multi-model short-range ensemble forecasting at spanish met agency (aemet) j. a. garcía-moya, a....
TRANSCRIPT
Multi-model Short-Range Ensemble Forecasting at
Spanish Met Agency (AEMET)
J. A. García-Moya, A. Callado, P. Escriba, C. Santos, D. Santos, J. Simarro
Spanish Met Service AEMET
Training Workshop onNOWCASTING TECHNIQUESBuenos Aires, August 2013
15 August 2013 T-Note Workshop 2
OutlineIntroductionEPS for short-range forecastSREPS system at AEMETVerification exercise
Against analysisAgainst synoptic observationsAgainst climate network observations
Probabilistic scoresSynoptic variables10 m surface wind speed6h accumulated precipitation24h accumulated precipitation
Time-Lagged Super-ensembleComparison with ECMWF EPSMulti-model predictabilityConclusions
15 August 2013 T-Note Workshop 3
Meteorological Framework
Main Weather Forecast issues are related with Short-Range forecast of extreme events.Convective precipitation is the most dangerous weather event in the Mediterranean.
15 August 2013 T-Note Workshop 4
Geographical Framework
Western Mediterranean is a close sea rounded by high mountains.In autumn sea is warmer than air.Several cases of more than 200 mm/few hours occurs every year.Some fast cyclogenesis like “tropical cyclones” also appears from time to time (“medicanes”).
15 August 2013 T-Note Workshop 55
Geographical Framework
15 August 2013 T-Note Workshop 66
15 August 2013 T-Note Workshop 77
15 August 2013 T-Note Workshop 8
Ensemble for Short Range
Surface parameters are the most important ones for weather forecast.Forecast of extreme events (convective precip, gales,…) is probabilistic.Short Range Ensemble prediction can help to forecast these events.Forecast risk (Palmer, ECMWF Seminar 2002) is the goal for both Medium- and, also, Short-Range Prediction.
15 August 2013 T-Note Workshop 9
Errors of short-range forecast
Due to model formulation.Due to simplifications in parameterisation schemes.Due to uncertainty in the initial state.Special for LAMs, due to errors in lateral boundary conditions.Due to uncertainties in soil fields (soil temperature and soil water content, …).
15 August 2013 T-Note Workshop 10
SREPS I
Multi-model approach (Hou & Kalnay 2001).Stochastic physics (Buizza et al. 1999).Multi-boundaries:
From few global deterministic models.From global model EPS (ECMWF).SLAF technique (Ebisuzaki & Kalnay 1991).
15 August 2013 T-Note Workshop 11
SREPS II
Different assimilation techniques:Optimal Interpolation.Variational (3D or 4D).
Perturbed analysis:Singular vectors (ECMWF, Palmer et al. 1997).Breeding (NCEP, Toth & Kalnay 1997).Scaled Lagged Average Forecasting (SLAF, Ebisuzaki & Kalnay 1991).EnKF, ETKF, LETKF, PO
15 August 2013 T-Note Workshop 12
Multi-model
Hirlam (http://hirlam.org).
HRM from DWD (German Weather Service).
MM5 (http://box.mmm.ucar.edu/mm5/).
UM from UKMO (UK Weather Service).
LM (COSMO Model) from COSMO consortium (http://www.cosmo-model.org).
WRF (NOAA – NCEP) – work in progress
15 August 2013 T-Note Workshop 13
Multi-Boundaries
From different global deterministic models:ECMWFUM from UKMO (UK Weather Service)
GFS from NCEPGME from DWD (German Weather Service)
CMC from SMC (Canadian Weather Service)
15 August 2013 T-Note Workshop 14
SREPS at AEMET
Mummub: Multi-model Multi-boundaries72 hours forecast two times a day (00 & 12 UTC).Characteristics:
5 models.5 boundary conditions.2 latest ensembles (HH & HH-12).
25 member ensemble every 12 hoursTime-lagged Super-Ensemble of 50 members every 12 hours.
15 August 2013 T-Note Workshop 15
15 August 2013 T-Note Workshop 16
Post-processing
Integration areas 0.25 latxlon, 40 levelsInterpolation to a common area
~ North Atlantic + EuropeGrid 380x184, 0.25º
SoftwareEnhanced PC + LinuxECMWF Metview + Local developments
OutputsDeterministicEnsemble probabilistic
15 August 2013 T-Note Workshop 17
Monitoring in real time
Intranet web serverDeterministic outputs
Models X BCs tables Maps for each couple (model,BCs)
Ensemble probabilistic outputsProbability maps: 6h accumulated precipitation, 10m wind speed, 24h 2m temperature trendEnsemble mean & Spread mapsEPSgrams (work in progress)
Verification: Deterministic & ProbabilisticAgainst ECMWF analysisAgainst observations
15 August 2013 T-Note Workshop 18
Monit: all models X bcs
Whole Area
Zoom over Spain
15 August 2013 T-Note Workshop 19
Monit: Spread – Ensemble mean maps
Spread at key
mesoscale areas
15 August 2013 T-Note Workshop 20
Case Study 06/10/2006 at 00 UTC
More than 15 mm/6 hours
15 August 2013 T-Note Workshop 21
Verification
Verification exercise, April-June 2006:
Calibration: with synoptic variables Z500, T500, PmslResponse to binary events: reliability and resolution of surface variables 10m surface wind, 6h and 24h accumulated precipitation
15 August 2013 T-Note Workshop 22
Interpolation
15 August 2013 T-Note Workshop 23
Obs verification - Probabilistic scoresEnsemble calibration:
Synoptic variables:Z500, T500, Pmsl
Scores:Rank histogramsSpread-skill
Response to binary events:Surface variables:
10m surface wind (10,15,20m/s thresholds)6h accumulated precipitation (1,5,10,20mm thresholds)24h accumulated precipitation (1,5,10,20mm thresholds)
Scores:Reliability, sharpness (H+24, H+48)ROC, Relative Value (H+24, H+48)BSS, ROCA with forecast length
15 August 2013 T-Note Workshop 24
Summary of Probabilistic Verification
15 August 2013 T-Note Workshop 25
Synoptic parameters
Using ECMWF Analysis as reference:MSLP
Over all FC lengths H+00 .. H+72:Spread-skill
H+72:Rank histograms
Using Synoptic observations as reference:MSLP
Over all FC lengths H+00 .. H+72:Spread-skill
H+72:Rank histograms
15 August 2013 T-Note Workshop 26
MSLP-ECMWF
15 August 2013 T-Note Workshop 27
MSLP - Obs
15 August 2013 T-Note Workshop 28
Binary events
Binary events X = {0,1} at every pointAccumulated precipitation in 24 hours >= 5mmUseful to decompose the forecast in thresholdsPerformance computed using contingency tables (CT’s)
15 August 2013 T-Note Workshop 29
Contingency tables It is the best way to characterize a binary event
fc(X)={1,0}ob(X)={1,0}
ob
1 0
fc1 a b a+b
0 c d a+d
a+c
b+d
a+b+c+d = N
15 August 2013 T-Note Workshop 30
Contingency tables: scores
Several scores can computed from CT’s
Base Rate s ( a + c ) / n
HitRate HIR a / ( a + c )
False Alarm Rate
FAR b / ( b + d )
False Alarm Ratio
FARatio
b / ( a + b )
Proportion Correct PC ( a + d ) / n
…
ob
1 0
fc1 a b a+b
0 c d a+d
a+c
b+d
a+b+c+d = N
15 August 2013 T-Note Workshop 31
Example: Every pdf threshold has its own CT
ob
1 0
P(X) >= 0/n1 a0 b0
0 c0 d0
P(X) >= 1/n1 a1 b1
0 c1 d1
P(X) >= 2/n1 a2 b2
0 c2 d2
…P(X) >= n/n
1 an bn
0 cn dn
HIR0 FAR0
HIR1 FAR1
HIR2 FAR2
HIRn FARn
15 August 2013 T-Note Workshop 32
ReliabilityObservation frequency conditioned to forecast probabilityReliability diagrams
( i/N , Δai/(Δai+Δbi) ) i = 0…NNear diagonal ~ reliable
Sharpness histograms( i/N , Δai+Δbi ) i = 0…NU shape ~ sharp (discriminating binary event)
15 August 2013 T-Note Workshop 33
Precipitation
Using ECMWF Deterministic Model as reference:
6 hours accumulation – 24 hours forecast length24 hours accumulation – 54 hours forecast lengthThresholds 1, 5, 10 y 20 mm
Using Synoptic observations as reference:6 hours accumulation – 24 hours forecast length24 hours accumulation – 54 hours forecast lengthThresholds 1, 5, 10 y 20 mm
15 August 2013 T-Note Workshop 34
ECMWF
Reliab. - 24 h Acc. Precip H+54 (1,5,10,20) mm
Reliab. - 6 h Acc. Precip H+24 (1,5,10,20) mm
15 August 2013 T-Note Workshop 35
ObservationsReliab. - 6 h Acc. Precip H+24 (1,5,10,20) mm
Reliab. - 24 h Acc. Precip H+54 (1,5,10,20) mm
15 August 2013 T-Note Workshop 36
Resolution
Based on Signal Detection TheoryROC (Relative Operating Characteristics)
( FARi , HIRi ) i = 0…N
ROCArea ~ Resolution or binary event discrimination
Forecast conditioned by observation
15 August 2013 T-Note Workshop 37
ROC curves – 24 h Acc Precip (1, 5, 10 & 20 mm)
ECMWF Observations
15 August 2013 T-Note Workshop 38
ECMWF - AnalysisR
OC
10 m wind H+24 (10,15,20) m/s
Reliab
ilit
y
15 August 2013 T-Note Workshop 39
Observations10 m wind H+24 (10,15,20) m/s
Reliab
ilit
yR
OC
15 August 2013 T-Note Workshop 40
Joint
15 August 2013 T-Note Workshop 41
Reliability & Sharpness
Good reliability according tothresholds (base rate)forecast length
NoUnder-samplin
g
15 August 2013 T-Note Workshop 42
Resolution
Good resolutionROC AreasBSSs
Good RV curves0
1
0.5
1
15 August 2013 T-Note Workshop 43
Time-Lagged Super-ensemble
How much predictability can be added by a time-lagged super-ensemble?40 members super-ensemble (SE-SREPS) with the last two runs of SREPS ( HH & HH-12).Verifications against observationsCheap in terms of computer resourcesJust a different post-process
15 August 2013 T-Note Workshop 44
Spread-skill
15 August 2013 T-Note Workshop 45
Rank histogram
15 August 2013 T-Note Workshop 46
Reliability diagrams
15 August 2013 T-Note Workshop 47
Comparison with ECMWF EPS
Period: April to June 2006European Synop obs: H+72.
Mslp / v 10m / Precipitation
European climate precipitation network: H+54 (longest SREPS period matching observations).
24 hours accumulated precipitation (from early morning to early morning).
15 August 2013 T-Note Workshop 48
Climate Obs – 24 h. Precip (1, 5, 10 & 20 mm)
15 August 2013 T-Note Workshop 49
Climate Obs – 24 h. Precip (1, 5, 10 & 20 mm)
15 August 2013 T-Note Workshop 50
Climate Obs – 24 h. Precip (1, 5, 10 & 20 mm)
Brier Skill Score (BSS) Decomposition – 106 realizations
Precip (mm)
Climatolog. Frequency
Uncertainty
MumMub ECMWF 21 ECWMF 51
1 0.24 0.18 0.35 0.28 0.28
5 0.11 0.09 0.27 0.22 0.22
10 0.05 0.05 0.18 0.18 0.19
20 0.01 0.01 0.01 0.07 0.08
Precip (mm)
MumMub ECMWF 21 ECWMF 51
1 0.06 0.13 0.14
5 0.03 0.08 0.08
10 0.03 0.03 0.03
20 0.07 0.02 0.02Precip (mm)
MumMub ECMWF 21 ECWMF 51
1 0.59 0.59 0.59
5 0.70 0.70 0.70
10 0.79 0.79 0.78
20 0.91 0.91 0.90
15 August 2013 T-Note Workshop 5151
24 hours accumulated precipitation from European climate network upscaled to 0.25 deg. Resolution - BSS
15 August 2013 T-Note Workshop 52
Why Multi-model?
Better representation of model errors (SAMEX, Hou & Kalnay 2001, and DEMETER).Consistent set of perturbations of initial state and boundaries.Better results than any single model ensemble (SAMEX, DEMETER, Arribas et al., 2005).
15 August 2013 T-Note Workshop 53
Single model Ensembles (4 members each)
15 August 2013 T-Note Workshop 54
BMA is a statistical method for postprocessing ensembles based
on standard method for combining predictive distributions from
different sources.
The BMA predictive PDF of any quantity of interest is a weighted
average of PDFs centred on the individual bias-corrected
forecasts, where the weights are equal to posterior probabilities
of the models generating the forecasts and reflect the models’
relative contributions to predictive skill over the training period.
Raftery et al, 2005
BMA INTRO
15 August 2013 T-Note Workshop 55
BMA FUNDAMENTALS
F2
F1
F4
F5
F3
BMA predictive PDF
15 August 2013 T-Note Workshop 56
BMA calibration using TEMP and SYNOP observations over whole area :
500 hPa Temperature (T500)500 hPa Geopotencial (Z500) 10m Wind speed (S10m)
3, 5 and 10 days of training period
3 months of calibration (April, May and June of 2006) for Z500 and T500 and 1 month for S10m (April 2006)
24, 48 and 72 hours forecast for Z500 and T500 and
24 and 72 hours forecast for S10m
BMA EXPERIMENTS
15 August 2013 T-Note Workshop 57
RESULTS
MULTIMODEL
BMA 3 T. DAYS
BMA 5 T. DAYS
BMA 10 T. DAYS
15 August 2013 T-Note Workshop 58
RESULTS
MULTIMODEL
BMA 3 T. DAYS
BMA 5 T. DAYS
BMA 10 T. DAYS
15 August 2013 T-Note Workshop 59
RESULTS
MULTIMODEL
BMA 3 T. DAYS
BMA 5 T. DAYS
BMA 10 T. DAYS
15 August 2013 T-Note Workshop 60
HRM T500 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T HAV
HEC
HGM
WEIGHT TIME SERIES
HIRLAM T500 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATEW
EIG
HT
IAV
IEC
IGM
IUK
LM T500 td 3 H+72
0,000
0,100
0,200
0,300
0,400
0,500
0,600
0,700
0,800
0,900
1,000
DATES
WE
IGH
TS LAV
LEC
LGM
LUK
MM5 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
MAV
MEC
MGM
MUK
UM T500 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
UAV
UEC
UGM
UUK
HRM HIRLAM LOKAL MODEL
MM5 UM
15 August 2013 T-Note Workshop 61
HIRLAM td 3 H+48
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
HIRLAM T500 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
WEIGHT TIME SERIES
HIRLAM td 3 H+24
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
HIRLAM
15 August 2013 T-Note Workshop 62
HIRLAM T500 td 3 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
WEIGHT TIME SERIES
HIRLAM td 5 H+72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
HIRLAM td 10 h+ 72
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
HIRLAM
15 August 2013 T-Note Workshop 63
WEIGHT TIME SERIES
HIRLAM td 3 H+24
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WE
IGH
T
IAV
IEC
IGM
IUK
Z500 td 3 H + 24
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
DATE
WEI
GH
T
IAV
IEC
IGM
IUK
HIRLAM
15 August 2013 T-Note Workshop 64
Summary I
A Multi-model-Multi-boundaries Short Range Ensemble Prediction System (MMSREPS), has been developed at the AEMET-SpainWe show here 3months verification results (April-June 21006), against both synoptic and climate observations and ECMWF analysis and model:
Response to binary events: reliability and resolution of surface variables 10m surface wind, 6h and 24h accumulated precipitation
These results look promising:Verification against ECMWF analysis and model shows very good resultsVerification against observations shows quite good results
Ensemble is a bit under-dispersive Good response to binary events (synoptic and climate observations)
15 August 2013 T-Note Workshop 65
Summary II
A Time-Lagged Super-ensemble with 40 members had shown better performance than the Multi-model SREPS alone.Multi-model technique gives much better spread than any other single-model techniqueBias correction and calibration through a Bayesian Model Averaging scheme is under development (first results show better performance).