Seasonal prediction of boreal winter stratosphereA. Portal (UNIMIB)+ , P. Ruggieri (UNIBO), F. Palmeiro, J. Garcia-Serrano (UB), D. Domeisen (ETH Zürich), S. Gualdi (CMCC)
+ Contact information: Alice Portal, University of Milan-Bicocca (DISAT), [email protected]
Motivations:I. Measuring the winter predictability of the stratosphere in 5 Copernicus seasonal prediction
models initialised in November1.II. Assessing model reproduction of the coupling between lower-stratosphere wave forcing
(LSWF) and stratosphere polar vortex (SPV). See schematics of Figure 1.III. Determining the impact of local-LSWF forecast skill on SPV skill.
Coupling between LSWF and SPV, diagnostics:
LSWF ←→ 〈v ′T ′〉100, 100-hPa Eliassen–Palm flux(vertical component)average in 40-80◦ N
SPV ←→ U10[55-70N], 10-hPa zonal-mean zonal windaverage in 55-70◦ N
where ′ : deviation from zonal mean· : zonal mean〈·〉 : meridional mean in mid latitudes
SPV
tropopauseLSWF 100 hPa
10 hPa
stratopause
60 N 90 NEq.
Figure 1: Waves propagating upwards from thetroposphere (LSWF) break in the SPV anddecelerate it. Latitude–pressure perspective.
1See description of the models in “Additional material 1”Alice Portal Seasonal stratosphere prediction 1 / 10
Seasonal predictability of the boreal stratosphere
0 20 40 60 80latitude
10
0
10
20
30
40m
/sa) Climatology of DJF U10
0 20 40 60 80latitude
0
5
10
15
20
m/s
b) Ens-mean standard deviation (-)mean ensemble spread (--)
ERA-I interannualstandard deviation
0 20 40 60 80latitude
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0c) Anomaly correlation coefficient
15-30 Nov persist fc
1-15 Dic persist fc
ERA-I CMCC DWD MF ECMWF UKMO
SPV
QBO-inducedvariability SPV
variability
lowintermodelspread
highintermodelspread
Figure 2: Variability and predictability of December-January-February (DJF) stratosphere, function of latitude. DJF averages of zonal-mean zonal windat 10 hPa are used to compute: a) 1993/94-to-2016/17 climatology; b) ensemble-mean or interannual standard deviation (solid), mean spread of themodel ensemble (dashed); c) correlation coefficient between ensemble-mean and ERA-Interim anomalies.
Seasonal variability and predictive skill are affected by tropical QBO and extratropical SPV.The tropics are well predicted due to the longer QBO timescale compared to the seasonal scale.The extratropics are less predictable due to low signal-to-noise ratio (ens. mean std/mean ens. spread,Figure 2b). Observe the wide inter-model spread in the extratropical predictive skill. The skill does notdepend on SPV model bias (cf. panels a and c in Figure 2).Low predictive skill in the subtropics may affect the propagation of the tropical QBO in the extratropics(Holton–Tan mechanism).
Stratospheric polar vortex:
1994/95 1998/99 2002/03 2006/07 2010/11 2014/15
2
1
0
1
2
U10
[55-
70N]
(std
dv)
CMCC[0.48]ECMWF[0.12]DWD[0.37]
MF[0.07]UKMO[0.36]MMM[0.40]
CMCC,UKMO,DWD[0.47]ERA-I
Figure 3: Ensemble-mean/interannual variability of the DJF SPV, seasonal model ACC in label.
N D J F M A0.2
0.0
0.2
0.4
0.6
0.8
1.0 ACC monthly U10[55-70N]persist fc 15-30 Novpersist fc 1-15 Dec
Figure 4: Monthly ACC and its standard deviation(shading, from bootstrap). Full dots evidencesignificant skill (p≤0.05).
Alice Portal Seasonal stratosphere prediction 2 / 10
Seasonal predictability of the boreal stratosphere
0 20 40 60 80latitude
10
0
10
20
30
40m
/sa) Climatology of DJF U10
0 20 40 60 80latitude
0
5
10
15
20
m/s
b) Ens-mean standard deviation (-)mean ensemble spread (--)
ERA-I interannualstandard deviation
0 20 40 60 80latitude
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0c) Anomaly correlation coefficient
15-30 Nov persist fc
1-15 Dic persist fc
ERA-I CMCC DWD MF ECMWF UKMO
SPV
QBO-inducedvariability SPV
variability
lowintermodelspread
highintermodelspread
Figure 2: Variability and predictability of December-January-February (DJF) stratosphere, function of latitude. DJF averages of zonal-mean zonal windat 10 hPa are used to compute: a) 1993/94-to-2016/17 climatology; b) ensemble-mean or interannual standard deviation (solid), mean spread of themodel ensemble (dashed); c) correlation coefficient between ensemble-mean and ERA-Interim anomalies.
Stratospheric polar vortex:
1994/95 1998/99 2002/03 2006/07 2010/11 2014/15
2
1
0
1
2
U10
[55-
70N]
(std
dv)
CMCC[0.48]ECMWF[0.12]DWD[0.37]
MF[0.07]UKMO[0.36]MMM[0.40]
CMCC,UKMO,DWD[0.47]ERA-I
Figure 3: Ensemble-mean/interannual variability of the DJF SPV, seasonal model ACC in label.
Three models (CMCC, DWD,UKMO) show significant winterpredictive skill for the SPV.We look at monthly skill values...
Stratospheric polar vortex:
1994/95 1998/99 2002/03 2006/07 2010/11 2014/15
2
1
0
1
2
U10
[55-
70N]
(std
dv)
CMCC[0.48]ECMWF[0.12]DWD[0.37]
MF[0.07]UKMO[0.36]MMM[0.40]
CMCC,UKMO,DWD[0.47]ERA-I
Figure 4: Ensemble-mean/interannual variability of the DJF SPV, seasonal model ACC in label.
N D J F M A0.2
0.0
0.2
0.4
0.6
0.8
1.0 ACC monthly U10[55-70N]persist fc 15-30 Novpersist fc 1-15 Dec
Figure 5: Monthly ACC and its standard deviation(shading, from bootstrap). Full dots evidencesignificant skill (p≤0.05).
Alice Portal Seasonal stratosphere prediction 2 / 10
Seasonal predictability of the boreal stratosphere
0 20 40 60 80latitude
10
0
10
20
30
40m
/sa) Climatology of DJF U10
0 20 40 60 80latitude
0
5
10
15
20
m/s
b) Ens-mean standard deviation (-)mean ensemble spread (--)
ERA-I interannualstandard deviation
0 20 40 60 80latitude
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0c) Anomaly correlation coefficient
15-30 Nov persist fc
1-15 Dic persist fc
ERA-I CMCC DWD MF ECMWF UKMO
SPV
QBO-inducedvariability SPV
variability
lowintermodelspread
highintermodelspread
Figure 2: Variability and predictability of December-January-February (DJF) stratosphere, function of latitude. DJF averages of zonal-mean zonal windat 10 hPa are used to compute: a) 1993/94-to-2016/17 climatology; b) ensemble-mean or interannual standard deviation (solid), mean spread of themodel ensemble (dashed); c) correlation coefficient between ensemble-mean and ERA-Interim anomalies.
Stratospheric polar vortex:
1994/95 1998/99 2002/03 2006/07 2010/11 2014/15
2
1
0
1
2
U10
[55-
70N]
(std
dv)
CMCC[0.48]ECMWF[0.12]DWD[0.37]
MF[0.07]UKMO[0.36]MMM[0.40]
CMCC,UKMO,DWD[0.47]ERA-I
Figure 3: Ensemble-mean/interannual variability of the DJF SPV, seasonal model ACC in label.
N D J F M A0.2
0.0
0.2
0.4
0.6
0.8
1.0 ACC monthly U10[55-70N]persist fc 15-30 Novpersist fc 1-15 Dec
Figure 4: Monthly ACC and its standard deviation(shading, from bootstrap). Full dots evidencesignificant skill (p≤0.05).
Alice Portal Seasonal stratosphere prediction 2 / 10
Relation between SPV and LSWF
We define a reconstruction of the SPV-wind anomaly (∆U10) on the basis of pre-vious LSWF anomalies 2
∆U∗10(t) ≡ −B∫ t
t0∆〈v ′T ′〉100(tf ) e−(t−tf )/τ dtf ,︸ ︷︷ ︸
integral of time-weighed LSWF anomalies
(1)
with τ = 45 days, B constant.
Following, we assess the relation between SPV and LSWF on daily and seasonal timescale: the effective SPV-wind anomaly, ∆U10, is displayed against the correspondentanomaly reconstructed from LSWF, ∆U∗10. Slope and correlation values, computedfor each scatter plot, facilitate the comparison between model LWSF–SPV couplingand ERA-Interim coupling.
2derived from “Relation between the 100-hPa Heat Flux and Stratospheric PotentialVorticity”, Hinssen and Ambaum (2010)
Alice Portal Seasonal stratosphere prediction 3 / 10
Relation between SPV and LSWFReconstruction of the SPV-wind anomaly (∆U10) from previous LSWF anomalies 3
∆U∗10(t) ≡ −B∫ t
t0
∆〈v ′T ′〉100(tf ) e−(t−tf )/τ dtf
Figure 5: Daily assessment. Density of the scatter plot between daily values of SPV windanomaly, ∆U10, and of its reconstruction from LSWF anomalies, ∆U∗10.
60
40
20
0
20
40
U10
(m/s
)
r=0.86b=0.171 K 1
ERA-I
r=0.83b=0.149 K 1
CMCC
r=0.83b=0.166 K 1
ECMWF
200 0 200U *
10 (m/s K)60
40
20
0
20
40
U10
(m/s
)
r=0.81b=0.142 K 1
DWD
200 0 200U *
10 (m/s K)
r=0.75b=0.142 K 1
MF
200 0 200U *
10 (m/s K)
r=0.84b=0.146 K 1
UKMO0
5
10
15
20
25
0
200
400
600
800
0
100
200
300
400
0
100
200
300
400
0
50
100
150
200
250
0
50
100
150
200
250
300
350
Discrepancies from ERA-Interimdistribution are found mainly inCMCC (too narrow) and MF (toowide).In MF, LSWF is not as determinantfor the SPV as is found in reanalysis.This is evidenced by reduced modelcorrelation (r).All models display weaker couplingbetween LSWF and SPV compared toreanalysis (see values of b and r).
3derived from “Relation between the 100-hPa Heat Flux and Stratospheric Potential Vorticity”, Hinssen and Ambaum (2010)
Alice Portal Seasonal stratosphere prediction 4 / 10
Relation between SPV and LSWFReconstruction of the SPV-wind anomaly (∆U10) from previous LSWF anomalies 4
∆U∗10(t) ≡ −B∫ t
t0
∆〈v ′T ′〉100(tf ) e−(t−tf )/τ dtf
Models with higher skill in predicting the DJF SPV alsoshow stronger interannual variability of LSWF (∆U∗10).One model (DWD) reproduces a weaker link betweeninterannual variability of LSWF and of SPV, as shown bymodel rDJF.ERA-Interim seasonal coupling between LSWF and SPV isgenerally stronger than in models (see values of bDJF andrDJF).
Figure 6: Seasonal assessment. Scatter plot between DJF(ensemble-mean) values of SPV wind anomaly, ∆U10, and ofits reconstruction from LSWF anomalies, ∆U∗10.
75 50 25 0 25 50 75 100
10
0
10
U10
,DJF
(m/s
)
rDJF=0.91, bDJF=0.18 K 1
ERA-I
40 20 0 20 40
5
0
5
U10
,DJF
(m/s
)
rDJF=0.91, bDJF=0.15 K 1
rDJF=0.74, bDJF=0.11 K 1
rDJF=0.89, bDJF=0.16 K 1
Models w DJF skill 'significant'CMCCDWDUKMO
40 20 0 20 40U *
10, DJF (m/s K)
5
0
5
U10
,DJF
(m/s
)rDJF=0.84, bDJF=0.19 K 1
rDJF=0.82, bDJF=0.14 K 1
Models w DJF skill 'not significant'ECMWFMF
4derived from “Relation between the 100-hPa Heat Flux and Stratospheric Potential Vorticity”, Hinssen and Ambaum (2010)
Alice Portal Seasonal stratosphere prediction 5 / 10
Predictability of local LSWF (v ′T ′), links with ENSO
300
300
300
ERA-I
60
60
60
CMCC
60
MF
60
60
60
180
ECMWF
60
6060
60
180
180
DWD
60
60
60
6060
UKMO
90
60
30
0
30
60
90v'
T' (m
/s K
)
252015105
0510152025
Cova
rianc
e (m
/s K
)
November
300
300
ERA-I30
30
30
CMCC
30
30
MF
30
90 90
ECMWF
30
30
30
90
DWD
3030
30
3090
UKMO
90
60
30
0
30
60
90
v'T'
(m/s
K)
5
0
5
Cova
rianc
e (m
/s K
)
December-January-February
Figure 7: Left: ERA-Interim climatologies of v′T ′ (green–purple). Right: Covariance between model ensemble mean v′T ′ and standardised ERA-Interimv′T ′ (blue–red, note different scaling in November and DJF). Stippling shows significant ACC (p≤0.05). Black contours show model ensemble-meanvariance at 60, 120, 180, 300, 600 (m/s K)2 in November; at 30, 60, 90, 180 (m/s K)2 in DJF; reanalysis interannual variance at 300, 600, 900 (m/s K)2.
The predictive skill for November is good in all models exept MF.Residual DJF model skill is found mainly in the Pacific sector. It is lower and model-dependent overEurasia.
We inquire the local-LSWF response to ENSO in models and in reanalysis, in order to understand whetherseasonal Pacific predictability is related to ENSO.The impact of other potential sources of seasonal stratospheric predictability (QBO, Arctic sea-ice extent andEurasian snow cover) is considered in “Additional material 2”.
ERA-I MMM EM: |r|>0.6
12
8
4
0
4
8
12
regr
essio
n slo
pe
CMCCECMWFDWDMFUKMO
Figure 8: Impact of ENSO on local LSWF. regression ofv′T ′ on niño3.4 index—DJF averages—for ERA-Interim(left) and for multi-model mean (MMM, center). In thelatter, total model agreement in the sign of the regression ishighlighted with dark red contours. Correlation (r) betweenmodel ensemble-mean v′T ′ and niño3.4 is displayed forvalues greater/smaller than 0.6/-0.6 (EM, right).
Alice Portal Seasonal stratosphere prediction 6 / 10
Predictability of local LSWF (v ′T ′), links with ENSO
300
300
300
ERA-I
60
60
60
CMCC
60
MF
60
60
60
180
ECMWF
60
6060
60
180
180
DWD
60
60
60
6060
UKMO
90
60
30
0
30
60
90v'
T' (m
/s K
)
252015105
0510152025
Cova
rianc
e (m
/s K
)
November
300
300
ERA-I30
30
30CMCC
30
30
MF
30
90 90
ECMWF
30
30
30
90
DWD
3030
30
3090
UKMO
90
60
30
0
30
60
90
v'T'
(m/s
K)
5
0
5
Cova
rianc
e (m
/s K
)
December-January-February
Figure 7: Left: ERA-Interim climatologies of v′T ′ (green–purple). Right: Covariance between model ensemble mean v′T ′ and standardised ERA-Interimv′T ′ (blue–red, note different scaling in November and DJF). Stippling shows significant ACC (p≤0.05). Black contours show model ensemble-meanvariance at 60, 120, 180, 300, 600 (m/s K)2 in November; at 30, 60, 90, 180 (m/s K)2 in DJF; reanalysis interannual variance at 300, 600, 900 (m/s K)2.
ERA-I MMM EM: |r|>0.6
12
8
4
0
4
8
12
regr
essio
n slo
pe
CMCCECMWFDWDMFUKMO
Figure 8: Impact of ENSO on local LSWF. regression ofv′T ′ on niño3.4 index—DJF averages—for ERA-Interim(left) and for multi-model mean (MMM, center). In thelatter, total model agreement in the sign of the regression ishighlighted with dark red contours. Correlation (r) betweenmodel ensemble-mean v′T ′ and niño3.4 is displayed forvalues greater/smaller than 0.6/-0.6 (EM, right).
Alice Portal Seasonal stratosphere prediction 6 / 10
Relation between regional LSWF and SPVPredictability of regional LSWF:
West- and East- North Pacific LSWF displays seasonal skill, mainly thanks to ENSO (cf.Figure 7 and 8);Eurasian LSWF shows high interannual variance, but is poorly predictable in the seasonalrange (Figure 7).
We estimate the importance of Pacific and Eurasian LSWF for the seasonal variability of the SPV.Model results are compared with reanalysis.
Table 1: Strength of the link between SPV and regional LSWF. Correlation between DJF anomalies of SPV winds (∆U) and DJF wind anomaliesreconstructed from regional LSWF (∆U∗reg, see Eq. 1). The latter are calculated using v′T ′ from the regions selected in the figure on the right. Boldvalues are significant at 95% confidence level.
40-80◦ N W Pacific E Pacific Pac sec Eurasia
CMCC 0.91 -0.05 0.75 0.40 0.85MF 0.82 -0.33 0.61 0.55 0.51ECMWF 0.84 -0.22 0.53 0.26 0.71DWD 0.74 -0.35 0.45 0.07 0.69UKMO 0.89 -0.40 0.68 0.19 0.64
ERA-I 0.91 0.14 0.01 0.11 0.7240-80 NPac secE Pacific
W PacificEurasia
Alice Portal Seasonal stratosphere prediction 7 / 10
Summary of the results
I. In the boreal stratosphere:I winter variability and predictability are affected by QBO in the tropics and by the presence of the
SPV in the extratropics (Figure 2);I we find a wide inter-model spread in the winter predictability of the SPV, which does not depend
on bias. Three models exhibit significant DJF skill (Figures 3, 4).
II. Daily coupling between LSWF and SPV is well reproduced by 4 models (Figure 5). Onemodel, MF, shows substantial departures from reanalysis.Seasonal anomalies of LSWF explain ∼ 80% of interannual SPV variability (Figure 6). Onemodel, DWD, shows weaker seasonal coupling.
III. Good predictability of the local LSWF is found within a month from initialisation (Nov).Over the Pacific model skill persists into DJF, it is weaker over Atlantic and Eurasia.Model SPV skill is favoured by:
I strong interannual variability of the winter LSWF (Figure 6);I November predictability of local LSWF (Figure 7);I DJF predictability of Eurasian LSWF (Table 1), lacking in models that exhibit no SPV skill in the
seasonal range (Figure 7).Model SPV shows a strong link with regional Pacific LSWF (Table 1) forced by strongSPV–ENSO teleconnection (Figure 8). This feature is not present in reanalysis.
Alice Portal Seasonal stratosphere prediction 8 / 10
Additional material 1
General description of the 5 seasonal prediction systems.
Models Resolution ∗ InitialConditions ∗∗
EnsembleSize
CMCC(system 3)
1◦ lat/long46 L 1st November 40 members
MF(system 6)
TL35991 L
20th, 25th October1st November
2×12 members1 member
ECMWF(SEAS5)
TCO31991 L 1st November 25 members
DWD(system 2)
T12795 L 1st November 30 members
UKMO(GloSea5, system 13)
N21695 L
25th October1st, 9th November
7 membersfor start date
∗For vertical resolution we indicate the number of vertical levels (L).∗∗Model simulations start on the 1st of November, in years 1993 to 2016. Differently, UKMO simu-lations start on three separate dates between the end of October and the beginning of November—see initial-conditions dates.
Alice Portal Seasonal stratosphere prediction 9 / 10
Additional material 2Impact of seasonal winter signal on LSWF and SPV:
ENSO QBO A sice EA snowc
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8
slope
LSWF regression
ENSO QBO A sice EA snowc
2
1
0
1
2
3
SPV regression
ERA-I CMCC ECMWF DWD MF UKMO
Figure 9: Regressions on sources of seasonal variability. On the left, the linear regression slope between DJF LSWF and seasonal signals, i.e. DJF ENSO,DJF QBO, October-November Arctic sea ice, October-November Eurasian snow cover. The same for DJF SPV on the right. The 5th and 95th percentileof model slope are represented with errorbars—relative distribution is calculated using bootstrap. ERA-Interim regressions characterised by significantcorrelation according to one-tailed t-test (p≤0.05) are indicated with full black circles on top of black crosses.
Alice Portal Seasonal stratosphere prediction 10 / 10
Top Related