Impacts of Running-In-Place on LETKF analyses and forecasts of the 24 May 2011 outbreak Impacts of Running-In-Place on LETKF analyses and forecasts of.
Post on 13-Dec-2015
215 Views
Preview:
Transcript
Impacts of Running-In-Place on LETKF analyses and forecasts of the 24 May 2011 outbreak
Corey K. Potvin1,2, Louis J. Wicker2, and Therese E. Thompson1,3
1Cooperative Institute for Mesoscale Meteorological Studies
2NOAA/OAR National Severe Storms Laboratory3School of Meteorology, University of Oklahoma
WoF Challenge: “Spin-Up” Problem
• During first several analysis cycles of storm, convective-scale state & error covariances often poor
• Initial radar DA thus inefficient
• Goal: accelerate spin-up longer forecast lead time
• Spin-up not just at start of DA – new storms can form after initial development
Running-In-Place (RIP; Kalnay & Yang 2010)
• Under ideal conditions, obs should be used once• During spin-up, re-assimilating obs extracts additional
information• Step 0: Regular LETKF update• Step 1: Use LETKF weights at current analysis time tn to
update xa(tn-1) - “no-cost” smoother
• Step 2: Covariance inflation at tn-1
• Step 3: Integrate ensemble tn-1 tn • Repeat until RMS difference between obs & forecasts
converge (or max # iters reached)
Previous Work
• Kalnay and Yang (2010) – RIP rapidly spins up idealized QG model
• Yang et al. (2012a) – RIP helpful even given strong nonlinearity (Lorenz-63 model)
• Yang et al. (2012b) –perfect-model typhoon OSSEs (WRF)
• Yang et al. (2013) – RIP applied to real typhoon• Wang et al. (2013) – iterative EnSRF; perfect- and
imperfect-model supercell OSSEs (WRF)• No published real-supercell experiments
EnKF Configuration
• NSSL-LETKF – uses Miyoshi’s LETKF core• WRF v3.4.1 -- Δ=3km, 170 × 170 × 51 points,
Thompson microphysics• 36 members; 5-min analysis cycles• 3 WSR-88D’s (Δ=6km); objectively QC’d• No mesonet assimilation• Additive noise (Dowell & Wicker 2009) + adaptive
multiplicative inflation (Miyoshi 2011; Hunt et al. 2007)• GEFS-NME-based ensemble initial condition
1845 UTC Experiments
• First storm echoes• IC very poor• Best forecasts obtained: – with 3 vs. 1 RIP iterations– stopping RIP after 19 UTC
MRMS 1850 UTC
Multi-Radar/Multi-Sensor (MRMS) reflectivity mosaic
at 2 km AGL
1910 & 1920 UTC AnalysesRIP greatly accelerates spin-up of dBZ, w, ζ
2 km
AG
L dB
Z
1905 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 1905 Z and 2005 ZDots = interpolated 19-20 UTC NSSL rotation tracks > .005 s-1, .010 s-1, .015 s-1
CNTL RIP
1910 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 1910 Z and 2010 ZDots = interpolated 19-20 UTC NSSL rotation tracks > .005 s-1, .010 s-1, .015 s-1
RIPCNTL
1915 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 1915 Z and 2015 ZDots = interpolated 19-20 UTC NSSL rotations > .005 s-1, .010 s-1, .015 s-1
CNTL RIP
2000 UTC Experiments
• Begin DA once storms already mature• Mesoscale background storms displaced
2005 NME prior MRMS
Cold pool problem
• Default RIP (1 or 3 iters) generates too-cold cold pools• Likely from repeated assimilation of large-innovation dBZ in
same locations – analysis increments not retained during forecast cycle, as in Dowell et al. (2011)
• What helped: – Use only 1 RIP iteration– Don’t update θ – but colds pools still too cold, indicating other
covariances problematic– Only assimilate Vr, and only update u, v, w – still too cold!– Increase obs error estimates (“gentle” approach)– mitigates cold
pool bias as well as no-theta update
• Final solution: 1 iter with σVr = 4, σdBZ = 8
RIP_gentle RIP_default
CNTL MRMS
2010 UTC Analyses
• Spin-up in gentler approach nearly as fast as in default
• w, ζ improved faster than dBZ
2015 UTC AnalysesRIP_gentle RIP_default
CNTL MRMS
• Spin-up in gentler approach nearly as fast as in default
• w, ζ improved faster than dBZ
2005 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 2005 Z and 2105 ZDots = interpolated 20-21 UTC NSSL rotation tracks > .005 s-1, .010 s-1, .015 s-1
CNTL RIP
2010 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 2010 Z and 2110 ZDots = interpolated 20-21 UTC NSSL rotation tracks > .005 s-1, .010 s-1, .015 s-1
CNTL RIP
2015 UTC 1-h Forecasts
Red = neighborhood (3 × 3) ensemble prob ζ > .005 s-1 somewhere over lowest 3 kmPink = tornado damage pathsContours = 2 km AGL 40 dBZobs at 2015 Z and 2115 ZDots = interpolated 20-21 UTC NSSL rotation tracks > .005 s-1, .010 s-1, .015 s-1
CNTL RIP
Conclusions
• RIP can accelerate spin-up in radar DA• Added forecast value restricted to 2-3 analysis
cycles in this case• Inflated error variances may be useful, at least
when mean IC very poor
RIP - Outstanding Questions
• Better to restrict to large-innovation regions?• Substantial improvement over simply re-
assimilating obs (“poor-man’s RIP”) or Quasi-Outer Loop?
• Better or worse than directly forcing updrafts (e.g., thermal bubbles, dBZ-based w nudging)?
• How does impact change given less favorable environment?
top related