You are on page 1of 12

Efficient Technique for Pipe Roughness Calibration and

Sensor Placement for Water Distribution Systems


Richárd Wéber 1 and Csaba Hős 2
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

Abstract: The need to improve the accuracy of mathematical models describing hydraulic networks (e.g., water distribution systems) poses
numerous challenges regarding the problems of parameter calibration and optimal sampling. In this paper, a comparison is made between
different sensor placing strategies, including a novel, direct (iteration-free) algorithm. The approach used is based on maximizing both the
sensitivity and the so-called hydraulic distance of the sampling points without the need for optimization. Based on these sampled measure-
ments, the pipe roughness coefficient of each pipe is calibrated individually (without grouping) by solving the resulting underdetermined
linear system with singular value decomposition. The efficiency of the approach is demonstrated on case studies, including artificial
(e.g., Anytown and C-town) and real-life water distribution systems through detailed statistical comparison. DOI: 10.1061/(ASCE)
WR.1943-5452.0001150. © 2019 American Society of Civil Engineers.

Introduction as the weighted shortest path between two sampled nodes. This
approach ensures that the pressure loggers are located at nodes that
Water distribution systems (WDSs) are essential infrastructures are (a) sensitive and (b) “far” from each other, i.e., widespread over
providing sufficient quantity and quality of drinking water for both the network.
civilian and industrial customers. The complexity of these systems The calibration process adjusts the friction parameter of each
increases quickly with the number of demand (consumption) points pipe individually. The advantage of this approach is that no addi-
(that is, the size of the area served), resulting in the challenging task tional data (e.g., age or pipe material) is needed, which is usually
of instrumentation and supervision of such large systems. One of used for the grouping of the pipes and might be inaccurate or
the emerging challenges is that because the number of measure- even missing. However, the resulting problem is underdetermined,
ment devices—pressure loggers, flow meters—is far smaller than i.e., the number of parameters to be adjusted is (far) greater than
the number of potential measurement points (typically fire hydrants the available measurement data. Here, the use of singular value de-
for pressure), one must optimize (in some sense) the layout of the composition (SVD) is proposed to solve such problems, which re-
measurements. Another issue with the increasing complexity is that sults in a parameter vector closest to the initial guess (least-norm
the number of parameters to be calibrated in a WDS computer solution).
model escalates with system size and the calibration process re- This paper is organized as follows. The next section includes a
quires both high computational effort and highly sophisticated literature overview in which, even though the focus is on friction
algorithms. parameter calibration, a brief overview of demand calibration and
The focus of the present study is on hydraulic systems with sampling approaches is also provided. Then the mathematical tools
significant frictional pressure drop, and the aim is to calibrate that are used, such as the mathematical background of hydraulic
the pipe roughness values in such a way that the model output simulation, sensitivity analysis, SVD, and distance metrics on
(nodal pressures) is as close to the “real values” as possible. Such graphs, are summarized. Then the main results, i.e., a deterministic,
calibration requires two steps: (1) designating the sampling points, iterative calibration and a direct, iteration-free sensor placement
i.e., the nodes at which pressure values are measured and (2) a technique, are described. The next section includes a comparison
calibration algorithm that minimizes the discrepancy between of the proposed layout design with five standard approaches from
the measured and the computed pressure values by adjusting the the literature on six different networks of increasing size and com-
pipe friction coefficients. The novelties of the proposed techniques plexity, including the the Anytown, C-town, and KY2 networks, in
are as follows. addition to a small artificial network and two real-life networks. In
The sampling design is based on a quantity combining the sen- the case of the simple systems, a full enumeration of the possible
sitivity and the “hydraulic distance” of the sampling points, defined sampling layouts (that is, for a given number of sampling points, all
possible combinations were evaluated) is performed, while for the
1 larger systems a MATLAB version R2018a built-in optimization
Ph.D. Student, Dept. of Hydrodynamic Systems, Faculty of Mechan-
ical Engineering, Budapest Univ. of Technology and Economics, Budapest package is used. For each case, 100 calibration processes were
1111, Hungary (corresponding author). ORCID: https://orcid.org/0000 run with random initial parameter sets, and a detailed statistical
-0002-2556-3841. Email: rweber@hds.bme.hu comparison is provided. The final section concludes the study.
2
Associate Professor, Dept. of Hydrodynamic Systems, Faculty of
Mechanical Engineering, Budapest Univ. of Technology and Economics,
Budapest 1111, Hungary. ORCID: https://orcid.org/0000-0002-1930-515X.
Email: cshos@hds.bme.hu Literature Review
Note. This manuscript was submitted on November 8, 2018; approved
on May 22, 2019; published online on November 12, 2019. Discussion Hydraulic models of WDSs consist of a large number of nodal mass
period open until April 12, 2020; separate discussions must be submitted balances and pressure difference–flow rate relationships, with the
for individual papers. This paper is part of the Journal of Water Resources former being linear and the latter being nonlinear, due to the inher-
Planning and Management, © ASCE, ISSN 0733-9496. ently nonlinear dependence of frictional pressure losses on flow

© ASCE 04019070-1 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


rates and pump head curves. Hence, when calibrating a hydraulic Gao (2017) defined a multistage calibration process using a
model, one must cope with a (possibly large number of) nonlinear weighted least-squares method. Zhang et al. (2018) also introduced
algebraic equations. One of the first attempts to support the manual a numerical technique that is capable of adjusting nodal demands
calibration of parameters (namely, roughness coefficients and nodal and pipe roughness coefficients. Overall, the mentioned methods
demands) in a nonlinear mathematical model that could describe follow the widespread idea of minimizing (using a sophisticated
a hydraulic system was reported in Walski (1983) based on fire algorithm) the difference between the measurements and the out-
hydrant tests on low and high flow rates. Up to the present day, come of the hydraulic model.
the main idea behind the calibration process has been to define an Finding the optimal measurement layout was approached from
objective scalar function that indicates the difference between mea- several directions to make the hydraulic model calibration more
sured values and the output of the simulation and then to use an accurate. Yu and Powell (1994) formulated this problem in a multi-
optimization method (iterative or heuristic) to find the parameter objective way in which the accuracy was maximized, and the total
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

set that provides the lowest value of the objective function. The metering cost was minimized using decision trees. Bush and Uber
first algorithmic method can be found in Ormsbee (1989), where (1998) introduced an idea in which the propagated measurement
nonlinear optimization was used. error should be minimized. This theory was also applied to build
Studies have been published on the calibration of the roughness a three-step method in Lansey et al. (2001), and further develop-
coefficient itself without the adjustment of consumption values. ment was carried out where the uncertainty of the model output was
Greco (1999) used the sensitivity matrix with linear and nonlinear considered. Kapelan et al. (2003) used this theory by taking into
programming to increase the accuracy of the model. To avoid the consideration the total sampling design cost and solved the lem
(potentially inaccurate) computation of the derivatives, Kapelan as multiobjective optimization. Single and multiobjective genetic
et al. (2007) presented a method called shuffled complex evolution algorithms were used in Kapelan et al. (2005), where the latter
metropolis that could also calculate uncertainties without the wide- was capable of determining a whole set of (Pareto) optimal solu-
spread first-order second method (FOSM). Koppel and Vassiljev tions. In Behzadian et al. (2009) the authors combined a multiob-
(2009) presented a technique in which the number of parameters jective genetic algorithm with an adaptive neural network to
was decreased by grouping the pipes based on their age, and then minimize uncertainties. However, completely different approaches
they used the Levenberg–Marquardt algorithm for optimization. were presented in de Schaetzen et al. (2000), where the methods
Khedr et al. (2015) showed that a sophisticated optimization algo- were based on the shortest path algorithm, sensitivity, and Shannon
rithm could increase the accuracy of the calibration compared to entropy. Morosini et al. (2014) performed sensitivity analysis to
the case of adjusting the roughness coefficients manually. Simi-
prioritize the influences of nodes and pipes, but it was not empha-
larly, Meirelles et al. (2017) also compared the standard calibration
sized that the most sensitive nodes and pipes were often in the
process to optimization using particle swarm optimization and
vicinity of each other or even adjacent ones. Moreover, if one
concluded that the latter results in higher accuracy. Todini and
wishes to place several sensors incrementally (which is typically
Pilati (1987) introduced the Global Gradient Algorithm for solving
the case), it is not clear how to locate the subsequent sampling
the basic hydraulic equations, which was further developed for
node. Klapcsik et al. (2017) demonstrated the performance of sev-
calibration in Giustolisi and Berardi (2011). This algorithm also
eral of the aforementioned techniques plus a novel one based on
inspired a direct inversion algorithm [published in Du et al. (2018)]
graph theory using real-life water distribution networks.
combining all unknowns in a system into a single global vector,
Another typical problem of optimal sampling involves early
including unmeasured pipe flows, nodal pressures, and rough-
warning systems for detecting contamination, accidental or inten-
ness coefficients, and then solving them simultaneously using a
Newton-based linear method. Ostfeld et al. (2012) summarize the tional. Numerous studies were carried out to address the issue of
competition known as the Battle of Water Calibration Networks how to minimize the effects of pollution by locating an optimal set
whose purpose was an objective comparison of different calibration of sensor layout. One of the first attempts was made by Lee and
techniques using the C-town network. Deininger (1992) based on maximizing the “covered nodes” in
A different approach to increasing the accuracy of the hydraulic the network using integer linear programming. The origin tracking
model is to calibrate nodal demands or demand patterns. For this algorithm was presented in Laird et al. (2005); it is capable of de-
purpose Kang and Lansey (2009) used a weighted least-squares termining contamination spread in a system without the need for
method that was also capable of predicting the chlorine concentra- discretizing pipelines. Later, Berry et al. (2006) introduced the min-
tion in the examined system. The SVD theory was applied to imization of the expected impact calculating numerous different
determine the search direction where the golden section search scenarios and optimized with mixed integer programming and heu-
was used to obtain the optimal nodal demands in Cheng and He ristic method. A competition, called the Battle of the Water Sensor
(2011). Kun et al. (2015) presented an efficient inversion model Networks, was organized in 2006 and its results were published
that avoided the repeated computation of a Jacobian matrix. Do later in Ostfeld et al. (2008). The main conclusions were that plac-
et al. (2016) developed a genetic algorithm for the calibration of ing sensors at water sources (e.g., pump stations, tanks) is not nec-
the demand multipliers where the number of measurements is lower essary and the clustering of sensors is inefficient. Xu et al. (2008)
than the number of unknowns, i.e., it is underdetermined. SVD was presented a method based on graph theory where no additional data
also applied in Sanz and Pérez (2015) to reduce the parameters by apart from the topology of the network were needed, which seems
grouping nodal demands into several patterns. to be uneconomical given that hydraulic data are already available
Another frequent idea of calibration is to take into consideration but not utilized. A technique can be found in Krause et al. (2008)
both the demands (or demand patterns) and roughness coefficients. that is capable of handling huge networks (21,000 nodes) while
A two-step weighted least-squares method was introduced for maintaining rigorous approximation and low CPU running time.
the adjustment of parameters in Kang and Lansey (2011). Dini Hart and Murray (2010) published a comprehensive review on
and Tabesh (2014) used the ant colony optimization method for the topic, also suggesting directions for further research, e.g., effi-
simultaneous calibration. In Xie et al. (2017) a Bayesian-based, ciently solving large-scale problems or comparing new methods
two-level Markov chain Monte Carlo–Particle Filter method was to existing ones. Later, Zheng et al. (2018) published a paper com-
used that can also handle the uncertainties of calibrated values. paring different sensor placement strategies, using the quantities

© ASCE 04019070-2 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


presented in Ostfeld et al. (2008), but considering expected and Mathematical Tools
worst-case scenarios.
To sum up the relevant literature, calibration and sensor place- Hydraulic Simulation
ment problems are usually solved by optimization techniques
requiring significant computational effort and sophisticated solvers. In this section, the fundamental equations of one-dimensional hy-
Table 1 contains all the important properties of calibration ap- draulic modeling and the solution strategy are introduced. Even
proaches found in the literature. As that literature shows, often sev- though an in-house C++ software (STACI) was used for simula-
eral parameter sets (e.g., pipe roughness and demand) are calibrated tions, the underlying equations are the same as those implemented
at once. Also, the calibration problem is usually transformed into an in EPANET2 (Rossman 2000). The mass conservation law is
overdetermined problem by grouping the parameters in such a way solved for every node
that the final number of aggregated parameters is smaller than the X X
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

number of available measurements—resulting in a least-squares Qi − Qi ¼ d ð1Þ


solution. Moreover, it seems that nowadays, heuristic techniques i∈in i∈out
are more common than iterative methods. Finally, concepts from
graph theory (e.g., distance, segmentation) seem to have begun where it is assumed that the fluid is incompressible (i.e., liquid).
to infiltrate WDS analysis. Q denotes the volume flow rate, the term with “in” indicates in-
The primary motivation of this work is to create an objec- flows, while that with “out” indicates outflows, and d is the nodal
tive method for comparing different sensor placement strategies demand (consumption). The other set of equations describes the
(the ones available in the literature and a novel one presented energy balance through the edges (e.g., pipe, pump, valve) and
in this paper) in terms of the “goodness” of the pipe roughness is formally given by
calibration and in terms of computation times. To keep the anal-
ysis focused, other (possibly also important) effects are ignored, ps − pe ¼ fðQÞ ð2Þ
e.g., unmodeled leaks or the uncertainty of the nominal consump-
tion values, even though the extension of the proposed methods to where ps and pe = pressure at two ends of edge; and f function =
these cases is straightforward. A novel approach is presented that nonlinear function of volume flow rate. As an example, for a pipe
calibrates all pipe roughness values independently (i.e., without connecting two nodes with geodetic heights hs and he , one has
grouping) and a sensor placement strategy that is highly efficient.
Then, these novel techniques and those available in the literature L ρ QjQj
ps − pe ¼ ρgðhe − hs Þ þ λ ð3Þ
are compared in an objective way, using several networks, from D 2 A2
simple artificial ones to real-life cases. The comparison is based on
a statistical evaluation of the “goodness” of the calibration (yet to where λ = friction coefficient; ρ = density of water; g = gravita-
be defined precisely) and the required CPU time. Even though tional acceleration; and A, D, and L = area of cross section, pipe
roughness calibration problems are typically solved offline, the au- diameter, and length, respectively. To estimate the actual value of
thors think that the required computational effort (CPU time) is still the friction coefficient λ, one can use e.g., the Darcy–Weisbach
relevant because the size of the hydraulic models tends to increase (where λ is a function of the Reynolds number and the relative wall
quickly nowadays (e.g., Gao 2014). It is not uncommon to use roughness) or the Hazen–Williams equations (where the C factor
models with 10,000þ pipe segments; thus, the use of efficient defines λ).
numerical techniques can significantly increase the potential for Combining (and rearranging) all (1) mass and (2) energy equa-
even larger hydraulic models in the future. tions for a WDS results in N nodes þ N edges equations, which are to

Table 1. Basic properties of calibration methods


Objectivity Optimization Parameter
Reference Single Multi Direct Heuristic Roughness Demand Grouping
Walski (1983) X — Manual X X X
Ormsbee (1989) X — X — X X X
Greco (1999) X — X X — —
Kapelan et al. (2007) X — — X X — X
Kang and Lansey (2009) X — X — — X X
Koppel and Vassiljev (2009) X — X — X — X
Cheng and He (2011) X — X — — X —
Giustolisi and Berardi (2011) — X X X — X
Kang and Lansey (2011) X — X — X X X
Dini and Tabesh (2014) X — — X X X X
Kun et al. (2015) X — X — — X X
Khedr et al. (2015) — X — X X — X
Sanz and Pérez (2015) X — X — — X X
Do et al. (2016) X — — X — X —
Gao (2017) X — X X X X
Meirelles et al. (2017) X — — X X X X
Xie et al. (2017) X — — X X X X
Du et al. (2018) X — X — X — X
Zhang et al. (2018) X — X — X X X
Proposed method X — X — X — —

© ASCE 04019070-3 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


be solved for nodal pressures and edge flow rates. Formally, one (that is, closest to the initial guess of the parameters) is of greatest
can write interest. Such systems (and overdetermined or even mixed systems)
are easily solved by SVD. SVD decomposes a matrix A ∈ Rn×m
FðxÞ ¼ 0 ð4Þ with a rank r into the product of three special matrices
as A ¼ UΣV T , where U ∈ Rn×n and V ∈ Rm×m are column-
where x ¼ ½p; QT ¼ ½p1 ; p2 ; : : : ; pN nodes ; Q1 ; Q2 ; : : : ; QN edges T con-
orthogonal matrices (i.e., UUT ¼ VV T ¼ I, where I is the identity
tains the unknown hydraulic variables. The preceding system is matrix), and Σ ∈ Rn×m is a diagonal matrix. U and V T contain the
solved iteratively via Newton’s technique, that is left and right singular vectors (eigenvectors of AAT and AT A, re-
Jðxn Þðxnþ1 − xn Þ ¼ −Fðxn Þ ð5Þ spectively, if m ¼ n) and Σ the singular values (square root of ei-
genvalues of AAT , if m ¼ n). There are no restrictions on Matrix A
which is a linear set of equations, with J being the Jacobian, that is, i.e., it can even be singular. If an underdetermined system Ay ¼ b
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

J i;j ¼ ∂Fi =∂xj . Note that this is a sparse system with mostly two is solved by SVD, the solution vector will have the smallest norm of
or three nonzero elements per row, so the package UMFPACK all possible solutions, that is, kyk → Min. However, the SVD is
(Davis 2004) is used to solve it efficiently and avoid an expensive also computationally favorable since the overall cost is Oðmn2 Þ,
inversion of the Jacobian. where n ≪ m, meaning, in the present case, it will grow linearly
with the size of the network.
Another useful feature of this technique is that it can determine
Pressure Sensitivity Analysis
the rank of a matrix, i.e., the number of linearly independent rows
It is useful to compute the variation of the solution (that is, pressure or columns. In theory, the singular values are always positive, and
or flow rate values) with respect to (wrt) a parameter, for example, the number of nonzero singular values gives the rank of the matrix.
the pipe friction coefficient λ. That is, the variation of the pressure However, in reality, in numerical cases the out-of-rank singular val-
at node i wrt to a small change in the friction coefficient of ues are never precisely zero (but, instead, some small finite num-
pipe j is ber). Therefore, a widely used approach is that 90% of the overall
energy (information) should be retrained. This means that once Σ
∂pi
Si;j ¼ ð6Þ (which was a diagonal matrix including the singular values) has
∂λj been computed, one deletes the smallest singular values in such
a way that the sum of the squares of the remaining (nondeleted)
To compute this quantity, Eq. (4) is differentiated wrt to the singular values should provide at least 90% of the sum of the
parameters [and keeping in mind that if the parameter is varied, squares of all singular values (Leskovec 2014).
so is the solution, i.e., pðλÞ], giving After determining the SVD of a matrix, its inverse (or pseudoin-
∂F ∂p ∂F verse) can also be easily found by Aþ ¼ VΣþ UT , where Σþ is a
þ ¼0 ð7Þ diagonal matrix with the reciprocal of the retained (nondeleted) sin-
∂p ∂λi ∂λi
gular values. Finally, the method is capable of evaluating the de-
where the first term (the Jacobian) was already computed during the terminant of a singular, nonsquared matrix; the determinant is
Newton iteration, the second term is one column from the sensi- simply the product of the retained singular values. For additional
tivity matrix S pursued to compute, and the third term is easily features on SVD, the reader may consult Leskovec (2014). These
evaluated from either Eq. (1) or (2). Eq. (7) is again a linear set computations were performed using the C++ library Eigen (Gael
of algebraic equations, which has to be solved as many times as and Benoi 2010), which ensures computational efficiency.
there are parameters. However, because the linear coefficient matrix
(the Jacobian) is constant, the use of lower–upper (LU) decompo- Graph Theory: Distance
sition makes this process efficient and inexpensive. It should be
noted that the same process could be employed to compute the sen- The final mathematical tool that is needed is borrowed from graph
sitivity of any output variable wrt to any parameter (for example, theory. A water distribution system can be interpreted as a graph
the sensitivity of the pipe flow rates wrt to demand); however, this where the pipelines, valves, pumps, and so forth are the edges
study focuses on the sensitivity of nodal pressures to the pipe fric- (links), and their intersections (nodes) are the vertices. In graph
tion parameters. theory, a path is defined as a route through links between two nodes
The roughness coefficients of every pipeline affect the pressure (possibly far away from each other) (e.g., Barabási 2016). Clearly,
at each node; the local or nodal sensitivity can be defined by taking in a looped network there could be several different paths between
the absolute value of every element and summing the rows of the two particular nodes and, in the simplest case, the shortest path or
matrix: distance is the one with the minimum number of edges. For exam-
ple, in Fig. 1 the shortest path between Nodes 1 and 4 is three (if
X
N edges just the number of edges defines the length of the path), as high-
Snodal;i ¼ jSfi;jg j ð8Þ lighted by the solid thick line. However, the distance or length can
j¼1 also be defined in the case of a weighted graph (Barabási 2016). In
this case, the shortest path is the one along which the sum of the
This equation forecasts the increase (or decrease) of a nodal weights is minimal. In this sense, in Fig. 1, the distance between
pressure in cases where the roughness of every pipe is (equally) Nodes 1 and 4 is 10 þ 10 þ 5 þ 10 ¼ 35 (dashed line). For com-
increased. puting this quantity, the C++ library igraph was used (Csardi and
Nepusz 2012).
Singular Value Decomposition
The calibration technique introduced in the next section results in a Proposed Calibration and Sensor Placement Method
nonsquare linear system; that is, there are fewer equations than un-
knowns. Such a linear system is underdetermined, and there are In the following sections, the model calibration and the sensor
infinitely many solutions of which the one with the smallest norm placement problem both focus on the accuracy of the model output,

© ASCE 04019070-4 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


For now, it is assumed that the measurement was performed
at nominal demand values.
5. Solve the linear system Sr Δλ ¼ Δp using the previously pre-
sented SVD method for an update of the roughness coefficient
distribution.
6. Update the parameter vector and repeat from Point 2 until Δp is
sufficiently small. To determine the end of the cycle, an absolute
threshold can be prescribed for Δp, e.g., 10−3 m, or the subse-
quent values of Δλ can be monitored, and a threshold can be
defined for the relative difference, e.g., 0.1%.
Mathematically speaking, the solution of the linear equation
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

system at Point 5 is far from trivial. On the one hand, the linear
system is underdetermined, i.e., the number of equations (number
of sampling points) is significantly lower than the number of var-
iables (number of pipes). On the other hand, it is possible that the
Fig. 1. Visualization of definition of distance. rows (the measurements) are not linearly independent, resulting in
almost identical rows in the sensitivity matrix. This can easily hap-
pen if the sampling points are close to each other (e.g., they are two
i.e., the purpose is to minimize the discrepancy between the “real” nodes of the same, short pipe). As presented earlier, both issues can
pressure distribution and the output of the calibrated model by ad- be addressed if SVD is used to compute the pseudoinverse of the
justing the roughness coefficients of the model. Note that the cal- reduced sensitivity matrix (Sþ r ). Because the system is underdeter-
ibration and sampling techniques are independent of each other. mined, there are usually multiple possible solutions for the prob-
First, the proposed calibration technique is explained using prede- lem, but the SVD finds the one with the least norm, that is, this
fined sampling locations, then an efficient sensor placing strategy is iteration gives the closest set of parameters to the initial estimation
described. at the first step and the closest to the previous one in the following
steps; i.e., it is a step-by-step least-norm solution. To speed up
the convergence of the subsequent parameter improvement steps,
Calibration Technique it is useful to apply relaxation, i.e., Δλ ¼ ωSþ r Δp, where
0.1 ≤ ω ≤ 0.8.
There are two important features of the proposed calibration tech-
In this paper, the method is used for only one hydraulic state
nique. First, it assigns an independent roughness coefficient to each
with nominal values, but the technique itself can be used straight-
pipeline i.e., here they are not grouped as in, e.g., Mallick et al.
forwardly for multiple operational points. Assume that there are
(2002). The advantage of grouping similar pipes (based on material
measurements available with N op different operational loads. In this
or age, for example), and then assigning a single friction coefficient
case, in Step 2, N op hydraulic simulation needs to be performed and
to all pipes within a group reduces the problem to a large extent, but
it might require more input data that might not be available or N op sensitivity matrices calculated ; constructing the reduced ma-
be uncertain. Second, it is iterative; thus, no heuristic optimization trix is similar, and it has N op × N sampled rows. The rest of the
(e.g., genetic algorithm, simulated annealing) is required. This method is unchanged.
method could use any friction model, e.g., Darcy-Weisbach, Another property of the method is that in the case of fully in-
Hazen-Williams: in the first case, the parameter to be adjusted dependent measurements (mathematically speaking, if the rank of
is the relative roughness, while in the latter case, it is the C-factor. the reduced sensitivity matrix is not lower than its smaller size), the
For the time being it is assumed that the sampled nodes have al- technique is capable of removing the difference fully between the
ready been chosen (the sensor placement techniques will be dis- measurements and the model output, i.e., Δp ¼ 0, see Eq. (9).
cussed subsequently). The calibration technique consists of the However, if there are linearly dependent measurements, the result-
following steps. ing linear system Sr Δλ ¼ Δp is ill-conditioned because this means
1. The sampling nodes, the measured values at these nodes, and the that the left-hand side of the linear equations are almost identical,
estimated parameter vector are input values. If additional data while the corresponding right-hand sides are different. This can
(e.g., age or material) are available for the pipelines, they should easily happen, e.g., if two sampled nodes are too close to each other
be used to increase the accuracy of the calibration. If such data and there is uncertainty in the consumption values or in the pressure
are unavailable, some “default” value needs to be assigned to measurement. In such cases, the proposed calibration technique can
each pipe. still be utilized and will converge robustly. However, the difference
2. Run a standard hydraulic simulation with the estimated rough- between the measurement and the model output would not
ness values and with the nominal (average) demand values. vanish, i.e., Δp ≠ 0.
Then calculate the sensitivity matrix.
3. Extract the reduced sensitivity matrix Sr including only the sen- Sensor Placement Based on Hydraulic Distance with
sitivity of the sampled pressures, i.e., delete the rows corre- Sensitivity
sponding to unsampled nodes. The original sensitivity matrix
is of size N nodes × N edges and, in the case of N sampled number Intuitively, a “good” sensor placement fulfills two requirements.
of pressure transducers, the size of the reduced sensitivity matrix First, sensitive nodes are used for sampling. As a counterexample,
Sr is N sampled × N edges . it is inefficient to deploy pressure loggers close to reservoirs, where
4. Compute the pressure difference between the hydraulic model the pressure is primarily defined by the reservoir water level. How-
output (obtained with the estimated parameters) and the mea- ever, sensitive nodes tend to accumulate at certain locations within
surements (“real” values): a WDS, so simply choosing the most sensitive nodes often results
in the accumulation of sampling points close to each other (de
Δp ¼ pmod − pmeas ð9Þ Schaetzen et al. 2000; Klapcsik et al. 2017). The problem with

© ASCE 04019070-5 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


Table 2. Basic data of networks This leads us to the concept of hydraulic distance, which is the
No. of Overall pipe Total distance of the actual node from the nodes of prescribed pressures
Name pipes length (m) consumption (m3 =h) (e.g., reservoirs, tanks) or sampled pressures, measured by weight-
ing the edges by frictional pressure loss, i.e., the pressure difference
4 × 4 Grid 24 2,400 2,610
Anytown 35 69,590 1,454 without the geodetic head difference. (For the mathematical defi-
C-town 452 56,723 980.7 nition of distance, see the section “Graph Theory—Distance.”)
Sopron Network I (SN-I) 287 7,673 2.11 Reference is made to the nodes from which distance is measured
Sopron Network II (SN-II) 920 13,084 20.1 as “master nodes.” The hydraulic distance is defined as the mini-
KY2 1,124 152,248 329 mum value of the distances between the actual node and the
master nodes.
A node can then be characterized by the product of its nodal
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

sensitivity and the hydraulic distance. Note that if a sensitive node


close or neighboring samplings is simply that the measurements are is designated for measuring, the nodes surrounding it (probably
not independent of each other. Thus, the second requirement is that also with high sensitivity) become of low importance as they
the measurement points should be “far enough” from each other to are close to a master node. Similarly, a node far from the master
be “as independent as possible.” nodes but with low sensitivity is not suitable for calibration. This

Fig. 2. Layout of networks. Anytown, Grid, C-town, and KY2 are artificial networks, while the other two are different zones of Sopron (Hungarian
city). For basic data of the water distribution networks, see Table 2.

© ASCE 04019070-6 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


technique is sequential in the sense that it adds sampling points one two methods are based on the minimization of the propagation of
by one; after a node is designated for sampling, only the distances measurement uncertainty to the output of the calibrated model, as
must be recomputed. Nevertheless, the strategy is straightforward explained, e.g., in Kapelan et al. (2003), by minimizing the trace
and free of iteration or heuristic optimization. (A-optimality) or the determinant (D-optimality) of the covariance
The HDS sensor placement method consists of the follow- matrix, that is
ing steps.
1. Estimate the roughness values, if possible using the available C ¼ Covðpmod Þ ¼ SSþ þ T T
r Covðpmeas ÞðSr Þ S ð11Þ
data (e.g., pipe material, age) to improve accuracy.
and then minimize the trace or the determinant
2. Calculate the sensitivity matrix and the nodal sensitivities.
3. For each node, determine the distance from the master nodes NX
1 nodes
and the product of the distance and nodal sensitivity. FCov−A ¼ C1=2
i;i → Min! ð12Þ
N nodes
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

4. Find the maximum value and designate the corresponding node i¼1
as a sampled node, i.e., add it to the master nodes.
5. Repeat from Step 3 until a sufficient number of nodes has been FCov−D ¼ ðdetðCÞÞð1=2NÞ → Min! ð13Þ
selected.
Note that the computation of the hydraulics and the sensitivity is A similar method, based on a sensitivity matrix, can be found in
required only once, and only the distances are recalculated at each de Schaetzen et al. (2000). The idea is to maximize the sensitivity
step. This ensures a low computational time, even in the case of that the measured nodes cover, mathematically
large networks.
X
N edges
FSmax ¼ aj ; aj ¼ maxðSi;j Þ → Max! ð14Þ
i
j¼1
Comparison of Measurement Layouts
which means the sum of the columnwise maximum values, and this
This section presents a detailed comparison of different sensor basically maximizes the sensitivity of the measured nodes. In de
placement strategies, including the proposed technique and other Schaetzen et al. (2000) the entropy of the sensitivity is also used:
approaches available in the literature. The previously presented
calibration technique is used for every method during the compari- X
N edges
aj
son to ensure low CPU time. This ensures that once the sampling FShannon ¼ − bj lnðbj Þ; bj ¼ P → Max! ð15Þ
nodes are designated, the calibration process will be the same for all j¼1 j aj

methods. The analysis that follows uses six WDSs with increasing
complexity, whose main parameters are listed in Table 2 and the While in de Schaetzen et al. (2000) one objective function
layouts of the networks are depicted in Fig. 2. The steps of the was defined by taking the weighted sum of the two method, the
analysis are as follows. present analysis deals with these theories separately, similarly to
• Fix the number of pressure measurement points: N sampled . Klapcsik et al. (2017).
• Perform a hydraulic analysis with nominal friction values and Finally, in de Schaetzen et al. (2000) graph theory is used
average demands. Save all nodal pressure data. This simulates (similar to the proposed technique); the weights in the definition
the measurement, and the corresponding pressure values are de- of distance are the lengths of the pipelines. With this approach,
noted by pmeas. the next sensors are located as far away from each other and the
• By means of a sampling layout design technique, choose sources as possible (in a geographical sense). This approach adds
N sampled sampling nodes. the sensors one by one. Hence, there is no need for optimization as
• Perturb all nominal friction parameter values by a random in the case of the first four approaches. The mentioned techniques
number with normal distribution (0 expected value and 20% are summarized in Table 3.
standard deviation). Perform calibration using the perturbed
frictions as initial values and evaluate the error, that is, the norm
Test Networks
of the difference between the nominal and calibrated model
pressure at all nodes, mathematically In this section, six networks are used for comparison. The first net-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi work is entirely artificial, and it is a regular 4 × 4 grid network. The
PN nodes 2

i¼1 ðpmod;i − pmeas;i Þ
next two systems are well known from the literature: the Anytown
ep ¼ ð10Þ and C-town networks. While the fourth one, KY2, is still artificial,
N nodes

• Repeat this process 100 times, then compute the average and
standard deviation of the ep errors. For a given number of sam- Table 3. Sensor placement methods for roughness calibration in the
pling nodes, one layout is superior to another one if the average literature
of ep is smaller. It is expected that the calibration error will van- Name Short notation Optimization Reference
ish if all nodes are sampled.
Covariance Cov-A/Cov-D Yes Lansey
Note that in a real-life case, ep is not available since only the
(A-optimal/D-optimal) et al. (2001)
sampled pressures are known. However, the authors think that this Sensitivity Smax Yes de Schaetzen
quantity gives an objective way of comparing different sampling maximization et al. (2000)
strategies, at least for artificial test cases. Shannon entropy Shannon Yes de Schaetzen
et al. (2000)
Shortest path Spath No de Schaetzen
Standard Sampling Methods et al. (2000)
The standard sampling methods found in the literature are briefly Hydraulic distance HDS No Presented
with sensitivity in this paper
described; for a complete description, see the cited papers. The first

© ASCE 04019070-7 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


Grid network Anytown
Cov-A Cov-A
1 Cov-D 1 Cov-D
Smax Smax
Shannon Shannon
Spath Spath
0.8 HDS 0.8 HDS
ep / max(ep ) [-]

ep / max(ep ) [-]
0.6 0.6

0.4 0.4
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

0.2
0.2

0
0 0 5 10 15 20
0 2 4 6 8 10 12 14 16
no. of sensors
no. of sensors
Fig. 4. Average error of calibration (with standard deviation of average
Fig. 3. Average error of calibration (with standard deviation of average
as error bars) as function of number of sensors for Anytown network.
as error bars) as function of number of sensors for 4 × 4 Grid network.

the final two networks are real-life WDSs near Sopron, Hungary. determining the objective function for every possible scenario (di-
The network layouts can be found in Fig. 2 while some of the es- rect enumeration) was impossible (the number of potential layouts,
sential characteristic data are provided in Table 2. e.g., for five sensors only, is approximately 8.2 × 1010 ). Thus,
As mentioned, the first system for the comparison is a simple, MATLAB’s built-in discrete genetic algorithm was used to find
4 × 4 grid network, for which all parameters (e.g., pipe length, no- a solution close to optimal for all techniques except the Spath
dal demand) are identical. The pressure source is located at the top and HDS methods, which do not require any optimization. The re-
left corner and a prescribed amount of water is fed into the system sults for C-town can be seen in Fig. 5, where Smax outperforms the
at the bottom right corner, which is equal to the total nominal con- other techniques, HDS performs second best, and Spath is slightly
sumption, i.e., there is no water entering at the prescribed pressure behind them. In the case of KY2 (Fig. 6) four techniques converge
point. The second system is the Anytown network described in similarly, but HDS is slightly ahead of Smax. In Fig. 9 the optimal
Walski et al. (1987). In the case of these networks (Grid and Any- measurement layout depicts the cases corresponding to sampling
town), due to the small number of nodes, it was possible to perform with less than 20% error. In the case of C-town, the Smax method
full enumeration, that is, for a given sensor number, evaluate all requires 16 sensors, while HDS requires 29, but the other tech-
possible combinations of the sampling nodes and compare the niques could not achieve this goal within 30. On the other hand,
objective function (FCov−A , FCov−D , FSmax , FShannon ) results and in the case of the KY2 network, HDS needs 15 sensors. Again,
choose the optimal layout (minimum or maximum) for each the sensors are well spread out across the network, and they are
method. mostly found close to the end of the networks or at pressure zone
The results for the Grid network can be seen in Fig. 3 and those borders.
for the Anytown system in Fig. 4. The average pressure error de-
fined by Eq. (10) (including the pressure error of the unsampled
nodes) is depicted against the number of sensors. In addition, C-town
the standard deviation of the average, based on 100 calibrations,
Cov-A
is depicted using error bars. For the Grid network, the Spath 1 Cov-D
and Shannon techniques provide slow convergence, and the Smax
Shannon
Cov-A technique gives slightly (but not significantly) slower con- Spath
0.8 HDS
vergence than the rest of the techniques. In the case of the Anytown
system, Smax and the HDS techniques provide the fastest conver-
ep / max(ep ) [-]

gence. The number of sensors needed to decrease the error of the 0.6
initial “uncalibrated” error to its 20% is also highlighted. These
sensor placement layouts are depicted in Fig. 9. It is easy to see
that these “optimal” layouts are far from the pressure sources (res- 0.4
ervoirs), but they seem to be located close to each other. However,
these sampling points are far from each other in the “hydraulic”
sense in that they are separated by pipes with large frictional pres- 0.2
sure drop due to the high velocity.
The next networks in this comparison are the C-town (Ostfeld
0
et al. 2012) and KY2 (Hwang and Lansey 2017) (see Fig. 2 and 0 5 10 15 20 25 30
Table 2 for basic data). These are still hypothetical networks, no. of sensors
but they are already an order of magnitude larger than the previous
Fig. 5. Average error of calibration (with standard deviation of average
ones and contain several pumps, pressure-reducing valves, and so
as error bars) as function of number of sensors for C-town network.
forth. Since even the C-town network includes over 400 nodes,

© ASCE 04019070-8 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


KY2 Sopron Network II
Cov-A Cov-A
1 Cov-D 1 Cov-D
Smax Smax
Shannon Shannon
Spath Spath
0.8 HDS 0.8 HDS
ep / max(ep) [-]

ep / max(ep ) [-]
0.6 0.6

0.4 0.4
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

0.2 0.2

0 0
0 5 10 15 0 1 2 3 4 5 6 7 8
no. of sensors no. of sensors
Fig. 6. Average error of calibration (with standard deviation of average Fig. 8. Average error of calibration (with standard deviation of average
as error bars) as function of number of sensors for KY2 network. as error bars) as function of number of sensors for Network II from
Sopron Waterworks.

Real-Life WDSs
The real water distribution networks are derived from Sopron, a city are widely spread out and are mostly located around the edge of the
in Hungary where several tens of thousands of people live and the network.
examined systems are in different zones. Their technical data can be Finally, it is essential to highlight the CPU time that is required
found in Table 2, and the layouts are depicted in Fig. 2. Similarly to to find an optimal layout. Table 4 contains the runtime values in
the C-town case, for the evaluation of the optimization methods, seconds for both Sopron networks and C-town. It can be seen that
MATLAB’s built-in genetic algorithm function was used. The re- the covariance-based methods (Cov-A and Cov-D) require a mas-
sults can be seen in Figs. 7 and 8. At SN-I HDS, the Smax and sive amount of CPU time since they contain expensive matrix op-
Spath techniques are relatively close to each other, the latter slightly erations (e.g., pseudoinverse, determinant). In contrast, the HDS
outsmarts the others, while at SN-II Shannon is also among the best and Spath techniques are computationally inexpensive.
ones. It is interesting to note that these networks only require four
(SN-I) and two (SN-II) transducers to achieve a sufficiently cali- Conclusions
brated model. Moreover, the placement of additional sensors does
not increase the accuracy significantly. The optimal sensor place- This paper introduced an objective comparison of different sensor
ments are depicted in Fig. 9. Similarly to C-town, the transducers placement techniques for roughness calibration. A novel calibration
technique for hydraulic systems was introduced that is capable of
adjusting the roughness coefficients of each pipe individually with
low CPU time, creating the possibility of statistical comparison be-
Sopron Network I tween sampling strategies.
Cov-A The proposed calibration technique is based on the least-norm
1 Cov-D solution of underdetermined linear systems utilizing SVD and
Smax
Shannon pseudoinverse. This approach is deterministic in the sense that,
0.8
Spath even though the parameter vector is improved in several steps,
HDS
no heuristic optimization method is used. A novel iteration-free
pressure sensor placement strategy was also proposed based on
ep / max(ep ) [-]

0.6 the combination of nodal sensitivities and hydraulic distance, en-


suring that the sampling nodes are both sensitive to parameter
variation and are independent of each other, i.e., they are widely
0.4 spread over the WDS.
The proposed technique was compared to existing sensor layout
strategies utilizing statistical comparison on several WDSs of
0.2
varying size and complexity. The evaluation criterion was the dis-
crepancy between the initial “real” pressure distribution and the
0
calibrated ones at every node. For the two small networks (Grid
0 1 2 3 4 5 6 7 8 layout and Anytown) it was possible to perform full enumeration,
no. of sensors that is, the objective function was evaluated for every sensor place-
ment method (that required optimization) and for every possible
Fig. 7. Average error of calibration (with standard deviation of average
sensor layout, an optimal one was chosen for each case.
as error bars) as function of number of sensors for Network I from
It was found that the performance of the proposed technique
Sopron Waterworks.
is competitive against the existing approaches. For most hydraulic

© ASCE 04019070-9 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

Fig. 9. “Ideal” sensor placements for each network with minimum number of sensors required to decrease initial (uncalibrated) error to 20%.

systems, the Spath, Smax, and the presented HDS technique gives The main advantage of the proposed technique lies in its com-
the most accurate calibration for a given number of sensors. petitive performance and its computational efficiency. Apart from
However, the Spath technique struggled with the Grid layout the C-town problem, the HDS technique gave similar results to the
(consisting of many identical pipes). It was also interesting to Smax approach within a fraction of the CPU time (Table 4). The
see that the tree-like C-town system turned out to be a “challeng- authors speculate that this advantage becomes more pronounced
ing” problem for all approaches, clearly because of its lack of topo- with increases in network size. Furthermore, HDS is free from
logical connectivity. any unpredictability found in heuristic optimization techniques.
However, how uncertainties of other model parameters (e.g., de-
mand, pipe diameter) affected the presented HDS technique is a
Table 4. CPU time in seconds to calculate optimal measurement layouts question for further research.
for different methods Considering the “best layouts” for calibration (Fig. 9), it can be
concluded that the pressure transducers should be located far from
Network Cov-A Cov-D Smax Shannon Spath HDS
points of prescribed pressure; moreover, they were also found to be
C-town 6,750 3,200 40 20 4 <1 “hydraulically far away” from each other, which was observed
KY2 4,000 11,800 32 30 6 <1 regardless of the actual sensor deployment technique. However,
SN-I 400 100 10 10 3 <1 in the case of tree-like networks, it is important to cover as many
SN-II 8,750 2,066 20 14 10 <1
branches as possible.

© ASCE 04019070-10 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


Data Availability Statement Giustolisi, O., and L. Berardi. 2011. “Water distribution network calibration
using enhanced GGA and topological analysis.” J. Hydroinf. 13 (4):
The source code of the in-house hydraulic solver (STACI) can be 621. https://doi.org/10.2166/hydro.2010.088.
obtained from GitHub (https://github.com/hoscsaba/staci). More- Greco, M. 1999. “New approach to water distribution network calibration.”
over, it has a graphical interface that is also available (http://www J. Hydraul. Eng. 125 (8): 849–854. https://doi.org/10.1061/(ASCE)
.hds.bme.hu/staci_web/). This requires a user account that can be 0733-9429(1999)125:8(849).
requested from the authors. The WDNs presented in the paper, the Hart, W. E., and R. Murray. 2010. “Review of sensor placement strategies
additional C/C++ codes, and the MATLAB codes are also available for contamination warning systems in drinking water distribution sys-
tems.” J. Water Resour. Plann. Manage. 136 (6): 611–619. https://doi
from the authors by request. The figures and tables presented in the
.org/10.1061/(ASCE)WR.1943-5452.0000081.
article are available upon request from the authors.
Hwang, H., and K. Lansey. 2017. “Water distribution system classification
using system characteristics and graph-theory metrics.” J. Water
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

Resour. Plann. Manage. 143 (12): 04017071. https://doi.org/10.1061


Acknowledgments /(ASCE)WR.1943-5452.0000850.
Kang, D., and K. Lansey. 2009. “Real-time demand estimation and
The research reported in this paper was supported by the Higher confidence limit analysis for water distribution systems.” J. Hydraul.
Education Excellence Program of the Ministry of Human Capaci- Eng. 135 (10): 825–837. https://doi.org/10.1061/(ASCE)HY.1943
ties within the framework of the Water Science & Disaster Preven- -7900.0000086.
tion research program of the Budapest University of Technology Kang, D., and K. Lansey. 2011. “Demand and roughness estimation in
and Economics (BME FIKP-VÍZ). water distribution systems.” Water Resour. Plann. Manage. 137 (1):
20–30. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000086.
Kapelan, Z. S., D. A. Savic, and G. A. Walters. 2003. “Multiobjective sam-
References pling design for water distribution model calibration.” J. Water Resour.
Plann. Manage. 129 (6): 466–479. https://doi.org/10.1061/(ASCE)
Barabási, A. L. 2016. Network science. Cambridge, UK: Cambridge 0733-9496(2003)129:6(466).
University Press. Kapelan, Z. S., D. A. Savic, and G. A. Walters. 2005. “Optimal sampling
Behzadian, K., Z. Kapelan, D. Savic, and A. Ardeshir. 2009. “Stochastic design methodologies for water distribution model calibration.” J. Hy-
sampling design using a multi-objective genetic algorithm and adaptive draul. Eng. 131 (3): 190–200. https://doi.org/10.1061/(ASCE)0733
neural networks.” Environ. Modell. Software 24 (4): 530–541. https:// -9429(2005)131:3(190).
doi.org/10.1016/j.envsoft.2008.09.013. Kapelan, Z. S., D. A. Savic, and G. A. Walters. 2007. “Calibration of water
Berry, J. W., W. E. Hart, C. A. Phillips, J. G. Uber, and J.-P. Watson. 2006. distribution hydraulic models using a Bayesian-type procedure.” J. Hy-
“Sensor placement in municipal water networks with temporal integer draul. Eng. 133 (8): 927–936. https://doi.org/10.1061/(ASCE)0733
programming models.” J. Water Resour. Plann. Manage. 132 (4): 218– -9429(2007)133:8(927).
224. https://doi.org/10.1061/(ASCE)0733-9496(2006)132:4(218). Khedr, A., B. Tolson, and S. Ziemann. 2015. “Water distribution system
Bush, C. A., and J. G. Uber. 1998. “Sampling design methods for water calibration: Manual versus optimization-based approach.” Procedia
distribution model calibration.” J. Water Resour. Plann. Manage. Eng. 119 (1): 725–733. https://doi.org/10.1016/j.proeng.2015.08.926.
124 (6): 334–344. https://doi.org/10.1061/(ASCE)0733-9496(1998) Klapcsik, K., R. Varga, and C. Hős. 2017. “Optimal pressure measurement
124:6(334). layout design in water distribution network systems.” Periodica Poly-
Cheng, W., and Z. He. 2011. “Calibration of nodal demand in water technica Mech. Eng. 62 (1): 51–64. https://doi.org/10.3311/PPme
distribution systems.” J. Water Resour. Plann. Manage. 137 (1): 31–40. .11409.
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000093. Koppel, T., and A. Vassiljev. 2009. “Calibration of a model of an opera-
Csardi, G., and T. Nepusz. 2012. Igraph reference manual, 812. tional water distribution system containing pipes of different age.” Adv.
Davis, T. A. 2004. “Algorithm 832: UMFPACK V4.3, an unsymmetric- Eng. Software 40 (8): 659–664. https://doi.org/10.1016/j.advengsoft
pattern multifrontal method.” ACM Trans. Math. Software 30 (2): 1–4. .2008.11.015.
de Schaetzen, W., G. Walters, and D. Savic. 2000. “Optimal sampling Krause, A., J. Leskovec, C. Guestrin, J. VanBriesen, and C. Faloutsos.
design for model calibration using shortest path, genetic and entropy 2008. “Efficient sensor placement optimization for securing large water
algorithms.” Urban Water 2 (2): 141–152. https://doi.org/10.1016 distribution networks.” J. Water Resour. Plann. Manage. 134 (6): 516–
/S1462-0758(00)00052-2.
526. https://doi.org/10.1061/(ASCE)0733-9496(2008)134:6(516).
Dini, M., and M. Tabesh. 2014. “A new method for simultaneous calibra-
Kun, D., L. Tian-Yu, W. Jun-Hui, and G. Jin-Song. 2015. “Inversion model
tion of demand pattern and Hazen-Williams coefficients in water dis-
of water distribution systems for nodal demand calibration.” J. Water
tribution systems.” Water Resour. Manage. 28 (7): 2021–2034. https://
Resour. Plann. Manage. 141 (9): 04015002. https://doi.org/10.1061
doi.org/10.1007/s11269-014-0592-4.
/(ASCE)WR.1943-5452.0000506.
Do, N. C., A. R. Simpson, M. Asce, J. W. Deuerlein, and O. Piller. 2016.
“Calibration of water demand multipliers in water distribution systems Laird, C. L., L. T. Biegler, B. G. van Bloemen Waanders, and A. R. Bartlett.
using genetic algorithms.” J. Water Resour. Plann. Manage. 142 (11): 2005. “Contamination source characterization for water networks.”
4016044. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000691. J. Water Resour. Plann. Manage. 131 (2): 125–134. https://doi.org/10
Du, K., R.-Y. Ding, Z.-H. Wang, Z.-G. Song, B.-F. Xu, M. Zhou, Y. Bai, .1061/(ASCE)0733-9496(2005)131:2(125).
and J. Zhang. 2018. “Direct inversion algorithm for pipe resistance Lansey, K. E., W. El-Shorbagy, I. Ahmed, J. Araujo, and C. T. Haan. 2001.
coefficient calibration of water distribution systems.” J. Water Resour. “Calibration assessment and data collection for water distribution net-
Plann. Manage. 144 (7): 04018027. https://doi.org/10.1061/(ASCE) works.” J. Hydraul. Eng. 127 (4): 270–279. https://doi.org/10.1061
WR.1943-5452.0000948. /(ASCE)0733-9429(2001)127:4(270).
Gael, G., and J. Benoi. 2010. “Eigen v3.” Accessed October 08, 2019. Lee, B. H., and R. A. Deininger. 1992. “Optimal locations of monitoring
http://eigen.tuxfamily.org. stations in water distribution system.” J. Environ. Eng. 118 (1): 4–16.
Gao, T. 2014. “Efficient identification of segments in water distribution https://doi.org/10.1061/(ASCE)0733-9372(1992)118:1(4).
networks.” J. Water Resour. Plann. Manage. 140 (6): 04014003. Leskovec, J. 2014. Dimensionality reduction, 384–414. Cambridge, UK:
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000395. Cambridge University Press.
Gao, T. 2017. “Roughness and demand estimation in water distribution net- Mallick, K. N., I. Ahmed, K. S. Tickle, and K. Lansey. 2002. “Determining
works using head loss adjustment.” J. Water Resour. Plann. Manage. pipe groupings for water distribution networks.” J. Water Resour.
143 (12): 04017070. https://doi.org/10.1061/(ASCE)WR.1943-5452 Plann. Manage. 128 (Apr): 130–139. https://doi.org/10.1061/(ASCE)
.0000845. 0733-9496(2002)128:2(130).

© ASCE 04019070-11 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070


Meirelles, G., D. Manzi, B. Brentan, T. Goulart, and E. Luvizotto. 2017. Walski, T. M. 1983. “Technique for calibrating network models.” J. Water
“Calibration model for water distribution network using pressures esti- Resour. Plann. Manage. 109 (4): 360–372. https://doi.org/10.1061
mated by artificial neural networks.” Water Resour. Manage. 31 (13): /(ASCE)0733-9496(1983)109:4(360).
4339–4351. https://doi.org/10.1007/s11269-017-1750-2. Walski, T. M., et al. 1987. “Battle of the network models: Epilogue.”
Morosini, A. F., F. Costanzo, P. Veltri, and D. Savić. 2014. “Identification J. Water Resour. Plann. Manage. 113 (2): 191–203. https://doi.org/10
of measurement points for calibration of water distribution network .1061/(ASCE)0733-9496(1987)113:2(191).
models.” Procedia Eng. 89: 693–701. https://doi.org/10.1016/j.proeng Xie, X., H. Zhang, and D. Hou. 2017. “Bayesian approach for joint esti-
.2014.11.496. mation of demand and roughness in water distribution systems.”
Ormsbee, L. 1989. “Implicit network calibration.” J. Water Resour. Plann. J. Water Resour. Plann. Manage. 143 (8): 04017034. https://doi.org/10
Manage. 115 (2): 243–257. https://doi.org/10.1061/(ASCE)0733-9496 .1061/(ASCE)WR.1943-5452.0000791.
(1989)115:2(243). Xu, J., P. S. Fischbeck, M. J. Small, J. M. VanBriesen, and E. Casman.
Ostfeld, A., et al. 2008. “The battle of the water sensor networks (BWSN): 2008. “Identifying sets of key nodes for placing sensors in dynamic
Downloaded from ascelibrary.org by "Indian Institute of Science, Bangalore" on 07/12/21. Copyright ASCE. For personal use only; all rights reserved.

water distribution networks.” J. Water Resour. Plann. Manage.


A design challenge for engineers and algorithms.” J. Water Resour.
134 (4): 378–385. https://doi.org/10.1061/(ASCE)0733-9496(2008)
Plann. Manage. 134 (6): 556–568. https://doi.org/10.1061/(ASCE)
134:4(378).
0733-9496(2008)134:6(556).
Yu, G., and R. S. Powell. 1994. “Optimal design of meter placement in
Ostfeld, A., et al. 2012. “Battle of the water calibration networks.” J. Water water distribution systems.” Int. J. Syst. Sci. 25 (12): 2155–2166.
Resour. Plann. Manage. 138 (5): 523–532. https://doi.org/10.1061 https://doi.org/10.1080/00207729408949342.
/(ASCE)WR.1943-5452.0000191. Zhang, Q., F. Zheng, H.-F. Duan, Y. Jia, T. Zhang, and X. Guo. 2018.
Rossman, L. A. 2000. Vol. 38 of EPANET 2: Users manual. Cincinnati: “Efficient numerical approach for simultaneous calibration of pipe
Cincinnati US EPA National Risk Management Research Laboratory. roughness coefficients and nodal demands for water distribution sys-
Sanz, G., and R. Pérez. 2015. “Sensitivity analysis for sampling design and tems.” J. Water Resour. Plann. Manage. 144 (10): 04018063. https://doi
demand calibration in water distribution networks using the singular .org/10.1061/(ASCE)WR.1943-5452.0000986.
value decomposition.” J. Water Resour. Plann. Manage. 141 (10): Zheng, F., J. Du, K. Diao, T. Zhang, T. Yu, and Y. Shao. 2018. “Investigat-
1–9. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000535. ing effectiveness of sensor placement strategies in contamination
Todini, E., and S. Pilati. 1987. “A gradient method for the analysis of pipe detection within water distribution systems.” J. Water Resour. Plann.
networks.” In Proc., Int. Conf. on Computer Applications for Water Manage. 144 (4): 06018003. https://doi.org/10.1061/(ASCE)WR.1943
Supply and Distribution. Leicester, UK: Leicester Polytechnic. -5452.0000919.

© ASCE 04019070-12 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2020, 146(1): 04019070

You might also like