an integrated approach to simulate stream water quality...
TRANSCRIPT
1
An Integrated Approach to Simulate Stream Water Quality for Municipal Supply Under Changing Climate
Erin Towler: National Center for Atmospheric Research, Boulder, CO, USA
Balaji Rajagopalan: Department of Civil, Environmental and Architectural Engineering,
University of Colorado at Boulder, Boulder, CO, USA; Co-operative Institute for Research in
Environmental Sciences, University of Colorado, Boulder, CO, USA
David Yates: National Center for Atmospheric Research, Boulder, CO, USA
Alfredo Rodriguez: Aurora Water, Aurora CO, USA
R. Scott Summers: Department of Civil, Environmental and Architectural Engineering,
University of Colorado at Boulder, Boulder, CO, USA
Link to paper in the ASCE Civil Engineering Database:
http://cedb.asce.org/cgi/WWWdisplay.cgi?312536
Abstract
To better plan for potential changes to stream water quality under climate change, an
integrated approach to simulate paired streamflow and water quality under a range of climate
scenarios is developed. Several stochastic nonparametric simulation techniques are integrated to
create an end-to-end approach for comprehensive planning, with three steps: (i) develop a
relationship between streamflow and water quality, (ii) simulate streamflow ensembles under
climate change scenarios, and (iii) simulate water quality ensembles using the streamflow
ensembles in conjunction with the developed relationship. The framework is demonstrated on a
municipal water provider developing a new water supply source – but variations of salinity
2
concentrations with streamflow pose limits to its use. For current climate, the simulations
accurately reproduce all of the relevant distributional and threshold statistics of flow and water
quality, providing confidence in their use in long term planning. Under climate change, reduced
streamflow scenarios result in ensembles with higher salinity concentrations, which can be used
in risk management and impact assessments. The approach is general and extends to other water
quality variables associated with hydroclimate.
Subject Headings: water quality; salinity; simulation; water supply; climate change; municipal
water
Introduction
For many municipalities, population and economic growth have spurred increased water
demands, requiring the development of new water sources. This can be challenging, as most
reliable and abundant supplies are often already allocated, and remaining options can be limited
due to issues ranging from investment costs to engineering limitations to public concerns. As
such, new development projects are often carefully weighed and go through extensive planning.
However, few consider the potential effects that a changing climate will have on these new
sources. This is critical as there is mounting evidence that climate change will continue to strain
water supplies, especially in the western United States (US) (Barnett et al. 2004). To date, most
efforts to understand changes have focused on water quantity and have largely overlooked
changes to water quality. This presents another level of complexity for drinking water managers,
but understanding changes to water quality characteristics is critical for comprehensive new
source water management and treatment planning.
The potential impacts of climate on water quality have been broadly identified (see
Murdoch et al. (2000) and Whitehead et al. (2009) and references therein), and are often built
3
upon relationships established between water quantity and quality. For instance, pollutant
concentrations that impact water quality are often associated with streamflow fluctuations
(Johnson 1979; Manczak and Florczyk 1971; Stow and Borsuk 2003), and variability in low
flows can be of particular concern when balancing water quality protection and pollutant
discharges (Saunders and Lewis 2003; Saunders et al. 2004). These relationships have been
exploited to assess water quality, especially recently in predicting total maximum daily loads
(Borsuk et al. 2002), simulating the likelihood of exceeding a turbidity standard using seasonal
forecasts (Towler et al. 2010b) and climate change projections (Towler et al. 2010a), and for
estimating salinity in streamflows using parametric (Mueller et al. 1988) and nonparametric
statistical methodologies (Prairie et al. 2005; Prairie and Rajagopalan 2007). However, these
types of water quality evaluations are rarely incorporated with quantitative evaluations of climate
change. Clearly, there is a need to fully couple climate, streamflow, and water quality assessment
components to ensure that planning strategies are informed by the full range of potential impacts.
To this end, this paper puts forth an integrated stochastic framework that includes
simulation techniques that can be used to develop paired projections of streamflow and water
quality under a range of climate scenarios. This contribution is unique in that it combines
simulation techniques for water quantity and quality to create a new end-to-end approach for
comprehensive planning. The proposed integrated methodology has three steps: (i) develop a
relationship between streamflow and water quality, (ii) simulate streamflow ensembles under
climate change scenarios, and (iii) simulate water quality ensembles using the streamflow
ensembles from (ii) in conjunction with the relationship developed in (i). The approach is
demonstrated on a municipal water provider in Colorado, US, that is developing a new source of
water supply to meet its burgeoning demand – but salinity concentrations, as measured by total
4
dissolved solids (TDS), vary with streamflow and pose limits to its use. Using climate scenarios,
streamflow and salinity simulations are generated that are relevant to the water provider’s
planning, and this application to water quality for municipal water supply is a distinctive aspect
of this study. Further, the potential for using this technique for other water quality variables is
also illustrated. In short, the goal of the study is to develop a general tool that can be used to
characterize water quality variability in current climate, as well as under climate change
scenarios.
Background
Study Area
The approach is demonstrated for the city of Aurora, Colorado, US, a growing suburb of
Denver, which is served by the municipal water provider, Aurora Water. To help meet rising
demand and to reduce drought vulnerability (Pielke et al. 2005), the utility has pursued demand
management options (Kenney et al. 2008) and new source development. The recent Prairie
Waters Project (PWP) is an example of the latter, and is an effort that utilizes water rights that
Aurora Water already owns to increase their supply by 20% (American Water Works
Association (AWWA) 2010). The PWP pumps water out of the South Platte River near the US
Geological Survey (USGS) Henderson streamflow gage (Figure 1 and see upcoming Data
section), which is downstream of a treated wastewater discharge location and is influenced by
return flows from agricultural sources, causing it to have a higher concentration of contaminants
and solutes. As such, the water undergoes multiple processes to ensure public health and
compliance with all Environmental Protection Agency (EPA) primary standards. The treatment
process for the new water source has been designed with multiple barriers, and includes two
main purification steps: first, the water flows through a riverbank filtration system and is pumped
5
into a protected aquifer basin for recharge and recovery at the PWP North Campus, and second,
it undergoes advanced treatment at the Peter Binney Water Purification Facility. The advanced
treatment includes softening, ultraviolet oxidation, filtration, and activated carbon adsorption
(AWWA 2010). However, none of the processes are specifically designed to remove TDS,
which have been observed to range from 200 to 900 mg/L (milligrams per liter; Figure 2). The
secondary EPA standard for salinity (as measured by TDS) is 500 mg/L. The TDS at the intake
location varies throughout the year, and regularly exceeds the EPA standard (Figure 2). To reach
the desired TDS level, the PWP treated water will be blended with water from the nearby Aurora
Reservoir, which is a water supply that has a more constant TDS level of about 250 mg/L.
Clearly, the ability to simulate the TDS in the South Platte under climate change conditions will
be extremely useful to Aurora Water, as increased TDS values will constrain the use of this new
water source and affect their blending planning.
Climate Change Scenarios
Aurora Water draws from surface water supplies primarily from three river basins: the
Arkansas, Platte, and Colorado. Understanding the potential changes to runoff in these basins is
of great interest for Aurora Water, as well as for the many other users who rely on them, and has
been a subject of extensive study (e.g., see Ray et al. 2008 and references therein). Further, the
recent Joint Front Range Climate Change Vulnerability Study (JFRCCVS) quantified streamflow
sensitivity to climate change specifically for drinking water suppliers along the Front Range of
Colorado, including Aurora Water (Woodbury et al. 2011). The study examined several General
Circulation Model-based climate change scenarios in conjunction with two hydrologic models,
the Water Evaluation and Planning (WEAP) model from Stockholm Environment Institute
(Yates et al. 2005a; Yates et al. 2005b), and the Sacramento model used in the National Weather
6
Service River Forecast System. An outcome of this effort was a dataset of simulated streamflow
at specific gage locations throughout the study area, including the South Platte River. For this
study, the simulations for the South Platte River gage at Henderson, which is an approximate
location for the PWP intakes (see upcoming Data section), were examined. Results showed that
except for one climate scenario examined, the majority of scenarios and both hydrologic models
projected annual streamflow reductions in the range of 5 to 36% (Woodbury et al. 2011 or see
Table 1 in Towler et al. 2012). Given these findings and concerns over decreased runoff, it was
determined that four flow volume reduction scenarios in the range of 0% to 30% would be
selected for the approach presented in this paper, which represent likely trajectories under a
changing climate. The first flow scenario, CC0, represents baseline conditions or natural
variability, i.e., the flow is reduced by 0%. Three additional scenarios, CC10, CC20, and CC30,
are subjected to 10, 20, and 30% annual flow volume reductions, respectively. The streamflow
data used to create these scenarios are detailed in the next section.
Data
The two datasets used in this analysis are described below. All years indicated refer to
water years, where the water year spans from October 1 to September 30.
(i) Undepleted streamflow data for the USGS South Platte River at Henderson gage
06720500, which is an approximate location for the PWP intakes, were obtained from Denver
Water, a neighboring municipal water provider, for the period 1947 – 1991. “Undepleted” flows
are calculated to represent what the gage flow would have been if the effects of management –
such as diversions, reservoirs, and return flows – were removed. While an estimate of natural
flow, the term undepleted flow recognizes the fact that certain changes to flow that are unknown
7
or unquantifiable are not accounted for. Nonetheless, undepleted flows are valuable in that they
can provide a baseline for assessing the implications of climate change (Woodbury et al. 2011).
To develop a longer time series, the daily undepleted flow record was reconstructed for
the period of 1992-2008. This was done by utilizing the observed daily streamflow data from the
Henderson gage, or the depleted flows (also referred as ‘gage flows’), which were available from
1926-2008, in conjunction with a relationship identified between the gage flow and undepleted
flow for the overlap period of 1976-1991. The flows for the overlap period exhibit a strong
positive correlation (ρ=0.92, figure not shown). The reader is referred to Towler (2010) for
procedural details, but in general, the functional relationship was applied to the unpaired gage
flows from 1992-2008 to reconstruct the corresponding undepleted flows at the Henderson gage.
This resulted in a final time series of undepleted flows for sixty-two years, 1947 - 2008, where
the daily value matrix, w, is of size 62x365, and the annual value matrix, W, is of size 62x1.
Several measures were tested to evaluate the predictive skill of the model, which indicated that
the model does well in a predictive mode (figures not shown, see Towler (2010)).
(ii) TDS concentrations were calculated from specific conductance data obtained from
the SPCURE (South Platte Coalition for Urban River Evaluation) monitoring network
(http://spcure.org/). The station is identified as METRO SP-124 (South Platte River at 124th
Avenue), which is the approximate location of the Henderson gage, and is available through
EPA’s STORNET database (http://www.epa.gov/storet/). Data was available at a sub-monthly
frequency from 1991-2008, with a sample size of 257. Specific conductance was multiplied by a
factor of 0.64 to obtain TDS values (Snoeyink and Jenkins 1980).
Integrated Framework
8
The integrated methodology involves simulating the streamflow first and subsequently
the stream water quality. The three main steps of the proposed approach are: (1) Develop flow
and TDS relationship, (2) Simulate undepleted flows, and (3) Simulate stream water quality
(TDS). These steps are detailed below.
Step 1: Develop Flow and TDS Relationship
A functional relationship based on local polynomials is developed between the daily TDS
and undepleted flows available for the 1991-2008 period. The scatterplot is shown in Figure 3
and a strong inverse relationship (ρ=-0.68), characteristic of a dilution curve, can be seen. The
relationship is smoothed using a local polynomial technique (Loader 1999), which is a
nonparametric regression technique that ‘‘locally’’ evaluates the function at each desired point
(grey line in the figure). Two parameters need to be estimated for this approach: first the degree
of the polynomial (1 or 2), as well as a smoothing parameter, alpha, which indicates the fraction
of data points included in each estimation (0≤alpha≤1). The parameter combination is selected
by minimizing an objective criteria, the generalized cross validation function (see Loader 1999).
Here, the function was estimated at each point using a second order polynomial (i.e., degree =2)
and all of the data points (i.e., alpha = 1). To do this, the Locfit library (http://cran.r-
project.org/web/packages/locfit/index.html) in the statistical package R (http://www.r-
project.org/) was used. Though the relationship between streamflow and salinity may be
qualitatively known a priori, it is important to fit a model that has the ability to effectively
simulate quantitative estimates. The dynamic nature of nonparametric models provides valuable
flexibility in capturing any arbitrary underlying feature (i.e., linear or nonlinear). A step-by-step
overview of this technique is provided by Prairie et al. (2005).
Step 2: Simulate Undepleted Flows
9
The second part in the framework is to stochastically simulate undepleted flows based on
natural variability (i.e., CC0) and plausible climate change scenarios (i.e., CC10, CC20, CC30).
This is accomplished in two-parts: (1) annual streamflow values are simulated and (2) they are
disaggregated to daily flow values such that they sum to the aggregate annual flow. Below
descriptions of each part are provided.
(1) Annual Streamflow Simulation
A well-known key to water resource management is a strong understanding of
streamflow variability, and there is a rich history of stochastic simulation efforts that offer ways
to develop additional synthetic sequences of hydrology. Linear time series modeling is a well-
developed and widely applied approach (Chatfield 2004; Salas 1985), though it suffers from
several drawbacks, including the assumption of a normal distribution of data and errors, as well
as only being able to model underlying linear features (Rajagopalan et al. 2005). As such,
nonparametric techniques have been explored as a means of providing a more flexible and
general approach.
Nonparametric methods are appealing in that they do not make any prior assumptions
about the underlying structure of the time series. Methods have evolved from the simple index-
sequential method (Kendall and Dracup 1991) to nearest neighbor bootstrap resampling methods
(Efron and Tibshirani 1993; Lall and Sharma 1996) and kernel methods (Sharma et al. 1997), as
well as to more advanced local polynomial models with residual resampling approaches (Prairie
et al. 2006) and incorporation of paleoreconstruction information (Prairie et al. 2008). For a
detailed overview of parametric and nonparametric methods of precipitation and streamflow
simulation, the reader is referred to the review by Rajagopalan et al. (2010).
10
Here, a lag-1 nearest neighbor bootstrapping approach is employed, similar to what is
used by Lall and Sharma (1996), for its simplicity and ease of implementation. Further, this
nonparametric approach has the ability to capture all the relevant statistics, and has been
successfully applied to generate ensembles of daily weather (Rajagopalan and Lall 1999; Yates
et al. 2003; Buishand and Brandsma 2001), streamflows (Lall and Sharma 1996, Grantz et al.
2005; Prairie et al. 2006), and water quality (Prairie et al. 2005; Towler et al. 2009). Though the
reader is referred to Lall and Sharma (1996) for details, an overview of the main steps of the
algorithm is given here:
(i) Randomly resample one of the reconstructed annual flow values, Wi.
(ii) Calculate the scalar distance between this and all the annual flows in the record.
(iii) Sort the distances calculated in (ii) and select K-nearest neighbors. There are
several methods for selecting K , but the heuristic rule, where K is calculated as the square root
of the sample size, or in this case 62=K , with its theoretical justifications (Fukunaga 1972;
Lall and Sharma 1996) has worked well in generating streamflow ensembles (Lall and Sharma
1996, Grantz et al. 2005; Prairie et al. 2006).
(iv) A probability metric is used to assign weights to each of the K-nearest neighbors
given as:
∑=
= K
ii
jjp
1
1
1
for all j = 1,2,…,K.
This results in the closest neighbor receiving the highest weight and the furthest (i.e., the Kth
neighbor) receiving the lowest weight. The cumulative sum of these weights provides a
cumulative distribution function as:
11
∑=
=i
jji pcp
1
for all i = 1,2,….,K.
Other weight functions (e.g., bisquare function) can also be used, though it has been shown that
simulations are robust to the choice of weight function (Lall and Sharma 1996).
(v) One of the K neighbors (i.e., one of the historical years) is resampled, say year t,
using the cumulative weight function described in (iv), and the annual flow corresponding to
year t+1 is the simulated value.
(vi) Steps (ii) through (v) are repeated to generate ensembles (i.e., 250 in this case), each
of 61-years in length to simulate the time horizon of 2010-2070.
The simulated annual values based on the historical streamflows provide a robust
characterization of the natural variability, i.e., CC0. As previously mentioned, three additional
streamflow reduction scenarios are examined in this study: CC10, CC20, and CC30, which
represent annual streamflow reductions of 10, 20, and 30 percent, respectively. Simulations
were developed for a time horizon that spanned from 2010 to 2070 (sixty-one years). For each
scenario, the appropriate linear trend was gradually imposed on each 61-year CC0 flow
simulation. This methodology has also been used to study the water supply risk in the Colorado
River Basin due to climate change (Barnett and Pierce 2008; Rajagopalan et al. 2009).
(2) Disaggregation to Daily Streamflows
The simulated annual streamflows from (1) need to be disaggregated to daily values.
Stochastic disaggregation techniques have been developed by hydrologists and widely used in
basin-wide flow simulation, traditionally using linear approaches (Grygier and Stedinger 1988;
Stedinger and Vogel 1984; Valencia and Schaake 1973) and improved nonparametric techniques
(Prairie et al. 2007; Tarboton et al. 1998). However, disaggregation to finer time scales (i.e.,
12
daily) is computationally challenging, especially when using traditional methods. Recently,
Nowak et al. (2010) developed a simple method based on resampling historical proportion
vectors, which we adopted for this study. The reader is referred to Nowak et al. (2010) for
procedural details and method validation, but here the main steps are provided in brief:
(i) For each year of the historic record, the observed daily streamflow values, w, are
converted to a proportion of the year’s total annual flow. The resulting matrix P, will have
dimensions 62 x 365, where 62 is the number of years of observed data.
(ii) For each simulated annual flow, Z, K-nearest neighbors are identified and one of
them (i.e., one of the historical years, say year y) is selected using the approach outlined in (1)
above. The corresponding proportion vector (py) is applied to the simulated value to obtain the
daily flow vector (zy), such that:
Zpz yy =
(iii) Repeat step (ii) for all the annual flows. Thus, ensembles of daily streamflow
sequences are generated.
Step 3: Simulate Stream Water Quality ( TDS )
The daily simulated flows from the previous step (i.e., Step 2) are used with the
developed functional relationship (i.e., Step 1) to simulate the daily TDS. To characterize the
variability, a nearest neighbor residual resampling technique is used, where residuals are
resampled (i.e., bootstrapped) within a neighborhood of the point estimate and then added to the
mean estimate from the local regression. This method for uncertainty quantification is detailed
in Prairie et al. (2005). This creates daily TDS sequences for each climate change scenario –
CC0, CC10, CC20, and CC30.
Results
13
Streamflow and Salinity: Historic Validation
For streamflow validation, the simulation technique was used to generate ensembles of
daily streamflow of the same length as the historical data (i.e., 62 years) and a suite of
distributional and threshold exceedance statistics were computed at the annual and daily time
scales. The stochastic disaggregation approach to daily streamflow simulation has been well
tested (Nowak et al. 2010), but a sampling from the method validation to this data is provided
here.
The probability density function (PDF) of the historical and simulated annual streamflow
PDF is shown in Figure 4 (left), and it can be seen that the historical PDF is very well captured
by the simulations. This would indicate that all of the annual distributional statistics (such as
mean, variance, skew, lag-1 correlation etc.) are also well reproduced (figures not shown).
The historical PDF of daily streamflows for the month of May and those from the
simulations are shown in Figure 4 (right) – the highly skewed PDF is very well described. For
each month, it was also found that all the distributional properties of daily streamflow (e.g.,
mean, variance, skew, maximums, minimums) are faithfully simulated (figures not shown).
From both a water quantity and quality perspective, drought and threshold exceedance
statistics are of great importance. As such, the annual minimum 7-day-average flow was
computed, which is important for setting pollutant discharge permits, as well as the maximum
number of consecutive days below a flow threshold. Results are shown using the 33rd quantile of
the historical daily flows (Q33) as the threshold, though other thresholds were validated as well
(figures not shown). Box plots of these statistics from the simulations and the corresponding
historical values are shown in Figure 5. The threshold statistics are not all guaranteed to be
reproduced in the model, but the figures show that they are very well captured. It is quite
14
remarkable that a single annual value disaggregated into 365 daily values is able to reproduce the
statistics at all time scales – from daily to annual.
The simulated daily streamflows were used to simulate TDS from the functional
relationship (i.e., Figure 3). However, it is not straightforward to provide a validation of the
TDS simulations similar to the streamflow described above. This is due to the fact that TDS
observations are “snapshots” in time, in that observations are not recorded daily like streamflow.
For instance, the month of June has only sixteen TDS observations for the entire record, whereas
flows are recorded every day. However, one of the key benefits of the proposed stochastic
technique is its ability to use a long record of streamflow data with limited water quality data to
simulate a rich variety of water quality. The box plots of TDS simulations alongside those of the
historical observations for two representative months, January (a relatively dry time of the year)
and June (a wet time of the year due to snowmelt runoff), are shown in Figure 6. As would be
expected, there is more variability from the simulations (black boxes), compared to the historical
variability (grey boxes). Similar results were found for the other months of the year (figures not
included).
The above results demonstrate that the coupled streamflow and TDS simulation
technique is useful, and provides a strong capability to generate rich variety of simulations and
reproduce the historical variability faithfully.
Streamflow and Salinity: Climate Change Simulation
Following the validation efforts, daily flow scenarios were simulated for the selected
climate change scenarios: CC10, CC20, and CC30. Annual streamflows were generated for the
future period of 2010 – 2070, and to this the appropriate linear reduction trend was applied (i.e.,
10%, 20% and 30%, respectively). Then, annual flows were disaggregated to generate daily
15
streamflow sequences corresponding to these reduction trends. The average PDF of the
simulated daily streamflows for the year 2070 for CC0 and CC30 shows the shift towards lower
flows under the flow reduction scenario (Figure 7, left).
The simulated daily streamflows from CC10, CC20, and CC30 were then translated to
climate change scenarios of daily TDS. As expected, the PDFs for CC0 and CC30 show that the
TDS shifts towards higher values with flow reductions (Figure 7, right), which is consistent with
the inverse relationship with streamflow. By computing the area under the curves from zero to
select thresholds, the PDFs can be used to calculated exceedance probabilities (Table 1). In terms
of the EPA secondary standard, the threshold of interest is 500 mg/L; for this there is a 72%
chance of exceedance in the 0% flow reduction case (i.e., natural climate variability). This
increases to 81% with a 30% flow reduction due to climate change. The higher threshold
exceedance probability indicates that in order to meet a constant finished TDS target, there will
be an elevated demand for the lower TDS source water. This is important for planning and
management, and has been found to significantly affect both utility treatment expenses and
residential costs (see Towler et al. 2012).
Extension to Other Water Quality Variables
Though the framework was developed and demonstrated for salinity, it should be pointed
out that the framework is portable to other water quality variables that are related to flow. For
this study site, additional water quality parameters are available through the SPCURE website,
and for illustrative purposes the strong, nonlinear relationship between streamflow and both
nitrogen (N) and phosphorus (P) can be seen (Figure 8). These nutrients play a critical role in
the primary productivity of water bodies, and algal blooms during the growing season are
predicted when total P levels are greater than 0.01 mg/L and/or total N level are greater than 0.15
16
mg/L (Gibson et al. 2000). The utility currently plans for the PWP water to go directly into
treatment, but if raw water storage is ever considered, nutrient simulations would be useful. It is
straightforward to use the proposed integrated approach to simulate ensembles of N and P
(figures not shown), and consequently the probability of exceeding select levels (Table 2). From
the table, it is clear that both thresholds (i.e., 0.01 and 0.15 mg/L for P and N, respectively) are
consistently exceeded, more than 99% of the time for all scenarios. For P, the 1 mg/L level is
regularly exceeded (i.e., 66% in CC0 and 77% in CC30), which would require removal or
dilution of at least 90% to meet the threshold. Similarly for N, the 1.5 mg/L level is exceeded
83% and 88% of the time for CC0 and CC30, respectively, again requiring removal or dilution of
at least 90% to achieve the threshold.
Discussion
The availability of historic streamflow records and water quality observations makes
stochastic simulation an attractive tool for developing future projections. The approach offers an
efficient method to develop ensembles, which effectively characterize the accompanying
uncertainty, and have enormous potential for robust decision-making and impact assessments
(e.g., Grantz et al. 2007; Towler et al. 2012). One limitation is the stationarity assumption of the
functional relationships – namely between the daily flows and TDS values. While the
relationships can be updated with additional data, the assumption is that the historical
relationships remain valid for future climate. This assumption was tested for the historic period,
where an analysis that split the paired data into two halves indicated that the flow-TDS
relationship and the distributions of flow and TDS did not significantly change between the first
and second periods (figures not shown). One advantage of process-based models is that they can
explicitly account for disturbances – such as land use changes or reservoir construction – that
17
might affect the historical relationship. Here, climate scenario selection was informed by two
conceptual hydrologic models (i.e., the aforementioned Sacramento and WEAP models).
Though other watershed-based models have shown promising results for estimating hydrology
and water quality (e.g., Ficklin et al. 2009; Tu 2009; Yoshimura et al. 2009), water quality
modeling remains a challenging task (Rode et al. 2010), and the stochastic approach presented
here provides a straightforward and informative alternative to quantifying water quality changes.
Moreover, the complexity of environmental systems suggests that there may not be a single
“best” modeling approach, and the benefits of combination dynamic-statistical models (Block
and Rajagopalan 2009) and multi-model ensembles (Regonda et al. 2006) have been found.
Further, we point out that other data-driven approaches, such as artificial neural networks, have
been applied successfully in a range of hydro-environmental case studies (e.g., Dawson and
Wilby 1998; Chau et al. 2005; Taormina et al. 2012; Muttil and Chau 2006). For future study, it
would be interesting to compare the results from this study using different data-driven
techniques, such as in Wu et al. (2009).
The method does not explicitly model changes in the timing of spring runoff from
melting snow, which is expected to occur earlier in Colorado due to climate change (Ray et al.
2008; Woodbury et al. 2011). This could prove to be an important attribute of climate change
with profound water quality and management implications. Though not demonstrated here, it
would be easy to shift the flow simulations in this framework to be consistent with projected
changes in peak timing by modifying the nearest neighbor scheme to consider both annual flow
and peak timing in the selection of the proportion vector to be used in the annual flow
disaggregation. The approach can also be modified to simulate multi-site streamflow (Nowak et
al. 2010) and water quality variables that capture the spatial correlations.
18
Conclusions
This paper presents an integrated approach to jointly simulate streamflow and water
quality variables under climate scenarios. The methodology was demonstrated for a new source
water being developed in Aurora, Colorado, where this type of assessment is relevant to their
treatment and blending planning. The approach was validated for natural variability (current
climate), and results show that the simulations accurately reproduce all of the relevant
distributional and threshold statistics of the flow and water quality, providing confidence in their
use in long term planning. Climate change projections suggest reduced flow for this new source
water, and the streamflow reduction scenarios result in a shift of the salinity distribution towards
higher values. The salinity ensembles also provide the probability of exceeding different water
quality values, which can be used to manage risk. These have been used as inputs for a
companion study that examines cost impacts to the utility and residential customers from the
changing salinity (Towler et al. 2012). The potential for using this technique with other water
quality variables that are associated with streamflow is also illustrated.
The paper is distinctive in that it combines and builds on several techniques that
previously have only been used separately in water resources. Further, it extends the influence of
these techniques to understanding water quality variability for new source development for
municipal supply. In short, the integrated framework proposed here offers a simple and robust
planning tool for water utility managers to use and alter specific to their individual needs.
Acknowledgements
The authors would like to acknowledge Water Research Foundation project 3132,
“Incorporating climate change information in water utility planning: A collaborative, decision
19
analytic approach”, the National Water Research Institute (NWRI) through a NWRI fellowship
to the senior author, and the U.S. EPA through a STAR fellowship to the senior author for partial
financial support on this research effort. This publication was developed under a STAR
Research Assistance Agreement No. F08C20433 awarded by the U.S. Environmental Protection
Agency. It has not been formally reviewed by the EPA. The views expressed in this document
are solely those of the authors and the EPA does not endorse any products or commercial
services mentioned in this publication. The first author acknowledges the National Center for
Atmospheric Research (NCAR); NCAR is sponsored by the National Science Foundation.
20
References
AWWA (American Water Works Association) (2010). "Industry news - Colorado's Prairie
Waters project dedicated." Jour. AWWA, 102(11), 88.
Barnett, T., Malone, R., Pennell, W., Stammer, D., Semtner, B., and Washington, W. (2004).
"The effects of climate change on water resources in the West: Introduction and overview."
Clim. Change, 62(1-3), 1-11.
Barnett, T. P., and Pierce, D. W. (2008). "When will Lake Mead go dry?" Water Resour. Res.,
44(3), W03201.
Block, P., and Rajagopalan, B. (2009). "Statistical-dynamical approach for streamflow modeling
at Malakal, Sudan, on the White Nile River." J. Hydrol. Eng., 14(2), 185-196.
Borsuk, M. E., Stow, C. A., Reckhow, K. H. (2002). "Predicting the frequency of water quality
standard violations: A probabilistic approach for TMDL development." Environ. Sci. Technol.,
36(10), 2109-2115.
Buishand, T. A., and Brandsma, T. (2001). “Multisite simulation of daily precipitation and
temperature in the Rhine basin by nearest- neighbor resampling.” Water Resour. Res., 37, 2761–
2776.
21
Chatfield, C. (2004). The analysis of time series: An introduction, 6th Ed., Chapman &
Hall/CRC, Boca Raton.
Chau, K.W., Wu, C.L., Li, Y.S. (2005). “Comparison of several flood forecasting models in
Yangtze river.” J. Hydrol. Eng., 10(6), 485–491.
Dawson, C.W., and Wilby, R. (1998). “An artificial neural network approach to rainfall- runoff
modelling.” Hydrolog. Sci. J., 43(1), 47–66.
Efron, B., and Tibshirani, R. (1993). An introduction to the bootstrap, Chapman & Hall, New
York.
Ficklin, D. L., Luo, Y., Luedeling, E., and Zhang, M. (2009). "Climate change sensitivity
assessment of a highly agricultural watershed using SWAT." J. Hydrol., 374(1-2), 16-29.
Fukunaga, K. (1972). Introduction to statistical pattern recognition, Academic Press, New York.
Gibson, G., Carlson, R., Simpson, J., Smeltzer, E., Gerritson, J., Chapra, S., Heiskary, S., Jones,
J., and Kennedy, R. (2000). "Nutrient criteria technical guidance manual: Lakes and reservoirs,
first edition,” EPA-822-B00-001, Environmental Protection Agency, Washington, DC.
22
Grantz, K., Rajagopalan, B., Clark, M., and Zagona, E. (2005). “A technique for incorporating
large-scale climate information in basin-scale ensemble streamflow forecasts.” Water Resour.
Res., 41, W10410, doi:10.1029/2004WR003467.
Grantz, K., Rajagopalan, B., Zagona, E., and Clark, M. (2007). "Water management applications
of climate-based hydrologic forecasts: Case study of the Truckee-Carson River basin." J. Water
Res. Pl.-ASCE, 133(4), 339-350.
Grygier, J. C., and Stedinger, J. R. (1988). "Condensed disaggregation procedures and
conservation corrections for stochastic hydrology." Water Resour. Res., 24(10), 1574-1584.
Johnson, A. H. (1979). "Estimating solute transport in streams from grab samples." Water
Resour. Res., 15(5), 1224-1228.
Kendall, D. R., and Dracup, J. A. (1991). "A comparison of index-sequential and Ar(1)
generated hydrologic sequences." J. Hydrol., 122(1-4), 335-352.
Kenney, D. S., Goemans, C., Klein, R., Lowrey, J., and Reidy, K. (2008). "Residential water
demand management: Lessons from Aurora, Colorado." J. Am. Water Resour. Assoc., 44(1),
192-207.
Lall, U., and Sharma, A. (1996). "A nearest neighbor bootstrap for resampling hydrologic time
series." Water Resour. Res., 32(3), 679-693.
23
Loader, C. (1999). Local regression and likelihood, Springer, New York.
Manczak, H., and Florczyk, H. (1971). "Interpretation of results from the studies of pollution of
surface flowing waters." Water Res., 5(8), 575-584.
Mueller, D. K., and Osen, L. L. (1988). “Estimation of natural dissolved-solids discharge in the
upper Colorado River basin, Western United States.” Water Resources Investigation Rep. No.
87–4069, United States Geological Survey, Denver, CO.
Murdoch, P. S., Baron, J. S., and Miller, T. L. (2000). "Potential effects of climate change on
surface-water quality in North America." J. Am. Water Resour. Assoc., 36(2), 347-366.
Muttil, N., and Chau K.W. (2006). "Neural network and genetic programming for modelling
coastal algal blooms." International Journal of Environment and Pollution 28 (3-4), 223-238.
Nowak, K., Prairie, J., Rajagopalan, B., and Lall, U. (2010). "A nonparametric stochastic
approach for multisite disaggregation of annual to daily streamflow." Water Resour. Res., 46(8),
W08529.
Pielke, R. A., Doesken, N., Bliss, O., Green, T., Chaffin, C., Salas, J. D., Woodhouse, C. A.,
Lukas, J. J., and Wolter, K. (2005). "Drought 2002 in Colorado: An unprecedented drought or a
routine drought?" Pure Appl. Geophys., 162(8-9), 1455-1479.
24
Prairie, J. R., Rajagopalan, B., Fulp, T. J., and Zagona, E. A. (2006). "Modified K-NN model for
stochastic streamflow simulation." J. Hydrol. Eng., 11(4), 371-378.
Prairie, J. R., Rajagopalan, B., Fulp, T. J., and Zagona, E. A. (2005). "Statistical nonparametric
model for natural salt estimation." J. Environ. Eng.-ASCE, 131(1), 130-138.
Prairie, J., Nowak, K., Rajagopalan, B., Lall, U., and Fulp, T. (2008). "A stochastic
nonparametric approach for streamflow generation combining observational and
paleoreconstructed data." Water Resour. Res., 44(6), W06423.
Prairie, J., Rajagopalan, B., Lall, U., and Fulp, T. (2007). "A stochastic nonparametric technique
for space-time disaggregation of streamflows." Water Resour. Res., 43(3), W03432.
Prairie, J. R., and Rajagopalan, B. (2007). "A basin wide stochastic salinity model." J. Hydrol.,
344(1-2), 43-54.
Rajagopalan, B., and Lall, U. (1999). “A K-nearest-neighbor simulator for daily precipitation and
other weather variables.” Water Resour. Res., 35, 3089-3101.
Rajagopalan, B., Salas, J. D., and Lall, U. (2010). "Stochastic methods for modeling precipitation
and streamflow." Advances in Data-Based Approaches for Hydrologic Modeling and
Forecasting, World Scientific, Singapore.
25
Rajagopalan, B., Grantz, K., Regonda, S. K., Clark, M., and Zagona, E. (2005). "Ensemble
streamflow forecasting: Methods and applications." Advances in Water Science Methodologies,
Taylor and Francis, Netherlands.
Rajagopalan, B., Nowak, K., Prairie, J., Hoerling, M., Harding, B., Barsugli, J., Ray, A., and
Udall, B. (2009). "Water supply risk on the Colorado River: Can management mitigate?" Water
Resour. Res., 45, W08201.
Ray, A. J., Barsugli, J. J., and Averyt, K. B. (2008). Climate change in Colorado, Colorado
Water Conservation Board, Denver, CO.
Regonda, S. K., Rajagopalan, B., Clark, M., and Zagona, E. (2006). "A multimodel ensemble
forecast framework: Application to spring seasonal flows in the Gunnison River Basin." Water
Resour. Res., 42(9), W09404.
Rode, M., Arhonditsis, G., Balin, D., Kebede, T., Krysanova, V., van Griensven, A., and van der
Zee, S. (2010). "New challenges in integrated water quality modelling." Hydrol. Process., DOI:
10.1002/hyp.7766.
Salas, J. D. (1985). "Analysis and modeling of hydrologic time series." Handbook of hydrology,
McGraw-Hill, New York, 19.1-19.72.
26
Saunders, J. F., Murphy, M., Clark, M., and Lewis, M. L. (2004). "The influence of climate
variation on the estimation of low flows used to protect water quality: A nationwide assessment."
J. Am. Water Resour. Assoc., 40(5), 1339-1349.
Saunders, J. F., and Lewis, W. M. (2003). "Implications of climatic variability for regulatory low
flows in the South Platte River Basin, Colorado." J. Am. Water Resour. Assoc., 39(1), 33-45.
Sharma, A., Tarboton, D. G., and Lall, U. (1997). "Streamflow simulation: A Nonparametric
approach." Water Resour. Res., 33(2), 291-308.
Snoeyink, V. L., and Jenkins, D. (1980). Water chemistry, Wiley, New York.
Stedinger, J. R., and Vogel, R. M. (1984). "Disaggregation procedures for generating serially
correlated flow vectors." Water Resour. Res., 20(1), 47-56.
Stow, C. A., and Borsuk, M. E. (2003). "Assessing TMDL effectiveness using flow-adjusted
concentrations: A case study of the Meuse River, North Carolina." Environ. Sci. Technol.,
37(10), 2043-2050.
Taormina, R., Chau, K. W., and Sethi R. (2012). "Artificial neural network simulation of hourly
groundwater levels in a coastal aquifer system of the Venice lagoon", Engineering Applications
of Artificial Intelligence, 25(8), 1670-1676.
27
Tarboton, D. G., Sharma, A., and Lall, U. (1998). "Disaggregation procedures for stochastic
hydrology based on nonparametric density estimation." Water Resour. Res., 34(1), 107-119.
Towler, E. (2010). "Understanding and modeling the impacts of climate change on source water
quality and utility planning." Ph.D. dissertation, Univ. of Colorado at Boulder, Boulder, CO.
Towler, E., Rajagopalan, B., Gilleland, E., Summers, R. S., Yates, D., and Katz, R. W. (2010a).
"Modeling hydrologic and water quality extremes in a changing climate: A statistical approach
based on extreme value theory." Water Resour. Res., 46, W06511, doi:10.1029/2009WR007834.
Towler, E., Rajagopalan, B., Seidel, C., and Summers, R. S. (2009). “Simulating ensembles of
source water quality using a K-nearest neighbor resampling approach.” Environ. Sci. Technol.,
43, 1407-1411.
Towler, E., Rajagopalan, B., Summers, R. S., and Yates, D. (2010b). "An approach for
probabilistic forecasting of seasonal turbidity threshold exceedance." Water Resour. Res., 46,
W11504, doi:10.1029/2009WR008876.
Towler, E., Raucher, B., Rajagopalan, B., Rodriguez, A., Yates, D., and Summers, R.S. (2012).
"Incorporating climate uncertainty in a cost assessment for new municipal source water." J.
Water Res. Pl.-ASCE, 138:396-402, doi:10.1061/(ASCE)WR.1943-5452.0000150.
28
Tu, J. (2009). "Combined impact of climate and land use changes on streamflow and water
quality in Eastern Massachusetts, USA." J. Hydrol., 379(3-4), 268-283.
Valencia, D., and Schaake, J. C. (1973). "Disaggregation processes in stochastic hydrology."
Water Resour. Res., 9(3), 580-585.
Whitehead, P. G., Wilby, R. L., Battarbee, R. W., Kernan, M., and Wade, A. J. (2009). "A
review of the potential impacts of climate change on surface water quality." Hydrolog. Sci. J.,
54(1), 101-123.
Woodbury, M., Yates, D., Baldo, M., and Kaatz, L. (2011). Front range climate change
vulnerability group: A streamflow sensitivity study, Water Research Foundation, Denver, CO.
Wu, C.L., Chau, K.W., and Li, Y.S. (2009). "Predicting monthly streamflow using data-driven
models coupled with data-preprocessing techniques." Water Resour. Res., 45, W08432,
doi:10.1029/2007WR006737.
Yates, D., Gangopadhyay, S., Rajagopalan, B., and Strzepek, K. (2003). “A technique for
generating regional climate scenarios using a nearest-neighbor algorithm.” Water Resour. Res.,
39, 1199.
29
Yates, D., Purkey, D., Sieber, J., Huber-Lee, A., and Galbraith, H. (2005a). "WEAP21 - A
demand-, priority-, and preference-driven water planning model part 2: Aiding freshwater
ecosystem service evaluation." Water Int., 30(4), 501-512.
Yates, D., Sieber, J., Purkey, D., and Huber-Lee, A. (2005b). “WEAP21 - A demand-, priority-,
and preference-driven water planning model part 1: Model characteristics.” Water Int., 30(4),
487-500.
Yoshimura, C., Zhou, M., Kiem, A. S., Fukami, K., Prasantha, H. H. A., Ishidaira, H., and
Takeuchi, K. (2009). "2020s Scenario analysis of nutrient load in the Mekong River Basin using
a distributed hydrological model." Sci. Total Environ., 407(20), 5356-5366.
30
Figures
Figure 1 Schematic of study area (not to scale).
31
Figure 2 Box plots show the observed total dissolved solids (TDS) values by season from 1991-
2008, where OND represents aggregated observations from October, November, and December.
Sample sizes are 51 (OND), 62 (JFM), 62 (AMJ), and 82 (JAS). Horizontal line is the EPA
secondary standard, grey triangles are seasonal averages.
32
Figure 3 Scatterplot between daily undepleted flow and total dissolved solids (TDS) from 1991-
2008. Grey line is local smoother.
Figure 4 Probability density function (PDF) of annual (left) and May daily (right) volumes in
thousand acre-feet (TAF), for all 250 simulations (box plots) and historic observed (grey line).
33
Figure 5 Average minimum 7-day flow (left) and longest consecutive flows below Q33 (33rd
quantile, right) for all 250 simulations (box plots) and observed (grey triangle).
34
Figure 6 Total dissolved solids (TDS) validation between observed (grey box plots) and
simulated (black box plots), and associated sample size (n).
35
Figure 7 Probability density function (PDF) of average daily undepleted flow (left) and total
dissolved solids (TDS; right) for 2070 time slice for natural variability (CC0) and the 30%
reduction (CC30) scenario.
Figure 8 Scatterplot of flow and nitrogen (left, sample size = 570) and phosphorus (right, sample
size = 650). Grey line is local smoother.
36
Tables
Table 1 Probability of Exceeding Select Total Dissolved Solids (TDS) Levels for the Year 2070 for Each Streamflow Reduction Scenario
Probability of Exceeding TDS Level Scenario Reduction (%)
TDS Level
(mg/L) 0 10 20 30 300 92 93 95 96 400 84 86 88 90 500 72 75 77 81 600 47 50 53 59 700 13 14 16 19 800 2.3 2.5 2.5 2.7
37
Table 2 Probability of Exceeding Select Phosphorus and Nitrogen Levels for the Year 2070 for Each Streamflow Reduction Scenario
Probability of Exceeding Level Scenario Reduction (%) Water
Quality Variable
Level (mg/L)
Threshold Reduction
Requirement (%) 0 10 20 30
0.01* - 99.7 99.7 99.8 99.9 0.1 90.0 99.3 99.3 99.5 99.7 1 99.0 65.8 69.5 71.8 76.6 2 99.5 17.0 18.4 19.7 21.9
Phosphorus** (mg/L)
3 99.7 3.70 4.00 4.20 4.50 0.15* - 99.7 99.7 99.8 99.9
1.5 90.0 82.8 83.8 85.8 88.2 3 95.0 58.0 60.6 63.1 67.9 5 97.0 23.4 25.1 26.9 30.4
Nitrogen*** (mg/L)
7 97.9 1.58 1.81 2.09 2.49 * Water quality threshold ** Measured orthophosphate and phosphorus as P *** Measured nitrite and nitrate as N