role in the infrared radiative transfer in the stratosphere. Thus, accurate prediction of ozone is important to both numerical weather prediction as well as long-term climate predictions and global change studies.
The previous operational MRF model at used a very crude representation of ozone by prescribing a zonal and seasonal mean climatology. This assumption was especially inaccurate for long-range and climate prediction. Furthermore, a reliable first guess of ozone is now necessary for an accurate forward model from model space to radiance space for the assimilation of TOVS- and GOES-measured radiances. Thus including ozone as a prognostic variable will help the assimilation/forecast system in two ways. First, by providing more accurate radiative heating of the atmosphere, it can provide improved first guess fields for the analyses. Secondly, the predicted ozone itself is a necessary first guess field in the direct assimilation of radiances. Also, accurate prediction of spatial variability of ozone can help in improving the prediction of UV index (which gives a measure of harmful ultra-violet radiation reaching the earth's surface), thus improving our ability to warn the public.
The new model includes ozone as a three-dimensional prognostic variable. The continuity equation for ozone can be written as
dX/dt = P - LX + Xd
where X is the ozone mixing ratio in kg/kg, d/dt is the material derivative, P is the photochemical production rate (source), L is the destruction rate per unit ozone mixing ratio (sink), and Xd is weak horizontal and vertical diffusion of ozone. The ozone advection is treated exactly in the same manner as that for specific humidity, with Eulerian spectral approach. The source and sink terms are parameterized based on the zonal average, ten-day mean climatological data obtained from the GSFC two-dimensional model (R. Rood, A. Douglas, and M. Cerniglia, personal communication). The continuity equation is solved by time splitting. First, a provisional value of X is obtained after applying horizontal and vertical advection. Then the diffusion is applied in an implicit manner, the horizontal diffusion in spectral space and the vertical diffusion in grid-point space. Finally, the photochemical effects are applied, photochemical destruction being treated implicitly (Rood et al., 1991). Inclusion of prognostic ozone adds about 10% to the CPU time of the model.
Rood, R., A. R. Douglas, J. A. Kaye, M. A. Geller, C. Y. Chen, D. J. Allen, E. M. Larsen, E. R. Nash, J. E. Nielsen, 1991: Three-dimensional simulations of wintertime ozone variability in the lower stratosphere. J. Gephys. Res., 96 , # D3, 5055-5071.
4.1 Improved time interpolation.
In the previous version of the analysis, the 6-hour forecast, used as a background
(first guess), was interpolated between the zero and 6-hr forecast to the observation
time if it was prior to the analysis time. If the observation time fell after the analysis
time, the 6-hr forecast was used at the observation time.
Since the development of the previous version of the analysis, the 3 and 9 hr
forecasts have become available to the analysis. Since all observations fall between
the 3 and 9 hr forecast times, it is now possible to linearly interpolate to all
observation times with the largest interpolation being 1.5 hours. This change
resulted in small reductions in the fit of the observations to the background, with the
largest changes (but still small) being for satellite radiances. In most cases the
impact of this change should be small, with the largest changes for small scale
rapidly moving systems.
4.2 Nonlinear analysis
The new SSI analysis system has been modified in two ways to incorporated
nonlinearities in the analysis. First, the minimization algorithm has been changed to
be a nonlinear minimization algorithm. At this time, the only nonlinear components
in the minimization are the SSM/I wind speeds and in the limitation on the negative
and supersaturated moisture described below. For the SSM/I wind speeds, the
introduction of the nonlinear minimization algorithm allows the use of the wind
speed data without the creation of a direction. In the previous version of the
analysis system, the wind direction was created using a local analysis of the wind
direction. Since the SSM/I winds were often located far from any conventional
data, the created wind direction often reflected the background wind direction. This
resulted in the background wind direction being given too much weight and the near
surface wind analysis was slightly degraded. The most significant impact of this
change was in the southern hemisphere where a slight positive impact was noted.
The second change that introduces nonlinearities is the external iteration. The
external iteration is a loop around the entire analysis procedure which allows the
inclusion of weak nonlinearities in the observation and balance constraints which
would be too expensive to include directly in the minimization algorithm. For
example, the radiance calculation for the moisture channels is nonlinear. However,
the nonlinearity is weak enough that it can be included by relinearizing around the
solution every so often. Experiments indicate that the nonlinearities included in the
external iteration could be well accounted for by relinearizing about 5 times.
However, the computational expense of the external iteration is currently too large
to perform 5 external iterations. Therefore in the current implementation there are
only 2 external iterations (i.e., calculated observational increments, linearize,
perform 37 minimization iterations and then repeat). This is not ideal, but allows a
significant enhancement in the inclusion of nonlinearities in the analysis system and
has allowed a significant increase in the weight given the moisture channels for the
4.3 Limiting supersaturation and negative moisture.
In the previous version of the analysis, the supersaturation and negative values were
limited only at the end of the analysis. The limitation on the supersaturation and
negative values was such that it could not exceed supersaturation or become
negative except if the guess was supersaturated or negative. In this case, the values
were limited by the guess values. This formulation of the limitation on the
supersaturation and negative moisture was allowing areas of negative moisture to
grow in the analysis and not properly distributing integrated moisture quantities (i.e.,
total precipitable water, satellite radiances) properly in the vertical.
In the new version of the analysis, the inclusion of the nonlinearities in the analysis
allows a direct penalization of supersaturated and negative values in the
minimization. Since the constraint is not exact, it does allow negative and
supersaturated values (which is necessary for the grid to spectral transformation),
but it makes it much more unlikely to get supersaturated or negative values. This
change should make persistent negative areas unlikely in the analysis and will allow
the projection of integrated moisture information more properly in the vertical.
4.4 Reformulated background error covariance.
A new formulation for the background error covariance, which determines the
spatial and multivariate distribution of information in the analysis, has been
incorporated in the SSI analysis system. This formulation is similar to that
documented in Derber and Bouttier (1998) and Bouttier et al. (1997) and applied in
the ECMWF forecast model. The new formulation allows a specification of a
spatial variance field, reformulates the balance equation, and defines the spectral
statistics as a function of total wave number. This reformulation allows much greater
flexibility and should result in future improvements to the analysis system. In the
current implementation of the new formulation, the spatial variance fields are only
allowed to vary by lattitude. In future implementations, this field will vary with the
synoptic situation and include information on the presence of hurricanes. The major
impact of the new background error covariance is to improve the tropical analyses
and forecasts and to better model the lattitudinal variation in the moisture field.
4.5 3-D ozone.
Previously, the ozone analysis was a 2-D total ozone analysis performed separately
from the temperature, moisture and wind analysis. A 3-D ozone analysis has
become desirable in order to use radiances in the ozone analysis and to include
predicted ozone in the forecast model. Also, it has became desirable to combine the
analyses together. There is little impact on the thermodynamic variables except
through the radiance forward operator, but greatly simplifies the analysis code.
Because of the 3-D nature of the ozone analysis, it was necessary to include ozone
profile information. SBUV data was made available on an orbit-by-orbit basis and
these data has been incorporated in the ozone analysis. Note, because of this change,
radiances from the polar orbiters and the geostationary satellites are directly
influencing the ozone analysis and NESDIS total ozone retrievals are no longer
being used. The ozone analysis suffers from a lack of data and a lack of a long
history of improvements. While the basic features of the analysis and forecast are
reasonable, the ozone analysis should not yet be considered mature. Evaluation of
the new 3-D ozone analyses and forecasts are ongoing with Climate Prediction
Center and future significant improvements to the ozone analysis can be expected.
4.6 Changes in the use of polar orbiting level 1-b data.
Several changes to the use of the polar orbiting data have been made. The changes
can be divided into two categories; changes in the input data set, and changes to the
radiative transfer calculations. The changes to the input data set have primarily
involved the inclusion of additional satellite data. First, the radiances from NOAA-11
(both HIRS and MSU) have been incorporated into the analysis. The inclusion
of these data with those from the NOAA-12 MSU and the NOAA-14 HIRS and MSU
means that data from 3 different polar orbiting satellites are currently being used in
the analysis system. In addition, less thinning of the MSU and the HIRS data have
been done, resulting in additional satellite data being used for all satellites.
The radiative transfer used for the satellite data has undergone extensive changes.
The transmittance calculation in the radiative transfer has been changed to a version
from NESDIS called OPTRAN (McMillin et al., 1995). Substantial changes to the
OPTRAN and the radiative transfer code was necessary to reduce computational
cost, to allow the use of different satellites and to remove coding errors. The
surface emissivity calculations have also been modified for both the microwave and
the infrared channels. For the microwave channels, the effects of surface
polarization have been included for each channel and a frequency dependence has
been introduced in the surface emissivity. For the infrared, a simple model for the
surface emissivity as a function of satellite zenith angle, frequency and wind speed
based on Masuda et al. (1988) has been included. The effects of the infrared
emissivity model are more important for the geostationary radiances discussed
below because of the larger scan angles, but they do have a nontrivial impact also
for the polar orbiting data.
4.7 Inclusion of GOES radiances.
With the inclusion of the OPTRAN transmittance calculation, it became possible to
directly use radiances from GOES-8 and GOES-9 over the oceanic regions. The
radiances used are supplied by NESDIS operations and are only available where
NESDIS has determined there are clear fields of view. Several quality control
checks were introduced into the analysis. Most importantly, the ability to simulate
channel 8 within 10C is used to eliminate those field of views that are slightly cloud
or land contaminated. For this check, the inclusion of the surface emissivity
calculation mentioned in the previous section was particularly important. Also any
profile which is created from less than 3 field of views (of a 5x5 block) has been
eliminated. Other quality control procedures similar to those used in the polar
orbiting data are also applied. A bias correction similar to the bias correction used
for the polar orbiting data has also been included (McNally et al., 1998 previous tpb
ref.). The observational errors assigned to the GOES radiances are similar to those
from the polar orbiting data. Only data which is observed within +/- 1.5 hrs of the
analysis time are used. The primary impact of these data appears to be on the
moisture fields over the Eastern Pacific and the Western Atlantic.
4.8 T170 42 levels.
With the increase in model resolution, the analysis resolution was also increased.
The increased resolution has two primary effects on the analysis. First the better
resolution of the orography decreases the differences between the observed and
modeled elevation and improves the resolution of local orographic effects. Because
of this, the simulated observations from the background field will fit the
observations better. The improved vertical resolution primarily impacts the satellite
radiances resulting in improved forward calculations for the radiances.
4.9 Y2K compliance
The analysis system was modified to the fullest extent possible to be Y2K
compliant. The only component not yet Y2K compliant is the reading of the BUFR
input data set containing the conventional observations since the file is not yet
converted to Y2K format. This final conversion will occur soon.
Bouttier, F., J. Derber, and M. Fisher, 1997: The 1997 revision of the Jb term in
3D/4D-var. ECMWF Research Department Tech. Memo. No. 238, available from
ECMWF, Shinfield Park, Reading RG29AX, UK..
Derber, J. and Bouttier F., 1998: A reformulation of the background error
covariance in the ECMWF global data assimilation system. Submitted to Tellus.
McNally, A. P., J.C. Derber W.-S. Wu, B.B. Katz, 1998:Previous TPB
Masuda, K. T. Takshima and Y. Takayama, 1988: Emissivity of pure and sea waters
for the model sea surface in the infrared window regions. Remote Sens. Environ.,
McMillin, L. M., L. J. Crone, and T.J. Kleespies, 1995: Atmospheric transmittance
of an obsorbing gas. 5. Improvements to the OPTRAN approach. Applied Optics,
5. PARALLEL TESTING AND EVALUATION (G.H. White and P. Caplan)
The new implementation reduces several biases in the global forecasts.
shows the temperature bias in 5-day forecasts by the operational and
new models each verified against its own system's analyses for a two-week period in late
March and early April. The new model substantially reduces a cold bias in the
tropical and midlatitude tropopause, although it has somewhat more warm bias in
the polar regions in the tropopause and more cold bias in the stratosphere. The
reduced cold bias in much of the atmosphere reflects greater diabatic heating due to
the new radiation and soil physics. The new model's cloudiness also appears in
better agreement with independent estimates of cloudiness from satellite
observations, although oceanic regions of low-level stratus now appear to have too
little cloudiness. The new system also has warmer low-level temperatures
over land, correcting a cold bias in the operational system.
One longstanding problem in the operational model has been a tendency for
the forecasts to show less variability than the the analyses. The problem is
which shows the zonal mean transient eddy kinetic energy at
250 mb in five-day forecasts and verifying analyses for a two-week period in late
March and early April. This is the energy associated with daily deviations from the
two-week mean. Five-day forecasts with the operational model show subtantially
less day-to-day variability at nearly all latitudes than the verifying analyses.
Five-day forecasts with the new model display nearly the same variability as the verifying
analyses as can be seen in
Over the period in which the new analysis/forecast system was run daily in parallel the objective scores (each model vs. its own analysis) showed clear improvement in the rms vector errors in the jet-level winds over North America and the Northern Hemisphere, especially in the first few days of the forecasts (Fig. 5.3). The winds in the tropics showed substantial improvement (Fig. 5.4). The 500-mb anomaly correlations for geopotential in the extratropics were essentially the same for the new and old models, as shown in (Fig. 5.5). Evaluation of precipitation forecasts w done over the first 12-36 hour period in the form of threat and bias scores over the U.S. Improvement in the former can be seen, especially in the light amounts, while the bias is considerably better for all amounts, as shown in Fig. 5.6a and Fig. 5.6b. A useful extension of this threat and bias technique of evaluating precipitation forecasts can be obtained by directly applying it to windspeeds. This was done for 24-h forecasts of 200-hPa winds, with the results shown in Fig. 5.7a and Fig. 5.7b for four regions. Of particular interest for aviation are N. America and the Northern Hemisphere where for all windspeeds out to 60 m/sec the threat score in the new model(T170) is better than in the old (MRF). For the bias, the improvement is especially striking: on the order of half of the negative bias of the windspeed is removed. For example, for wind speeds over 50 m/sec over N. America, the MRF had a bias of .83, meaning that the area enclosed by the 50 m/sec isotach was too small by 17%. In the T170, the bias was about 0.92, too small by 8%. A limited sample of evaluations of model winds and temperatures against rawinsonde and aircraft observations was available. The results (not shown) confirmed the findings of a low-level warm bias over land, but also showed a reduced the operational cold bias in the tropical lower troposphere. The negative wind speed bias was largely removed in the upper troposphere, but at the price of some increase in the rms vector wind errors there, a result inconsistent with the verifications against analyses shown above in Fig. 5.3. This disagreement will be the subject of future research.
6.0 CHANGES IN MODEL OUTPUT FORMAT
Information about file changes is no longer available.
An extensive package of changes has been implemented in the MRF/GDAS system, including increases in the horizontal and vertical resolution, and improvements in the physics, data assimilation and analysis. The new model produces less precipitation over the U.S., substantially reducing the old model's wet bias. Improvements were also noted in precipitation threat scores here and in evaluations of tropical winds at 200 and 850 hPa and in the winds at 200 hPa in the extratropics when verified against analyses. The new model atmosphere was better able to maintain its transient eddy kinetic energy, but there was some apparent loss of skill of upper tropospheric winds where verified against observations. The new model is warmer in most areas correcting the old model's cold bias (however, the bias is overcorrected and a warm bias introduced in the lower atmosphere over land). For tropical disturbances, the new model damps out much of the low-level noise present in the old model, but only partially eliminates the problem of the generation of spurious disturbances.
8.0 LOOKING AHEAD
Due to contraints on computer resources, it was not possible to run two T170 systems in parallel for complete tuning of the large package of changes described in this TPB. When the installation and testing of a new Class 8 computer are complete early in 1999, we will be able to tune the new system more thoroughly. Our plans for the immediate future include:
9.0 POST-IMPLEMENTATION CHANGES
Because of problems found in the weeks following this implementation the following changes were made to the system:
On 24 June at 1200Z a correction was made to the analysis system to speed the rate of convergence of the iterations. This is expected to reduce the level of noise in both the analysis and the forecast]
Additional changes were made to the system on 1200Z Tues 21 July as follows:
Changes (1) and (2) were tested in a parallel system at T170 resolution, while the changes in (3) (which are not resolution-dependent) were tested at T62. Then, the entire package was tested in the T170 parallel system from July 17 through July 21. Sample maps from these last tests are available on the following website:
http://sgi62.wwb.noaa.gov:8080/STATS/maps/opnl/opnl.html or ../parl/parl.html
The map files are labeled in the form [mdd]f[hh].gif, where m=month, d=day, and hh= forecast hour, e.g., 717f12.gif is the 12-h forecast from 17 July. All forecasts originate at 00Z.