Environmental Modelling & Software 51 (2014) 12e25

Contents lists available at ScienceDirect

Environmental Modelling & Software

journal homepage: www.elsevier.com/locate/envsoft

Drinking water distribution systems contamination management to

reduce public health impacts and system service interruptions
Amin Rasekh*, Kelly Brumbelow
Zachry Department of Civil Engineering, Texas A&M University, 3136 TAMU, College Station, TX 77843-3136, USA

a r t i c l e i n f o a b s t r a c t

Article history: Decisions on protecting public health against drinking water systems contamination threats should be
Received 2 September 2012 made with careful consideration of credibility of threat observations and adverse impacts of response on
Received in revised form system serviceability. Decision support models are developed in this study to prepare water utility op-
23 September 2013
erators for making these critical decisions during the intense course of an emergency. A pressure-
Accepted 24 September 2013
dependent demand model is developed to simulate the system hydraulics and contaminant propaga-
Available online 17 October 2013
tion under pressure-deficit conditions that emerge after the response actions are executed. Contrary to
conventional demand-driven models, this hydraulic analysis approach prevents potential occurrence of
Keywords:
Water distribution system
negative pressures during the simulation and may identify better response protocols through exploring a
Contamination larger search space. Response mechanisms of contaminant containment and discharge are optimized
Emergency management using evolutionary algorithms to achieve public health protection with minimum service interruption.
Evolutionary computation Sensitivity analyses are conducted to assess optimal response performance for varying response delay,
Multiobjective number of hydrants, and intrusion characteristics. Different methods for quantifying impacts on public
Pressure-dependant demand analysis health and system serviceability are explored and the sensitivity of the optimal response plan to these
different formulations is investigated. The simulation-optimization schemes are demonstrated and
discussed using a virtual water distribution system.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction risk of health impacts has returned to pre-event levels). An emer-

Aging drinking water infrastructure and increased risks of
terrorism have intensified concerns for vulnerability of water dis-
tribution systems (WDS) to accidental and intentional contamina-
tion. A contamination event may cause health and sociopolitical
impacts, erode public trust, and interrupt system operation. To
effectively cope with these threats, there is a need to prepare
contamination emergency management plans that describe the
actions a drinking water utility should take in preparation for and in
response to a contamination threat or incident. An emergency
management plan should be based upon careful risk assessments
and cover the four phases of hazard mitigation, emergency pre-
paredness, emergency response, and disaster recovery (Lindell
et al., 2006).
A contamination emergency response phase is initiated with an
actual (or potential) release of contaminant that spreads across a
WDS, and it extends until the situation is stabilized (i.e., when the
* Corresponding author. Tel.: þ1 979 739 9652; fax: þ1 979 862 1542.
E-mail addresses: aminrasekh@tamu.edu (A. Rasekh), kbrumbelow@
civil.tamu.edu (K. Brumbelow).
gency response protocol explains actions that managers may take
in response to the perceived state of the system after the emer-
gency begins, and it considers how best to achieve managers’
multiple objectives. These response actions can be classified as
“assessment,” “preventive,” or “protective” actions, depending on
whether they collect information about the state of the system,
operate on the system to decrease impacts, or require action by the
public to reduce exposure, respectively (Perry and Lindell, 2007).
Title IV of the Public Health Security and Bioterrorism Prepared-
ness and Response Act of 2002 (United States Congress, 2002) re-
quires all community water systems serving a population greater
than 3300 in the United States to prepare or revise emergency
response plans. The Response Protocol Toolbox (RPT) has been pre-
pared by the United States Environmental Protection Agency (USEPA,
2003) to help water utilities meet this requirement. It provides
general guidelines on how response decisions should be made at the
various stages of a contamination event as more information is
collected. Because this toolbox is essentially a qualitative document,
however, it does not provide specific guidance on how appropriate
response strategies should be devised for a particular WDS. This
study develops quantitative simulation-optimization models to
prepare emergency response protocols that specify functional

1364-8152/$ e see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.envsoft.2013.09.019
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
contaminant containment and flushing operation rules for the
achievement of conflicting response objectives.
2. Problem statement and solution approach
Emergency response is a progressive, interactive, and adaptive
process that includes parallel activities of assessing unusual
contamination observations and making appropriate emergency
response decisions. As more information is obtained about
contamination, emergency management progresses through three
threat stages of “possible,” “credible,” and “confirmatory” (as
described in RPT) accompanied by an increase in seriousness of the
threat impacts and magnitude of response decisions. While public
health protection is the primary response focus, emergency man-
agement should carefully consider other potential consequences on
infrastructure serviceability due to response implementation, spe-
cifically in the early stages of the process where the attack credibility
level is relatively low. At this stage, a multiobjective response plan is
useful to identify the balance between actions taken to protect
public health against a potential threat and limiting overaction that
adversely impacts the ability of the system to meet multiple aspects
of its overall mission. Nevertheless, if streaming threat information
and observations corroborate occurrence of a contamination event,
minimizing potential health impacts becomes the sole primary
objective that should be sought for. Under these circumstances, the
large size of multiobjective optimization results may not be deci-
pherable for making timely emergency decisions and use of a single-
objective model may become preferable.
To date, limited research has addressed this multicriteria nature
of the contamination emergency response problem (Preis and
Ostfeld, 2008; Alfonso et al., 2010). Multiobjective frameworks
proposed so far have only considered hydrant and valve locations as
decision variables and have not optimized the operation timing.
While these studies have considered the number of operational
actions as an emergency response criterion, the important criterion
of system service interruption has not been explicitly addressed.
Moreover, previous single and multiobjective studies have used
demand-driven analysis (DDA) to simulate WDS behavior, and this
assumption inevitably limits the optimization search space to
response protocols that do not cause excessively low pressure in
the WDS. This may unfavorably filter out many possible response
protocols with a high potential for reducing the health impacts.
In the light of these needs, this study develops and integrates a
hydraulic pressure-dependant demand model (PDDM) and evolu-
tionary optimization schemes to find the optimal emergency
response protocols with explicit consideration of two important
response criteria: (1) public health impacts, and (2) system service
interruption. Emergency response is treated as both single and
multiobjective optimization problems to address utility managers’
needs under different conditions. Operational rules for contami-
nant containment and discharge locations and timing are explicitly
treated as optimization decision variables. Sensitivity analyses are
performed to provide insight into effective response protocols and
assess sensitivity of optimal protocols to different parameters such
as response delay. Different formulations for quantifying impacts
on public health and service availability are examined with the help
of the PDDM and an exposure model. Performance of the proposed
schemes is investigated using the WDS of Mesopolis virtual city,
which resembles the interdependency and interconnectedness of
realeworld complex water distribution networks.
3. Model development
Contamination scenario and the simulation and optimization
models will be described in this section. Different attributes of a
13
contamination scenario which is treated as a model input here are
explained and a series of existing contaminant source identification
methods are introduced. This is followed by a detailed description
of the pressure-dependant demand and exposure models for the
simulation of contamination events under pressure-deficit condi-
tions. Optimization objective functions and decision variables are
then explained along with an overview of the evolutionary-
computation-based optimization algorithms used in this study.
3.1. Contamination scenario
Optimization of emergency response protocols needs to be
performed for a given WDS contamination scenario, which is
defined by a set of attributes including: (1) site(s) of contaminant
intrusion, (2) contaminant type, (3) contaminant mass, (4) time of
year, (5) the time of day the contamination event is started, and (6)
the intrusion duration (Rasekh and Brumbelow, 2013). The
simulation-optimization models developed here treat the
contamination scenario as input information. This scenario could
be a potential critical and base scenario for which the response
protocols need to be optimized before an emergency occurs
(Perelman and Ostfeld, 2010; Rasekh and Brumbelow, 2013).
Alternatively, it can be a scenario that is occurring and being
characterized by applying a contaminant source identification
model (Preis and Ostfeld, 2006; Zechman and Ranjithan, 2009;
Hart et al., 2009; Liu et al., 2011a; Rasekh and Brumbelow, 2012;
Gugat, 2012; Shen and McBean, 2013) linked with the sensor
network (Ostfeld et al., 2008; Janke et al., 2009). Therefore, the
simulation-optimization models developed here can be employed
after the potential critical or design basis scenarios have been
characterized during the emergency preparedness phase or a real
contamination scenario has been identified during a contamination
emergency.
3.2. Hydraulic simulation under pressure-deficit conditions
Behavior of a WDS under normal operating conditions is most
commonly simulated using standard DDA models like EPANET
(Rossman, 2000) for design, operation, and rehabilitation purposes.
DDA models are formulated on the premise that nodal water de-
mands are known and completely met during the simulation
period so that nodal pressure and pipe flows can be calculated by
solving a system of quasi-linear equations (Wu et al., 2006). Pro-
jections of network behavior that are based on DDA are reasonably
accurate under normal conditions when pressures are sufficiently
high. Under abnormal pressure-deficit conditions caused by
emergency response actions, however, DDA may illustrate a dis-
torted image of the true system behavior.
A standard DDA solves a system of energy and continuity
equations to calculate unknown nodal heads and pipe flow rates
(Todini and Pilati, 1988; Rossman, 2000). This system of equations
may be extended to include pressure-dependant demand functions
(PDDF) as well to relate pressurized water availability to existing
nodal head under pressure-deficit conditions (Laucelli et al., 2012).
Full pressure-driven analysis approaches solve this extended sys-
tem of equations simultaneously to determine unknown nodal
demands, nodal heads, and pipe flow rates (Laucelli et al., 2012).
Alternatively, the standard DDA model may be iteratively run and
nodal demands are updated sequentially after each iteration using
the PDDFs until a satisfactory convergence is achieved (Liu et al.,
2011b; Jun and Guoping, 2013; Kanta and Brumbelow, in press).
This modeling approach is used in this paper. A major advantage of
this approach is that it allows benefiting from the computational
efficiency and robustness of hydraulic and quality simulators of the
well-established EPANET software. The fact that many existing
14 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25

Fig. 1. Flowcharts of simulation and optimization models.

WDS models are developed in this modeling environment and

8
most water utilities are familiar with EPANET further highlights this > 0 Hi < Himin
>
>
advantage. >
<  1=2
Several PDDFs have been proposed in previous studies that may Qi ¼ nom Hi Himin (1)
> Qi Hiser Himin
Himin < Hi < Hiser
be used for a pressure-driven hydraulic simulation (Shirzad et al. in >
>
>
:
press). This study uses the PDDF proposed by Wagner et al. (1988), Qinom Hi > Hiser
which is among the most popular formulations (Laucelli et al.,
2012) and is the best suited for describing water demand sup- A general framework of the simulator is shown in Fig. 1. Every
plied to customer (Giustolisi and Walski, 2011). Shirzad et al. (in demand node in the WDS belongs to one of the three categories
press) recently conducted a comprehensive laboratory study to defined by Eq. (1) at every iteration and its category may switch
compare different pressure-discharge relations with application in depending on their nodal head at the next iteration. Interested
pressure-driven WDS simulation. They concluded that Wagner’s readers may refer to Appendix A of Kanta and Brumbelow (in press)
PDDF best matched experimental data since it closely resembles for a detailed description of the iterative PDDM used here and step-
the orifice relation. The Wagner’s model is based on the rationale by-step implementation instructions.
that hydraulic laws for flows through devices indicate that the flow The results of PDDM are saved as an output hydraulic file,
is proportional to the square root of head. It considers a parabolic which is then used by the EPANET quality analysis model to
relationship between the service head Hiser for node i and the simulate the propagation of contaminant throughout the WDS.
minimum head Himin needed to satisfy the nominal demand Qinom : The contaminant transport is simulated as a perfect tracer: density
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
effects, decay, and reaction with pipe wall materials and other
existing species are not considered. EPANET-MSX (Multi-Species
eXtension) (Shang et al., 2008) may be used to enable EPANET to
model complex reaction schemes between multiple chemical and
biological species. The obtained WDS quality analysis results,
represented as spatial and temporal concentration of contaminant
in the system, are exported to a human exposure model for eval-
uation of public health impacts.
3.3. Human exposure model
Adverse health impact, defined here as either the number of
sicknesses or the total ingested contaminant mass resulting from a
contamination event, is estimated using the PDDM and water
quality simulation model coupled with an exposure model. The
quantity of contaminant ingested by individuals during a
contamination event depends on water ingestion pattern, time-
varying concentration of contaminant, and availability of drink-
ing water under pressure-deficit conditions. An individual is
assumed to become sick if the cumulative amount of contaminant
ingested during a contamination event exceeds a known toxic
dose. The timing-of-ingestion model selected for this study as-
sumes that tap water is ingested at the common starting times for
the three major meals (7:00, 12:00, and 18:00) and times halfway
between these meals (9:30 and 15:00), given that there is water
available for drinking at a consumer’s demand node. The tap water
intake rate used in the exposure model is central to accuracy of
estimated impacts and is set to 0.93 L/day here based on USEPA
(2004).
3.4. Optimization objective functions
Quantification of emergency response criteria is the first step for
the preparation of the optimization model. Different measures may
be formulated to quantify two important objectives of reducing
public health impacts and system service interruptions. This study
investigates two distinct formulations for each of these criteria and
analyses the sensitivity of optimal response plan to each quantifi-
cation method.
3.4.1. Public health impacts
Some studies have quantified the health impacts in terms of
contaminant concentration in system nodes either as total
contaminant concentration (Baranowski and LeBoeuf, 2008) or
total number of nodes with concentration above a specified
threshold (Alfonso et al., 2010). This approach does not account for
the ingestion timing and rate which can significantly influence
estimated impacts (Davis and Janke, 2008) and does not consider
the critical fact that the connections serving higher populations are
comparatively more vulnerable. In general, health impacts metrics
may be differentiated depending upon whether or not they are
based on a threshold (i.e., number of people/nodes that experience
a concentration or contaminant ingested mass above a threshold)
(Perelman and Ostfeld, 2010; Zechman, 2011; Zechman et al., 2011)
or sum of concentration or consumed mass for all nodes/people
(Baranowski and LeBoeuf, 2008; Preis and Ostfeld, 2008). Accord-
ingly, using the developed exposure model, the health impact cri-
terion is quantified here as either the total number of sicknesses fH1
or the total contaminant mass fH2 ingested by all people during the
whole course of the event:
X
Np
fH1 ¼ ai ; ai ¼ 1 if mi > md ; ai ¼ 0 otherwise
i¼1
15
X
Np
fH2 ¼ mi (3)
i¼1
where Np ¼ the population size served by the WDS; ai ¼ a binary
index; md ¼ known toxic dose; mi ¼ total contaminant mass
ingested by individual i during the whole course of event that is
calculated by the PDDM and exposure model:
X
NI
mi ¼ Vi;j  Ci;j (4)
j¼1
where NI ¼ number of water ingestion events for consumer i,
Vi,j ¼ volume of water ingested by individual i at ingestion event j,
and Ci,j ¼ concentration of contaminant in water volume ingested
by consumer i at ingestion event j. Vi,j ¼ 0.186 liter (calculated
through dividing daily water ingestion rate of 0.93 by 5 daily in-
gestions times) if water is available for individual i at ingestion
event j, and zero, otherwise. Water availability is affected by the
flushing and isolation actions taken by the managers during the
emergency.
Function fH1 explicitly represents the health impacts in terms of
morbidity while fH2 projects the health consequences in the more
implicit form of ingested mass. However, fH2 does not require that
toxic dose be known in advance, which greatly simplifies the
analysis since this value is often difficult to determine and subject
to high variance among individuals. When information is available
to construct a probabilistic doseeresponse model (Chick et al.,
2001), the population fH1 affected by a particular scenario may be
estimated probabilistically with consideration of the dosee
response uncertainties (Ostfeld et al., 2008).
3.4.2. WDS service interruption
Interruption to system serviceability is also mathematically
expressed using two different measures. The first measure is the
total number of hours the volume of water supplied to consumers is
below a certain percentage of their demand. Alternatively, the
second measure sums up the difference between water demand
and supply for all consumers during the whole course of the
emergency. Mathematically,
X
tend X
Nc
fS1 ¼ bti ; bti ¼ 1 if ðSti < lDti Þ; bti ¼ 0 otherwise
t¼0 i¼1
(5)
X
tend X
Nc
fS2 ¼ ðDti  Sti Þ (6)
t¼0 i¼1
where tend ¼ duration of emergency response phase in hours;
Nc ¼ total number of consumers; bti ¼ a binary index;
Sti ¼ volume of water supplied to consumer i at time step t
calculated by the PDDM model; Dti ¼ water demand of consumer
i at time step t; l ¼ supply deficit threshold that ranges from 0 to
100%. The function fS1 treats all consumers equally while fS2
places more emphasis on consumers with higher water demand.
Both these metrics are always unrealistically projected as zero if
DDA is used.
3.5. Optimization decision variables
Fig. 2 shows a schematic arbitrary timeline of contamination
(2) emergency period and contaminant containment and discharge
operations. The optimal order of response actions and their timing
16 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Response Flushing
delay duration
CVS closure
duration
Time
Hydrant CVS closure
open time time
Intrusion duration
Fig. 2. A schematic timeline of a contamination emergency period. The variables that
are optimization decision variables (i.e., elements of a response protocol) are shown in
italics.
are identified by the single and multiobjective optimization
models, which will be described in detail later in Section 3.6.
Contaminant containment through isolation valve operations is
a preventive response mechanism implemented to prevent
contaminant spread to uncontaminated regions of a network and to
preclude consumers from withdrawing contaminated water
(USEPA, 2003). Decisions on combination of valves to be closed and
timing must be made carefully and implemented quickly to be
effective and minimize accompanying impacts on non-
consumptive uses in isolated regions. The decisions include the
time at which closure valve sets (CVS) should be closed after the
response delay (ts¼{ts,1, ts,2,., ts,ns}) and the duration that the CVSs
should remain closed (Dts¼{Dts,1,Dts,2,.,Dts,ns}), where ns is the
total number of CVSs. The optimization model selects a subset of
the whole set of ns CVSs to be operated. When it returns Dts ¼ 0 for a
particular CVS, it means that CVS should not be operated.
The set of CVSs should be configured before the optimization is
performed. While every single pipeline may be theoretically consid-
ered as a possible closure location, this would result in a tremen-
dously large decision space that may not be handled by the optimizer
practically. A large set would also lead to elaborate response protocols
that may be burdensome to execute in practice. CVSs that include
fewer valves will minimize the number of operational actions needed
to isolate an area and are thus more preferable. Assessments need to
Fig. 3. Configuration of closure valve sets (CVS) in Mesopolis.
be performed to identify the most reasonable configuration of these
CVSs. In this study, it is based upon WDS layout, extensive hydraulic
simulations, WDS designers’ understanding of the system behavior,
and computation intensity considerations. Future research may
enhance and automate the configuration process using systematic
optimization (Giustolisi and Savic, 2010), network clustering
(Perelman and Ostfeld, 2011), and minimum-cut set approaches
(Yannopoulos and Spiliotis, 2013), among others.
Contaminant flushing is the other preventive response mecha-
nism that is executed through opening hydrants to discharge a
large volume of contaminated water. System flushing should be
planned and implemented carefully so that it is performed at the
sections of the system where contaminant concentration is higher.
Otherwise, it will worsen the situation by further spread of
contaminated water to uncontaminated areas as it can considerably
alter flow regime (USEPA, 2003; Haxton et al., 2012). The decisions
include selecting the subset hydrants that should be opened (h¼
{h1,h2,.,hnh}) among the complete set of hydrants to flush
contaminated water, time when hydrants should be opened after
the response delay (th¼{th,1,th,2,...,th,nh}), and time period when
hydrants should be remained opened (Dth¼{Dth,1,Dth,2,.,Dth,nh}). A
maximum number of hydrants,nh, may be opened during the
simulation depending on personnel and equipment availability.
Sensitivity analyses are performed on nh to assess how this
parameter affects the response performance.
Implementation of response actions should account for the
response delay times between when intrusion is initiated and
when injection of contaminant in the network is identified as likely
or confirmed through multiple contamination trigger events. It is
also during this time period that the contamination scenario is
identified using a source characterization model as explained in
Section 3.1. Sensitivity analyses are conducted on response delay in
this study to understand how delays influence the effectiveness of
response actions.
3.6. Optimization algorithms
The objective functions described in Eqs. (2)e(6) are minimized
through optimization of preventive response mechanisms of
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Fig. 4. Water consumption patterns for three representative types of consumers in
Mesopolis.
contaminant containment and flushing. Emergency decisions that
should be optimized in response to a contamination threat are
diverse and discrete, and the relationship between decisions and
performance is complex. Evolutionary algorithms have been
demonstrated as flexible and powerful tools for solving such
complicated optimization problems in the field of water resources
management (Reed et al., 2013; Afshar et al., 2009; Rasekh et al.,
2010; Savic et al., 2011; Laucelli and Giustolisi, 2011; Takbiri and
Afshar, 2012; Labadie et al., 2012). A genetic algorithm (GA) is an
adaptive search heuristic that mimics the process of genetic and
natural evolutions (Holland, 1975). It uses strings of decisions var-
iables (also referred to as chromosomes) to represent a possible
solution to the problem. A GA starts with a population of chro-
mosomes and performs recombination and random perturbation to
successively generate better solutions. Solutions with higher fitness
values have a higher chance of survival and being transmitted to the
next generation. Fitness progressively improves through successive
generations and the algorithm converges to an optimal solution.
Single and multiobjective genetic algorithms are developed here
and coupled with the PDDM and exposure model for identification
of emergency response protocols (Fig. 1).
The single-objective model uses an elitist real-coded genetic
algorithm with roulette wheel selection, simulated binary cross-
over (SBX) (Deb and Agrawal, 1995), and polynomial mutation
(Deb, 2001). In the roulette wheel selection, the probability that a
solution will be selected is given by the ratio of its fitness to the
total fitness of other members of the current population. The
offspring hydrants are generated through performing crossover and
mutation operations on longitude and latitude coordinates of the
parent hydrants and picking the hydrant nearest to the resulting
coordinates.
In contrast to the single-objective optimization approach that
searches for a single solution with the best scalar fitness value,
multiobjective optimization seeks a set of trade-off solutions which
together define the best multiobjective alternatives surface called
the Pareto-optimal front (Deb, 2001). Since GAs are population-
based algorithms, they offer a means of finding the complete
Pareto-optimal front in a single optimization run. This study em-
ploys Non-dominated Sorting Genetic Algorithm II (NSGA-II) (Deb
Table 1
Contamination scenarios.
Scenario Location Loading (kilograms) Demand multiplier
West West WTP 300 1.00
East East WTP 300 1.00
17
Fig. 5. Accumulation of total ingested contaminant mass over time. The ultimate value
at the end of the set emergency period (i.e., 6 days) denotes fH2 (see Equation (3) and
Table 1).
et al., 2002), an elitist evolutionary algorithm that benefits from a
fast non-dominated sorting strategy and does not need any user-
defined parameter for preserving diversity in Pareto optimal
surface.
Performance of NSGA-II is evaluated using the hypervolume
measure (Zitzler and Thiele, 1998). Hypervolume is the non-
overlapping volume in objective space covered by members of a
non-dominated ensemble of solutions with respect to a reference
point, set as the worst value for each objective. The hypervolume is
applied to evaluate the convergence of the non-dominated front as
it evolves. Increasing values of the hypervolume necessarily indi-
cate an improvement in the uniformity and proximity to a true
Pareto optimality, which are characteristics that are indicative of
the quality of a non-dominated set of solutions.
4. Application example
Mesopolis virtual city (Johnston and Brumbelow, 2008) is used
here to demonstrate and discuss performance of the proposed
simulation-optimization schemes for the identification of optimal
response protocols. In what follows, a general description of the
Mesopolis WDS, two considered contamination scenarios, and
model settings are provided. The single-objective optimization
model is then applied to find the minimum-health-impacts hy-
drant operation rules and sensitivity analyses are performed on the
number of hydrants and response delay. The multiobjective opti-
mization is finally performed to optimize simultaneous contami-
nant containment and discharge practices for the identification of
the non-dominated response protocols with varying trade-offs
between the response objectives of public health protection and
service interruption prevention.
4.1. WDS, scenarios, and settings
The WDS of Mesopolis is comprised of one reservoir, two water
treatment plants (WTP), 13 tanks, 65 pumps, 876 hydrants, and
2062 water mains (Fig. 3). Demands are exerted at 706 residential,
Start time Duration (hour) fH1 (people) fH2 (grams)
18:00 6 33,944 304.08
19:00 5 54,638 397.28
18 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Fig. 6. Single-objective optimization convergence history for five GA runs. It is asso-
ciated with the East Scenario, a 6-h response delay, and 5 hydrants.
industrial, and commercial/institutional nodes, representing a mid-
size city of nearly 147,000 residents. The diurnal demand patterns
for three representative consumer types are illustrated in Fig. 4.
This study accounts for the spatial and temporal variability of de-
mands but consideration of demand uncertainties requires appli-
cation of optimization algorithms beyond those used here (Jin and
Branke, 2005; Kasprzyk et al., 2012).
Response plans are optimized here for two particular contami-
nation scenarios that are listed in Table 1. The demand multiplier
associated with each scenario is representative of aggregate water
demand for a WDS that typically varies throughout the year. The
contaminant agent is arsenic with a toxic dose of md ¼ 3.5 mg for a
body weight of 70 kg as reported by White (1999). The contami-
nation zone associated with West and East Scenarios is their major
distinction. Contamination of West WTP leads to contamination of
a limited western region, while the entire Mesopolis network is
contaminated if the contamination occurs at the East WTP. Both
contamination scenarios occur in the third day of simulation, which
is when the system has reached dynamic equilibrium. Temporal
accumulation of total ingested contaminant mass as the emergency
proceeds is illustrated in Fig. 5 for both contamination scenarios. It
is observed that the health impacts practically reach a maximum
intensity after 6 days and this duration is thus set as the model
simulation period.
Nine CVSs are considered here as shown in Fig. 3 based upon
WDS connectivity, hydraulic simulations, WDS developers’ under-
standing of the system behavior, and computation intensity limi-
tations. CVS 1, 2, and 3 are located on long mains that transmit
water to far sections of WDS. CVS 4 allows isolation of a large
Fig. 7. Percentage reduction in health impacts for varying number of hydrants for a 6-h response delay using two formulations of (a) total number of sicknesses fH1, and (b) total
ingested mass of contaminant fH2.
eastern area through closing only two mains. Closure of CVS 5 and 6
shuts down the East and West WTPs, respectively. CVS 7 isolates
the highly populated residential area in the western part of the city
from both WTPs by closing only two mains. CVS 8 allows the
western area to be supplied by only West WTP and is thus helpful if
the East WTP is contaminated. Finally, CVS 9 includes three mains
and completely disconnects the eastern network from central and
western areas.
Model emitter discharge coefficients for hydrants are set to
166.5 gpm/psi0.5 (associated with a 3-inch diameter connection fire
hydrant) to calculate pressure-dependant outflow in the PDDM.
Values of Hiser and Himin for the PDDM are based on engineering
design standard of the cities of Bryan and College Station, Texas
(Cities of Bryan and College Station, 2005). Under normal condi-
tions, a design head of 35 psi should be maintained throughout the
system. Water outflow at a connection is assumed to be unavailable
if its pressure drops below 20 psi, the minimum allowed during fire
flow conditions. Because the PDDM developed here involves an
iterative process, a convergence criterion should be devised. It is
considered to be when the percentage of demand nodes whose
current category has been switched falls below a preset value. It is
set to 1% in this study. Under the very rare occasions when the
iteration does not converge after a preset maximum time period,
which is two minutes here, the solution is discarded. The optimi-
zation results show that it occurs to only nearly 0.2% of all solutions.
4.2. Single-objective optimization
The single-objective model is applied first to identify optimal
subset of hydrants and their operation timing that minimize the
health impacts formulated as fH1 and fH2. Appropriate genetic al-
gorithm settings are determined based on sensitivity analyses.
Crossover and mutation rates are 0.80 and 0.05, respectively. SBX
crossover distribution and polynomial mutation indices are set to
15 and 5, respectively. Population size is set to 40, and the model is
run for 200 generations. Fig. 6 shows optimization convergence
during the evolution process of five random GA runs for the East
Scenario, nh ¼ 5, a response delay of 6 h. The observation that all the
randomly seeded runs lead to a similar reduction in health impacts
may be experimentally interpreted as the model’s acceptable
robustness and efficiency in finding optimal response protocols.
Complete demonstration of the model’s performance, however, is
subject to availability of the global optimal response protocols.
Optimal reduction in health impacts with increasing number of
hydrants is illustrated in Fig. 7 for both contamination scenarios
and health impact formulations. Increasing number of flushing
locations leads into greater reduction in health consequences,
which is not surprising. The consistently decreasing slope of the
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Fig. 8. Percentage reduction in health impacts for varying response delay for 5 hydrants using two formulations of (a) total number of sicknesses fH1, and (b) total ingested mass of
contaminant fH2.
curves, however, suggests that opening an additional hydrant is not
as effective when a larger number of hydrants are opened.
Degradation of optimal response protocols performance with
increasing response delay is observed in Fig. 8. Prolonged public
exposure and extended contamination area are majorly responsible
for this decreasing efficiency of emergency response practices. The
response mechanism of contaminant flushing is reasonably effec-
tive for short response delays but its performance rapidly di-
minishes with increasing delay. Figs. 7 and 8 indicate that
contamination would be noticeably better controlled when it
happens at the West WTP than the East WTP. This observation may
be attributed to the smaller impact area associated with the West
Scenario.
Public exposure to contaminated water varies for different
population segments across the city once a contamination event
occurs. The understanding of this variance and how it is affected by
implementing response actions can be improved through analyzing
the cumulative distribution curves of ingested contaminant mass
presented in Fig. 9. The vertical axis indicates the percentage of
total population that has ingested a mass of contaminant below the
corresponding value on the horizontal axis. fH1 is the population on
the vertical axis above the cross point of toxic dose line with each
cumulative curve, while fH2 is the area between the curve and
vertical axis after it is transformed to absolute population values
(i.e., multiplying the percentage values by the total population of
Mesopolis).
Fig. 9 shows the response distribution curves that are based
on the health impact formulation of fH2 are above those based on
fH1 when ingested contaminant mass is larger than the toxic dose
of 3.5 mg. This observation indicates that the optimal response
protocols that are obtained through minimizing fH1 better protect
Fig. 9. Distribution of health impacts for minimum-health-impact protocols for (a) West Scenario and (b) East Scenario. Number of hydrants and response delay are 5 and 6 h,
respectively, for both scenarios.
19
the population sectors that are at high exposure risk (i.e.,
ingested contaminant mass above the toxic dose). Fig. 9(b) shows
that the protocols that are based on fH1 may even magnify the
exposure severity for the population who are at a very high risk
(ingested contaminant mass above nearly 10 mg). These obser-
vations may be explained by the fact that every person in the
population that ingest a mass of contaminant greater than the
toxic dose contribute equally to fH1 measure, no matter how
much greater. This is, however, not true for the fH2 measure.
Considering the population at low risk (ingested mass below the
toxic dose), it is similarly observed that the response protocols
based on fH2 are superior when the East Scenario occurs. When
the West WTP is contaminated, however, this observation does
not hold.
The optimal response protocols associated with the two
contamination scenarios, fH2 health impacts formulation, and
response delays of 6 and 12 h are graphically illustrated in Fig. 10. It
is observed that the optimal hydrant locations are concentrated in a
limited western area when the West WTP is contaminated whereas
the optimal flushing locations are scattered over a larger eastern
region if the East WTP is the contamination site. This is linked to the
impact areas associated with the West and East contamination
scenarios. Provision of these optimal response maps for finer dis-
cretization of response delay and other vulnerable nodes would
serve as a supportive visual aid for utility managers to respond to
the contamination in a more timely and efficient manner.
Fig. 10 also shows the optimal timing associated with the
flushing locations for different scenarios and response delays. It
is clearly observed that the optimal flushing locations move
farther from the contamination site as the response delay in-
creases and the contaminant plume moves further throughout
20 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25

Fig. 10. The optimal discharge locations and timing associated with minimum-health-impact protocols using fH2 formulation.

the system. Moreover, the hydrants would be better opened
immediately after the response delay to minimize the health
impacts, which is not surprising. The optimal values of flushing
durations, however, might not be as predictable; the optimal
durations vary significantly over different discharge locations. An
analysis inspired by this observation shows that setting the du-
rations to maximum possible values leads to degradation of
obtained optimal response protocols performance and this is
consistent for all protocols. This rather unexpected observation
might be linked to the fact that while hydrant operation can
discharge contaminated water, it also changes contaminant
plume and spread across the system. A prolonged flushing may
not thus necessarily lead to an increased performance as it might
divert the contaminant plume to originally safe regions across
the systems.
4.3. Multiobjective optimization
Response mechanisms of contaminant containment and
discharge are next optimized simultaneously to minimize the im-
pacts on public health and system service. The multiobjective
optimization algorithm coupled with the PDDM identifies the most
efficient subset of hydrants and CVSs and their optimal operation
timing for minimizing the health impacts and service interruptions.
The response protocols are optimized for the East Scenario, a
response delay of 6 h, and nh ¼ 5. The total number of decision
variables is thus 33 (15 for contaminant discharge and 18 for CVSs
operation). NSGA-II algorithm population size is 100, and the
optimization evolution process is continued for 200 generations.
Crossover and mutation rates are set to 0.80 and 0.08 and SBX
crossover distribution and polynomial mutation indices are 15 and
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25 21

Fig. 11. Multiobjective optimization convergence history: (a) hypervolume evolution for three NSGA-II runs, (b) non-dominated front evolution corresponding to the hypervolume
curve plotted as red solid dash type. It is associated with the East Scenario, a 6-h response delay, and 5 hydrants.

10, respectively. The tournament size for the NSGA-II selection
operator is set to 3. Additionally, the reference point for calculating
the hypervolume convergence measure is fH1 ¼ 54,638 people and
fS1 ¼ 10,000 block-hours.
Convergence of the hypervolume measure during the evolution
process for three randomly-seeded NSGA-II runs is shown in
Fig. 11(a). This is associated with the East Scenario, response
objective formulations of fH1 and fS1, five hydrants, and a 6-h
response delay. Increasing value of hypervolume as the number of
generations increase is an indicative of convergence to better non-
dominated sets of response protocols as illustrated in Fig. 11(b).
Obtained frontier migrates away from the worst point towards the
coordinates reference point and spreads uniformly across both
axes. The search practically converges and stabilizes after approx-
imately 100 generations.
The obtained Pareto-optimal fronts for four possible combina-
tions of objective functions formulations, which are represented by
Eqs. (2)e(6), are illustrated in Fig. 12. Illustrated Pareto fronts
demonstrate that there are significant trade-offs between the two
response criteria of public health protection and system service-
ability. In light of the fact that impacts on public health are
considered more crucial, one may pose the question whether we
should be concerned about the trade-offs at all. In the trade-off
curves found here, each individual response protocol is associated
with a level of reduction in health impacts and accompanied sys-
tem interruption. A central point to remember is that, while
implementation of such a protocol would lead to system inter-
ruption for certain, reduction in health impacts is conditioned on
the credibility of threat observations (i.e., the RPT sequencing of
possible, credible, and confirmatory threat stages). For the trade-off
analysis to be rational, the horizontal axis should be multiplied by
the probability that the contamination has actually occurred in
order for it to be consistent with the vertical axis representing
unconditioned system interruption. In reality, however, such

Fig. 12. Pareto-optimal fronts for minimizing public health impacts (fH1 and fH2) and system service interruption (fS1 and fS2) for the East Scenario, a 6-h response delay, and 5
hydrants.
22 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25

Fig. 13. The minimum-health-impact protocols using fH1 and fH2 formulations for East Scenario and a 6-h response delay.

probability values are not known exactly and must be inferred from
an ensemble of uncertain sensor triggers and unusual observations.
Thus, the trade-off curves must be understood through this filter of
threat uncertainty.
The maximum reduction in health impacts when the hydrants
are operated solely is marked in Fig. 12. This helps to understand
how much the response performance is improved when the CVSs
are operated as well. It is observed that this leads to a significant
improvement in public health protection. For the fH1efS1 combi-
nation, for instance, the percentage reduction in health impacts is
almost doubled (from 40% to 79%) when the CVSs are operated
together with the hydrants.
Fig. 12 illustrates the non-dominated response protocols in the
objective space. Further analysis has been performed on two
selected minimum-health-impact protocols to understand their
behavior in the decision space and the effect of health impacts
formulation used. Both protocols are highlighted in Fig. 12 and are
associated with the fS2 service interruption formulation and. Fig. 13
illustrates the optimal decision variables values. Spatial spread and
cumulative distribution curves of the health impacts associated
with each of the two response protocols are also shown in Figs. 14
and 15, respectively.
Optimal hydrants and CVSs operation instructions are depicted
in Fig. 13. The hydrants should be opened immediately after the
response delay and flushing duration varies for different hydrants,
which is in agreement with the single-objective optimization
findings. Immediate closure of CVS4 for both protocols blocks
further spread of contaminant to the eastern regions of the city.
Closure of CVS6 is the other common action for the two response
protocols. This creates a pressure gradient, which may guide the
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Fig. 14. Spatial distribution of health impacts for Min-fH1 and Min-fH2 protocols for
East Scenario and a 6-h response delay. Pipes are removed from the map for better
readability.
contaminant plume to the central area that is then partly flushed
out through opening a number of hydrants there.
When it comes to protection of the western Mesopolis from the
contaminant spread, however, the two response protocols based on
the number of sicknesses and total ingested contaminant mass
behave very differently. The min-fH2 protocol requires closing three
CVSs in the western area for a relatively long duration to prevent
Fig. 15. Distribution of health impacts for the minimum-health-impact protocols using
formulations fH1 and fH2 for East Scenario, a 6-h response delay, and 5 hydrants.
23
the contaminant plume from reaching the western population.
Opening the hydrants in the isolated central region will flush out
the contaminant and protects the population in this area from
contaminated water as well. As Fig. 15 also indicates, this isolation
and flushing essentially reduces ingested contaminant mass to zero
for nearly 40% of the total population.
For the min-fH1 protocol, however, similar mechanics of isola-
tion and discharge are not clearly observed; while closure of CVS4
would protect the eastern region, no effective action is taken to
block further propagation of contamination across the western
region as Fig. 13 indicates. Identification of this particular
minimum-health-impact response protocol by the optimization
model is motivated by the expression of health impacts based upon
a preset and fixed threshold. Instead of attempting to protect the
whole population, the model tries to only decrease the number of
people who have ingested a mass of contaminant above that
threshold. This is partially achieved through distributing the
injected mass across a larger area to lower the concentration.
Fig. 14 indicates that the western population does not become
sick even when no response action is taken. Therefore, this popu-
lation does not contribute to the health impacts objective function
when fH1 formulation is used. This clearly explains why no attempt
is made to protect people in the western region for the min-fH1
protocol and health impacts practically remain unchanged. A
similar conclusion may be made for the people residing right on the
northeast downstream of the East WTP, where the health impacts
are noticeably more reduced when the min-fH2 protocol is
executed.
5. Conclusions
This study showed the formulation used for the quantification of
health impacts may have significant effect on the structure and
performance of optimal response protocols. A metric that is based
on a pre-specified exposure threshold has the advantage of being
understandable and explicit as it communicates the impacts
sensibly in terms of number of sicknesses or casualties when
thresholds are known accurately. However, the results of this study
indicated that using the alternative metric of total ingested
contaminant mass may lead to more effective emergency response
recommendations even when exposure thresholds can be esti-
mated accurately. In this sense, quantification of health impacts as
total mass of contaminant ingested by whole population may be
preferable. This formulation also has the advantage that the
optimal response plan found for a specific contamination scenario
is still optimal for any other similar attack scenario with different
contaminant type and loading. This is because this metric is a linear
function of injected mass, whereas the number of sicknesses and
deaths vary nonlinearly as injected mass changes. Ideally, dosee
response curves can be utilized after optimization is performed
based upon total ingested contaminant mass metric to provide a
more explicit and understandable presentation of adverse health
consequences.
It was shown that integrated response of contaminant
containment and discharge may effectively reduce the contami-
nation health impacts. The CVSs were located and configured
manually in this study. Future research may apply systematic ap-
proaches to automate and enhance this process for preparation of
more efficient optimal response protocols. Future work should
examine the incremental effectiveness of valve and hydrants
operation when they are executed in combination with other
response mechanisms such as public warnings disseminated
through news media. Public health can be best protected through a
comprehensive response plan that includes multiple response
strategies.
24 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
Although simultaneous contaminant containment and
discharge operations may significantly reduce the health conse-
quences, it is observed that their performance exponentially de-
grades with delay time. This limits the direct application of the
computationally intensive simulation-optimization schemes dur-
ing an emergency. Instead, these schemes may more appropriately
used to find optimal response protocols for an ensemble of repre-
sentative contamination scenarios called the design basis threats.
Clustering methods based on pattern analysis approaches are
currently being developed for the identification of such represen-
tative scenario. The set of threat-response information e also called
the Response Planning Matrix in RPT (USEPA, 2003) e may be also
used directly after an event happens in real-time or being fed as
input to dynamic simulation-optimization models for adaptation to
the emerging conditions as an emergency proceeds. Efforts are
underway to develop these dynamic decision support models.
Acknowledgements
This material is based upon work supported by the National
Science Foundation under Grant Number CMMI-0927739. Any
opinions, findings, or recommendations expressed in this paper are
those of the authors and do not necessarily reflect the views of the
National Science Foundation. We also gratefully acknowledge the
constructive and helpful comments of three anonymous reviewers
that significantly improved this paper.
References
Afshar, A., Rasekh, A., Afshar, M.H., 2009. Risk-based optimization of large
flood-diversion systems using genetic algorithms. Eng. Optim. 41 (3), 259e
273.
Alfonso, L., Jonoski, A., Solomatine, D., 2010. Multiobjective optimization of oper-
ational responses for contaminant flushing in water distribution networks.
J. Water Res. Plan. Manag. 136 (1), 48e58.
Baranowski, T.M., LeBoeuf, E.J., 2008. Consequence management utilizing optimi-
zation. J. Water Res. Plan. Manag. 134 (4), 386e394.
Chick, S.E., Koopman, J.S., Soorapanth, S., Brown, M.E., 2001. Infection transmission
system models for microbial risk assessment. Sci. Total Environ. 274 (1), 197e
207.
Cities of Bryan and College Station, 2005. B/CS Unified Design Guideline Manual,
Water, Sewer, and Streets. Bryan and College Station, Available www.bcsunited.
net/2005Files/DesignManual2005.pdf.
Davis, M.J., Janke, R., 2008. Importance of exposure model in estimating impacts
when a water distribution system is contaminated. J. Water Res. Plan. Manag.
134 (5), 449e456.
Deb, K., 2001. Multiobjective Optimization Using Evolutionary Algorithms. Wiley,
Chichester.
Deb, K., Agrawal, R.B., 1995. Simulated binary crossover for continuous search space.
Complex Sys. 9, 115e148.
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multi-objective
genetic algorithm NSGA-II. IEEE Trans. Evol. Comput. 6 (2), 182e197.
Giustolisi, O., Savic, D., 2010. Identification of segments and optimal isolation valve
system design in water distribution networks. Urban Water J. 7 (1), 1e15.
Giustolisi, O., Walski, T.M., 2011. Demand components in water distribution
network analysis. J. Water Res. Plan. Manag. 138 (4), 356e367.
Gugat, M., 2012. Contamination source determination in water distribution net-
works. SIAM J. Appl. Math. 72 (6), 1772e1791.
Hart, D.B., McKenna, S.A., Klise, K.A., Wilson, M.P., Murray, R., 2009. CANARY User’s
Manual and Software Upgrades. U.S. Environmental Protection Agency, Wash-
ington, D.C.. EPA/600/R-08/040A.
Haxton, T., Murray, R., Klise, K., 2012. Examining the Application of Modeling Tools
to Identify Effective Flushing Locations. World Environmental and Water Re-
sources Congress, Palm Springs, CA.
Holland, J.H., 1975. Adaptation in Natural and Artificial Systems. University of
Michigan Press, Ann Arbor, Mich.
Janke, R.J., Murray, R., Haxton, T.M., Taxon, T., Bahadur, R., Samuels, W., Berry, J.,
Boman, E., Hart, W., Riesen, L., Uber, J., Morley, K., 2009. Threat Ensemble
Vulnerability Assessment-Sensor Placement Optimization Tool (TEVA-spot)
Graphical User Interface. U.S. Environmental Protection Agency, Washington,
DC. EPA/600/C-08/010, 2009.
Jin, Y., Branke, J., 2005. Evolutionary optimization in uncertain environments-a
survey. IEEE Trans. Evol. Comput. 9 (3), 303e317.
Johnston, G., Brumbelow, K., 2008. Developing Mesopolis: A ‘virtual City’ for
Research in Water Distribution Systems and Interdependent Infrastructures.
https://ceprofs.tamu.edu/kbrumbelow/Mesopolis/Developing%20Mesopolis.
pdf.
Jun, L., Guoping, Y., 2013. Iterative methodology of pressure-dependent demand
based on EPANET for pressure-deficient water distribution analysis. J. Water
Res. Plan. Manag. 139 (1), 34e44.
Kanta, L., Brumbelow, K., 2013. Vulnerability, risk, and mitigation assessment of
water distribution systems for insufficient fire flows. J. Water Res. Plan. Manag..
http://dx.doi.org/10.1061/(ASCE)WR.1943-5452.0000281 (in press).
Kasprzyk, J.R., Reed, P.M., Characklis, G.W., Kirsch, B.R., 2012. Many-objective de
Novo water supply portfolio planning under deep uncertainty. Environ. Model.
Software 34, 87e104.
Labadie, J.W., Zheng, F., Wan, Y., 2012. Optimal integrated operation of reservoir-
assisted stormwater treatment areas for estuarine habitat restoration. Envi-
ron. Model. Software 38, 271e282.
Laucelli, D., Berardi, L., Giustolisi, O., 2012. Assessing climate change and asset
deterioration impacts on water distribution networks: demand-driven or
pressure-driven network modeling? Environ. Model. Software 37, 206e216.
Laucelli, D., Giustolisi, O., 2011. Scour depth modelling by a multi-objective evolu-
tionary paradigm. Environ. Model. Software 26 (4), 498e509.
Lindell, M.K., Prater, C.S., Perry, R.W., 2006. Fundamentals of Emergency Manage-
ment. Federal Emergency Management Agency, Hyattsville, MD.
Liu, L., Ranjithan, S.R., Mahinthakumar, G., 2011a. Contamination source identifi-
cation in water distribution systems using an adaptive dynamic optimization
procedure. J. Water Res. Plan. Manag. 137 (2), 183e192.
Liu, J., Yu, G., Savic, D., 2011b. Deficient-network simulation considering pressure-
dependent demand. In: International Conference on Pipelines and Trenchless
Technology (ICPTT), Beijing, China.
Ostfeld, A., et al., 2008. The battle of the water sensor networks (BWSN): a design
challenge for engineers and algorithms. J. Water Res. Plan. Manag. 134 (6), 556e
568.
Perelman, L., Ostfeld, A., 2010. Extreme impact contamination events sampling
for water distribution systems security. J. Water Res. Plan. Manag. 136 (1),
80e87.
Perelman, L., Ostfeld, A., 2011. Topological clustering for water distribution systems
analysis. Environ. Model. Software 26 (7), 969e972.
Perry, R.W., Lindell, M.K., 2007. Emergency Planning. John Wiley, Hoboken NJ.
Preis, A., Ostfeld, A., 2006. Contamination source identification in water systems: a
hybrid model treeselinear programming scheme. J. Water Res. Plan. Manag. 132
(4), 263e273.
Preis, A., Ostfeld, A., 2008. Multiobjective contaminant response modelling for
water distributions systems security. J. Hydroinf. 10 (4), 267e274.
Rasekh, A., Afshar, A., Afshar, M.H., 2010. Risk-cost optimization of hydraulic
structures: methodology and case study. Water Res. Manag. 24 (11), 2833e
2851.
Rasekh, A., Brumbelow, K., 2012. Machine Learning Approach for Contamination
Source Identification in Water Distribution Systems. World Environmental and
Water Resources Congress, Palm Springs, CA.
Rasekh, A., Brumbelow, K., 2013. Probabilistic analysis and optimization to char-
acterize critical water distribution system contamination scenarios. J. Water
Res. Plan. Manag. 139 (2), 191e199.
Reed, P.M., Hadka, D., Herman, J.D., Kasprzyk, J.R., Kollat, J.B., 2013. Evolutionary
multiobjective optimization in water resources: the past, present, and future.
Adv. Water Res. 51, 438e456.
Rossman, L.A., 2000. EPANET 2.0, User’s Manual. National Risk Management
Research Laboratory, U.S. EPA, Cincinnati, OH.
Savic, D.A., Bicik, J., Morley, M.S., 2011. A DSS generator for multiobjective optimi-
sation of spreadsheet-based models. Environ. Model. Software 26, 551e561.
Shang, F., Uber, J.G., Rossman, L., 2008. EPANET Multi-species Extension Software.
U.S. Environmental Protection Agency, Washington, DC. EPA/600/C-10/002.
Shen, H., McBean, E.A., 2013. Application of parallel computing in data mining for
contaminant source identification in water distribution systems. Can. Water
Res. J. 38 (1), 34e39.
Shirzad, A., Tabesh, M., Farmani, R., Mohammadi, M., 2012. Pressure-discharge re-
lations with application in head driven simulation of water distribution net-
works. J. Water Res. Plan. Manag.. http://dx.doi.org/10.1061/(ASCE)WR.1943-
5452.0000305 (in press).
Takbiri, Z., Afshar, A., 2012. Multi-objective optimization of fusegates system under
hydrologic uncertainties. Water Res. Manag. 26 (8), 2323e2345.
Todini, E., Pilati, S., 1988. A gradient algorithm for the analysis of pipe networks. In:
Coulbeck, B., Orr, Chun-Hou (Eds.), Computer Applications in Water Supply.
Wiley: Research Studies Press, pp. 1e20.
United States Congress, 2002. Public Health Security and Bioterrorism Preparedness
and Response act of 2002. Public Law No. 107-188, H.R. 3448, 107th Congress.
Government Printing Office, Washington, D.C.
United States Environmental Protection Agency (USEPA), 2003. Response Protocol
Toolbox: Planning for and Responding to Drinking Water Contamination
Threats and Incidents. USEPA, Washington, D.C.
United States Environmental Protection Agency (USEPA), 2004. Estimated per capita
water ingestion in the United States. USEPA, Office of Water, Washington, D.C.
Wagner, J.M., Shamir, U., Marks, D.H., 1988. Water distribution reliability: simulation
methods. J. Water Res. Plan. Manag. 114 (3), 276e294.
White, J.W., 1999. Hazards of Short-term Exposure to Arsenic-contaminated Soil.
Office of Environmental Health Assessment Services, Olympia, WA.
Wu, Z.Y., Wang, R.H., Walski, T.M., Yang, S.Y., Bowdler, D., Baggett, C.C., 2006. Effi-
cient pressure dependent demand model for large water distribution system
 A. Rasekh, K. Brumbelow / Environmental Modelling & Software 51 (2014) 12e25
analysis. In: 8th Annual Water Distribution Systems Analysis Symposium,
Cincinnati, Ohio.
Yannopoulos, S., Spiliotis, M., 2013. Water distribution system reliability based on
minimum cuteset approach and the hydraulic availability. Water Res. Manag.
27 (6), 1821e1836.
Zechman, E.M., Brumbelow, K., Lindell, M., Mumpower, J., Rasekh, A., Shafiee, M.,
2011. Including consumer behavior and system dynamics in vulnerability
analysis for water distribution system contamination events. In: NSF CMMI
Engineering Research and Innovation Conference, Atlanta, GA.
25
Zechman, E.M., 2011. Agent-based modeling to simulate contamination events and
evaluate threat management strategies in water distribution systems. Risk Anal.
31 (5), 758e772.
Zechman, E.M., Ranjithan, S., 2009. Evolutionary computation-based methods for
characterizing contaminant sources in a water distribution system. J. Water Res.
Plan. Manag. 135 (5), 334e343.
Zitzler, E., Thiele, L., 1998. Multiobjective Optimization Using Evolutionary Algo-
rithms e a Comparative Case Study. Parallel Problem Solving from
NaturedPPSN V. Springer, Berlin/Heidelberg.