Comparison of RAMS results to those of other models
Results from RAMS on the previous page showed
a prediction of weather for Texas for January 13, which was one of many
forecasts that were produced at the AMS 99 conference. Results from
other models for the local area were also presented at AMS on an operational
basis. Chief among them was the Advanced
Regional Prediction System (ARPS), developed by the Center
for Analysis and Prediction of Storms (CAPS) at the University
of Oklahoma. ARPS was run remotely from the conference at the
National Center for Supercomputer Applications
(NCSA) with results posted
on the web. Operationally, CAPS utilized a large supercomputer (128-cpu
SGI Origin 2000) at NCSA
to generate one 36-hour forecast at 32 km resolution (initialized at 0Z)
and another 12-hour forecast (at 12Z) each day for the continental United
States. A 5-member ensemble of 32-km forecasts (initialized at 0Z)
were also produced each day. A nest for a 15-hour forecast at 9 km
resolution was produced once each day (typically initialized at 15Z).
Another nest was defined for 6-hour forecasts at 3 km resolution produced
twice each day. The outer nest was initialized with the results of
an assimilation process (ARPS
Data Analysis System), which is somewhat similar to LAPS. All
of the results were posted as a set of static two-dimensional graphical
images on the web. Although the CAPS-NCSA collaboration resulted
in an impressive capability, it is in stark contrast to what was provided
by IBM at the AMS conference. RAMS was run locally at the conference,
not remotely, on a small inexpensive supercomputer (IBM
SP), which in fact had an obsolete configuration (six 120 MHz P2SC
thin nodes for computation and one 135 MHz P2SC wide node for I/O).
Although that machine when new cost a small fraction of the system employed
at NCSA, a current technology system with the equivalent computational
power is even cheaper (i.e., current nodes are over 2.5 times faster, yet
cost less). In addition, highly interactive, two-, 2-1/2- and three-dimensional
visualizations as well as animations of RAMS, including computation tracking
were available, as opposed to static plots of fixed content.
In an attempt to assess the differences between these two operational
forecasting efforts, a qualitative comparison of RAMS results at 8 km resolution
with ARPS results at 3 and 9 km resolution is discussed below. The
comparison is limited at best for a number of reasons, including:
-
Different domains are used in each forecast (resolution [grid spacing horizontally
and vertically], extent [thrtee-dimensional mesh size], centroid [center
of the domain], and projection [cartographic coordinate system])
-
Weather variables in the plots from ARPS are expressed in different parameters
and/or units and are laid out in physical distance coordinates rather than
geographic.
-
Each forecast has a different initialization time
-
Observations were not available for initialization of RAMS. The same
data source (ETA-32) had to be use for both boundary and initial conditions.
-
The only available information on ARPS results are static plots.
There is no access to actual data that could be compared. In addition,
only limited time steps common across the various forecasts were available.
(This was not a problem for RAMS since all of the data were available.)
-
Simultaneous animation cannot be used as another comparison technique because
each set of hourly plots from a particular ARPS forecasts uses different
scales for contours, color, etc.
-
Observations or even plots of the various parameters were unavailable to
assess the relative quality of the forecasts let alone model output statistics
from either model.
For each of the two sets of comparisons below, the ARPS results for each
time step from each forecast are presented as a set of three four-panel
plots. After each plot from a 9 km resolution forecast, is a set
of images produced by the RAMS slicer as a rough approximation of the some
of the same content of those plots. Given the aforementioned caveats,
the comparison is far from exact. For the first comparison, the RAMS
results are followed by results from a 3 km resolution ARPS forecast.
In each case, the RAMS results are from the forecast discussed on the previous
page. On any of the images from RAMS
data below, a higher-resolution version can be viewed by clicking on each
one.
The first comparison uses ARPS results from a 15-hour forecast at 9
km resolution and a 6-hour forecast at 3 km resolution, both initialized
at 18Z, and a RAMS run at 8 km initialized at 0Z on 13 January. The
output from 0Z on 13 January is used for comparison. Since this is
the initial time step from the RAMS run, conclusions that can be drawn
will be limited since 32 km ETA was used for both initial and boundary
conditions.
The general distribution of surface Td
(dew point temperature) is similar between RAMS and ARPS. However,
the values are much lower from ARPS in the northwestern portions of the
RAMS domain (i.e., upper elevations and drier) than from RAMS. Surface
T (temperature) and MSLP (mean sea level pressure) have similar distributions
and values. The structure and definition of the front that was observed
to pass through the domain was more clearly defined in the wind data from
RAMS than from ARPS.
The general distribution of LI (lifted index [temperature]) is somewhat
similar between RAMS and ARPS. qe
(equivalent potential temperature) has much smaller range and more uniform
structure than that of ARPS. Since predicted reflectivity is about
0 across the domain from both models, there is no basis for comparison.
The general distribution of height is quite different at each of the
four pressure levels between RAMS and ARPS. However, the dynamic
range at each pressure level is very similar. (Note: the values
shown in the ARPS plots are in decameters while the results in the RAMS
plots are in meters.) There is very little correspondence in wind
velocity between the two sets of plots.
The second comparison uses ARPS results from a 15-hour forecast at 9
km resolution initialized at 8Z, and a RAMS run at 8 km initialized at
0Z on 13 January. The output from 15Z on 13 January is used for comparison.
The general distribution of surface Td
(dew point temperature) is similar between RAMS and ARPS. Surface
T (temperature) and MSLP (mean sea level pressure) are also roughly the
same in distributions and magnitude. The structure and definition
of the front that was observed to pass through the domain was more clearly
defined in the wind data from RAMS than from ARPS.
The general distribution of LI (lifted index [temperature]) and qe
(equivalent potential temperature) are somewhat similar between RAMS and
ARPS. In both sets of plots, there is an indication of a gradient
in the respective fields, illustrattive of the location of a front.
That location is different between RAMS and ARPS, with the latter showing
it more to the southeast, almost as if there was a temporal bias between
the sets of results. Other results from RAMS shows the front passing
through the same area, several hours later, which appears to be more consistent
with some of the available observations.
There is no agreement in predicted reflectivity from both models.
The general distribution of height is quite different at each of the
four pressure levels between RAMS and ARPS. However, the dynamic
range at each pressure level is very similar. (Note: the values
shown in the ARPS plots are in decameters while the results in the RAMS
plots are in meters.) There is very little correspondence in wind
velocity between the two sets of plots.
Given the aforementioned caveats, it is difficult to draw very specific
conclusions in trying to compare the forecasts produced by RAMS and ARPS
at AMS 99. One can make a few observations with a grain of salt.
Greater forecasting skill should be expected from ARPS given the significant
investment in data assimilation for initialization and the vastly greater
amount of computing power dedicated to running multiple nests. Surprisingly,
this has not been proven since some of the details of the front
that passed through the areas appear to be better captured by RAMS (surface
results), in what is arguably a very small computer configuration without
complete initialization assembled in a rather inexpensive fashion in a
few days. Several other surface parameters show rough agreement despite
the differences in applied resources. The large differences in the
upper air data really cannot be addressed without actual observations to
compare with numerical results from both models.
lloydt@watson.ibm.com
