1. Motivation, Background and Goal
Short Range Ensemble Forecast(SREF) provides probabilistic forecast
meso-scale, and has undergone development by EMC/NCEP since 1996 (Du, et al, 2003). The SREF system
has been run operationally at NCEP for two years. The SREF Aviation project, a subset of the NCEP SREF
system, seeks to further improve the utility of SREF products by applying ensemble techniques to aviation
Basically, an ensemble
is composed of a set of models or "ensemble members" run at the
time. The members can be made from the same model using different initial/ boundaryconditions or physics
schemes, from a multi-model system, or from some combination of the above. The output from the various
ensemble members is used to perform statisticalcomputations to create probabilistic forecast information
such as an ensemble mean, spread (standard deviation), probability of occurence, etc.
The aviation weather
requires additional forecast products than are typically used for
weather forecasting. Aviation forecasting focuses on short time scales. Substantial efforts have been made
to improve the aviation weather forecast, such as utilizing more sophisticated meteorological models and
employing more complex algorithms to compute aviation weather variables such as icing and turbulence.
Traditionally, deterministic weather forecast approaches were used. It has been recognized that the
computer-generated weather system does not always result in a perfect forecast because of the chaotic
characteristics of the atmosphere. Thus, probabilistic or ensemble-based forecast from the SREF system
can provideadditionalinformation in terms of forecast confidence and uncertainty, and can be useful to
aviation weather forecasters.
Please keep in mind that
SREF-aviation products are still under development and very
As the SREF system is updated and SREF ensemble outputs are changed, SREF aviation products will be
improved or make use of new algorithms as well. Also, SREF aviation products will be upgraded based on user
feedback and requirements. We are trying to provide a useful, efficient, and user-friendly tool to aviation
Also please keep in mind that the
goal of SREF system is not only improve the forecast"preciseness",
but more important, to quantify the uncertainties in the initial conditions and the models or confidence interms
of a set of statistics based on the ensemble members. It gives out information on what the most likely outcome
The SREF documentation is located here. Additional ensemble training information can be found here.
2. Ensemble models and members
SREF runs at NCEP twice a day; first at 09Z,
at 21Z. The forecasts run out to 63 hours,and data is output
at 3-hour intervals.
The current SREF ensemble is made up of 15 members,
on 32km Eta model and 40km RSM models.
|| Betts-Miller-Janic (BMJ)
|| Control + 1 pair
|| Control + 1 pair
||BMJ-SAT (Saturated moisture profile)||
||KF-DET(Full cloud detrainment)
||Simple Arakawa Shubert (SAS)
|| Control + 1 pair
||Relaxed Arakawa Shubert (RAS)
While the current SREF system has 15 members, not
aviation product is made from all 15 SREF
members. The Eta and RSM have different post-processed data output, and don'talways contain the
variables needed for a particular product. The table below lists the variousaviation products and which
of the models/physics schemes were used to derive them:
3. Multiple initial conditions and boundary conditions
The current SREF system uses a breeding method to
different ensemble members. That is, using a
random perturbation field, we add or subtract from the original initial condition to create a pair of new perturbed
The boundary conditions come from NCEP's global ensemble output.
4. Model timelines, Data display, Output format and Product update times
The SREF-aviation production is divided into 3 parts:
(1) Computation: Copy SREF files and run models, 30 minutes total
09Z cycle: 9:30-10:00Z
21Z cycle: 21:30-22:00Z
(2) Image creation: Produce image (GIF) files from model output, about 30minutes
Send GIF files to the NCEP server, 30 minutes. This also includes
updating the html files' time stamp on the server
The total time used to produce
new forecast is one and half hours, finishing at about 11:00Z for
the 09Z run
and 23:00Z for the 21Z run.
The images are produced by
directly from output files in GRIB format. Each weather parameter
22 images so that the results can be animated. Thus, your browser needs to download all 22 of the GIF files before
a new forecast product can be displayed. Those who have low-speed Internet connections or less powerful
computers will notice some delay in viewing SREF products.
The SREF-aviation computation data are
in GRIB-extension format. The extension is NCEP's special
addition to regular GRIB format for storing unique ensemble data. Because of the limits to NCEP's on-line system
storage, SREF-aviation computation results are kept only 3 days. If users want these data, we can send them files
in GRIB-extension format only. A list of the weather parameters available in the SREF-aviation GRIB-extension
file can be found here. Background information on the NCEP GRIB-extension format is also available on our
SREF BUFR files (for specific locations) are not
5. Ensemble variables
variables used in the ensemble forecast to create the probabilistic
of mean, spread,
"spaghetti", and probability of occurrence. In SREF-aviation, only mean, spread and probability are used. The
"spaghetti plots" of individual ensemble members are not shown.
Definition-ensemble mean: The average value of all members.
The standard deviation with respect to ensemble mean. The spread of a
forecast diversity of the ensemble mean among all the members.
For an event, out of all members, in how many members does the event
of aevent (variable) is expressedas a percentage.
For example, a probability of
at FL180 equals to 60 % means that 9 of the 15 members, or 60% of the
(9/15 x 100 = 60 %), have icing at that level.
In the images, the mean is
by contours, spread by colored shading, and probability can be denoted
by contours or shading.
6. Definition and method of computation for each variable
(1) Icing (In-Flight) probability
In the real world, icing is a
in which ice particles or water droplets freeze on an aircraft,
the wings. The icing forecast predicts where the kind of environment favorable for icing events is likely to develop.
The icing environment depends
on temperature, humidity and cloud water content at certain levels.
are many algorithms to predict such an environment, ranging from the simple to the very complex. Since the current
RSM lacks cloud water content,we use a T-RH (temperature-relative humidity) algorithm to predict icing instead
of a method using cloud water. Thus, current icing products can't distinguish between frozen ice and supercooled
large droplets (SLD). The new version of RSM, which has cloud water content in its post-processed output, will
be soon be implemented at NCEP. Then we can use more sophisticated icing algorithms.
The range of T and RH for icing event is:
-10 C < T < 0 C and RH > 70%
Icing is computed at 8 flight levels, from low to high:
FL030: 3000 ft (900mb)
FL060: 6000 ft (800mb)
FL090: 9000 ft (725mb)
FL120: 12000 ft (650mb)
FL150: 15000 ft (575mb)
FL180: 18000 ft (500mb)
FL240: 24000 ft (400mb)
and model gridpoints, if temperature and relative humidity both meet
icing criteria above,
then the icing condition is "yes". If M ensemble members out of all N members predict icing conditions at the same
level and gridpoint, then the icing probability is
M/N x100 %
(2) Clear Air Turbulence (CAT) Probability
CAT is defined as a
event occuring at high, cloudless altitudes. Like icing, there
morethan a dozen
methods to predict the CAT intensities. The SREF aviation turbulence product uses the Ellrod(Gary P. Ellrod, 1992)
The Ellrod algorithm
CAT intensity into 3 categories: LIGHT to MODERATE, MODERATE, and
MODERATE to SEVERE, and uses a turbulence intensity index to determine which categorythe event belongs to.
The Ellrod turbulence
index is defined as a function of stretching deformation (DST),
tion (DSH), vertical wind shear (VWS) and convergence (CVG). These four factors implicitly reflect the effects of
wind gradients in both the horizontal and vertical directions, and the temperature gradient as well. In detail,
Index = VWS * [ DEF + CVG ]
VWS = sqrt(dU*dU + dV*dV) / dz, where U , V are wind u,v components (in m/s), respectively
DEF = sqrt(DST*DST + DSH*DSH)
DST = dU/dx - dV/dy
DSH = dV/dx + dU/dy
CVG = - (dU/dx + dV/dy)
As pointed out by
defining the 3-category index values for different CATintensities
For different horizontal resolutions, Ellrodsuggested using different index values. For the SREF-aviation CAT, all
computations use the following index values:
the turbulence intensity index is computed between 2 levels, i.e.
a layer instead of at a
single level. In SREF-aviation products, the turbulence intensity is evaluated within 8 continuous layers, from low
FL240-FL210 or 400-450mb
FL270-FL240 or 350-400mb
FL300-FL270 or 300-350mb
FL330-FL300 or 275-300mb
FL360-FL330 or 225-275mb
FL390-FL360 or 200-225mb
FL420-FL390 or 175-200mb
(3) Sky (Cloud) Cover Probability
Sky or cloud
is divided into 4 categories, clear, scattered, broken, and overcast.
category has a
range of cloud amount in fractions or percentages, as listed below:
Eta and RSM models
cloud cover as a percentage instead of a category or
products use the seamless range of percentages shown in theabove table. So the probability of a particular cloud
cover category is the percentage of ensemble members predicting a cloud cover percentage that falls in the
category. For example, when the probability of a clear sky at one gridpoint is 50%, that means half the ensemble
members predict clear (0% cloud cover) at that gridpoint.
(4)Flight Restriction (LIFR, IFR, MVFR, VFR) probability
LIFR: Low Instrument Flight
IFR: Instrument Flight Rules
MVFR: Marginal Visual Flight Rules
VFR: Visual Flight Rules
Restriction categories are defined by cloud base
(ceiling), surface visibility and cloud amount category, as
Here is the pseudo code that describes the computation:
For all members N, find how
cloud amounts are broken and overcast, say M.
VFR probability = N/N=100 %, i.e. no broken and overcast clouds
(N-M) members belong to VFR, some of M belong to either LIFR, or IFR, or MVFR, or VFR,
Do the following:
Within M, find how many members L1 with cloud base <500 ft or Visb <1 mile,
LIFR probability = L1/N
Within M, find how many members L2 with 500<=cloudbase<1000 ft or 1<=Visb<3 mile,
IFR probability = L2/N
Within M, find how many members L3 with 1000<=cloudbase<3000 ft or 3<=Visb<5 mile,
MVFR probability = L3/N
Then the Left in M, i.e. M-L1-L2-L3 members are belong to VFR,
VFR probability = [(M-L1-L2-L3) + (N-M)] / N = (N-L1-L2-L3) / N
Note that in the computation, "AND/OR" means
"AND" or "OR". When using "AND" the evaluation is
more conservative. We use "OR" in our product.
Ceiling and visibility are also computed and listed for reference. (See sections 9 & 10).
(5) Jet Stream probability
Jet stream probability is computed at 3
4,500 ft, 18,000 ft and 34,000 ft with 3 wind speed thresholds at
At 4,500 ft, the
are 20, 40, 60 knots
At 18,000 and 34,000 ft the thresholds are 60, 80, 100 knots
e.g., at 18,000 ft
a threshold of 80 knots, the image will show the probability of wind
speed > 80 knots at 18,000 ft.
height: mean and spread in feet
(II) Temperature at tropopause height level: mean and spread in C
The first height above the ground where air temperature is below 0 C: mean and height spread.
(8)Total Cloud cover
(I) Mean and
averaged cloud total amount (%) and spread (%).
example: cloud contour = 80 and shade color is yellow, which indicates that the average cloud amount is
80%, and standard deviation (diversity among members) is 20 %
(II) Max cloud amount: maximum cloud amount from all ensemble members.
(III) Min cloud amount: minimum cloud amount from all ensemble members.
example: Min cloud amount at a region = 60%, which indicates the
ensemble mean cloud coverage for
this region is AT LEAST 60 %.
the horizontal visibility at the surface. Visibility is not a model
but is computed from post
processing output, based on surface humidity, rain, etc. Two notes:
(a) Eta and RSM do
consider the haze directly, but haze forms in high humidity. Thus
implicitly include low visibility caused by haze.
(b) Dust storms
be detected in the Eta or RSM.
defined as the height of the lowest layer of cloud when the sky cover
overcast; the mean and spread are computed in feet. If no cloud or sky cover is less 50%, then there is no
ceiling, then its value is defined as model default 14200 m or 46580 feet. The model cloud is defined at each
grid point and level and is not continuously distributed in the vertical. The cloud base height is computed
from the lowest level of cloud at each gridpoint in each ensemble member. But after averaging all the
ensemble members, the cloud base may no longer correspond exactly to any of the model levels.
The height of the highest layer of cloud; but currently not available from the ensemblemodels.
SREF is measured by the model convective cloud amount. Convective
cloud is created each
ensemble member from its convective scheme (BMJ, KF, or SAS). This definition is not the standard used in
aviation, where convection is defined as the mass convergence ordivergence. In the SREF product, the location of
convection is expressed in terms of convective cloudamount mean and spread.
(13)Convection Speed and direction
Defined as convective storm's
field. Mean (contour), spread (filled) and vector flow
plotted in one picture.
(14)Low level wind shear (LLWS)
LLWS here is defined by
the vector change in the layer between the surface and 2000 feet level.
Since winds are
defined at fixed pressure levels with 25mb intervals, the winds at different location may be at different pressure levels.
Furthermore, 2000feet level will not just at a pressure level. Therefore, to compute the LLWS, we must first find the
2000 feet level at each grid, then using 10m wind and the 2000 feet wind to get the wind change.
First serch 2000 feet + 10m height (ie 2030
feet) from the surface upward level by level, after find it, say
between level-5 and level-6,
then compute the U and V components of the wind at this level by linear interpolation of level-5 wind and level-6 wind.
Second step is computing the wind vector change over 10m and 2030 feet. The LLWS is defined as
LLWS = SQRT [ (U2030
- U10)2 + (V2000
- V10)2 ]
Where 2030 means 2030 feet, 10 means 10
The LLWS mean and spread thus can be obtained from
15 ensemble members.
The severe LLWS is defined by (1) LLWS
> 20 knotes over 2000 feet, or
(2) LLWS > 0.16 over any 200 feet layer withn the 2000 feet
The probability of severe LLWS is also computed, but please note that, the pressure level intervals are
fixed 25 mb, or 600 feet, much larger than 200 feet,
so, in current case, only case(1) is the threshold for severe LLWS here.
Probabilities of three kinds of precipitation are defined: rain, snow and freezing rain are calculated.
(16)Surface wind speed and direction at 10 m
The mean of surface wind
speed is computed from average of total wind speeds Wi
Wi = SQRT(Ui2+Vi2)
The products include
Wind speed mean, spread (wind direction is also depicted in the same image)
The probability for wind speed exceeding 10 knotes
The probability for wind speed exceeding 20 knotes
The probability for wind speed exceeding 30 knotes
The calculations of the mean
and spread of wind direction(North is reference) are a little bit
complex. First, compute
the mean and spread of U and V components respectively, then the mean and spread of wind direction can be expressed as:
mean of wind direction = arctg ( Umean /
spread of wind direction = arctg (Wspread /Wmean)
Wspread = SQRT ( Uspread2 + Vspread2)
Wmean = SQRT ( Umean2 + Vmean2)
(17) Fog occurence probability
Fog is detected at a grid point in an ensemble model if following cloud base and
cloud top threshold is satisfied:
Cloud base < 10 m And Cloud top < 400 m
In general, the top of ground fog or radiation fog is less than 200m which is
mainly controlled by surface cooling and weak turbulence, while sea fog or advection
fog is wind-driven and can be thicker than radiation fog. So 400 m is used for the
cloud top threshold.
Currently, low level wind shear product has been verified, and showing
that, the skill decreases with the
forecast time, but it i sstillful uptime 63 hours. Please see here fore detail.
The verification of other SREF aviation products has not yet been
observed data, either service
or pilot reports, are not available to NCEP. We are planning to work with AWC and Aviation Service Branch to do
this work. We will also cooperate with FSL using its RTVS (Real Time Verification System) to verify some SREF
can be done, probabilistic forecastvariables (mean, spread, and
converted from the NCEP GRIB-extension format into a regular data format such asGRIB and BUFR. Until that
happens, there is no simple way to store the ensemble data in a standard, easy to use format. AWC, RTVS,
Aviation Service Branch and EMC are working togetherto solve this problem.
8. Plans for the Near-Future
system has been initiated and has been making good progress. But
it is still in the deve-
lopment phase. It will be updated along with the SREF system. Here are some upcoming improvements to the
(1) Add Alaska domain.
(2) Update ensemble members to include more,
physics schemes. Currently, breeding pertubations for
initial conditions is the main method used to differentiate between the ensemble members. An ensemble with
many different physics schemes and pertubations will create further diversity among the different members, and a
larger spread in results. Our study already showed better skill than with a breeding-only ensemble.
(3) Increase Eta and RSM resolution from the current operational 48 km to 32 km.
9. Existing issues
(1) Ensemble members
There are several possible
to create different ensemble members, with the goal of creating the most
diversity among the ensemble members. We are still experimenting with different techniques to see which will
result in the highest skill.
(2) Variables and Algorithms
(i) Icing is currently calculated using the
T-Rh algorithm. We will switch to a more complex one if the
basic data is available and computing resources permit. Our goal is not to investigate which algorithm is best,
but to try instead to use the ensemble technique for aviation.
(ii) Ellrod turbulence also is relatively simple algorithm, which may be a candidate for replacement.
(iii) SREF aviation products are created from
forecast variables, which aren't always the same as those
used in aviation. For example, in aviation convection is defined differently than in the usual meteorological
(iv) Convection speed and direction are
an issue. In SREF, convection speed and direction are computed
from storm U-V fields. They might not match the actual convection speed and direction.
(v) Vertical wind shear is
pressure levels of 1000, 950, 900 mb. For mountain regions like the
states, we can't compute wind shear on pressure surfaces that intersect or are below the surface.
perform verification on SREF aviation products is most important
problem. The current
difficulty is output files in NCEP's GRIB format extension, not the standard GRIB format. RTVS and AWC can't
read the NCEP format extension & can't help with verification. We are working on this issue. Storing probabilistic
data in the BUFR format is also a problem.
(4) Ensemble domain restricted to CONUS; Alaska domain must be added.
(5) Probabilistic data format
products are grid-based and not available for specific cities or
since the single station
post-processing uses BUFR files (not currently output by SREF). If some users want specific stations, please
contact us. We can do it for you separately.
(6) 90-minute update time may be too slow to meet the requirement of aviation users.
(7) Machine production problems
Because SREF-aviation products are still in the testing phase they are
not NCEP "official" products. By
NCEP policy, the products must run on the NCEP development machine instead of the production machine.
Since the recent upgrade, the development machine is often unstable goes down from time to time. When that
happens, the SREF cron jobs will be re-run manually, which will delay product delivery.
10. Comments and Feedback
We welcome any
suggestions and feedback from users. If users have new requests,
we will satisfy
them if we can.
1. Binbin Zhou, et al,
2004: An introduction to NCEP Aviation Project,
Preprints, AMS 11th Conference on Aviation,Range and Areospace, Oct 4-8, Hyannis, MA, Amer Meteor. Soc.,
2. Binbin Zhou, et al,2005: Ensemble Forecast and Verification of Low Level Wind Shear by
NCEP Short Range Ensemble Forecast (SREF) System
Preprints,AMS 17th Conference on NWP, August, 2005, Washington DC, Amer. Meteor. Soc.,
3. Federal Meteorological handbook, No. 1 (FMH-1),
4. NWS Instruction 10-813 of TAF, 2004
J. et al, 2004: The NOAA/NWS/NCEP short-range ensemble forecast (SREF) system: Acknoledgement Following persons are thanked for their technical
discussions and valuable comments .
Fred Mosher, AWC/NCEP?NOAA
Also thanks to Mary Hart for her review and editing
Evaluation of an initial condition vs multi-model physics ensemble approach.
Preprints,AMS 16th Conference on Numerical Weather Prediction, Seattle, Washington, Amer. Meteor. Soc.,
Steven Silberberg, AWC/NCEP/NOAA
Mark Andrews, Aviation Service Branch, NWS//NOAA
Michael Graf, Aviation Service Branch, NWS/NOAA
Jennifer Mahoney, FSL/NOAA
Kevin Baker, ?/NOAA
Kelvin L JohnStone, NWS/NOAA
Following persons are thanked for their technical support, discussions and valuable comments .
Fred Mosher, AWC/NCEP?NOAA
Also thanks to Mary Hart for her review and editing this document