Ensemble-Kalman Filter (Evensen 1994)
Post on 02-Feb-2016
62 Views
Preview:
DESCRIPTION
Transcript
Mesoscale Ensemble-Based Data Assimilation
Fuqing Zhang Texas A&M University, College Station, Texas
Collaborators: Ellie Meng, Altug Aksoy, Chris Snyder, David Dowell, Jenny Sun and John Nielsen-Gammon
• Regional scale [O(1 day) & O(1000km)]: assimilating sounding and surface
observations using a mesoscale model (MM5)
• Storm-scale [O(1 hour) & O(100km)]: assimilating radar observations using
a cloud-resolving model
• Thermally-forced circulation: assimilating only surface observations for
simultaneous state and parameter estimation using a 2D sea-breeze model
Ensemble-Kalman Filter (Evensen 1994)
Use ensemble forecast to estimate flow-dependent background error covariance
t=t0-t t=t0
ensemble forecast
x1a
xNa
obs y x1a
EnKF
xNa
x1f
xNf
xa = xf + BHT(HBHT+R) -1(y-Hxf)
Kalman Filter (Kalman 1960)
Uses all available information in order to produce the most accurate possible description of the state of the flow. Also
provides the uncertainty in the state of the flow resulting from the uncertainties in the various sources of information.
ensemble forecast
x1f
xNf
t=t0+t
Vertical velocity at 5km(colored) and surface cold pool (black lines, every 2K)
Storm-scale EnKF with Simulated Radar OBS (Snyder and Zhang 2003; Zhang, Snyder and Sun 2004)
Assimilating Vr if dBZ>12 every 5 minutes; no storm in initial ensemble
Truth
EnKF
Black curves: EnKF analyses at the tower location Gray curves: Independent observations from the instrumented tower Open circles: Samples from the dual-Doppler analysis.
Storm-Scale EnKF with Real Radar OBS (Dowell, Zhang, Wicker, Snyder and Crook 2004)
Assimilating Vr from one radar verify against dual-doppler analysis and tower data
Experimental Design: Regional-scale EnKF(Zhang, Meng and Aksoy 2004; Meng and Zhang 2004)
• Forecast model: MM5, 30-km grid spacing over CONUS domain (190x120x27)
• Case in study: the “surprise” snowstorm of 24-26 January 2000
• A 20-member ensemble: initiated at 00Z 24 Jan with random but balanced
perturbations using MM5 3Dvar background error statistics (Barker et al. 2003)
• Perfect-model OSSE: truth as one of the ensemble members; no model error
• OBS type: sounding obs of u, v, T from truth run at (300 km)2 spacing, every 12h
surface obs of u, v, T from truth run at (60 km)2 spacing, every 3h
• OBS error: 1 K for T and 2 m/s for u&v; uncorrelated
• Square-root sequential EnKF: OBS assimilated one by one; OBS not perturbed
• Radius of influence: 1800 km with Gaspari and Cohn (1999) cutoff
• Variance relaxation: mixing prior and posterior variances (Zhang, Snyder and Sun 2004)
Forecast Experiment: Truth (above) vs. Forecast (below) Model-derived reflectivity (colored) and MSLP (blue lines, every 2 hPa)
EnKF Performance: Forecast Error (above) vs. Analysis Error (below)
Errors in MSLP (every 0.5hPa) and surface winds (full barb, 5m/s)
EnKF Performance: Forecast Error (above) vs. Analysis Error (below)
RMS error of difference total energy (every 2m/s); 2 2 21 1
N 2RM_DTE = (u' + v' + kT' )
EnKF Performance: Forecast Error (dotted) vs. Analysis Error (solid)
vertical error distributions at 0 (green), 12 (red), 24 (blue) and 36 h (black)
EnKF Performance: Time Evolution of EnKF Analysis Error (solid) vs. Forecast Error (dotted) and ensemble spread (gray)
EnKF Performance: Spectral Analysis of Forecast Error (dotted) vs. Analysis Error (solid) at 0 (green), 12 (red), 24 (blue) and 36 (black) h
Sensitivity to OBS Accuracy, Coverage and Availability Analysis Error (solid) vs. CNTL (gray) and Forecast Error (dotted)
Sensitivity to Model Errors (Meng and Zhang 2004) Analysis Error (blue) vs. CNTL (green) and Forecast Errors (dotted)
Resolution Error: truth is produced with 10-km grid spacing; 30km used for the ensembles
Parameterization Error: truth is produced with Grell CPS; KF CPS used for the ensembles
Summary of Regional-scale EnKF
• EnKF assimilation of sounding and surface observations proved
effective for meso- and regional scales
• EnKF can reduce the analysis errors by as much as 80% for u, v, T
and p; it is relatively less efficient for w and q whose errors have
stronger smaller-scale components
• EnKF is most effective in correcting errors in larger-scale growing
structures; less effective in correcting errors in smaller, marginally
resolvable scales which also have faster error growth and shorter
predictability
• EnKF performance can be significantly degraded if an imperfect model
is used, suggesting a need for the explicit treatment of model errors
2
2
22
2
( ) ,
( ) .b
bu u w
t x z x z
b b b bu u w N w Q
t x z z
• Model: Two-dimensional, irrotational, incompressible flow with prognostic variables (perturbation) temperature (b′) and vorticity (η′)
• Explicit heating function:
0/10
0
tan cos ( ) ,2
z zxQ A e t t
x
• Parameters estimated: Mean horizontal wind; vertical mixing coefficients;
Static stability; Heating amplitude; Heating depth
Simultaneous State and Parameter EstimationExplicit treatment of model error in a 2D sea-breeze model
(Aksoy, Zhang, Nielsen-Gammon and Epifanio 2004; Aksoy, Zhang and Nielsen-Gammon 2004)
EnKF Performance: Thermally-forced (Sea-Breeze) Circulation Assimilating only perturbation temperature over land surface
Bu
oya
ncy
Vo
rtic
ity
Prior
Prior
Posterior
Posterior
u b
2N 0A 0z
EnKF Performance: Explicit treatment of Model ErrorsEstimating simultaneously six imperfect model parameters and the state
Concluding Remarks
• EnKF demonstrated to be promising for convective-, meso-/regional scale
data assimilation
– Why: ensemble forecasting provides the best estimate of the state, the
associated uncertainty, and the flow-dependent background error
covariance
• EnKF demonstrated to be promising for explicit treatment of model error
through simultaneous state and parameter estimation
• Remaining issues: ensemble generation, complex model errors, boundary
conditions, large data volume, correlated OBS error, …, etc.
top related