Download Using Model Output: Uncertainties and Probabilities

Using Model Output: Uncertainties and
Probabilities
Comparing model results to observations can only
take
us so far - how do we evaluate the robustness of
future climate change simulations?
Two key aspects:
1) What are the uncertainties? How do they manifest
and propagate?
2) Can we estimate probabilities of occurrence? How
robust are these PDF
PDF’s?
s?
Uncertainty Propagation
•
How do uncertainties in the large-scale GCM forcing propagate
into the much smaller scale RCM?
•
y
Over time RCM solution turns from initial value to boundary
value problem; this implies nested models represent an ill-posed
mathematical boundary value problem (inconsistencies between
large and small scale)
•
Strong scale dependency - flow regime within RCM may
become inconsistent with large-scale
g
flow
•
Realism of large-scale flow is extremely important!
Dealing with Uncertainty Propagation
•
Evaluate spread in GCM ensembles: Is the regional response
consistent across the ensemble members? Or do they vary in differing
ways? What about systematic biases?
•
Perform multi-decadal, ensemble RCM runs - allows for exploration of
uncertainties in mean state and higher-order statistics (but also may
constrain spatial resolution; depending on computational resources)
•
Nudging of the RCM solution can help
help, but can also hide biases - can
also be problematic when nudging back to GCM solution (as opposed
to observations) - this can just accentuate GCM biases
Quantifying Uncertainty: Issues
•
Overall assessment from IPCC: ‘Most sources of uncertainty at regional
scales are similar to those at the g
global scale’
•
Two exceptions:
p
1)) inhomogeneities
g
in land use cover and land use
changes; 2) aerosol forcing
•
Uncertainties are not uniform across climate parameters; e.g., generally
larger and more poorly posed for precipitation than for temperature smaller signal to noise ratios
•
Different parameterizations of sub-grid processes yield different results;
this is also scale
scale-dependent
dependent
Quantifying Uncertainty: Techniques
•
Use of multi-model ensembles to generate quantitative
measures of spread across the model simulations (intra(intra and
inter-model)
Not a very robust approach,
approach in particular not a representative
sample for generating PDFs Does not provide unbiased
sample of the phase space of possible change
•
One possible modification - weight model results according to
biases in control simulations (compared to observed climate) Employ robust observational constraints
•
Temperature anomalies with respect to 1901 to 1950 for six
continental-scale regions for 1906 to 2005 (black line) and as simulated
(red envelope) by MMD models incorporating known forcings; and as
projected
j t d ffor 2001 tto 2100 b
by MMD models
d l ffor th
the A1B scenario
i
(orange envelope). The bars at the end of the orange envelope
represent the range of projected changes for 2091 to 2100 for the B1
scenario ((blue),
), the A1B scenario (orange)
(
g ) and the A2 scenario (red).
( )
The black line is dashed where observations are present for less than
50% of the area in the decade concerned.
Combined Uncertainties
•
Another technique uses the concept of ‘combined uncertainties’
•
The emission scenarios that estimate future GHG forcing; the GCM
that simulates the large-scale response to the GHG forcing; and the
regional downscaling techniques each have their own uncertainties that
must be combined in some manner
•
Myriad of possible methodologies; key aspect - are the uncertainties
linearly additive,
additive or do they have strong non-linear interactions?
•
A number of projects addressing this approach: PRUDENCE,
NARCCAP CREAS
NARCCAP,
CREAS, CLARIS
•
Temperature anomalies with respect to 1901 to 1950 for three
Central and South American land regions for 1906 to 2005
(black line) and as simulated (red envelope) by MMD models
incorporating known forcings; and as projected for 2001 to 2100
by MMD models for the A1B scenario (orange envelope). The
bars at the end of the orange envelope represent the range of
projected
p
j
changes
g for 2091 to 2100 for the B1 scenario ((blue),
),
the A1B scenario (orange) and the A2 scenario (red). The black
line is dashed where observations are present for less than 50%
of the area in the decade concerned.
•
MMD ensemble annual mean surface air temperatures in South
America compared with observations. a) observations from the
HadCRUT2v data set (Jones et al., 2001); b) mean of the 21 MMD
models;
d l c)) diff
difference b
between
t
th
the multi-model
lti
d l mean and
d th
the
HadCRUT2v data. Units ｰC.
•
As above,
above but for precipitation
precipitation. Observations (CMAP) are an update of
Xie and Arkin (1997). Units mm/day.
Regional Model Comparison
Projects
PRUDENCE
NARCCAP
CREAS
CLARIS
PRUDENCE
(Prediction of Regional scenarios and Uncertainties for Defining EuropeaN
Climate change risks and Effects)
•
PRUDENCE will provide high-resolution climate change
scenarios for 2071-2100
2071 2100 for Europe using dynamical
downscaling methods (i.e., regional climate modelling).
•
The variability and level of confidence in these scenarios will be
characterised as a function of uncertainties in model
formulation, natural/internal climate variability, and alternative
scenarios
sce
a os o
of future
utu e at
atmospheric
osp e c co
composition.
pos t o
•
These scenarios will be used to explore changes in the
frequency and magnitude of extreme weather events
events.
Annual change over Europe
in the precipitation amount
greater than 20mm/day
(2071-2100 vs. 1961-1990)
for the A2 scenario simulated
with the regional climate
model HIRHAM nested in the
global ECHAM4/OPYC3
model. Green colours indicate
an increase, yellow little
change. Orange colours over
the globe indicate relative
temperature change
NARCCAP
(North American Regional Climate Change Assessment Program)
•
NARCCAP is an international program to produce high resolution climate
change simulations in order to investigate uncertainties in regional scale
projections of future climate and generate climate change scenarios for use in
impacts research.
•
The AOGCMs have been forced with the SRES A2 emissions scenario for the
21st century.
Simulations with these models were also produced for the current (historical)
period. The RCMs are nested within the AOGCMs for the current period 19712000 and for the future period 2041-2070.
•
As a preliminary step to evaluate the performance of the RCMs over North
America, the RCMS are driven with NCEP Reanalysis II data for the period
1979-2004. All the RCMs are run at a spatial resolution of 50 km.
NARCCAP Time Slices
•
NARCCAP also includes two timeslice experiments at 50 km resolution
using the GFDL atmospheric model (AM2.1) and the NCAR CCSM
atmospheric model (CAM3).
•
In a timeslice experiment, the atmospheric component of an AOGCM is
run using observed sea surface temperatures and sea ice boundaries
for the historical run
run, and those same observations combined with
perturbations from the future AOGCM for the scenario run.
•
Omitting the coupled ocean model saves considerable computation and
allows the atmospheric model to be run at higher resolution
• NEXT FIGURE - - SOME TYPICAL
NARCCAP RESULTS
NARCCAP 50 KM domain
CREAS
(Regional Climate Change Scenarios for South America)
•
High resolution scenarios in South America for raising awareness
among government officials and policy-makers
•
Uses three regional climate models: two from the UK Hadley Center,
and a version of the NCEP Eta model
•
Focus on A2 and B2 emission scenarios
•
UKMet HADCM3 Global GCM provides large-scale climate change
f i
forcing
•
50 KM resolution for regional models
CLARIS
A Europe-South
p
America Network for Climate Change
g Assessment and Impact
p
Studies
•
The CLARIS project aims at strengthening collaborations between
Europe and South America to develop common research strategies on
climate change and impact issues in the subtropical region of South
America through a multi-scale integrated approach (continentalregional-local).
•
The CLARIS framework will facilitate the participation of European
researchers to IAI (Inter American Institute) projects and the
submission of new common research proposals. Moreover, its opening
towards stakeholders (e.g. agriculture, reinsurance, hydroelectricity),
associated to the project through an expert group, will promote future
initiatives on climate impact analysis, thus, contributing to related
sustainable development strategies.
The Flip-Side of Uncertainty is
Probability - Can We Assess
and Quantify
y Likelihood of
Occurrence
Constructing and Evaluating
Probabilities
•
No single model is ‘the best’ - a multi-model approach should be used
•
This leads to the possibility of probabilistic assessments of climate
projections from diverse models
•
Problem of enormous complexity due to high dimensionality of climate
model, plus sparse and limited length of observations
Common approach to minimize problems - use regional averages
g of regional
g
assessments and
But does this constrain the range
potential effects?
Two Major Issues
• Poor convergence (heavy-tailed probability
distributions) – tends to minimize probabilities of tails
and exaggerate in middle
• MODEL BIAS!!! Bias must be minimized or
eliminated – one cannot make a robust pdf from a
biased dataset – leads to Bayesian approaches
Probabilities, cont.
Differing approaches can yield very different results and interpretations
•
Tebali et al. (2004): Bayesian framework; uses model bias and
convergence criteria to define the shape and width of pdf’s
pdf s for
temperature and precipitation
•
Incorporates expert judgment in the form of how much weight to place
on each criteria; assessed over regional averages
•
Greene et al. (2006): Bayesian framework based on extension of
methods used for ensemble seasonal forecasting
•
Calibrated for the historic period (1902-1998) and then applied to future
projections
•
Two reasons differs from Tebaldi et al.:
•
1) Att
Attributes
ib t large
l
uncertainty
t i t tto models
d l th
thatt poorly
l representt historical
hi t i l
climate trends
•
2) Strong stationarity assumption required to extrapolate from historic
record to future trends
•
Presumed responsible for the smaller warmings and wider PDF’s
(i l di a ffew negative
(including
ti values)
l
)
•
Tends to ‘highlight’ low latitude regions due to weak trends in
observations, and/or worse model performance
• PDFS NEXT FIGURE - MULTI-MODALITY;
LONG TAILS
• BIAS AND CONVERGENCE
•
Map comparing PDFs of change in temperature (2080 to 2099
compared to 1980 to 1999) from Tebaldi et al. (2004a,b) and Greene et
al (2006) as well as the raw model projections (represented by shaded
al.
histograms) for the Giorgi and Francisco (2000) regions. Areas under
the curves and areas covered by the histograms have been scaled to
equal unity. The scenario is SRES A1B and the season is NH winter
(
(DJF).
)
Other Techniques for Constructing
PDFs
•
Apply simple pattern scaling to multi-model GCM ensembles
•
Modulate (normalized) regional patterns of change by global mean
temperature changes obtained from a simple
simple, probablistic EBM (e
(e.g.,
g
MAGICC)
•
Perturbed physics ensembles - allows for characterization of
uncertainties arising from poorly constrained model parameters
•
Problems worse for climate parameters with nonlinear response (e.g.,
precipitation) - not so bad for those with linear response (e
(e.g.,
g surface
temperature)
Techniques,
q
continued
• Use robust observational constraints on model simulations
- this plays off observed historic changes over the region of
interest versus global changes
• Use expert understanding of relevant processes and how
they should unfold rather than the probabilistic methods
(which are in their infancy and do not yet provide
definitive results)
• Greater use of multi-variate ‘process-oriented’ analyses
based on (presumed) key mechanisms
•
Robust findings on regional climate change for mean and
extreme precipitation, drought, and snow. This regional
assessment is based upon AOGCM based studies, Regional
Climate Models, statistical downscaling and process
understanding. More detail on these findings may be found in
the notes below, and their full description, including sources is
given in the text. The background map indicates the degree of
consistency between AR4 AOGCM simulations (21 simulations
used) in the direction of simulated precipitation change.
•
Results from the perturbed physics ensemble of Harris et al.
(2006) showing evolution in the median, and 80%, 90%, and
95% confidence ranges for annual surface temperature change,
for a 1% per annum increase in CO2 concentration for 150
years, for all 24 regions described by Giorgi and Francisco
(2000).
Probabilistic Thresholds
•
Reliabilityy Ensemble Averaging
g g ((REA)) - used byy Giorgi
g and Mearns
(2002) to calculate the probability of regional climate change exceeding
given thresholds based on ensembles of different model simulations.
•
Probabilities of surface air temperature and precipitation change
calculated for 10 regions of subcontinental scale spanning a range of
latitudes and climatic settings.
•
They conclude the REA method can provide a simple and flexible tool
to estimate probabilities of regional climate change from ensembles of
model simulations for use in risk and cost assessment studies.
•
Limited sensitivity analysis of regional climate change probabilities for
the 21st centuryy
•
Dessai et al. examined the sensitivityy of regional
g
climate change
g
probabilities to various uncertainties.
•
Used a simple probabilistic energy balance model that sampled
uncertainty in greenhouse gas emissions, the climate sensitivity, the
carbon cycle, ocean mixing, and aerosol forcing.
•
Propagated global mean temperature probabilities to General Circulation Models
(GCMs) through the pattern-scaling technique.
•
Devised regional skill scores for each GCM, season (DJF, JJA), and climate
variable (surface temperature, and precipitation) in 22 world regions, based on
model performance and model convergence.
•
A range of sensitivity experiments carried out with different skill score schemes,
climate sensitivities, and emissions scenarios.
•
For temperature change probabilities,
probabilities emissions scenarios uncertainty tends to
dominate the 95th percentile whereas climate sensitivity uncertainty plays a
more important role at the 5th percentile.
•
The sensitivity of precipitation change probabilities to the tested uncertainties
are region specific, but some conclusions can be drawn.
At the 95th percentile, the uncertainty that tends to dominate is emissions
scenarios closely followed by GCM weighting scheme and the climate
scenarios,
sensitivity.
At the 5th percentile, GCM weighting scheme uncertainty tends to dominate for
JJA but for DJF all uncertainties have similar proportionate influence
JJA,
influence.
•
•
Development of probability distributions for regional climate change from uncertain global
mean warming and an uncertain scaling relationship
B. Hingray, A. Mezghani, T.A. Buishand (KNMI),
•
To produce probability distributions for regional climate change in surface temperature
and precipitation, a probability distribution for global mean temperature increase has been
combined with the probability distributions for the appropriate scaling variables, i.e. the
changes in regional temperature/precipitation per degree global mean warming.
•
Each scaling variable is assumed to be normally distributed.
•
The uncertainty of the scaling relationship arises from systematic differences between the
regional changes from global and regional climate model simulations and from natural
variability.
•
The contributions of these sources of uncertainty to the total variance of the scaling
variable are estimated from simulated temperature and precipitation data in a suite of
regional climate model experiments conducted within the framework of the EU-funded
project PRUDENCE, using an Analysis Of Variance (ANOVA).
•
For the area covered in the 2001–2004 EU-funded project SWURVE, five case study
regions (CSRs) are considered: NW England, the Rhine basin, Iberia, Jura lakes
(Switzerland) and Mauvoisin dam (Switzerland).
•
The resulting regional climate changes for 2070–2099 vary quite significantly between CSRs,
between seasons and between meteorological variables.
•
For all CSRs, the expected warming in summer is higher than that expected for the other seasons.
This summer warming is accompanied by a large decrease in precipitation.
•
The uncertainty of the scaling ratios for temperature and precipitation is relatively large in summer
because of the differences between regional climate models
models.
•
Differences between the spatial climate-change patterns of global climate model simulations make
significant contributions to the uncertainty of the scaling ratio for temperature.
•
However, no meaningful contribution could be found for the scaling ratio for precipitation due to the
small number of global climate models in the PRUDENCE project and natural variability, which is
often the largest source of uncertainty.
•
In contrast,
contrast for temperature
temperature, the contribution of natural variability to the total variance of the scaling
ratio is small, in particular for the annual mean values.
•
Simulation from the probability distributions of global mean warming and the scaling ratio results in a
wider range of regional temperature change than that in the regional climate model experiments.
•
For the regional change in precipitation, however, a large proportion of the simulations (about 90%) is
within the range of the regional climate model simulations.