GSD Participation in Warn on Forecast 2012-2013David Dowell Curtis AlexanderStan Benjamin John BrownMing Hu Haidao LinEric James Brian JamisonPatrick Hofmann Joe OlsonTanya Smirnova Steve Weygandt

Assimilation and Modeling BranchNOAA/ESRL/GSD, Boulder, CO, USA

HRRR-CONUS

RapidRefresh

1. Case Studies: Storm-Scale Radar-Data Assimilation and Ensemble Forecasting• 27 April 2011• VORTEX2 cases

2. Website Development to Enhance Collaboration

3. Real-Time, Hourly-Updated Model Development• candidates for nested WoF systems• RAP and HRRR• NARRE and HRRRE

4. 2013 Priorities / Wish List

Outline

1. Test drive “state-of-the art” radar DA methods for a large number and variety of retrospective cases

2. Examine forecast ensemble spread resulting from storm-scale perturbations (and other sources) for real cases• complementary to OSSE work by Corey Potvin• ensemble sensitivity analysis

Retrospective Storm-Scale Ensemble Radar DAand Forecasting: Goals

Tuscaloosa, AL tornadosource: CBS 42 Birmingham, AL

27 April 2011 SupercellTornado Outbreak

Experiment Summary: 4/27/2011 Tornado Outbreak

45-member ARW ensembles (x=3 km) initialized from NAM and RAP600-km domain for these preliminary experiments

Velocity and reflectivity data assimilated every 3 min for 1 hKBMX, KDGX, KGWX, KHTX ; simple, automated quality controladditive storm-scale pert. during cycled radar DA -- only source of ensemble spreadWRF-DART ensemble adjustment Kalman filter

Ensemble forecast produced after radar DA

ensemble experiments

control experimentsKDGX

KGWX

KHTX

KBMXensemble

forecast

19Z 20Z 21Z 22Z 23Z 0Z

radar

DA

deterministic

forecast

19Z 20Z 21Z 22Z 23Z 0Z

NAM/

RAP

init.

NAM/

RAP

init.

Probability of Rotating Updrafts(2-5 km updraft helicity > 25 m2 s-2)

2000-2100 UTC

control experiment(no radar DA,

deterministic forecast)

radar DA, 0-1 hensemble forecast

NSSL CompositeReflectivity

2000 UTC

2100 UTC

500 km

Probability of Rotating Updrafts(2-5 km updraft helicity > 25 m2 s-2)

2000-2100 UTC

control experiment(no radar DA,

deterministic forecast)

radar DA, 0-1 hensemble forecast

NSSL CompositeReflectivity

2000 UTC

2100 UTC

removing spurious storms from analysis and forecaststill a challenge for radar DA

500 km

Probability of Rotating Updrafts(2-5 km updraft helicity > 25 m2 s-2)

2000-2100 UTC

control experiment(no radar DA,

deterministic forecast)

radar DA, 0-1 hensemble forecast

NSSL CompositeReflectivity

2000 UTC

2100 UTC

radar DA reorganizes storms in region where mesoscale environment (observed and simulated) was alreadysupportive of convective storms

500 km

Probability of Rotating Updrafts(2-5 km updraft helicity > 25 m2 s-2)

2000-2100 UTC

control experiment(no radar DA,

deterministic forecast)

radar DA, 0-1 hensemble forecast

NSSL CompositeReflectivity

2000 UTC

2100 UTC

radar DA introduces viable storms where they were needed; (CI enhanced through radar DA, maintenance supported

by mesoscale environment in model)

500 km

2100 UTC

2200 UTC

Probability of Rotating Updrafts(2-5 km updraft helicity > 25 m2 s-2)

2000-2100 UTC

control experiment(no radar DA,

deterministic forecast)

radar DA, 1-2 hensemble forecast

NSSL CompositeReflectivity

some storms introduced by radar DA persist; probabilities vary among storms

500 km

2145 UTC

NSSL CompositeReflectivity

Ensemble Forecast (105 min)Composite Reflectivity

Mean Spread

2145 UTC

NSSL CompositeReflectivity

Ensemble Forecast (105 min)Composite Reflectivity

Mean Spread

southern storm: high mean, low spread

2145 UTC

NSSL CompositeReflectivity

Ensemble Forecast (105 min)Composite Reflectivity

Mean Spread

southern storm: high mean, low spread

northern storm: low mean, high spread

Ensemble Sensitivity Analysis (work in progress)

Mean Spread

ensemble-based correlations between initial conditions (and/or model parameters) and forecast metric

method applied previously to larger scales (Hakim and Torn 2008)

ongoing work to apply to convective scale•What types of storm-scale perturbations resulted in the northern storm persisting in the forecast?

1. 18 May 2010 Dumas, Texas Supercell• collaboration with Texas Tech University (Chris Weiss, Tony

Reinhart, Pat Skinner)

2. 5 June 2009 Goshen County, Wyoming Supercell• collaboration with Penn State University (Jim Marquis et al.)

foci: assimilation of radar and surface observations into high-resolution models, diagnosis of severe storm processes

VORTEX2 Case Studies

photo by David Dowell for VORTEX2

18 May 2010 Dumas, TX Storm: Observations and Analysis

assimilation of KAMA, SR1, DOW6, and DOW7 data

verification ofsurface fields

KAMA

DOW7 DOW6SR1

StickN

et

KAMA

80 km

StickN

et

, lowest model level

2 m AGL (StickNet) 8 m AGL (simulation)

2 K contour interval

Perturbation* Temperature (K) 2300 UTCStickNet (circles) and Ensemble Mean (outside circles)

* relative to model’s base state

10 km

2 m AGL (StickNet) 8 m AGL (simulation)

2 K contour interval

Perturbation* Temperature (K) 2344 UTCStickNet (circles) and Ensemble Mean (outside circles)

* relative to model’s base state

10 km

2 m AGL (StickNet) 8 m AGL (simulation)

3 m s-1 contour interval

Westerly (u) Wind Component 2328 UTCStickNet (circles) and Ensemble Mean (outside circles)

10 km

Web Graphics for Warn-on-Forecast Experiments

WRF NetCDF

GRIB2

.png

Unipost

NCL

web display

tool for enhancing our collaboration,leveraging community software and scripts developed previouslyfor RAP-HRRR (acknowledgments: Brian Jamison, Susan Sahm)

quick, easy sharing of results from retrospective and real-time

experiments

rough drafts of websites:rapidrefresh.noaa.gov/WoFMeso/rapidrefresh.noaa.gov/WoFSS/

http://rapidrefresh.noaa.gov/WoFMeso/

http://rapidrefresh.noaa.gov/WoFMeso/

http://rapidrefresh.noaa.gov/WoFMeso/

http://rapidrefresh.noaa.gov/WoFSS/

http://rapidrefresh.noaa.gov/WoFSS/

• Rapid Refresh

• High-Resolution Rapid Refresh

• NARRE• hourly-updated ensemble, 10-12 km, hybrid/EnKF DA• 2015-2016?

• HRRRE• hourly-updated ensemble, 3 km• 2017-2018?

All are candidate models for nested WoF systems.

Hourly-UpdatedNOAA NWP Models

13km Rapid Refresh

3km HRRR

RAP and HRRR Changes 2011-2012

ModelData

Assimilation

RAP-ESRL(“RAP v2”)

(13 km)

WRFv3.3.1+ Numerics changes: (w-damp upper bound conditions, 5th-order vertical advection)Physics changes: (microphysics, land-surface, PBL) MODIS land use, fractional 3005 min shortwave radiation New reflectivity diagnostic

GSI merge with trunkSoil adjustment Temp-dep radar- hydrometeor buildingPW assim modsCloud assim modsTower/nacelle/sodar observationsGLD360 lightning

HRRR (3 km)

WRFv3.3.1+ Numerics changes: (w-damp upper bound conditions, 5th-order vertical advection)Physics changes: (microphysics, land-surface, PBL) MODIS land use, fractional 3005 min shortwave radiation New reflectivity diagnostic

Eastern US, Reflectivity > 25 dBZ11-21 August 2011

MUCH reduced bias for HRRR 2012, similar CSI

40 km CSI( x 100)

2011 HRRR2012 HRRR

Optimal

HRRR Verification 2011 vs 2012

13 km bias (x 100)

Model Data Assimilation

RAP-ESRL(13 km)

WRFv3.4.1+ incl. physics changes (convection, snow-radiation fix)Numerics changes: 6th-order diffusion near surfacePhysics changes: MYNN PBL scheme 9-layer RUC LSM (from 6-layer) Modified roughness length RRTMG short/longwave radiation Thompson microphysics update

Merge with GFS trunk

GFS ensemble background error cov

Cloud fraction assimilation Full column cloud buildingImproved hydrometeor analysis

Radiance bias correction

Reduced observation error(sharper inversions, low-level thermo)

HRRR (3 km)

WRFv3.4.1+ incl. physics changes (convection, snow-radiation fix)Numerics changes: 6th-order diffusion near surfacePhysics changes: MYNN PBL scheme 9-layer RUC LSM (from 6-layer) Modified roughness length RRTMG short/longwave radiation Thompson microphysics update

3 km/15 min reflectivity assimilation3 km cloud cycling3 km land-surface cycling

RAP and HRRR Changes 2013

3-km Interp

2013: Cycled Reflectivity at 3 km

GSI 3D-VAR

Cloud Anx

DigitalFilter

1 h

r fc

st

18 hr fcst

3-km Interp

GSI 3D-VAR

Cloud Anx

DigitalFilter

1 h

r fc

st

18 hr fcst

GSI 3D-VAR

Cloud Anx

DigitalFilter

18 hr fcst

3 km HRRR

13z 14z 15z13 km RAP

15 hr fcst1 hr pre-fcst

Refl Obs

1-hr Reduction In Latency for 14z HRRR

Additional Positive Contribution to HRRR (3-km) Forecasts from Reflectivity DA in HRRR

14-day June 2011 retrospective periodverification over eastern half of US (widespread convective storms)

Critical Success Index (CSI) for 25-dBZ Composite Reflectivity

upscaled to 40-km grid

reflectivity DA in RAP + HRRR (for 1 h)reflectivity DA in RAP only

Time-lagged ensembleModel InitTime Example: 13z + 2, 4, 6 hour

HTPF

Forecast Valid Time (UTC)

11z 12z 13z 14z 15z 16z 17z 18z 19z 20z 21z 22z 23z

13z+212z+311z+4

13z+412z+511z+6

13z+612z+711z+8

HTPF2 4 6

18z

17z

16z

15z

14z

13z

12z

11z

Model runs used

model has 2hr latency

The HCPF and HTPFHRRR Convective Probabilistic Forecast (HCPF)

HRRR Tornadic Storm Probabilistic Forecast (HTPF)

Use time-lagged ensemble to estimate liklihood of

convection and tornado production

Identification of updraft rotation using model forecast fields:• Intensity – Maximum Updraft Helicity 2-5 km AGL ≥ 25 m2 s-2

• Time – Two hour search window centered on valid times• Location – Searches within 45 km (15 gridpoints) of each point

for each member• Members – Three consecutive HRRR initializations

HTPF = # grid points matching criteria over all members

total # grid points searched over all members

Tornadic Storm Probability (%)

Reflectivity (dBZ)

13z + 09hr fcstValid 22z 27 April 2011

27 April 2011 Storm Reports

Observed Reflectivity22z 27 April

Example: 27 April 2011

Tornado = Red Dots

Valid 1200-1200 UTC 28 Apr

1. National quality-controlled WSR-88D datasets – including Doppler velocity – for retrospective and real-time radar DA experiments• nonmeteorological data removal utilizing polarimetric

information

2. Collaboration on regional storm-scale radar DA and ensemble forecasting for retrospective periods ~1 week• parameter space: multiple radar DA methods, multiple

resolutions, …• common radar observations for input, model configuration,

forecast verification• 3-km ensemble for background

Priorities / Wish List for 2013

RAP and HRRR Model Details

Model Version Assimilation Radar DFI Radiation Microphysics Cum

Param PBL LSM

RAP-ESRL

WRF-ARW

v3.3.1+GSI-3DVar Yes RRTM/

GoddardThompson

v3.3.1G3 +

Shallow MYJ RUCv3.3.1

HRRRWRF-ARW

v3.3.1+

None: RAP I.C. No RRTM/

GoddardThompson

v3.3.1 None MYJ RUCv3.3.1

Model Domain Grid Points

Grid Spacing

Vertical Levels

Boundary Conditions Initialized

RAP-ESRL

North America

758 x 567 13 km 50 GFS Hourly

(cycled)

HRRR CONUS 1799 x 1059 3 km 50 RAP-ESRL Hourly

(no-cycle)

HRRRRAP

observations assimilated with GSI (3DVar) into experimental RAP at ESRLrawinsonde; profiler; VAD; level-2.5 Doppler velocity; PBL profiler/RASS; aircraft wind, temp, RH; METAR; buoy/ship; GOES cloud winds and cloud-top pres; GPS precip water; mesonet temp, dpt, wind (fall 2012); METAR-cloud-vis-wx; AMSU-A/B/HIRS/etc. radiances; GOES radiances (fall 2012); nacelle/tower/sodar

diabatic digital filter initialization with radar-reflectivity and lightning (proxy refl.) data

Positive Contribution to HRRR (3-km) Forecastsfrom Reflectivity DA (DDFI) in Parent (13-km) RAP

11-20 August 2011 retrospective periodverification over eastern half of US (widespread convective storms)

Critical Success Index (CSI) for 25-dBZ Composite Reflectivity

upscaled to 40-km grid

HRRR with RAP reflectivity DA (real time)

HRRR without RAP reflectivity DA

Latent Heating (LH) Specification

-60

-45

-30

-15

0

-60 -45 -30 -15 0Model Pre-Forecast Time (min)

Temperature Tendency (i.e. LH) = f(Observed Reflectivity)LH specified from reflectivity obs applied in four 15-min periodsThe observations are valid at the end of each 15-min pre-fcst periodNO digital filtering at 3-kmHour old mesoscale obsLatency reduced by 1 hr

LH = Latent Heating Rate (K/s)p = PressureLv = Latent heat of vaporizationLf = Latent heat of fusionRd = Dry gas constantcp = Specific heat of dry air at constant pf[Ze] = Reflectivity factor converted to

rain/snow condensatet = Time period of condensate formation

(600s i.e. 10 min)

Forward integration, full physicsApply latent heating from radar reflectivity, lightning data

Diabatic Digital Filter Initialization (DDFI) -20 min -10 min Init +10 min

RR model forecast

Backward integration,no physics

Obtain initial fields with improved balance, vertical circulations associated withongoing convection

The model microphysics temperature tendency is replaced with a reflectivity-based temperature tendency. Dynamics respond to forcing.

Analysis noise is reduced by digital filtering.

Anticipated Progression of RAP and HRRR Radar DA

now: radar DA in RAP (13 km) only

near future (proposed): continued radar DA in RAP (13 km);short period of radar DA in HRRR (3 km) before HRRR forecast begins

future: cycling with all obs (including radar) on HRRR (3-km) grid3DVar and reflectivity-based temperature tendencyhybrid / ensemble DA and forecasting

RAP13 km fcst

DDFI

obs

radardata

fcst

HRRR3 km fcst

t02 h t01 h t0

interpolation

radardata

radardata

radardata

radardata

3DVar + cloud analysis

… …

HRRR reflectivity DA•same formulation of reflectivity-based temperature tendency as in RAP•no digital filter

ExperimentComparison

(2) HRRRinitialized

“with 3-kmradar DA”

RAP13 km fcst fcst

DDFI

obs

radardata

HRRR3 km fcst

t02 h t01 h t0

interpolation

3DVar + cloud analysis

… …

RAP13 km fcst

DDFI

obs

radardata

fcst

HRRR3 km fcst

t02 h t01 h t0

interpolation

radardata

radardata

radardata

radardata

3DVar + cloud analysis

… …

(1) HRRRinitialized

“without 3-kmradar DA”

CompositeReflectivity

2300 UTC11 June 2011

1-h fcstwith 3-kmradar DA

1-h fcstwithout 3-km

radar DA

mature convective systems benefit particularly from

subhourly radar DA

observations

1000 km

Model and Data Assimilation

WRF model run as “cloud model”homogeneous base state; no parameterizations for PBL, surface layer, radiation, …x = 1000 m, z = 50 to 500 mLin et al. (1983) precipitation microphysics, configured for hail and large raindrops

n0(hail) = 4 × 104 m-4 n0(rain) = 1 × 106 m-4

weaker cold pool (Gilmore et al. 2004) than for default scheme, but still strong…

Radar data assimilated every 2 min for 3 hoursData Assimilation Research Testbed (DART) ensemble Kalman filterKAMA reflectivity and Doppler velocity throughout periodmobile Doppler velocity (SR1, DOW6, DOW7) when available

60-member ensemblevariability from (1) random perturbations to base-state wind profile and (2) random

local perturbations to horizontal wind, temperature, and humidity (dewpoint)“analysis” (“simulation”) is prior ensemble mean

Verification of model surface fields with StickNet observationsfirst, determine if the model is capable of simulating the storm and environmental

features of interest in radar-DA-only experimentslater, assimilate surface (StickNet and MM) observations into “final” analysis

2 m AGL (StickNet) 8 m AGL (simulation)

3 m s-1 contour interval

Westerly (u) Wind Component 2314 UTCStickNet (circles) and Ensemble Mean (outside circles)

10 km

Summary of Surface Verification

Overall patterns are reasonable; differences involve the details.

Cooling in the downshear precipitation core is too weak in the simulation.

consistent with perceived errors in single-moment microphysics schemes

The main body of the simulated cold pool is generally too cold and too widespread.

consistent with perceived errors in single-moment microphysics schemes

StickNet winds (2 m AGL) are generally weaker than model winds (8 m AGL).

implications for diagnosis of “baroclinic”, “barotropic”, and “friction-induced” contributions to mesocyclone rotation

model lower boundary condition currently free slip; more realistic surface and boundary layer needed for simulation and data assimilation

Challenges of Storm-Scale DA and NWP

Large radar datasets in need of quality control

Large model grids1000’s of km wide, grid spacing ~1 km

Model error and predictabilityunresolved processes: updraft, downdraft, precipitation microphysics, PBL, …predictability time scale ~10 min for an individual thunderstormforecast sensitivity to small changes in initial conditions (e.g., water vapor)

Flow-dependent background-error covariancesno quasi-geostrophic balance on small scalesretrieving unobserved fields

Verifying forecasts (to improve future ones)unobserved fields, isolated phenomena

All tasks (preprocessing and assimilating obs, producing forecasts) must occur quickly for the forecast to be useful in real time!

within an hour for some applicationswithin minutes for warning guidance

190 radars

volumes every10 min or less