Short%20Range%20Ensemble%20Forecasts - PowerPoint PPT Presentation

About This Presentation
Title:

Short%20Range%20Ensemble%20Forecasts

Description:

Bill Bua and Stephen Jascourt (NCEP/COMET/UCAR) Zoltan Toth (NCEP/EMC) Acknowledgements ... Hurricane Michelle. 980, SD~14. SD - hPa. Uncertainty in. intensity ... – PowerPoint PPT presentation

Number of Views:87
Avg rating:3.0/5.0
Slides: 113
Provided by: bill57
Category:

less

Transcript and Presenter's Notes

Title: Short%20Range%20Ensemble%20Forecasts


1
Short Range Ensemble Forecasts
  • A NEW NWP TOOL FOR THE 0-3 DAY TIME RANGE

JUN DU NCEP/EMC WASHINGTON, DC USA
Where America's Climate and Weather Services
Begin
2
Acknowledgements
  • Steve Tracton (NCEP/EMC)
  • Bill Bua and Stephen Jascourt (NCEP/COMET/UCAR)
  • Zoltan Toth (NCEP/EMC)

3
What we will cover?
  • Definition of an ensemble forecast system
  • Rational/justification for ensemble forecasting
  • Atmosphere as a chaotic system
  • Initial and boundary condition uncertainty
  • Model errors/uncertainty
  • NCEP Short-Range Ensemble Forecast (SREF) System
    pronounced as serf
  • Interpretation of SREF products
  • Product examples
  • Case studies

4
What is an ensemble?
  • A set of multiple predictions valid at the
    same time generated from reasonably different
    initial conditions and/or with various credible
    versions of models, the objective being to
    improve skill through ensemble averaging, which
    eliminates non-predictable components, and to
    provide reliable information on forecast
    uncertainties (e.g., probabilities) from the
    spread (diversity) amongst ensemble members

5
Example An Ensemble of Three NWP Models
If youve used more than one model in the
forecast process, youve done ensemble
forecasting!
  • AVN, UKMET, ECMWF
  • All initialized 12Z 6 May 2002
  • 72-hr forecasts valid 12Z 9 May 2002

6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
ENSEMBLE FORECASTING
SPAGHETTI
CLUSTERS
  • SLIGHTLY DIFFERENT INITIAL CONDITIONS OR MODEL
    FORMULATIONS PRODUCE A NUMBER OF POSSIBLE
    FORECASTS AND THE SPREAD OF THESE FORECASTS
    QUANTIFIES UNCERTAINTY

10
What is the rationale for ensemble prediction?
  • Atmosphere is essentially a chaotic system
  • Result Small perturbations (errors) in the
    atmosphere can potentially grow into large
    differences in atmospheric evolution
  • Uncertainties (perturbations) in analyses
    inevitable now and forever gt even a perfect NWP
    model can yield large errors (uncertainties) in
    forecasts
  • NWP models not perfect only estimate dynamical
    and physical behavior gt added source of error
    (uncertainties) in forecasts

11
How chaotic can the atmosphere be?
8 day forecast
12
Chaos can reign even in the short-range!
3.5 day forecast
13
Schematic of Ensemble Prediction The initial
probability PDF(D) rep-resents the initial
uncertainties. From the best estimate of the
initial state a single deterministic forecast
(blue solid curve) is performed. This single
deterministic forecast fails to predict correctly
the future state (red dotted curve). An ensemble
of perturbed forecasts (thin blue solid curves)
start- ing from perturbed initial conditions
designed to sample the initial un- certainties
can be used to estimate the probability PDF(Dn)
at future time, Dn.
14
What do we want an ensemble prediction system
(EPS) to do?
  • Encompass the case dependent range of possible
    forecast scenarios by region, circulation system,
    sensible weather elements, etc.
  • Provide the most skillful forecast probability
    distribution (PDF) within the range of
    possibilities
  • Facilitate the communication of forecast
    uncertainty - probabilistic forecast products -
    to the end-users (public, emergency managers,
    government agencies, etc.)

15
How is an EPS made?
  • In principle, can use one or a combination of
  • Different initial atmospheric states
  • Different models or variations on the same model
    (perturbed physics, dynamics, numerics)
  • Different lateral or lower boundary conditions

16
Initial condition ensembles
  • Ideally
  • Sufficient ensemble members to adequately
    prescribe the PDF of the initial condition
    uncertainty
  • A sufficiently skillful model to reliably predict
    the PDF of all possible outcomes

17
Initial condition ensembles
  • Realistically
  • Can only have a relatively few ensemble members,
    because of computational and operational time
    constraints
  • We have imperfect NWP models, so even if we get
    the initial condition uncertainty right, we may
    not get the right forecast PDF (spread,
    probability of events) from them

18
So, whats an EPS developer to do?
  • Find the initial condition uncertainties that
    matter (not all errors in analyses are important)
  • Make use of the initial conditions that project
    onto the most rapidly growing atmospheric modes
  • This gives us the largest spread given the
    limited number of individual ensemble members
    possible.

19
How do we find the initial condition errors
that will grow?
  • Singular vectors (ECMWF)
  • Seeks out non-linear growing atmospheric modes
  • Breeding method for initial condition
    perturbations (NCEP, Toth and Kalnay, 1993)
  • Works out mathematically and practically to be
    roughly equivalent to singular vector method, but
    at a much lower cost

20
How do we breed perturbations?
  • Begins with analysis/forecast cycle which differs
    only in initially prescribed random distribution
    ("seed") of analysis errors.
  • Initially random perturbations added and
    subtracted from the control (not operational, but
    at less resolution) analysis (applies to all
    model parameters!)
  • Each breeding cycle generates a pair of perturbed
    analyses (10 in all).
  • Goal Lets see what grows!

21
How do we breed perturbations?
  • Continue process of breeding cycles over
    consecutive periods.
  • Procedure
  • Find differences between the ensemble members and
    the control forecast
  • Scale them back so they are about the size of
    analysis errors
  • Add and subtract from control analysis and then
    run ensemble members again

22
How do we breed perturbations?
  • After several cycles, the growing modes dominate
  • Scaled ensemble perturbations at this point (and
    in the future) estimate the analysis errors that
    result in the most rapidly growing modes

23
(No Transcript)
24
EMC
25
NCEP Global EPS Configuration
  • Current ensemble configuration (all
    perturbations developed from breeding method)
  • 00Z
  • T170 high resolution control out to 7 days, then
    out to 16 days at T62 resolution
  • T126 control that is started from truncated T170
    analysis truncated to T62 after 84 hr
  • 10 perturbed forecasts at T126 out to 84 hrs,
    then T62

26
NCEP Global EPS Configuration
  • 12Z
  • T170 control out to 126 hrs ("Aviation
    forecast"), then T62 out to 16 days
  • 10 perturbed forecasts generated and run the same
    way as at 00Z
  • Results in 23-member ensemble over 24 hours

27
Global model ensemble products
  • On web
  • Spaghetti diagrams
  • NH ensemble mean/spread
  • NH control fcst and normalized spread
  • Relative measure of predictability
  • Probabilistic precipitation forecasts

28
NCEP Global EPS Configuration
  • Future plans
  • Add 06Z and 18Z ensembles (10 perturbations and
    control)
  • Would give 45 ensemble members per day
  • Increase ensemble control and member resolution
    concurrent with increases in operational model
    resolution
  • First 5 days or so only, no demonstrated benefit
    beyond the medium range

29
  • Why We Need Ensembles for Mesoscale)
  • Deal with uncertainties in analyses and model
    formulation
  • But Requires tradeoffs when computer resources
    limited (e.g., model resolution)
  • But But Mesoscale predictability often
    substantially controlled by synoptic
    predictability (and uncertainties therein)
  • gt Subjective or statistically based
    downscaling possible to get
    uncertainties in mesoscale weather
  • Ideal Ensembles with highest resolution
    justifiable (Issue point of diminishing return?)
  • Compromise Combination of single (or few) high
    resolution and coarser resolved ensemble

30
CONSIDER!!!
  • High Resolution Mesoscale models
  • allow us to see features not in coarser models
  • But even small timing and placement errors can
    be significant in attempt to accurately forecast
    details (see Mass, et al., 3/02 BAMS!!!).
  • But But Forcaster judgement could mitigate
  • One model (even with forecaster input)
  • is an all or nothing proposition gt

One detailed mesoscale model based
forecast could allow the user to make
highly specific and detailed inaccurate
forecasts. (after Grumm)
31
  • NWS Vision 2005 Goals
  • NWS will move towards adding probabilistic
    forecast products
  • Move from subjective model to model
    comparison to the use of more objective
    ensemble prediction systems
  • American Meteorological Society (AMS)
    Statement - Enhancing Weather Information with
    Probability Forecasts (3/02 BAMS!)
  • The AMS endorses probability forecasts and
    recommends their use be substantially
    increased.

32
(No Transcript)
33
ROLE IN SEAMLESS PRODUCT SUITE Together,
Medium Range Ensemble Forecasting (MREF) and
Short Range Ensemble Forecasting (SREF) systems
are viewed as integral in a seamless suite of
products that enable estimates in the forecast
confidence of specific weather threats - first,
in the context of the larger-scale circulation
patterns and associated weather at medium ranges
(3-10 days) and, then, in the details of
specific weather systems and sensible weather at
short ranges (0-3days).
34
Current Short Range Ensemble Forecast (SREF)
System
  • Meso Eta Model (4 members 1 control)
  • Eta coordinate, 48-km resolution, 45 layers
  • Updated in January 21, 2002 to have the same
    model physics as the operational Eta-12
  • Regional Spectral Model (RSM, 4 members 1
    control)
  • Sigma coordinate, 48-km equivalent resolution,
    42 layers
  • Has old AVN/MRF large scale precipitation and
    convective schemes (inferred clouds, SAS using
    tallest convective cloud)
  • Moving toward having same precipitation/convection
    physics as operational AVN/MRF
  • Other physics essentially the same as operational
    AVN
  • /- perturbations from separate Meso Eta and RSM
    breeding cycles

35
Current Short Range Ensemble Forecast, Cont.
  • Forecasts initialized from 21z and 09z so that
    SREF output available by time 00z and 12z Meso
    Eta Model (12 Km) run completed
  • SREF guidance available for use with the Meso Eta
    run
  • Domain is full North American continent
  • GRIB files at 40-km resolution (Grib-212)
  • Forecast range is 0-63 hours.

36
Short Range Ensemble Forecast (Near-)Future
plans
  • Eta ensemble runs with Kain-Fritsch (KF)
    convection, all else left the same, now being
    tested, operational by end of June 2002
  • Control and 4 ensemble perturbations
  • A model physics-perturbed ensemble of 10 when
    combine with SREF Eta with Betts-Miller-Janjic
    (BMJ)
  • Multi-Model - initial condition and physics
    perturbed ensemble of 15 when combined with SREF
    Eta-BMJ and RSM

37
(Longer Term) Future Work
  • System Development
  • Test and evaluate possible enhancements
    (relative contributions and any necessary
    tradeoffs)
  • PERTURB PHYSICS (e.g.,
    Convection, BL processes)
  • ADDITIONAL MEMBERS (40-50)
  • ADDITIONAL MODEL DIVERSITY
    (E.G., RUC, ETA/KF)
  • INCREASED MODEL RESOLUTION
  • TRANSITION TO WRF (LONGER
    TERM)
  • Product Development
  • Generic
  • User Specific
  • Statistical Post Processing (Generic
    Ensemble MOS)
  • Extensive Forecaster/User Training and Education

38
Ensemble Products on Web
  • Challenge
  • Avoid information overload!
  • Condense information from inidvidual ensemble
    members to useful and user friendly form
  • Assure users make appropriate interpretations of
    SREF products

http//lnx48.wwb.noaa.gov/SREF/SREF.html
39
General SREF (and Global Ensemble) Product
Categories
  • Mean and spread
  • Spaghetti diagrams of one contour line from all
    ensemble members
  • Probability charts (based on how many ensemble
    members meet or exceed certain thresholds)

40
Mean and Spread Interpretation
41
3 day forecast from 00 UTC 11/2/01
Depth uncertainty - how strong will trough be?
Phase uncertainty - where will the trough axis be?
SD - meters
42
Uncertainty in strength of system
43
3 day forecast from 00 UTC 11/2/01
Uncertainty in intensity
Uncertainty in forward speed/location
980, SD14
SD - hPa
Hurricane Michelle
44
Mean and Spread Advantages
  • Compact communication of ensemble forecast
    information
  • Can see field over entire domain
  • Ensemble mean on average has greater skill than
    any individual member
  • Spread (sample standard deviation) quantifies the
    degree of uncertainty

45
Mean and Spread Limitations
  • Assumes normal distribution of forecasts (bell
    curve with maximum likelihood at mean)
  • Mean may hide important details
  • Bi- or multi-modal solutions
  • Timing problems in prediction of features
  • Precipitation forecasts (particularly where
    convective precipitation is expected to be
    important)
  • Can use spread as guide to where mean may not be
    communicating the correct information, and use
    additional tools to make further assessments

46
(No Transcript)
47
Ridge/Trough Highly predictable
Unpredictable, Strong gradient
48
010519/0000V63 SREFX-CMB 500MB 5820M
49
3-day forecast from 00 UTC 11/2/01, spaghetti
diagram for ensemble
Uncertain location of incoming western trough
Uncertain amplitude of eastern trough
From CDC web site http//www.cdc.noaa.gov/map/im
ages/ens
50
Spaghetti Diagram Interpretation
Phase uncertainty
Amplitude uncertainty
51
Spaghetti Diagram Interpretation Clustering
Clustering
Ensemble mean
52
Spaghetti Diagrams Interpretation
53
Spaghetti Diagrams Advantages
  • Avoids the assumption of normally distributed
    data and pitfalls thereof
  • Can tell if ensemble mean, if present, is
    representative of the ensemble as a whole
  • Shows spread among ensemble members and whether
    there is clustering of members around two or more
    forecasts
  • Shows mode (I.e. most frequently occurring
    solution)
  • Indicates outliers which may overly influence the
    ensemble mean and spread

54
Spaghetti Diagrams Limitations
  • Limited to one or only a few contours
  • Cannot see full field of interest over the full
    domain
  • May not choose the right contour (use ensemble
    mean/spread to make the best choice)

55
Sequence provides info on envelope of storm tracks
010519/0000V63 SREFX-CMB SFC LOWS
56
010519/0000V63 SREFX-CMB 24HR PQPF OF .25
57
010519/0000V63 SREFX-CMB SHADED, IN AT LEAST
60 OF MEMBES
58
Probability charts combine info from Spaghetti,
but later permits view of individual solutions
010519/0000V63 SREFX-CMB SPGHETTI .25
59
010519/0000V63 SREFX-CMB LIFTED INDEX PROB 0F lt
-4
60
010519/0000V63 SREFX-CMB CAPE gt 2000J/KG
61
Probability charts Advantages
  • Depicts probabilities for exceeding critical
    value in a compact manner
  • Variable of interest is seen over the full domain
  • Uses actual distribution of data from ensemble
    members to determine probabilities

62
Probability charts Limitations
  • Do not get information on full PDF
  • Only know percentage of ensemble members that
    exceed the value (sampling problem of limited
    ensemble size)
  • Need to use several threshold values for complete
    picture
  • Does not depict maximum value

63
EXAMPLE OF POINT DATA BOX AND WHISKER PLOTS
0 6 12 18 24 30
36 42 48 54 60
FHRgt
The blue boxes represent values from the 25
quartile (bottom of the box) to the 75 quartile
(top of the box) with the median of the ensemble
as a horizontal line in the box. The whiskers
extend to both the max and min values supplied
by the EPS output.. This allows you to instantly
ascertain the uncertainty and the median (not
the mean) in one view.
64
SREF Behavior Current Configuration
  • RSM and Eta tend to group into separate clusters
    (especially for QPF)
  • SREF Eta tends to have smaller spread than the
    RSM
  • Illustration in next three slides
  • YellowRSM
  • GreenSREF Eta
  • Bold contours RSM and Eta means

65
(No Transcript)
66
(No Transcript)
67
outliers
68
SREF Behavior The Eta-KF Ensemble Subset
  • Increased ensemble spread
  • Could be result of using KF/mass-flux scheme or
    use of 4th order (less damping) diffusion scheme
  • Sharper gradients, including in stability
    parameters
  • Tendency for larger instability
  • Consistent with later triggering of convective
    scheme than for BMJ, particularly in weakly
    capped regions

69
SREF Eta/BMJ SLP mean/spread
70
SREF Eta/KF SLP mean/spread
Increased spread
71
SREF Eta/BMJ CAPE
72
SREF Eta/KF CAPE
73
Same Deterministic Model with Different
Convection Schemes Results In Different
Precipitation Forecasts
74
EMC
75
Focus on Winter Weather

ER/NCEP Winter Weather Experiment
  • Goals
  • Improve Winter Weather Services to the public
    through coordination of the winter weather
    watches/warnings with National guidance products
  • Test short range ensemble for their applications
    to winter weather forecasting
  • Results Encouraging

76
WHY??
  • Recent Snowstorms Creating the challenge for a
    more focused NWS effort
  • January 25, 2000
  • December 30, 2000
  • March 4-6, 2001
  • HPC-WFO Interaction
  • Need for coordination as forecast capabilities
    are extended to medium range
  • Need for a more unified message to the media
    (quantify uncertainty)

77
MAJOR SNOWSTORM AMBUSHES WASHINGTON
Not Good- especially when effecting DC (just
after announce-ment of new Super Computer by
NWSHQ
78

79
NCEP Winter Weather Experiment
  • Time line Nov 1 May 1
  • Participants
  • NCEP EMC
  • Provide Short Range Ensemble (SREF) Guidance
    (operational in May 2001)
  • NCEP HPC
  • Provide SREF- and MRF ensemble- based Winter Wx
    guidance
  • Collaboration with WFOs (Chat Room Technology)
  • WFOs (Eastern Region)
  • Mt. Holly, State College, Sterling, Wakefield
  • Use graphical guidance from NCEP to produce
    coordinated Winter Storm Watches/Warnings

80
WWE Products/Services
  • 2 shifts per day
  • 630AM/PM 430PM/AM
  • Issuing 3 graphics
  • Watch/Warning Guidance Graphic
  • Storm Tracks Graphic
  • Conditional Probability Graphic
  • Using Chat Room software
  • As WFO coordination tool

Low track valid 12Z/06 00Z/08
81
Winter Weather Products directly fromSREF
(graphical)
  • Probability of freezing rain for each 3, 6, 12
    and 24 hour period
  • Joint probability of freezing rain and PQPF
    exceeding specified criteria
  • Mean, maximum and minimum snow amounts and
    freezing rain for 3, 6, 12 and 24 hour periods

82
SREF Conditional Probability Product
Example of product directly from SREF that is
available to HPC as guidance for generating
collaborative winter weather products. Storm
tracks include SREF and all other models
83
Watch/Warning Guidance Graphic
  • Amounts compared to threshold values from ERH
  • Color coded to indicate percentage of threshold
    met
  • Graphic produced for Day 1 and Day 2
  • Available at T5hrs from ensemble start (14Z/01Z)
  • Posted to chat room for comments by 1445Z/0245Z
  • Distributed at 15Z/03Z
  • Second chat session at 19Z/07Z if necessary

84
24 Hour Snow Thresholds
85
24 Hour Freezing Rain Thresholds
86
Watch/Warning (WSW) Guidance Graphic
87
Ensemble Performance January 31, 2002
Prob Snow
Prob Freezing Rain
27 hour forecasts valid 12Z January 31
9AM Radar Jan. 31, 2002
88
Ensemble Performance January 31, 2002
Dominant precipitation type 27 hour forecast
valid 12Z January 31
9AM Radar January 31
89
9AM Satellite January 31, 2002
Now thats a Mesoscale feature!!
Whose knife was used to slice this?
90
CASE STUDIES WINTER
Jan 25-26, 2000 East Coast Surprise
Snowstrom Dec 29-30, 2000 East Coast
Millennium Snowstorm WARM SEASON April
7-8, 2002 Texas Squall Line May 3-4, 2002
Southeast U.S Precip Tropical Storm
Barry (August 2001)
91
Based on Op models, official forecasts for DC
as late as 21Z 24 Jan called for only
40 chance of light snow!
92
Ensemble provides a clear heads up on morning
of 24th for the possibility of a major snow
event, especially when considered in context of
independent information from satellite imagery
and radar that suggested storm track closer to
coast and precip further inland than available
operational models were indicating
93
Wide range of solutions (CTL vs Best) in precip
and Storm Track (next) gt Deterministic (yes/no
forecast) very risky!
94
12 hr (top) and 24 hr (bottom) MSLP from 12Z 24
Jan for worst and best SREF members.

95
Official forecast gt winter storm warn-ing
with of 3-6 inches predicted in DC and 5-10
inches in Balt. Reality DC and Balt woke up
on the morning of the 30th with sunny skies and,
surprise, no snow.
Eta from 12 GMT 29 Dec 24 hour accuml precip
ending 00GMT 31 Dec 2000
96
SREF 24hr spaghetti from 12Z 29 Dec for .50 12
hr precip ending 12 GMT 30 Ensemble indicated
a 30-40 chance of signif- icant snow thus, in
this case SREF gave a heads up (60) for the
chance of no snowstorm
97
KEY POINTS
January 24/25, 2000 DC Snowstorm Ensembles
gave heads up for snowstorm in face of
deterministic model and official forecasts of
no snow December 30, 2000 DC
Non-Snowstorm Ensemble gave heads up of no
snow in face of deterministic model (Eta)
predicted and official forecasts of snowstorm
98
EMC
99
WARM SEASON CASE MAY 4-5,02
RSM
Eta KF
Eta BMJ
100
WARM SEASON CASE MAY 4-5,02
Op Eta (12km)
101
WARM SEASON CASE APRIL 7-8,02
102
WARM SEASON CASE APRIL 7-8,02
Eta KF
Eta BMJ
OpEta (12km)
RSM
103
WARM SEASON CASE APRIL 7-8,02
Eta BMJ
Eta KF
RSM
OpEta (12 km)
104
(No Transcript)
105
60 hr Eta vt 12Z Aug 6, 2001
106
63 hr SREF vt 12Z Aug 6, 2001
107
(No Transcript)
108
(No Transcript)
109
Summary Due to observational and model
limitations, the ensemble method offers a
scientifically sound basis to utilize over the
standard deterministic approach for NWP.
Further, EPS output supports operational
forecasters providing both deterministic or
probabilistic forecast products. In
principle, EPS output facilitates an opportunity
to provide ranges of scenarios that may occur, as
well as helping to identify the most likely
result Research indicates that an ensemble
comprised of different models will likely offer
better results than an EPS from one model gt
combination of uncertainties related to
dynamically amplifying analysis errors and model
formulation
110
Summary (cont.) There a a number of other ways
to display data EPS output, and each user can
decide which method will benefit them most gt
learning process and continual interaction
between forecasters and EPS developers An EPS
comprised of members lower in resolution can
provide more useful information and user value
than a single higher resolution deterministic
run. Future Directions This is an evolving
science and as computing power and verification
efforts increase and change, so too will the
techniques to produce and visualize ensemble
output ? Stay Tuned!!
111
EMC
112
Additional Ensemble Links
http//www.wmo.ch/web/www/DPS/WS-S/PROCEEDINGS/
Introduction.htm http//www.euromet.met.ed.ac.uk/
ucisa/teachers/english/nwp/n7400/ n7400003.htm
http//eyewall.met.psu.edu/mos/index.html
http//sgi62.wwb.noaa.gov8080/ens/training/ncep
wks/ncepwks.html http//sgi62.wwb.noaa.gov8080/
ens/enshome.html http//lnx48.wwb.noaa.gov/SREF/
SREF.html http//eyewall.met.psu.edu/ensembles2/
index.html http//eyewall.met.psu.edu/SREF/index
.html http//www.meteo.psu.edu/gadomski/ewall.h
tml http//www.met.utah.edu/jhorel/html/models/m
odel_ens.html http//www.wmo.ch/web/www/DPS/ET-E
PS-TOKYO/Documentation-plan.html
http//www.hpc.ncep.noaa.gov/ensembletraining
http//www.ecmwf.int/newsevents/training/rcourse
_notes/GENERAL_CIRCULATION/CHAOS/Chaos.html
Write a Comment
User Comments (0)
About PowerShow.com