Storm-Scale Modeling with HRRR toward Warn-On-Forecast NOAA Earth System Research Laboratory GSD/AMB Stan Benjamin, Curtis Alexander, David Dowell, Steve Weygandt, Ming Hu, Tanya Smirnova, John Brown, Patrick Hofmann, Eric James, Ed Szoke, Technical Workshop on Numerical Guidance to Support WoF 5 February 2013
82
Embed
Storm-Scale Modeling with HRRR toward Warn-On-Forecast
Technical Workshop on Numerical Guidance to Support WoF 5 February 2013. Storm-Scale Modeling with HRRR toward Warn-On-Forecast. NOAA Earth System Research Laboratory GSD/AMB Stan Benjamin, Curtis Alexander, David Dowell, Steve Weygandt, Ming Hu, Tanya Smirnova , John Brown, - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Storm-Scale Modeling with HRRR toward Warn-On-Forecast
NOAA Earth System Research LaboratoryGSD/AMB
Stan Benjamin, Curtis Alexander, David Dowell, Steve Weygandt, Ming Hu, Tanya Smirnova, John Brown,
Patrick Hofmann, Eric James, Ed Szoke, Haidao Lin
Technical Workshop on Numerical Guidance to Support WoF5 February 2013
RUC to Rapid Refresh transition
NCEP Production Suite Review 4-5 December 2012Rapid Refresh / HRRR 2
• North American domain • WRF-ARW Model
• GSI –• Unipost
Gridpoint Statistical Interpolation
• CONUS domain • RUC Model
• RUC 3dvar• RUC post
13km Rapid Refresh
13km RUC
3km HRRR
Hourly updated models
WRF-ARW / GSI enhancements for RAP
NCEP Production Suite Review 4-5 December 2012Rapid Refresh / HRRR 3
• GSD contributions to WRF-ARW – Grell 3-D cumulus scheme (updated each version)– RUC (Smirnova) land-surface model (“ “ “)– Diabatic Digital Filter Initialization (with NCAR)– DFI-radar-latent heat reflectivity assimilation– Modified MYNN PBL (v3.5)– Latent heat limit to allow much longer time steps
RAP version 2 ChangesGFS ensemble background error covariance
RAPdev1 shows less 10m wind bias compared to RAP-primary, near 0 in daytime, from 0.8 m/s down to ~0.5 at night (an improvement)
RAPdev1 shows more accurate 2m temp diurnal bias, less exaggerated diurnal cycle than evidence in RAP-primary (RR1h)
RUC LSM with 9 levels – implemented in RAP-dev1 on 18 Oct12:Zs= 0, 1, 4, 10, 20, 40, 100, 160, 300 cm RUC LSM with 6 levels –in RAP primary ESRL, RAP-oper-NCEP:Zs= 0, 5, 20, 40, 160, 300 cm
3h fcst - 10m wind bias vs. METARs – e. US
3h fcst – 2m temp bias vs. METARs – e. US
GSI 3D-VarHMObs
RAP-2012 – using GSI using fixed background error covariance
GSI 3D-Var
Obs
Cloud Anx
DigitalFilter
HMObs
ReflObs18 hr fcst
GSI 3D-Var
Obs
Cloud Anx
DigitalFilter
1 hr
fcst
HMObs
ReflObs18 hr fcst
Obs
Cloud Anx
DigitalFilter
ReflObs18 hr fcst
13z 14z 15z13 km RAP
1 hr
fcst
GSI HybridHMObs
RAP-2013 GSI hybrid using Global Ensemble
GSI Hybrid
Obs
Cloud Anx
DigitalFilter
HMObs
ReflObs18 hr fcst
GSI Hybrid
Obs
Cloud Anx
DigitalFilter
1 hr
fcst
HMObs
ReflObs18 hr fcst
Obs
Cloud Anx
DigitalFilter
ReflObs18 hr fcst
13z 14z 15z13 km RAP
1 hr
fcst
80-member GFS EnKF Ensemble forecast valid at
15Z (9-h fcst from 6Z)
Available 4 times a day valid at 03, 09, 15, 21Z
Real-time test for RAP hybrid usingbkg error cov from GDAS hybrid DA ensemble
RMS profile for 3h forecasts
• Compare RAP development with GSI hybrid to RAP primary cycle with GSI-3dvar– 30-day real-time test from 22 Nov – 22 Dec 2012 – GSI hybrid with half static BE and half BE from GFS
3 km/15 min reflect assimilation3 km cloud cycling3 km land-surface cycling
RAP and HRRR Changes 2013
BackgroundRadar
Specification of Hydrometeors
Scale at which Latent Heating
is appliedDimensionality Updated
2012 HRRR model
initialization13-km RAP No 13-km 3-D Hourly
2013 HRRR model
initialization13-km RAP yes
3km in 60min spin-up (also using 13km
radar-LH-DFI)3-D Hourly
Rapidly Updating
Analysis (RUA-HRRR)
3-km HRRR1 hr fcst Yes None 3-D Hourly
Real-TimeMeso Analysis (RTMA-HRRR)
3-km HRRR1 hr fcst No None 2-D
Hourly(15 min
planned)
Time-Lagged HRRR (HCPF)
3-km HRRRFcsts No Same as HRRR 2-D Hourly
HRRR Real-Time Exper Products
RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 38
ESRL GSI applications at 3km1800 x 1060 (x 50 levs)
Application CPU cores (zeus) Run time
3km HRRR (3d, radar, sat only, no var solver) 240 8 min
3km RUA (no var) 84 6-7 min
3km RTMA (2d, sfc obs only) 60 7 min
Full 3km 3dvar (full RAP obs data set) 300 20-30 min
3-km Interp
2013: Cycled Reflectivity at 3 km
GSI 3D-VAR
Obs
Cloud Anx
DigitalFilter
1 hr
fcst
HMObs
ReflObs18 hr fcst
3-km Interp
GSI 3D-VAR
Obs
Cloud Anx
DigitalFilter
1 hr
fcst
HMObs
ReflObs18 hr fcst
GSI 3D-VAR
Obs
Cloud Anx
DigitalFilter
HMObs
ReflObs18 hr fcst
3 km HRRR
13z 14z 15z13 km RAP
15 hr fcst1 hr pre-fcst
Refl Obs1-hr Reduction In Latency for 14z HRRR
Latent Heating (LH) Specification
Obse
rved
3-D
Rad
ar R
eflec
tivity
Ti
me
(min
)
-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)RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 41
Many options for weighting LH specification vs model LHTwo approaches including:(a)100% specification for the entire pre-forecast hour (b) Time-varying with linear ramp down to 0% specification at 1 hr
Option (b) permits more “free model” behavior before additional DA
Latent Heating (LH) Specification
RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 42
0-h fcstwithout 3-km
radar DA
Obs0000 UTC11 June
2011
0-h fcstwith 3-kmradar DA(fixed)
0-h fcstwith 3-kmradar DA(ramp)
Benefit from pre-forecast 3-km model integration
1-h fcstwithout 3-km
radar DA
Obs0100 UTC11 June
2011
1-h fcstwith 3-kmradar DA(fixed)
1-h fcstwith 3-kmradar DA(ramp)
Convective systems more mature even by 1-hr
3-day retrospective period 09-12 June 2011 (36 runs)Forecasts every 2 hours
> 25 dBZ Composite ReflectivityEastern half of US
Upscaled to 40-km grid
With 3-km fixed radar DAWith 3-km ramp radar DAWithout 3-km radar DA
Bias = 1.0
Native3-km grid
Greatly improved CSI and BIAS between 0-1 fcst hr
Benefit persists until 4 hrsVery similar skill at longer lead times
HRRR Reflectivity Verification
14-day retrospective period June 2011 (160 runs)Forecasts every 2 hours
> 25 dBZ Composite ReflectivityEastern half of US
Upscaled to 40-km grid
With 3-km ramp radar DAWithout 3-km radar DA
Bias = 1.0
Native3-km grid
Greatly improved CSI and BIAS between 0-1 fcst hr
Benefit persists until 4 hrsVery similar skill at longer lead times
building All levels with sufficiently high obs refl (> 0 dBZ)
All levels (same as for Tsfc < 5C)
None None Specify qs at all levels with sufficiently high reflectivity (> 0 dBZ)
Set to max diagnosed qs only at that obs max refl level.
clearing None None None Clear except allowdiagnosed qs to be retained at max obs refl level
Clear qr/qs/qg in any volumes with obs refl < 0 dBZ
Clear qr/qs/qg in any volumes with obs refl < 0 dBZ
Precip hydrometeor (qr/qs/qg) clearing/building from 3-D radar
reflectivity data
2m temps below are clearly < 4 deg C,In fact, usually < -10 deg C
Observed lake effect snow band is clearly over Lake Ontario
But the lake effect snow band in the HRRR 0h is located over land (NY state) south of Lake Ontario
With snow building/clearing
Without snow building/clearing
Improved snowfall/snowcover1- hr Forecasts valid 14 UTC 25 Jan 2013
Improved 0-hr reflectivity analysisIntroduced snow building/clearing inESRL RAPv2 (primary) and implicitlyESRL HRRR (primary and dev1)At 06 UTC 29 January 2013
0-hr13-km> 25 dBZ
0-hr20-km> 25 dBZ
Improved reflectivity forecasts
1. 0-1 hr forecasts for HRRR-primary and ESRL-RAP-primary are now performing fairly well2. Addition of 3km/15min assimilation (HRRR_dev1) helps further still at 0-2 hrs
20-km> 30 dBZ
20-km> 35 dBZ
BackgroundRadar
Specification of Hydrometeors
Scale at which Latent Heating
is appliedDimensionality Updated
2012 HRRR model
initialization13-km RAP No 13-km 3-D Hourly
2013 HRRR model
initialization13-km RAP yes
3km in 60min spin-up (also using 13km
radar-LH-DFI)3-D Hourly
Rapidly Updating
Analysis (RUA-HRRR)
3-km HRRR1 hr fcst Yes None 3-D Hourly
Real-TimeMeso Analysis (RTMA-HRRR)
3-km HRRR1 hr fcst No None 2-D
Hourly(15 min
planned)
Time-Lagged HRRR (HCPF)
3-km HRRRFcsts No Same as HRRR 2-D Hourly
HRRR Real-Time Exper Products
RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 55
3-km Interp
Hourly HRRR RTMA
15 hr fcst 3-km Interp
15 hr fcst 3-km Interp
15 hr fcst
3 km HRRR
13z 14z 15z
GSI2D-VAR
1 hr
fcst
GSI2D-VAR
1 hr
fcst
ObsObs
3-km HRRR RTMA
HRRR AnxRTMA
1-hr HRRR Fcst(Background)Valid 19 UTC30 Nov 2012
0-hr HRRR AnalValid 19 UTC30 Nov 2012
10 m Winds
Analysis Increments
RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 57
3-km Interp
Hourly HRRR Rapidly Updating Anx
15 hr fcst 3-km Interp
15 hr fcst 3-km Interp
15 hr fcst
3 km HRRR
13z 14z 15z
Cloud Anx
HMObs
1 hr
fcst
Cloud Anx
HMObs
1 hr
fcst
3-km HRRR RUA
1-hr HRRR Forecast (Background)Valid 22 UTC
03 November 2012
0-hr HRRR Analysis (RUA)Valid 22 UTC
03 November 2012
Obs22 UTC
03 Nov 2012
GSICloud Anx
RapidlyUpdatingAnalysis(RUA)
Specifies HydrometeorsFrom Radar Observations
RUA Meeting 25 Jan 2013High-Resolution Rapid Refresh 59
10-11 hr fcst
09-10 hr fcst
08-09 hr fcst
11-12 hr fcst
10-11 hr fcst
09-10 hr fcst
Forecasts valid 21-22z 27 April 2011 Forecasts valid 22-23z 27 April 2011
All six forecastscombined to formprobabilities valid22z 27 April 2011
– Major improvement over 2011 HRRR, storm coverage/accuracy
• HRRR – 2013– 3km/15min radar assimilation– Initialized from RAP-2013– Available 45 min earlier, much more
accurate 0-15h storm forecasts, more reliable 2-computer
NWS-NCEP - operational
• Implemented 1 May 2012
• RAPv2 - Scheduled to be implemented in spring 2014
• HRRR – estimated 2015
2015-2016?North American Rapid Refresh
ENSEMBLE (NARRE)• NMMB (from NCEP) & ARW (from ESRL) dynamic cores• Common NAM parent domain at 10-12 km• Initially ~6 member ensemble made up of equal numbers of
NMMB- & ARW-based configurations• Hourly updated with forecasts to 24 hours• EnKF/hybrid data assimilation for ensemble member
initialization• Use 3 different physics suites for ensemble members
but not for data assimilation– RAP physics (Thompson microphysics, Grell-3d cu, MYJ or
MYNN PBL)– NMM physics (Ferrier microphysics, BMJ/Janjic cu, MYJ PBL)– NCAR physics (let MMM recommend - WSM,YSU,KF?)– Consider stochastic physics for a single physics suite
69Sept 2012 – modifications from Stan B. and Geoff D.
NARRE hyb/EnKF assim options – p.1• Separate 1h ensemble data assimilation cycle
– Testing already performed by• CAPS – Ming Xue, Kefeng Zhu – 40km/3h RAP EnKF assimilation• GSD – Ming Hu – 13km/40km/1h RAP EnKF assimilation
• Use of GFS (global) EnKF-produced background error covariance– Only updated every 6h, not available at 1h freq– Current promising testing by
• Dave Parrish and Wan-shu Wu at NCEP in 6h NAM cycle.• ESRL (Ming Hu) with alternative RAP now running since Nov
2012– Clearly outperforms use of fixed background error covariance
(current NCEP NAM and NCEP RAP)
NARRE hybrid/EnKF assimilation options – p2• Separate 1h ensemble data assimilation cycle
• Jan 13 update – RAP testing using GSI-hybrid DA with GFS-DA-ensemble fully successful, will be promoted to primary ESRL RAP (initializing HRRR) in February 2013
Assuming 1h EnKF assimilation for NARRE
• Resolution for data assimilation ensemble– 30-40km may be sufficient (and perhaps higher-res
cannot be afforded)• Model for data assimilation EnKF
– Choose single dynamic core (ARW or NMMB), do not mix cores
• Physics for data assimilation EnKF– Apply stochastic physics to a single set
• Successful inflation/localization in RAP ensemble-RAP testing by CAPS, NSSL, U.OK
73
2017-2018?High Resolution Rapid Refresh
ENSEMBLE (HRRRE)• Each member of NARRE contains 3 km nests
– CONUS, Alaska, Hawaii & Puerto Rico/Hispaniola nests– The two control runs initialized with radar data & other hi-res obs
• This capability puts NWS/NCEP[+OAR/ESRL] in a position to – Provide NextGen Enroute AND Terminal guidance (FWIS-like)– Provide PROBABILITY guidance with full Probability Density
Function specified, hence uncertainty information too– Provide a vehicle to improve assimilation capabilities using hybrid
(EnKF+4DVar) technique with current & future radar & satellite– Address Warn-on-Forecast as resolutions evolve towards ~1 km
• NAM nests are extensions of the 00z, 06z, 12z & 18Z runs.• HRRRE requires an increase in HPCC funding over
and above that required for the NARREFrom Geoff DiMego, Dec 2011, NCEP Model Review
DRAFT Storm Prediction Center Desired Numerical Guidance Attributes
2015 2017 2022•North America Short-Range Mesoscale Ensemble: 12-15 km ~30 member multi-model/multi-physics/multi-IC ensemble with EnKF/hybrid DA, run every 6 hrs with forecasts to 84 hrs
•North America Short-Range Mesoscale Ensemble: 12-15 km ~40 member multi-model/multi-physics/multi-IC ensemble with EnKF/hybrid or newer-state-of-the-art DA, run every 6 hrs with forecasts to Day 4*
•Global Mesoscale Ensemble Forecast System: 10-12 km global domain 40-50 member ensemble with advanced state-of-the-science DA, run every 6 hrs with forecasts to Day 7-10*
•CONUS Storm Scale Ensemble Forecast (SSEF): 4 km CONUS ~10 member multi-model/multi-physics/multi-IC storm scale ensemble with EnKF/hybrid DA, issued every 6-hrs with forecasts to 48-60 hrs*
•CONUS Storm Scale Ensemble Forecast (SSEF): 3 km CONUS ~15 member multi-model/multi-physics/multi-IC storm scale ensemble with EnKF/hybrid or newer state-of-the-art DA, issued every 6-hrs with forecasts to 48-60 hrs*
•CONUS SSEF: 2 km CONUS 30-50 member multi-model/multi-physics/multi-IC storm scale ensemble with advanced state-of-the-science DA, issued every 6-hrs with forecasts to 48-60 hrs*
•Movable domain update SSEF: 2-3 km movable regional domain ~7 member storm scale ensemble with EnKF/hybrid run every 2-4 hrs with forecasts to 15-18 hrs*, focused on “severe weather of the day” areas
•Movable domain update SSEF: 2 km movable regional domain 10-15 member storm scale ensemble with EnKF/hybrid or newer state-of-the-art DA ,run every 1-2 hrs with forecasts to 18-24 hrs*, focused on “severe weather of the day” areas
•Movable domain update SSEF: 1 km movable regional domain 20-30 member storm scale ensemble with advanced state-of-the-science DA run every 1 hr with forecasts to 18-24 hrs*, focused on “severe weather of the day” areas
•Stormscale 3D Analysis: 4 km EnKF/hybrid DA, CONUS storm scale analysis updated every 1-hr.
•Stormscale 3D Analysis: 2 km EnKF/hybrid or newer state-of-the-art DA, CONUS storm scale analysis updated every 15 min
•Stormscale 3D Analysis: 2 km EnKF/hybrid or newer state-of-the-art DA, CONUS storm scale analysis updated every 5 min
•Appropriate RAP (HRRR), Global deterministic / ensemble, and CFS capacities
•Appropriate RAP (HRRR), Global deterministic / ensemble, and CFS capacities
•Stretch Goal - Warn-on-Forecast Prototype Ensemble: Develop <1 km multi-WFO domain 10-20 member nested SSEF, advanced state-of-the science DA, run as needed every 5-10 min with forecasts to 1-2 hrs*
*Model and data assimilation (DA) complexity will be impacted strongly by High Performance Computing capabilities.
Desired Attributes of NOAA Operational Numerical Guidance System
74
DRAFT Storm Prediction Center Desired Numerical Guidance Attributes
2015 2017 2022•North America Short-Range Mesoscale Ensemble: 12-15 km ~30 member multi-model/multi-physics/multi-IC ensemble with EnKF/hybrid DA, run every 6 hrs with forecasts to 84 hrs
•North America Short-Range Mesoscale Ensemble: 12-15 km ~40 member multi-model/multi-physics/multi-IC ensemble with EnKF/hybrid or newer-state-of-the-art DA, run every 6 hrs with forecasts to Day 4*
•Global Mesoscale Ensemble Forecast System: 10-12 km global domain 40-50 member ensemble with advanced state-of-the-science DA, run every 6 hrs with forecasts to Day 7-10*
•CONUS Storm Scale Ensemble Forecast (SSEF): 4 km CONUS ~10 member multi-model/multi-physics/multi-IC storm scale ensemble with EnKF/hybrid DA, issued every 6-hrs with forecasts to 48-60 hrs*
•CONUS Storm Scale Ensemble Forecast (SSEF): 3 km CONUS ~15 member multi-model/multi-physics/multi-IC storm scale ensemble with EnKF/hybrid or newer state-of-the-art DA, issued every 6-hrs with forecasts to 48-60 hrs*
•CONUS SSEF: 2 km CONUS 30-50 member multi-model/multi-physics/multi-IC storm scale ensemble with advanced state-of-the-science DA, issued every 6-hrs with forecasts to 48-60 hrs*
•Movable domain update SSEF: 2-3 km movable regional domain ~7 member storm scale ensemble with EnKF/hybrid run every 2-4 hrs with forecasts to 15-18 hrs*, focused on “severe weather of the day” areas
•Movable domain update SSEF: 2 km movable regional domain 10-15 member storm scale ensemble with EnKF/hybrid or newer state-of-the-art DA ,run every 1-2 hrs with forecasts to 18-24 hrs*, focused on “severe weather of the day” areas
•Movable domain update SSEF: 1 km movable regional domain 20-30 member storm scale ensemble with advanced state-of-the-science DA run every 1 hr with forecasts to 18-24 hrs*, focused on “severe weather of the day” areas
•Stormscale 3D Analysis: 4 km EnKF/hybrid DA, CONUS storm scale analysis updated every 1-hr.
•Stormscale 3D Analysis: 2 km EnKF/hybrid or newer state-of-the-art DA, CONUS storm scale analysis updated every 15 min
•Stormscale 3D Analysis: 2 km EnKF/hybrid or newer state-of-the-art DA, CONUS storm scale analysis updated every 5 min
•Appropriate RAP (HRRR), Global deterministic / ensemble, and CFS capacities
•Appropriate RAP (HRRR), Global deterministic / ensemble, and CFS capacities
•Stretch Goal - Warn-on-Forecast Prototype Ensemble: Develop <1 km multi-WFO domain 10-20 member nested SSEF, advanced state-of-the science DA, run as needed every 5-10 min with forecasts to 1-2 hrs*
*Model and data assimilation (DA) complexity will be impacted strongly by High Performance Computing capabilities.
Desired Attributes of NOAA Operational Numerical Guidance System
75
Possible alternative NWP 2-member suite - 2022
Global ensemble – as shown by EMC
• 10-12km global 40-50 member ensemble• runs to Day 14 twice daily (to 7d at 06z, 18z)• hourly assimilation cycle with 24h runs init hourly
Single regional model/assimilation ensemble - SSEF• SSEF (equivalent to HRRRe) – 2-3km• Initialize hourly (or subhourly), run to 24h• Every 6h, run out to 48-60h• 2-3km EnKF/hybrid/other DA• Multi-species microphysics with at least 2-moment rain• Include aerosols/fire/smoke for all runs with aerosol-
aware microphysics• No separate HWRF or fire-weather runs or AQ/dust run• 0.5-1.0km multi-WFO nest where needed (WoF)• SSEF analysis = RUA, no separate products
76
Global Continental Local Nesting Strategy
• Discussion items– Nesting necessary to save computing; current NAM domain could be expanded to global 10 km– Advantages/disadvantages of same model for global and regional– Applicability (DA) of a multi-model ensemble-based covariance estimate with multiple bias characteristics– Data assimilation at resolution of 3 km or less poses challenges
• Advanced assimilation of radar reflectivity• Consistent integration of radar and satellite info• Proper dynamically balanced analysis increments at multiple high resolution scales (10, 3, 1 km)
77
DCMIP goal:Identify global model numerical deficiencies, especially for inadvertent or intentional diffusion
March 2, 2012• Major Tornado Outbreak – Ohio and Tennessee Valleys
• Model design for HRRR – most accurate numerics possible necessary for identification of storm structure (line vs. supercells vs. cell vs. MCS variations)– ARW, no 6th order diffusion, 5th-order vertical advection
• WoF options – discussion this week with NSSL, ESRL, EMC, U.OK, SPC partners– direct nest inside HRRR/HRRRe within same executable– Separate system using HRRR/HRRRe lateral BCs/background 82