Water Resour Manage (2011) 25:1219–1228

DOI 10.1007/s11269-010-9674-0

A Security-Oriented Manual Quality Sampling

Methodology for Water Systems

Demetrios G. Eliades · Marios M. Polycarpou ·

Bambos Charalambous

Received: 16 November 2009 / Accepted: 20 May 2010 /

Published online: 10 June 2010
© Springer Science+Business Media B.V. 2010

Abstract The management and security of water resources will be a key challenge in
the years ahead. This work investigates the water quality manual sampling scheduling
problem in drinking water distribution networks. This work contributes to research
by presenting a problem formulation based on the underlying substance propagation
dynamics, coupled with the impact dynamics describing the “damage” caused by
a contamination in a drinking water distribution network. The proposed solution
methodology optimizes a risk-objective, in order to compute a manual sampling
scheme comprised of sampling nodes and times, using evolutionary computation
techniques. To illustrate the methodology, we present simulation results on a real
water distribution network.

Keywords Water distribution networks · Manual sampling · Fault diagnosis ·

Security

This work is partially supported by the Research Promotion Foundation (Cyprus) and the
University of Cyprus.
D. G. Eliades (B) · M. M. Polycarpou
KIOS Research Center for Intelligent Systems and Networks, ECE Department,
University of Cyprus, Nicosia, Cyprus
e-mail: eldemet@ucy.ac.cy

B. Charalambous
Water Board of Lemesos, 3602 Lemesos, Cyprus
1220 D.G. Eliades et al.

1 Introduction

Water utilities are responsible for the delivery of drinking water to consumers and
for the regular water quality monitoring at various locations in the network, so as
to assure that the water quality is within certain prespecified bounds. In addition,
water quality monitoring is important for detecting contamination problems which
may occur within the water distribution network.
According to the World Health Organization, a free chlorine residual concentra-
tion (i.e. the amount of chlorine which was dissolved in water in the form of chlorine
gas and which is available to react and inactivate a number of organic substances)
must exist in drinking water distribution systems, with minimum target concentra-
tion 0.2 mg/L at point of delivery (WHO 2008). The chlorine concentrations in
a distribution network can be monitored using online sensors; however, these are
usually expensive equipment and so far have not been widely adopted, especially by
small water utilities. When these chlorine concentration sensors are not available,
the standard approach is for the water utility to perform manual water sampling for
quality analysis. Water quality may be checked at a few nodes within a network, and
for a few times during the day. Sampling locations may be selected by water utility
personnel in an arbitrary fashion based on his/her experience, which could be subject
to human bias, or may be chosen using certain regulatory requirements, depending
on consumer distributions and historical data. In practice, due to the large-scale
nature of water distribution networks as well as the partial knowledge of the time-
varying, hydraulic and quality dynamics, it is difficult to optimally identify the best
locations and times to conduct manual sampling, or install on-line quality sensors. In
addition, each node in the network has certain characteristics, such as the outflow
water pattern and the number of customers, which makes the selection problem non-
trivial.
Contamination faults in a drinking water distribution network can occur any-
where, due to natural or accidental events, or even malicious attacks. The impact, in
terms of losses or damages, caused due to such a contamination depends on the time
and place the fault occurs. For example, a contamination occurring at an end-pipe
in the network cannot propagate to consumers other than those served by that pipe;
on the other hand, a contamination near the water supply source may propagate to
the majority of consumers served. Due to the time-varying nature of demands, water
flows in pipes fluctuate during each day, and that influences the damage rate of a
contaminant substance in the water distribution system.
Various methodologies have been proposed for identifying a set of locations to
install quality concentration sensors which can detect contaminations. The primary
assumption is that quality concentration sensors are permanently installed at certain
locations in the network and can detect contaminants above a certain concentration
threshold. The general location problem has been examined in various research
disciplines. One of the first mathematical formulations is the “Set Covering” com-
binatorial problem, in which the underlying system is expressed as a graph, where
nodes represent locations and arcs represent functional or geographical relations
between location nodes (Christofides and Paixão 1993). A similar problem is the
“Maximal Covering”, which, instead of full coverage, it searchers for solutions which
maximize a certain serving distance or time (Church and ReVelle 1974). Population
weights are assigned to each node in the graph while distance weights are assigned to
A Security-Oriented Manual Quality Sampling Methodology for Water Systems 1221

arcs. The optimization program maximizes the number of people served, constrained
by a fixed number of facilities and a limiting serving distance. The problem of
selecting locations for water sampling in drinking water distribution networks was
first formulated as a problem of “Maximal Covering” (Lee and Deininger 1992).
The objective was to find the locations from where to measure water quality, so
that the maximum of water demand is checked. Using the hydraulic data, the
authors computed the “contribution” of water from one node to another. In addition
to quality monitoring in water distribution networks, similar problems have been
examined in other water resources research, such as the problem of valve placement
in water distribution networks (Liberatore and Sechi 2009), and the groundwater
quality monitoring problem (Mogheir et al. 2009).
Other mathematical formulations have been developed for solving the water
distribution network sensor placement problem by using various impact metrics, such
as the time of detection, population affected, or amount of consumed water (Berry
et al. 2006a; Hart et al. 2007). Most research on the optimization problem of sensor
placement and manual sampling, formulates it as an integer (or binary) optimization
program with constraints. This problem is known to be NP-hard, i.e. there exists no
algorithm which may solve this problem in polynomial time. Therefore, the larger the
network, the more difficult it is to optimally compute a solution to these problems.
Based on the above work, Boccelli and Hart (2007) have examined the quality
monitoring problem by extending the formulation for computing sensor locations
which maximize the coverage, constrained by a desired water age distribution. Within
the water management community, significant research interest on security issues
in drinking water distribution systems was instigated after the design competition
entitled “Battle of the Water Sensor Network” (Ostfeld et al. 2008), where the
objective was to select locations for installing quality sensors in two real water
distribution networks, in order to minimize a number of objectives.
The problem of manual sampling was discussed in Berry et al. (2006b), where
the authors examined the problem of scheduling manual sampling for contaminant
detection. They proposed a mixed integer program for computing the sampling route;
i.e. the location and time to take samples, while considering certain real conditions
such as utility working hours, the time required for sampling and the traveling
time between nodes. Computational difficulties in solving large-scale problems by
using this formulation have been expressed in Berry et al. (2006a, b) for the sensor
placement and manual sampling scheduling problem respectively.
A comparison of the use of online sensors with the use of manual sampling has
been examined in a previous study (Eliades et al. 2009). It was shown that on-
line sensors offer increased security, when compared with manual sampling. On
the other hand, manual sampling offers more flexibility for selecting the node to
be monitored in real time. Manual sampling can be considered as a special case
of the sensor placement problem, with time constraints and instantaneous (instead
of continuous) monitoring. With respect to the solution methodology, the manual
sampling scheduling problem is different from the quality-sensor placement problem
in that the computed solution is not a set of nodes, but a set of nodes coupled with
time, i.e. when and where to perform sampling, under certain constraints.
In the present work we concentrate on the mathematical formulation of the
manual sampling scheduling problem, when considering certain risk-objectives. In
Section 2, we present the problem formulation and in Section 3 the solution
1222 D.G. Eliades et al.

methodology; in Section 4 we demonstrate the methodology with a real example.

Finally, in Section 5, we present some final conclusions and future work.

2 Problem Formulation

Consider a drinking water distribution network, comprised of pipes, junctions and

water storages. The topology of this network can be represented as a graph, with
edges corresponding to pipes and m nodes corresponding to junctions and water
storages. For modelling purposes, each pipe is assumed to be segmented into a
number of finite volume cells, by using a suitable algorithm; in total m0 finite volume
cells are considered. Let n be the number of all junctions, water storages and finite
volume cells in all pipes, such that n = m + m0 . The set of all network indices,
corresponding to each cell, is defined as V = {v1 , ..., vn }; in addition we define and
the set of all node indices V0 , comprised of vi ∈ V, i = 1, ..., m.
Let xi (k) denote the contaminant concentration at discrete time k within finite
volume cell vi ; the advection and reaction dynamics describing the propagation of
the substance in the water distribution network, can be expressed in a state-space
formulation

x(k + 1) = A(k)x(k) + g(x(k)) + φ(k; θ), (1)

where the vector x(k) = [x1 (k), ..., xn (k)] is the state of chlorine dynamics; A(k)
is an n × n matrix which characterizes the advection dynamics and g(x(k)) is a
function which describes the reaction dynamics of the contaminant. Function φ(k; θ)
describes the change in the contaminant concentration at the contamination source;
the location of the contamination and the characteristics of the input signal depend
on the fault parameter θ .
When a contamination fault occurs, the contaminant is propagated in the network,
and eventually it may reach consumers who will consume the outflow water. The
purpose of a security-oriented manual sampling should be to reduce the risk of
“damage” caused by the contaminant consumption; therefore, it is necessary to
consider metrics which measure the impact of a contamination fault on the system.
Let W ⊂ V0 be the set of Nd node indices corresponding to locations which outflow
water to consumers, based on demand requests. For each “demand” node wi ∈
W, i = 1, ..., Nd , an impact value ξi (k) can be computed at each time step k. The
impact of a fault depends on the volume and contaminant concentration of the
contaminated water consumed.
In the present work, we define ξi (k) to be the volume of contaminated water
consumed at demand node wi ∈ W, until time k. The state-space representation of
the impact dynamic is given by

ξi (k + 1) = ξi (k) + f (xwi (k), dwi (k)) (2)

where dwi (k) is the outflow demand, measured in m3 /s, at demand node wi . Function
f (·) is a non-negative function which computes the impact increase, i.e. the volume
of polluted water consumed, at each time step, and is given by

dwi (k)t if xwi (k) > η
f (xwi (k), dwi (k)) = (3)
0 if xwi (k) ≤ η
A Security-Oriented Manual Quality Sampling Methodology for Water Systems 1223

where η ≥ 0 is the concentration threshold above which the consumed volume is

assumed to be polluted and t is the length of each time-step. Furthermore, we
define ψ(k) as the output vector, given by ψ(k) = f (ξ(k)) where the function f (·)
characterizes the total impact, or “damage”, which has caused due to a contamination
fault. For the contaminated water consumption volume considered in the present
work, the total impact is given by the aggregated volume of consumed contaminated
water, i.e. f (ξ(k)) = 1 ξ(k), where 1 is a vector of “ones” of appropriate dimension.

3 Methodology

Given a network topology with certain hydraulic dynamics which are subject to un-
certainties, let S ⊆ V0 be the set of Ns candidate sampling nodes, i.e. locations where
water sampling is feasible, such that S = {s1 , ..., s Ns }. From a practical viewpoint, we
consider that sampling can be performed at certain times within a day (e.g. due to
working hours); we define T as the finite set of time steps at which sampling can
be conducted within a period with discrete length T H , i.e. T = {τ1 , ..., τ Nτ }, τi ≤ T H .
Hydraulic dynamics are considered approximately periodic, with a 24-h period, due
to the periodic nature of consumer demands. The set Q of all feasible sampling
locations and times, is given by Q = {(i, j)|i ∈ S, j ∈ T}, such that Q = {q1 , ..., q Nq }
and Nq = Ns · Nτ .
The manual sampling scheduling problem can be defined as finding the set Y ⊂ Q
of M sampling locations and times (where |Y| = M < Nq ), such that one or more
security-oriented risk objectives are optimized.
An approach to solving the sampling problem is to consider a number of
“representative” quality faults which can occur within a water distribution system.
Assuming that the hydraulics are periodic with period of one day, we can construct
a finite set of fault parameters, = {θ1 , ..., θ Nθ }, which correspond to Nθ fault
functions φ(k; θ), i = 1, ..., Nθ . For simplicity, the elements in can be obtained
through grid sampling, for faults that have start time within one period. Let ts (θi )
be a function which computes the start time of the fault characterized by parameter
θi . From a practical perspective, the detection time of a fault is bounded by a finite
constant, because even if it is not detected during sampling, it will eventually be
detected after some time instance tmax by other methods such as customer complaints
or increased hospitalizations.
Let x(k; θi ) and ψ(k; θi ) correspond to the state x(k) and output impact vector
ψ(k) when a fault with parameter θi occurs. We define the overall-impact matrix
of size Nθ × Nq , such that its (i, j) element, (i, j) , corresponds to the impact of fault
with parameter θi , when manual sampling q j ∈ Q, q j = (μ j, λ j) is conducted, i.e. at
the node μ j and time λ j. In specific, (i, j) is given by
⎧
⎪
⎨ψ(λ j; θi ) if ts (θi ) ≤ λ j and xμ j (λ j; θi ) ≥ 
(i, j) = ψ(λ j + T H ; θi ) if ts (θi ) > λ j and xμ j (λ j + T H ; θi ) ≥  (4)
⎪
⎩
ψ(ts (θi ) + tmax ; θi ) otherwise
where  is non-negative concentration threshold above which detection is possible
and T H is discrete length of one period considered.
In our problem formulation, we consider minimizing the average and maximum
impact, with respect to the various contamination faults in . Assume that the
1224 D.G. Eliades et al.

task is to select M sampling locations and times, out of Nq candidates, and let
L = {1, ..., Nq } be the set of sampling indices. We define X as the set of all M-
 
combinations from the set L, of size |X| = NMq , such that X = {X | X ∈ L M , Xi =
X j ⇔ i = j}.
The general single-objective optimization problem is formulated as
Y = argmin { f (X; ) | |X| = M} (5)
X∈X

where M is the number of samples taken within one period, and is the overall-
impact matrix. The function f (X; ) is the objective and its formulation depends on
the risk-objectives considered; in the present work, the average and the maximum
impact are considered.
The average impact metric is suitable for optimizing reliability, when a contam-
ination event can occur at any node with equal probability. For the overall-impact
matrix whose (i, j) element is (i, j) , the average impact objective for solution X is
given by
Nθ
1
f (X; ) = min (i, j) . (6)
Nθ i=1
j∈X

The maximum impact is used to reduce the effect of the most extreme contamina-
tion fault, in term of causing the most damage and is given by
f (X; ) = max min (i, j) (7)
i=1,...,Nθ j∈X

4 Application

In this section we illustrate the formulation and the solution methodology for the
manual sampling scheduling problem in a real drinking water distribution network,
depicted in Fig. 1, operating under realistic conditions. The network is comprised
of 321 pipes connected to m = 198 junctions, 100 of which are used for water
consumption. The main source of disinfected water is a storage tank which is located
in the lower part of the figure. The structural characteristics are assumed to be
known, i.e. pipe length, diameters and pipe roughness coefficients, node elevations
and daily average consumption volume at each node. In addition, historical flow-data
are provided, measured at the supply node, and are assumed to describe the normal
operation over all nodes. We consider the set of nodes V0 = {v1 , ..., v198 } and a t = 5
minute time step for computing the hydraulic and quality dynamics; each 24-h period
is therefore comprised of T H = 288 time steps.
There are Nd = 100 nodes in the network which outflow water to consumers; the
set of demand nodes is W = {w1 , ..., w100 }, and W ⊂ V0 . We further assume that
the concentration threshold η above which the consumed volume is assumed to be
polluted, is η = 0. In this example, we consider that the set of candidate sampling
locations is the same as the demand nodes, i.e. S ≡ W. Regarding the finite set
T comprised of time steps at which sampling can be conducted, we assume that
sampling can be performed every 30 mins, thus there are maximum Nτ = 48 time
instances for manual sampling within one day. From these, the set Q of all feasible
sampling locations and times has size Nq = 4,800. Therefore, a contamination fault
A Security-Oriented Manual Quality Sampling Methodology for Water Systems 1225

Fig. 1 Real water distribution

system with 321 pipes and 198
junctions

may occur anywhere and at anytime in the distribution network; this is a common
assumption when considering malicious contamination attacks.
For the construction of the fault parameter set , we consider faults which start
the contamination at a certain location in the network, and continue to contaminate
the water at that location until detection. In specific, a fault with parameter θ ∈ is
modeled as a step-function of 1 mg/L, which starts at a certain time within a certain
period 1 ≤ ts (θ) ≤ 288 and is terminated when the contamination is detected through
sampling, or after a certain number of time-steps has passed, in specific tmax = 288.
We construct the set of fault parameters characterizing faults which occur at one
demand node every half hour within one day; 48 faults are therefore considered for
each node; has size Nθ = 4,800.
For constructing the overall-impact matrix , we assume that the time-varying
matrix A(k) is known for all times, that no reactions occur in the network g(x(k)) = 0
and that the outflow dw (k) for all demand nodes w ∈ W are known. We repetitively
simulate the operation of the system corresponding to duration of 48 h, applying one
fault θi at a time, for i = 1, ..., 4,800. Within one simulation, we measure the impacts
based on the various sampling schemes q j, for j = 1, ...4,800. Finally, the overall-
impact matrix of size 4,800 × 4,800 is constructed.
We examine the single-objective optimization problem, for computing a solution
with M = {1, 2, 3} sampling events within one period, i.e. a day, for minimizing
the average (and maximum) volume of polluted water consumed. The optimization
problem described by (5)–(7) is non-linear, and for solving it we utilize an integer
1226 D.G. Eliades et al.

evolutionary optimization algorithm. Genetic algorithms have proven suitable for

these types of optimization problems with constraints, in computing optimal or near-
optimal solutions.
In each problem examined, the evolutionary algorithm iterated for 50 epochs,
and for each epoch a population of 10,000 individual solutions was build. The
following scenarios were considered: to optimize the average/maximum consumption
of polluted water for various numbers of samples, under various sampling periods
constraints (e.g. during work hours).
The results are summarized in Table 1. Each element in the solution corresponds
to a node and a time when sampling should be performed. For instance, for the
first row, if only one sampling is to be conducted within a day, by sampling at node
“74” at 9 a.m., the average consumption of polluted water consumed is minimized;
furthermore, for this solution, the objective function, i.e. the average consumption
of polluted water is 7.20% (with respect to the maximum impact of the worst
contamination fault). When an all-day sampling period is considered, some solutions
may correspond to non-working hours; this however can be adjusted by imposing
constraints to the optimization algorithm. As expected, the constrained solutions
are equal or worse to the all-day sampling solutions. For the network examined,
it appears that node “74” is quite important from a average impact objective. By
minimizing the maximum polluted water consumption, however, the solutions may
be different; for example, for two-samplings per day, examining node “58” at 9 a.m.
and node “68” at 6 p.m. can reduce the maximum impact by 20.85% (with respect to
the maximum impact of the worst contamination fault).
The three histograms in Fig. 2 illustrate the effectiveness of the proposed algo-
rithm of selecting manual sampling locations and times, in reducing the volume of
polluted water consumed for each contamination fault-scenario, over the case when
no manual sampling is conducted. The first histogram is computed by simulating
each contamination fault scenario from the set of all contamination fault scenarios
, under a nominal consumption model, and by measuring the volume of polluted
water consumed if no detection ever occurs.
The second histogram depicts the fault-scenario distribution with respect to their
maximum impact when an optimal solution is used from Table 1; in specific, the
solution methodology was used to compute two samples per day (at node ‘74’), within
working hours (at 8 a.m. and 4:30 p.m.) and by minimizing the average impact i.e.
the volume of polluted water consumed. It is assumed that when a contamination is

Table 1 Solutions for various sampling schemes and objectives

Samples Objective Period Solution Objective
function (%)
1 Mean All day (74, 9 a.m.) 7.2
2 Mean All day (74, 6 a.m.), (74, 2:30 p.m.) 4
3 Mean All day (74, 1 a.m.), (74, 10 a.m.), (74, 4:30 p.m.) 2.78
2 Mean 8 a.m.–5 p.m. (74, 8 a.m.), (74, 4:30 p.m.) 4.13
3 Mean 8 a.m.–5 p.m. (74, 8 a.m.), (77, 12 a.m.), (74, 5 p.m.) 3.21
1 Max All day (58, 9:30 a.m.) 41.91
2 Max All day (58, 9 a.m.), (68, 6 p.m.) 20.85
3 Max All day (76, 5 a.m.), (72, 11:30 a.m.), (78, 6 p.m.) 15.78
A Security-Oriented Manual Quality Sampling Methodology for Water Systems 1227

Fig. 2 Histogram of the normalized impact for all the faults considered

detected, the system stops and no damage is further caused. What we can observe by
comparing the first two histograms is that, for the 4,800 fault-scenarios considered,
the impact, i.e. the volume of contaminated water consumed, is significantly reduced.
In fact, the worst-case fault scenarios in the first histogram can cause less than 40%
of their maximum impact if manual sampling at node ‘74’ at times 8 a.m. and 4:30 p.m.
is conducted.
To illustrate the effectiveness of the proposed methodology over other sampling
schemes, the third histogram depicts the frequency distribution of manual sampling
at the node ‘166’, which has been arbitrary selected based on the fact that it is located
in a central location in the distribution location, as seen in Fig. 1. Sampling times
have been arbitrary selected to be the same as in the computed solution described
in the second histogram. By comparing the second and the third histogram, i.e. the
optimally computed and the arbitrary selected solution, we confirm that although
both manage to reduce the impact for all fault-scenarios, the solution computed by
using the proposed methodology is significantly better than the arbitrary one.
In general, the larger the number of nodes in the network and the larger the
number of fault-scenarios considered, the larger the space and time requirements
in computing the overall-impact matrix. By applying expert knowledge, however,
it may be possible to sufficiently reduce the size of this matrix, in accordance
to the available computational space. The genetic algorithm used for solving the
proposed optimization problem does not require high computational power (the
parent selection algorithm utilized in our work is known to have O(n) complexity).
Therefore, it is possible to run the optimization algorithm on standard computers.

5 Conclusions and Future Work

Quality sampling is important in drinking water distribution networks, and is receiv-

ing increasingly importance, along with security issues arising due to the fact that
they are considered as part of the “critical infrastructure”. In this paper we addressed
certain issues regarding the selection of locations and times for manual sampling, in
order to examine the quality of water, i.e. whether sufficient quantity of chlorine
residual exists, or whether the concentration of a substance exceeds the pre-specified
1228 D.G. Eliades et al.

limits. The selection of location and time for manual sampling in this work is driven
by security-oriented criteria, for minimizing the possible damage caused due to a
contamination event, which may occur accidentally or even intentionally.
The mathematical formulation describing the dynamics involved has been pre-
sented, and a solution methodology has been proposed for solving the problem. The
problem is formulated into a single objective optimization problem, utilizing certain
risk-objectives such as the average and maximum impact. A simulation example
has been presented on a real drinking water distribution network. The optimization
problem was solved using a single-objective evolutionary algorithm.
In this work, no prior assumptions are made in relation to the probability of a
contamination fault occurring. In fact, the contamination fault scenario construction
makes no assumptions on the likelihood of a contamination occurring at some nodes;
this is true in the case when considering malicious contamination attacks, which may
occur anywhere and at anytime in the network. In future work we will examine the
use of probabilistic models in the fault-scenario construction.

References

Berry J, Hart W, Phillips C, Uber J, Watson J (2006a) Sensor placement in municipal water networks
with temporal integer programming models. ASCE J Water Resour Plan Manage 132(4):218–224
Berry J, Lin H, Lauer E, Phillips C (2006b) Scheduling manual sampling for contamination detection
in municipal water networks. In: ASCE proc. water distribution systems analysis symposium, vol
247, no 40941, pp 1–27
Boccelli DL, Hart WE (2007) Optimal monitoring location selection for water quality issues. In:
ASCE proc. of world environmental and water resources congress, vol 243, no 40927, pp 522–
527
Christofides N, Paixão J (1993) Algorithms for large scale set covering problems. Ann Oper Res
43(5):259–277
Church R, ReVelle C (1974) The maximal covering location problem. Pap Reg Sci 32(1):101–118
Eliades DG, Polycarpou MM, Charalambous B (2009) Security issues in drinking water distribution
networks. J Inf Assur Secur 4:500–508
Hart W, Berry J, Riesen L, Murray R, Phillips C, Watson J (2007) SPOT: a sensor placement
optimization toolkit for drinking water contaminant warning system design. Technical report
SAND2007-4393 C, Sandia National Laboratories
Lee B, Deininger R (1992) Optimal locations of monitoring stations in water distribution system.
ASCE J Environ Eng 118(1):4–16
Liberatore S, Sechi GM (2009) Location and calibration of valves in water distribution networks
using a scatter-search meta-heuristic approach. Water Resour Manag 23(8):1479–1495
Mogheir Y, de Lima JLMP, Singh VP (2009) Entropy and multi-objective based approach for
groundwater quality monitoring network assessment and redesign. Water Resour Manag
23(8):1603–1620
Ostfeld A, Uber JG, Salomons E, Berry JW, Hart WE, Phillips CA, Watson JP, Dorini G,
Jonkergouw P, Kapelan Z, di Pierro F, Khu ST, Savic D, Eliades D, Polycarpou M, Ghimire SR,
Barkdoll BD, Gueli R, Huang JJ, McBean EA, James W, Krause A, Leskovec J, Isovitsch S, Xu
J, Guestrin C, VanBriesen J, Small M, Fischbeck P, Preis A, Propato M, Piller O, Trachtman GB,
Wu ZY, Walski T (2008) The battle of the water sensor networks (BWSN): a design challenge
for engineers and algorithms. ASCE J Water Resour Plan Manage 134(6):556–568
WHO (2008) Guidelines for drinking-water quality, vol 1, 3rd edn. Incorporating 1st and 2nd
addenda. World Health Organization