Wei Wang

1
Real-Time WRF Testing at NCAR

• Wei Wang
• Mesoscale and Microscale Meteorology Division
• NCAR
• Contributors Jim Bresch, Chris Davis, Jimy
  Dudhia, Dave Gill, Joe Klemp, Kevin Manning, Bill
  Skamarock, Morris Weisman, and Mike Baldwin (NSSL)

2
Outline

• Why are we doing real-time runs?
• Objectives.
• What have we been doing at MMM for real-time WRF
  testing since WRF first became available?
• Forecast examples, and precipitation verification
  scores, performance of mass and height models.
• What have we learnt from the real-time testing?
• What are we going to do next?

3
Why real-time runs?

• Real-time experiment is an integral part of the
  WRF model development. It is complimentary to
  other tests done during the development, such as
  case studies, idealized tests, etc..
• Real-time runs provide test beds for evaluating
  new developments.

4
Objectives

• Test robustness of the modeling system for
  various domain configurations (grid sizes, number
  of vertical levels), dynamical cores and physics
  options
• Debugging, and examine systematic errors
• Examine the performance of the WRF model, and
  compare with other models, such as Eta and MM5.
• Subjective evaluation
• Statistical evaluation precipitation
  verification
• Precipitation climatology

5
What have we been doing?
6
What have we been doing? (cont)
7
Current Configurations

• CONUS 22 km grid
• Initialized from 40 km Eta data, 6 hourly BC
• Physics includes KF, NCEP 3-class ice
  microphysics, MRF PBL and 5-layer soil model
• Time step 120 sec, 260x164x28
• 2 h 18 m for 48 h fcst on Alpha ES40 (4
  processors)
• Central US 10 km grid
• Initialized from 40 km Eta data, 3 hourly BC
• Physics includes KF, NCEP 3-class microphysics,
  MRF PBL and OSU LSM
• Time step 60 sec, 244x214x35
• 82 min for 30 h fcst on 32 Alpha processors

8
Current Configurations

• CONUS 10 km grid
• Same physics as CONUS 22 km
• Number of grid points 571x361x35, time step 60
  sec.
• Run on Alpha cluster at FSL.

9
Current regional 10 km domain
10
Current 22 km domain
11
Forecast Examples

• A recent heavy precipitation case
• Heavy precipitation case
• Severe convective case of 24 Oct 2001 as
  simulated by a 12 km WRF model
• Fine-scale features.

12
Observed 24 h rainfall verifying at 1200 UTC
June 16, 2002
13
Observed
Eta 12 km
WRF 22km
WRF 10 km
14
WRF10
Threat Score
ETA12
Eta
WRF22
WRF10
Bias Score
15
Storm Reports for Oct 24, 2001
16
24h WRF 12 km forecast valid 00Z 25 Oct 2001
WRF vertically integrated cloud water
Satellite cloud image
17
24h WRF 12 km forecast valid 00Z 25 Oct 2001
WRF 700 mb vertical velocity
Radar summary
18
Precipitation verification scores

• Two verification sites http//www.nssl.noaa.gov/e
  takf/verf
• -- verifies against gridded precipitation
  analysis
• since May 2001 for 22 km WRF
• http//www-ad.fsl.noaa.gov/fvb/rtvs/index.html
  
• -- verifies station precipitation, but only
  for
• short-range forecasts (up to 12 h) since
  Fall 2001 for
• 22 km and 10 km (regional) WRF

19
Threat and Bias Score 36 h Forecast, 24 h Accum
May 2002
June 2002
22 km BMJ cumulus
30 km BMJ cumulus
20
Threat and Bias Score 48 h Forecast, 3 h Accum
for May 2002
0.01 in
0.25 in
22 km BMJ cumulus
21
Comparison between mass and height models

• Real-time experiments were set up to run mass and
  height models in parallel for three months.
• Results were examined for individual cases as
  well as statistically by precipitation
  verification scores.

22
TS Barry 48 h SLP Verifying at 0000 UTC Aug. 6,
2001
WRF mass
WRF height
23
Threat and Bias Score 36h Forecast, 24h
Accum Comparison between mass and height models
October 2001
November 2001
24
What we learnt from real-time WRF?

• WRF modeling system is robust
• WRF model has shown good performance for many
  cases, particularly for those of severely
  convective and heavy-precipitation producing
  cases.
• WRF 24 h precipitation scores show comparable
  performance to those produced by 22 and 12 km
  Eta, and Eta-KF. At high precipitation
  thresholds, WRF shows some skill higher threat
  score, but bias remains closer to 1.25.
• WRF model is capable of producing finer
  structures when compared with models of similar
  resolution.

25
What we learnt from real-time WRF?

• Individual cases as well as statistical
  performance (in terms of precipitation
  verification) between mass and height models are
  nearly identical.
• Have identified a number of problems and resolved
  them significant improvement

26
Whats next?

• Continue to do real-time runs, and use real-time
  runs as test beds for new development this has
  been proved to be very important process for the
  model development.
• Begin to do careful investigation on certain
  cases to understand WRF performance, and physics
  in the model critical for further improvement
  to the overall performance of the model as a NWP
  tool.