Get documents about "presentation source Comet
Shared by: liaoqinmei
-
Stats
- views:
- 0
- posted:
- 11/7/2012
- language:
- English
- pages:
- 92
Document Sample
Meso- and Storm-Scale NWP:
Scientific and Operational
Challenges for the Next Decade
COMET Faculty Course on NWP
9 June 1999
Boulder, Colorado
ns f
Kelvin K. Droegemeier
School of Meteorology and
Center for Analysis and Prediction of Storms
University of Oklahoma
1
What Are Operational Models Predicting?
Global and synoptic flow patterns
Precipitation via crude parameterizations that are
unable to resolve individual clouds
Topographic forcing
Coastal and lake
influences
Crude linkages
between the land
surface and
atmosphere
ns f
2
What Do Forecasters Use?
Single forecasts
Output frequency of 3 to 12 hours
Accumulated precipitation and other traditional fields
Graphical overlays of model, radar, satellite
GETTING THIS
FROM THIS
ns f
3
What Do We Need to Predict?
Individual thunderstorms and squall lines
Lake effect snow storms
Down-slope wind storms
Convective initiation
Seabreeze convection
Stratocumulus decks off the coast
Cold air damming
Post-frontal rainbands
ns f
4
Why?
Local high-impact weather causes economic losses in
the US that average $300 M per week
Over 10% of the $7 trillion US economy is impacted
each year
Commercial aviation losses are $1-2 B per year (one
diverted flight costs $150K)
Agriculture losses exceed $10 B/year
Other industries (power utilities, surface transport)
About 50% of the loss is preventable!
ns f
5
Pielke Jr. (1997)
What is Needed?
Models that
– run at high spatial resolution (1-3 km)
– utilize high-resolution observations (e.g., from the
WSR-88D network)
– handle terrain well
– represent important physical
processes, especially microphysics
and land-surface interactions
Physical/theoretical understanding
Tools for integrating model
output, observations
ns f
6
Role of the University Community
Educating students about NWP -- a whole new
ballgame!
– Physical processes
– Data sets & observing platforms
– Numerical models & methods
– Data assimilation & predictability
Research in all facets of NWP
Running models in in real time
– More than 25 universities do this today!
– Major change from 20 years ago!
– Academia is driving operational NWP
Collecting data
– GPS, WSR-88D, other
ns f
7
Trends in Large-Scale
Forecast Skill
ns f
8
Predictability: Hitting the Wall
For global models, the predictability increases for all
resolvable scales as the spatial resolution increases
(quasi 2-D dynamics)
– The improvement is bounded
– Going beyond a few 10s of km gives little payoff
The next quantum leap in NWP will come when we
start resolving explicitly the most energetic weather
features, e.g., individual convective storms (3-D)
60 km 30 km
30 km 10 km
10 km 2 km
ns f
9
Center for Analysis and
Prediction of Storms (CAPS)
One of first 11 NSF Science and Technology Centers
established in 1989
STCs were designed to attack problems of fundamental
research that eventually would yield important benefits
to society
Mission of CAPS: To demonstrate the practicability of
numerically predicting local, high-impact storm-scale
spring and winter weather, and to develop, test, and
help implement a complete analysis and forecast system
appropriate operational, commercial, and research
applications
ns f
10
The Key Scientific Questions
Can value be added to present-day NWP and radar-
based nowcasting by storm-resolving models?
Which storm-scale events are most predictable, and will
fine-scale details enhance or reduce predictability?
What physics is required, and do we understand it well
enough for practical application?
What observations are most critical, and can data from
the national NEXRAD Doppler radar network be used to
initialize NWP models? Can this be done in real time?
What networking and computational infrastructures are
needed to support high-resolution NWP?
How can useful decision making information be
generated from forecast model output?
ns f
11
Prediction Targets
Somewhat problematic
For 1-3 km resolution grids, location to within
– 200 km 6 hours in advance
– 100 km 4 hours in advance
– 50 km 2 hours in advance
– 10 km 1 hour in advance
Initiation
Movement
Intensity
Duration
ns f
12
Meso-scale NWP
The prediction of the general characteristics
associated with mesoscale weather phenomena
WSR-88D CREF (02 UTC 30 Nov 1999)
6-hour ARPS Forecast at 9 km resolution
ns f
13
Storm-scale NWP
The prediction of explicit updraft/downdrafts and
related features (e.g., gust fronts, meso-cyclones)
NEXRAD Radar Observations
ARPS 90 min Forecast (3 km)
ns f
14
Economic Impact
Negative Consequences of a Bad Forecast
2000-2010
Breadth of Application
1990’s
1980’s
1970’s
Model Spatial Resolution
ns f
15
Present NWS Operations
CONUS RUC and Eta Models (32 & 40 km)
NCEP
Central
Operations
ns f
16
NWS Forecast Offices
ns f
17
Small-Scale Weather is LOCAL!
Rain and
Snow Fog Rain and
Snow
Snow and
Intense Freezing
Turbulence Rain
Severe
Thunderstorms
ns f
18
The Future of Operational NWP
20 km CONUS Ensembles
10 km
3 km
1 km
ns f
19
The Future of Operational NWP??
ns f
20
Principal Differences Between
Large- and Small-Scale NWP
Large-scale: Rawinsondes observe “everything” that
is needed to initialize a model (T, RH, u, v)
Small-scale: Doppler radar observes only the radial
wind and reflectivity in precipitation regions; clear-
air PBL data available in some situations
Large-scale: Well-known balances can be applied to
reconcile wind and mass fields (e.g., geostrophy,
balance equation)
Small-scale: Only simple balances available (mass
continuity); otherwise, it’s the full equations!!
ns f
21
Large-scale: Forecasts are of sufficient duration to
be produced and disseminated in reasonable time
frames
Small-scale: Forecasts are of very short duration and
thus are highly perishable
Large-scale: Observing network is mature and errors
and natural variability are understood
Small-scale: Key observing system (WSR-88D) is
new; only a few links exist for providing base data in
real time
ns f
22
Large-scale: Dynamics and predictability limits are
fairly well understood; model physics and numerics are
reasonably mature
Small-scale: Dynamics fairly well understood, but
predictability limits have not been established; model
physics still evolving; physical processes complicated
(addition of detail a double-edged sword)
Large-scale: Conventional data assimilation techniques
work well; large-scale features evolve slowly
Small-scale: Conventional data assimilation techniques
not applicable; events are spatially intermittent and
evolve rapidly; how to remove an incorrect
thunderstorm and insert the correct one???
ns f
23
Large-scale: Computing power reasonably sufficient
Small-scale: Need 100 to 1000 times more computing
power than is now available commercially
Large-scale: No lateral boundary conditions to worry
about for global and hemispheric models
Small-scale: Lateral boundaries in limited-area models
exert a tremendous influence on the solution;
compromise between high spatial resolution and
domain size
ns f
24
Recipe for a Storm-Scale NWP
System
Advanced numerical model with appropriate physics
parameterizations
High-resolution observations (WSR-88D, profilers,
satellites, MDCRS) and appropriate ways for using
them
Powerful computers and networks
A way to retrieve quantities that cannot be observed
directly
Strategies for converting output to useful decision
making information
ns f
25
The CAPS Advanced Regional Prediction
System (ARPS)
Lat eral bo undar y cond itions
from large -scale model s ARPS D ata Assi milation S yste m (ARPS DAS )
Gri dded first gu ess
Mo bile M esone t Da ta Acquisition Pa ram eter Retr ieva l and 4DDA
Raw inson des & Analysis
S i ngle-Doppler Velocity
Incoming
AC ARS
ARPS D ata Anal ysis Re trieval (S DVR)
data
CL ASS
SA O S ystem (AD AS )
Sat ellite – In gest 4-D -
Variati onal Vel
P ro filers – Q uality contro l Variati onal oci ty Adjustment
AS OS/AW OS
– O bjecti ve ana lysis Data & Thermo-
Ok lahoma Meso net
– A rchiva l As simi lation dynami c Re trieval
WS R-88D Wide band
Pr oduc t Ge nera tion and
Da ta Support Syste m
Foreca st Gener ation
ARPS PLT and ARPS VIEW
ARPS N umerical Model – P lots an d imag es
– M ulti-s cale no n-hyd rostat ic pred iction – A nimati ons
mode l with compr ehens ive ph ysics – D iagno stics a nd stat istics
– F orecas t evalu ation
ns f
26
NEXRAD Doppler Radar Data
ns f
27
Single-Doppler Velocity Retrieval (SDVR)
real
We observe ... wind
– one (radial) wind component
– reflectivity
We need ... observed
– 3 wind components component
– temperature
– humidity
– pressure
– water substance (6-10 fields)
SDVR solves the inverse problem
– control theory (adjoint), simpler methods
– computationally very intensive
ns f
28
Sample SDVR Result
Dual-Doppler SDVR-Retrieved
Weygandt (1998)
ns f
29
Sample SDVR Result
Dual-Doppler SDVR-Retrieved
Weygandt (1998)
ns f
30
Sample SDVR Result
Dual-Doppler SDVR-Retrieved
Weygandt (1998)
ns f
31
5 April 1999 - Impact of Radar Data
Initial 700 mb Vertical Initial 700 mb Vertical
Velocity Using NIDS Velocity Using Level II
Data and SDVR
12 Z Reflectivity
ns f
32
5 April 1999 - Impact of Radar Data
3 hr ARPS CREF 3 hr ARPS CREF Forecast
Forecast (9 km) Using (9 km) Using Level II
NIDS Data Data and SDVR
Valid 15Z Valid 15Z
15 Z Reflectivity
ns f
33
The Lahoma, OK Hailstorm
Conway et al. (1996)
ns f
34
ns f
35
Availability of Base Data
CAPS has been using Level II (base) NEXRAD data
in case study predictions down to 1 km resolution
and Level III data (NIDS) in its daily operational
forecasts
Although NIDS data are available in real time from
all radars, they are insufficient in many cases for
storm-scale NWP
– Precision is degraded via value quantization
– Only the lowest 4 tilts are transmitted
No national strategy yet exists for the real time
collection and distribution of Level II data
An example of universities leading the way!!
ns f
36
Real Time Test Bed for Acquiring WSR-88D
Base Data (Project CRAFT)
Approval Pending
DDC ICT
INX
AMA TLX KFSM
LBB
Radars Online
FWS
ns f
37
CRAFT Phase I
ns f
38
Regional Collection Concept
Must await
open-RPG
Great
opportunity
for
universities!
ns f
39
The CAPS Vision
Regionalization and Customization of NWP
CONUS Forecasts (20 km resolution)
Regional (5 k m resolution)
Local
(0.5-1.0 k m
resolution)
Sub-regional
(2 k m resolution)
ns f
40
Real Time Testing
Daily operation of experimental forecast models is
critical for
– involving operational forecasters in R&D
– evaluating model performance under all conditions
– testing new forecast strategies (e.g., rapid model updates,
forecasts on demand, re-locatable domains)
– developing measures of skill and reliability based on a long-
term data base of model output
– learning how to integrate new forecast information into
operational decision making
Over 25 groups around the US are running models
in real time in collaboration with NWS Offices or
NCEP Centers; few are assimilating observations
ns f
41
CAPS’ Real Time Testing
Daily operational forecasts with full-physics at spatial
resolutions down to 3 km
Assimilation of high-resolution observations
consistent with the model high spatial resolution
– WSR-88D Level II (base) data
– WSR-88D Level III (NIDS) data
– GOES satellite data for quantitative vapor/cloud/precip
– MDCRS commercial aircraft T and V
– Surface mesonets
More than 2000 products produced each hour and
posted on the web (http://hubcaps.ou.edu)
Execution on the 256-node Origin 2000 at NCSA
ns f
42
ARPSView Decision Support System
ns f
43
1999 Special Operational Period
5-Member, 30 km Ensemble
9 km
3 km
WSR-88D Base Data Being Ingested
WSR-88D Base Data Pending
ns f
44
ARPS 32 km Forecast - AR Tornadoes
ARPS 12-hour, 32
km Resolution
Forecast CREF Valid
at 00Z on 1/22/99
Radar
Radar
(Tornadoes
in Arkansas)
ns f
45
Proprietary
ARPS 9km Forecast - AR Tornadoes
Radar
ARPS 6-hour, 9 km
Radar Forecast CREF Valid
(Tornadoes at 00Z on 1/22/99
in Arkansas)
ns f
46
Proprietary
ARPS 3km Forecast - AR Tornadoes
Weather Channel Radar ARPS 6-hour, 3 km
at 2343 Z Forecast CREF Valid at 00Z
ns f
47
6 January 1999
GOES Visible Image ARPS 12 h Forecast Visibility
1745Z, 6 Jan 99 (27 km) Valid 18Z, 6 Jan 99
ns f
48
9-10 May 1999
NCEP Eta 12-hour Forecast Valid 00 Z
Composite Radar Valid 2347 Z on Monday, 10 May 1999
Sunday, 9 May 1999
ns f
49
9-10 May 1999
ARPS 4-hour, 3 km CREF Forecast
Composite Radar Valid 0344 Z on Valid 04 Z Monday, 10 May 1999
Monday, 10 May 1999
ns f
50
1 June 1999
KFWS CREF Valid 00 Z on
Tuesday, 1 June 1999
ARPS CREF Initial Condition
Valid 00 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
51
1 June 1999
KFWS CREF Valid 01 Z on
Tuesday, 1 June 1999
ARPS CREF 1-hour Forecast
Valid 01 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
52
1 June 1999
KFWS CREF Valid 02 Z on
Tuesday, 1 June 1999
ARPS CREF 2-hour Forecast
Valid 02 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
53
1 June 1999
KFWS CREF Valid 03 Z on
Tuesday, 1 June 1999
ARPS CREF 3-hour Forecast
Valid 03 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
54
1 June 1999
KFWS CREF Valid 04 Z on
Tuesday, 1 June 1999
ARPS CREF 4-hour Forecast
Valid 04 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
55
1 June 1999
KFWS CREF Valid 05 Z on
Tuesday, 1 June 1999
ARPS CREF 5-hour Forecast
Valid 05 Z on Tuesday, 1 June 1999
(3 km resolution with Level II data from
KTLX and KFWS + NIDS)
ns f
56
3 June 1999
KAMA CREF Valid 00 Z on 3 June
1999 ARPS 3-hour 3 km Forecast
Valid 00 Z on 3 June 1999
(without NEXRAD base data)
ns f
57
3 June 1999
KAMA CREF Valid 03 Z on 3 June
1999 ARPS 6-hour 3 km Forecast
Valid 03 Z on 3 June 1999
(without NEXRAD base data)
ns f
58
3 June 1999
KAMA CREF Valid 04 Z on 3 June
1999 ARPS 7-hour 3 km Forecast
Valid 04 Z on 3 June 1999
(without NEXRAD base data)
ns f
59
3 June 1999
KAMA CREF Valid 05 Z on 3 June
1999 ARPS 8-hour 3 km Forecast
Valid 05 Z on 3 June 1999
(without NEXRAD base data)
ns f
60
3 June 1999
KAMA CREF Valid 06 Z on 3 June
1999 ARPS 9-hour 3 km Forecast
Valid 06 Z on 3 June 1999
(without NEXRAD base data)
ns f
61
Numerical Forecasts of the May 3 Tornadic Storms
5:30 pm
NEXRAD Radar Observations
ARPS Prediction Model
(1/2 hour forecast)
ns f
62
Numerical Forecasts of the May 3 Tornadic Storms
6:00 pm
NEXRAD Radar Observations
ARPS Prediction Model
(1 hour forecast)
ns f
63
Numerical Forecasts of the May 3 Tornadic Storms
6:30 pm
NEXRAD Radar Observations
ARPS Prediction Model
(1 1/2 hour forecast)
ns f
64
Numerical Forecasts of the May 3 Tornadic Storms
7:00 pm
NEXRAD Radar Observations
ARPS Prediction Model
(2 hour forecast)
ns f
65
Numerical Forecasts of the May 3 Tornadic Storms
7:00 pm
Moore, OK
Tornadic
Storm
Moore, OK
Tornadic
Storm
NEXRAD Radar Observations
ARPS Prediction Model
(2 hour forecast) 12-hour Eta
Forecast
66
Numerical Forecasts of the May 3 Tornadic Storms
ARPS With and Without NEXRAD Base Data
7:00 pm
WITHOUT
WITH
(3 ARPS hour forecast)
ARPS Prediction Model
(2 hour forecast)
NEXRAD Radar Observations 67
How Good are the Forecasts?
Forecast Verification
D/FW Airport
40 km for 3 Hour Forecast
ns f
68
How Good Are the Forecasts?
ns f
69
The Issues
Traditional skill measures (e.g., threat score or
“overlap” agreement) not appropriate for
intermittent storm-scale phenomena
Specific character of storms (intensity, motion,
initiation, decay) important for operational
forecasters
QPF is critical!
Problem: We forecast more things than we can
observe/verify (how to verify 500 mb height fields
that contain thunderstorms?)
Point verification is rather meaningless
ns f
70
Approaches
Qualitative (by hand) verification
– location, speed, timing, duration, intensity, orientation,
mode
– “With 4 hours of lead time, the location of storms was
within 30 km of observed 80% of the time”
– “The model predicted storms 10% of the time when none
were observed”
Phase-shifting verification
– maximize spatial correlation
– generates a shift vector
Will eventually have to consider cost-benefit and
reliability
ns f
71
Quantitative Forecast Evaluation for
May 3 Forecasts (3 km)
Hourly analysis of echoes by county in Oklahoma (N=77)
Statistics
– Hit Rate: The fraction of correct forecasts (best=1, worst=0)
– Critical Success Index (Threat Score): The hit rate after removing
the correct forecasts of no echoes (best=1, worst=0)
– False Alarm Ratio: The fraction of forecasts that are incorrect
(best=0, worst=1)
– Probability of Detection: The fraction of forecasts that are correct
(best=1, worst=1)
– Bias: A measure of the tendency to overforecast or
underforecast. (bias=1 is optimal)
ns f
72
May 3 ARPS Forecast with WSR-88D Level II Data
Initialized 22 UTC Using 3 km Spatial Resolution
Existence of Echoes for all 77 Oklahoma Counties
2
HR
CSI
FAR
POD
Averages
1.5 Bias
HR = 0.911
CSI = 0.621
1
FAR = 0.233
POD = 0.798
0.5 BIAS = 1.110
0
2200 2230 2300 2330 0000 0030
Time (UTC)
ns f
73
May 3 ARPS Forecast with WSR-88D Level II Data
Initialized 22 UTC Using 3 km Spatial Resolution
CREF Echoes of 50 dBz +/- 10 dBz by County
2
HR
CSI
1.5
FAR
POD
Averages
Bias
HR = 0.940
1
CSI = 0.511
FAR = 0.258
0.5
POD = 0.633
BIAS = 0.939
0
-0.5
2200 2300 0000 0100 0200
Time (UTC)
ns f
74
May 3 ARPS Forecast with WSR-88D Level II Data
Initialized 22 UTC Using 3 km Spatial Resolution
CREF Echoes of 50 dBz +/- 5 dBz by County
1.2
1
Averages
0.8
HR = 0.919
0.6 CSI = 0.398
FAR = 0.324
0.4
POD = 0.489
0.2
HR
BIAS = 0.771
CSI
0 FAR
POD
Bias
-0.2
2200 2300 0000 0100 0200
Time (UTC)
ns f
75
t = 1 hour t = 2 hours
Truth
The
Importance
of
Phase Errors
Forecast
Zhang (1999)
ns f
76
Standard
Threat Score
ns f
77
Zhang (1999)
Phase-Shifted
Threat Score
Average Phase
Shift Error (km)
ns f
78
Zhang (1999)
Lessons Learned
Getting the larger-scale features correct is the easy part --
getting the reflectivity correct is tough!
– But does it matter?
– These models are not reflectivity generators!
Solution sensitivity (surface characteristics, soil moisture)
Initial conditions are the critical aspect -- much work needed in
data assimilation and parameter retrieval
Model physics seem adequate (QPF needs work, though)
How good is good enough?
Fine resolution gives more detail but also greater uncertainty
and sensitivity (e.g., caps, outflow boundaries)
Forecasters easily overwhelmed by zillions of new products
More experience needed with ensemble forecasting
ns f
79
Traditional Truth
Forecasting
Methodology
Initial State
Uncertainty
Single Forecast
ns f
80
Ensemble
Forecasting
Mean
Initial State
Uncertainty
Truth
t critical
Deterministic
Forecast
Probabi li sti c
Forecast
ns f
81
Ensemble Forecasting
Advantages
– Ensemble mean is generally superior
– Ensembles provide
a measure of expected skill or confidence
a quantitative basis for probabilistic forecasting
a rational framework for forecast verification
information for targeted observations
Limitations/Challenges
– Not clear how to optimally specify the initial
conditions (singular vectors, breeding, perturbed
observations)
– Requires more computer resources
ns f
82
Storm and Mesoscale Ensemble
Experiment (SAMEX)
Collaborative effort among CAPS, NCAR, AFWA, NCEP
and NSSL
Performed during May, 1998
Goal: Examine the value of coarse-resolution, multi-
model ensemble forecasts versus single high-resolution
deterministic forecasts
Expose operational forecasters in real time to both
types of output
ns f
83
SAMEX Domains
km)
CAPS (32 km), NCEP (32 km)
NCAR (30 km)
AFWA (27 km)
NSSL (32 km) Ensemble P roduct Domain
CAPS (9 km), NCEP (10 km)
NCAR (10 km)
AFWA (9 km)
AFWA
(3 km) CAPS (3 km)
ns f
84
ns f
85
ns f
86
Oops!!
3-hour Observed Precipitation
25-Member Ensemble
POP > 0.1 inch/hour
ns f
87
Explicit 9 km Prediction
3-hour Accumulated Precipitation 9 km, 15-hour ARPS
Forecast Reflectivity
ns f
88
500 mb Errors
ns f
89
m
20-30 k Resolution Ensemble Domain
Pacific
Northwest
Great Lakes
Central and
Inter-Mountain Southern
California
Coast Great Plains
Southeast
US
Florida
Coast
ns f
90
Summary
Storm-scale NWP is a significant scientific and
technological challenge
Predictability appears plausible at storm scales
More work needed in
– data assimilation, especially from satellite, GPS, WSR-88D
– physics parameterizations (especially cloud microphysics,
radiation, and land-atmosphere exchanges)
– fundamental predictability and sensitivity
Transition to operations will be a major challenge
– centralized versus distributed?
– verification techniques
– creation of useful products
– forecaster interpretation and utilization
NWS FO involvement in R&D will be critical
ns f
91
Some Key Scientific Issues
Predictability of storm-scale flows and application of ensemble
strategies and forecast verification techniques at 1-3 km resolution
Data impact/sensitivity, especially land-atmosphere interactions
Advanced data assimilation techniques (3DVAR, 4DVAR): most
everything boils down to the initial conditions!
Feedback of cloud-scale NWP to global and regional climate
Use of cloud-scale forecasts in hydrologic models
Application of new remote sensing technologies (e.g., GPS, phased-
array radars, polarization-diversity radars, MDCRS)
Linkages between high-impact local weather and local ecosystems,
biodiversity, and health
Intelligent distributed computing and networking: learning how to
create and deliver the information
Economic and societal impacts and mitigation: learning how to use
the information
ns f
92
