Page 1© Crown copyright 2004 Development of probabilistic climate predictions for UKCIP08 David Sexton, James Murphy, Mat Collins, Geoff Jenkins, Glen.

Slides:



Advertisements
Similar presentations
RT1 Development of the Ensemble Prediction System Aim Build and test an ensemble prediction system based on global Earth System models developed in Europe,
Advertisements

The new German project KLIWEX-MED: Changes in weather and climate extremes in the Mediterranean basin Andreas Paxian, University of Würzburg MedCLIVAR.
© UKCIP 2006 © Crown copyright Met Office Probabilistic climate projections from the decadal to centennial time scale WCRP Workshop on Regional Climate,
Climate changes in Southern Africa; downscaling future (IPCC) projections Olivier Crespo Thanks to M. Tadross Climate Systems Analysis Group University.
CMIP5: Overview of the Coupled Model Intercomparison Project Phase 5
Februar 2003 Workshop Kopenhagen1 Assessing the uncertainties in regional climate predictions of the 20 th and 21 th century Andreas Hense Meteorologisches.
Theme D: Model Processes, Errors and Inadequacies Mat Collins, College of Engineering, Mathematics and Physical Sciences, University of Exeter and Met.
© Crown copyright Met Office ACRE working group 2: downscaling David Hein and Richard Jones Research funded by.
Scaling Laws, Scale Invariance, and Climate Prediction
What is the point of this session? To use the UK’s experience to give ideas about creating and using climate change scenarios in other countries and situations.
A statistical method for calculating the impact of climate change on future air quality over the Northeast United States. Collaborators: Cynthia Lin, Katharine.
© Crown copyright Met Office Regional/local climate projections: present ability and future plans Research funded by Richard Jones: WCRP workshop on regional.
Applying probabilistic scenarios to environmental management and resource assessment Rob Wilby Climate Change Science Manager
Climate case study. Outline The challenge The simulator The data Definitions and conventions Elicitation Expert beliefs about climate parameters Expert.
Earth Systems Science Chapter 6 I. Modeling the Atmosphere-Ocean System 1.Statistical vs physical models; analytical vs numerical models; equilibrium vs.
Progress in Downscaling Climate Change Scenarios in Idaho Brandon C. Moore.
16 March 2011 | Peter Janssen & Arthur Petersen Model structure uncertainty A matter of (Bayesian) belief?
Uncertainty and Climate Change Dealing with uncertainty in climate change impacts Daniel J. Vimont Atmospheric and Oceanic Sciences Department Center for.
WFM 6311: Climate Risk Management © Dr. Akm Saiful Islam WFM 6311: Climate Change Risk Management Akm Saiful Islam Lecture-4: Module- 3 Regional Climate.
Introduction to Numerical Weather Prediction and Ensemble Weather Forecasting Tom Hamill NOAA-CIRES Climate Diagnostics Center Boulder, Colorado USA.
Development of a combined crop and climate forecasting system Tim Wheeler and Andrew Challinor Crops and Climate Group.
Water Management Presentations Summary Determine climate and weather extremes that are crucial in resource management and policy making Precipitation extremes.
NCPP – needs, process components, structure of scientific climate impacts study approach, etc.
The Scenarios Network for Alaska and Arctic Planning is a collaborative network of the University of Alaska, state, federal, and local agencies, NGOs,
Page 1GMES - ENSEMBLES 2008 ENSEMBLES. Page 2GMES - ENSEMBLES 2008 The ENSEMBLES Project  Began 4 years ago, will end in December 2009  Supported by.
OUCE Oxford University Centre for the Environment “Applying probabilistic climate change information to strategic resource assessment and planning” Funded.
ESA DA Projects Progress Meeting 2University of Reading Advanced Data Assimilation Methods WP2.1 Perform (ensemble) experiments to quantify model errors.
© Crown copyright Met Office Climate Projections for West Africa Andrew Hartley, Met Office: PARCC national workshop on climate information and species.
Applications of Bayesian sensitivity and uncertainty analysis to the statistical analysis of computer simulators for carbon dynamics Marc Kennedy Clive.
WP4.1: Feedbacks and climate surprises ( IPSL, HC, LGGE, CNRM, UCL, NERSC) WP4.1 has two main objectives (a) to quantify the role of different feedbacks.
© Crown copyright Met Office Using a perturbed physics ensemble to make probabilistic climate projections for the UK Isaac Newton workshop, Exeter David.
Towards determining ‘reliable’ 21st century precipitation and temperature change signal from IPCC CMIP3 climate simulations Abha Sood Brett Mullan, Stephen.
EFFICIENT CHARACTERIZATION OF UNCERTAINTY IN CONTROL STRATEGY IMPACT PREDICTIONS EFFICIENT CHARACTERIZATION OF UNCERTAINTY IN CONTROL STRATEGY IMPACT PREDICTIONS.
Downscaling and its limitation on climate change impact assessments Sepo Hachigonta University of Cape Town South Africa “Building Food Security in the.
© Crown copyright Met Office Providing High-Resolution Regional Climates for Vulnerability Assessment and Adaptation Planning Joseph Intsiful, African.
Recent Advances in Climate Extremes Science AVOID 2 FCO-Roshydromet workshop, Moscow, 19 th March 2015 Simon Brown, Met Office Hadley Centre.
World Climate Research Programme Climate Information for Decision Making Ghassem R. Asrar Director, WCRP.
Model dependence and an idea for post- processing multi-model ensembles Craig H. Bishop Naval Research Laboratory, Monterey, CA, USA Gab Abramowitz Climate.
Why it is good to be uncertain ? Martin Wattenbach, Pia Gottschalk, Markus Reichstein, Dario Papale, Jagadeesh Yeluripati, Astley Hastings, Marcel van.
The Tyndall Centre comprises nine UK research institutions. It is funded by three Research Councils - NERC, EPSRC and ESRC – and receives additional support.
Center for Radiative Shock Hydrodynamics Fall 2011 Review Assessment of predictive capability Derek Bingham 1.
Status of the Sea Ice Model Testing of CICE4.0 in the coupled model context is underway Includes numerous SE improvements, improved ridging formulation,
(Mt/Ag/EnSc/EnSt 404/504 - Global Change) Climate Models (from IPCC WG-I, Chapter 10) Projected Future Changes Primary Source: IPCC WG-I Chapter 10 - Global.
Ben Kirtman University of Miami-RSMAS Disentangling the Link Between Weather and Climate.
The Tyndall Centre comprises nine UK research institutions. It is funded by three Research Councils - NERC, EPSRC and ESRC – and receives additional support.
Experiences in assessing deposition model uncertainty and the consequences for policy application Rognvald I Smith Centre for Ecology and Hydrology, Edinburgh.
Development of Climate Change Scenarios of Rainfall and Temperature over the Indian region Potential Impacts: Water Resources Water Resources Agriculture.
The evolution of climate modeling Kevin Hennessy on behalf of CSIRO & the Bureau of Meteorology Tuesday 30 th September 2003 Canberra Short course & Climate.
1 Isaac Newton Workshop on Probabilistic Climate Prediction University of Exeter Sep 2010 Professor David B. Stephenson Exeter Climate Systems Mathematics.
Working with climate model ensembles PRECIS workshop Tanzania Meteorological Agency, 29 th June – 3 rd July 2015.
© Crown copyright Met Office Downscaling ability of the HadRM3P model over North America Wilfran Moufouma-Okia and Richard Jones.
© Crown copyright Met Office Uncertainties in the Development of Climate Scenarios Climate Data Analysis for Crop Modelling workshop Kasetsart University,
Reducing the risk of volcanic ash to aviation Natalie Harvey, Helen Dacre (Reading) Helen Webster, David Thomson, Mike Cooke (Met Office) Nathan Huntley.
© Crown copyright Met Office Working with climate model ensembles PRECIS workshop, MMD, KL, November 2012.
The Tyndall Centre comprises nine UK research institutions. It is funded by three Research Councils - NERC, EPSRC and ESRC – and receives additional support.
Presented by LCF Climate Science Computational End Station James B. White III (Trey) Scientific Computing National Center for Computational Sciences Oak.
Downscaling Global Climate Model Forecasts by Using Neural Networks Mark Bailey, Becca Latto, Dr. Nabin Malakar, Dr. Barry Gross, Pedro Placido The City.
Using the past to constrain the future: how the palaeorecord can improve estimates of global warming 大氣所碩一 闕珮羽 Tamsin L. Edwards.
Page 1© Crown copyright 2004 Data Assimilation at the Met Office Hadley Centre, Met Office, Exeter.CTCD Workshop. 8 th Nov, 2005 Chris Jones.
NOAA Northeast Regional Climate Center Dr. Lee Tryhorn NOAA Climate Literacy Workshop April 2010 NOAA Northeast Regional Climate.
Introduction to emulators Tony O’Hagan University of Sheffield.
RAL, 2012, May 11 Research behaviour Martin Juckes, 11 May, 2012.
Ruth Doherty, Edinburgh University Adam Butler & Glenn Marion, BioSS
GFDL Climate Model Status and Plans for Product Generation
Presentation transcript:

Page 1© Crown copyright 2004 Development of probabilistic climate predictions for UKCIP08 David Sexton, James Murphy, Mat Collins, Geoff Jenkins, Glen Harris, Kate Brown, Robin Clark, Penny Boorman, Simon Brown, Richard Jones, Jason Lowe, Ben Booth, B. Bhaskaran, David Hassell, Ruth McDonald, Tom Howard, Lizzie Kennett UEA, October 19, 2007

Page 2© Crown copyright 2004 Content  UKCIP08  Probabilistic climate prediction system  Modelling uncertainty and perturbed physics ensembles  Weighting with observations  Time Scaling  Other components of Earth System  Downscaling  Assumptions

Page 3© Crown copyright 2004 UKCIP ‘02  Based on the state-of-the-art at the time - HadCM3, HadAM3H time-slice, 50km HadRM3 experiments  Used by many private and public-sector organisations to make decisions and spend money  “Scenario” based with no quantification of uncertainties (although plenty of caveats pointing this out)

Page 4© Crown copyright 2004 Emission scenarios Effects of internal variability Modelling of Earth system processes Uncertainties in model projections … which includes how informative are models about reality

Page 5© Crown copyright 2004 Modelling uncertainty  Set of international climate models are all ‘tuned’ to observations  But there is no guarantee these are the actual optimal models  Other choices of values for model input parameters could have provided equally plausible simulations of observations whilst providing a wide range of responses in the future  So tuning could affect the decisions planners make based on climate predictions

Page 6© Crown copyright 2004 UKCIP08 – Probabilistic predictions  To provide joint probability distribution functions (pdfs) of predicted changes in a selection of key UK climate variables at 25km resolution for , ,…,  Results will be presented for each variable by month  We aim to deliver the final report and the pdfs October 2008

Page 7© Crown copyright 2004 UKCIP08 Products  Report  Three types of output:  Probabilistic PDF  Weather Generator (change factors from PDFs)  Raw daily data from 17 regional climate models  Web-based data delivery package (UI)  Will produce nice graphics  Provide some analysis  Provide some guidance  Documentation on guidance  Preparatory workshops

Page 8© Crown copyright 2004 Probabilistic climate predictions are …  It is not a probability distribution from which the real world samples what it does  So not an ensemble weather forecast for the future.  It is just a representation of the degree to which each possible future climate is plausible given the evidence (climate models and observations). As the evidence changes so will the prediction.  Underlying value is to reduce the risk of a user making a bad decision  So instead of giving a policy maker all our modelled and observed data we give them a summary statement of the extent to which various possible future climates are consistent with the evidence.

Page 9© Crown copyright 2004 Production of UKCIP08 predictions EBM Time-scalingDown- scaling Perturbed physics ensemble Ocean PPE Aerosol PPE Carbon cycle PPE No computer in world is big enough to run many variants of a 25km Earth system model so we have developed a framework to combine lots of pieces (Murphy et al, Phil. Trans. Royal Society, 2007).

Page 10© Crown copyright 2004 Perturbed physics ensembles

Page 11© Crown copyright use “perturbed physics ensembles” to sample systematically a space of possible model configurations Relatively large ensembles designed to sample modelling uncertainties systematically within a single model framework Executed by perturbing model input parameters controlling key model processes, within expert-specified ranges Key strength: Allows greater control over experimental design cf multi-model “ensembles of opportunity” Key limitation: does not sample “structural modelling uncertainties”, e.g. changes in resolution, or in the fundamental assumptions used in the model’s parameterisation schemes – need to include results from other models to account for these.

Page 12© Crown copyright 2004 First steps Take one climate model (in this case version 3 of the Hadley Centre model) Specify distributions for multiple uncertain model parameters controlling atmospheric physical processes Run an ensemble of simulations horizontal resolution) of the equilibrium response to doubled CO 2

Page 13© Crown copyright gives a large (~300 member) sample of possible changes (e.g. summer UK rainfall)

Page 14© Crown copyright 2004 Making probabilistic climate predictions for 2xCO2 response

Page 15© Crown copyright 2004 Bayesian prediction – Goldstein and Rougier  Aim is to construct joint probability distribution p(X, m h, m f,y,o,d) of all uncertain objects in problem.  Input parameters (X)  Historical Model output (m h )  Model prediction (m f )  True climate (y h,y f )  Observations (o)  Model imperfections (d)  It measures how all objects are related in a probabilistic sense

Page 16© Crown copyright 2004 Best-input assumption  Physical and dynamical processes in a climate model are controlled by numbers called model input parameters.  We assume that one choice of these values, x*, is better than all others True climateDiscrepancy Model output of best choice of parameter values x*

Page 17© Crown copyright 2004 Best-input assumption  We only know the probability that any combination of parameter values is the best- input model. But that means we need millions of model variants.  That is too expensive - can only afford hundreds of runs but they have to sampled in a way that is consistent with your beliefs about where the best model is.  Need a cheap alternative..

Page 18© Crown copyright 2004 Emulators e.g. climate sensitivity Ensemble member Sqrt(climate sensitivity) Dots – actual runs Lines – 95% credible interval from emulator Emulators are statistical models, trained on ensemble runs, designed to predict model output at untried parameter combinations

Page 19© Crown copyright 2004 Sampling different model variants with emulator

Page 20© Crown copyright 2004 Climate sensitivity – before weighting with observations FOCUS ON BLACK CURVE The Prior

Page 21© Crown copyright 2004 Parameter Constraints due to weighting

Page 22© Crown copyright 2004 Weighting different model variants

Page 23© Crown copyright 2004 Weighting different model variants

Page 24© Crown copyright 2004 Climate sensitivity “Truncation level” = amount of independent information from observations FOCUS ON RED CURVE The Posterior

Page 25© Crown copyright 2004 Climate sensitivity “Truncation level” = amount of independent information from observations FOCUS ON RED CURVE

Page 26© Crown copyright 2004 Weighting models with observations and discrepancy

Page 27© Crown copyright 2004 Physics/dynamics matter…  Compare models against several observational variables – with just one variable you can simulate climate well for the wrong reasons  Will compare with present-day mean climate - Indirect assessment of key processes for our climate prediction but adds confidence to our prediction of one-off event  We are not going to assume models are perfect so using better models has an impact

Page 28© Crown copyright 2004 Best-input assumption  Physical and dynamical processes in a climate model are controlled by numbers called model input parameters.  We assume that one choice of these values, x*, is better than all others True climateDiscrepancy Model output of best choice of parameter values x*

Page 29© Crown copyright 2004 Comparing models with observations  Use likelihood function i.e. skill of model is likelihood of model data given some observations V = obs uncertainty + emulator error + discrepancy Discrepancy is ‘distance’ between real system and ‘best’ choice of input parameters Truncation level = dimensionality of m, o

Page 30© Crown copyright 2004 Discrepancy – a schematic of what it does Avoids observations over-constraining the pdfs. Avoids contradictions from subsequent analyses when some observations have been allowed to constrain the problem too strongly.

Page 31© Crown copyright 2004 Specifying discrepancy  Use multimodel ensemble from AR4 and CFMIP  For each multimodel ensemble member, find emulated model variant that is closest to that member  There is a distance between climates of this multimodel ensemble member and this “best” emulated model variant i.e. effect of processes not explored by slab model variants.  Pool these distances over all multimodel ensemble members

Page 32© Crown copyright 2004 Four types of data…

Page 33© Crown copyright 2004 Errors in predicting multimodel ensemble Each dot is a member of multimodel ensemble Grey shading represents 95% confidence interval from internal climate variability A choice: select 10 as this is as large as possible whilst still providing a robust estimate Number of observable quantities in cost function used to find ‘best input’

Page 34© Crown copyright 2004 Climate sensitivity “Truncation level” = amount of independent information from observations FOCUS ON RED CURVE

Page 35© Crown copyright 2004 Joint probabilities

Page 36© Crown copyright 2004 Time scaling

Page 37© Crown copyright 2004 Production of UKCIPnext predictions EBM Time-scalingDown- scaling Equilibrium PPE Ocean PPE Aerosol PPE Carbon cycle PPE For A1B, B1, A1FI scenarios…

Page 38© Crown copyright 2004 Coupled Atmosphere-Ocean Ensembles  Smaller ensembles of HadCM3 because of spin- up issues  Perturbations to atmosphere- model parameters with equivalent HadSM3 versions  Flux adjustments used to keep models stable and reduce SST biases Observations Historical + A1B forcing Collins et al. 2006

Page 39© Crown copyright 2004 Pattern Scaling to Produce Pseudo-Transient Ensembles - Methodology 

Page 40© Crown copyright 2004 Some plumes…Wales August temperature No carbon cycle feedback yet

Page 41© Crown copyright 2004 Other components of Earth System

Page 42© Crown copyright 2004 Production of UKCIPnext predictions EBM Time-scalingDown- scaling Equilibrium PPE Ocean PPE Aerosol PPE Carbon cycle PPE For A1B, B1, A1FI scenarios…

Page 43© Crown copyright 2004 Uncertainties in the transient response of global mean surface temperature Ocean parameters perturbed Sulphur Cycle parameters perturbed Atmosphere parameters perturbed Ocean parameter perturbation experiments (17 member ensemble) run to quantify effects of uncertainties in ocean transport processes Sulphur cycle parameter perturbation experiments (another 17 member ensemble) also run

Page 44© Crown copyright 2004 Impact of terrestrial uncertainties on CO2 Standard HadCM3, 16 variants of terrestrial carbon cycle Black crosses - observations Total atmospheric CO2 concentration

Page 45© Crown copyright 2004 Downscaling

Page 46© Crown copyright 2004 Production of UKCIPnext predictions EBM Time-scalingDown- scaling Equilibrium PPE Ocean PPE Aerosol PPE Carbon cycle PPE

Page 47© Crown copyright 2004 Downscaling Have also run a 17-member 25km resolution ensemble of perturbed physics regional model versions. Driven by boundary forcing from the HadCM3 A1B transient simulations ( ). We will construct regression relationships between the 17 GCM and 17 RCM simulations of future climate. Use these to create regional response pdfs at 25km scale. Will add further uncertainty to the regional responses.

Page 48© Crown copyright 2004 Downscaling uncertainty 16 realisations of the difference in response of the regional model relative to its driving global model, for January precipitation (% change for relative to ).

Page 49© Crown copyright 2004 Downscaling relationships…

Page 50© Crown copyright 2004 Assumptions

Page 51© Crown copyright 2004 What are the main assumptions we cannot test  Local feedbacks between atmosphere and other components of Earth System (carbon cycle, aerosol chemistry and ocean) are of second order importance to effects linked to global temperature change.  Structural model uncertainty is a good proxy for difference between HadCM3 family of models and real system  Pattern scaling, downscaling relationships applicable across parameter space  Multimodel members have equal contribution to discrepancy

Page 52© Crown copyright 2004 THE END  ANY QUESTIONS?

Page 53© Crown copyright 2004 UKCIPnext (Hadley Centre contribution) – Aims and Objectives  To provide joint probability distribution functions (pdfs) of predicted changes in a selection of key UK climate variables at 25km resolution for each decade during the 21st century  Results will be presented for each variable by month indicating mainly mean outcomes but also extremes for e.g. max/min temperature, precipitation  We aim to deliver the pdfs and final report summer 2008

Page 54© Crown copyright 2004 Sensitivity to prior – climate sensitivity Before observational After observational constraint constraint

Page 55© Crown copyright 2004 Sensitivity to prior - %ΔUK summer rainfall Before observational After observational constraint constraint

Page 56© Crown copyright 2004 Monte Carlo Sampling Emulated Samples EmulatedDistributions

Page 57© Crown copyright 2004 Reducing uncertainty  Improve observational uncertainties  Improve model i.e. reduce discrepancy  Run larger ensembles  Use more observational constraints independent of the ones used already  Remove pattern scaling and downscaling steps  Remove assumptions about linking sub- modules

Page 58© Crown copyright 2004 Weather Generators  We will make probabilistic predictions for the variables that are inputted into the weather generator  Weather Generators will be used to generate time series consistent with probabilistic predictions  If need spatially coherent time series at high temporal and spatial resolution, can use output from 17 regional climate model runs

Page 59© Crown copyright 2004 Ideal for future UKCIPs  Run with fully coupled Earth System Models perturbing parameters in all components simultaneously and then downscale  That is, no equilibrium runs, no ensembles on individual components  Would need other climate centres to run this experiment for their standard model and ideally they would have these downscaled.

Page 60© Crown copyright 2004 Response surface predicted by emulator Climate Sensitivity as a function of two parameters according to mean prediction of the emulator – note emulator also predicts uncertainty of response surface

Page 61© Crown copyright 2004 Summer UK % precipitation change Another choice: what truncation level to choose… FOCUS ON RED CURVE

Page 62© Crown copyright 2004 Probabilistic climate prediction  Probabilistic prediction is a function of  Model  Observations  Choices  Assumptions  Choices guided by principle that we think it is important to model the Earth System correctly.

Page 63© Crown copyright 2004 Bayesian framework by Goldstein and Rougier: some terms Murphy et al., 2004, Nature, 430, histogram of “perturbed physics” ensemble “emulated” prior distribution posterior distribution

Page 64© Crown copyright 2004 Ensemble Simulations  “Bedrock” provided by a relatively large ~300 member ensemble of HadSM3 (atmosphere-slab ocean) run at 1x and 2xCO 2  Results sensitive to how you select parameter combinations Murphy et al., 2004 Webb et al., submitted Stainforth et al., 2005

Page 65© Crown copyright 2004 Weights As truncation level increases, have to be luckier to land on a quality point in parameter space

Page 66© Crown copyright 2004 Precision of percentile estimates Number of Monte Carlo samples million Precision of 95 th percentile estimate CHOOSE THIS ONE!

Page 67© Crown copyright 2004 Emulators are statistical models, trained on ensemble runs, designed to predict model output at untried parameter combinations Emulators

Page 68© Crown copyright 2004 Monte Carlo sampling of parameters combined with an emulator overcomes dependency on sampling strategy to produce prior prediction (blue line) consistent with beliefs about where the best input lies. Prior distribution – prediction before any observations used Emulators and priors

Page 69© Crown copyright 2004 Discrepancy on future variable  Model not perfect so there are processes in real system but not in our model that could alter model response by an uncertain amount.  Places extra uncertainty on prediction variable in form of a variance

Page 70© Crown copyright 2004  Where is the ‘best’ input?  Observations reduce uncertainty about which points are best in parameter space  Most effective if a strong relationship exists Constraining predictions

Page 71© Crown copyright 2004 Standard carbon cycle, 3 versions of atmosphere GCM Dashed – no carbon cycle Solid – with carbon cycle

Page 72© Crown copyright 2004 Estimating discrepancy  Four ways I can think of…  Elicitation  Observations  Super-parameterised models  Ensemble of international climate models