Convection-permitting forecasts initialized with continuously-cycling
limited-area 3DVAR, EnKF and “hybrid”data assimilation systems
Craig Schwartz and Zhiquan LiuNCAR/NESL/MMM
NCAR is sponsored by the National Science Foundation
Introduction• Convection-permitting forecasts have
commonly been initialized from operational analyses (e.g., GFS, NAM)– Example: Interpolate GFS analysis onto WRF
domain
• Continuously cycling mesoscale data assimilation systems can produce initial conditions for convection-permitting forecasts– Dynamically consistent analysis/forecast
system
A few data assimilation approaches• Three-dimensional variational (3DVAR)
– Background error covariances (BECs) typically fixed/time-invariant
– May yield poor results when actual flow differs from that encapsulated within the fixed “climatology”
• Ensemble Kalman filter (EnKF)– Time-evolving, “flow-dependent” BECs
estimated from a background ensemble
• “Hybrid” variational/ensemble– Incorporates ensemble background errors
within a variational framework – Combination of fixed and
time-evolving background errors
A few data assimilation approaches
75% squirrel25% cat
Experimental design•Full-cycling (6-hr period) between May 6 – June 21, 2011
•Data assimilation/cycling on a 20-km domain
•Three experiments assimilating identical observations:
•Pure 3DVAR•Pure EnKF•Hybrid
•0000 UTC analyses initialized 36-hr 4-km forecasts
•EnKF: 4-km forecasts initialized from mean analyses
•Control: Interpolate 0000 UTC GFS analyses directly onto the domain and run forecasts
•GFS initialized from 3DVAR analyses in 2011
Cycling data assimilation: Hybrid/EnKF flowchart
Computational domain
WRF settings and physics•Forecast model: WRF-ARW (version 3.3.1)
•57 vertical levels, 10 hPa top
•Physics:
•Morrison double-moment microphysics
•RRTMG longwave and shortwave radiation
•MYJ PBL
•Tiedtke cumulus parameterization (20-km domain
only)
•NOAH land surface model
•Aerosol, ozone climatologies for RRTMG
Selected data assimilation settings•NCEP’s Gridpoint Statistical Interpolation (GSI) data assimilation system:
-GSI-3DVAR -GSI-hybrid -Ensemble square root Kalman filter (EnSRF)
•50 ensemble members
•Hybrid: 75% of the background errors from the ensemble, 25% from the static contribution
•Used posterior inflation for EnSRF and localization in both EnSRF and hybrid
Observation snapshot (0000 UTC 25 May)
Precipitation verification
•Focus on 4-km precipitation forecasts
•NCEP Stage IV observations as “truth”
•Verified hourly precipitation forecasts
•All precipitation statistics shown are aggregated over 44 4-km forecasts
•Fractions skill score (FSS) quantifies displacement errors
Precipitation BiasAggregated hourly over the
first 12 forecast hrsAggregated hourly
over18-36-hr forecasts
FSS: The first 12-hrs
0.25 mm/hr 1.0 mm/hr
5.0 mm/hr 10.0 mm/hr
FSS: Forecast hours 18-36
0.25 mm/hr 1.0 mm/hr
5.0 mm/hr 10.0 mm/hr
For more information…
• All of the previous material was summarized in this publication:
Schwartz, C. S., and Z. Liu, 2014: Convection-permitting forecasts initialized with continuously-cycling limited-area 3DVAR, ensemble Kalman filter, and “hybrid” variational-ensemble data assimilation systems. Mon. Wea. Rev., 142, 716–738, doi: 10.1175/MWR-D-13-00100.1.
Preview of new work• Recently, the exact same experiments
were performed but over a new period:– May 4 – June 30, 2013– 55 4-km forecasts
• Near identical configuration as before, except used Thompson microphysics
• Also performed dual-resolution hybrid analyses with a 4-km deterministic background and 20-km ensemble
Cycling data assimilation: Hybrid/EnKF flowchart
4-km
20-km
FSS: The first 12-hrs2013 experiments: FSS aggregated over
55 forecasts
0.25 mm/hr 1.0 mm/hr
5.0 mm/hr 10.0 mm/hr
FSS: The first 12-hrs2013 experiments: FSS aggregated over
55 forecasts
0.25 mm/hr 1.0 mm/hr
5.0 mm/hr 10.0 mm/hr
Dual-resolution hybrid: 4-km analyses and subsequent forecasts
FSS: Forecast hours 18-362013 experiments: FSS aggregated over
55 forecasts
0.25 mm/hr 1.0 mm/hr
5.0 mm/hr 10.0 mm/hr
Summary• Precipitation bias characteristics similar
in the cycling experiments• Differences in precipitation placement
evident– Hybrid and EnSRF performed best– Shows the benefit of flow-dependent
background errors
• Further improvement possible with high-resolution analyses
Example forecast
6-hr forecast initialized 0000 UTC 24 May 2011
Example forecast
30-hr forecast initialized 0000 UTC 24 May 2011