EnKF Assimilation of Simulated HIWRAP Radial Velocity Data
description
Transcript of EnKF Assimilation of Simulated HIWRAP Radial Velocity Data
EnKF Assimilation of Simulated HIWRAP Radial Velocity Data
Jason Sippel and Scott Braun - NASAs GSFC
Yonghui Weng and Fuqing Zhang - PSU
Global Hawk flown for 2012-2014 hurricane seasons– 26-h flight time; 19-km altitude– ‘Over-storm’ payload:
• HIWRAP Doppler radar (Ku & Ka)• HAMSR sounder (Tb for qv & T above
precip)• HIRAD radiometer (sfc. winds)
– ‘Environmental’ payload:• Dropsondes (u, v, qv, T)• S-HIS sounder (clear-air radiances for
qv & T)• Cloud Physics Lidar (aerosols)• (maybe) TWILITE Doppler Lidar
Motivation: NASA’s HS3 CampaignBackground
Global Hawk at NASA’s Dryden hangar
Motivation: Past WRF-EnKF success• Same ensemble Kalman filter
(EnKF) setup as used for:– WSR-88D Vr assimilation in
Hurricane Humberto (Zhang et al. 2009, MWR)
– P3 Vr assimilation in Hurricane Katrina (Weng and Zhang 2011, MWR)
– Penn. State University HFIP real-time WRF/EnKF
• Proven successful, even with difficult storms
Background
Humberto: Obs vs. EnKF analysis
Objectives1. Generate a 48-h ensemble without data assimilation
2. Select ‘truth’ realizations for simulated data experiments
3. Assimilate simulated HIWRAP observations with an ensemble Kalman filter (EnKF)
4. Assess quality of analyses and forecasts
WRF-EnKF system• EnKF from Zhang et al. (2009)
• WRF-ARW V3.1.1, 27/9/3 km
• 30-member ensemble, IC/BCs from WRF-VAR + GFS
• Ensemble integrated 12 h to generate mesoscale covariance
• WSM-6 mp for assimilation; GSFC for truth (model error)
Methods
Model domains
Selecting ‘truth’ realizations
Realizations selected to test EnKF performance in face of:
• Small error of the priorHow much improvement does the EnKF offer when the forecast is already pretty good? (NODA1)
• Large error of the priorHow well can the EnKF correct when the truth is unlike most of the prior? (NODA2)
Methods
‘Truth’ realizations and NODA forecasts
‘Truth’ simulation flight tracks• Instantaneous scans every ~28
km; observation cones slightly overlap at surface
• Data grouped into 1-h flight segments from same output time; ~1900 obs/hr
• Add 3 m/s random error, only assimilate when attenuated dBZ > 10
Methods
50°
~ 3 km
~ 3 km
19.0 km
Observations every ~3 km on cone surface
Results: CTRL analyses evolutionOSSE Results
• Intensity metrics: Generally small corrections due to small initial error in NODA
• Position: more noticeable correction, particularly for CTRL2
Results: Analysis error reductionOSSE Results
• EnKF reduce RM-DTE > 80% after 13 cycles in both cases [DTE = 0.5 × (u`u` + v`v` + Cp/Tr × T`T`), prime is difference from truth]
• CTRL2 has larger and more widespread error reduction than CTRL1
Comparison of RM-DTE differences
Results: Deterministic forecastsOSSE Results
• Forecast error is reduced relative to NODA in both cases, particularly from 36-48 h
• NODA2 needs more time to produce better analyses (i.e., that produce ‘good’ forecasts)
Results: Ensemble forecastsOSSE Results
Similar to deterministic forecast results…
Note huge benefit of cycling
Truth
13 HIWRAP cycles12 HIWRAP cycles
+ 1 HIWRAP/HIRAD cycle
Benefit of multiple data sources
Improved analysis of inner-core winds and wavenumber 1 asymmetry with complement of simulated HIRAD data
OSSE Results
SummaryHIWRAP data appears to be useful for EnKF analyses and subsequent forecasts of a hurricane
• Strong analysis error reduction, particularly for a poor first guess
• Notable forecast improvements after just one assimilation cycle in CTRL1
• A longer assimilation window (i.e., Global Hawk time scale) appears to benefit forecast more when the first guess is poor
• Future work will assess impact of multiple platforms at once as well as multiple instruments (e.g., HIRAD and dropsondes)
Results: Re-examining flight tracks
Compare two different flight track philosophies:
• CTRL: 360-km legs that approximately span entire precipitation region
• VLEG: Single 360-km pattern followed by two complete patterns with 180-km legs, then repeat (i.e., focus more on center)
Methods (again)
Results: CTRL/VLEG comparisonOSSE Results
• Average CTRL RM-DTE near center is ~3-3.5 m/s
• VLEG systematically reduces near-center error more but error farther out less
Results: CTRL/VLEG comparisonOSSE Results
• All CTRL forecasts better than NODA• Differences between CTRL and VLEG forecasts on par
with differences as a result of obs error (i.e., AERR)