Transcript ppt file
Track Forecasting of 2001
Atlantic Tropical Cyclones
Using a Kilo-Member
Ensemble
8:30 AM April 30, 2002
Jonathan Vigh
Master’s Student
Colorado State University
Department of Atmospheric Science
Acknowledgments
Wayne Schubert, graduate advisor
Mark DeMaria
Scott Fulton
Rick Taft
Funding
– Significant Opportunities in Atmospheric
Research and Science (SOARS) Program,
fellowship
– American Meteorological Society, fellowship
What causes track error?
Inaccurate spatial and temporal sampling ->
analysis uncertainty
Incomplete representation of physical processes
Discretization and truncation error
Atmosphere’s inherent chaotic nature
– Instability
– Nonlinear interactions between various spatial scales
(Leslie et al., 1998)
What is an ensemble?
A collection deterministic realizations
obtained by varying:
– Model (numerics, resolution, or physics)
– Perturbing model parameters
– Initial analysis fields
Time of analysis
Generation method (adjoint or 4DVAR methods)
Stochastic or bred perturbations
Benefits of an ensemble
With many realizations over properly perturbed
initial conditions, the subspace of dynamical
pathways can be sampled
Ensemble mean is generally more accurate than
a single deterministic forecast (Leith, 1974)
Ensemble allows estimation of higher moment
statistics of the forecast
– Forecast uncertainty
– Bifurcation of dynamical pathways
– Probability density functions
MBAR: Multigrid Barotropic Model
Modified barotropic vorticity equation
Finite difference, multigrid methods (Fulton,
2001)
3 Nested grids on a square 6000 km domain
h1 = 125 km, h2 = 63 km, h3 = 31 km
Efficient and accurate multigrid methods makes
a kilo-scale operational model feasible
– Accuracy comparable to LBAR in only 1/38 of the
computing time
– Each 120-hr track forecast takes ~2.5 seconds
– A 1980 member ensemble takes approximately 1.3
hours on a 1 GHz Intel PC
Perturbations in initial and
background fields
Operational MRF ensembles
5 independent breeding cycles are used in
the analysis cycle to estimate subspace of
fastest growing analysis errors (Toth and
Kalnay, 1997)
Adding and subtracting these vectors from
the analysis yields 10 + control initial and
background fields for MBAR
00Z at 2 degree resolution
Perturbations to vertical averaging
Four deep layer vertical averages of wind
field simulate uncertainties in steering
layer depth
Pressure weighted averages of following
layers (mb):
– Shallow
– Medium
– Deep
– Entire
(850 - 700)
(850 – 350)
(850 – 200)
(1000 – 100)
Perturbations to vertical
decomposition of vertical modes
In 2D barotropic vorticity models, ultra-long
Rossby waves experience excessive
retrogression (Wiin-Nielsen, 1959)
Inclusion of inverse Rossby radius in the
prognostic equation can fix this
Uncertainties in the vertical decomposition of the
tropical atmosphere are handled by perturbing
equivalent phase speed
– 50 ms-1
– 150 ms-1
– 300 ms-1
Perturbations to vortex
size/strength
Simulates uncertainties in the size and
strength of the vortex
– Weak or small TS (vmax = 15 ms-1)
– Weak or medium sized hurricane (wmax = 35
ms-1)
– Strong or large hurricane (vmax = 50 ms-1)
For a barotropic model, the size of the
outer circulation is important factor in the
track forecast
Perturbations to storm motion
vector
Simulates uncertainties in the initial storm
location and direction
Motion vector added to wind field of bogus
vortex with exponentially decaying blending
radius
–
–
–
–
–
No motion perturbation
Fast and to right
Slow and to right
Fast and to left
Slow and to left
Cross-multiplication across the five
perturbation classes
11 initial and background fields (180)
4 deep layer averages
(495)
3 vertical decompositions
(660)
3 vortex sizes/strengths
(660)
5 motion vectors
(396)
--------------------------------------------1980 ensemble members
26 sub-ensembles
So what does the kilo-ensemble
look like?
Chantal
Dean
Erin
Iris
Michelle
Olga
Chantal
August 17 and 18
Well handled by the ensemble
A fairly weak storm embedded in trade
flow
For first several days of forecast, a tight
envelope, indicating high confidence
Dean
August 23
Example of the challenges of recurvature
off the East Coast
Total ensemble mean lagged behind
actual path
Ensemble ‘swarm’ stretched out in the
direction of recurvature
Erin
September 3
A storm which weakened to a tropical
depression, then later strengthened to a
strong hurricane
Ensemble mean successfully predicted
path, although significant cross-track
spread developed
Iris
October 5
Another example of a tight envelope early
on, suggesting high forecast confidence
Ensemble spreads out at end, but actual
track still contained in envelope
A minority of members experience
recurvature
Michelle
Storm tracked along edge and then
outside of envelope – an example of the
ensemble’s failure to accurately span all
dynamical pathways
An example of rapid growth in ensemble
spread with time, suggesting low
confidence in ensemble mean forecast
Ensemble mean caught in middle of
bifurcation -> large errors
Olga
Forecasts for 11/24/02 inaccurately
predict recurvature – large spread
develops
Forecasts for 11/25/02 catch onto the
correct path. Large spread but ensemble
means are accurate
Forecasts for 11/26/02 show an example
of one sub-ensemble correctly picking the
storm path
Average error of sub-ensembles based on vertical averaging
800
700
Average track error (nm)
600
500
CLP5
A5K0
ZTOT
400
SLY4
SLY5
SLY6
SLY7
300
200
100
0
0
12
24
36
48
60
72
Forecast period (hrs)
84
96
108
120
Average track error for sub-ensembles over ceqv
700
600
Forecast period (hrs)
500
CLP5
400
A5K0
ZTOT
SGM1 (50 m/s)
300
SGM2 (100 m/s)
SGM3 (300 m/s)
200
100
0
0
12
24
36
48
60
72
Average track error (nm)
84
96
108
120
Average track error of sub-ensembles based on vortex size/strength
700
600
Average track error (nm)
500
CLP5
400
A5K0
ZTOT
SVM1 (15 m/s)
SVM2 (30 m/s)
300
SVM3 (50 m/s)
200
100
0
0
12
24
36
48
60
72
Forecast period (hrs)
84
96
108
120
Average xbias of sub-ensembles based on vertical averaging
50
0
0
12
24
36
48
60
72
84
96
108
120
Average xbias (nm)
-50
-100
CLP5
A5K0
ZTOT
-150
SLY4
SLY5
SLY6
SLY7
-200
-250
-300
-350
Forecast period (hrs)
Average xbias over sub-ensembles based on ceqv
50
0
Average xbias (nm)
1
2
3
4
5
6
7
-50
8
9
10
11
CLP5
A5K0
ZTOT
SGM1 (50 m/s)
SGM2 (100 m/s)
SGM3 (300 m/s)
-100
-150
-200
Forecast period (hrs)
Average ybias over sub-ensembles based on ceqv
150
100
Average ybias (nm)
50
CLP5
0
A5K0
0
12
24
36
48
60
72
84
96
108
120
ZTOT
SGM1 (50 m/s)
SGM2 (100 m/s)
-50
SGM3 (300 m/s)
-100
-150
-200
Forecast period (hrs)
Average ybais of sub-ensembles based on vertical averaging
80
60
Average ybias (nm)
40
20
CLP5
A5K0
ZTOT
0
SLY4
1
2
3
4
5
6
7
8
9
10
11
SLY5
SLY6
SLY7
-20
-40
-60
-80
Forecast period (nm)
Average xbias over sub-ensembles based on vortex size/strength
50
0
0
12
24
36
48
60
72
84
96
108
120
-50
Average xbias (nm)
-100
-150
CLP5
A5K0
ZTOT
-200
SVM1 (15 m/s)
SVM2 (30 m/s)
-250
SVM3 (50 m/s)
-300
-350
-400
-450
Forecast period (hrs)
Average ybas of sub-ensembles based on vortex size/strength
100
50
Average ybias (nm)
0
0
12
24
36
48
60
72
84
96
108
120
CLP5
A5K0
ZTOT
-50
SVM1 (15 m/s)
SVM2 (30 m/s)
SVM3 (50 m/s)
-100
-150
-200
Forecast period (hrs)
Average spread across perturbation class
1800
1600
Total average spread (nm)
1400
1200
SBF
SGM
1000
SLY
SMV
800
SVM
ZTOT
600
400
200
0
12
24
36
48
60
72
Forecast time (hrs)
84
96
108
120
Conclusions
Swarm diagrams can lead to useful estimates of
forecast confidence
Ensembles are useful for spotting bifurcations in
possible future tracks
The ensemble mean isn’t more accurate than
control
But great utility of ensembles is the estimation
of higher moments, such as forecast spread ->
estimates of forecast reliability
Future Work
Tune ensemble perturbation classes to reduce bias of
ensemble mean
Tune ensemble spread to be the ‘right’ size
Calculate probability density functions of storm location
from ensemble output
Use fuzzy logic/adaptive weighting/neural network to
select more accurate custom sub-ensembles
Automate for operational use, web output
Create ensemble toolbox for forecasters allowing effect
of perturbations in parameters on the forecast track to
be easily seen and quantified
References
Fulton, S. R., 2001: An adaptive multigrid barotropic
cyclone track model. Mon. Wea. Rev., 129, 138-151.
Leith, C. E., 1974: Theoretical skill of Monte Carlo
forecasts. Mon. Wea. Rev., 102, 409-418.
Leslie, L. M., Abbey, R. F., and Holland, G. J., 1995:
Tropical Cyclone Track Predictability. Meteorol. Atmos.
Phys., 65, 223-231.
Toth, Z., and E. Kalnay, 1997: Ensemble forecasting at
NCEP and the breeding method. Mon. Wea. Rev., 125,
3207-3310.
Wiin-Nielsen, A., 1959: On barotropic and baroclinic
models, with special emphasis on ultra long waves. Mon.
Wea. Rev., 87, No. 5, 171-183.