1
Development and Testing of a Regional GSI-Based EnKF-Hybrid System for the Rapid Refresh
Configuration
Yujie Pan1, Kefeng Zhu1, Ming Xue1,2, Xuguang Wang1,2, Jeffrey S. Whitaker3, Stanley G. Benjamin3 and Stephen S. Weygandt3 and Ming Hu3
Center for Analysis and Prediction of Storms1 and School of Meteorology2
University of Oklahoma, Norman Oklahoma 73072NOAA Earth System Research Laboratory3, Boulder, Colorado
5th EnKF WorkshopAlbany, New York
May 2012
2
OutlinePart 1: Introduction to the regional GSI-based EnKF-hybrid data assimilation system
Part 2: Single observation tests
Part 3: Comparison of hybrid with GSI and pure EnKF
EnKF-Hybrid 1 way interactive EnKF-Hybrid 1 way with multi-physics EnKF EnKF-Hybrid 2 way interactive Verification of precipitation forecasts on 13 km grid
''1''1
2'1
1'11
211'1
21
21
21
,
HxyHxyαCαxBx
αx
oToTT
oe JJJJ
R
K
k
ekk
1
'1
' xαxx
3
B 3DVAR static covariance; R observation error covariance; K ensemble size; C correlation matrix for ensemble covariance localization; e
kx kth ensemble perturbation; '1x 3DVAR increment; 'x total (hybrid) increment; 'oy innovation vector;
H linearized observation operator; 1 weighting coefficient for static covariance;
2 weighting coefficient for ensemble covariance; α extended control variable.
Extended control variable method (Lorenc 2003) in 3D GSI hybrid (Wang 2010, MWR):
Extra term associated with extended control variable
Extra increment associated with ensemble
GSI-Hybrid: Method
4
EnKF
EnKF—RR
RUC
EnKF Domain207x207 grid points~40 km, 51 levelsPrecip. Forecast Domain532x532 grid points~13 km, 51 levelsPrecip. Verification DomainRUC Domain as indicated
Ensemble members 40
Experiment Domains
5
Observations assimilated
Sounding and profiler Surface data from land stations and ships
Aircraft Satellite retrieve winds
6
Single Observation Tests (Comparing GSI, Hybrid and EnKF)
3DVAR
Different weight for the static covariance in Hybrid
Solid line: Height at 600 hPa (background)Shading: Temperature increment
EnKF
Weight=1 Weight=0Weight=0.5
Hybrid
Half staticHalf flow-dependent
7
Hybrid GSI-EnKF DA system: 1 way coupling
control forecast Hybrid
control analysis
control forecast
data assimilation First guess forecast
EnKF analysis k
member 1 forecast
member 2 forecast
member k forecast
EnKF
EnKF analysis 2
EnKF analysis 1
member 1 forecast
member 2 forecast
member k forecastEnsemble
covariance
……
……
……
Wrf-DFL0 20m 40m
Wrf-DFL0 20m 40m
Wrf-DFL0 20m 40m
GSI
observations
Innovation
EnK
FH
ybrid
8
Experiments Horizontal localization
(KM)
Vertical localization
(ln(p))
Fix inflation Adaptive inflation
EnKF 1000 KM (height
dependent)
1.1/1.6 (height
dependent)
0.1 0.9
Hybrid 1way 1000 KM 1.1
……Time (UTC)3hr fcst00 03
obs obs Obs
12 3hr fcst3hr fcst
BackgroundFields
EnKF & hybrid
Analysis Fields
EnKF & hybrid EnKF & hybrid
2010-05-08 00:00
obs
21 3hr fcst
EnKF & hybrid
2010-05-17 21:00
…………
Interpolation
13 KM 12 hr Fcst
Interpolation
13 KM 12 hr Fcst
Hybrid And EnKF Configuration
9
Surface Variables Verification (RMSE; 3-18 hr Forecasts)
Hybrid 1way
EnKF
GSI 3dvar
18h3h
3-18 hour forecasts verification against surface data.
10
Verifications Against Soundings (RMSE)
Hybrid 1way
EnKF
GSI-3dvar
11
Verifications Against Soundings (RMSE)
12
group Long wave Short wave Surface layer PBL Cumulus
p1 rrtm scheme (1) Goddard short wave (2)
Monin-Obukhov (Janjic) scheme (2)
Mellor-Yamada-Janjic TKE scheme (2)
Grell 3D ensemble
scheme (5)
p2 rrtm scheme (1) Dudhia scheme (1) Monin-Obukhov scheme (1)
YSU scheme (1) Kain-Fritsch (new Eta)
scheme (1)
p3 rrtm scheme (1) Goddard short wave (2)
MYNN surface layer (5)
MYNN 2.5 level TKE scheme (5)
Grell-Devenyi ensemble
scheme (3)
p4 GFDL (Eta) longwave (99)
GFDL (Eta) short wave (99)
Monin-Obukhov (Janjic) scheme (2)
Mellor-Yamada-Janjic TKE scheme (2)
Grell 3D ensemble
scheme (5)
p5 rrtm scheme (1) Goddard short wave (2)
Monin-Obukhov (Janjic) scheme (2)
Mellor-Yamada-Janjic TKE scheme (2)
Grell-Devenyi ensemble
scheme (3)
Multi-physics GSI-EnKF Hybrid System Configuration
13
Surface variables verification (RMSE; 3-18 hr Forecasts)
When Multiple-physics schemes were employed for EnKF, hybrid was also improved .
Multi-hybrid
Single-hybrid
GSI 3dvar
14
Verifications Against Soundings (RMSE)
Multi-hybrid
Single-hybrid
GSI 3dvar
15
Sensitivity Tests To Covariance Weight
Experiment Weight to static covariance
Hybrid 00 0.0
Hybrid 01 0.1
Hybrid05 0.5
Hybrid 09 0.9
GSI 3dvar 1.0
Hybrid main parameters:Horizontal localization : ~1100 KMVertical localization : 1.1 ( ln(p) )
Verifications Against Soundings
1100 KM horizontal localization improve the performance of hybrid at jet level
16
Hybrid GSI-EnKF DA system: 2 way coupling
member 1 forecast
member 2 forecast
member k forecast
control forecast GSI-ECV
EnKF
control analysis
EnKF analysis k
EnKF analysis 2
EnKF analysis 1
member 1 analysis
member 2 analysis
member k analysis
member 1 forecast
member 2 forecast
data assimilation
control forecast
Ensemble covariance
Re-center EnSR
analysis ensemble
to control analysis
…… ……
……
……
First guess forecast
GSI
observations
Innovation
member k forecast
Wrf-DFL0 20m 40m
Wrf-DFL0 20m 40m
Wrf-DFL0 20m 40m
Wrf-DFL0 20m 40m
Wang et al. 2011
17
Surface Variables Verification (RMSE)
Hybrid 2way
EnKF
GSI-3dvarSingle-physics EnKF was used.
18
Verifications Against Soundings (RMSE)
Hybrid 2way
EnKF
GSI-3dvar
19
Verifications Against Soundings (RMSE)
20
OBS (NCEP Stage IV) GSI EnKF
2010051111
2010051305
11 hr forecast started from 2010051100
5 hr forecast started from 2010051300
Hybrid2way
Hourly Precipitation Forecasts on 13 km Grid
21
Verification of Hourly QPF on 13 km Grid
Hybrid 2way
EnKF
GSI
22
Conclusions• The GSI-based hybrid (run at 40 km grid spacing for RAP data set and
model), with either 1-way or 2-way interaction with a single-physics EnKF and using equal weight for static and flow-dependent covariances, outperforms the GSI and pure EnKF for most verified variables (relative humidity, temperature, wind), except surface temperature. The advantage lasts up to the 18 hour forecast time.
• The hybrid with half static covariance is better than the one without static covariance, indicating the benefit of including static covariance for the current application.
• EnKF and hybrid predict more accurate precipitation pattern and location on a 13 km grid than GSI, which is also demonstrated by ETS score. But hybrid doesn’t improve the precipitation forecasts as much as EnKF.
• The performance of the EnKF system is noticeably improved when multiple physics schemes are used in the ensemble forecast, especially for temperature and moisture fields.
23
Future Plan (in collaboration with GSD and EMC)
• Use height-dependent localization for flow-dependent covariance in the hybrid – found helpful within EnKF
• Use well tuned multi-physics EnKF within 2-way hybrid. • Test the impact of the strong constraint available in GSI• Add satellite data.• Implement and test dual-resolution (40/13 km) hybrid• Test the system with hourly cycles• Eventual quasi-operational testing of hourly cycled, two-way interactive
EnKF/hybrid system for RAP including radar data.• Long term: Hybrid system applied to NARRE (North America Rapid
Refresh Ensemble) and HRRRE (High-Resolution Rapid Refresh Ensemble)
• Nesting CAPS’s Storm-Scale EnKF within (see Youngsun Jung’s talk)
24
Thank you!!
state-dependent covariance inflation
• Fix inflation
• Adaptive inflation
• Final inflation
' '1
1 (ln( ) / )
a a
a
sfc
x xpb taper lncutp
' '2
2 2
2 2 1
a a
f a
a
x x
c
925ap hPa
700ap hPa
500ap hPa
300ap hPa
1 2
1000 , 6sfcp hPa lncut
tape
r(r)
Pres
sure
(hPa
)
Vertical smoothing-scale (vz) in GSI
Step1: vz*( log( p(k-1)/psf )-log( p(k+1)/psf) )/2
Step2: vz=vz/1.5
Vertical smoothing scales in GSI
p(k): average pressure at the k-th model levelpsf: average surface pressure
Convert to vertical grid units
loc = loc*coefficent
0 0.2 0.4 0.6 0.8 1
100
200
300
400
500
600
700
800
900
1000
vert
Pres
sure
(hPa
)
WindRH & T
1 1.1 1.2 1.3 1.4 1.5
100
200
300
400
500
600
700
800
900
1000
hori
Pres
sure
(hPa
)
loc = loc*coefficent