My Background, Previous Work, and Future Plan

1
Experiments of Hurricane Initialization with WRF
Variational Data Assimilation System
Qingnong Xiao NCAR/MMM, Boulder, CO
80307-3000 _________________________________ Ackn
owledgment Xiaoyan Zhang, James Done, Zhiquan
Liu, Wei Wang, Chris Davis, Jimy Dudhia, and Greg
Holland
2
Introduction

• WRF Weather Research and Forecasting (WRF) Model
• Developed by NCAR, NCEP, and several US
  universities and DOD labs.
• Two cores
• ARW - Advanced Research WRF, led by NCAR and the
  university community
• NMM - Nonhydrostatic Mesoscale model, led by NCEP
  and in operational application
• WRF-Var WRF Variational (WRF-Var) Data
  Assimilation System

3
(No Transcript)
4
Why WRF hurricane initialization?

• WRF ARW improved track and intensity over
  official forecast beyond 36 h.
• Short-term forecasts (lt 2 days) show a rather
  poor skills in WRF ARW, due to model spin-up
  problem.
• An improved hurricane initialization, using
  advanced data assimilation technique, can augment
  the skills of short-term forecasts.

WRF hurricane forecast in 2005 (Orange), Davis et
al. 2008
5
Why WRF-Var for hurricane initialization?

• WRF-Var is an advanced data assimilation system
  based on the variational technique.
• It includes WRF 3D-Var, 4D-Var, and
  ensemble/variational hybrid (En3D-Var, En4D-Var).

• It can assimilate all observational data,
  including satellite and radar data.
• It is robust, and facilitates research and
  real-time applications.

6
WRF-Var data assimilation system
Background constraint (Jb)
Observation constraint (Jo)
obs

• xb model background (former information)
• H(x) observation operator (simulating
  observations from model)
• y H(x) innovation vector (new information)
• Minimum of the cost function J(x), (analysis)
  updates the background with new information from
  observations.

Jo
former forecast
Analysis
Jo
Background
obs
corrected forecast
Jb
Jo
xa
obs
9h
12h
15h
Assimilation window
With hypotheses, the analysis estimates the true
state of the atmosphere (in terms of max
likelihood).
7
WRF-Var data assimilation system
Theoretically, the gradient of cost funbction
should be zero at the minimum
8
WRF-Var data assimilation system
However, it is very difficult to calculate the
gradient of the cost function

• B matrix is usually huge, B-1 is nonexistent or
  difficult to calculate.
• (?xH)T, adjoint of observation operators and
  adjoint model (in 4D-Var), is difficult to
  develop and needs significant computation time.

9
WRF-Var data assimilation system
Technically, the analysis xa, is iteratively
calculated with a pre-defined minimum criterion.
10
WRF-Var Flow Chart
xb
Cycling
NCEP Analysis
WPS
TC Vortex Relocation
WRF REAL
Regular Obs
Satellite Obs
Observation Preprocessor
Forecast
yo
WRF-Var (3/4D-Var or En-Var)
xa
Radar Obs
TC Bogus Obs
B
Verification and Statistics
Background Error Calculation
11
WRF-Var Hurricane Initialization

• Vortex relocation in background fields
• If cycling, vortex relocation in background
  fields is important.
• Synthetic vortex (bogussing/relocation) in
  observation data
• Similar to JMAs scheme, see Xiao et al. (2006)
• Assimilation of regular observations
• WMO GTS
• Dropsonde data from reconnaissance
• Bogus data assimilation
• The algorithm is described in Xiao et al. (2006)
• Satellite data assimilation
• Raw data - brightness temperatures
• Retrieved data
• Radar data assimilation
• Ground-based Doppler radar data
• Airborne Doppler radar data

12
Case studies with BDA

• BDA - Bogus data assimilation
• BDA is a technique we proposed for hurricane
  initialization when I worked at FSU. It combines
  traditional vortex bogussing with data
  assimilation. Its initial application was with
  MM5 4DVAR (Xiao et al. 2000 (Mon. Wea. Rev.) Zou
  and Xiao 2000 (J. Atmos. Sci.)
• With the WRF data assimilation development, I
  includes the capability in WRF-Var

13
(No Transcript)
14
Hurricane Katrina track
15
Hurricane Katrina intensity
16
Comparison with GFS ICs

• Green without BDA, Red with BDA
  (statistics from 21 cases in 2004 and 2005
  seasons, Xiao et al. 2008)
• It is clearly shown that BDA improves hurricane
  track and intensity.
• More improvements are seen in the forecast of
  intensity than track.

17
Case studies with airborne Doppler radar data
assimilation

• Hurricane Jeanne (2004)
• Flight at around 1800 UTC 24 September 2004
• Data include wind and reflectivity

Airborne Doppler winds and reflectivity at 2.5 km
AMSL
18
Hurricane initialization
ADR-DA
NO-DA
GTS-DA
19
Hurricane forecast (reflectivity)
GTS plus radar wind plus
reflectivity
24-hr
36-hr
20
Hurricane track
Black Observation Red NO-DA Blue GTS-DA Green
GTS ADR wind DA Cyan GTS _ ADR wind and
reflectivity DA
21
Hurricane intensity
Black Observation Red NO-DA Blue GTS-DA
Green GTS ADR wind DA Cyan GTS _ ADR wind and
reflectivity DA
22

• Real-time hurricane forecasts in 2007
• Initialization 3D-Var analysis
• Observations
• All conventional data TEMP, SYNOP, METAR, PILOT,
  AIREP, SHIPS, BUOY, etc.
• Satellite-retrievals QUIKSCAT and GOES WINDS,
  GPS PW and REFRACTIVITY
• Satellite radiances AMSU-A and AMSU-B from
  NOAA-15, 16, and 17
• Synthetic observations CSLP and winds (bogus
  observations)
• First-guess GFS analysis

23
(No Transcript)
24
(No Transcript)
25
(No Transcript)
26
(No Transcript)
27

• Real-time hurricane forecasts in 2007
• Model WRF V2.2
• Domain Configuration
• 3 domains,
• 2-way moving nest of domain 2 and 3,
• 35 vertical layers,
• dimensions of 424X325 (domain1),
• 202X202 (domain 2),
• 241X241 (domain 3),
• grid-spacings of 12, 4, and 1.333km.
• Physics WSM5 microphysics,
• YSU PBL,
• Kain-Fritsch cumulus for Domain 1,
• Forecast 3 days

Moving nest
28

• Real-time hurricane forecasts in 2007
• Visualization
• Track and intensity display,
• Animation of
• SLP and surface temperature,
• Surfacde wind,
• Accumulated rainfall,
• Column max reflectivity,
• 700 hPa vertical velocity,
• Wind and temperature at
• different levels,
• 1.5 - 5 km shear,
• 1.5 - 12 km shear,
• 100-500 hPa thickness,
• Cloud-top temperature,
• etc.

29
Track Forecasts for Hurricane Dean (2007)
IC 3D-Var using GFS analysis as
first-guess Initialization time 0000 UTC, each
day Forecast time 3 days
30
3-day forecasts for Hurricane Dean (2007) from
0000 UTC daily

• The general intensifying and decaying trend of
  the forecasts is good
• The landfall time and location is pretty good
• It over-predicts the intensity when Dean is weak,
  and under-predicts it when Dean becomes strong
• 3D-Var analyses are not well balanced with model,
  so there is initial adjustment

31
3-day forecast of Humberto (2007) by WRF
initialized with GFDL analysis at 1200 UTC 12
September 2007
32
3-day forecast of Humberto (2007) by WRF
initialized with 3D-Var analysis at 1200 UTC 12
September 2007
33
3-day forecast of Humberto (2007) by WRF
initialized with 3D-Var analysis at 1200 UTC 12
September 2007
Best track till 2100 UTC 14 September 2007
34
3-day forecasts for Humberto from 1200 UTC
September 2007

• The intensification from tropical storm to
  category I hurricane just before landfall is
  predicted well
• The landfall time and location is pretty good
• The trend of weakening after landfall is
  predicted. However, it over-predicts its strength
  inland.

35
Track verification of Hurricane (2007) forecasts
(3DVAR HI GFDL)
Black HI with 3DVAR Red WPS using GFDL
36
CSLP verification of Hurricane (2007) forecasts
(3DVAR HI GFDL)
Black HI with 3DVAR Red WPS using GFDL
37
MWS verification of Hurricane (2007) forecasts
(3DVAR HI GFDL)
Black HI with 3DVAR Red WPS using GFDL
38
Verification of hurricane forecasts in 2007
season (3DVAR HI GFDL)
Black HI with 3DVAR Red WPS using GFDL
39
Conclusions

• The hurricane initialization program using
  WRF-Var is designed. It includes assimilation of
  all available observations (in-situ and
  remote-sensing) and BDA (bogus data
  assimilation).
• Case studies demonstrate positive impact of the
  hurricane initialization scheme on the hurricane
  forecasts (track and intensity).
• Statistics from 21 cases in 2004 and 2005
  hurricane seasons indicates that hurricane track
  and intensity forecasts are improved compared
  with the forecasts using the NCEP/GFS-interpolated
  initial conditions.
• Airborne Doppler radar data assimilation has
  great potential to improve hurricane vortex
  initialization and forecasts of hurricane
  structure and intensity.
• The WRF-Var hurricane initialization scheme was
  implemented in real time runs in the 2007
  hurricane season. It ran smoothly and robustly.
  The results are comparable with the runs from
  GFDL initial conditions.

40
Future Plan

• Develop a regional coupled ocean-atmosphere model
• Atmosphere model WRF ARW
• Ocean model ROMS or HYCOM
• Develop a data assimilation system for the
  regional coupled ocean-atmosphere model
• 3D-Var (initially)
• 4D-Var (after 3D-Var works properly)
• En3/4D-Var (hybrid with EnKF technique)
• Hurricane initialization and modeling
• Assimilate atmospheric data (especially satellite
  data and radar data)
• Assimilate ocean data
• Research and real-time applications

41
Thank you!