Operational Implementation of the WRFVar System

1
Operational Implementation of the WRF-Var System
and Typhoon Data Assimilation
National Center for Atmospheric Research
(NCAR) Central Weather Bureau, Taiwan Presentatio
n for 2006 NCAR-CWB project annual review 29
November 2006
2
Tasks for 2006 CWB project
Task-1. Establish WRF-Var/NFS operational system
on CWBs newly procured computer A. Successfully
compile all of the system codes, and improve the
efficiency to meet the operational
requirement B. Increase of the WRF-Var/NFS model
resolution to 15-km for improving typhoon
forecast C. Derivation of the new background
error statistics (cv5) based on the CWB NFS
forecast data D. Develop WRF-Var-based
observation verification package Task-2.
Enhancement of the WRF-Var System A.
Assimilation of more observations, such as
QuikSCAT, AWS, and GPSRO data in BUFR format B.
Test the WRF-Var FGAT technique with asynopic
observations C. Assimilation of ground-based GPS
PW data Task-3. Continued Interaction on WRF Var
http//box.mmm.ucar.edu/people/guo/individual_guo/
CWB/CWB_Project_2006.html
3
TASK1Establish WRF-Var/NFS operational system on
CWB IBM computer

• Successfully compile all of WRF-Var/NFS codes in
  CWB IBM computer and improve the efficiency to
  meet the operational requirements
• This task was most completed by CWB staff
• The new machine is IBM p5-575 Cluster 1600 with
  2496 CPUs (156 nodes, 13 Frames), and the machine
  has SMT Dual core, Super Architecture, peak of
  6GFlops.
• NCAR provided the consultants and updated the
  OBS_FGGE_PROC, NFS2WRF, 3DVAR_OBSPROC, WRF-Var,
  and WRF2NFS programs. Now all of these programs
  are running OK in CWB new IBM machine
  (communication with Eric Chiang). The flow chat
  of the WRF-Var/NFS system is shown the next
  slide, and the yellow boxes indicate the
  update/new-developments in year of 2006.
• CWB staffs have conducted many kinds of tests,
  including
• the monthly pre-operational run from 15 August
  to 15 September 2003
• typhoon track forecast for Dujuan, Krovan,
  Maemir(2003), Haitang (2005), Bilis, Kaemi,
  Shanshan (2006).

4
WRF-Var/NFS System in CWB new IBM Computer
NCAR provide Default BES and Software to derive
BES
From CWB
COSMIC/CDAAC
CWB_FGGE data Decoder QuikSCAT AWS,mesonet
CWB Bogus data Decoder
CWB First guess (s) (6-h forecast) to WRF (h)
Converter
Background Error Statistics (BES) Derivation
COSMIC GPSRO BUFR Decoder OBS error
WRF-Var(3D-Var)
CV_options5 special for CWB Seasonal (summer
and winter) BES High resolution BES Multiple
domain (45/15-km) BES Tuning and testing
CWB Ground-based GPSPW
WRF 3D-Var netCDF (h) to CWB DMS (s) Converter
CWB NFS Model
6-h forecast cycling
Verification package (CAA)
Yellow boxes indicate the new developments
in year of 2006
48-h forecast
5
The monthly pre-operational run from 15 August to
15 September 2003
48hr?? 3dvar??? ???H?????, ??RC02???
(Adopted from CWBs results)
6
Typhoon Haitang forecast experiments
With 15-km resolution initial time at 2005071600Z
45-km
15-km
OP operational, 3DV 3D-Var, ob black
(Adopted from CWBs results)
7

• Increase of the WRF-Var/NFS model resolution to
  15-km for improving typhoon forecast
• In addition to WRF-Var/NFS Exps in CWB, NCAR
  staff have done many WRF-Var/WRF experiments in
  NCAR IBM (power-5, bluevista) computer for
  Typhoon Dujuan (2003) and Haitang (2005).
• The experiments are 45-/15-km 6-h cycling runs.
• We conducted the Exps with and without CHAMP
  GPSRO data assimilation

Table1. The 24h Average Increment of Track error
( without GPSRO with GPSRO ) for Typhoon
Dujuan (August 2003) (km)
Addition of assimilating the GPS RO soundings
with all other available observations improved
the Typhoon Dujuan track forecast (5 7km)
during the 3days cycling period (2003082812Z to
2003083112Z).
8
The track and central pressure forecast error
increments (NCEPAVN-3DVar) Positive ?3DVar
better, Negative ?NCEPAVN better
Table2. The 24h Average Increment of Track error
(km) for Dujuan
Table3. The 24h Average Increment of
Pressure error (hp) for Dujuan

This improvements may most be from the BOGUS data
assimilation.
9
Forecast experiments with and without WRF 3DVar
for Typhoon Haitang (200507)coldst_si
initiated from NCEP AVN, coldst_cv5
initiated from WRF 3DVar with cv5 BES
2005071600Z
2005071612Z
2005071700Z
Central pressure forecast
10
Haitang Track forecast errors for different
initial times
From experiments on both Typhoon Dujuan (200308)
and Haitang (200507), the forecast of the Typhoon
track and intensities (central pressure) with the
WRF 3DVar initialization are significantly
improved over the initialization directly from
NCEP AVN analysis.
11

• Derivation of the new background error statistics
  (BES, cv5) based on the CWB NFS forecast data
• The new background error statistics files are
  derived based on the NFS forecast datasets
• Winter 2005010100Z to 2005013112Z 6-h
  cycling WRF-Var(cv3)/NFS run
• Summer 2003081500Z to 2003091512Z 6-h
  WRF-Var(cv3)/NFS cycling run
• For CWB WRF domains, the BES files are
  generated by Eric Chiang and
• For the NCAR WRF domains, they are generated
  by Hui-Chuan Lin using the gen_be code
  in WRF-Var with bin_type1.
• The BES interpolation capability was
  developed in WRF-Var 2.1 code by NCAR staff.
• The interpolated BES could be regarded as the
  first approximation of the Background Error
  Statistics estimates used to run the WRF-Var, not
  need to derive the BES by using gen_be code
  based on a period of the forecast data prior to
  run WRF-Var.
• The some of the BES tuning experiments were
  conducted for Typhoon Haitang by CWB
• and NCAR staff.

12
SOUND RMS errors for 72-h forecast for wind,
temperature, and specific humidity With the cv3
BES, CV5 BES, and CV5 BES with tuning factor
of 0.5
The CV5 BES with the tuning factor of 0.5 reduced
the forecast errors
(Adopted from Eric Chiang, CWB)
13
In NCAR, more experiments were conducted for
Typhoon haitang with the different tuning
technique for background error statistics, i.e.
multiple outer loops with the different tuning
factors of variances and scale-lengths for CV5
BES. This technique gave the improved typhoon
analysis and forecast for both track and central
pressure. The details can be found from the
extended abstract downloaded from
http//www.mmm.ucar.edu/events/2006wrfusers/agend
a.php by clik P4.2 Or see http//box.mmm.ucar.edu/
people/guo/individual_guo/CWB/CWB_Project_2006.htm
l under the section D, Typhoon Haitang, click
WRF2006_POSTER_guo. The experiment design and
results for this study are briefly shown in the
next 3 slides.
14
Forecast experiments with WRF-Var/WRF

• Exp1 -- Cold_SI Cold-start initiated by WRF_SI
  based on the NCEP GFS analysis
• Exp2 -- C3DVCV3 Warm-start (cycling) runs
  initiated by WRF-Var with CV3 BES
• Exp3 -- C3DVCV5 Warm-start (cycling) runs
  initiated by WRF-Var with CV5 BES
• Exp4 -- C3DVCV5E3 Warm-start (cycling) runs
  initiated by 3 External loops WRF-Var with 3
  different tuning factors to CV5 BES for both
  domains
• Exp5 -- C3DVCV5E2 Warm-start (cycling) run
  initiated by 1 External loop WRF-Var for domain1
  (45-km) and 2 External loops with different
  tuning factors to CV5 BES for domain2 (15-km)

15
Tuning factors for Exp.4 and Exp.5

• Exp4 -- 3DVCV5E3 CV5 BES with 3 sets of tuning
  factors for 3 outer loops as below,
• VAR_SCALING1 1.50, 1.00, 0.50, (y)
• VAR_SCALING2 1.50, 1.00, 0.50, (c_u)
• VAR_SCALING3 1.50, 1.00, 0.50, (T_u)
• VAR_SCALING4 1.00, 1.00, 0.50, (rh)
• VAR_SCALING5 1.50, 1.00, 0.50, (Psfc_u)
• Larger Small
• scale scale
• LEN_SCALING1 1.00, 0.50, 0.25, (y)
• LEN_SCALING2 1.00, 0.50, 0.25, (c_u)
• LEN_SCALING3 1.00, 0.50, 0.25, (T_u)
• LEN_SCALING4 1.00, 0.50, 0.50, (rh)
• LEN_SCALING5 1.00, 0.50, 0.20, (Psfc_u)
• Exp5 -- 3DVCV5E2 CV5 BES with 2 sets of tuning
  factors for 2 outer loops. The second set of the
  tuning factors was removed.

16
Track forecast errors during the 72-h forecast
period
17

• Develop WRF-Var-based observation verification
  package

• Modify the wrfvar code for the special option
• ANALYSIS_TYPE QC-OBS
• Modify the wrfvar code for the special option
• ANALYSIS_TYPE VERIFY
• Develop the utility program to split the WRF
  model outputted historic files
• Develop an utility program to process the wrfvar
  output statistics files
• Develop a flexible shell script

• Advantages
• A variety of the observations consistently
  quality-controlled (filtered) by WRF-Var (wrfvar)
  code
• The verifications can be done not only against
  the conventional data but also the
  non-conventional data
• The exactly same observations are verified
  against for all the experiments
• It is easy to be used for the comparison of the
  results from the different experiments.

18
TASK2 Enhancement of the WRF-Var System

• Assimilation of more observations, such as
  QuikSCAT, AWS, and GPSRO data in BUFR format
• QuikSCAT data in CWB operational database
• OBS_FGGE_PROC convert it to LITTLE_R file, wind
  speed and direction errors stored in the fields
  of the U- and V-component.
• The minimum wind speed errors are allowed to be
  1.0 m/s.

Examples of the QuikSCAT data during Typhoon
Haitang period
2005071520Z
2005071608Z
2005071621Z
19
AWS and mesonet data in Taiwan area Decoder
program was added to OBS_FGGE_PROC In LITTLE_R
file, they are assigned as the type of SYNOP.
These mesoscale surface observations have the
high temporal resolution. With the FGAT
technique, there are 587 SYNOP data assimilated
within 15-km domain, but only 158 SYNOP data used
in normal assimilation.
20
GPSRO BUFR format data As LEO satellites have
been launched on April 15, 2006 with
FORMOSAT-3/COSMIC project, the GPSRO soundings
will be available for operational use in BUFR
format. The procedure of GPSRO BUFR data
processing
21
NCAR staff has used the GPSRO data for Hurricane
Florence forecast to assess its impact because
there are number of the GPSRO soundings available
for this case. The work of experiment for Typhoon
Shanshan is under going.
Black obs, Red NCAR real-time, Green GPSRO
22
Track forecast errors at the different initial
times for Florence
Red numbers indicate the smaller errors between
Exp.RT and 3DVAR. 3DVAR with GPSRO assimilated
improved the track forecast.
23

• Test the WRF-Var FGAT technique with asynopic
  observations
• For all types of the asynoptic observations,
  QuikSCAT, AWS, and GPSRO, etc., the FGAT
  technique in WRF-Var may be more suitable for
  assimilation. Several experiments have been
  conducted in NCAR for Typhoon Haitang.
• In these experiments, 12 types of observations,
  SYNOP, METAR, SHIP, BUOY, SOUND, AIREP, PILOT,
  SATOB, GPSREF, QSCAT, SATEM, and BOGUS, are
  assimilated.
• Cold-start with SIWRF-Var at 2005071400Z, then
  for each of 6-h cycle, 7 hourly forecast from 3h
  to 9h from the previous cycle are used as the
  first guesses.

24
With the FGAT technique, much more
non-conventional observations, such as SATOB,
GPSREF, SATEM, QSCAT, and AWS, are ingested by
WRF-Var. From the Table below, FGAT technique, in
general, gave the improved Typhoon track forecast
skill except the initial time of 2005071612Z.
25

• Assimilation of ground-based GPS PW data
• There 107 ground-based GPS sites in Taiwan GPS
  list.
• To retrieve the PW data, the surface
  meteorological parameters, pressure and
  temperature must be available. So there are about
  50 PW data available in interval of 30 minutes.
• At 2005071612Z, there are 51 PW data in Taiwan,
  but there are 88 ZTD data over west Pacific,
  Taiwan, and mainland China area.

26
CWB provided the one-month (July 2005) raw
ground-based GPS data. UCAR/COSMIC processed the
data with software Bernes to get PW/ZTD text
file. NCAR/MMM developed the decoder program to
generate the LITTLE_R file as input to WRF-Var
system. Eric Chiang (CWB) visited NCAR in May to
October 2005 conducted the GPS PW assimilation
experiments for Typhoon Haitang from 2005071400Z
to 2005072012Z. The RMS errors verifying against
the observed PW are improved by 1.8.
(0.777-0.763)/0.777 1.8 (adopted from Chiang,
CWB)
Details of the results can be found from NCAR-CWB
web page http//box.mmm.ucar.edu/people/guo/indiv
idual_guo/CWB/CWB_Project_2006.html under the
section B, Progresses, 2) Status of 2006 CWB
project
27
TASK3 Continued Interaction on WRF Var

• Update and improve the CWB project web pages on
  both NCAR/NCAR and CWB sides
• The WRF-Var system and related codes are
  frequently updated on web page
• http//box.mmm.ucar.edu/people/guo/individual_guo
  /CWB/CWB_Project_2006.html
• The bug fixes, new developments, and the reports
  of the progresses all are posted on this web site
  timely.
• The standard version of all the codes are
  provided to CWB staffs.
• Two main updates completed recently
• 1) A beta version of WRF-Var
• It should be noted that WRF and WRF-Var need to
  be compatible in terms of versions. Therefore,
  WRF-Var 2.2 can accept first guess fields from
  both WRF2.2 or WRF2.1. However, the output from
  WRF-Var 2.2 can only be used to initialize WRF2.2
  (not WRF2.1).
• 2) The 3DVAR_OBSPROC was modified to
  produce both ASCII and NCEP prebufr format
  observation files, and the GPSRO data can be
  used for both local and non-local operators.

28

• Site visit
• Eric Chiang from CWB vested to NCAR for six
  months, and work successfully with NCAR staff.
  His work mainly focused on the Background error
  statistic tuning and GPSPW assimilation.

29

• END
• THANK YOU !