IMPACT OF CLIMATE CHANGE ON WATER DEMAND
Luther Uthayakumaran 1, Frank Spanninks 1, Adrian Barker 2, Andrew Pitman 2, Jason P. Evans 2
1. Sydney Water Corporation, Sydney, NSW
2. University of New South Wales, Sydney, NSW
.
KEYWORDS
Climate Change, Weather, Water Demand,
Resilience, Modelling
which are increasing in severity and frequency
ABSTRACT
with climate change. In 2016 Sydney Water
entered into a research partnership with the
Climate change impacts weather and changes
Centre for Climate Change of the University of
in weather is the single largest factor influencing
New South Wales to bring together the
fluctuations in water demand. Therefore, it
expertise, knowledge and resources of both
would be natural to expect climate change to
organisations, to develop a method of
impact on demand, quantifiable by integrating
quantifying the impact of climate change on
the outputs of climate models with Sydney
demand. As a result of this, Sydney Water has
Water’s demand forecasting model. However,
now able to successfully integrate the effect
most climate models typically do not produce a
climate change has on demand forecasts, and
single forecast, but rather produce an ensemble
for the first time has incorporated this into
of equally likely scenarios. The high variation
pricing recommendations made to the
between these projections lead to different
Independent Pricing and Regulatory Tribunal,
estimates of demand, leaving decision makers
and other infrastructure planning processes
with the task of selecting one projection to use.
such the Metropolitan water planning process.
On the other hand, there exists no scientific way
of doing this, as they are all equally likely,
The methodology used, involves integrating a
meaning selecting one would be as good as the
mathematical demand forecast model
other. The paper describes the way Sydney
developed by Sydney Water with climate
Water approached the problem, which was to
projections of the New South Wales / Australian
use a two-staged process; developing a
Capital Territory Regional Climate Modelling
mathematical solution to integrate the output of
(NARCLiM), via a Stochastic Weather
global models with Sydney Water’s demand
Generator developed by the University of New
forecasting model, and developing an approach
South Wales. The integrated model estimates,
based on the risk tolerance levels of the
in addition to the interactive effect of the
decision it informs to select the most
predictor variables including population growth,
appropriate output from a climate model, to
dwelling type mix and climate change, the
arrive at the most appropriate estimations of the
marginal impact of each of these factors on
marginal-impact of climate change on demand
water demand. Thus, it is also possible to
as, 2.14% in the near-future (2020-2040) and
quantify the relative impact of climate change in
4.42% in the far-future (2060-2080).
comparison to population growth and dwelling
type mix.
INTRODUCTION
For a complete technical explanation of the
Weather is the single most significant factor that
methodology refer Barker, et al (2018a).
influences day-to-day variations in water
demand. Weather driven variations in Sydney is
The purpose of this paper is to give an overview
estimated to be as high as 50 Giga Litre; most
of the approach and how some of the
of which is the result of extreme weather
challenges, both to do with the science, and
conditions such as heat waves and droughts,
policy development were overcome.
1
The key challenges involved in estimating the
Sydney Water Consumption Model
impact of climate change on long term demand
are;
The Sydney Water Consumption Model is a
statistical-forrecasting-model based on a
dynamic panel regression methodology
Climate models typically predict averages,
(Woolridge, 2010; Bun and Sarafidis, 2015)
whereas demand is mostly driven by extreme
based on the work of Abrahams Et al (2012).
events (severity and frequency)
The model forecasts metered water use - which
constitutes 90% of all consumption (the rest is
Climate Models (NARCLiM) does not produce
either leakage or undermetered usage, which
amounts to approximately 57 GL a year) -
single projection, but generate 12 member
based on population change, dwelling types,
ensembles, based on –
whether or not a property has water efficiency
-
four climate models and,
programs such as BASIX, and five weather
-
three downscaling methodologies
variables, which are;
1. average daily precipitation,
the difficulty of estimating the uncertainty
2. number of days in a month when
precipitation exceeds 2 mm,
surrounding secondary impacts (eg, demand
3. average daily maximum temperature,
driven by the use of water in mitigation
4. number of days in a month when
programs such as urban cooling).
temperature exceeds 300 C
5. average daily pan evaporation.
The first challenge is overcome by building a
Historic recordings from 12 weather stations
‘weather generator’ which develops Monte-
operated by the Bureau of Meteorology were
used in developing the model. For full technical
Carlo simulations of weather scenarios based
details of SWCM, please refer Barker et al
on climate average projections of NARCLiM,
(2018a)
and integrating the weather generator with
Sydney Water demand model to produce
demand forecasts taking climate change into
New South Wales / Australian Capital Territory
account.
Regional Climate Modelling Project (NARCLiM).
The NARCLiM project provides temperature and
The second has more to do with decision
precipitation data based on four different global
making than science; the challenge here is to
climate models (GCMs) for the present (1990-
have the right decision-making framework for
2010), near-future (2020-2040) and far-future
incorporating climate change into government
(2060-2080). The GCMs used are CCCMA3.1,
decision-making and policy development. We
CSIRO MK 3.0, ECHAM5 and MIROC 3.2. All
propose a risk based approach; selecting an
simulations were based on Intergovernmental
Panel on Climate Change’s Specoial Report on
ensemble member of the model, based on the
Emission Scenarios, scenario A2.. Data is
risk tolerance profile of the decision it informs.
available on 10km x 10 km grids for the whole of
Eastern Australia including Sydney Water’s areas
Secondary impacts of climate change on
of operation. Three runs are produced for each
demand, likely to be mainly constituted by
period/GCM, with each run generated using
demand from outdoor irrigation, greening and
different physics assumptions for the downscaling
urban cooling has not been properly estimated
process.
at this stage and is out of the scope of this
The NARCliM data was bias corrected, so that
paper.
the temperature and precipitation have the same
yearly averages as the Australian Water
SIMULATION /EXPERIMENT
Availabilty Project (AWAP) data, over the same
period.
The scientific component involves integrating
the NARCLiM projections with the Sydney
Stochastic Weather Generator.
Water Consumption Model (SWCM), by
developing a series of weather scenarios using
A stochastic weather generator developed by
a Stochastic Weather Generator and using them
Barker et al. (2018a) was used for the generation
as inputs to SWCM.
of weather scenarios as inputs for the SWCM. A
2
weather generator was used to overcome the
the present, near future or far future. We
problem that each NARCliM member only
therefore undertook three analyses, each for the
produces a single realisation of a stochastic
present, near and far-future:
process (i.e. weather). The weather generator
enables multiple (in this case 100) realisations to
1. Isolate the effect of climate change on water
be generated, each consistent with a NARCliM
consumption. Here, population is held at
ensemble member, to examine the statistical
2019/20 levels and the dwelling type mix
distribution of weather and water consumption
uses the population estimates;
forecasts.
2. Isolate the effect of population change on
water consumption. Here, population is
allowed to vary from 2014/15-2024/25, with
For each period/GCM/run combination, the
a dwelling type mix associated with that
stochastic weather generator was calibrated to
population change.
produce weather scenarios with statistical
3. Isolate the effect of dwelling type mix. Here,
properties similar to those of the NARCliM data.
population varies from 2014/15-2024/25 and
NARCliM weather data from the closest grid point
the dwelling type mix varies between the
to each of the 12 weather stations was used to
dwelling type mix estimate, simulations with
calibrate the stochastic weather generator. Each
no single dwellings, and simulations
weather scenario contains data for the 11 financial
assuming all single dwellings.
years from 2014/15 to 2024/25 and 100 weather
scenarios were generated for each period/climate
However, for the purpose of this paper, we will
model/run combination.
discuss only (1); the marginal impact of climate
change.
In total 13 sets of 200 years of data were generated
for each time period (present, near future, far
The process was repeated for each NARCliM
future) allowing quantification of the variance due to
period, GCM and run leading to 12 forecasts
changing weather. All weather variables were
altogether.
assumed not to have a yearly trend within the 20
year NARCliM period. It should be noted that
RESULTS
estimates of water demand by SWCM requires pan
evaporation, a variable not generated by most
The results, (Figure 1, Table 1 & Table 2),
weather and climate models including the NARCliM
suggest a large variation between the different
project. Instead, the evaporation model described
ensemble members of NARCLiM. However,
by Barker et al. (2018a) was used to generate
there is no direct scientific way of selecting one
evaporation data as a function of precipitation and
of these forecasts over the other, because the
maximum temperature.
different GCM outputs which lead to the
different demand forecasts, are all considered
The output from this was used as the five weather
equally likely. They are scenarios, rather than
related inputs to the SWCM.
members of a probability distribution. This
means using a central tendency such as a
Experiments performed
median or average would be equally useless, in
mathematical tems, these would not yield
The consumption forecast generated through
anything more accurate than arbitarily choosing
the above reflects changes in population,
one of the ensemble outputs.
dwelling types, and changes in weather patterns
that are likely to occur in relation to climate
change.
The population inputs were approved by the
NSW Government. The weather data
associated with a forecast is taken from a
stochastic weather generator simulation based
on data from a NARCliM ensemble member in
one of the present, near or far future periods.
We can vary the NARCliM ensemble member
and time period represented, with weather
reflecting the present, near or far future.
Therefore, we can examine the consumption
Figure 1: Annual Demand
forecasts for combinations of populations
Red CCMA3.1, Green CSIRO-MK 3.0, Blue ECHAM5,
Brown-MIRCO3.2 (reproduced from Barker et al (2018b)
between 2014/15 and 2024/25 with weather for
3
need to use a single scenario to inform all
decisions. Planning against different scenarios
depending on the risk profile of the decision
involved, is commonly used across government
Table 1 Total Annual Consumption (2019/2020)
- Period: 2020-2040
and industry. For example, governments
Model run
Median
% increase
typically plan for high impact situations such as
demand
from base
national security or preventing deadly diseases
GL
year
against worst case scenarios, while they plan
CCCMA3.1
488.7
0.93
for things like economic turns against medium
CCCMA3.1
488.7
0.75
term scenarios. We propose something similar,
CCCMA3.1
489.0
0.72
choosing a projection scenario, based on;
CSIRO-MK 3.0
493.0
1.79
CSIRO-MK 3.0
490.9
1.32
the impact of error, i.e, the impact of ‘getting
2
it wrong’
CSIRO-MK 3.0
492.1
1.62
ECHAM5
492.9
1.90
potential to respond to an error in terms of
ECHAM5 1
493.6
2.14
capability and speed to respond
ECHAM5
492.5
1.69
how equitably the impact of error is
MIROC3.2
484.4
-0.06
distributed among stakeholders.
MIROC3.2
485.6
0.27
MIROC3.2
487.4
0.51
We demonstrate this through the example of the
following two decisions.
1. Determining the optimal retail price of water
to recommend to IPART.
Table 2 Total Annual Consumption (2019/2020)
2. Plan/determining investments in building
- Period: 2060-2080
future infrastructure.
Model run
Median
% increase
demand GL from base
Table 3 Risk profile of decision frameork
year
CCCMA3.1
498.0
2.86
CCCMA3.1
499.8
3.06
CCCMA3.1
498.4
2.65
CSIRO-MK 3.0
504.3
4.12
CSIRO-MK 3.0 2 500.5
3.31
CSIRO-MK 3.0
500.0
3.24
ECHAM5
502.2
3.82
ECHAM5 1
504.6
4.42
ECHAM5
502.9
3.85
In the case of the retail price determination,
MIROC3.2
494.9
2.10
over-forecasting and under-forecasting, more or
MIROC3.2
497.7
2.75
less have the same impact. Over forecasting
MIROC3.2
494.7
2.00
can lead to lower retail prices (as prices are
1 Suggested for infrastructure invetment
usually set to achieve retail targets) and hence
2 Suggested for price determination
loss of revenue to Sydney Water, whereas
under forecasting can lead to higher cost to
consumers. However, the degree of loss or gain
DISCUSSION
is more or less the same in both cases.
Therefore, the most appropriate forecast to use
The differences between the outputs from the
would be the ensemble member in the middle of
different ensembles is so large that it would be
the range. In the case of determining
investments in future infrastructure, the
hard, if not impossible to plan for all of them,
consequences of under-forecasting, which could
Which, makes it a necessity to select one.
lead to supply constraints or shortages, most
would agree is of graver consequence than
That solution to this does not lie in the science,
wasted investments, building unnecessary
but in the way the science is incorporated into
excess capacity. Further, responding to under-
decision making. We argue that there is no
forecasting would be harder as building
4
infrastructure takes time, whereas excess
Barker, A., Pitman, A. J., Evans, J. P.,
capacity in infrastructure could be absorbed in
Spaninks, F., and Uthayakumaran, L. (2018b).
time as natural growth takes place. It could also
Drivers of future water demand in Sydney
be argued that the cost of error would be felt
Australia: examining the contribution from
more by consumers than Sydney Water.
population and climate change. Submitted to
Journal of Climatology.
Therfore, in the case of determining investments
in infrastructure, the ensemble member
Abrahams, B.,Kumaradevan,S.,Spaninks,F.,and
providing the worst case projection scenario
Sarafidis,V. (2012). An econometric assessment
should be used. Thus, given informing long-term
of pricing Sydney’s residential eater use.
The
planning decisions is the purpose of the long-
Economic Record, 88:89-105.
term demand forecasting, we recommend
ECHAM5, as appropriate for this purpose.
Bun,M.J.G. and Sarafidis, V. (2015) Dynamic
panel data models. In Baltagi. B.H., editor The
Therefore, percentage increase in the near term
Oxford Handbook of Panel Data. Oxford
is, 2.14% and percentage increase in the far
University Press.
term is 4.42%
Wooldridge. J.M, (2010)
Econometric Analysis
REFERENCES
of Cross Section and Panel Data. MIT Press,
2nd edition.
Barker, A., Pitman, A. J., Evans, J. P.,
Spaninks, F., and Uthayakumaran, L. (2018a).
Evans, J. P., Ji, F., Lee, C., Smith, P., Argueso,
Probabilistic forecasts for water consumption in
D., and Fita, L. (2014). Design of a regional
Sydney, Australia from stochastic weather
climate modelling projection ensemble
scenarios and a panel data consumption model.
experiment - NARCliM. Geoscientific Model
Submitted to Water Resources.
Development, 7:621-629.
5
