CLIMATE IMPACTS

Rainfall-runoff modelling of future climate

Watershed Modelling – Exercise 3

Fall Semester 2020

Conduct a climate impact study on your assigned watershed in the Bern Oberland region with the
conceptual watershed model that you have modified and calibrated. Fit a daily stochastic weather
generator model, simulate an ensemble which consists of multiple years representing present and
future climates, and analyze the impacts of climate change on the streamflow.

The tasks in the exercise are:

1. Setting the WeaGETS weather generator model (WEEK 1)

• Decide on the setup to run the WeaGETS model for precipitation and air temperature
simulation that best reproduces the present climate (observed data).

• Run your watershed model with an ensemble of climate data produced by the weather
generator and with the observed climate data. Compare the model performance (streamflow
outputs).

2. Assess the changes to streamflow statistics in a future climate (WEEK 2)

• Generate a climate ensemble consisting of multiple realizations (simulations) following 6

climate trajectories (climate models) and 2 different GHG emission scenarios.

• Simulate future streamflow using the climate ensemble and discuss the changes in flow
statistics (e.g. mean and extremes).

• Partition the hydrological (streamflow) uncertainties predicted in a future climate into their
climatic sources – emission uncertainty, climate model uncertainty and stochastic
uncertainty.

REPORT

Write a synthesis report describing your work in English or German, fit into 5 pages, single-spaced
A4, including figures. You will be evaluated in your report on (1) your results, (2) how you interpret
the results, and (3) the clarity and presentation of your work.

Upload your report in PDF format by the deadline date using the Moodle link for the course. Deadline
for the submission is Friday, 27th November.

DO NOT EMAIL PDF TO INSTRUCTOR OR TEACHING ASSISTANT

1. Objectives
Global warming is already affecting everyday life. Climate changes are foreseen to accelerate in the
years to come, potentially effecting, for example, the energy sector (including hydro-power), water
availability and demand (e.g. agriculture sector), and the insurance sector (e.g. rising losses due to
flooding and other natural hazards). As a result, policymakers and stakeholders are keen to know
how climate changes may affect the hydrology. Future changes to the climate are, however, not fully
known and depend on many different factors, such as future demographic changes, advances in
technology in the energy sector and future climate mitigation strategies managed by the
governments. As a result, assessments of future changes of the hydrology are associated with large
uncertainties.

The overarching aim of this exercise is to learn how to conduct a hydrological climate impact study.
The goal is to gain experience in setting future climate trajectories to be used as an input into
hydrological modelling, assessing changes in the hydrological components, exploring the hydro-
climatic uncertainties, and reporting the changes to the hydrology with uncertainty ranges.

2. The WeaGETS daily weather generator model

To generate a climate ensemble, you will use a simple weather generator model, named WeaGETS,
developed by Chen et al. (2012). The WeaGETS model produces daily precipitation and maximum
and minimum daily temperatures. In the following paragraphs, a concise description of the model is
given. Note that the Matlab code you received in this exercise is a reduced-complexity version of the
model. The full model code can be found online (link to code), and is documented in the paper by
Chen et al. (Chen, J., Brissette, F.P. and Leconte, R., 2012. WeaGETS–a Matlab-based daily scale
weather generator for generating precipitation and temperature. Procedia Environmental Sciences,
13, pp.2222-2235; link to paper).

WeaGETS provides three options to generate precipitation occurrence, four options to produce
precipitation intensity and a conditional scheme to simulate T max and Tmin. The precipitation
occurrence parameters include the transition probabilities of first-, second- and third-order Markov
chains (you need to choose one of the three options). For precipitation intensity, there is one
parameter for the Exponential distribution, two parameters for the Gamma distribution, and three
parameters for the Skew Normal distribution and Mixed Exponential distribution (you need to choose
one of the distributions that match your data). The parameters are computed on a biweekly basis (26
estimations over the whole year – input is always a year with 365 days). Because of climate variability
and the finite length of the historical records, the variation from one 2-week period to the other will
not be smooth. The estimation of the parameters is done on a biweekly basis to capture the
seasonality in the climate (rainfall and air temperature).

2.1 Input data

The following input is required:

▪ Daily precipitation P (mm)
▪ Daily minimum and maximum air temperatures Tmin, Tmax (°C)
The provided input data is the same for the three catchments. Data cover 34 years period between
1986 and 2019. It was extracted from the Adelboden climate station located in the Hondrich
catchment at 1325 m asl.

2.2 Generation of precipitation occurrence

WeaGETS provides three options including first-, second- and third-order Markov chain models to
simulate precipitation occurrence. The first-order Markov process is the simplest and most widely
used. The probability of precipitation on a given day depends on whether the previous day was wet
or dry. Therefore, we obtain two transition probabilities, P01 and P11:

P01 = P{precipitation on day t | no precipitation on day t-1}

P11 = P{precipitation on day t | precipitation on day t-1}

2
Since precipitation either occurs or does not occur on a given day, the two complementary transition
probabilities are P00 = 1 - P01 and P10 = 1 - P11.

Letting Rt = 0 if day t is dry, and Rt = 1 if day t is wet, the probabilities above can be extended to the
second and third-order Markov chains, i.e. looking two and three days back in time:

Pijk = P{ Rt = k | Rt-1 = j | Rt-2 = i }

Phijk = P{Rt = k | Rt-1 = j | Rt-2 = i | Rt-3 = h}

where h, i, j and k are 0 or 1.

The number of parameters required to characterize precipitation occurrence increases exponentially

with the order of Markov process (i.e., two, four and eight parameters). Note that first-order Markov
chains may not be adequate for generating long dry or wet spells. Higher-order Markov models
usually perform better.

2.3 Generation of precipitation intensity

For a predicted rainy day, four probability distribution functions are available to produce the daily
precipitation intensity. The first is the one-parameter Exponential distribution, with the probability
density function given by:

f (x) = e−x
where x is the daily precipitation intensity and λ is the distribution parameter (equal to the inverse of
the mean).

The second function is the two-parameter Gamma distribution. The probability density function for
this distribution is given by:
−𝑥
𝑥
( )𝛼−1 𝑒 𝛽
𝛽
𝑓(𝑥) =
𝛽Γ(𝛼)

where α and β are the two distribution parameters, and Г(α) indicates the gamma function evaluated
at α. This method often performs better than the exponential distribution and is widely used to
generate daily precipitation intensities.

The other distributions, especially the Mixed Exponential distribution, were found better in
reproducing daily precipitation extremes in comparison to the Exponential and Gamma distributions
(for details see, e.g. link to paper).

2.4 Generation of maximum and minimum temperatures

The WeaGETS model uses a first-order linear autoregressive model to generate Tmax and Tmin. The
observed time series is first reduced to residual elements by subtracting the daily means and dividing
by the daily standard deviations. The means and standard deviations are conditioned on the wet or
dry status. The residual time series are then generated by:

𝑋𝑝,𝑖 (𝑗) = 𝐴𝑋𝑝,𝑖−1 (𝑗) + 𝐵𝜖𝑝,𝑖 (𝑗)

where Xp,i(j) is a 2×1 matrix for day i of year p whose elements are the residuals of T max (j=1) and
Tmin (j=2), εp,i(j) is a 2×1 matrix of independent random components that are normally distributed with
a mean of zero and a variance of unity. A and B are 2×2 matrices whose elements are defined such
that the new sequences have the desired auto and cross-correlation coefficients. During the
simulation, the residuals are added to the observed daily mean and multiplied by the observed daily
standard deviation to get the time series which are input into the hydrological model.

3
2.5 Running the model

The version of the model you received contains a graphical user interface. Simply run the script
RUN_WeaGETS in Matlab.

WEEK 1 ASSIGNMENT - Setting the WeaGETS weather generator model

First, decide on the setup of the WeaGETS model to use to generate the climate for your catchment.
There are different strategies you can apply to decide on the setup. For example, you can simulate
30 years (i.e. same length as the observed data) for each of the options the model has (12 options
in total: 3 Markov orders for the rainfall occurrence x 4 distributions for the rainfall intensities). It is
suggested to run each of the options more than once, to capture the natural variability (e.g. create
an ensemble with a total of 60 realizations: 12 options x 5 realizations of 30 years). Then, compare
the statistics of the simulated rainfall and temperature vs. the observed data and choose the model
option best suitable for your data. For example, compare the monthly number of wet days (Fig. 1),
CDFs of the rainfall intensity (Fig. 2), and monthly mean and standard deviation of the temperature
(Fig. 3). Also, compare statistics for extreme rainfall and temperature. Report which option you chose
and support your decision with figures and tables (there is no need to report the statistics for all the
options, only the ones that will be used later in your work).

Second, generate an ensemble representing the current climate using WeaGETS. To capture the
natural variability, it is recommended that the ensemble will be composed of at least 10 realizations
of 30 years. Then, run your watershed model with the observed data from this exercise (the 30-year
data) and with the simulated ensemble as inputs, and compare the results. You can follow for the
comparison similar strategies (figures, tables) as used for the validation of the model in Exercise 2.
Report the outcome – what is the level of similarity between the simulated streamflow using the
observed and simulated data?

Fig 1. Example of an analysis of the number of wet days per month. Blue boxplots represent the
observed data (natural variability of the number of wet days) and the red line represents the
simulated mean monthly number of wet days.

4
 Fig 2. Example of a comparison between the empirical cumulative distribution functions of the
observed (blue) and simulated (red) data. Extreme rainfall at the 99th percentile is 38.5 mm for the
observed data and 36.7 mm for the simulated data.

Fig 3. Example of a comparison between the observed (blue dots) and simulated (red line) air
temperature for its monthly mean, and standard deviation (x symbols and a dashed line).

5
WEEK 2 ASSIGNMENT - Assess the changes to streamflow statistics in a future climate

In this assignment, you will assess future changes to streamflow statistics in your catchment forced
by the changes to the climate.

Details on the factors of change (FC) computed for this location for mean temperature and
precipitation for two emission scenarios (RCP 4.5 and RCP 8.5, from CMIP5) and 6 climate models
are summarized in the table below. The FC are computed for the period 2070-2099 relative to the
current climate (control run) 1981-2010. The FC for precipitation is in % of the change in the mean,
the FC for temperature is an absolute difference in the mean. Models are from the Euro-CORDEX
regional climate assessments (at 11-km resolution) and are derived by GCMs from the CMIP5
initiative.

RCP4.5 RCP8.5
Climate model
Precipitation Temperature Precipitation Temperature
CLMCOM-CCLM4-ECEARTH 5.7% 1.9°C 10.2% 3.8°C
CLMCOM-CCLM4-HADGEM 5.3% 3°C 11.2% 5.3°C
DMI-HIRHAM-ECEARTH 1.1% 1.5°C 4% 3.4°C
MPICSC-REMO1-MPIESM 3% 1.7°C -1.9% 3.5°C
SMHI-RCA-IPSL -1.2% 2.9°C -3.8% 5.3°C
MPICSC-REMO1-MPIESM 3% 1.7°C -1.9% 3.6°C

First, generate a climate ensemble which consists of multiple realizations of a 30-year period for the
6 climate models and the 2 emission scenarios listed above. The climate ensemble should consist
of at least 120 realizations (10 realizations x 6 climate models x 2 emission scenarios). We will follow
a simple delta change approach for the simulations, i.e. generate the climate using the WeaGETS
model and then apply the FC on the temperature (additive, simply add the relevant FC value to the
generated temperature time series) and on the precipitation (multiplication, simply multiply the
generated precipitation time series with the relevant FC value).

Then, simulate the future streamflow with your conceptual water balance model using the generated
climate ensemble as input. Compare the statistics of the streamflow between the simulations of the
present and future climates. In your report, consider the changes to the mean streamflow (and
seasonality) and the extreme high and low flows. Discuss and summarize the results using
figures and tables, and remember to also report the uncertainty range of your assessments. The
statistics can include, for example, the changes to mean monthly streamflow, changes to maximum
annual streamflow, changes to the minimum 7-day flow, etc. In your report, also discuss the reasons
for the change as revealed by your hydrological model. For example, are the changes in temperature
affecting snowmelt amounts and timing, and how these modifications affect the streamflow statistics?
You can also consider other factors, such as changes to ET and the effects on streamflow, etc. in
your report.

Last, partition the hydrological uncertainties into their climatic sources. This will be done at the
annual scale, so first aggregate (sum) the streamflow in the ensemble.
• The total uncertainty will be computed as the 5-95th percentile range of all the annual
streamflow from the simulated future climate ensemble, i.e. all uncertainties combined.
• The emission scenario uncertainty will be computed as the difference of the mean of all
streamflow simulations between the two RCPs, divided by the total uncertainty.
• The climate model uncertainty will be computed as the difference between the mean
streamflow of the highest (maximum) and lowest (minimum) trajectories emerging from the
climate model. This will be conducted separately for each emission scenario. Then, average
the ranges of the two and divided by the total uncertainty.
• The stochastic uncertainty will be computed for the median climate trajectory for each
emission scenario by estimating the 5th – 95th percentile range for the 10 (or more)
realizations you simulated. Also, average the ranges of the two emission scenarios and
divide this range with total uncertainty.
Report the relative contribution of each source of climatic uncertainty to the future assessments of
streamflow. Discuss its meaning – what can be done to reduce the uncertainties in streamflow
predictions?