View
11
Download
0
Category
Preview:
Citation preview
Post-processing climate model output toobtain accurate high-resolution climateprojections
Thordis ThorarinsdottirNorwegian Computing Centerwww.nr.no/~thordis
Bern
November 9, 2018
Joint with Q Yuan, W K Wong, S Beldring, S Huang and C-Y Xu
Post-processing climate model output toobtain accurate high-resolution climateprojections & why uncertainty matterseven if the answer is just a number
Thordis ThorarinsdottirNorwegian Computing Centerwww.nr.no/~thordis
Bern
November 9, 2018
Joint with Q Yuan, W K Wong, S Beldring, S Huang, C-Y Xu and
Peter Guttorp
(IPCC, 2013)
Objective
Data set Resolution Historical Futureperiod period
seNorge 1× 1 km X ?EURO-CORDEX RCM 12× 12 km X X
I Bias-correction: Correct errors in RCM output
I Downscaling: Recover spatial and temporal variability at thefine scale
I Get results within a feasible computation time
6 / 34
Methods for bias-correction/downscaling
I Model output statistics: Apply a statistical transfer functionbetween model and observed data
I Perfect prognosis: Link large-scale predictors and local-scalepredictands in a regression framework
I Weather generators: Stochastic models that explicitelymodel marginal and higher order structures
(Maraun and Widmann, 2018)
7 / 34
seNorge_hist
RCM_hist
Our general framework
seNorge_up
seNorge_hist
RCM_hist
upscale
Our general framework
seNorge_up
seNorge_hist
RCM_hist_corr
RCM_hist
correct bias
upscale
Our general framework
RCM_future_corr
seNorge_up
seNorge_hist
RCM_hist_corr
RCM_histkeep climate change signal
correct bias
upscale
Our general framework
RCM_future_corr
seNorge_up
seNorge_hist
RCM_hist_corr
Variability at 1x1
RCM_histkeep climate change signal
simulate
correct bias
upscale
Our general framework
RCM_future_corr
seNorge_up
seNorge_hist
RCM_hist_corr
Variability at 1x1
RCM_hist
Final outputat 1x1
keep climate change signal
downscale
simulate
correct bias
upscale
Our general framework
Study region and data
I Daily temperature data
I Training set: 1957-1986
I Test set: 1987-2005
I 9 catchments, 150-3000 km2
14 / 34
Bias-correction
Yst ∼ N(µst , σ2st)
where
µst = f µ1 (xs) + f µ2 (t) + f µ3 (t)
log(σst) = f σ1 (xs) + f σ2 (t)
with
f1: baseline
f2: seasonality
f3: trend
15 / 34
Modelling fine scale variability
We simulate fine scale variability from a separable space-timemodel estimated using historical residuals at 1× 1 km:
Zst =Yst − µstσst
= ηt + νst
ηt ∼ autoregressive process
Wst = Zst − ηt ∼ N(0,Σt)
Cov(Wst ,Ws′t) = θ0t1{‖s − s ′‖ = 0}+ θ1t exp(−‖s − s ′‖/θ2t)
Simulating 19 years of daily fields for 5000 grid points: 4.5 hours
16 / 34
Z·t does not have Gaussian marginals
Z·t ∼ TPN(µt , σ1t , σ2t), Ut = Φ−1(FTPN(Z·t)) ∼ AR(p)
17 / 34
Predictive marginal performance per catchment1987-2005
Full distribution Middle part
Gaulfo
ss
Aamot
Krinsv
atn
Oeyun
gen
Trang
en
Vera
vatn
Dillfos
s
Hoegg
aas_
Bru
Kjelds
tad_
i_Gar
berg
elva
Gaulfo
ss
Aamot
Krinsv
atn
Oeyun
gen
Trang
en
Vera
vatn
Dillfos
s
Hoegg
aas_
Bru
Kjelds
tad_
i_Gar
berg
elva
0.00
0.01
0.02
0.03
0.04
IQD
dsMethod eqm Xstar XshiftTrd XshiftTrdSigma rcm CNRM−CM5_CCLM MPI_CCLM
IQD(F ,G ) =
∫(F (x)− G (x))2ω(x)dx
18 / 34
Predictive marginal performance per catchment1987-2005
Lower part Upper part
Gaulfo
ss
Aamot
Krinsv
atn
Oeyun
gen
Trang
en
Vera
vatn
Dillfos
s
Hoegg
aas_
Bru
Kjelds
tad_
i_Gar
berg
elva
Gaulfo
ss
Aamot
Krinsv
atn
Oeyun
gen
Trang
en
Vera
vatn
Dillfos
s
Hoegg
aas_
Bru
Kjelds
tad_
i_Gar
berg
elva
0.000
0.003
0.006
0.009
IQD
dsMethod eqm Xstar XshiftTrd XshiftTrdSigma rcm CNRM−CM5_CCLM MPI_CCLM
IQD(F ,G ) =
∫(F (x)− G (x))2ω(x)dx
19 / 34
Autocorrelation functions per catchment 1987-2005
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
0.7
0.8
0.9
1.0
0 5 10 15
rawRCM
obs
eqm
sdm
20 / 34
Spatial correlation per catchment 1987-2005
21 / 34
Spatial correlation per catchment 1987-2005
22 / 34
Why uncertainty matters even if theanswer is just a number
Sea level will rise in Bergen on Norway’s west coast
1950 2000 2050 2100
−20
020
4060
8010
012
0
RCP 8.5
Year
Ano
mal
y (c
m)
24 / 34
Previous report assessed feasibility, consequencesand costs of several adaptation options
1. Outer barrierI > 30 billion NOKI Large
environmental andeconomicconsequences
2. Inner barrier atVagenI 500 million NOKI Limited benefits
3. Inner barrier atDamgardssundetI 500 million NOKI Limited benefits
Regional Havstigning Prosjektrapport
Bergen, 2009-3-26
Grieg Foundation Visjon Vest G. C. Rieber Fondene
25 / 34
Our questions
I Are these adaptation options appealing from a cost/benefitperspective?
I If we should adapt, when would be the best time?
I What are the effects of the associated uncertainties on thecost/benefit analysis?I Sea level rise is uncertainI Total yearly damage in each year is uncertainI Change in the total yearly damage due to sea level rise is
uncertain
https://github.com/eSACP/SeaLevelDecisions
(T. et al., Water Resources Research, 2017)
(Guttorp and T., Significance, 2018)
26 / 34
Our questions
I Are these adaptation options appealing from a cost/benefitperspective?
I If we should adapt, when would be the best time?
I What are the effects of the associated uncertainties on thecost/benefit analysis?I Sea level rise is uncertainI Total yearly damage in each year is uncertainI Change in the total yearly damage due to sea level rise is
uncertain
https://github.com/eSACP/SeaLevelDecisions
(T. et al., Water Resources Research, 2017)
(Guttorp and T., Significance, 2018)
26 / 34
Our questions
I Are these adaptation options appealing from a cost/benefitperspective?
I If we should adapt, when would be the best time?
I What are the effects of the associated uncertainties on thecost/benefit analysis?I Sea level rise is uncertainI Total yearly damage in each year is uncertainI Change in the total yearly damage due to sea level rise is
uncertain
https://github.com/eSACP/SeaLevelDecisions
(T. et al., Water Resources Research, 2017)
(Guttorp and T., Significance, 2018)
26 / 34
Local sea level projections
We relate git-corrected Bergen sea level to global sea levelseries of Church and White (2011), then use the method of Bolinet al. (2014) to model the relationship between global annualmean temperature and global annual mean sea level rise.
27 / 34
Local sea level projections
1950 2000 2050 2100
−20
020
4060
8010
012
0
RCP 8.5
Year
Ano
mal
y (c
m)
1950 2000 2050 21000
5010
0
RCP 8.5
Year
Ano
mal
y (c
m)
28 / 34
Changes in damage costs due to sea level rise
●●
●
Sea level anomaly (cm)
Rel
ativ
e m
ean
annu
al d
amag
e
0 10 20 30 40
15
100
150
200
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
● ● ●
Sea level anomaly (cm)
Rel
ativ
e m
ean
annu
al d
amag
e
−100 −50 0 50 100
0.00
51
200
400
600
800
1000
● ● ●●●
●
● ●
●
● ●
●
●●
●
● ● ●● ● ●●
●
●
● ●●
● ●
●
● ●●
● ●●
●●
●
● ●
●
● ●
●
Hallegatte et al. (2013) investigate global changes in damagecosts under 20 and 40 cm sea level rise. We extrapolate theirresults for 15 European cities and use the results as an ensembleprediction for the changes in damage costs in Bergen.
29 / 34
Annual damage costs
Cost (million NOK)
Dens
ity
0 5 10 15 20 25
0.0
0.5
1.0
1.5
2.0
2.5
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●
●
●
● ●
●
●
Estimated quantiles
Obs
erve
d qu
antile
s
0 5 10 15 20 25
05
1015
2025
The Norwegian Natural Perils Pool publishes annual damage costsdue to storm surges on county level. We fit a Burr distribution tothe 1980-2015 data from Hordaland and Rogaland counties.
30 / 34
Optimal adaptation timing depends on thedecision-maker’s loss function/risk aversion
2020 2040 2060 2080 2100
01
23
4
Year of adaptation measure
Tota
l cos
t 201
6−21
00 (
billi
on N
OK
)
Adaptation option: Build two inner barriers
31 / 34
Including the uncertainty is vital; uncertainty in thedamage costs has the largest effect
Total cost 2016−2100 (million NOK)
10 20 50 150 500 1500 5000
RCP 8.5
RCP 4.5
RCP 2.6 No uncertaintyDamage uncertaintyEffect uncertaintySLR uncertaintyFull uncertainty
32 / 34
Excluding uncertainty in sea level rise projectionsdelays the assessed timing for adaptation by adecade
2020 2040 2060 2080 2100
01
23
4
Year of adaptation measure
Tota
l cos
t 201
6−21
00 (
billi
on N
OK
)
Adaptation option: Build two inner barriers
33 / 34
Conclusions
I Statisticians have a lot to contribute in climate science andclimate-related decision-making (Benestad et al., NCC, 2017)I Stochastic models are needed to realistically describe fine scale
variability in statistical downscalingI Uncertainty assessments and propagation of uncertainty are
vital
I Communication and messaging are key components
34 / 34
Recommended