Application of Generalized Extreme Value Theory to

Download Report

Transcript Application of Generalized Extreme Value Theory to

Application of Generalized Extreme Value
theory to coupled general circulation
models
Michael F. Wehner
Lawrence Berkeley National Laboratory
[email protected]
SAMSI Climate Change Workshop
February 17-19, 2010
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Outline
 GEV results in assessment reports
 Uncertainty in temperature extremes
 Model fidelity and precipitation extremes
 A few points for the discussion session
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
GEV results in assessment reports
“Rare events will become commonplace”
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
 Simulations for 2090-2099 indicating how currently rare extremes (a 1-
in-20-year event) are projected to become more commonplace. a)
Temperature - a day so hot that it is currently experienced once every
20 years would occur every other year or more by the end of the
century. (Units:years)
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Sources of uncertainty in estimating return values
 20 year return value of annual maximum
daily mean surface air temperature
 GEV parameters (Short sample size)
 Unforced internal variability
 Multi-model differences
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
GEV parameter uncertainty
 Following the bootstrapping method of Hosking and Wallis
1. Fit GEV parameters to sample
2. Generate 50 random samples distributed by the GEV distribution
3. Calculate return values and their standard deviation
CCSM3.0
a) 20 years
b) 40 years
c) 100 years
d) Average over
land (CMIP3
models)
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Internal variability
1. Divide long control run into 40 year segments
2. Calculate return value for each segment and s
CCSM3.0
600 years
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Multi-model variation
 Fifteen CMIP3 forty year control runs
 Intermodel standard deviation
Color scale is
5 times the
previous two
slides
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Multi-model variation
 Fifteen CMIP3 forty year control runs
 Sam as previous except remove mean state bias
Color scale is
5 times the
previous two
slides
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Model resolution and extreme precipitation
 Typical CMIP3 models are too coarse to simulate
rare intense storms.
 Horizontal resolution study with fvCAM2.2
 200km (B mesh)
 100km (C mesh)
 50km (D mesh)
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Model resolution and extreme precipitation
 20 year return value of annual maximum daily total precipitation (mm/day)
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
from the CCSP3.3 report
 Simulations for 2090-2099 indicating how currently rare extremes (a 1-in-
20-year event) are projected to become more commonplace. (b) daily total
precipitation events that occur on average every 20 years in the present
climate would, for example, occur once in every 4-6 years for N.E. North
America. (Units:years)
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Conclusions
 IPCC AR5 will contain far more about extremes than AR4
 Largest source of uncertainty is inter-model difference
 Uncertainty in the fit of GEV is about the same as
unforced internal variability and is small!
 Extreme precipitation requires high resolution.
 At least over land.
 Makes it hard to make projections with the CMIP3
models.
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Discussion
 GEV distribution fits climate data
very well
 Cells that fail the Anderson Darling
test at 5% level
 Surface air temperature annual
maximum
 Arctic failure is due to clustering
at freezing point. Not very
interesting, return value is 0oC.
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Discussion
 Detection & attribution of changes in extreme weather
events
 Zwiers et al GEV methodology: let location parameter
be time dependent. Scale and shape parameters be
static. Test whether time dependence is significant.
• Temperature: Is this trivial if mean temperature changes have
been detected and attributed? How does the difference between
a return value and the mean change?
• Precipitation: Widely believed to be more detectible due to
Clausius-Clayperon relationship. But changes may not be of the
same sign. May not be as severe as mean precipitation
changes
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
Projected 1990-2090 DRV minus DTmean
 SRES A1B (4 models)
 Change is confined to land and fairly small (<2.5K)
 Should we expect to detect this change in distribution shape?
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
1990-2090 wintertime precipitation changes
SRES A1B
20 year return value
C
O
M
P
U
T
A
T
I
O
N
A
L
mean
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N
x

Thank You!
C
O
M
P
U
T
A
T
I
O
N
A
L
R
E
S
E
A
R
C
H
D
I
V
I
S
I
O
N