Download Extremes (M. Wehner, LLNL and UC Berkeley)

Extreme weather in a changing climate
CMIP Tutorial
NCAR
August 19, 2016
Michael F. Wehner
Lawrence Berkeley National Laboratory
[email protected]
Extreme Weather in a changing climate: Storms
Extreme Weather in a changing climate: Impacts
Boulder Creek
(behind the hotel)
Folsom Lake, California
“snowmageddon”
Texas
Outline of my talk
• Simple metric of extremes
– ETCDDI indices
• Extreme Value Theory results
– Block Maxima
– Peaks over Threshold.
• Storm Tracking
• Are CMIP5 models fit for purpose?
– Projections
– Detection and Attribution
– Individual Extreme Event attribution
Expert Team on Climate Change Detection Indices
(ETCCDI)
• 27 indices intended for detection and attribution of trends
• Tailored to available observational data
– often spatially sparse.
• Not descriptive of very rare events.
– Some are not really extreme at all.
• Gridded observations are available here:
– http://www.climdex.org/datasets.html
• Model results and code to calculate are available here:
– http://etccdi.pacificclimate.org/
Useful ETCCDI temperature indices
Txx=annual maximum hottest daytime temperature (hot days)
Tnn=annual minimum coldest nighttime temperature (cold nights)
Txn=annual maximum coldest nighttime temperature (warm nights)
Tnx=annual minimum hottest daytime temperature (cool days)
Useful ETCCDI temperature indices
Txx=annual maximum hottest daytime temperature (hot days)
Tnn=annual minimum coldest nighttime temperature (cold nights)
Txn=annual maximum coldest nighttime temperature (warm nights)
Tnx=annual minimum hottest daytime temperature (cool days)
End of century RCP8.5 changes. Hot days/cold nights
Useful ETCCDI temperature indices
rx1day=annual maximum daily precipitation
rx5day=annual maximum pentad precipitation
cdd=consecutive dry days
Useful ETCCDI temperature indices
rx1day=seasonal maximum daily precipitation
rx5day=seasonal maximum pentad precipitation
cdd=consecutive dry days
• Pentad precipitation is preferred over daily.
• Big storms often last longer than a day, but rarely more than 5
• 00:00 GMT may not be a convenient time
• Seasonal is preferable over annual
• Winter storms are different than summer storms!
• Consecutive dry days is not really extreme.
• A interesting measure of the dry season in some locales.
Extreme value statistics
See Dan Cooley’s talk for more details.
Sampling for extremes
Block maxima Generalized Extreme Value (GEV) distribution
Many of the ETCCDI indices are block maxima.
Peaks over Threshold Generalized Pareto Distribution (GPD)
Sample from a parent high frequency data set.
.
• Interesting quantities
• Return value for a fixed period change in magnitude for fixed rarity.
• Return period for a fixed value change in frequency for a fixed threshold.
Extreme value statistics: some guidance
• A “natural” block size is a season.
• May be too short in dry regions for precipitation.
• Winter is different than summer. Annualizing can lead to
mixing the underlying seasonal distributions. Not a good
thing.
• Thresholds.
• Arbitrary. 85%, 90%, 95% 99%?
• A balance between being in the tail vs. large enough
datasets.
• Covariates are extremely useful.
• EV theory generally presumes stationarity but climate
change is not
• Covariates can remove that restriction by adding more fitting
parameters
• Useful covariates (are usually linear):
– CO2 (anthropogenic climate change)
– ENSO, NAO, PDO (natural modes of variability)
– Blocking
– This is new. Experience is limited.
Source: wikipedia
Are models “fit for purpose” for extremes?
Seasonal high temperatures
Angelil et al. (2016) J. Clim
2013 was an extreme year for Australian temperatures.
The distribution of this model (CAM5.1) is comparable to two different observations.
Part of an attribution study. Yes, there is a big human influence.
Are models “fit for purpose” for extremes?
20 year return
value of the winter
daily precipitation
(mm/day)
CMIP5 class
models are not high
enough resolution.
Picture is much
worse for summer
due to defects in
convective
parameterizations.
What does the IPCC have to say about temperature extremes?
• AR5 ES:
– It is virtually certain that, in most places, there
will be more hot and fewer cold temperature
extremes as global mean temperatures
increase
– Under RCP8.5 it is likely that, in most land
regions, a current 20-year high temperature
event will occur more frequently by the end of
the 21st century (at least doubling its
frequency, but in many regions becoming an
annual or two-year event) and a current 20year low temperature event will become
exceedingly rare.
IPCC AR5 Figure 12.14
• Changes in 20 year
return values of the
annual hottest and
coldest day.
• Used time dependent
GEV statistics.
• “Today’s rare hot events
become commonplace”
• Cold extremes increase
more than hot extremes.
What does the IPCC have to say about precipitation extremes?
• AR5 ES:
– Globally, for short-duration precipitation events, a shift to
more intense individual storms and fewer weak storms is
likely as temperatures increase.
– Regional to global-scale projected decreases in soil
moisture and increased risk of agricultural drought are likely
in presently dry regions and are projected with medium
confidence by the end of this century under the RCP8.5
scenario.
Models were judged to not be of high enough quality to justify
presenting changes in long period return values.
But there was pressure to present something…
IPCC AR5 Figure 12.27: Rare precipitation extremes
• CMIP5 models are limited by their coarse resolution to reproduce severe storms
• Return value and period expressed as the amount per oC local temperature
increase rather than give the scenario projections
– Results consistent with Clausius-Clapeyron scaling.
• A deliberate choice to downplay the attention paid to this figure.
3rd US National Climate Assessment
• “Frequency change multiplier” = “Risk Ratio”
= RP2000 / RP2100
High resolution global climate modeling
• High resolution is required to accurately simulate intense storms.
– CMIP6 subproject HIRESMIP will provide 5-10 models in the 25km range.
– Extremely computationally expensive.
– Database is much larger. Sub-daily becomes much more interesting!
• ~25km global Community Atmospheric Model (CAM5.1)
– Able to simulate hurricanes up to Category 5.
– Far superior extreme precipitation statistics.
– Thanks to the large US DOE investment in high performance computing.
•
Was 2 wall-clock days per simulated year.
•
Now is 1 day/year.
High resolution global climate modeling
Tropical
Storm
Cat1
Cat2
Cat3
Cat4
Cat5
Figures by Prabhat
Cat1
Cat2
Cat3
Cat4
Cat5
Figures by Prabhat
Cat4
Cat5
Figures by Prabhat
Storm Tracking
• With the advent of tropical cyclone permitting models, extreme storms are
much more interesting.
• In order say anything about storms, you have to find them first.
• Large datasets. Computationally expensive
The Toolkit for Extreme Climate Analysis (TECA2.0) is a general purpose tool
built upon the MapReduce algorithm to efficiently track storms contained in
very high resolution simulation in a highly parallel manner.
• Currently tropical cyclone tracking is ready for release
• Scaled to 750000 processors. 10000 is more typical.
• Reduces time to solution from weeks to hours
• Extra-tropical cyclone (ETC) tracking is beta
• Atmospheric River (AR) identification is pre-beta
Supervised machine learning has been demonstrated
as a non-parametric storm tracker (frontal systems).
Tropical Cyclones in a warmer world
Do CMIP5-class models have any value in projections of future
tropical cyclone frequency?
All storms
1o
0.25o
1o fvCAM5 makes more tropical cyclones in the warmer world
0.25o fvCAM5 make fewer tropical cyclones in the warmer world but more intense TCs.
NO! Not by direct tracking anyways.
Limited by tracking of weak storms.
Tropical Cyclones in a warmer world
Do the bulk characteristics of CMIP5-class models have value in projections of
future tropical cyclone statistics? Part 1. Intense Hurricanes
Emanuel’s Maximum Potential Intensity (MPI)
Direct tracking of high resolution models in a warmer world indicates the strongest
storms get stronger consistent with DMPI.
DMPI is quite similar between coarse and high resolution
Yes, CMIP5-class models can predict the behavior of TC relevant fields.
Tropical Cyclones in a warmer world
Do the bulk characteristics of CMIP5-class models have value in projections of
future tropical cyclone statistics? Part 2. Cyclogenesis
•Hi-res model produces fewer TC in a
warmer world.
Emanuel’s Genesis Potential Index (GPI)
•GPI correctly reflects the change in
the bulk cyclogenesis-relevant fields.
•But GPI is missing any information
about changes in short term instabilities.
•The warmer worlds are more stable.
•Fewer opportunities for TC to be
realized even though cyclogenesis
potential is more
favorable.
No, GPI fails.
Actual tracks
Storm tracking in CMIP5/6
Data transfer of sub-daily data is a serious problem.
Case study: Track ETC in all historical, rcp8.5 and piControl CMIP5 experiments
160TB of model output. It took a network professional 3 months to download it.
This will improve when
Globus is fully enabled.
Avoid wget if at all possible.
HIRESMIP data for TC tracking ~10TB/decade/model
Storm tracking in CMIP5/6
Data transfer of sub-daily is a serious problem.
Case study: Track ETC in all historical, rcp8.5 and piControl experiments
160TB of model output. It took a network professional 3 months to download it.
Netflix to my house 
.
Extreme Event Attribution
• When extreme weather happens, the public wants to know
– “Is this climate change?”
• Not quite the correct question, better to ask:
– “How has the risk of this event changed because of climate change?”
Or
– “How did climate change affect the magnitude of this event?”
• We approach these questions in two different ways.
1. Use extreme value statistics and the existing CMIP5 “ensemble of
opportunity”.
2. Design our own ensembles of climate model simulations tailored to event
attribution.
– Climate of the 20th Century C20C (~50-400 ensemble members)
– climateprediction.net (~1000+ ensemble members)
Extreme event attribution: CMIP statistical analysis
• Consider three different summer heat wave events
– Europe 2003 (~70,000 excess deaths)
– Russia 2010 (~50,000 excess deaths, massive fires)
– Texas 2011 (lots of dead cows, massive drought, $8BN to ag industry)
• These are very rare events. We are interested in how the rarity of these
events has changed.
• We calculated the change in risk by comparing the extreme value
statistics in these regions from realistic historical simulations to those in
the pre-industrial simulations and the observations.
Pardeep Pall, Michael Wehner, Dáithí Stone (2014) Probabilistic Extreme Event Attribution in Dynamics and
Predictability of Large-Scale, High-Impact Weather and Climate Events, Richard Grojahn, Jianping Li, Richard
Swinbank, Hans Volkert, editors. Cambridge University Press. 37-46, ISBN 978-1-107-07142-1.
Fractional attributable risk (FAR)
• Real World: with industrialized humans
• Not Real World: without industrialized humans
Risk _ ratio =
Preal
PNot _ real
PNot _ real
FAR= 1 Preal
• Fractional attributable risk is often used to determine liability.
Extreme event attribution: CMIP statistical analysis
The risk of each of these events has least doubled since the preindustrial era
Event
Risk Change at
time of event
Change in risk
2023
Change in risk
2040
Europe 2003
~2X
35X
154X
Russia 2010
2-3 X
2.5-4 X
5-8 X
Texas 2011
1.5-4 X
2-5 X
4-10 X
Extreme event attribution: other datasets
• CMIP provides prescribed emissions or prescribed concentration scenarios.
• Fully coupled atmosphere, ocean, land, ice models
•
A more constrained scenario is “prescribed temperature”
• AMIP is the traditional case where sea surface temperature and ice
concentration are prescribed by historical estimates.
• C20C+ is a factual/counterfactual experiment designed for extreme event
attribution.
– Counterfactual world preserves observed interannual variations but
anthropogenic changes are removed.
– 5-10 models. Ensemble sizes from 50 to 400 over the period 1959-2015
– http://portal.nersc.gov/c20c/
• HAPPI1.5/HAPPI2.0
• Half a degree Additional warming, Projections, Prognosis and Impacts
• Stabilized 1.5 & 2K above preindustrial prescribed SSTs
• In response to the Paris COP21 agreement
• Currently underway.
Multivariate extremes
• Real world systems are multivariate.
• Impacts can depend on the combinations
– Hot, dry, windy  wildfires
– Hot, moist, stagnant  human health
• Past and future statistics also depends on the combinations.
– Mechanisms of changes vary.
• Our project brings climate analysts, impacts scientists and statisticians
together
– Impacts scientists help us define what is “extreme”.
– Statisticians are developing non-asymptotic methods.
– Climate analysts design targeted numerical experiments
Useful tools
Read this book!
Coles, S. (2001), An Introduction to Statistical Modeling of Extreme Values,
Springer, London.
ExtRemes: R based package with many useful EV functions so you don’t have to
code up Coles(2001).
http://www.ral.ucar.edu/~ericg/extRemes/
climextRemes: A python based subset of ExtRemes useful for non-stationary
problems of interest to me. (projections & attributions)
https://cran.r-project.org/web/packages/climextRemes/index.html
TECA2.0. Parallel storm tracking tool (soon to be released to the public). With
python and c++ interfaces
First release. Tropical Storms
Later: Extratropical Cyclones, Atmospheric Rivers, Jet Stream
Conclusions
• A wide variety of research into extreme weather is enabled by CMIP5.
• CMIP6 will be considerably better because of the HIRESMIP.
• Biases in extreme temperature and precipitation are generally much worse than
for seasonal means.
• Some extreme weather phenomena should not be analyzed with CMIP5-class
models.
• Model evaluation is critical. Are the models “fit for purpose”?
• Analysis of seasonal extremes is limited by the smaller ensembles of CMIP5
• Large CESM ensemble is of use.
• Resist the temptation to annualize data.
• Daily and subdaily dataset sizes are challenging.
• Most analyses are embarrassingly parallel in some dimension.
• Parallel processing can substantially increase throughput.
• I/O can be a bottleneck in parallel codes.
Thank you!
[email protected]