Transcript Jouni Räisänen

Climate in the near future – results
from a simple probabilistic method
Jouni Räisänen and Leena Ruokolainen
Department of Physical Sciences, Division of Atmospheric
Sciences, University of Helsinki, Finland
What will I show and what is it good for?
• A ”resampling ensemble” method for deriving
probabilistic estimates of climate change
– uses existing multi-model ensembles of climate change
simulations (IPCC AR4 data set)
– first-order representation of both modelling uncertainty and
natural variability
– related to pattern scaling – but no intention to remove noise
• Best suited for projections of near-term climate change
– sample size
– for longer-term projections, the unknown ability of multi-model
ensembles to capture the actual modelling uncertainty
becomes a larger headache
Annual mean Temp and Prec changes at
(60ºN, 25ºE), from 1971-2000 to 2011-2020,
as simulated by 21 models under the A1B scenario
~95% probability of
warming, ~95%
probability of increasing
precipitation?
However,
21 is a small sample
for estimating probabilities
Resampling ensembles
• Patterns of simulated climate change remain quasi-constant
in time, when the forcing is dominated by increasing GHGs and
internal variability is filtered out e.g. by averaging over a large
number of models.
* Same 21-model mean global warming (0.62C) in both cases.
* Regional differences much smaller than differences between
individual simulations (rms difference = 0.11C)
multi-model
global mean T
Work hypothesis
“1900”
P4
P3
P2
P1
time
“2100”
If the multi-model mean global mean temperature change
is the same from period P3 to P4 as from P1 to P2, then
the probability distribution of regional climate changes
should also be approximately the same in the two cases.
multi-model
global mean T
P3
P4
P2
P1
time
Resampling ensembles for the climate change from P1 to P2
(e.g., 1971-2000 to 2011-20) are formed by taking the climate
changes in “all” pairs of periods P3  P4 with the same multimodel mean global warming as plausible realisations of the
change from P1 to P2.
Cross verification* indicates that the increased sample size
(as compared with only using P1 and P2) outweighs eventual
biases caused by the methodology, for both T and Precip
*Räisänen and Ruokolainen (2006, Tellus 58A, 461-472)
Technical details
• Data set
–
–
–
–
IPCC AR4 simulations
21 models for A1B scenario
one transient simulation (1901-2098) per model
also some analysis with constant-forcing control
simulations
• Resampling with 5-year interval in “P4”
– nominal sample size for forecasts from 1971-2000 to
2011-2020 = 420 (20 pairs of periods × 21 models)
– 21 << effective sample size << 420
Annual mean Temp and Prec changes at
(60ºN, 25ºE), from 1971-2000 to 2011-2020: the
resampling ensemble
95% probability of
warming, 80% probability
of increasing precipitation?
Sample size >> 21
 these estimates are
likely to be more reliable
than the ones (95% and
95%) obtained directly from
the 1971-2000 and
2011-2020 data.
Annual and seasonal T and P changes at
(60ºN, 25ºE), from 1971-2000 to 2011-2020
Seasonal means have a wider pdf than annual means
(for temperature change, particularly in winter),
and monthly means even more so.
Note: Gaussian shape is used for illustration only (although it seems to
be a good approximation)
Temp and Prec changes at (60ºN, 25ºE)
from 1971-2000 to 2011-2020, A1B scenario
Temperature change
Median
estimate
Prob. of
warming
Precipitation change
Median
estimate
Prob. of
increase
DJF
1.2C
90%
5%
73%
JJA
0.6C
89%
3%
63%
Ann
0.9C
95%
4%
80%
“Best-guess” warming: winter > summer
Probability of warming: winter ≈ summer
Lower signal-to-noise ratio makes forecasts of precipitation
change less certain than those of temperature change
Annual mean T and P changes at (60ºN,25ºE),
from 1971-2000 to later decades (A1B scenario)
The pdf widens with time, as model differences become
increasingly important with increasing forcing
Best-guess annual mean warming versus
probability of warming, as estimated from the
models (from 1971-2000 to 2011-2020)
°C
High probability of warming almost everywhere
Particularly high probability of warming in tropical latitudes,
where internal variability is small!
%
Recent climate changes:
1991-2000 vs. 1961-1990
Observed annual mean temperature
change from 1961-90 to 1991-2000
(Tyndall Centre / CRU)
C
How usual / unusual is this in simulations
- with no external forcing
- with increasing GHG concentrations?
Probability of
below-observed
temperature change,
simulations with no external
forcing
< 5%: nowhere
>95%: 58% of land
The same, in (greenhouse
gas etc.) forced simulations
< 5%: 3%
>95%: 5%
Changes from 1961-90 to 1991-2000
• Observed temperature changes
– in many areas, too large to be reasonably explained by internal
climate variability (as estimated from the models)
– consistent with a combination of anhtropogenic climate change
and internal variability
• Observed precipitation changes (not shown)
– Within the 5-95% range of the model-based distributions in 83%
of all land – both for the unforced and the forced simulations
• Similar conlusions (impact of greenhouse gas forcing
clearly detectable in temperature, but not in precipitation) are
obtained with more advanced detection-attributionmethods
”Variance correction”
• Resampling ensemble method in its basic form assumes that
the magnitude of natural variability is correctly simulated by
models
• If not – the pdfs may become systematically too narrow or too
wide (particularly important for short-term forecasts, in which
uncertainty is dominated by natural variability)
• Direct evaluation of interdecadal variability virtually
meaningless (small sample sizes)
• Ruokolainen and Räisänen (2007)* implemented a variance
correction scheme based on a comparison of simulated and
observed interannual variability
• Cross verification suggests that the correction makes more
good than harm
*Tellus 59A, 309-320
Annual mean Temp and Prec changes at
(60ºN, 25ºE) – without and with variance
correction (1971-2000 to 2011-2020)
95% (95%) probability of
warming, 75% (80%)
probability of increasing
precipitation?
Models tend to underestimate
interannual precipitation
variability (at this location)
 variance correction results in
a slightly wider distribution of
precipitation changes.
In general, the variance correction appears to have only
relatively modest effects (but P is affected more than T).
Strengths and limations of the method
• Strengths
– Simple
– Efficient way of extracting probabilistic information from long
transient simulations
– Applicable to both multi-model and perturbed-parameter
ensembles
• Limitations
– ”Signal” assumed to be fully determined by multi-model average
global mean warming (not exactly true)
– Biases in simulated variability may affect width of the pdfs
(although this may be partially corrected in post-processing)
– No attempt to use observational constraints to weight or scale
model-simulated climate changes (but how much would this
change projections of near-term change?)
Another short story: climatic nowcasting?
• March 2007 was extremely warm in Helsinki:
(Tmean = 3.1C – previous record = 2.0C)
• How unusual was this
– In the context of the 20th century climate?
– In the present ”AD 2007” climate?
• Question answered by estimating a pdf for the
”AD 2007” March temperature
– starting point: observations for 1901-2000
– -change approach, taking into account (i) observed global
mean warming and (ii) AR4-model-simulated changes in
March mean temperature and interannual variability
– details to be documented…
Resulting probability
distributions
2007
1901-2000
2 3
Return period estimates
≥ 2.0C
≥ 3.0C
Observed (1901-2000) climate
~ 60 yr
~ 700 yr??
Present (AD 2007) climate
~ 14 yr
~ 80 yr?
Probability of
below-observed
precipitation change,
simulations with no external
forcing
< 5%: 9%
>95%: 8%
The same, in (greenhouse
gas etc.) forced simulations
< 5%: 10%
>95%: 7%
Cross verification – in brief
1.
2.
3.
Choose one model simulation as “truth”, against which
forecasts derived from other models are verified
Calculate a verification statistics (and average over the
global domain)
Repeat 1-2 for all choices of the verifying model, and
average the verification statistics
Cross verification gives no absolute measure of forecast
performance in the real world, but it is a useful tool for
comparing the potential performance of different
forecast methods.
Cross verification results:
annual mean T and P change
CRPS = continuous ranked probability score.
Perfect deterministic forecast : CRPS = 0.
Temperature
Precipitation
Standard
Ratio
Standard
Ratio
2011-2020
0.188ºC
0.963
4.34%
0.964
2041-2050
0.285ºC
0.980
5.60%
0.973
2071-2080
0.417ºC
0.987
7.04%
0.982
CRPS increases with
time: long-term forecasts
are less accurate than
short-term forecasts
Resampling method yields lower
CRPS scores than the standard method
(in which each simulation is used only once).
This suggests that resampling improves
the forecasts
Quantile plots of climate change from 1971-2000
to 2011-2020: impact of “variance correction”
Basic resampling method
Resampling with variance correction
Where and when simulated interannual variability is smaller than the
observed variability, variance correction tends to make the derived
probability distribution of climate change wider (and vice versa).
In most cases, the effect is not dramatic.