Linking probabilistic climate scenarios with downscaling methods for impact studies

1
Linking probabilistic climate scenarios with
downscaling methods for impact studies

• Dr Hayley Fowler
• School of Civil Engineering and Geosciences
• University of Newcastle, UK
• With Contributions from
• Claudia Tebaldi (NCAR)
• Stephen Blenkinsop, Andy Smith (Newcastle
  University)

2
Aim

• Develop a framework for the construction of
  probabilistic climate change scenarios to assess
  climate change impacts at the
• regional (100,000 to 250,000 km2)
• river basin (10,000 to 100,000 km2)
• catchment (1000 to 5000 km2) scales

3
Motivation

• Different GCMs produce different climate change
  projections, especially on a regional scale
• Therefore no one model provides a true
  representation
• Most probabilistic scenarios to date have been
  produced for large regions or globally
• Regional scale studies more relevant for impacts
• How can we combine probabilistic climate
  scenarios with downscaling methods to study
  impacts at the catchment scale?

4
How can we combine probabilistic climate
scenarios with downscaling methods to study
impacts at the catchment scale?

• Examining how well different RCMs simulate
  different statistical properties of current
  climate in their control climates
• Do different RCM-GCM combinations produce
  different future projections?
• How can we combine the estimates of different
  models to produce probabilistic scenarios?

5
Case-study Locations
1 British Isles 2 Eden 3 Ebro 4 Gallego 5 Meuse 6
Dommel 7 Brenta 8 Scandinavia 9 Eastern Europe
6
Method RCMs WG
?
7
Data available for UK
RCM data 50km x 50km Control 1961-90 Future
SRES A2 2070-2100
Interpolated observations 5km x 5km
8
Data Observations Models
Observed series - Aggregated 5km interpolated
precipitation dataset Regional Climate Models
PRUDENCE (http//prudence.dmi.dk/)
9
How well do RCMs represent the seasonal cycle?
10
How well do RCMs represent the seasonal cycle?
11
How well do RCMs represent the seasonal cycle?
12
Summer Skewness Coefficient
13
UK Regions
14
Method RCMs WG
?
15
Model weighting (a la Tebaldi)

• Bayesian statistical model delivers a fully
  probabilistic assessment of the uncertainty of
  climate change projections at regional scales
• Based on
• Reliability Ensemble Average method (Giorgi and
  Mearns, 2002)
• Summary measures of regional climate change,
  based on a WEIGHTED AVERAGE of different climate
  model responses

16
Model weighting (a la Tebaldi)

• Weights account for
• BIAS - the performance of GCMs when compared to
  present day climate ( i.e. results from model
  validation)
• CONVERGENCE - the degree of consensus among the
  various GCMs responses/

17
Model weighting (a la Tebaldi)

• pdf of change in temperature and precipitation
  fitted using area-averages of the model output
• Prior pdfs are assumed to be uninformative
• Data from regional models/observation
  incorporated through Bayes theorem, to derive
  posterior pdfs
• Model-specific reliabilities parameters
  estimated as a function of model performance in
  reproducing current climate (1961-1990) and
  agreement with the ensemble consensus for future
  projections
• These are standardised and applied as weights in
  the downscaling step

18
(No Transcript)
19
(No Transcript)
20
NWE Seasonal Mean ?
Precipitation
Temperature
21
Method RCMs WG
?
22
EArWiG

• EA Weather Generator
• Developed for EA for catchment scale Decision
  Support Tool models
• Generates series of daily rainfall, T, RH, wind,
  sunshine and PET on 5km UK grid
• Observed and climate change based on UKCIP02
  scenarios
• Collaborative with CRU, UEA

23
EArWiG

• Map viewer interface developed
• Can select catchments, time periods and different
  UKCIP02 scenarios

Toolbar
Catchments tab
Map window
Model tab
Catchment finder
OSGB locator
OSGB pointer coords
24
Neyman-Scott Rectangular Pulses Rainfall Model
Storm origins arrive in a Poisson process with
arrival rate ? Each storm origin generates C
raincells separated from the storm origin by time
intervals exponentially distributed with
parameter ß Raincell duration is exponentially
distributed with parameter ? Raincell intensity
is exponentially distributed with parameter ?
Rainfall intensity is equal to the sum of the
intensities of all the active cells at that
instant
25
Weather Generator

• Depending on whether the day is wet or dry, other
  meteorological variables are determined by
  regression relationships with precipitation and
  values of the variables on the previous day
• Regression relationships maintain both the cross-
  and auto-correlations between and within each of
  the variables

26
Change factor fields

• Change factor fields are applied to the fitted
  rainfall model statistics
• Mean
• Variance
• PD
• Skewness Coefficient
• Lag 1 Autocorrelation
• Change factor fields are applied to the weather
  generator statistics
• Mean temperature
• Temperature SD

27
CF Summer mean temperature
28
CF Winter mean precipitation
29
CF Spring PD
30
(No Transcript)
31
Method RCMs WG
?
32
Rainfall-runoff model

• ADM model, simplified version of Arno
• Calibrated for Eden catchment on observed data
• R20.73, 0.78
• Each simulated climate used to produce simulated
  flow series (30 years) for each climate model
  using P and PET

33
EARWIG run for each RCM
3
2
4 1000
1

• Had_P
• RCAO_E
• Control
• Each series is 30 years in length

2071-2100
1961-1990
34
NWE Seasonal Mean ?
Precipitation
Temperature
35
Re-sampling

• Monte-Carlo re-sampling technique used to weight
  models according to ? values from Bayesian
  weighting
• Random numbers used to choose a control and
  future run for a particular RCM, then seasonal
  statistics of change in mean flow, SD flow, 5th
  and 95th percentiles calculated.
• If seasonal ?0.14 then random number generator
  produces 140 resamples from a particular RCM
• Generates total of 1000 change statistics for
  each season pdf fitted used kernel density

36
2080s
37
2020s
38
(No Transcript)
39
(No Transcript)
40
(No Transcript)
41
Questions for the audience

• Should we weight models (CG)?
• Should we be weighting on statistics other than
  mean?
• If so, what?
• Should we be looking at weighting by some spatial
  bias measure rather than a simple regional
  average? Makes the statistics harder
• Models may produce reasonable mean statistics and
  get higher order statistics important for impact
  studies wrong