Sei sulla pagina 1di 21

Aims Bias and Convergen

e The model Results Con lusions

Probabilisti Estimation of Climate Change in

Venezuela using a Bayesian Approa h

Alexis Durán1 and Lelys Guenni2

June 2010

1
Universidad Experimental Ezequiel Zamora, Venezuela
2
Universidad Simón Bolívar, Venezuela

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Contents

1 Aims
2 Bias and Convergen e
3 The model
4 Results
5 Con lusions

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Aims

Analize temperature and pre ipitation hanges proje ted by


the 16 GCMs used in the First Venezuelan Climate Change
Communi ation
Find the posterior probability distributions of present and
future proje tions of pre ipitation and temperature at a
regional s ale (5◦ × 5◦ grid size)using a Bayesian approa h
Provide a reliability measure of GCMs limate hange
proje tions for present (1960-1989) and future limates (30
years entered in 2025, 2050 and 2100). Two riteria are used:
Bias and Convergen e. This is an appli ation of the Reliability
Ensemble Average method proposed by Tebaldi et al. (2005)

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

The Data

Outpus of 16 OAGCMs models from the MAGICC /


SCENGEN software (Wigley et al., 2003) with medium limate
sensitivity and a SRES2 IPCC s enario (heterogeneous world;
in reasing population; e onomi growth regionally oriented
with low globalization; low and fragmented per apita
e onomi growth and te hnologi al hanges)
Pre ipitation and temperature data were obtained from the
R-Hydronet network data set for South, Central Ameri a and
the Caribbean for years 1960 to 1989. These data were
re al ulated for the nine 5◦ × 5◦ grids overing the whole
ountry as shown in the next Figure.

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

The OAGCM Models

Table: OAGCM models onsidered by MARN (2005)


Model Number Name Country
M1 HadCM2 England
M2 UKTR England
M3 CSIRO-TR Australia
M4 ECHAM4 Germany
M5 UKHI-EQ England
M6 CSIRO2-EQ Australia
M7 ECHAM3 Alemania
M8 UIUC-EQ USA
M9 ECHAM1 Germany
M10 CSIRO1-EQ Australia
M11 CCC-EQ Canada
M12 GFDL-TR USA
M13 BMRC-EQ Australia
M14 CGCM1-TR Canada
M15 NCAR-DOE USA
M16 CCRS/NIES Japan

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

The Grids used in the analysis

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Bias

Aim: To evaluate the existing un ertainty of the dierent GCMs in


proje ting the present limate
Denition of Bias: Dieren e between ea h model output and a
weighted mean between model outputs and observed data
(weighted by the posterior pre isions):
X0 ∗ τ0 + P ni=1 Xi ∗ τi
P
µ̃ = n (1)
τ0 + i =1 τ i
Biasi = Xi − µ̃
Subs ript i varies from 1 to n; n = 16 is the total number of
limate models used; Xi is the average of the limati variable for
present limate for model i with pre ision τi ; X0 is the observed
data average with pre ision τ0 and µ̃ is the mean (weighted) for
present limate. (Pre ision is the inverse of the varian e).
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Convergen e

Aim: To evaluate the un ertainty among the GCMs in future


limate proje tions. In this ase a model is more reliable if its
ability to make future limate proje tions is similar to signals and
magnitudes of the remaining model proje tion hanges. In
mathemati al form onvergen e for model i an be expressed as:

i =1 Yi ∗ λi
Pn
ν̃ = P n (2)
i =1 λi
Convergen ei = Yi − ν̃
Subs ript i varies from 1 to n; n = 16 is the total number of
limate models used; Yi are the average future limate proje tions
for model i with pre ision λi and ν̃ is the mean (weighted) for
future limates.

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

The model

We onsider Xij and Yij pre ipitation and temperature simulated for
present and future limates respe tively, for the dierent GCMs.
i = 1, . . . , 16 and j goes from 1 to 9 for ea h grid ell. The analysis
is arried out at a monthly level, with pre ipitation and temperature
quarterly averaged (DJF, MAM,JJA,SON).
We onsider two ases:
1 Present and future limate are independent
2 Present and future limate are not independent

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Independent Case

Model Data likelihood:


Xij ∼ N (µj , τij−1 ) (3)
Yij ∼ N (νj , (θτij )−1 ) (4)
The parameter of interest is: ∆j = νj − µj
Observed Data likelihood:
X0 ∼ N (µj , τ0−1 ) (5)
Parameters of models 3 y 4 are estimated using a Bayesian
approa h. Prior distributions should be provided for the parameter
set (µj , νj , τij , θ ). Diuse gamma priors are proposed for θ and τij ;
uniform priors are proposed for µj and νj restri ted to nite values
with a wide range to onsider all possible parameter values.
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Posterior distribution: independent ase

The posterior probability distribution for the parameter ve tor at


ea h grid ell j has the form:

p(µ, ν, τi , θ|X0 , P
X, Y ) ∝
na + −1
exp{ ni=1 [bτi − τ2 ((Xi − µ)2 + θ(Yi − ν)2 )] (6)
n

τi θ 2 i

−[d θ + 2 (X0 − µ)2 ]}


τ 0

where a, b, , d are known prior hiperparameters.


Inferen e from the posterior distribution is a hieved by using
MCMC methods

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Dependent ase

Model Data likelihood:


Xi = µ + ηi (7)
ε
Yi = ν + β(Xi − µ) + √i (8)
θ
1
Yi |ν, θ, µ, β, Xi , τi θτi exp{− θτi (Yi − ν − β(Xi − µ))2 } (9)
p

2
In this model a linear relationship between (Yi − ν) and (Xi − µ) is
dened by the parameter β .
If β takes the zero value the model redu es to the independent
ase. As in the previous ase the parameter of interest is
∆j = νj − µj , whi h is the limate hange at grid ell j .

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Posterior Distribution: dependent ase

The same prior distribution as in the independent ase is assumed


for all parameters. A non-informative uniform prior for β is
assumed in the interval (−1, 1).
After applying the Bayes theorem the posterior distribution for
parameter ve tor (µ, ν, τi , β, θ) is proportional to:
p(µ, ν, τi , β, θ|X0 , P
X, Y) ∝
τina+1−1 θ + 2 exp{ ni=1 [−b τi − X − µ)2 + θ(Yi − ν − β(Xi − µ))2 )]
n
τ
2 (( i
i

−[d θ + τ20 (X0 − µ)2 ]}


(10)
where a, b, , d are known prior hiperparameters.
Inferen e from the posterior distribution is a hieved by using
MCMC methods

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Temperature: Independent ase

DEF MAM

3.0
3.0
Deltha (ºC)

Deltha (ºC)

2.5
2.5
2.0

2.0
1.5

1.5
grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

JJA SON

4.0
3.5
3.0
Deltha (ºC)

Deltha (ºC)

3.0
2.5

2.5
2.0
2.0

1.5
grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

Figure: Boxplots of simulated ∆ for temperature in the independent ase


for 30 years entered in the year 2100
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Dependent Case: Temperature

Table: Summary statisti s of 5000 simulated β by grid, entered in the


year 2100
Grid Per entile 5% Mean Median Per entile %95
1 7.65302 9.61837 9.68360 11.36844
2 8.24154 10.15864 10.23301 11.88699
3 6.55086 8.55633 8.51311 10.68685
4 9.19478 10.38972 10.43891 11.39611
5 8.80463 10.06858 10.08505 11.27856
6 8.34223 9.83922 9.837431 11.34465
7 8.75636 10.75940 10.82612 12.51741
8 8.79315 10.25467 10.29290 11.57691
9 9.37951 10.51142 10.52238 11.58817

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Temperature hanges: dependent ase

DEF MAM

4.0

3.5
3.0
Deltha (ºC)

Deltha (ºC)
3.0

2.5
2.0
2.0

1.5
1.0

grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

JJA SON

4.0
3.5
3.0

3.0
Deltha (ºC)

Deltha (ºC)
2.5

2.0
2.0
1.5

1.0
grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

Figure: Boxplots of simulated (∆)for the temperature in the dependent


ase for 30 years entered in the year 2100
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Temperature pre ision: dependent ase

DEF MAM

500

500
Presición

Presición
300

300
0 100

0 100
M1 M3 M5 M7 M9 M12 M15 M1 M3 M5 M7 M9 M12 M15

Modelo Modelo

JJA SON
600

500
Presición

Presición
400

300
200

0 100
0

M1 M3 M5 M7 M9 M12 M15 M1 M3 M5 M7 M9 M12 M15

Modelo Modelo

Figure: Temperature pre ision obtained from 5,000 simulations of the


posterior distribution for 30 years entered at 2100, dependent ase
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Best models for temperature proje tions

Table: Best models for temperature proje tions, dependent ase


Quarter Model Grid
DEF GFDL-TR All minus 4
UKTR 4
MAM CCC-EQ All minus 7
UKTR 7
JJA UIUC-EQ 4 - 5 - 6 - 7 - 8 - 9
BMRC-EQ 1-2-3
SON UKTR 1 - 2 - 3 - 4 - 5 - 6
CSIRO1-EQ 7-8-9

TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Pre ipitation hanges: dependent ase

DEF MAM

20

600
10
Deltha (mm)

Deltha (mm)

400
0
−10

200
−30

0
grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

JJA SON
100

20
Deltha (mm)

Deltha (mm)

0
0

−20
−100
−200

−60
grilla1 grilla3 grilla5 grilla7 grilla9 grilla1 grilla3 grilla5 grilla7 grilla9

Figure: Boxplots of simulated ∆ for pre ipitation in the dependent ase


for 30 years entered in the year 2100
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Pre ipitation pre ision: dependent ase

DEF MAM

400

200
300

150
Presición

Presición
200

100
100

50
0

0
M1 M3 M5 M7 M9 M12 M15 M1 M3 M5 M7 M9 M12 M15

Modelo Modelo

JJA SON

120
300
Presición

Presición

20 40 60 80
200
100
0

0
M1 M3 M5 M7 M9 M12 M15 M1 M3 M5 M7 M9 M12 M15

Modelo Modelo

Figure: Pre ipitation pre ision obtained from the posterior simulations for
the data entered around the year 2100, dependent ase
TIES 2010
Aims Bias and Convergen e The model Results Con lusions

Con lusions and future work

1 Observed trends for the dependent and independent ase:


Temperature will in rease in the oming years ountrywide.
2 For pre ipitation there were dieren es for the dependent and
independent ase. For the dependent ase results indi ate
de reasing rainfall for the rst 60 years of proje tions, followed
by a onsiderable in rement in the following years. In the
independent ase a de rease of pre ipitation was onsistently
observed.
3 The low posterior pre isions for pre ipitation and the high bias
and onvergen e values for most models indi ate a lower
onden e in the pre ipitation proje tions in omparison with
temperature.

TIES 2010

Potrebbero piacerti anche