Title: Downscaling of GCM Outputs for Flood Frequency Analysis in the Saguenay River System
1Downscaling of GCM Outputs for Flood Frequency
Analysis in the Saguenay River SystemDesaggrega
tion spatio-temporelle des sorties des GCM pour
lanalyse frequentielle des crues dans le bassin
du Saguenay
Research Group McMaster University (Dr. Y.
Dibike, S. Khan, B. Sawatzky, V.
Arnold) Universite Laval (F. Anctil, N.
Lauzon) ALCAN (B. Larouche)
Financial Support Climate Change Action Fund
(CCAF), EC ALCAN Company, Jonquiere, Quebec
2Contents
Project overview
DANN Downscaling Approach
Progress Results
3-
- Evaluate stochastic and statistical downscaling
methods - Develop dynamic neural network downscaling
methods - Inter-comparison study
- Hydrologic impact of climate change in the
Saguenay watershed - Flood regime analysis Magnitude Frequency
- Uncertainty analysis
Project objectives
4Simulations
Hydrologic Models
Downscaling Methods
1 2 3
WATFLOOD HEC-HMS CEQEAU HBV96 (ANN)
SDSM1 LARS-WG2 DANN3
1 2 3
1 2 3
1 2 3
Flood Regime Analysis
Project overview
5Statistical prediction / estimation
Linear regression (Box Jenkins)
Nonlinear regression (Sigmoid, ANN)
6Artificial Neural Networks (ANNs)
Inputs
neuron
Cell body
connection
nucleus
synapse
outputs
dendrites
axon
Simplified natural neurons
Artificial neurons
7Artificial Neuron
Neuron i
Xj Xn-1 Xn
Output
Wij
?
G
YG(?)
bi
Win
??wijxj bi
Inputs
8 Time Delay Line
SiWiX
X(t)
wi(0)
D
X(t-1)
G
Si
D
Yi(t)
X(t-2)
bi
D
wi(3)
X(t-3)
Neuron i
Delay Line ( order p3 )
9 MLP (FNN) RNN
Input Variables
Hidden Layer
Output layer
XQ
?
XP
?
XN
Y
XTmx
?
XTm
10MLP --gt IDNN / TDNN
XQ(t-1)
D
D
Hidden Layer MLP
Output layer
Y(t)
XmxT(t)
X(t)
D
RNN --gt TDRNN
D
11ANN Fondamental Elements
Data (input selection)
Topology (layers neurons)
Structure (Type link)
Algorithm Training
Degree of importance for Generalization
Degree of difficulty
12DANN models
- IDNN (TDNN)
- RNN (Elman)
- Jordan RNN
- Generalized RNN
13Study
Saguenay-Lac-Saint-Jean (SLSJ) Watershed
14The Study Area
Study
- The Saguenay Lac-Saint-Jean (SLSJ) hydrologic
system in northern Quebec - The total area is about 73,800 km2
- It extends between 70.5o - 74.3o West and between
47.3o - 52.2o North. - Saguenay is a well known flood prone area as many
Canadians still remember the year-1996 flood of - this river
- Only one of the SLSJ sub-basins namely
Chute-du-Diable is considered in the first phase
of this study
15Data Collection
Study
- Historical (observed) daily meteorological data
(such as daily precipitation, maximum and minimum
temperature) - ALCAN meteorological network
- Environment Canada (METDAT CDROM)
- Historical (observed) daily hydrologic data
(streamflow and reservoir inflow) - ALCAN hydrometric network
- Environment Canada (HYDAT CD-ROM)
- Observed daily data of large-scale predictor
variables representing the current climate
condition (1960 2000) - The Reanalysis dataset of the National Centers
for Environmental Prediction (NCEP) - GCM output of large-scale predictor variables
- The Canadian Climate Impacts and Scenarios (CCIS)
project website.
16GCM Data
- The data is extracted from 201-year simulations
with the Canadian Global Coupled Model-1 (CGCM1) - Uses the IPCC "IS92a" forcing scenario
- The change in greenhouse gases (GHG) forcing
corresponds to that observed from 1900 to 1990
and increases at a rate 1 per year thereafter
until year 2100. - The direct effect of sulphate aerosols (A) is
also included.
indicates p_, p5 or p8 which represent the
variable values near surface, at 500 hPa height
or 850 hPa height, respectively.
17Application Chute-du-Diable
- Chute-du-Diable watershed 9,700 km2
- Variables to be downscaled (predictands)
- daily precipitation
- daily Max and Min temperature
- The period between 1961 till 2000 is identified
to represent the current climate condition - The future climate change simulations (CGCM1) at
the coordinate 50o N latitude and 71o W longitude
were extracted for three distinct periods - the 2020s (2010 and 2039),
- the 2050s (2040-2069) and
- the 2080s (2070-2099)
18Downscaling experiment
- Case 1 The predictand is observed data from a
single station - Chute du Diable
- Chute des Passes
- Case 2 The predictand is observed data averaged
over the basin - From 25 meteorological stations with
precipitation and Tmax and Tmin measurements - Model calibration and validation
- 30 years (1961-1990) are used for calibration
- 10 years of data (1991-2000) are used for
validation
19Selection of predictors
- Selecting predictor variables
- Very important step
- Correlation analysis and scatter plots DANN
sensitivity analysis - Identified variables must be physically sensible
- Summary of the most relevant large-scale
predictor variables identified
temp mslp p500 p850 sphu s500 p__u p5_u p8_u p__v p8_v p_zh p5zh p8zh
SDSM x x x x x
TDNN x x x x x x x x
Elman x x x x x x x x
Jordan x x x x x x x x
GRN x x x x x x x x
20Model performance criteria
- Performance criteria
- Precipitation
- Mean daily precipitation and daily precipitation
variability for each month, - Monthly average dry and wet-spell lengths
- Residuals
- RMSE, R2, r
- Tmax and Tmin
- Monthly means and variances
- Residuals
- RMSE, R2, r
21Validation results SDSM, LARS-WG, TDNN1, TDNN3
SDSM
LARS-WG
TDNN1
TDNN2
22Residuals
SDSM
LARS-WG
TDNN1
TDNN2
23Downscaling results for the current and future
condition
SDSM
LARS-WG
TDNN1
TDNN2
24Conclusions
- Even though SDSM LARS-WG models indicate an
increasing trend in mean daily temperature, SDSM
resulted in a relatively higher increase than
that of LARS-WG. The TDNNs indicate a lower
increasing trend in mean daily temperature than
the SDSM. - SDSM output shows on average an increase in mean
daily temperature by about 4.5 oC, while LARS-WG
output indicates an average increase in mean
daily temperature by about 2.5 oC. - TDNN1 and TDNN2 indicate on average an increase
in mean daily temperature by about 3 oC and 3.5
oC respectively. - Both the SDSM and the TDNNs output shows an
increasing trend in the daily precipitation and
their variability. - LARS-WG results do not show any obvious trend in
both the daily precipitation and their
variability.
25Current and Future Work
- Application of different hydrologic models
(CEQEAU, HEC-HMS and WATFLOOD, HBV96, ANN) for
flow simulation in the river basin - Development of a dynamic neural-network based
downscaling method - -- with an adaptive module to facilitate model
transferability - Downscaling the GCM outputs for each of the
remaining sub-basins in SaguenayLac-Saint-Jean
river system - Perform flood frequency analysis in the river
system corresponding to the present and predicted
future flow regimes - Assess the hydrologic impact of future climate
change in the Saguenay river system as a whole.
26Merci
Thanks !