ww.sciencedirect.com

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3

Available online at w

ScienceDirect

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

Optimal sensor placement for detectingorganophosphate intrusions into waterdistribution systems

Ziv Ohar, Ori Lahav, Avi Ostfeld*

Faculty of Civil and Environmental Engineering, Technion e Israel Institute of Technology, Haifa 32000, Israel

a r t i c l e i n f o

Article history:

Received 6 October 2014

Received in revised form

14 January 2015

Accepted 16 January 2015

Available online 28 January 2015

Keywords:

Water distribution systems

Sensor placement

Organophosphates

Water security

EPANET-MSX

Genetic algorithms

* Corresponding author. Tel.: þ972 4 8292782E-mail address: ostfeld@tx.technion.ac.il

http://dx.doi.org/10.1016/j.watres.2015.01.0240043-1354/© 2015 Elsevier Ltd. All rights rese

a b s t r a c t

Placement of water quality sensors in a water distribution system is a common approach

for minimizing contamination intrusion risks. This study incorporates detailed chemistry

of organophosphate contaminations into the problem of sensor placement and links

quantitative measures of the affected population as a result of such intrusions. The sug-

gested methodology utilizes the stoichiometry and kinetics of the reactions between

organophosphate contaminants and free chlorine for predicting the number of affected

consumers. This is accomplished through linking a multi-species water quality model and

a statistical doseeresponse model. Three organophosphates (chlorpyrifos, malathion, and

parathion) are tested as possible contaminants. Their corresponding by-products were

modeled and accounted for in the affected consumers impact calculations. The method-

ology incorporates a series of randomly generated intrusion events linked to a genetic

algorithm for minimizing the contaminants impact through a sensors system. Three

example applications are explored for demonstrating the model capabilities through base

runs and sensitivity analyses.

© 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Water distribution systems (WDS) deliver water from water

sources to water consumers via a large network of pipes,

water treatment plants, pumping stations, storage devises

and many other infrastructure components. Given their large

size and numerous access points water distribution systems

can be vulnerable to accidental or deliberate contamination

events.

Detection by water quality sensors has the potential to

mitigate the effects of contamination events and improve

water security. This understanding had led to considerable

; fax: þ972 4 8228898.(A. Ostfeld).

rved.

amount of work conducted over the past decade on sensor

network placement strategies and optimization in WDS,

resulting in over 100 published papers (Hart andMurray, 2010).

A large number of contaminants can be introduced into a

WDS. These can be further transported by thewater to various

locations in the network. Three members of the organophos-

phates group, chlorpyrifos (CP), parathion (PA) and malathion

(MA), which are commonly used as pesticides, were examined

here as possible contaminants. The objective of this study is to

increase the reliability of water quality sensor locations in

WDS methodologies by incorporating detailed reaction

chemistry between possible contaminants, in this case or-

ganophosphates, and free chlorine containing water. The

Table 1 e Oragnophosphats pesticide and daughtercompounds solubility and oral LD50.

Organophosphatspesticide

Solubility(mg/l)

Oral LD50

(mg kg�1)

Chlorpyrifos (CP) 2 a135

Chlorpyrifos oxon (CPO) 140 b2

3,5,6-trichloro-2-pyridinol (CPH) 36,000 NA

Malathion (MA) 148 a2100

Malaoxon (MAO) 38,000 c158

Parathion (PA) 12.4 a13

Paraoxon (PAO) 880 d1.8

p-nitrophenol (PAH) 3200 e202

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3194

methodology developed herein consists of a multi-species

water quality model based on the suggested reactions re-

ported by Schwartz et al. (2014), and a genetic algorithm

optimization process to find optimal sensor locations. The

objective function is a modification of Ostfeld et al. (2008).

Indexes were added to the original formulas in order to extend

it for a multi-species framework. By doing so, the effect of the

contaminants and their by-products on the consumers' couldbe quantified. Three example applications, with increasing

complexity, were used to demonstrate the methodology per-

formance and limitations through base runs and several

sensitivity analyses.

a From WHO, 2009.b From Lockridge et al., 2005.c From Lewis 1996.d From Tsang et al., 2004.e From Isayev et al., 2006.

Fig. 1 e Solution scheme.

2. Literature review

Thework presented in this paper forms part of a field aimed at

enhancing the security of water supply systems by using

water quality sensors for early detection of contamination

events. The general field includes studies on optimal sensor

layout, contamination source identifying (Tryby et al., 2010;

Kumar et al., 2012), response modeling and decision making

following contaminant detection (Preis and Ostfeld. 2008a;

Alfonso et al., 2010; Rasekh and Brumbelow. 2014) and event

detection models (Murry et al., 2010; Oliker and Ostfeld. 2014),

which are based on the statistical analysis of a data series in

order to learn the behavior of water quality parameters and to

identify deviation from normal behavior.

Over 100 studies have already been published in the field of

optimal sensor placement. A review of the vast majority of

these can be found in Hart and Murray (2010). In order to

compare several algorithms a “battle of the water sensor

networks” (BWSN) was conducted (Ostfeld et al., 2008) in

which 15 different approaches were compared. The compar-

ison consisted of four objectivemeasures: (1) expected time of

detection; (2) effected population prior to detection; (3) con-

sumption of contaminated water prior to detection; and (4)

maximization of detection likelihood. A winner was not

declared due to the multi-objective nature of the above

targets.

In most cases researchers tackle the problem of optimal

sensor locations by assuming ideal sensors (Hart and Murray,

2010). In contrast, Preis and Ostfeld (2008b) formulated a

methodology for contamination source identification using

information from imperfect sensors. Ostfeld and Salomons

(2004) generated a matrix of compromised nodes due to

random contamination events; and a genetic algorithm was

used to obtain the optimal coverage of the pollution matrix.

Selectionmethods of the contamination events are additional

sub-field of the sensors location problem. Perelman and

Ostfeld (2010, 2012) and Rasekh and Brumbelow (2013)

focused on finding extreme impact events, whereas Davis

and Janke (2011) aimed at establishing pattern behavior of

potential impact associated with contamination events. The

above studies are engineering oriented, i.e. chemical aspects

of the optional contaminants were not addressed.

Organophosphates are widely used pesticides, which

makes them relatively accessible. Consumption of organo-

phosphate can result in malfunction of the nervous system or

even death, if consumed in large quantities. Chlorine, which is

the most commonly used disinfectant for potable water, is

known to oxidize organophosphates to form their corre-

sponding oxon (Wu and Laird. 2003), which is typically much

more toxic than the original contaminant. This property is

shown in Table 1 which specifies the oral lethal reference

doses (LD50) of the organophosphate species modeled in this

work and that of their oxidation by-products. Organophos-

phates have the potential to pose a significant threat to water

consumers in a contamination event. No toxic cumulative

relations were found in the literature between mother and

daughter compounds, therefore in the incidents calculations

described below, themaximal potential incidents between the

two was used.

Durik et al. (2009) investigated the fate of several organo-

phosphates in chlorinated water including chlorpyrifos (CP),

parathion (PA) and malathion (MA). It was found that these

organophosphates have similar degradation patterns when

reacted with free chlorine. Organophosphates (OP) are rapidly

oxidized by hypochlorous acid (HOCl) to form the corre-

sponding oxon (OPO) which is more toxic than the parent

pesticide. Both OP and OPOmay undergo further hydrolysis to

form OPH, a stable and less toxic group of end products. Hy-

pochlorite ion (OCl�) was found not to oxidize OP pesticides

but rather to act as a nucleophile, accelerating hydrolysis at

high pH values. This work is discussed further in the water

quality model formulation section.

Table 2 e Reaction stoichiometry and rate coefficients for the Organophosphate degradation.

Reaction stoichiometry Rate/equilibrium coefficient at 25 (�C)

1 5HOCl þOP %kHOCl;OP

OPOþ 5Hþ þ 5Cl� þ SO2�4 kHOCl;OP ¼ Table 3

2 OP %kh;OP

OPH kh;OP ¼ kN;OP þ kB;OP,½OH��kN;OP ¼ Table 3

kB;OP ¼ Table 3

3 OPO %kh;OPO

OPH kh;OPO ¼ kN;OPO þ kB;OPO,½OH��kN;OPO ¼ Table 3

4 OPþOCl� %kOCl;OP

OPH kOCl;OP ¼ Table 3

5 OPO þOCl� %kOCl;OPO

OPH kOCl;OPO ¼ Table 3

6 HOCl /Hþ þOCl� pKa ¼ 7:5

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3 195

As mentioned, none of the previous studies has incorpo-

rated detailed chemical reactions as part of the solution to the

problem of optimal sensor locations. The incorporation of the

detailed chemistry provides the ability to estimate the con-

centration of both the contaminant and its degradation by

products throughout the WDS, and accordingly to provide an

assessment of the potential number of incidents according to

their toxicity values of either.

3. Model formulation

As a preliminary step the water quality model was run

without contamination injection for numerous demand cy-

cles until a periodic quality behavior was observed. By doing

so, stable quality periodic conditions were ensured

throughout the network and initial quality conditions could be

set. Additional discussion on similar water quality model

initial conditions can be found in Ohar and Ostfeld (2014).

The main part of the methodology is comprised of two

stages: (1) construction of affected population matrix (APM);

and (2) minimization of expected affected population prior to

detection using a genetic algorithm (GA). The suggested so-

lution scheme is described in Fig. 1.

3.1. Stage I e construction of affected population matrix

The first stage was the construction of an APM. The APM is an

ðn�mÞ double variable matrix, containing the calculated

affected population due to random contamination events, in

which the rows ðnÞ refer to the water distribution system

consumer nodes and the columns ðmÞ refer to any given time

step during the contamination scenario. The random

contamination event had three degrees of freedom: injection

location, injection time and the injected contamination

Table 3 e Chlorination and hydrolysis rate coefficients for orgatransformation products.

Organophosphate pesticide kN,OP (h�1) kB,OP (M�1 h�

Chlorpyrifos 0.000372 37.1

Chlorpyrifos oxon 0.00213 230.2

Malathion 0.0000792 1980

Malaoxon

Parathion 0.000266 4.3

Paraoxon 0.0002 46.1

substance (CP, MA or PA). The injected contaminant concen-

tration was the contaminant solubility limit, since higher

concentrations would increase water turbidity and would be

detected by the consumers. In addition to the random events

matrix the construction of the APM had two phases, the first

was a water quality simulation, applied in order to calculate

the constituents' concentrations throughout the WDS and

time, and the second, calculation of the affected population

due to the above concentrations and due to water

consumption.

3.2. Water quality model formulation

Each contamination scenario was simulated in EPANET-MSX

(Shang et al., 2008) resulting in contaminant and contami-

nant reaction by-products concentrations throughout the

WDS and time. EPANET-MSX is a program which enables the

modeling of complex reaction schemes between multiple

chemical species in WDS, based on the hydraulic engine of

EPANET. The MSX input is a set of equilibrium and ordinary

differential equations. For the case of organophosphates the

above mentioned set of equations are based on the organo-

phosphates degradation kinetics suggested by Duirk et al.

(2009) and Schwartz et al. (2014), which enhanced the former

for WDS conditions.

The OP water quality model simulates the fate of OP

following reaction with free chlorine. In addition to the

degradation equations a set of equations was added in order

to calculate measurable parameters, i.e. alkalinity, pH and

residual chlorine concentration, which can assist in detecting

the event by water quality sensors (Schwartz et al., 2014). The

proposed model also takes into account several conservative

constituents and parameters which are needed in order to

calculate the effect of mixing of different water sources. The

reaction mechanism, stoichiometry and kinetics used in the

simulation are described in Tables 2 and 3 and Eqs. (1)e(3).

nophosphates pesticides and their corresponding oxon

1) kHOCl,OP (M�1 h�1) kOCl,OP or kOCl,OPO (M�1 h�1)

1720000 990

1340

1720000 382

565

2200000 37

48

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3196

d½OP�dt

¼ �kHOCl;OP½HOCl�½OP� � kh;OP½OP� � kOCl;OP

�OCl�

�½OP� (1)

d½OPO�dt

¼ kHOCl;OP½HOCl�½OP� � kh;OPO½OPO� � kOCl;OPO

�OCl�

�½OPO�(2)

d½OPH�dt

¼ kh;OP½OP� þ kOCl;OP

�OCl�

�½OP� þ kh;OPO½OPO�þ kOCl;OPO

�OCl�

�½OPO� (3)

where: ½OP�, ½OPO� and ½OPH� are the mother and daughter

organophosphate compounds; kHOCl;OP and kOCl;OP are the

intrinsic rate coefficients for HOCl and OCl� reacting with

each OP pesticide; kOCl;OPO is the intrinsic rate coefficients

assisting in the hydrolysis of the OPO; and kh;OP; kh;OPO are the

hydrolysis rate coefficients of ½OP� and ½OPO� to ½OPH�.Since HOCl is significantly more reactive than OCl�, as

previously observed with OP pesticides and other anthropo-

genic chemicals (Duirk et al., 2009), the effect of OCl� on the

total free chlorine degradation can be neglected, resulting in:

d½Cl2T�dt

¼ �5kHOCl;OP½HOCl�½OP� � kOCl;OP

�OCl�

�½OP�� kOCl;OPO

�OCl�

�½OPO� � kb½Cl2T� (4)

where: kb½Cl2T� represent normal observed first order decay of

free chlorine with kb ¼ 1 ð1=dÞ.From the above stoichiometry the rate equation for alka-

linity can be derived as:

dðalkÞdt

¼ �kHOCl;OP½HOCl�½OP� � kOCl;OP

�OCl�

�½OP�� kOCl;OPO

�OCl�

�½OPO� (5)

In additional calculations, ionic strength was calculated

using Kemp's equation as a function of TDS (Schwartz et al.,

2014). The ions' activity coefficients were calculated as a

function of the ionic strength, temperature and the valence of

each specific ion using Davis equation, in order to determine

the equilibrium constants for the hypochlorous acid, carbon-

ate and water weak-acid systems.

Fig. 2 e EPANET Net 3 example application layout with the

proposed sensors location of the base run.

The pH value was calculated from Eq. (6). All species con-

centrations in this equation are a function of the equilibrium

constants and the concentrations of the various systems in

the water.

alk ¼ �OH��þ �

HCO�3

�þ 2$�CO2�

3

�þ �OCl�

�� �Hþ� (6)

The species concentrations and other water quality

parameter values vary throughout the simulations as the

contaminant flows in the network and undergoes both

degradation and formation of the by-products. The

contaminant concentration results are used for the APM

calculations and the contamination fronts are used as

indication for the contaminant reaching a node in the

optimization stage.

3.3. Calculation of affected population

Consumers can be affected either from the ingestion of the

original contamination injected to the WDS or from ingestion

of the byproduct species. Therefore, the expected population

affected from a contamination event in a particular scenario

is a function of the ingested mass of the original contami-

nation or its degradation by products. The following expected

affected population calculation is based on the Z2 formula

presented by Ostfeld et al. (2008), modified in order to ac-

count for a multi-species model. The mass ingested for a

particular species s by any individual at network node i until

time t is:

Ms;i;t ¼ 4DtXN

k¼1

cs;i;krik (7)

where: 4 ¼mean amount of water consumed by an individual

(L/day/person); Dt ¼ evaluation time step (days); N ¼ number

of evaluation time steps up to time t; cs,i,k ¼ contaminant

species s concentration at node i and time step k generated in

the water quality simulations (mg/L); rik ¼ “dose rate multi-

plier” (Murray et al., 2006) for node i and time step k (unitless)

and calculated by:

rik ¼ qik

�qi ck2N (8)

where: qik ¼water demand for time step k and node i; and qi ¼average water demand at node i.

In order to express the probability that any person

ingesting mass Ms,i,t would be affected (i.e., becomes symp-

tomatic) a doseeresponse model (Chick et al., 2001, 2003) was

used:

Ri;t ¼ maxs

�F�b log10

��Ms;i;t

�W

��D50;s

���(9)

where: Ri,t ¼ probability [0,1] that a person who ingests

contaminant mass Ms,i,t up to time t will become symptom-

atic; F ¼ standard normal cumulative distribution function;

b ¼ so-called probity slope parameter (unitless);

W ¼ assumed average body mass (kg/person); and

D50,s ¼ dose that would result in a 0.5 probability of becoming

symptomatic (mg/kg).

The population affected, Pa,i,t, at each node up to time t is:

Pa;i;t ¼ maxt

�Ri;tPi

�(10)

where: Pi ¼ population assigned to node i.

Fig. 3 e (a) Results of 20 consecutive genetic algorithm (GA) runs on Net3 base run; and (b) GA optimal solution progress.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3 197

3.4. Stage II e genetic algorithm (GA) for minimizing theexpected population affected prior to detection

Darwin's natural selection principles, combining an artificial

survival of the fittest with genetic operators abstracted from

nature are the cornerstone of a GA,a heuristic search proce-

dure (Holland, 1975; Goldberg, 1989). GAs have become over

the past decade one of the more successful robust optimiza-

tion techniques employed for water resources and environ-

mental engineering management (Nicklow et al., 2010).

The problem at hand deals with an integer optimization

problem since the sensor locations are the decision variables

of the optimization process. In this study, the optimization

was performed using MATLABTR GA toolbox and its specific

configuration for integer optimization. This process utilizes a

Laplace crossover and power mutation mechanisms intro-

duced by Deep et al. (2009), alongwith a penalty function (Deb,

2000). The GA was run using the following main parameters:

population ¼ 50 or 100 members for 5 sensors and 20 sensors

respectively; crossover fraction ¼ 0.75; max generation

number ¼ 500; and max stall generations ¼ 100.

Sensors were assumed to instantly detect any nonzero

contaminant concentration (Boolean sensors). Therefore, the

contaminant intrusion time of detection for the sensor

network, td, was defined as the minimum time among all

sensors until a contaminant front had reached the jth sensor

td ¼ minj

tj (11)

In the case that the injection was not detected by the

proposed sensor network, td would be equal to the injection

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3198

time plus 48 h, assuming there is human detection of the

event due to the symptoms caused by the contaminant over

this period of time.

Assuming no contaminant mass is ingested post the

detection time, due to actions taken to eliminate further

exposure, the affected population, Pa, for a particular

contamination scenario was calculated as follows:

Pa ¼XV

i¼1

Pa;i;td (12)

where: V ¼ total number of nodes.

The objective function to be minimized is the average

population affected over a given set of contaminated

scenarios:

Minimize

24XSc

j¼1

Pa

35 (13)

where: Sc ¼ total number of contamination scenarios;

Fig. 4 e EPANET Net 3 sensitivity analysis 2 sensors layout.

4. Applications

The above described methodology was applied on three

example applications with different characteristics and

complexity. In the application's base runs (BR) and sensitivity

analyses the following parameters were used: the OP in-

jections duration was 2 h, chlorine injection concentration at

the water sources was 4 mg/l (USEPA, 2009), and for purposes

of calculations the parameters described in the methodology

section were set to: 4 ¼ 2 L/d, b ¼ 0.34 (�), W ¼ 70 kg and

D50 ¼ LD50 as detailed in Table 1. The specific network was

confronted with 500 randomly generated events each as a

base run.

The water quality characteristics being delivered in the

following applications follow typical values of ground and

desalinated water streams in Israel (Schwartz et al., 2014).

4.1. EPANET example application: Net 3 results

EPANET (USEPA, 2013) example application Net 3 was

comprised of 92 nodes, 117 pipes, 3 tanks and 2 constant head

sources with water quality characteristics which correspond

to ground water and desalinated water and operates under a

24 h demand cycle while delivering water to 198,000 con-

sumers assuming demand rate of 300 L/d/capita.

The APM of Net 3 was constructed while calculating the

water quality values at 300 s time steps. In the BR five sensors

were located in the network resulting in average of 695 in-

cidents out of 7712 incidents in the case no sensors were

deployed (9%). The selected sensor locations are a direct

outcome of the randomly selected events, a different set of

events may yield a different optimal layout. The sensors

layout is illustrated in Fig. 2.

Since a GA does not guarantee a globally optimal solution

two analyseswere performed, using the sameAPM, in order to

assess the solution quality. In the first analysis the GAwas run

consecutively resulting in the same optimal solution 8 times

out of 20 runs. The results of the 20 consecutive runs and

example results of the best solution progress during a GA run

can be seen in Fig. 3. In the second sensitivity analysis (SA2)

two of the selected optimal locations were banned (nodes 181

and 213). This resulted in 822 average incidents indicating

again that the BR solution had the characteristics of an

optimal solution. SA2 sensor layout is illustrated in Fig. 4.

The simulation time step affects the constituents' con-centration throughout the model, influencing the expected

incidents. If we presume that small step sizeswill yield amore

accurate simulation, than, the results of a simulation with

larger step sizes that have different results, is considered to be

less accurate. On the other hand, increasing the time step

would reduce the computational burden and shorten the time

for building the APM. An example of the change in parathion,

parathion oxon, and affected population in a random event

due to different time step sizes is demonstrated on Fig. 5. In

order to evaluate the effect of using a larger time step, SA3was

performed with a 900 s time step. The APM was recalculated

with the same events as in the BR resulting in average of 780

incidents out of 7553 incidents and with similar sensor loca-

tions to the BR. The average evaluation time was reduced to

62 s compared to 353 s in the BR.

SA4 was performed in order to explore how deployment of

additional sensors will affect the optimization result. The

optimal layout, based on the calculated APM from the BR, was

calculated for 20 sensors and resulted in 258 incidents, which

is about 37% of the affected population of the BR. As expected,

and as previously observed (Murray et al., 2008) increasing of

the number of sensors will lower the potential damage.

In the final sensitivity analysis (SA5), an additional 500

events were generated for evaluating their influence on the

optimal sensor locations. The new APM constructed of 1000

events had 7507 average incidents and five sensors would

reduce the amount of affected consumers to 658.5. The

optimal sensor layout was similar to the BR except for one

sensor location which had been changed to node 237 instead

of node 213. This result shows that an APM based on 500

events in this network provides sufficient base for the opti-

mization. The results of Net 3, as well as the following appli-

cations results, are summarized in Table 4.

Fig. 5 e Parathion (PA) and parathion oxon PAO concentrations at Net 3 node 253 due to contamination event in node 206 at

36:00 and the number of affected population (APM) due to contaminant consumption in 300 s (a) quality time step and in

900 s (b) quality time step.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3 199

4.2. BWSN1 results

BWSN 1 was comprised of 126 nodes, 168 pipes, 2 tanks, and a

constant head source (ground water characteristics). The

system was subjected to a demand flow pattern of 48 h while

delivering water to about 16,400 consumers. The analysis

performed on this network had the same characteristics as

EPANET Net 3.

In the BR five sensors had reduced the average incidents

from 497 to 46.6 (9.3%). The BR sensors and the network layout

can be seen in Fig. 6.

BWSN1-SA3 was performed as Net3-SA3 and resulted,

when comparing to the BR results, in a similar average

possible incidents of 499 and 47.5 incidents (9.5%) after per-

forming the optimization. The layout, in this case, was similar

to the BR except for one change in location to the adjacent

node in one of the sensors. These results and the results of

SA3 in Net 3 show that when a sufficient amount of events are

performed the maximal average incidents gives similar re-

sults in both time step sizes, in spite of the differences in

concentration values, as shown in Fig. 5. These results yield

similar optimal sensor locations but differences in detection

time which result from the change in the time step sizes that

might influence the result after the optimization process. The

deployment of 20 sensors in SA4 lowered the average in-

cidents to 16.3, 37% in comparison to the BR run.

Table

4e

Sum

mary

ofexam

ple

applica

tionsresu

lts.

Example

applica

tione

analysisnam

eAnalysis'sm

ain

characteristics

Senso

rloca

tions

Average

inciden

tsMaxim

alaverage

inciden

ts(nose

nso

rs)

Consu

mers

Net3

-BR

5se

nso

rs111,119,181,201,213

695

7712

198,187

Net3

eSA2

5se

nso

rs,Nodes181&

213

are

notperm

itted

119,173,177,203,211

822.7

7712

198,187

Net3e

SA3

5se

nso

rs,900stimestep

111,119,181,201,213

780.6

7553

198,187

Net3

eSA4

20se

nso

rs35,61,103,111,117,120,123,143,149,169,

181,184,201,203,203,206,213,231,247,263

258

7712

198,187

Net3

eSA5

1000events,5se

nso

rs111,119,181,201,237

658.5

7507

198,187

BW

SN1e

BR

5se

nso

rs17,23,68,90,103

46.6

497

16,400

BW

SN1e

SA3

5se

nso

rs,900stimestep

17,24,68,90,103

47.5

499

16,400

BW

SN1e

SA4

20se

nso

rs17,23,27,34,41,46,48,58,65,69,77,82,

94,97,102,103,110,115,118,126

16.6

497

16,400

BW

SN1e

SA6

5se

nso

rs17,23,68,90,103

42.0

464

16,400

Dover-BR

5se

nso

rs2039,5680,6522,7197,9185

173

652

93,730

Dovere

SA4

20se

nso

rs1909,2030,2331,4779,5680,5882,6072,

6397,7206,7372,7839,8332,9031,9185,

9910,10178,11056,11473,11891,14407

97.7

652

93,730

Dovere

SA5

1000events,5se

nso

rs1910,5693,6078,7148,9185

184.6

622

93,730

Acronyms:

BR¼

Base

run;SA

¼Sensitivityanalysis.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3200

In order to evaluate the effect of lower chlorine level at the

water source, BWSN1-SA6 was conducted with free chlorine

level of 1.5 mg/l at the water source, andwith the same events

series as the BR. As a result from the lower residual chlorine

less OP's had reacted to form OPO. Since the concentration of

the toxic OPO was lower in this case the number of incidents

was decreased to 42.0 in comparison to the BR (464 incidents

without sensors). The sensor layout in this case remained the

same as in the BR. As can be seen from Fig. 7, SA6 layout did

not change since the relative number of incidents spread

along the network remained similar to the BR, resulting the

same optimal layout.

4.3. Dover results

The Dover network consisted of 6165 nodes, 6584 pipes, 5

tanks, 14 water sources (as negative consumers, with

ground water characteristics) driven by a 24 h demand cycle

while delivering water to 93,700 consumers. The Dover

network is less aggregated in comparison to BWSN1 and Net

3, which is shown by comparing the average consumers per

node values of 15, 130, and 2154 consumes per node for

Dover, BWSN 1 and Net 3, respectively. The complexity of

the Dover network requires high computational time,

therefore the APM calculation was performed with a time

step size of 900 s.

The BR resulted in average incidents of 173, about 25% of

the average incidents in the case that was run without

sensors. The low expected incidents relative to the overall

number of consumers demonstrate the difference between

a detailed and an aggregated model. In a detailed model a

contamination introduced in a relative high number of

nodes can reach only a limited amount of consumers. In

addition, the large number of nodes affects the sensors

ability to lower the potential number of incidents when

compared to aggregated models such as Net3 and BWSN1.

This is due to the spanning of the 500 events and the

complicated network topology which enables contaminant

injections in a larger amount of nodes downstream from the

sensors. The BR sensors and the network layout are shown

on Fig. 8.

The sensors lower ability to reduce the number of potential

incidents was observed again in SA4 where the deployment of

20 sensors reduced the average number of incidents by 56% in

comparison to the BR.

As in Net3, an additional 500 events were simulated in SA5

resulting in 184.6 incidents on average. In this case, the

spatial diversity of this network led to a different set of

optimal locations. Nevertheless, the selected regions of the

sensors are similar to the selected regions in the BR. This is

derived from the hydraulics of the network and the affect of

the original APM. This raises a question on what are satis-

factory base data for supporting the location decision in a

real scale network.

5. Conclusions

A methodology for optimal sensor location, which in-

corporates detailed chemistry of possible contaminants,

Fig. 6 e BWSN 1 example application layout with the proposed sensors location of the base run.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3 201

was demonstrated. This work modeled three organophos-

phates as possible contaminants and their affect on con-

sumers as a base for deciding where to place water quality

sensors.

In the case at hand, the potential hazard from organo-

phosphates contamination could not be fully assessed

without calculation of the organophosphates degradation

products, which are extremely more toxic. The usage of only

the original contaminant toxicity will result in permissive

incident values. In other potential contaminations, which are

not dealt with in this work, a non-detailed quality model

might lead to conservative results. Therefore, the incorpora-

tion of detailed chemistry to themodel improves the ability to

Fig. 7 e Average number of incidents relative to the total numb

(BR) and for sensitivity analysis 6 (SA6).

assess the potential damages from a contamination event and

can assist in the process of early detection system design.

The selected optimal layout is a direct function of the

randomly generated events. It was shown that the number of

simulated events affect on the optimal solution depends on

the network size. In a detailed network the widespread nature

of the events causes the sensors to exclude less incidents

since less significant events are created in comparison to

aggregated networks, as can be seen in the relation between

maximal average incidents and the number of consumers,

and since it is harder to detect the widely spread events. Due

to the computational effort which comes along with detailed

networks and multi-species models, incorporation of event

er of incidents in BWSN1 nodes illustrated for the base run

Fig. 8 e Dover example application layout with the proposed sensor location of the base run.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3202

selection methods could be viewed as complementary

research.

The results also show that enlarging the simulation time

step will reduce the computational burden, and can result in

similar sensor locations, while on the other hand, decrease

the accuracy of the simulation.

In this study the sensors were assumed to be ideal and

Boolean. Water quality models, such as the one implemented

in this work, can be utilized to better understand the behavior

of water parameters in the case of organophosphates

contamination events and can be used for event detection as a

preliminary stage for solving the problem of non-Boolean

sensors optimal placement.

Acknowledgments

This study was supported by the Technion Grand Water

Research Institute, by the Water Authority, by the Technion

Funds for Security research, by the joint Israeli Office of the

Chief Scientist (OCS) Ministry of Industry, Trade and Labor

(MOITAL), and by the Germany Federal Ministry of Education

and Research (BMBF), under project number GR 2443.

r e f e r e n c e s

Alfonso, L., Jonoski, A., Solomatine, D., 2010. Multiobjectiveoptimization of operational responses for contaminantflushing in water distribution networks. J. Water Resour. Plan.Manag. 136 (1), 48e58.

Chick, S.E., Koopman, J.S., Soorapanth, S., Brown, M.E., 2001.Infection transmission system models for microbial riskassessment. Sci. Total Environ. 274 (1), 197e207.

Chick, S.E., Soorapanth, S., Koopman, J.S., 2003. Inferringinfection transmission parameters that influence watertreatment decisions. Manag. Sci. 49 (7), 920e935.

Davis, M.J., Janke, R., 2011. Patterns in potential impactsassociated with contamination events in water distributionsystems. J. Water Resour. Plan. Manag. 137 (1), 1e9.

Deb, K., 2000. An efficient constraint handling method for geneticalgorithms. Comput. Methods Appl. Mech. Eng. 186 (2e4),311e338.

Deep, K., Singh, K.P., Kansal, M.L., Mohan, C., 2009. A real codedgenetic algorithm for solving integer and mixed integeroptimization problems. Appl. Math. Comput. 212 (2), 505e518.

Duirk, S.E., Desetto, L.M., Davis, G.M., 2009. Transformation oforgamophosphorus pesticides in the presence of aqueouschlorine: kinetics, pathways and structure e activityrelationships. Environ. Sci. Technol. 43, 2335e2340.

Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization,and Machine Learning. Addison-Wesley, New York, p. 412.

Hart, W.E., Murray, R., 2010. Review of sensor placementstrategies for contamination warning systems in drinkingwater distribution systems. J. Water Resour. Plan. Manag. 136(6), 611e619.

Holland, J.H., 1975. Adaptation in Natural and Artificial Systems.The University of Michigan Press, Ann Arbor.

Isayev, O., Rasulev, B., Gorb, L., Leszczynski, J., 2006. Structure-toxicity relationships of nitroaromatic compounds. Mol.Divers. 10 (2), 233e245.

Kumar, J., Brill, E.D., Mahinthakumar, G., Ranjithan, S.R., 2012.Contaminant source characterization in water distributionsystems using binary signals. J. Hydroinformatics 14 (3),585e602.

Lewis, R.J.S.R., 1996. Sax's Dangerous Properties of IndustrialMaterials, sixth ed. Wiley, Hoboken, NJ.

wat e r r e s e a r c h 7 3 ( 2 0 1 5 ) 1 9 3e2 0 3 203

Lockridge, O., Duysen, E.G., Voelker, T., Thompson, C.M.,Schopfer, L.M., 2005. Life without acetylcholinesterase: theimplications of cholinesterase inhibitor toxicity in AChE-knockout mice. Environ. Toxicol. Pharmacol. 19 (3), 463e469.

Murray, R., Uber, J., Janke, R., 2006. Model for estimating acutehealth impacts from consumption of contaminated drinkingwater. J. Water Resour. Plan. Manag. 132 (4), 293e299.

Murray, R., Baranowski, T., Hart, W.E., Janke, R., 2008. Riskreduction and sensor network design. In: Proc., WaterDistribution Systems Analysis 2008, ASCE, Reston, Va.

Murray, R., Haxton, T., McKenna, S.A., Hart, D.B., Klise, K.A.,Koch, M., Vugrin, E.D., Cutler, L., 2010. Water Quality EventDetection Systems for Drinking Water ContaminationWarning Systems: Development Testing and Application ofCANARY. U.S. Environmental Protections Agency, Office ofResearch and Development, National Homeland SecurityResearch Center, Cincinnati, Ohio, USA.

Nicklow, J., Reed, P., Savic, D., Dessalegne, T., Harrell, L., Chan-Hilton, A., Karamouz, M., Minsker, B., Ostfeld, A., Singh, A.,Zechman, E., 2010. State of the art for genetic algorithms andbeyond in water resources planning and management. J.Water Resour. Plan. Manag. 136 (No. 4), 412e432.

Ohar, Z., Ostfeld, A., 2014. Optimal design and operation ofbooster chlorination stations layout in water distributionsystems. Water Res. 58, 209e220.

Oliker, N., Ostfeld, A., 2014. Minimum volume ellipsoidclassification model for contamination event detection inwater distribution systems. Environ. Model. Softw. 57, 1e12.

Ostfeld, A., Salomons, E., 2004. Optimal layout of early warningdetection stations for water distribution systems security. J.Water Resour. Plan. Manag. Div. ASCE 130 (No. 5), 377e385.

(þ 34 co-authors) Ostfeld, A., et al., 2008. The battle of the watersensor networks: a design challenge for engineers andalgorithms. J. Water Resour. Plan. Manag. Div. ASCE 134 (6),556e568.

Perelman, L., Ostfeld, A., 2010. Extreme impact contaminationevents sampling for water distribution systems security. J.Water Resour. Plan. Manag. 136 (1), 80e87.

Perelman, L., Ostfeld, A., 2012. Extreme impact contaminationevents sampling for real-sized water distribution systems. J.Water Resour. Plan. Manag. 138 (5), 581e585.

Preis, A., Ostfeld, A., 2008a. Multiobjective contaminant responsemodeling for water distribution systems security. J.Hydroinformatics 10 (4), 267e274.

Preis, A., Ostfeld, A., 2008b. Genetic algorithm for contaminantsource characterization using imperfect sensors. Civ. Eng.Environ. Syst. 25 (1), 29e39.

Rasekh, A., Brumbelow, K., 2013. Probabilistic analysis andoptimization to characterize critical water distribution systemcontamination scenarios. J. Water Resour. Plan. Manag. 139(2), 191e199.

Rasekh, A., Brumbelow, K., 2014. Drinking water distributionsystems contamination management to reduce public healthimpacts and system service interruptions. Environ. Model.Softw. 51, 12e25.

Schwartz, R., Lahav, O., Ostfeld, A., 2014. Integrated hydraulic andorganophosphate pesticide injection simulations forenhancing event detection in water distribution systems.Water Res. 63, 271e284.

Shang, F., Uber, J.G., Rossman, L.A., 2008. “EPANET Multi-speciesExtension User's Manual.” EPA/600/S-07/021. U.S.Environmental Protection Agency, Cincinnati OH availableonline at: http://www-ext.lnec.pt/projects2013/saa/pdf/Manual_EPANET_MSX.pdf (accessed 05.10.14).

Tryby, M.E., Propato, M., Ranjithan, S.R., 2010. Monitoring designfor source identification in water distribution systems. J.Water Resour. Plan. Manag. 136 (6), 637e646.

Tsang, J.S.W., Neverov, A.A., Brown, R.S., 2004. La3þ catalyzedmethanolysis of O,O-diethyl S-(p-nitrophenyl)phosphorothioate and O,O-diethyl S-phenyl phosphorothioatemillions-fold acceleration of the destruction of V-agentsimulants. Org. Biomol. Chem. 2 (23), 3457e3463.

USEPA, 2009. National Primary Drinking Water Regulations.United States Environmental Protection Agency, EPA, 816-F-09e004.

USEPA, 2013. EPANET: USEPA Hydraulic and Water Quality WaterDistribution Piping Systems Program. Available online at:http://www.epa.gov/nrmrl/wswrd/dw/epanet.html (accessed25.03.14).

WHO, 2009. The WHO Recommended Classification of Pesticidesby Hazard and Guidelines to Classification. World healthorganization, Geneva.

Wu, J., Laird, D.A., 2003. Abiotic transformation of chlorpyrifos tochlorpyrifos oxon in chlorinated water. Environ. Toxicol.Chem. 22 (2), 261e264.