Predicting rainfall for the city of the future - SWITCH · Predicting rainfall for the city of the...

3rd SWITCH Scienti�c MeetingBelo Horizonte, Bresil

29 November � 4 December 2008

Predicting rainfall for the city of the future

Xavier Beuchat, Marc Soutter

Ecohydrology LaboratorySwiss Federal Institute of Technology (EPFL)

Lausanne, Switzerland

December 2008

Flood in western France10 Mars 2008

Fullscreen Acrobat Reader

Predictingrainfall forthe city ofthe future

Xavier

Beuchat,

Marc Soutter

Introduction

Climatechange:globalpredictionsanddownscaling

Modellingstrategy:theoreticalconsidera-tions

Case study:Geneva,Switzerland

Data

Results of thedownscalingat a dailytime step

Planning for the city of the future

- pollution- urbanisation- energy- ...

Economic crisis

Climate change

Strategies should be �exible or robust!


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Why climate change should be taken into account?

There is now evidence that climate change may perturb

global water cycle

Climate change prediction are uncertain

Classical stormwater managment may not be adequate

(neither �exible nor adaptable)

Alternative strategies BMP, SUDS may be more valuable

Objectives of the work

Modelling rainfall for the city of the futur, taking

uncertainties into account

Using predicted rainfall series as input in a rainfall-runo�

model

Testing di�erent strategies to mitigate impacts of climate

change


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Problem of scales!

Global Climate Models (GCMs)

prediction are coarse.

Typically:

2,5° latitude and 3,75°

longitude

Monthly time step

Urban hydrology

Spatial variability is especially important

Temporal resolution can be 10 minutes depending on the

phenomenon we want to model


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Downscaling


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Theoretical framework for producing probabilisticrainfall scenarios for the city of the future

Approach based on the pilot methodolgy developped within

the DEFRA project FD2113: �single and multi-site rainfall

generation with climate model uncertainty�(ended

september 2006)

Modelling and simulation of daily rainfall time series at a

single or multiple sites

Methods

Generalized Linear Models (GLMs) are used to downscale

rainfall at daily time step from GCMs outputs

Disagregation to hourly data is then achieved via methods

based on Poisson cluster models


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Using GLMs to downscale rainfall at a daily timestep (1)

Basically, GLMs derive relationships between a variable of

interest, Y = (Y1, . . . ,Yn)t , called the predictand and a set of

p temporally varying predictor variables, or covariates, whose

values can be arranged in a n × p matrix X. The relationships

between the predictand and the predictors are assumed to be

given by

g(µ) = Xβ ; µ = E(Y) (1)

Moreover, the distribution of each Yi is assumed to belong to

the exponential family. That is

fY (y ; θ, φ) = exp

[yθ − b(θ)

a(φ)+ c(y , φ)

](2)

where a, b and c are some speci�c functions and φ is called the

dispersion parameter.


Xavier

Beuchat,

Marc Soutter

Introduction




Data



Here,

Predictand ← rainfall series

Predictors ← terms relating to seasonality, terms relating

to the history of the series itself (autocorrelations, intensity

of previous rain events. . . ) and external covariates such as

coarse-scale atmospheric variables.


Xavier

Beuchat,

Marc Soutter

Introduction




Data



Models are �tted using historical data (rainfall and large

scale atmospheric variables derived from NCEP reanalysis)

to describe relationships.

Then, for future scenarios, one may simulate from the

�tted models driven by GCM-generated atmospheric

variables.

Actually, the modelling framework is composed of two

components: �rst, an occurrence model based on logistic

regression is used to model the pattern of wet and dry

days, and second, an amount model based on gamma

distributions allows the simulation of rainfall amounts on

wet days.

Simulating nonstationary rainfall sequences is achieved by

allowing some predictors to modulate the e�ect of other

predictors incorporated via interactions (alternative to,

e.g., �tting separate models in each month of the year).


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Case study: Geneva, Switzerland

Station Genève-Cointrin Genève-Aïre Croix-de-Rozon Jussy

Northings [km] 508 810 498 580 499 480 495 800Eastings [km] 120 310 122 320 111 080 116 900Altitude [m] 420 375 478 465Data availability 1900�. . . 1968�. . . 1974�2005 1900�. . .

Auto. since 1980;Type

TB beforeTB TB TB


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Evolution of surface air temperatures

−1

01

23

Low greenhouse gas emission scenario (SRESB1)

1950 1980 2010 2040 2070 2100

bccr_bcm2_0cnrm_cm3csiro_mk3_0csiro_mk3_5gfdl_cm2_0gfdl_cm2_1giss_model_e_riap_fgoals1_0_ginmcm3_0miroc3_2_hires

miroc3_2_medresmpi_echam5mri_cgcm2_3_2ancar_ccsm3_0ncar_pcm1ukmo_hadcm3

−1

01

23

4

Medium greenhouse gas emission scenario (SRESA1B)

1950 1980 2010 2040 2070 2100

bccr_bcm2_0cnrm_cm3csiro_mk3_0csiro_mk3_5gfdl_cm2_0gfdl_cm2_1giss_model_e_hgiss_model_e_riap_fgoals1_0_gingv_echam4

inmcm3_0miroc3_2_hiresmiroc3_2_medresmpi_echam5mri_cgcm2_3_2ancar_ccsm3_0ncar_pcm1ukmo_hadcm3ukmo_hadgem1

−1

01

23

4

High greenhouse gas emission scenario (SRESA2)

1950 1980 2010 2040 2070 2100

bccr_bcm2_0cnrm_cm3csiro_mk3_0csiro_mk3_5gfdl_cm2_0gfdl_cm2_1giss_model_e_ringv_echam4inmcm3_0miroc3_2_medres

mpi_echam5mri_cgcm2_3_2ancar_ccsm3_0ncar_pcm1ukmo_hadcm3ukmo_hadgem1


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Summary

Models were �tted using observed rainfall series and NCEP

predictors during the period 1982-2007.

Fitting and simulations were carried out simultaneously for

the all the four sites.

Three SRES scenarios were considered: A2 (high), A1B

(medium), B1 (low)

Around 20 di�erent GCMs were considered per scenario.

The period 1956-1981 has been used for validation.

The future period used for simulation is 2072-2098.


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Summary monthly statistics of 200 simulations forthe site Genève-Cointrin during the �tting period

2 4 6 8 10 12

12

34

56

Month

mm

Mean

2 4 6 8 10 12

24

68

1216

Month

mm

SD

2 4 6 8 10 12

0.2

0.3

0.4

0.5

0.6

0.7

Month

Pro

port

ion

Pwet

2 4 6 8 10 12

24

68

1216

Month

mm

CMean

2 4 6 8 10 12

05

1015

2025

Month

mm

CSD

2 4 6 8 10 12

050

150

250

Month

mm

Max

2 4 6 8 10 12

0.0

0.1

0.2

0.3

0.4

Month

Cor

rela

tion

ACF1

2 4 6 8 10 12

−0.

10.

10.

20.

30.

4

Month

Cor

rela

tion

ACF2

2 4 6 8 10 12

−0.

10.

00.

10.

20.

3

Month

Cor

rela

tion

ACF3

2 4 6 8 10 12

510

2030

Month

days

Max Dry Spell

2 4 6 8 10 12

050

150

250

Month

mm

Max 5−days tot

2 4 6 8 10 12

24

68

1012

Month

Skewness


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Summary monthly statistics of 200 simulations forthe site Genève-Cointrin during the validation period

2 4 6 8 10 12

1.5

2.5

3.5

4.5

Month

mm

Mean

2 4 6 8 10 12

46

810

Month

mm

SD

2 4 6 8 10 12

0.2

0.3

0.4

0.5

0.6

0.7

Month

Pro

port

ion

Pwet

2 4 6 8 10 12

46

810

12

Month

mm

CMean

2 4 6 8 10 12

46

810

14

Month

mm

CSD

2 4 6 8 10 12

050

100

200

Month

mm

Max

2 4 6 8 10 12

0.00

0.10

0.20

0.30

Month

Cor

rela

tion

ACF1

2 4 6 8 10 12−0.

100.

000.

100.

20

Month

Cor

rela

tion

ACF2

2 4 6 8 10 12−0.

100.

000.

100.

20

Month

Cor

rela

tion

ACF3

2 4 6 8 10 12

510

2030

Month

days

Max Dry Spell

2 4 6 8 10 12

050

150

250

Month

mm

Max 5−days tot

2 4 6 8 10 12

24

68

1012

Month

Skewness


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Summary monthly statistics of 200 simulations forthe site Genève-Cointrin during the future period(scenario A2)

2 4 6 8 10 12

02

46

810

Month

mm

Mean

2 4 6 8 10 12

05

1015

2025

Month

mm

SD

2 4 6 8 10 12

0.2

0.4

0.6

0.8

Month

Pro

port

ion

Pwet

2 4 6 8 10 12

05

1015

2025

Month

mm

CMean

2 4 6 8 10 12

05

1020

30

Month

mm

CSD

2 4 6 8 10 12

010

030

050

0

Month

mm

Max

2 4 6 8 10 12

−0.

10.

10.

30.

5

Month

Cor

rela

tion

ACF1

2 4 6 8 10 12

−0.

10.

10.

20.

30.

4

Month

Cor

rela

tion

ACF2

2 4 6 8 10 12

−0.

10.

00.

10.

20.

3

Month

Cor

rela

tion

ACF3

2 4 6 8 10 12

510

2030

Month

days

Max Dry Spell

2 4 6 8 10 12

020

040

060

0

Month

mm

Max 5−days tot

2 4 6 8 10 12

05

1015

20

Month

Skewness


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Annual total rainfall simulated by the models

1985 1995 2005

050

015

0025

00

Year

mm

Total year rainfall conditionnal

on NCEP outputs

1985 1995 2005

020

060

0

Year

mm

Total summer (JJA) rainfall conditionnal

on NCEP outputs

1985 1995 2005

020

060

0

Year

mm

Total winter (DJF) rainfall conditionnal

on NCEP outputs

2075 2085 2095

050

015

0025

00

Year

mm


on all SRESA1B outputs

2075 2085 2095

020

060

0

Year

mm



2075 2085 2095

020

060

0

Year

mm



2075 2085 2095

050

015

0025

00

Year

mm


on all SRESA2 outputs

2075 2085 2095

020

060

0

Year

mm



2075 2085 2095

020

060

0

Year

mm



2075 2085 2095

050

015

0025

00

Year

mm


on all SRESB1 outputs

2075 2085 2095

020

060

0

Year

mm



2075 2085 2095

020

060

0

Year

mm




Xavier

Beuchat,

Marc Soutter

Introduction




Data


Extreme value analysis

Independently of the SRES scenario considered, the

average forecasted 100-year return level for the future

period is higher than for the �tting period.

The higher the forcing is, the higher the average 100-year

return level is (SRESB1: 117.71 mm, SRESA1B: 124.40

mm, SRESA2: 133.97 mm).


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Perpsectives

Futur tasks make the methodology more robust and easy

to use

go on with subdaily simulationsfor a given city:

build a detailed rainfall-runo� modelassess impacts of perturbed rainfallsceanrios on stormwater mangementquantify these impacts against otherpotential threads (population increase. . . )test di�erent strategies to mitigate theseimpacts.

Demo cities on demand, provide them with scenarios of

rainfall for the future

I need for that at least 20 years of observed

rainfall series at (sub)daily time step, for one

or more stations.


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Example: a model for Belo Horizonte (1961-1999)


Xavier

Beuchat,

Marc Soutter

Introduction




Data


Example: a model for Belo Horizonte (2061-2099)

Predicting rainfall for the city of the future - SWITCH · Predicting rainfall for the city of the...

Documents

Transcript of Predicting rainfall for the city of the future - SWITCH · Predicting rainfall for the city of the...