The Risks and Rewards of High-Resolution and Ensemble Modeling Systems

1
The Risks and Rewards of High-Resolution and
Ensemble Modeling Systems

• David Schultz
• NOAA/National Severe Storms Laboratory
• Paul Roebber
• University of Wisconsin at Milwaukee
• Brian Colle
• State University of New York at Stony Brook
• David Stensrud
• NOAA/National Severe Storms Laboratory
• http//www.nssl.noaa.gov/schultz

2
Objectives of this Talk

• Discuss issues for operational weather
  forecasting in going to higher-resolution NWP.
• Briefly compare advantages and disadvantages of
  high-resolution simulations versus
  lower-resolution ensembles.
• Example 3 May 1999 Oklahoma tornado outbreak.
• Discuss unresolved scientific issues that will
  lead to improving predictability for operational
  forecasters.

3
High-Resolution NWP

• High resolution (lt 6 km) is now possible in real
  time due to increasing computer power and
  real-time distribution of data from National and
  International Modeling Centres.
• Many groups have demonstrated high-resolution
  real-time NWP (Mass and Kuo 1998).
• Small-scale weather features are able to be
  reproduced by high-resolution models (e.g., sea
  breezes, orographic precipitation, frontal
  circulations, convection).

4
But, . . .

• The use of models to study physical processes and
  to make weather forecasts are two distinctly
  different applications of the same tool.
• No guarantee that a high-resolution model will be
  more useful to forecasters than a model with
  larger grid spacing.
• Model errors may increase with increasing
  resolution, as high-resolution models have more
  degrees of freedom.
• High-resolution models may produce wonderfully
  detailed, but inaccurate, forecasts.

5
Ensemble Modeling Systems

• Ensembles of lower-resolution models can have
  greater skill than a single higher-resolution
  forecast (e.g., Wandishin et al. 2001 Grimit and
  Mass 2001).
• Ensemble forecasts directly express uncertainty
  through their inherently probabilistic nature.
• But, what is the minimum resolution needed for
  accurate simulations?
• How to best construct an ensemble?

6
The Forecast Process

• Hypothesis Formation
• Forecaster develops a conceptual understanding of
  the forecast scenario (problem of the day)
• Hypothesis Testing
• Forecaster seeks evidence that will confirm or
  refute hypothesis
• observations, NWP output, conceptual models
• Continuous process
• Prediction
• Forecaster conceptual model of forecast
  scenario(s)

(e.g., Doswell 1986 Doswell and Maddox 1986
Hoffman 1991 Pliske et al. 2003)
7
Intuitive Forecasters

• Defined by Pliske et al. (2003) as those who
  construct conceptual understanding of their
  forecasts on the basis of dynamic, visual images
  (as opposed to rules of thumb).
• Such forecasters would benefit from both
  high-resolution forecasts and ensembles.
• Show detailed structures/evolutions not possible
  in lower-resolution models
• Developing alternate scenarios from ensembles
• Construct probabilistic forecasts

8
3 May 1999 Oklahoma Outbreak

• 66 tornadoes, produced by 10 long-lived and
  violent supercell thunderstorms
• 45 fatalities, 645 injuries in Oklahoma
• 2300 homes destroyed 7400 damaged
• Over 1 billion in damage, the nations most
  expensive tornado outbreak

(Jarboe)
(Daily Oklahoman)
(Schultz)
9
Moore
Observed radar imagery (courtesy of Travis
Smith, NSSL) 2-km MM5 simulation initialized
25 hours earlier (no data assimilation) pink
1.5-km w (gt 0.5 m/s) blue 9-km
cloud-ice mixing ratio (gt0.1 g/kg)
Moore
0131 UTC
0100 UTC
0221 UTC
0200 UTC
10
Stage IV Radar/Gauge Precip. Analysis (Baldwin
and Mitchell 1997)
11
Modeled Storms as Supercells

• Identify updrafts(gt 5 m/s) correlated with
  vertically coherent relative vorticity for
  at least 60 minutes
• 22 supercells, 11 of which are on OKTX
  border

12
Observed vs Modeled Supercells
13
Ensembles (Stensrud and Weiss)

• 36-km MM5 simulations initialized 24 h ahead
• Six members with varying model physics packages
  3 convective schemes (KainFritsch,
  BettsMillerJanjic, Grell) and 2 PBL schemes
  (Blackadar, BurkeThompson)

14
Ensemble mean convective precipitation 2300
UTC 3 May to 0000 UTC 4 May (every 0.1 mm)
15
ensemble mean
ensemble spread
2000 J/kg
750 J/kg
ensemble maximum
ensemble minimum
2000 J/kg
1000 J/kg
Convective Available Potential Energy (J/kg)
16
ensemble mean
ensemble spread
75
200
ensemble maximum
ensemble minimum
200
200
Storm-Relative Helicity (m2 s2)
17
ensemble mean
ensemble spread
40
20
ensemble minimum
ensemble maximum
40
40
Bulk Richardson Number Shear (m2 s2)
18
Comparison

• Both the high-resolution forecast and the
  ensemble forecasts did not put the bulk of the
  precipitation in the right place in central
  Oklahoma.
• Both models indicated the potential for supercell
  thunderstorms with tornadoes in the
  OklahomaTexas region.
• Both models were sensitive to the choice of
  parameterization schemes (e.g., PBL).

19
Remaining Scientific Issues

• When should forecasters believe the model
  forecast as a literal forecast?
• What is the role of model formulation in
  predictability?
• What is the value of mesoscale data assimilation
  in the initial conditions?
• What constitutes an appropriate measure of
  mesoscale predictability?
• What is the appropriate role of postprocessing
  model data (e.g., neural networks,
  bias-correction techniques)?
• Other examples and further discussion will be
  found in a manuscript, currently in preparation.