RegionalScale EnKF in Comparison with WRF3Dvar Fuqing Zhang and Ellie Meng Texas A

1
Regional-Scale EnKF in Comparison with
WRF/3DvarFuqing Zhang and Ellie MengTexas AM
University
2
WRF EnKF vs. 3DVar over the Month of June 2003

• Observations standard soundings every 12 h QCd
  by WRF/3Dvar in D2
• Verifications against soundings both before and
  after assimilation
• Boundary conditions D1 updated by 12 hourly FNL
  analysis
• Reference forecast a single forecast for the
  period w/o assimilation

3
Month-long Performance of EnKF vs. 3Dvar with WRF
? Reference forecast ? EnKF ? 3DVar
(prior, solid posterior, dotted)

(day of the month)
(day of the month)

• All in terms of domain-averaged root-mean square
  error verifying against soundings
• Both the EnKF and 3DVar have smaller error than
  reference deterministic forecast
• EnKF is consistently better than 3Dvar in terms
  of 12-h forecast and posterior analysis

4
Month-long Performance of EnKF vs. 3Dvar with WRF
---- Reference forecast ? EnKF ? 3DVar
(prior, solid posterior, dotted)
Better performance of EnKF than 3DVar also seen
in both 12-h forecast and posterior analysis in
terms of root-mean square difference averaged
over the entire month
5
12h Fcst from EnKF vs. WRF/3Dvar vs. NCEP/FNL ICs
DF_EnKF 12h deterministic forecast initiated
from EnKF analysis DF_NCEP 12h deterministic
forecast initiated from FNL analysis

• DF_EnKF has smaller (larger) forecast error than
  3DVar (EnKF), suggesting EnKF does benefit from
  both using the ensemble mean for state estimation
  and using flow-dependent background error
  covariance.
• The forecast started from the NCEP FNL analysis
  has error larger than that of EnKF but smaller
  than that of 3DVar

6
EnKF4DVar Coupling EnKF with 4DVar in
Lorenz95Fuqing Zhang, Meng Zhang and Jim
Hansen
7
Comparison in Perfect model Scenario
(F8.0)80-variable Lorenz95 model with 20, 40
(above) vs. 10 (below) members
8
Comparison in Imperfect model Scenario (F8.5 but
truth is 8.0)80-variable Lorenz95 model with 20,
40 (above) vs. 10 (below) members
9
Imperfect model with moderate model error F8.5
but truth is 8.080-variable Lorenz95 model with
20, 40 (above) vs. 10 (below) members
10
Imperfect model with strong model error F9.0
but truth is 8.080-variable Lorenz95 model with
20, 40 (above) vs. 10 (below) members