The ECMWF Approach to Ensemble Prediction
Transcript of The ECMWF Approach to Ensemble Prediction
![Page 1: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/1.jpg)
The ECMWF Approach to Ensemble Prediction
Tim Palmer
University of Oxford, ECMWF.
![Page 2: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/2.jpg)
With particular thanks to
Jan Barkmeijer, Roberto Buizza, Philippe Chapelet, Dennis Hartmann, Renate Hagedorn, Martin Leutbecher, Jean-Francois Mahfouf, Martin Miller, Franco Molteni, Robert Mureau, Thomas Petroliagis, Kamal Puri, David Richardson, Glenn Shutts, Stefano Tibaldi, Joe Tribbia….
… and all of ECMWF RD and OD
![Page 3: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/3.jpg)
With thanks to Met Office staff
Doug Mansfield, James Murphy
1985: The First Semi-Operational Real-Time Ensemble Forecast
![Page 4: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/4.jpg)
Cf Miyakoda et al, 1982
![Page 5: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/5.jpg)
PNA+
PNA-
DAY 4 DAY 2 DAY 9 DAY 30
Integrations with a Barotropic Vorticity Equation Model
![Page 6: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/6.jpg)
If eigenvectors are not normal (operator not self-
adjoint), perturbation growth is not bounded by the
dominant eigenvalue
n0 = ax1 -bx2
g = 1cosq
q
g -PNA >> g +PNA
Zhang c. 1989
![Page 7: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/7.jpg)
From Buizza and Palmer, 1995
![Page 8: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/8.jpg)
Use SVs to initialise EPS. Why?
1. Meteorologically balanced perturbations
2. Initial-time perturbations with amplitudes focussed on sub-synoptic wavenumbers where analysis errors (observation error and model error) are likely to be significant, and final-time perturbations on synoptic and smaller wavenumbers where forecast error is dominant.
3. Perturbations with growth rates that would compensate for the overly dissipative nature of the model at large wavenumbers.
4. Since we don’t have a good quantification of initial error (esp impact of model error on initial error - see below), focussing on potential worse-case perturbations is no bad thing. The main problem with ensemble prediction (to the present day) is underdispersion -overconfidence.
![Page 9: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/9.jpg)
![Page 10: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/10.jpg)
![Page 11: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/11.jpg)
But now …. Perturbations produced by Ensemble Data Assimilation (EDA)…. Surely
no need for SVs any more??
![Page 12: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/12.jpg)
Simon Lang: Personal Communication
![Page 13: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/13.jpg)
Stochastic Physics
![Page 14: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/14.jpg)
Workshop on New
Insights and Approaches
to Convective
Parametrization
4-7 November 1996
Probability of an “on”cell
proportional to CAPE and
number of adjacent “on”
cells – “on” cells feedback
to the resolved flow
![Page 15: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/15.jpg)
Stochastically Perturbed Parameterisation Tendencies (SPPT)
A Holistic Scientifically Justified Approach to Representing Model Uncertainty
• Operational scheme in ECMWF’s ensemble prediction system
• Perturbations to total parametrised tendency of physical processes with multiplicative noise
for X={u,v,T,q}
• r is a uni-variate random number described through a spectral pattern generator which is smooth in space and time
• Spectral coefficients of r are described with an AR(1) process
• Gaussian distribution, truncated at ±2s
![Page 16: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/16.jpg)
![Page 17: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/17.jpg)
![Page 18: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/18.jpg)
0 1 2 3 4 5 6 7 8 9 100
5
10
15
lead time
en
se
mble
sp
rea
d
L63 st
L63 det
Spread L63 and L63 stoch.
Spread decreases (not increases!) with stochastic noise
Kwasniok, 2014
s=0
s=3.2
s=8
![Page 19: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/19.jpg)
0 10 20 30 40 50 60 70 80
m
12
34
56
78
9Forecast Day
Mean m
ethod: standard | Population: 691,688,686,685,682,681,679,2*671,670 (averaged)
oper_an ti enfo prod
Date: 20160101 00UTC to 20161231 12UTC
NHem Extratropics (lat 20.0 to 90.0, lon -180.0 to 180.0)
500hPa geopotential
Annual mean 2016
spread ECMW
F
EMrm
se ECMW
F
spread UKMO
EMrm
se UKMO
spread NCEP
EMrm
se NCEP
Martin Janousek: Personal Communication
ECMWF EPS spread/skill better balanced than other operational ensembles.
![Page 20: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/20.jpg)
The Future: More Interaction Between SVs and SPPT?
Use EDA to set SV amplitudes regionally?
![Page 21: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/21.jpg)
Ensemble Size vs Resolution
• We must not drop below 50 members. Needed to optimiseprobabilistic scores for high-impact weather (see Martin Leutbecher’s talk).
• Two complementary routes to developing high-resolution (e.g. 5km) 50 member ensembles….
![Page 22: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/22.jpg)
Stochastic Parametrisation
Triangular
Truncation
Partially Stochastic
Double precision
Single precision
Half precision?
Quarter precision?
1. State-dependent precision….
… a job for scientists
![Page 23: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/23.jpg)
25/08/17 00:00850hP wind speedT511L91
n = 0 52-bit significand
0 < n £160 11-bit significand
160<n £ 320 9-bit significand
320<n £ 511 7-bit significand
Matthew Chantry, Oxford – Peter Düben, ECMWF.
HurricaneHarvey
![Page 24: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/24.jpg)
2. Make the case for much greater computing capability, so ECMWF can remain the World No 1.
… a job for management
![Page 25: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/25.jpg)
Extra Slides
![Page 26: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/26.jpg)
The application of ensemble ideas to monthly (and seasonal) forecasting was not controversial.
What was more controversial is that ensemble forecast methods should be developed within what was referred to as “the limit of deterministic predictability”. Medium-range ensemble forecasting was treated with scepticism.
Instead of probabilistic prediction, the thought at the time was that we should be trying to “forecast the forecast skill” of the deterministic forecast.
This turned out to be a forlorn hope, though it led to some very important scientific spin offs, such as how to perturb the initial conditions in a medium-range ensemble forecast.
![Page 27: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/27.jpg)
Mesoscale Area of Interest1-km AGL Reflect. >40 dBZ
20160502-2016060300Z cycle; fh013-030
40-km Radius of Influence10 grid-point Gaussian smoothing parameter
Neighborhood Reflectivity Verification: CLUE Comparison to SSEO: Reliability
SSEO: 7-member multi-model
CLUE_M10: 5 ARW, 5 NMMB
CAPSEnKF: 9-member ARW
NCAREnKF:10-member ARW
no resolution
NOAA Hazardous Weather Testbed Spring
Forecasting Experiment (I. Jirak and A. Clark)
![Page 28: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/28.jpg)
x '(t) = L(t,t0 ) ¢x (t0 )
Let || x || denote the perturbation's energy.
Let (x;y) denote the corresponding inner product.
Then
maxx(t0 )¹0 (|| x(t) || / || x(t0 )) = s 1
where s1 is the largest eigenvalue for
(LL*)v i (t) = s i
2v i (t)
i.e. the largest singular value of L
The corresponding singular vectors, ranked by the size of s i
define the fastest growing perturbations
![Page 29: The ECMWF Approach to Ensemble Prediction](https://reader030.fdocuments.in/reader030/viewer/2022012104/616a1b4a11a7b741a34edfba/html5/thumbnails/29.jpg)
“If more realistic models with many
thousand variables also have the
property that a few of the eigenvalues
of AAT are much larger than the
remaining, a study based upon a small
ensemble of initial errors should..give a
reasonable estimate of the growth rate
of random error.” Lorenz (1965)
Lorenz (1965): A Study of the Predictability of a 28-Variable
Atmospheric Model, Tellus.