Title: Performance Characteristics of a Pseudo-operational Ensemble Kalman Filter
1Performance Characteristics of a
Pseudo-operational Ensemble Kalman Filter
Greg Hakim Ryan Torn University of
Washington http//www.atmos.washington.edu/hakim
April 2006, EnKF Wildflower Meeting
2Outline
- Issues for limited-area EnKFs.
- Boundary conditions.
- Nesting.
- Multiscale prior covariance.
- UW pseudo-operational system.
- Performance characteristics.
- Analysis of Record (AOR) test.
- Experiments using the UW RT data.
- Sensitivity targeting.
- Observation impact thinning.
3Boundary Conditions
- Obvious choice global ensemble, but
- Often ensembles too small.
- Undesirable ensemble population techniques.
- Different resolution, grids, etc.
- Flexible alternatives (Torn et al. 2006).
- Mean random draws from N(0,B).
- Mean scaled random draws from climatology.
- error boundary layer shallow due to obs.
4Nesting
- Grid 1 global ensemble BCs.
- E.g. draws from N(0,B) or similar.
- Grid 2 ensemble BCs from grid 1.
- One-way nesting straightforward.
- Cycle on grid 1, then on grid 2.
- Two-way many choices little experience.
- Note Hxb different on grids 1 and 2.
- Issues at grid boundaries.
5The Multiscale Problem
- Sampling error
- noise in obs est prior covariance.
- Ad hoc remedies
- localization
- Confidence intervals.
- Multiscale problem.
- Noise on smallest scales may dominate.
- Need for scale-selective update?
6Surface Temperature Covariance
7Mesoscale Example cov(V, qrain)
8Real Time Data Assimilation at the University of
Washington
9Objectives of System
- Evaluate EnKF in a region of sparse in-situ
observations and complex topography. - Estimate analysis forecast error.
- Sensitivity targeting thinning.
10Model Specifics
- WRF Model, 45 km resolution, 33 vertical levels
- 90 ensemble members
- 6 hour analysis cycle
- ensemble forecasts to t24 hrs at 00 and 12 UTC
- perturbed boundaries using fixed covariance
perturbations from WRF 3D-VAR
11Observations
Obs. Type Variables 00 UTC 06 UTC 12 UTC 18 UTC
Surface Altimeter 430 420 420 440
Rawindsonde u, v, T, RH 1000 0 1000 0
ACARS u, v, T 1650 1390 740 1860
Cloud Wind u, v 2030 1740 1670 1510
Total 5110 3550 3830 3810
12Probabilistic Analyses
sea-level pressure
500 hPa height
Large uncertainty associated with shortwave
approaching in NW flow
13Microphysical Analyses
20 February 2005, 00 UTC
model analysis
composite radar
14Ensemble Forecasts
Analysis
24-hour forecast
15Verification
16Temperature Verification
12 hour forecast
24 hour forecast
UW EnKF GFS CMC UKMO NOGAPS
ECMWF
17U-Wind Verification
12 hour forecast
24 hour forecast
UW EnKF GFS CMC UKMO NOGAPS
ECMWF
18Moisture Verification (Td)
12 hour forecast
24 hour forecast
UW EnKF GFS CMC UKMO NOGAPS
ECMWF
19No Assimilation Verification
Winds
Temperature
UW EnKF No Observations Assimilated
20Moving Toward the Mesoscale
21Analysis of Record
Hourly surface analyses. EnKF covariances. Availab
le t30 minutes. 15 km resolution.
22Hurricane Katrina at 10 km
23Sensitivity Analysis
- Basic premise
- how do forecasts respond to changes in initial
boundary conditions, the model? - Applications
- targeted observations network design.
- targeted state estimation (thinning).
- basic dynamics research.
24Adjoint approach
Given J, a scalar forecast metric, one can show
that
adjoint of resolvant
- Need to run an adjoint model backward in time.
- Complex code lots of approximations
- Does not account for state estimation or errors.
25Ensemble Approach
- Adjoint sensitivity weighted by initial-time
error covariance. - Can evaluate rapidly without an adjoint model!
- Can show this gives response in J, including
state estimation.
With Brian Ancell (UW)
26Sensitivity from the UW Real-time system
Case study removing one observation. Metric
average MSL pressure over western WA
27Sensitivity Demonstration
How would a forecast change if buoy 46036 were
removed?
28Overview of Case
29Overview of Case
30Overview of Case
31Overview of Case
32Overview of Case
3312 UTC 5 Feb Sensitivity
Sea-level pressure
850 hPa temperature
3412 UTC 5 Feb. Analysis Change
Analysis Change
Forecast Sensitivity
35Forecast Differences
- Assimilating the surface pressure observation at
buoy 46036 leads to a stronger cyclone. - Predicted Response 0.63 hPa
- Actual Response 0.60 hPa
36Summary of 10 Cases
37Observation Impact
- Adaptively sampling the obs datastream
- Thin by assimilating only high-impact obs.
38Observations Ranked by Impact
39Ob-Type Contributions to Metric
40Metric Prediction Verification
41Summary
- BCs flexibility weak influence.
- UW real-time system gov. center quality.
- Moisture field better than most.
- Surface AOR 10 km.
- Sensitivity analysis.
- Ensemble targeting easy flexible.
- Adaptive DA (thinning).
42(No Transcript)
43AOR Opportunities
- No propagate update
- nested high resolution single member.
- assimilate using coarse-grid stats.
- can be done now.
- Deterministic propagation
- as above, but evolve high-res state.
- Full filter
- evolve assimilate entire ensemble.
- 4DVAR with EnKF statistics.
- at least 3--5 years out.
44AOR Challenges
- True multiscale conditions (lt15 km).
- Scale-dependent sampling errors?
- Bias estimation and removal.
- EnKF allows state-dependent bias estimation.
- Model error estimation removal.
- Parameter estimation model calibration.
- Satellite radiance assimilation.
- Kalman smoothing.
45IR Temperature Analyses
30 March 2005, 12 UTC
model analysis
IR satellite image
46Global Perturbations
Randomly choose Ne draws and scale to desired
variance.
Create a number of draws from N(0,B)
Add to deterministic boundary condition and
calculate tendency
47Height Verification
12 hour forecast
24 hour forecast
UW EnKF GFS CMC UKMO NOGAPS
ECMWF
48Surface Obs. and Rawindsondes
49Observation Densities
aircraft obs.
cloud winds
50Ensemble inliers/outliers
inlier
outlier