1. North Carolina State University
How are tropical cyclones represented in operational model initial conditions? And why does it matter?
“Alberto”, mb
Gary Lackmann
North Carolina State University
5 July 2012
Contributions from Daryl Kleist (EMC), Mike Brennan (NHC), and John Brown (ESRL) and Briana Gordon (STI) are gratefully acknowledged
18 UTC Sunday 20 May 2012, GFS 95-km SLP + GOES Visible

2. Outline Background and Motivation B. Operational Models and TC IC
1.) Challenges of TC prediction and initialization
2.) Data Assimilation background and TC DA
3.) Hybrid DA in GFS and TC IC
B. Operational Models and TC IC
1.) GFS
2.) HWRF
3.) GFDL
4.) RAP
5.) NAM (briefly)
C. Conclusions and Questions
Some acronyms:
TC = Tropical Cyclone
DA = Data Assimilation
IC = Initial Conditions
EnKF = Ensemble Kalman Filter
BV = Bogus Vortex

3. Intensity related to interaction of multi-scale processes
Atlantic TC track prediction: Improving
Track related to large-scale steering flow; improvements in satellite data assimilation (DA), environmental recon sampling, NWP, human forecasting skill
Source:
Intensity prediction: Slower improvement, if any
Intensity related to interaction of
multi-scale processes

4. TC Intensity Forecasting
Why are intensity forecasts slow to improve?
What are challenges for numerical TC prediction?
Difficulty with initial conditions
Need to represent complex process interactions across spatial scales (e.g., eyewall replacements; resolution)
Difficulty representing physical TC processes (e.g., convection and swirling PBL over complex surface)
Incomplete understanding of physical processes
“Dynamically, the tropical cyclone is a mesoscale power plant with a synoptic-scale supportive system.” (Ooyama 1982)

5. Data Assimilation (DA) Overview (after Kalnay Fig. 5.1.2a)
Observations (+/- 3 h)
Background or first guess
Approach: Use ALL available information for best possible analysis
Observations + short-term forecast (“background”) + information about error + dynamical and physical relations, etc.
Global analysis (statistical interpolation and balancing, Quality Control [QC])
Initial Conditions
Global forecast model
6-h forecast
Operational forecasts

6. TC Data Assimilation Specific TC DA Challenges:
1.) Sometimes not enough information, esp. inner core
Rain contamination of some satellite-borne sensors
Few in-situ observations other than recon
2.) Much available information not used, esp. near TC
Obs, background can differ greatly near TC, QC eliminates obs
Data density issues (hurricane hunter radar coverage, drops)
Model resolution insufficient to capture inner core structure, observational representativeness challenge

7. Data Assimilation
Critical aspect: Relative weighting of observations & background (short-term model forecast) in analysis Accurate knowledge of error associated with background and observations determines weighting Static 3DVAR: Assume constant error statistics Ensemble Kalman Filter (EnKF): Use ensemble to provide flow-dependent background error information

8. TC Initialization: GFS
GFS links to NAM, RAP, to some extent GFDL, HWRF
Starting with 12Z run, 22 May 2012, new GFS hybrid DA system implemented
Hybrid: Blend of short-term ensemble and old (constant) information to define background error

9. Change to Analysis from Single Observation
850-mb Tv ensemble spread, 00Z 9/12/2008
Background T (contours), and change to analysis from assimilation of ob (shaded)
Tv observation
All static background error
All ensemble error (bf-1=0.0)
Hybrid, 50% ens, 50% static (bf-1=0.5)
Single 850mb Tv observation (1K O-F, 1K error)
Slide compliments of Daryl Kleist, EMC

10. GFS TC Initialization Information from: Daryl Kleist (personal communication 2012) and Kleist et al. 2011a,b
(1) NHC declares storm (TD strength or greater)
Does GFS 6-h forecast represent system?
Yes No
(2a) Vortex relocated to NHC position (in background field prior to DA)
(2b) Synthetic (bogus) wind “observations” generated for use in DA
(3) NHC storm information written to “TCvitals” file; system reads location, central pressure, used in DA process regardless of 2a or 2b

11. F06 (from 18 UTC), Valid 00 UTC 21 May 2012
Example with new GFS hybrid (parallel) DA system for TS “Bud” (now operational)
F06 (from 18 UTC), Valid 00 UTC 21 May 2012
Note weak representation of Bud; Tracker unable to “find” coherent system
GFS Parallel 5/21/12 00 UTC 850 , wind (6-h fcst)
GFSP ANALYSIS 00 UTC 21 May 2012:
Note radical change to Bud due to assimilation of synthetic wind observations (no relocation done in this case, since tracker didn’t find storm)
GFS Parallel 5/21/ UTC 850 mb , wind (Analys)
Slide compliments of Daryl Kleist, EMC

12. GFS: Vortex Relocation
4-step process: 1.) Locate hurricane vortex in background 2.) Separate TC from environmental field (filtering- from GFDL) 3.) Move hurricane vortex to NHC official position 4.) Data assimilation step includes MinSLP ob from NHC No relocation if storm center over major land mass, or if terrain elevation > 500 m See Liu et al. (2000) for more info on this process

13. GFS TC Initialization
Does GFS utilize recon data in Data Assimilation system? GFS uses some G IV and P3 data, but DA system makes limited use of in-situ observations in/near storm With old DA system, representativeness issues of inner-core obs, so these are flagged and most dropsonde data not assimilated GFS assimilation of NHC central pressure ob helps some (implemented in Kleist et al. 2011, WAF)

14. Operational GFS (T382) analysis
Operational GFS (T382) F72
Hanna (989 obs)
Ike (956 obs)
Control GFS (T574)
Control with MinSLP (T574)
Kleist et al. (2011 WAF)

15. GFS TC Initialization Information from: Daryl Kleist (personal communication 2012) and Kleist et al. 2011a,b
Due to coarse GFS resolution (effectively 27-km grid length), small and strong TCs will still be weaker in model IC than in reality; larger, weaker storms better represented New GFS Hybrid DA system, by using ensemble to measure background error, offers potentially major improvement, allows assimilated observation information to distribute in flow-dependent fashion (see following slides) Due to coarseness of ensemble, the former static part of error covariance is needed to represent small scales (static part of hybrid system uses higher-resolution background)

16. GFS: Single Observation
Slide compliments of Daryl Kleist, EMC
All static background error
All ensemble error (bf-1=0.0)
Hybrid, 50% ens, 50% static (bf-1=0.5)
Single Ps observation (-2mb O-F, 1mb error) near center of Hurricane Ike

17. GFS: Single Observation
Slide compliments of Daryl Kleist, EMC
All static background error
All ensemble error (bf-1=0.0)
Hybrid, 50% ens, 50% static (bf-1=0.5)
Single 850mb zonal wind observation (3 m/s O-F, 1m/s error) in Hurricane Ike circulation

18. GFS SLP, 10-m wind, 00 UTC 29 October 2012 (1 degree)
Sandy, 00 UTC/ 1:30 UTC 29 Oct 2012
GFS SLP, 10-m wind, 00 UTC 29 October 2012 (1 degree)
ftp://ftp.aoml.noaa.gov/hrd/pub/hwind/Operational/2012/AL182012/1029/0130/AL182012_1029_0130_contour08.png

19. GFS TC Initialization
New hybrid DA system (5/2012), and assimilation of MinSLP (2009) have improved TC IC for GFS Additional work is needed to better utilize observational information in/near TC core Resolution limitations remain an obstacle for full-strength initialization; larger, weaker storms better represented 2012: Preliminary Atlantic track error results (NHC) indicate GFS better than ECMWF at hours Any questions on GFS TC IC?

20. HWRF
Became operational in 2007 High-resolution (27/9/3 km domains) with moving inner domains for high-resolution TC prediction Utilizes high-resolution data assimilation Coupled with Princeton Ocean Model for air-sea feedbacks
Slide modified from Mike Brennan (NHC)

21. HWRF TC Initialization Information taken from: http://www. emc. ncep
1.) Define HWRF domain based on observed TC position
2.) Interpolate GFS analysis to HWRF grid
3.) Remove GFS vortex from analysis
4.) Insert high-resolution vortex:
- For 1st run or strength < 25 kt, composite bogus vortex:
- Used for initial HWRF run of any system of any intensity
Used for any HWRF run for systems of initial intensity < 25 kt
- Subsequent runs with initial intensity ≥ 25 kt:
Vortex from previous cycle 6-h forecast extracted
Storm location, size, and intensity corrected using TCVitals data
If first-guess vortex does not match the initial intensity specified by NHC, then portions of composite vortex added
5.) Run GSI (previous GFS DA system) with obs and vortex in DA cycle; GSI run separately for each domain
For 2012, vortex constructed on 3-km inner domain
Slide modified from Mike Brennan (NHC)

22. HWRF Bogus Vortex
Only used for “cold start” situations; ~once per storm
Bogus vortex created from 2D axisymmetric vortex from past model forecast of small, near-axisymmetric system
2D vortex includes perturbations of horizontal wind component, temperature, specific humidity and sea-level pressure
To create the bogus storm:
Wind profile of 2D vortex smoothed until its RMW / maximum wind speed matches observed values
Storm size and intensity are corrected following a procedure similar to that for cycled system
Vortex in shallow storms undergoes 2 final corrections: Vortex top set to 700 hPa, warm core structure removed
Slide modified from Mike Brennan (NHC)

23. HWRF Data Assimilation
Uses GSI DA system on outer domain and special 20°x20° “ghost” domain to assimilate conventional and satellite radiances
However, conventional data within 150 km of storm center not assimilated due to their negative impact on forecast
Largely due to static isotropic background error covariances
Testing 4DVAR and hybrid EnKF-Variational schemes with P3 tail Doppler radar data
Slide modified from Mike Brennan (NHC)

24. GFDL Operational since 1995 Triple nest, ~30, 10, and 5-km grid length
Coupled to Princeton ocean model
Uses “bogus vortex” plus asymmetries from previous 12-h forecast
Slide modified from Mike Brennan (NHC)

25. GFDL Initialization Taken from Bender et al. (2007)
Filters remove vortex from previous 12-h forecast
Azimuthal means computed for all prognostic variables, subtracted to get 3-D asymmetries, which are added to the initial axisymmetric vortex
Depth of storm adjusted based on NHC intensity analysis (depth of the storm increases as a function of NHC assigned intensity)
In 2002, filtering in upper-levels reduced to retain more of GFS analysis there
GFDL bogus vortex is available, can be used for local model initialization
Slide modified from Mike Brennan (NHC)

26. GFDL* Bogus Vortex Specification
Symmetric component (shown)
Created from axisymmetric version of model
Asymmetric component (not shown)
Added from 12-hr forecast of previous GFDL
model run
BV specified from observed location/intensity
Source: Kurihara et. al., 1993
*Geophysical Fluid Dynamics Lab (GFDL)

27. Bogus Vortex with TC Erika: PV on 2 Sept ’09, 00 UTC Analysis
GFS Only
PV ~ 2 PVU
GFDL+GFS
PV ~ 7.5 PVU

28. RAP 12Z run, 1 May 2012: RAP replaced RUC
RAP is WRF ARW model, with RUC-similar physics
Important changes in DA and some physics from RUC
RUC uses previous GFS GSI DA system (not hybrid)

29. RAP Initialization Information from: John Brown (NOAA ESRL, personal communication)
Similar to NAM, GFS information “injected” with “partial cycling” strategy
RAP: 03 and 15 UTC, 1-h partial cycle of RAP where GFS 3-h forecast used for background
After 3Z, 15Z analyses, DFI radar initialization applied, and IC for next 1-h forecast generated
Process repeated hourly until 09 and 21Z, when 1-h RAP forecast substituted into ongoing RAP

30. RAP Initialization Information from: John Brown (NOAA ESRL, personal communication)
Bottom line: RAP makes no unique provision for TC initialization
Utilizes information from GFS via partial cycling strategy (similar for NAM, with good results)
RAP system should improve on RUC, which would not have a TC unless one crossed in from lateral BC, formed in the RUC (rare), or “drawn for”

31. RAP Initialization Information from: John Brown (NOAA ESRL, personal communication)
Does RAP draw in recon or special TC obs?
If special in-situ obs in NAM then attempt to use in RAP
Radar and wind from P3 not used at this time
An advantage of RAP is radar-derived diabatic initialization; offshore in TC, this advantage less, but lightning used as proxy to help (GSD version)
Basic RAP DA system is based on previous GFS GSI 3DVar system. In future, use GFS-type hybrid?

32. NAM Initialization Information from: http://www. emc. ncsp. noaa
NAM Initialization Information from: (TC part: 25 May 2012)
TCVitals generated from NHC/FNMOC/JTWC
GFS first-guess with relocated storm also used as background to NDAS analysis
For all storms, NDAS process mimics GFS process for weak storms where vortex not found in background
TCVitals used for synthetic (bogus) wind profile obs for use in DA
Mass observations near storm flagged and omitted, ditto dropsondes
See

33. NAM Initialization Information from:
Uses 3-D Var, nothing special for TCs, but partial GFS cycling helps
See
Graphics courtesy NHC

34. Conclusions New hybrid GFS DA system cause for optimism
For NCEP operational models, GFS TC IC most important
GFS cycled in to NAM, RAP
GFS large-scale and BC data used in HWRF, GFDL
HWRF, GFDL have resolution advantage, but not fully available in AWIPS
High-resolution TC DA in HWRF has promise, but more computer power needed
DA systems in HWRF, NAM don’t use hybrid En-KF strategy of GFS

35. Model TC Initial Conditions
Storm initial intensity?
Weak storm? Better initialized
Strong storm? Model IC too weak
Storm size?
Larger storms better represented
Storm age?
Newly declared storms handled differently than “mature” storms in models

36. Acknowledgements Daryl Kleist (NOAA/NCEP/EMC)
Mike Brennan (NOAA/NCEP/NHC)
John Brown (NOAA/ESRL)
Briana Gordon (Sonoma Technology, Inc)
Wallace Hogsett (TSB NHC)
Stan Benjamin (NOAA/ESRL)
Brian Etherton (NOAA/ESRL)
NOAA CSTAR Grant #NA10NWS
Jonathan Blaes (NWS RAH)
COMET program for graphics and Operational Model Matrix

37. Questions?