PROCESS SYSTEMS ENGINEERING

Consensus Algorithm-Based Approach to Fundamental

Modeling of Water Pipe Networks
Shin Je Lee and Jong Min Lee
School of Chemical and Biological Engineering, Institute of Chemical Processes, Seoul National University,
Gwanak-gu, Seoul 08826, Republic of Korea
Jingbo Wu and Frank Allg€ower
Institute for Systems Theory and Automatic Control, University of Stuttgart, Stuttgart 70569, Germany
Jung Chul Suh
Samchully Corporation, Yeongdeungpo-gu, Seoul 07328, Republic of Korea
Gibaek Lee
Dept. of Chemical and Biological Engineering, Korea National University of Transportation, Chungju-si,
Chungbuk 27469, Republic of Korea

DOI 10.1002/aic.15760
Published online in Wiley Online Library (wileyonlinelibrary.com)

Modeling the flow dynamics of leaks in water pipe networks is an extremely difficult problem due to the complex
entangled network structure and hydraulic phenomenon. A mathematical model for leak dynamics in water pipe net-
works based on consensus algorithm and water hammer theory is proposed. The resulting model is a simple and linearly
interconnected system even though the dynamics of water pipe networks has considerable complexity. The model is then
validated using experimental data obtained from real pipe network. A comparative study demonstrates the proposed
model can describe the real system with high qualitative and quantitative accuracy and it can be used to develop
model-based leak detection and location algorithm based on state estimation. To show applicability of the proposed
model, we apply cooperative H1 estimation to the developed model. The results demonstrate the consensus-based
pipe model can be potentially used for leak detection and location with state estimation. V
C 2017 American Institute of

Chemical Engineers AIChE J, 00: 000–000, 2017

Keywords: mathematical modeling, water pipe networks, consensus algorithm, fault diagnosis, state estimation

Introduction networks must be immediately identified and repaired after the

Water pipe networks are one of the largest public infrastruc-
tures of industrial society and are used to efficiently transport
water resources throughout cities. As a vital municipal system
for water distribution, these networks require continuous strict
and thorough maintenance. Nevertheless, flow leaks in water
pipe networks frequently occur and cause serious problems in
terms of safety and cost. Regarding cost, such leaks cause the
loss of water, which results in an inevitable monetary loss and
unnecessary energy usage because additional pumping energy
is required to satisfy the specified carrying capacities. Further-
more, the high maintenance costs attributed to leaks, including
rehabilitation and replacement, lead to a huge financial waste
for society. Leaks are not only an economic issue but also a
safety and potentially a health issue. This is because leaks may
introduce contaminants into the pipeline at a relatively low
pressure when the incident occurs, thus deteriorating water
purity and quality. For these reasons, leaks in water pipe
Correspondence concerning this article should be addressed to J. M. Lee at
jongmin@snu.ac.kr.
C 2017 American Institute of Chemical Engineers
V pipe networks is a difficult problem because it involves
incident to avoid unnecessary loss of resources.
Governments and academic communities have devoted con-
siderable efforts to developing efficient leak detection and
location strategies to reduce leak accidents to an economically
optimal level in water pipe networks. There are numerous con-
ventional methods for detecting leaks, including transient-
based techniques using a negative pressure wave (NPW),
inverse-transient analysis (ITA), and frequency domain techni-
ques. Among them, the NPW-based method is one of the most
popular approaches.1–3
In recent years, model-based estimation techniques for leak
detection have been intensively studied, and most of them
have developed detection algorithms based on Kalman filter.4
Emara-Shabaik et al.5 proposed a nonlinear multiple model
state estimation scheme using a modified extended Kalman fil-
ter to detect and diagnose leaks. This estimation technique,
however, requires a dynamic process model of the system, and
its estimation and detection performances depend on the corre-
sponding model accuracy.
An appropriate leak dynamics model does not currently
exist. The mathematical modeling of leak dynamics in water

AIChE Journal 2017 Vol. 00, No. 00 1

extremely large system dimensions and complexity caused by
close-meshed networks and unpredictable events. From a prac-
tical perspective, this makes it difficult to generate a suffi-
ciently accurate and reliable model in an acceptable error
range. Therefore, the development of an appropriate model for
leak dynamics in water pipe networks is important for applica-
tion of the estimation to detect and locate leaks.
In general, water pressure is constantly maintained under
normal conditions, but when a leak occurs, the water pressure
suddenly decreases. Taking the pressure before the leak as a
reference criterion, the wave generated by such a leak is called
a NPW.6 When a leak occurs along the pipeline, it first
removes the confining pressure produced locally by the pipe
wall, allowing an outward flow from the pipe. Consequently, a
low-pressure water hammer wave is generated that propagates
from the location of the leak into the remainder of the system.
In water pipe networks, the NPW continually reflects and
attenuates after its formation by a leak. Therefore, the NPW
dynamics should be taken into account since it represents a
faulty state and the propagation of the leak signal in water
pipe networks. However, this is the most difficult aspect in
terms of model development.
Some early works have attempted to model hydraulic phe-
nomena of water pipe networks. Because of the considerable
complexity of water pipe networks, early mathematical
approaches typically relied on substantially simplified network
hydraulics by dropping all nonlinearities, which is often unac-
ceptable in practice. The model developed by Maschler and
Savic7 is linearized around a given point and redundant nodes
are eliminated with Gauss-Jordan elimination based on the
simplification range identified in the water network model.
Then, the remaining nodes are relinked with pipes according to
the structure of the reduced network model. Burgschweiger
et al.8 proposed detailed models suitable for nonlinear optimiza-
tion of daily network operations under reliable demand forecasts.
Boussafeur-Lamoudi9 proposed an automated simplification
method of water network models. These works were all focused
on simplifying water pipe network models that are mathemati-
cally sound but only applicable to small networks. For large net-
works, however, the models are still very difficult to solve and
not accurate for practical application.
In this work, we propose a dynamic model of flow and
NPW triggered by a pipe break in water pipe networks using a
consensus algorithm and water hammer theory. A consensus
means to reach an agreement regarding a certain quantity of
interest that depends on the state of all agents (or dynamic sys-
tems) in networks. A consensus algorithm (or protocol) is an
interaction rule that specifies the information exchange
between an agent and all of its neighbors on the network.10
Consensus algorithms have recently been extensively studied
to describe coordination tasks in various areas of science and
engineering, particularly in the context of cooperative control
of multiple autonomous vehicles, formation control, decentral-
ized task assignment, and sensor networks.11 The flow dynam-
ics in water pipe networks shows similar dynamics when
discrete positions within the network are modeled as dynamic
nodes that interact with their neighbors.
To describe the flow dynamics of a leak, we use water ham-
mer theory and modify it to represent a NPW since the water
hammer equations generally describe positive pressure waves.
The mass balance equation of the water hammer is used to rep-
resent the relationship between the fluid velocity and water
pressure in pipe networks. The significance of the proposed
2 DOI 10.1002/aic Published on behalf of the AIChE
model lies in that the flow dynamics of complex water pipe
networks can be represented by a simple and linear model
within an acceptable error range.
The developed model is validated using experimental data
obtained from field tests. A leak event is simulated by quickly
opening a hydrant valve installed in the network. From the val-
idation results, it is realized that the presented model can
effectively describe the leak flow dynamics of the real water
pipe network.
Using the proposed model and distributed state estimation
scheme, we develop a leak detection and location algorithm of
water pipe networks. The detection algorithm is based on dis-
tributed and cooperative H1 -estimation for large-scale inter-
connected linear systems proposed in Wu et al.12
Cooperative H1 -estimation combines the benefits of both
decentralized and distributed estimation in that cooperation
between the local estimators is used to deal with lack of local
detectability and the complexity of local estimators does not
grow with the total size of the system in contrast to the exist-
ing distributed estimations. An important requirement of
decentralized estimation approach is that the local subsystems
are detectable from local measurements. In a distributed esti-
mation, on the other hand, multiple estimators cooperate with
each other and create an estimate of subsystem without local
detectability. Instead, it gives rise to scalability issue where
the order of the estimators grows with the size of the network.
Therefore, the proposed method is referred to as Cooperative
estimation which overcomes local detectability and scalability
by only reproducing a desired subset of states for a local estima-
tor. It is an H1 -based design that provides guaranteed perfor-
mance with respect to model and measurement disturbances.
The methodology is generally applied to large-scale linear sys-
tems where subsystems may be physically interconnected and
thus also applied to the developed flow dynamics model of
water pipe networks.
The remainder of this paper is organized as follows. Prelim-
inaries are provided in the next section, and the procedure for
modeling the leak dynamics of water pipe networks is pre-
sented in the section entitled Modeling Procedure based on a
consensus algorithm and water hammer equation. The devel-
oped model is validated using experimental data in the section
entitled Validation of the Developed Model and leak detection
and location algorithm by application of distributed estimation
to the developed model is presented next, followed by con-
cluding remarks.
Preliminaries
In this section, we introduce some basic notations and sum-
marize useful concepts from graph theory to apply to water
pipe networks.
Network topology
A process network can be represented by a directed graph in
which the nodes of the graph correspond to process units and
junctions. The edges of the graph correspond to connecting
streams or pipelines. The mathematical model of the water pipe
network is based on a directed, unweighted graph G5ðN ; AÞ
that describes the network topology between individual nodes.
N is the set of nodes, including virtual nodes on the pipeline,
junctions sensors
2017 Vol. 00, No. 00 AIChE Journal
 Figure 1. Pipeline nodes k51;    ; N including a leak node kl with velocity .
N 5N node [ N junction [ N sensor 5fn1 ;    ; nN g
where nk 2 N ðk51;    ; NÞ represents the kth node. A  N
3N is the set of edges and represents pipe segments between
nodes
A5Apipeseg 5fðnj ; nk Þjk51;    ; N; j 2 N k g
where N k 5fj : ðnj ; nk Þ 2 Ag is the set of nodes that the node
k receives measurement information from and called the
neighborhood of the node k. Equation 2 models the informa-
tion or physical flow, i.e., the kth node is coupled to the jth
node if and only if ðnj ; nk Þ 2 A. The flow in an arc (nj, nk) is
defined to be positive when it is directed from j to k; other-
wise, it is negative or possibly zero. The flow is always non-
negative in arcs where only one direction is possible.
Modeling Procedure
In this section, we propose a fundamental model of flow
dynamics in water pipe networks when a pipe leak occurs. The
proposed model is based on the consensus algorithm and clas-
sical water hammer theory, which will be described in the fol-
lowing subsections. Then, the dynamics at the leak point is
described based on the concepts of the water hammer effect,
and the complete model is summarized in the last subsection.
Consensus algorithm
We use a consensus algorithm to model the flow dynamics
in water pipe networks because the nodes in water pipe net-
works reach a common value of state (on their own).
When multiple nodes (agents) agree on the value of a vari-
able of interest, they are said to have reached consensus. To
achieve consensus, there must be a shared variable of interest,
called the state, as well as appropriate algorithmic methods for
negotiating to reach consensus on the value of that variable,
called consensus algorithms. Consensus algorithms are
designed to be distributed assuming only neighbor-to-neighbor
interactions between nodes and limited information about
other parts of the system.11
If the effects of the interconnections are continuous-time,
then the state update of each node can be modeled using a dif-
ferential equation. A scalar state is updated by each node using
a first-order differential 11 Suppose that there are n nodes in
the network. The most common continuous time consensus
algorithm is given by
: X
n
xi ðtÞ52 aij ðtÞ½xi ðtÞ2xj ðtÞ; i51;    ; n
j51
where aij(t) is the consensus weight of the ith node between
node i and j and xi(t) is the state of the ith node at time t. Since
the pipe network topology is time invariant, the gains aij are
constant. Setting aij 5 0 indicates that the node i is not coupled
to the node j. A consequence of Eq. 3 is that the state xi(t) of
the node i achieves to the same state of its neighbors. In water
AIChE Journal 2017 Vol. 00, No. 00
(1) pipe networks, we use deviation variables of fluid velocity and
water pressure with the nominal values since the velocity and
pressure values are different for each pipeline.
Flow dynamics based on consensus algorithm
(2)
From the above consensus algorithm, we propose a flow
dynamics model of water pipe networks when a leak occurs.
The consensus algorithm can be applied to describe the propa-
gation of a NPW in water pipe networks because the phenom-
ena of the signal transfer to neighboring nodes are similar in
both systems. Therefore, the consensus algorithm can be
applied to fluid velocity as follows
dvk X
5Kkj ðvj 2vk Þ; k 2 N (4)
dt j2N k
where vk is the fluid velocity at node k and Kkj is the consensus
gain of node k. The gain Kkj can be varied with the network
characteristics.
When a leak occurs in the pipeline, the transient signal
propagates in the upstream and downstream, and it will affect
the velocities of adjacent nodes. Figure 1 shows an example of
a pipeline with a leak.
Equation 4 should be modified to reflect the leak velocity at
the nodes adjacent to the leak node, kl. When the water flows
from left to right in the pipeline, Eq. 4 is changed as follows:
at node k which is on the left side of the leak kl
dvk X
5Kkj ðvj 2vk 1Þ; k5kl 21 (5)
dt j2Nk
and at node k, which is on the right side of the leak kl
dvk X
5Kkj ðvj 2vk 2Þ; k5kl 11 (6)
dt j2N k
where  is the leak velocity. When a leak happens, the
fluid velocity decreases as much as  at the leak point. At the
(kl – 1)th node, the reduced velocity, , should be added as in
Eq. 5 because the flow at this point is affected by the leak as
much as 1 and the velocity at the neighbor node j would be
reduced by . Whereas, at the (kl 1 1)th node, the reduced leak
velocity should be subtracted as in Eq. 6 because the flow at
this point is affected by the leak as much as –.
(3) Water hammer equation
The water hammer phenomena have been extensively stud-
ied and well established since the early 1900s in this field.
With the ever increasing importance of water hammer phe-
nomena, many researchers have developed water hammer
theories from first principles, and their combined efforts have
resulted in the following classical mass and momentum equa-
tions for one-dimensional (1-D) water hammer flows13
Published on behalf of the AIChE DOI 10.1002/aic 3
 @V 1 @P dvkl
1 50 (7) 52Mðpkl 2pkl 21 Þ (16)
@x qa2 @t dt
@V 1 @P 4 where M51=qDx and pkl and vkl are the pressure and velocity
1 1 sw 50 (8)
@t q @x qD at the leak point, respectively.
where V is the cross-sectional average velocity, q is the fluid Complete model
density, a is the acoustic water hammer wave speed, P is the We summarize the complete model of the leak dynamics of
water pressure, D is the pipe diameter, sw is the shear stress at water pipe networks in this section. The developed model con-
the pipe wall, x is the spatial coordinate along the pipeline, sists of Eqs. 4–6 and 13–16 and it is formulated as a state
and t is the temporal coordinate. The water hammer wave space model as follows
speed can be calculated using the following formula
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi _
x5Ax1Bu1m (17)
K
a5
q
(9) y5Cx1g (18)
11/ KE De
where x5½p1 ; v1 ;    ; pkl ; vkl ;    ; pN ; vN > 2 R2N is the state
where K is the bulk modulus of elasticity of the fluid, E is of the system, u 5  is the unknown input, y 2 Rm is the mea-
Young’s modulus of elasticity, e is the pipe wall thickness, surement vector, A 2 R2N32N is the system matrix, B 2 R2N31
and / is a parameter that depends on the pipe anchoring. is the input matrix, C 2 Rm32N is the measurement matrix,
From the literature,13 we know that friction in the pipe and m 2 R2N and g 2 Rm are the noise vectors. The measure-
becomes negligible and sw can be safely set to zero in some ment C is defined by sensor number and location in the net-
cases. For example, wall friction is irrelevant as long as the works, with 1 for elements of Cð1; k1 Þ;    ; CðNs ; kNs Þ where
simulation time is significantly smaller than 4L=a. Then, the Ns is the sensor number and zero for all other entries. The
classical water hammer model given by Eqs. 7 and 8 becomes noise should be included in the model to portray the real sys-
@V 1 @P tem since the real water pipe network possesses many uncer-
1 50 (10) tainties. The system matrices, A and B, can be represented as
@x qa2 @t 2 3
@V 1 @P 0 L 0 2L 0  0
1 50 (11) 6 7
@t q @x 6 7
6 0 2K1 0 K1 0  0 7
6 7
By using the mass balance of Eq. 10, we can obtain the rela- 6 7
6 7
tionship between fluid velocity and water pressure in the pipe 6 .. 7
6 . 7
network. At node k, Eq. 10 is changed into the following form 6 7
6 7
6 7
@vk 1 @pk 6 0  0 L 0 2L 0    0 7
1 50; k2N (12) 6 7
@x qa2 @t A56 6
7
7
6 0  M 0 2M 0 0    0 7
The partial derivative of vk with respect to x in Eq. 12 should 6 7
6 7
be discretized to solve it, and we apply a backward finite dif- 6 7
6 .. 7
ference method (FDM) since the variable does not consider- 6 . 7
6 7
ably change in a short length in the pipeline. Using the first- 6 7
6 7
order FDM, Eq. 12 can be expressed for pressure as follows 60  0 L 0 2L 7
6 7
4 5
dpk
52Lðvk 2vk21 Þ; k2N (13) 0  0 KN 0 2KN
dt
(19)
where L5qa2 =Dx and Dx is the length between nodes.
When a leak occurs, Eq. 13 should be modified at the nodes and
adjacent to the leak node kl B5½ 0    0 Kkl 21 0 2Kkl 11 0    0 L 0 L 0    0 >
dpk (20)
52Lðvk 2vk21 2Þ; k5kl 61 (14)
dt
A and B can differ depending on the network topology and the
For the water pressure variable, it does not distinguish the location of the leak.
flow direction because the NPW is considerably faster than the It is worthwhile to note that the presented model has a form
fluid velocity. of simple and linear interconnected equations even though
the system has highly complex dynamics. This is attributed to
Dynamics at leak point the form of consensus algorithm applied to this system and the
first-order FDM of water hammer equation. The first-order
The flow dynamics at the leak point can be described by the
FDM is appropriate because the fluid velocity does not change
water hammer effect explained in the previous section because
much in the short length less than 10 m.
the propagation of a pressure wave is similar except for the
negativeness or positiveness of the pressure wave. Then, the
pressure and flow at leak node kl have the following dynamics Validation of the Developed Model
dpkl In this section, the developed dynamic model is validated
52Lðvkl 2vkl 21 Þ (15) using experimental data obtained from a real water pipe
dt

4 DOI 10.1002/aic Published on behalf of the AIChE 2017 Vol. 00, No. 00 AIChE Journal
Figure 2. Water pipe network of test area with 50 nodes and 48 links.
[Color figure can be viewed at wileyonlinelibrary.com]

network as explained in Ref. 2. First, we introduce geometry
information of the water pipe network and present steady-
state and unsteady-state water pressures measured from
sensors in the test area. Second, the developed dynamic
model is compared with the experimental data, and the
related residuals and errors are calculated in the following
subsection.
Experiments
The developed dynamic model was validated using the field
data obtained from a real water pipe network in South Korea.
Without any households or factories near the test area, the
experiment could be controlled for sensors to capture the leak
incident only.
The overall pipe network topology is shown in Figure 2.
The experimental pipe network of the test area includes three
pressure sensors and two fire hydrants. The total pipeline
length covered by these three sensors is 452 m. Although the
pipe network is not as complex as a real-life water pipe
network, it is sufficient to serve as a field test for validation
since it includes sufficient pipe length and two junctions to
characterize the network system.
Link over 10 m are divided by the node division strategy
proposed in Ref. 2. After the node division, the final node
matrix A0 has dimensions of 50 3 50, and the minimum and
maximum step sizes of arcs are 9.4841 and 9.6410 m, respec-
tively. The final 50 network nodes are marked on the network
displayed in the black box of Figure 2.
Here, the fire hydrants were used to generate a leak by open-
ing valves. This process could simulate a pipe leak by sud-
denly discharging water, leading to a NPW in the pipe
network. Figure 3 shows the pressure measurement station
used in the test and the water discharge of the hydrant while
conducting the experiment.
Pressure transients are measured by three pressure sensors,
named UNIK 5000 (GE). They are connected to a PC
equipped with 12 bit A/D and D/A converters. The sampling
frequency of the pressure sensor is 250 Hz; that is, the pressure

Figure 3. (a) Measurement station of pressure and (b) water discharge by opening hydrant valve (photographs pro-
vided by Hun-joo Cha of Samchully Co., Ltd.).
[Color figure can be viewed at wileyonlinelibrary.com]

AIChE Journal 2017 Vol. 00, No. 00 Published on behalf of the AIChE DOI 10.1002/aic 5
 obviously distinguished and they are commonly used for leak
detection and location of pipe networks.2,6,14
We obtained a total of ten experimental data sets, which
were generated at Hydrant 1. The water flow was discharged
at approximately 3 L/s at Hydrant 1, and the normal flow rate
of this network continuously varied from 4.17 to 8.33 L/s.

The proposed model

To simulate a complete model for this test area, we assign
network nodes, construct system matrices, A, B, and C,
according to the network topology and adjust model parame-
ters. The total node number N of this network is 50, so the sys-
tem dimension is R100 , and the leak velocity, , is calculated
as 0.38 m/s from cross-sectional area of the pipeline. In addi-
tion, the system and measurement disturbances, m and g, are
added as white Gaussian noise (WGN) to mimic the effect of
many random processes that occur in the network. Then, the
simulation results are compared with the experimental data.
As in the experimental data shown in Figure 4, the steady-
state and unsteady-state behaviors of the water pressure are
simulated using the proposed model, as shown in Figure 5.
The average pressure values of three sensors obtained from
the measurement are employed in the simulation as initial val-
ues. In Figure 5a, the water pressures are constantly main-
tained with background noise under totally normal state
without any leak as in Figure 4a.
In Figure 5b with the leak simulation, it is observed that
the pressure suddenly drops, and it is recovered with some

Figure 4. Experimental data of water pressure for

(a) steady state and (b) unsteady state.
[Color figure can be viewed at wileyonlinelibrary.com]

is sampled every 0.004 s, and one sensor collects 900,000 data

points per hour. The hardware equipment and data compres-
sion technique were uniquely designed to accommodate the
massive data set online in the test area.
During the field test, there was no other hydrant or pump
operating around the area. The valve opening action may
affect the pressure transient; however, it would be insignificant
compared with the influence of the sudden water release.
Additionally, hydraulic phenomena at pipe junctions or valves
may always exist; however, such phenomena appear as the
background noise of the system. Therefore, it is assumed that
the pressure transient changes in the experiment were caused
only by the sudden water discharge of the hydrant.
Figure 4 shows the steady-state and unsteady-state behav-
iors of water pressure for 30 s driven by the hydrant valve
opening. In Figure 4a, the water pressure was maintained at a
constant value with background noise under normal conditions
without any incidents. It is observed that the water pipe system
has a lot of noises and disturbances even under normal and
steady circumstances. The averaged pressures for Sensors 1, 2,
and 3 were 250, 267, and 245 kPa, and the variances were
3.46, 7.40, and 5.20, respectively.
The water pressure was maintained at a constant value at nor-
mal conditions, however, when the hydrant valve was opened,
it generated the NPW and sudden pressure drops due to the
NPW were clearly observed, as shown in Figure 4b. After the Figure 5. Simulation results of water pressure for
pressure dropped, the resulting pressure wave quickly disap- (a) steady state and (b) unsteady state.
peared after several fluctuations. The first two fluctuations are [Color figure can be viewed at wileyonlinelibrary.com]

6 DOI 10.1002/aic Published on behalf of the AIChE 2017 Vol. 00, No. 00 AIChE Journal
 Application of Cooperative Estimation to Water
Pipe Networks
Distributed and cooperative estimation proposed in Wu
et al.12 is applied to the developed model for fault detection
and location. The estimation is appropriate for large-scale
interconnected system such as the developed water pipe net-
work model. Especially, the fact that dimensions of local esti-
mators do not grow with the network size can be more useful
for larger system such as water pipe network. Furthermore,
another important point of this estimator is that it does not
require local detectability of subsystems and this is particu-
larly suitable for systems with insufficient sensors.
The cooperative estimator is designed to guarantee h1 -
performance for a class of linear interconnected systems where
system uncertainty can be modeled as an additive unknown
disturbance term in the dynamic equation. Therefore, it can
give asymptotically stable state estimates in application to the
developed water pipe network model that includes disturbance
terms.
The above points justify practical use of cooperative estima-
tion12 for water pipe networks.

System model and repartitioning the system

Figure 6. Enlarged pressure transient in the occurrence
A large-scale linear time-invariant system consists of N
of a leak for (a) experiment and (b) simulation.
interconnected subsystems that are each described by the dif-
[Color figure can be viewed at wileyonlinelibrary.com] ferential equations
X
N
x_ k 5Ak xk 1 Akj xj 1Bk u1mk (21)
fluctuations as in Figure 4b. However, the proposed model is j51

an ideal case, and there is a difference between realistic and X

N
ideal systems. yk 5Ck xk 1 Ckj xj 1gk (22)
If a sensor is located near the leak point, the pressure drop j51
would be more obvious than ones at other sensors. To show
that the proposed model can describe the time delay effect of for k51;    ; N, where xk 2 Rnk is the state variable, yk 2 Rrk
the NPW on sensors, the initial parts of the pressure drop for is the output, and mk ðtÞ; gk ðtÞ 2 q2 ½0; 1Þ are q2 -integrable
both the experimental data and simulation results are disturbance inputs of subsystem k. The subsystems can be
enlarged as shown in Figure 6. The pressure decreases in the incorporated in global interconnected system as written in
order of Sensors 1, 2, and 3 for both the experimental and Eqs. 17 and 18 with
simulation results, and it corresponds with the network topol-
ogy in Figure 2 because Hydrant 1 is close to sensors in the
order of Sensors 1, 2, and 3. Thus, the difference in accor-
dance with the sensor location can also be represented by the
proposed model.
We validated the model with 10 data sets obtained from the
experiments. Figure 7 shows comparative results of the devel-
oped model for the selected experimental data set. The resid-
ual errors for three sensors are less than 35, and the relative
error is calculated to be 16%. As shown in Figure 7, the fluctu-
ation part causes the most severe deviation between the model
and experiment. The average and maximum errors for 10
experimental data sets are calculated as 1.62% and 14.0%,
respectively. From this result, although the fluctuation part
generates a relatively large deviation from the experiment, the
developed model shows satisfactory results with high qualita-
tive and quantitative accuracy and it could be used to develop
a leak detection and diagnosis algorithm with estimation meth-
ods that require the dynamic model of the system. We develop
a leak detection and diagnosis algorithm using the proposed Figure 7. Residual of water pressure from experiment
model and cooperative state estimation technique in the subse- and simulation.
quent section. [Color figure can be viewed at wileyonlinelibrary.com]

AIChE Journal 2017 Vol. 00, No. 00 Published on behalf of the AIChE DOI 10.1002/aic 7
 2 3
A1 A12  A1N
6 7
6 7
6 A21 A2 ⯗ 7
6 7
6 7
A56 7 (23)
6 .. 7
6 ⯗ . ⯗ 7
6 7
4 5
AN1   AN
 >
B5 B>1  B>
N (24)
2 3
C1 C12  C1N
6 7
6 7
6 C21 C2 ⯗ 7
6 7
6 7
C56 7 (25)
6 .. 7
6 ⯗ . ⯗ 7
6 7
4 5
CN1     CN
> >
by using the stacked vectors x5½x>
1 ;    ; xN  ;
> > > > > >
m5½m1 ;    ; mN  , and g5½g1 ;    ; gN  .
The state should be repartitioned for designing local estima-
tors. We follow repartitioning policy presented in Wu et al.12
For all k51;    ; N, the global interconnected system of Eqs.
17 and 18 can be written as
2 ðkÞ 3 2 ðkÞ 32 ðkÞ 3 2 ðkÞ 3 2 ðkÞ 3
x_ AðkÞ A~ x B m
4 55 4 54 51 4 5u14 5
ðkÞ ðkÞ ðkÞ ðkÞ
x_ ðkÞ
c A~
c A
ðkÞ
c x c B c m c

(26)
ðkÞ ðkÞ
yk 5C x 1gk (27)

by permutation of the states. The rk-dimensional partial state

>
vector xðkÞ 5½x> >
nk ð1Þ ;    ; xnk ðrk Þ  is determined by a selection Figure 8. Three subsystems of water pipe network and
function nk ðÞ that determines which scalar components xj,i of the partial state vectors xðkÞ ; k51; 2; 3 for
the state xj are included in x(k) and the partial state vector xc
ðkÞ each subsystem.
(k) [Color figure can be viewed at wileyonlinelibrary.com]
consists of the remaining states after selecting x by nk ðÞ.
Some or all elements of the global state vector x may be cho-
sen that are relevant to subsystem k. For instance, xðkÞ may
contain xk, but it does not have to include all of them, if some Fault detection and location using state estimation
parts of its own state are not important. This represents a In this section, we propose Algorithm 1 to detect and locate
degree of freedom in the design of the estimators. It is required fault of water pipe networks by using the developed model
that all xj,i which contribute toward yk are included in x(k). As a and cooperative h1 -estimation.
result, every output yk can be equivalently expressed as Eq. The problem considered is to design a local estimator for
27. Also, the partial state vectors x(k) may overlap, e.g., x(1) every subsystem k that creates an estimate for the local partial
and x(2) may contain a common component xk where k5ðk; iÞ state variable x(k) using the local measurements yk described in
for k51;    ; N and i51;    ; nk . Eqs. 26 and 27. The estimator dynamics is proposed for each
We require the following assumption which is a sufficient subsystem as12
condition in the centralized case. ðkÞ X ðkÞ fðkÞ
Assumption 1. The global plant ðA; CÞ is observable or x^_ 5AðkÞ x^ðkÞ 1LðkÞ ðyk 2CðkÞ x^ðkÞ Þ1 ½A~ k x^k
ðkÞ
detectable. 0 k2Ic
1
This assumption is to make the state estimation problem X X (28)
@ ðjÞ ðkÞ
under consideration meaningful. 1K ðkÞ
en21 ðkÞ ð^
x k 2^x k ÞA
Figure 8 shows the divided subsystems in the corresponding k2Nk ðkÞ
k2Ic \I ðjÞ
water pipe network shown in Figure 2 and which scalar com-
ponents are selected by nk ðÞ and included in the partial state where the local estimate vector is denoted by
> ðkÞ
vector xðkÞ ðk51; 2; 3Þ. The overall pipe networks is divided x^ðkÞ 5½^
x> ^>
nk ð1Þ ;    ; x nk ðrk Þ  ; x
^k is the estimate for xk computed
into three subsystems and they receive measurement informa- ðkÞ ðkÞ
tion from their corresponding sensors and estimate the partial at subsystem k; ½A~ k is the column of A~ which corresponds
ðjÞ ðkÞ
states x(k). The junction node 22 is overlapped in all three sub- to xk ; and the unit vector en21 ðkÞ injects the difference x^k 2^xk
systems and its state components ½x1;43 ; x1;44 > are estimated to the rk-dimensional space. The estimator gain matrices L(k)
simultaneously in each subsystem as shown in Figure 8. and K(k) can be calculated by Theorem 1.12

8 DOI 10.1002/aic Published on behalf of the AIChE 2017 Vol. 00, No. 00 AIChE Journal
 2 ðjÞ 3
Theorem 1. The problem admits
 21a solution in theform21 of pj Pn21 0 0
estimators Eq. 28 with LðkÞ 5 PðkÞ GðkÞ and K ðkÞ 5 PðkÞ 6 k ðk1;j Þ 7
6 7
FðkÞ if the matrices F(k), G(k), and P(k) for k51;    ; N are a 6 7
ðkÞ 6
0 7
ðjÞ
solution of the following LMIs Uj 56 0 pj Pn21 7 (35)
6 k ðk2;j Þ 7
2 ðkÞ ðkÞ
3 6 7
QðkÞ 1W ðkÞ 2GðkÞ PðkÞ BðkÞ SðkÞ Tj1 ;k    Tjs ;k 4 5
6 k
7 ..
6 7 0 0 .
6 7
6 2GðkÞ> 2cI 0 0 0 0 0 7
6 7 where fkkj kj ðkÞ
\ I ðjÞ .
6 7 1 ; k2 ;   g5I
6  7 The proof is provided in the literature.12 The cooperative esti-
6 PðkÞ BðkÞ > 0 2xI 0 0 0 0 7
6 7 mator determines a collection of estimates x^ðkÞ ðtÞ; k51;    ; N,
6 7
6 7 that provides guaranteed h1 -performance in the sense that
6 S ðkÞ>
0 0 R ðkÞ
0 0 0 7
6 7 N ð1 N ð1

6 7 X X 
6 7 ðkÞ> W ðkÞ ðkÞ dt  x2 jjmjj2 1c2 jjgk jj2 dt1I0
6 ðkÞ> ðkÞ 7
6 Tj1 ;k 0 0 0 Uj1 ;k 0 0 7 0 0
6 7 k51 k51
6 7 (36)
6 7
6 .. 7
6 ⯗ 0 0 0 0 . 0 7 (k)
6 7 for a positive semidefinite weighting matrix W . Here, I0 5
4 5 PN ðkÞ> ðkÞ
ðkÞ>
Tj1 ;k 0 0 0 0 0 Uj1 ;k
ðkÞ k51 x0 PðkÞx0 is the cost due to the observer’s uncer-
tainty about the initial conditions of the subsystems and ðkÞ
<0
is the local estimation error vector defined as ðkÞ 5xðkÞ 2^ x ðkÞ
(29) > > > > > rk
x nk ð1Þ ;    ; xnk ðrk Þ 2^
5½xnk ð1Þ 2^ x nk ðrk Þ  2 R . When m 5 0 and
with fj1;k ; j2;k ;    ; jsk ;k g5Nk . The matrices are defined as g 5 0, the estimation errors exponentially decay so that ðkÞ
follows ! 0 for all k51;    ; N.
To apply the state estimation of Eq. 28, we first identify
where the leak occurs, i.e., leak node kl. By comparing pres-
! sure measurements with the proposed model for kl 51;    ; M
X X where M is the node number of the network, we can find the
ðkÞ
N 5 en21 ðkÞ e>
n21 ðkÞ
(30)
j2Nk
leak node kl such that the difference between the model and
k2I ðkÞ \I ðjÞ
 > measurement is minimized. Now we can identify all the states
QðkÞ 5PðkÞ AðkÞ 1AðkÞ> PðkÞ 2GðkÞ CðkÞ 2 GðkÞ CðkÞ of the network with cooperative and distributed estimation. If
 > faults occur in the pipe network, a NPW is generated and
2FðkÞ N ðkÞ 2 FðkÞ N ðkÞ propagates from the leak point to the entire network. There-
2 ðkÞ 3 fore, the NPW signals would appear in order of distance and
qðk; nk ð1ÞÞP1 0 0 this distance difference is also displayed on the pressure sen-
6 7
6 7 sors as in Figure 6. After all the states are estimated, the leak
6 7
6 . 7 location is easily identified since the node having the steepest
1a1 pk 6 0 . . 0 7 Pk
6 7 pressure drop should be nearest to the leak point. This method
6 7
4 5 does not require any fault location algorithm unlike the previ-
ðkÞ ous works.2,3 We set up edges between nodes are all less than
0 0 qðk; nk ðrk ÞÞPrk
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl} 10 m according to the node division strategy so that the loca-
(31) tion error range is within 10 m.
Now we propose the following algorithm.
ðkÞ (k)
where Pi 2 R is the ith diagonal element of P , pk and a Data: AðkÞ ; BðkÞ ; CðkÞ for subsystem xðkÞ ; k51;    ; N
are positive constants, and pðkÞ is the diagonal element of Result: Fault location
PðfðkÞÞ which corresponds to xk and Compute L(k) and K(k) offline;
 h i h i Read pressure measurements;
ðkÞ ðkÞ ~ðkÞ ðkÞ ~ðkÞ
S 5 P A ðkÞ
P A ðkÞ
   (32) if Fault occurs then
k1 k2 for j51;    ; M do
2
 3 Calculate leak dynamics x using Eqs. 17–20 for kl 5 j;
ðkÞ
pfðkðkÞ Þ p k1 0 0 P
Ns
6 1 7 Find j such that min j jjxl ðkl 5jÞ2ym;l jj2 ;
6 7
6
 7 l51
6 ðkÞ 7
RðkÞ 56 0 pfðkðkÞ Þ p k2 0 7 (33) end
6 1 7 kl j;
6 7
4 5 Implement estimator Eq. 28 and find x^ðkÞ ; k51;    ; N;
.. Give fault location;
0 0 .
else
ðkÞ ðkÞ ðkÞ Give no fault sign;
for fk1 ; k2 ;   g5Ic and
end
Algorithm 1. Fault detection and location algorithm
h i N and Ns is the subsystem and sensor numbers, respectively,
ðkÞ ðkÞ
Tj 5 F en21 FðkÞ en21 ðk2;j Þ  (34) and xl ðkl 5jÞ is the dynamics at sensor l when kl 5 j, and ym,l
k ðk1;j Þ k

AIChE Journal 2017 Vol. 00, No. 00 Published on behalf of the AIChE DOI 10.1002/aic 9
 the initial estimated state is not accurate compared to the case
where the error is zero. Should the initial estimation error be
greater than 2%, it would affect the leak identification result
as shown in Figure 10b; it is difficult to determine which node
is closest to the leak node because the slopes of the pressure
drop are similar due to the initial estimation error. However,
the initial estimation error is determined as the average value
of the nominal state obtained from the experimental data mea-
sured in the test area. It is found from the measurement data
that the initial estimation error does not exceed 2% of the
actual initial state in the presence of noise and model-plant
mismatch.
We conclude the estimation section with some remarks.
REMARK 1. It is found that (A(1), C(1)), (A(2), C(2)), and
(A , C(3)) defined in Eqs. 26 and 27 are undetectable. We
(3)

emphasize that the states of the network can be estimated

Figure 9. (a) Estimation results of all states in water
pipe network and (b) enlargement of estima-
tion results for 3 s.
[Color figure can be viewed at wileyonlinelibrary.com]
even without guaranteeing detectability of individual subsys-
tems. In fact, the proposed method can be applied to any
subsystem regardless of the structure of the subsystem even
if it does not have detectability if the subsystem has at least
one measurement sensor.
REMARK 2. In water pipe networks considered here, there
are three pressure measurements, subsystems, and local esti-
mators. In the proposed algorithm, local estimators only
produce a desired subset of state variables and the complex-
ity of these local estimators does not increase with the total
size of the water pipe network unless the size of each sub-
system change. Therefore, even if the size of the water pipe
networks increases, the size of local estimator will remain
the same because the local estimator estimates only the
desired states. In this sense, the estimation method presented
here is scalable. However, when the size of the water pipe

is the measurement of sensor l for l51;    ; Ns . First, the

matrices A(k), B(k), C(k) are determined according to the struc-
ture of the subsystem k and then the estimator gain matrices
L(k) and K(k) are computed offline. While reading pressure
measurements, if a fault occurs, an objective function should
be calculated to find which node is a leak node, i.e., kl 5 j
for j51;    ; M and model the leak dynamics using the pro-
posed model. After finding kl, estimator Eq. 28 is imple-
mented and fault location is identified.
Estimation results
The initial conditions of the system and the estimator are set
to 250 kPa. Estimation results of the water pipe network under
consideration are shown in Figure 9. Figure 9a shows all the
estimates of pressure at the nodes in the network can be esti-
mated using the cooperative distributed estimation and the
proposed model with only three measurement information.
The early part of the pressure drop by the leak is enlarged in
Figure 9b. The line with circle shows the steepest drop of all
the estimates and, thus, it is the nearest to the actual leak loca-
tion. Indeed, it is node 4 which is right next to the leak point
and the location error is 9.55 m. As the pressure wave propa-
gates from the leak to other part of the network, the pressure
drop appears in distance order as shown in Figure 9b.
In Figure 10a, the initial estimation error was changed to
2% to show the effect of initial estimation error on the leak Figure 10. State estimation results with initial estima-
location result. As shown in Figure 10a, the leak location can tion error (a) of 2% and (b) over 2%.
be identified with the initial estimation error of 2% although [Color figure can be viewed at wileyonlinelibrary.com]

10 DOI 10.1002/aic Published on behalf of the AIChE 2017 Vol. 00, No. 00 AIChE Journal
network increases, more pressure sensors are needed to
ensure scalability.
Concluding Remarks
The fundamental model of NPW dynamics due to leaks in
water pipe networks is developed based on a consensus algo-
rithm and water hammer theory. The leak dynamics model of
water pipe networks has hardly been considered in the litera-
ture thus far because of the complexity of the water pipe net-
work structure and hydraulics. The developed model has a
form of a simple and linear model of an interconnected net-
work system. Then, it is satisfactorily validated using real
experimental data obtained from a field test. The average and
maximum errors between the model and experiment are within
2 and 20%, respectively, and a large portion of these errors are
attributed to fluctuations of the NPW propagation through the
network, which is difficult to precisely model. Although the
model requires parameter adjustment to accurately simulate
the real water pipe network, it can be used for various applica-
tions, such as the development of a leak detection and location
algorithm. A leak detection and location algorithm based on
state estimation in water pipe networks is subsequently devel-
oped. We apply cooperative and distributed estimation scheme
to the developed model since its complexity does not grow
with the total size of the network and, thus, it can be applied to
large-scale systems where subsystems may be physically inter-
connected such as water pipe networks. We propose an algo-
rithm that can detect and locate the leak based on the
cooperative state estimators and from the estimation result, we
can find a possible leak location where the corresponding node
has the steepest pressure drop and the location error is 9.55 m.
Acknowledgments
This work was supported by the National Research Foun-
dation of Korea (NRF) grant funded by the Korean govern-
ment (MSIP) (NRF-2016R1A5A1009592) and Engineering
Development Research Center (EDRC). This work was sup-
ported by the Korea Ministry of Environment as a Project
AIChE Journal 2017 Vol. 00, No. 00 Published on behalf of the AIChE
for Developing Eco-Innovation Technologies (GT-11-G-02-
001-5).
Literature Cited
1. Colombo AF, Lee P, Karney BW. A selective literature review of
transient-based leak detection methods. J Hydro-Environ Res. 2009;
2:212–227.
2. Lee SJ, Lee G, Suh JC, Lee JM. Online burst detection and location
of water distribution systems and its practical applications. J Water
Resour Plann Manage. 2015;142:04015033
3. Kim Y, Lee SJ, Park T, Lee G, Suh JC, Lee JM. Robust leak detec-
tion and its localization using interval estimation for water distribu-
tion network. Comput Chem Eng. 2016;92:1–17.
4. Ye G, Fenner RA. Kalman filtering of hydraulic measurements for
burst detection in water distribution systems. J Pipeline Syst Eng
Pract. 2010;2:14–22.
5. Emara-Shabaik HE, Khulief YA, Hussaini I. A non-linear multiple-
model state estimation scheme for pipeline leak detection and isola-
tion. Proc Inst Mech Eng Part I: J Syst Control Eng. 2002;216:497–
512.
6. Ge C, Wang G, Ye H. Analysis of the smallest detectable leakage
flow rate of negative pressure wave-based leak detection systems for
liquid pipelines. Comput Chem Eng. 2008;32:1669–1680.
7. Maschler T, Savic DA. Simplification of water supply network mod-
els through linearisation. Cent Water Syst Rep 99. 1999;1:119.
8. Burgschweiger J, Gn€adig B, Steinbach MC. Optimization models for
operative planning in drinking water networks. Optim Eng. 2009;10:
43–73.
9. Boussafeur-Lamoudi K. Automated Simplification of Water Networks
Models. Schneider Electric Internship Report. 2013:1–40.
10. Olfati-Saber R, Fax JA, Murray RM. Consensus and cooperation in
networked multi-agent systems. Proc IEEE. 2007;95:215–233.
11. Ren W, Beard RW. Overview of consensus algorithms in coopera-
tive control. In: Distributed Consensus in Multi-Vehicle Cooperative
Control: Theory and Applications. London: Springer. 2008:3–22.
12. Wu J, Ugrinovskii V, Allg€ ower F. Cooperative H1 -estimation for
large-scale interconnected linear systems. In: 2015 American Control
Conference (ACC), Chicago, IL, USA. 2015:2119–2124.
13. Ghidaoui MS, Zhao M, McInnis DA, Axworthy DH. A review of
water hammer theory and practice. Appl Mech Rev. 2005;58:49–76.
14. Tian CH, Yan JC, Huang J, Wang Y, Kim D, Yi T. Negative pres-
sure wave based pipeline leak detection: challenges and algorithms.
In: 2012 IEEE International Conference on Service Operations and
Logistics, and Informatics (SOLI), Suzhou, China. 2012:372–376.
Manuscript received Dec. 1, 2016, and revision received Apr. 5, 2017.
DOI 10.1002/aic 11