probabilistic prediction at ecmwf
DESCRIPTION
Probabilistic prediction at ECMWF. Roberto Buizza European Centre for Medium-Range Weather Forecasts. Acknowledgements to Magdalena Balmaseda , Laura Ferranti, Martin Janousek , Martin Leutbecher, David Richardson, Tim Stockdale, Frederic Vitart, and the Predictability Division. Outline. - PowerPoint PPT PresentationTRANSCRIPT
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 1
Probabilistic prediction at ECMWF
Roberto BuizzaEuropean Centre for Medium-Range Weather Forecasts
Acknowledgements to Magdalena Balmaseda, Laura Ferranti, Martin Janousek, Martin Leutbecher, David Richardson, Tim
Stockdale, Frederic Vitart, and the Predictability Division.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 2
Outline
1. ECMWF medium-range probabilistic forecast systems: the EPS
2. ECMWF seasonal probabilistic forecast systems:
a) The operational S3 and EUROSIP
b) The forthcoming S4 (Q4-2011)
3. Conclusions
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 3
Delayed Ocean Analysis ~12 days
Real Time Ocean Analysis ~8 hours
HRESTL1279L91 (d0-10)
HRESTL1279L91 (d0-10)
S3TL159L62 (m0-7/13)
S3TL159L62 (m0-7/13)
EPSTL639L62 (d0-10)
TL319L62 (d10-15/32)
EPSTL639L62 (d0-10)
TL319L62 (d10-15/32)
Atmospheric model
Wave model
Ocean model
Atmospheric model
Wave model
1. Baseline operational systems (2011)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 4
1. The operational ECMWF EPS
The operational version of the EPS includes 51 forecasts with resolution:• TL639L62 (~32km, 62 levels) from day 0 to 10
• TL319L62 (~64km, 62 levels) from day 10 to 15 (32 at 00UTC on Thursdays).
Initial uncertainties are simulated by adding to the unperturbed analyses a combination of T42L62 singular vectors, computed to optimize total energy growth over a 48h time interval (OTI), and perturbations generated using the new ECMWF Ensembles of Data Assimilation (EDA) system.
Model uncertainties are simulated by adding stochastic perturbations to the tendencies due to parameterized physical processes (SPPT and SPBS schemes).
The EPS is run twice a day, at 00 and 12 UTC.
NH SH TR
Definition of the perturbed ICs
1 1 2 2 5050 5151…..
Products Products
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 5
1. Simulation on initial unc. with singular vectors
Since its implementation in 1992, initial uncertainties have been simulated in the EPS using singular vectors, i.e. perturbations with the fastest finite-time growth rate.
Perturbation growth can be measured in terms of total energy E with a total-energy norm, and a perturbation time evolution can be linearly approximated:
Once the adjoint of the tangent forward propagator with respect to the total-energy norm is defined, total-energy singular vectors can be computed by solving an eigenvalue problem:
time T
0)0,()( ztLtz
002
)0,(;)0,()();()( ztELztLtzEtztz
jjjELELE 221*21
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 6
Since 22 June 2010, differences between 6h perturbed forecasts from the ECMWF Ensemble Data Assimilation system have been combined with singular vectors to generate the EPS initial perturbations.
The TL399L91 EDA perturbed analyses are generated by randomly perturbing the observations and the SST, and by running the forecast model with the SPPT stochastic scheme. The random perturbations are defined by sampling a normal distribution with the observation error standard deviation.
1. Simulation of initial unc. using the EDA
It is worth to point out that since 18 May 2011, the EDA has been used to specify the background errors “of the day” in the high-resolution 4D-Var.
U850 background error standard deviationRandomization method (ope, left) – EDA (cy36r4, right)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 7
500 km6 h
1000 km3 d
2000 km30 d
1. Simulation of model unc.: SPPT(1L) & SPPT(3L)
Since Oct 1998, the EPS has included a stochastic scheme designed to simulate random model errors due to parameterized physical processes [SPPT(1L)].
Since Nov 2010, the scheme includes a multi-scale pattern generator to account for parameterization errors on multiple spatial and temporal scales [SPPT(3L)].
(from M Leutbecher)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 8
Since Nov 2010, a stochastic backscatter scheme (SPBS) is also used in the EPS: Rationale: a fraction of the dissipated energy is backscattered upscale and acts as
streamfunction forcing for the resolved-scale flow (Shutts & Palmer 2004, Shutts 2005, Berner et al 2009)
Streamfunction forcing in the latest formulation (M Steinheimer and G Shutts, 2010) is given by :
Recent improvements/updates include: Revised convective dissipation calculation Revised dissipation rate smoothing Changed pressure dependency of vertical correlations Option to force only part of the spectrum [reduces computational cost and avoids
problems seen with small scale forcing (dissipation – forcing feedback, etc.)]
Streamfunction forcing
Backscatter ratio
Total dissipation
rate
Pattern generat
or
1. Simulation of model unc.: SPBS
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 9
1. The EPS re-forecast suite
Some of the EPS products are bias corrected and calibrated using 450 EPS forecasts run for the past years:
18 years (1993 – 2010) 5 initial dates (ERA-I):
-14d, -7d, 0, +7d, +14d 5 members Same model version Same resolution Same fc length EDA initial perts from
corresponding dates of current year (2011)
Daily SVs
18y
51T639L62
51T319L62
2011
5 55 5
5 55 5
5 5
… 2 9 16 23 30 June …
2010
5 55 5
5 55 5
5 5
5 55 5
5 55 5
5 5
2009
5 55 5
5 55 5
5 5
2008
1993
…..
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 10
00UTC,12UTC,beginningNH Extratropics Lat 20.0 to 90.0 Lon -180.0 to 180.0
(12mMA = 12 months moving average)Ranked probability skill score
temperature 850hPaECMWF EPS verification
1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
T+72
T+120
T+168
T+72 12mMA
T+120 12mMA
T+168 12mMA
1. EPS performance: T850 v an, NH
The performance of the EPS has been improving continuously for upper level fields, as seen by looking at the RPSS for the t+72h, t+120h and t+168h probabilistic prediction of T850 over NH (verified against analyses).
Results indicate a predictability gain of ~ 2 days/decade.
(from M Janousek)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 11
1. EPS performance: sp/sk relationship NH
In Nov 2010, two major changes were introduced to further improve the EPS spread/skill relation:
The multi-scale SPPT(3L) replaced the original SPPT(1L), and SPBS was introduced The amplitude of the SV-initial perturbations was reduced by 50%
0
10
20
30
40
50
60
70
80
90
100
110
m
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Forecast Day
Mean method: standard
oper_an od enfo 0001
Date: 20101201 00UTC to 20110228 12UTC
N Hem Extratrop (lat 20.0 to 90.0, lon -180.0 to 180.0)
500hPa geopotential
Ensemble Mean
rmse djf2011spread djf2011
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
K
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Forecast Day
Mean method: standard
oper_an od enfo 0001
Date: 20101201 00UTC to 20110228 12UTC
N Hem Extratrop (lat 20.0 to 90.0, lon -180.0 to 180.0)
850hPa temperature
Ensemble Mean
rmse djf2011spread djf2011
(from D Richardson)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 12
1. EPS I-EFI for severe weather fc
The EPS-based I-EFI and related products (EPSgram, CDF) have become key products used to predict severe weather conditions.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 13
1. EPS & TIGGE: T850 NH, 2010, v ERA-int
CRPSS for T850 over NH for 2010 verified against own analysis of the 4 best TIGGE ensembles: ECMWF (red) NCEP (green) MetOffice (blue) CMC Canada (violet)
(from M Janousek)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 14
1. EPS & TIGGE: T/V850 NH, D10-JF11, v ERA-int
The skill of the TIGGE ensemble systems is continuously monitored.
This figures show the CRPSS for T850 (left) and WS850 (right) over NH for D10-JF11 verified against own analysis of the 4 best TIGGE ensembles: ECMWF (red), NCEP (green), MetOffice (blue) and CMC Canada (violet).
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Forecast Day
oper_an ti enfo prod | Mean method: fair | Population: 180,180,180,180,180,180,180,180,180,180,179,179,179,179,179
Date: 20101201 00UTC to 20110228 12UTC
N Hem Extratrop (lat 20.0 to 90.0, lon -180.0 to 180.0)
Continuous ranked probability skill score
850hPa temperature
ECMWF
NCEP
UKMO
CMC
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Forecast Day
oper_an ti enfo prod | Mean method: fair | Population: 180,180,180,180,180,180,180,180,180,180,179,179,179,179,179
Date: 20101201 00UTC to 20110228 12UTC
N Hem Extratrop (lat 20.0 to 90.0, lon -180.0 to 180.0)
Continuous ranked probability skill score
850hPa wind speed
ECMWF
NCEP
UKMO
CMC
(from D Richardson)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 15
1. EPS & TIGGE: TP NHx and WS850 TR
The left figure shows the CRPSS for 24h accumulated precipitation verified against synop observations for 2010-11.
The right panel shows the CRPSS of WS850 over the tropics for D10-JF11 verified against ERA-Interim.
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Forecast Day
oper_an ti enfo prod | Mean method: fair | Population: 180,180,180,180,180,180,180,180,180,180,179,179,179,179,179
Date: 20101201 00UTC to 20110228 12UTC
Tropics (lat -20.0 to 20.0, lon -180.0 to 180.0)
Continuous ranked probability skill score
850hPa wind speed
ECMWF
NCEP
UKMO
CMC
24-h Precipitation, +144 h, ExTrop (moving weekly avg)
JulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarApr
2010 2011
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
CR
PS
S
ECMWF( 0.098)JMA ( 0.016)UKMO (-0.015)NCEP (-0.041)
(from D Richardson)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 16
1. EPS & TIGGE: 2mT EU, D08JF09, obs
CRPSS for 2mT forecasts in DJF 2008/09 verified against observations at 250 European stations.
Results are shown for the four single-models (dotted lines with symbols) contributing to the TIGGE-4 multi-model (solid line without symbols) and the reforecast calibrated ECMWF model (dotted line with bullet symbols).
(from R Hagedorn)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 17
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1false alarm rate
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
hit
rat
e
Day 5-20041007-20041007ECMWF Monthly Forecast, 2-metein upper tercile , Area:Northern Extratropics
ROC score = 0.841ROC score = 0.676
0 0.2 0.4 0.6 0.8 1rel FC distribution
0.121 105
0.241 105
0.362 105
0.482 105
0.603 105
Forecast
Persistence
Persistence
Persistence
2mT
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1false alarm rate
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
hit
rat
e
Day 5-20041007-20041007ECMWF Monthly Forecast, Precipin upper tercile , Area:Northern Extratropics
ROC score = 0.686ROC score = 0.575
0 0.2 0.4 0.6 0.8 1rel FC distribution
97200.194 10
5
0.292 105
0.389 105
0.486 105
Forecast
Persistence
Persistence
Persistence
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1false alarm rate
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1h
it r
ate
Day 5-20041007-20041007ECMWF Monthly Forecast, Precipin upper tercile , Area:Northern Extratropics
ROC score = 0.856ROC score = 0.680
0 0.2 0.4 0.6 0.8 1rel FC distribution
0.123 105
0.247 105
0.370 105
0.493 105
0.616 105
Forecast
Persistence
Persistence
Persistence
1. 32d EPS scores over NH (ROCA)
MSLP TP
D5-11D12-18D19-25D26-32
D5-11D12-18D19-25D26-32
D5-11D12-18D19-25D26-32
Results based on 45 cases (ICs 4 times a year from ‘91 to ’02, 5m) indicate that 32d EPS weekly average probabilistic forecasts of 2mT, MSLP and TP over NH land points have some level of skill up to forecast day 32.
This is shown hereafter for the probabilistic prediction of occurrence of 2mT, MSLP and TP values upper tercile values.
(from F Vitart)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 18
04/05 05/06 06/07 07/08 08/09 09/10 10/110.4
0.5
0.6
0.7
0.8
ROC A
REA
04/05 05/06 06/07 07/08 08/09 09/10 10/110.4
0.5
0.6
0.7
0.8
ROC A
REA
1. EPS monthly extension: ROCA for 2mT over NH
Monthly Forecast d12-18
Persistence of day 5-11
Monthly Forecast d19-32
Persistence of day 5-18
Day 12-18 Day 19-32
The EPS forecast range is extended to 32-days twice-a-week (00UTC of Mon and Thu). The performance of EPS weekly-average forecasts has also been continuously improving up to fc-day 18 (left). The signal for longer fc days is weaker (right). This is shown here in terms of the ROC-area for the probabilistic prediction of 2mT in the upper tercile.
(from F Vitart)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 19
1. EPS d5-11 & d19-25 fcs: 2mT NA, D10J11
These plots show a time series of weekly average analysis (top row) and EPS ensemble-mean forecasts for d5-11 (middle row) and d19-25 (bottom row) 2mT forecasts over North America valid for the period 13 Dec ‘10-23 Jan ‘11.
Red/blue shading indicate areas where the ensemble-mean is significantly (at 10% level) warmer/colder than the model climatology.
13-19 Dec ‘10 20-26 Dec 27 Dec-2 Jan ’11 3-9 Jan 10-16 Jan 17-23 Jan
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 20
1. EPS d5-11 & d19-25 fcs: 2mT EU, D10J11
These plots show a time series of weekly average analysis (top row) and EPS ensemble-mean forecasts for d5-11 (middle row) and d19-25 (bottom row) 2mT forecasts over Europe valid for the period 13Dec ‘10-23 Jan ‘11.
Red/blue shading indicate areas where the ensemble-mean is significantly (at 10% level) warmer/colder than the model climatology.
13-19 Dec ‘10 20-26 Dec 27 Dec-2 Jan ’11 3-9 Jan 10-16 Jan 17-23 Jan
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 21
ERA40
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
10
-10
-10
-10
-10 -10-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanewku
10
-10
-10
-10
-10
-10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanex6i
-10
-10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
days
29/12
05/01
12/01
20/01
28/01
04/04
12/02
28R3
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
10
-10
-10
-10
-10 -10-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanelz5
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanek7h
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
10/04
31R1
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
10
-10
-10
-10
-10 -10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanes8c
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meaner98
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
09/06
32R2
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
10
-10
-10
-10
-10 -10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanewku
-10
-10
-10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanex6i
-10
-10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
06/07
32R3
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
10
-10
-1 0 -10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanewku
-10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanex6i
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/02
31/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/01
31/1230/1229/12
11/07
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanf4i5
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanf1xi
10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
33R1
06/08
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
ERA40
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanf46n
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
30
25
20
15
10
5
0
DA
YS
30E 90E 150E 150W 90W 30W
LONGITUDE
30E 90E 150E 150W 90W 30W
Ensemble meanf6ju
10
-10
-10
-10
15/0214/0213/0212/0211/0210/02 9/02 8/02 7/02 6/02 5/02 4/02 3/02 2/02 1/0231/0130/0129/0128/0127/0126/0125/0124/0123/0122/0121/0120/0119/0118/0117/0116/0115/0114/0113/0112/0111/0110/01 9/01 8/01 7/01 6/01 5/01 4/01 3/01 2/01 1/0131/1230/1229/12
35R3
09/09
1. MJO: OLR anomalies at forecast day 15
This plot shows Hovmoeller diagrams of OLR from ERA40 (left panel) and forecasts at day 15 from 6 successive versions of the ECMWF model since the monthly forecasting system became operational. Results show that there has been a considerable improvement in the amplitude of the MJO since the monthly forecasting system became operational, especially after the revision of the convection scheme in 2007 (cycle 32r3). Despite these improvements, the propagation of the MJO is too slow over the Indian Ocean in the last model cycles.
(from F Vitart)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 22
0 5 10 15 20 25 30 35 40 45Time Lag (days)
0
0.5
1
1.5
RM
S Er
ror
0 5 10 15 20 25 30 35 40 45Time Lag (days)
0
0.2
0.4
0.6
0.8
1
Cor
rela
tion
Bivariate Correlation Bivariate RMS error
1. EPS 32d (monthly) extension: MJO index
‘Perfect model’ skillEPS ens-mean
‘Perfect model’ skillEPS ens-meanModel climatology
To assess MJO skill, forecasts are projected into combined EOFs of U200, U850 and OLR (Wheeler & Hendon, 2004). The two dominant combined EOFs represent about 12% of the total variance each, and give a good representation of the MJO propagation. These plots show the skill of EPS forecasts of the MJO along these dominant EOFs.
(from F Vitart)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 23
Outline
1. ECMWF medium-range probabilistic forecast systems: the EPS
2. ECMWF seasonal probabilistic forecast systems:
a) The operational S3 and EUROSIP
b) The forthcoming S4 (Q4-2011)
3. Conclusions
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 24
2.a Seasonal prediction: S3 (2006-2011)
OASIS-2HOPE
~ 1.4 deg. lon1.4/0.3 d. lat.
TESSEL
IFS 31R11.1 deg. 62 levels
4-D variational d.a.
Multivar. O.I.
Gen. ofPerturb.
System-3CGCM
Initial Con. Ens. Forecasts
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 25
2.a S3 initial and model uncertainties
S3 simulates both initial and atmospheric model uncertainties:
Initial uncertainties: Ocean: initial uncertainties are simulated by using a 5-member ensemble ocean
assimilation system. Each analysis is created by using perturbed versions of the wind forcing.
SST: Prior to starting our coupled model forecasts, the ocean analyses are further perturbed by adding estimates of the uncertainty in the sea surface temperature to the ocean initial conditions.
Atmosphere: initial uncertainties are simulated using SVs
Model uncertainties: Atmosphere: model uncertainties are simulated by the SPPT scheme (as in the EPS)
System Ocean Coupling
Ocean SST Atmosphere
Ini unc Mod unc Ini unc Ini unc Mod unc
EPS >d10 -- -- yes EDA+SV SPPT(3TL)+SPBS
S3 d0 O-EDA -- yes SV SPPT(1TL)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 26
2.a S3 Ocean ICs: real-time analysis
The ocean analysis is performed over a 10 day window. All obs within a centered 10-days window are gathered and quality controlled. In the S3 HOPE-OI system, in addition to subsurface temperature, the scheme assimilates altimeter derived sea-level anomalies and salinity data.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 27
2.a S3 1-year ENSO outlook: 2009-2010
The tropics is the area where seasonal predictability is higher, controlled mainly by ENSO, a coupled ocean-atmosphere phenomenon centred over the tropical Pacific. S3 13m integrations are generated every quarter, based on an 11-member ensemble. They gave good predictions of the most recent El Nino events.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 28
2.a S3 2mT-anomaly fcs from 1 Dec: ave ROCA
These plots show a measure of the accuracy of seasonal probabilistic forecasts, the area under the Relative Operating Characteristics (ROCA) of the probabilistic forecasts of the 2mT anomaly being below the lower tercile forecasts with 1 Dec starting date. The ROCA has been evaluated considering S3 25-years (1985-2005) 11-member hindcasts.
The left panel shows the ROCA of the 2-4m average forecast, and the right panel the ROCA of the 5-7m average forecast.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 29
2.a S3 Accumulated Cyclone Energy (ACE) forecast
One of the operational seasonal forecasts is ACE (an index of storm activity defined by the sum of the square of the estimated maximum sustained velocity of every active tropical storm). The left panel shows the ACE forecast issued on 1 June 2010 for JASOND10.
The right panel shows that the accuracy of this product over WPAC (CC 52%).
(from F Vitart)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 30
2.a EUROSIP
EUROSIP is a multi-model seasonal forecasting system consisting of three independent coupled systems: ECMWF, Met Office and Météo-France (all integrated in a common framework). This table summarizes the configuration of the 3 seasonal systems that are currently used to generate EUROSIP operational products.
NCEP will be joining very soon. Germany and Canada has expressed interest to join as well.
Ocean Atmoshoriz resol
Atmosvert resol
ForecastMembers
Hindcast years
Hindcast Members
EC S3 HOPE (ORCA1Z42)
~ 120km (TL159) L62 < 0.5 41 (0÷7m)
5 (7÷13m)25
(‘81-‘05)11 (0÷7m)5 (7÷13m)
MF OPA (ORCA2Z42) ~ 300km L91 41 (0÷6m) 30
(‘81-‘10) 11 (0÷6m)
UK S5 NEMO (ORCA1Z75) ~ 120km (N96) L85 28 lagged
(0-6m)14
(‘89-’02) 12 (0÷6m)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 31
2.a S3 & EUROSIP 6m ENSO outlook: 1 Dec 08
EUROSIP is a multi-model seasonal forecasting system consisting of three independent coupled systems: ECMWF, Met Office and Météo-France (all integrated in a common framework). NCEP will be joining soon.
ECMWF (left) and EUROSIP 3-system SST anomaly forecasts for NINO3.4 area issued on 1 Dec 2008 and valid for 6 months. The observed SST anomaly lies at the edge of the ECMWF plum. Compared to the single ECMWF plum, the EUROSIP plum has a larger dispersion and gives a higher probability of cold conditions.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 32
2.a S3 & EUROSIP 6m ENSO outlook: 1 Oct 09
ECMWF (left) and EUROSIP 3-system SST anomaly forecasts for NINO3.4 area issued on 1 Oct 2009 and valid for 6 months.
In this case the observed SST anomaly lies outside the edge of the ECMWF plum. Compared to the single ECMWF plum, the EUROSIP plum has a larger dispersion and includes the observed SST within the forecast range.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 33
Pac: good distribution of temp anomalies Mid-lat Atl and Indian Ocean: colder than
fcs Austr: cold conditions were predicted
although the amplitudes underestimated Eur: cold temperatures were not predicted Greenland, West Canada and West Asia:
good predictions of warm/cold/warm anomalies
2mT analysis
(from L Ferranti)
2.a S3 & EUROSIP: 1 Nov fcs of 2mt for D10JF11
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 34
2.b The forthcoming S4 (Q4-2011): main features
New ocean model: NEMO v. 3.0 + 3.1 coupling interface, ORCA-1 configuration (~1-deg. resol., ~0.3 lat. near the equator) , 42 vertical levels, 20 levels with z < 300 m
3D variational ocean data assimilation (NEMOVAR) New physics package, including HTESSEL land-surface scheme, snow model (with EC-
Earth), new land surface initialization Prescribed sea-ice concentration with sampling from recent years
Ocean Atmos Horizontal resolution
Vertical resolution
Atmos mod uncertainty
ForecastMembers
Hindcast years
Hindcast Members
S3 HOPE 31r1 TL159 L62 < 0.5 1-lev SPPT 41 (0÷7m)5 (7÷13m)
25(‘81-‘05)
11 (0÷7m)5 (7÷13m)
S4 NEMO 36r4 TL255 L92 < 0.01 3-lev SPPTand SPBS
51 (0÷7m)15 (7÷13m)
30(‘81-‘10)
15 (0÷7m)15 (7÷13m)
System Ocean SST Sea ice Atmosphere
Ini unc Mod unc Ini unc Ini unc Mod unc
S3 O-EDA -- yes -- SV SPPT(1TL)
S4 O-EDA -- yes yes SV SPPT(3TL)+SPBS
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 35
2.b Operational coupled systems from Q4-2011
# fcs Hor resolution Vert lev Fc length Wave OceanHRES/DA 1 T1279 16 km 91 (<0.01 hPa, 75km) 0-10d WAM (0.25°, 30km) No
EPS 51T639 32 km
62 (<5 hPa, 35km)0-10d
WAM (0.5°, 60km)No
T319 64 km 10-15/32d NEMO 0.3-1.0 deg z42S4 51 T255 100km 91(<0.01 hPa, 75km) 0-13m WAM (1.0°, 120km) NEMO 0.3-1.0 deg z42
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 36
2.b The S4 hindcasts
Seasonal products are bias corrected and calibrated using 450 S4 forecasts run for the past years:
30 years (1981 – 2010) 15 members (up to 7m,
extended to 13m every quarter)
Same model version Same resolution ICs from ERA-I and
ORAS4 Same initial perts 30y
51*T255L91
2011
J F M A M …..
51*T255L91
51*T255L91
15*T255L91
15*T255L91
15*T255L91
15*T255L91
2010
15*T255L91
15*T255L91
15*T255L91
1981
….
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 37
2.b S4 ocean re-analysis (ORAS4)
Based on NEMO/NEMOVAR Model configuration: ORCA1, smooth coastlines, closed Caspian Sea Forced by ERA40 (until 1989) + ERA Interim (after 1989) Assimilates Temperature/Salinity from EN3 and altimeter data
Strong relaxation to SST (OI_v2) Online bias correction scheme
This plot shows the time evolution of the ocean heat content in ORAS4 (anomalies respect the period 1960-2010) for 3 different depth ranges. Around 2002, the warming stabilizes in the upper ocean, but continues in the deeper ocean. The implication is that the ocean may be absorbing more heat than anticipated.
(from M Balmaseda)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 38
S4 experiments (20years, 11-member ensembles) indicate that ocean data assimilation with NEMOVAR improves the SST forecast skill at different lead times and different regions.
2.b Impact of ocean DA on SST forecasts
(from M Balmaseda)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 39
S4 SST forecasts are re-scaled to correct for the too large variability, with re-scaling coefficients computed in cross-validation mode. SST forecast amplitudes are substantially too large in S4, and the re-scaling is important to give reasonable RMS error scores (and, more importantly, accurate forecast plumes).
2.b S4 re-scaled SST plume anomalies
(from T Stockdale)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 40
Preliminary results based on 18y 15m hindcasts indicate that S4 has more realistic spread. In terms of skill, over the EC-Pac the re-scaled S4 ensemble-mean forecasts are better (see plot for NINO3.4), but over West-Pac are comparable.Over the Atlantic, re-scaled S4 ensemble-mean forecasts are more skilful.
2.b S4 and S3 NINO3.4 and EQATL SST fcs
0 1 2 3 4 5 6 7Forecast time (months)
0.4
0.5
0.6
0.7
0.8
0.9
1
Ano
mal
y co
rrel
atio
n
wrt NCEP adjusted OIv2 1971-2000 climatology
NINO3.4 SST anomaly correlation
0 1 2 3 4 5 6 7Forecast time (months)
0
0.2
0.4
0.6
0.8
1
Rm
s er
ror
(deg
C)
95% confidence interval for 0001, for given set of start datesEnsemble sizes/corrections are 15/AS (0001) and 11/BC (0001)191 start dates from 19910201 to 20081201, various corrections
NINO3.4 SST rms errors
Fcast S4 Fcast S3 Persistence Ensemble sd
MAGICS 6.12 nautilus - net Wed May 18 10:37:36 2011
0 1 2 3 4 5 6 7Forecast time (months)
0.4
0.5
0.6
0.7
0.8
0.9
1
Ano
mal
y co
rrel
atio
n
wrt NCEP adjusted OIv2 1971-2000 climatology
EQATL SST anomaly correlation
0 1 2 3 4 5 6 7Forecast time (months)
0
0.2
0.4
0.6
Rm
s er
ror
(deg
C)
95% confidence interval for 0001, for given set of start datesEnsemble sizes/corrections are 15/AS (0001) and 11/BC (0001)191 start dates from 19910201 to 20081201, various corrections
EQATL SST rms errors
Fcast S4 Fcast S3 Persistence Ensemble sd
MAGICS 6.12 nautilus - net Wed May 18 10:37:36 2011
(from T Stockdale)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 41
2.b S4 and S3: comparison based on 18y hindcasts
Preliminary results based on hindcasts run for 18 years indicate that the S4 model climate is better (see plot that shows U925 model biases of S3 and S4)
S4 atmospheric scores are less well sampled at this stage, but are in the positive to neutral range
S4 forecasts are not better over the West Pacific mainly because of too strong easterlies in the equatorial Pacific
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
135°W
135°W 90°W
90°W 45°W
45°W 0°
0° 45°E
45°E 90°E
90°E 135°E
135°E
Global rms error: 0.876 NH:0.69 TR:0.993 SH:0.786925hPa zonal wind S4(15)-ERA Int 1991-2008 DJF
[m/s]
-4
-3
-2
-1
-0.5
0.5
1
2
3
4
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
135°W
135°W 90°W
90°W 45°W
45°W 0°
0° 45°E
45°E 90°E
90°E 135°E
135°E
Global rms error: 1.28 NH:1.32 TR:1.16 SH:1.47925hPa zonal wind S3(11)-ERA Int 1991-2008 DJF
[m/s]
-4
-3
-2
-1
-0.5
0.5
1
2
3
4
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
135°W
135°W 90°W
90°W 45°W
45°W 0°
0° 45°E
45°E 90°E
90°E 135°E
135°E
mean -0.0751 rms diff: 0.918U925 S4(15)-S3(11) 1991-2008 DJF
[m/s]
-2
-1.5
-1
-0.5
-0.25
0.25
0.5
1
1.5
2
(from T Stockdale)
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 42
Outline
1. ECMWF medium-range probabilistic forecast systems: the EPS
2. ECMWF seasonal probabilistic forecast systems:
a) The operational S3 and EUROSIP
b) The forthcoming S4 (Q4-2011)
3. Conclusions
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 43
3. Conclusions
Ensemble-based systems are used at ECMWF to estimate not only the most-likely scenario but also its uncertainty:
In the analysis with the Ensemble Data Assimilation system (EDA, since Jun 2010) In the medium-range (<1m) with the Ensemble Prediction System (EPS, since 1992) In the seasonal range (<13m) with seasonal System-3 (S3, since 1998)
The EPS uses a combination of EDA- and SV-based perturbations to simulate initial uncertainties, and stochastic schemes to simulate model uncertainties. It is based on 51 forecasts run with a 639v319L62 atmosphere model. It is coupled to the HOPE (1.0-0.3 deg) ocean model from day 10, and is run twice a day. Forecasts starting at 00UTC on Mon & Thu are extended to 32days. In Q3-2011 HOPE will be replaced by NEMO (with a 1.0-0.3 deg resolution, with initial conditions constructed with NEMOVAR).
The S3 uses a 159L62 old (31r1) version of the atmosphere model and the HOPE ocean model. In Q4-2011 the new S4, based on a more recent and higher-resolution 255L91 atmosphere model, and the NEMO ocean model will be introduced.
NCEP (6 Jun 2011) - Roberto Buizza: Probabilistic prediction at ECMWF 44
Conclusions
Work will continue to further improve the ECMWF probabilistic forecasting systems in different areas:
1. Increase the vertical resolution – ECMWF is planning to increase the vertical resolution of the HRES/DA, EDA and EPS (~L95) systems (2012). Work to assess the impact on the EPS will start soon.
2. Improve the model error simulation schemes – Work is progressing to assess the impact of using the SPBS scheme in the EDA, and to investigate potential upgrades of the stochastic schemes in the EDA and EPS.
3. Further exploit the link between the EDA and the EPS – The use of initial-time land-surface and SST perturbations generated using the EDA is under investigation. Work to assess the potential value of using 25 or 50 EDA members will start soon.
4. Include a dynamical sea-ice model and a mixed-layer model – Work has started to assess whether using the LIM2/LIM3 dynamical sea-ice model and a mixed-layer model can lead to further improvements.
5. Increase the resolution of the ocean model and test coupling from d0 – Work will start soon to assess the impact of using a higher-resolution ocean model (1/4 degree NEMO) in the probabilistic forecasting systems. For the EPS, the possibility to couple from initial time will also be investigated.