You are on page 1of 12

Water Research 217 (2022) 118416

Contents lists available at ScienceDirect

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

Bridging hydraulics and graph signal processing: A new perspective to


estimate water distribution network pressures
Xiao Zhou a, Shuming Liu a, *, Weirong Xu b, Kunlun Xin b, Yipeng Wu a, Fanlin Meng a
a
School of Environment, Tsinghua University, Beijing 100084, China
b
College of Environmental Science and Engineering, Tongji University, Shanghai 200092, China

A R T I C L E I N F O A B S T R A C T

Keywords: The low spatial density of monitored nodal pressures (nodal heads) has already become a bottleneck restricting
Data estimation the development of smart technologies for water distribution networks (WDNs). Inferring unknown nodal heads
Graph signal processing through available WDN information is an effective way to bypass data limitations, but an accurate and easy-to-
Nodal pressure
implement method is still absent. For general WDNs, the spatial distribution of nodal heads is approximately
Smart water
Water distribution network
‘smooth’ as there are few dramatic head changes. If heads can be divided into components with different spatial
varying speeds, then they can be approximated by a few slow varying components. On this basis, a graph-based
head reconstruction (GHR) method is proposed, which employs graph signal processing technologies to recon­
struct the slow varying parts to estimate unknown nodal heads. Four metrics are proposed to bridge WDN hy­
draulics and signal processing to quantify the similarity of adjacent nodal heads, which enhance the smoothness
of heads over the graph, and thus increase estimation accuracy. GHR was tested with different parameter settings
and compared with other head estimation methods. Results showed that GHR has less restrictive parameter
requirements compared with hydraulic simulation, and outperforms traditional data interpolation methods with
better accuracy. At a larger looped network under potential model uncertainties and measurement errors, GHR
still accurately estimated the heads for more than 10,000 unknown nodes, achieving a mean absolute error of
0.13 m using only 100 pressure meters. Thus the proposed method provides an efficient, robust, and convenient
way to estimate unknown nodal heads in WDNs.

1. Introduction acquisition and recording systems have reduced the cost of WDN
monitoring, due to the sheer size and complexity of WDNs, the number
Water distribution networks (WDNs) play a decisive role in the safety of required monitoring sensors is often much larger than what is actually
and stability of urban water supply. The rapid growth of city size and installed for affordability and/or practicality reasons for most of WDNs.
population (Mohammadi and Taylor, 2017) causes an increase in water Thus the low spatial density of available data has already become a
resources consumption and WDN load, urging the implementation of bottleneck restricting the development and application of intelligent
smart technologies and automatic solutions to improve the management technologies in WDNs. How to accurately infer as much unknown in­
of WDNs (Olsson, 2020), which requires large-scale data collection. For formation as possible from limited knowledge of the WDNs could be the
example, automatic diagnosis and forecasting of operating status (Guo key to solve the problem.
et al., 2018; Zhou et al., 2019; Xu et al., 2020) relies on analysis of the Nodal pressure is one of the most important indicators of the hy­
variation of measurements across many locations; and leakage control draulic status of a WDN. For simplicity, in the following discussion, we
and energy saving (Samir et al., 2017; Monsef et al., 2018) often require also use head (the summation of the pressure and elevation of a node) to
comprehensive knowledge of WDN operating status, such as nodal represent nodal pressure. In order to estimate unknown pressures, the
pressures throughout a WDN. Data availability is also the foundation for most widely used approach is to develop a hydraulic model to simulate
recently developed integrated solutions such as digital twin, smart the operation status of a real WDN and calculate unknown nodal pres­
water, and Internet of Things (Alshattnawi, 2017; Radhakrishnan and sures and pipe flow rates by solving continuity and energy equations
Wu, 2018; Conejos Fuertes et al., 2020). Although the progress in data (Rossman, 1999; Sharoonizadeh et al., 2016). Although hydraulic

* Corresponding author.
E-mail address: shumingliu@mail.tsinghua.edu.cn (S. Liu).

https://doi.org/10.1016/j.watres.2022.118416
Received 30 September 2021; Received in revised form 18 January 2022; Accepted 4 April 2022
Available online 5 April 2022
0043-1354/© 2022 Elsevier Ltd. All rights reserved.
X. Zhou et al. Water Research 217 (2022) 118416

models can estimate unknown WDN hydraulic status in a direct and interpolation or machine learning methods. For example, Mounce et al.
complete way, the model parameters often contain unavoidable un­ (2003) used triangle-based cubic interpolation to estimate pressure
certainties (Hutton et al., 2014) as they cannot be directly measured or drops during pipe burst; Soldevila et al. (2019) and Soldevila et al.
require unaffordable costs to be precisely determined (Zhou et al., (2021) employed Kriging spatial interpolation to estimate pressures of
2018). For example, pipe roughness coefficients can only be roughly unmeasured nodes, and Boatwright et al. (2018) estimated unknown
estimated based on pipe diameter and age, or indirectly inferred through pressures by a spatially-constrained inverse-distance weighted interpo­
the relationship between head loss and pipe flow rate. Nodal demands lation method. The above methods neglect the topological structure and
fluctuate over temporal and spatial scales (Herrera et al., 2010), and hydraulic features of WDN, which limit their performance. Traditional
accurate estimation of nodal demands is difficult due to unpredictable interpolation methods assume that spatially closed nodes will have
random customer usages. These model parameter uncertainties may similar values (Fig. 1a), however nodal heads determined by irregular
result in erroneous or misleading estimation results of WDN status. WDN topology and diverse physical components may suggest the
Model calibration and state estimation methods (Savic et al., 2009; opposite (Fig. 1b). Therefore, interpolation methods could lead to large
Preis et al., 2011; Tshehla et al., 2017; Zhou et al., 2018; Chu et al., errors, especially in branches and ends of a WDN. Some studies use
2021) have been proposed to improve the performance of hydraulic machine learning methods to ‘learn’ the relationships between known
models by reducing parametric model uncertainties via available mea­ measurements and unknown states in WDNs. Lima et al. (2017) built an
surements (Savic et al., 2009). However, there are usually thousands of artificial neural network to estimate unknown pressures using moni­
unknown model parameters to be estimated, whereas only a relatively tored inlet flow and pressures; Hajgat’o et al. (2021) used ChebNet to
small number of direct measurements are available. This leads to a reconstruct the nodal pressures and achieved high accuracy with the
highly underdetermined calibration problem that may result in an un­ examined WDNs. These approaches could be promising, but they are
stable solution, no unique solution, or no solution at all (Preis et al., difficult to thoroughly consider various operation conditions and hy­
2011). A possible way to solve the underdetermined problem is to draulic parameter uncertainties when training the model. Beside,
reduce the number of unknown parameters by grouping pipe roughness experienced setting of hyperparameters such as the number of layers and
coefficients (Mallick et al., 2002) and/or nodal demands (Kang and learning rate are often required when developing these models, which
Lansey, 2009; Zhou et al., 2018). This assumes that parameters in a also limit its wide-range application.
group are equal or perfectly correlated, which will result in potential It is necessary to consider the hydraulic features of a WDN to develop
uncertainties in calibrated models. It is also difficult to determine the more accurate and convenient approaches to estimate unknown nodal
appropriate grouping criteria and strategies (Savic et al., 2009; Kumar pressures through limited measurements. Recent developments of graph
et al., 2010). Some state estimation methods (Preis et al., 2011; Tshehla signal processing (GSP) techniques (Stanković et al., 2019) and the
et al., 2017; Chu et al., 2021) assume prior knowledge (also referred to successful applications in environmental monitoring (Jain et al., 2014),
as pseudo-measurements) of nodal demands to increase the number of brain science (Atasoy et al., 2016), and community analysis (Tremblay
constraints, then update hydraulic state estimations using real-time and Borgnat, 2014) provide a new perspective to bridge data processing
measurements. This could make the problem solvable, but the accu­ methods and WDN hydraulics. GSP extends classical signal processing
rate setting of pseudo-measurements is challenging, and the parameters tools to graph-structured signals (Lorenzo et al., 2018), and uses a
with low sensitivity to direct measurements could still be ill-posed and weighted graph to represent the spatial relationships of data in irregular
lead to unstable results, which limits the representativeness of model structures. This characteristic provides the ability to model and analyze
simulation results. Beside, as the model calibration and state estimation the complex interactions among the nodal pressures data of a WDN. The
methods are often complex in computation and experienced reason for choosing GSP to reconstruct unknown nodal pressures lies in
pre-processing is often required, an accurate and easy-to-implement that (1) as the WDN structure can be easily represented by a graph, GSP
method is still in its absence. Due to the difficulty and high cost, most is able to consider the topological relationship among WDN data; (2)
water utilities do not have access to well-calibrated high-precision GSP uses weights to represent the similarity between adjacent signals,
models especially for large complex WDNs, which limits their manage­ which provides the possibility to represent the WDN hydraulics by
ment capabilities. assigning different weights based on the WDN hydraulic features.
When a high-precision hydraulic model is unavailable, an alternative Graph-based methods have been used to solve WDN problems
way is to infer unknown pressures from field measurements using including resilience analysis (Ulusoy et al., 2018; Meng et al., 2018),
interrelated relationships and specific spatial distribution patterns in partitioning (Hajebi et al., 2015), and reconstruction of WDN dynamics
WDN nodal pressures, which has been tried by researchers using data (Wei et al., 2020), etc. In these studies, metrics such as potential energy

Fig. 1. Comparison between actual nodal heads and interpolation results.

2
X. Zhou et al. Water Research 217 (2022) 118416

loss (Ulusoy et al., 2018; Sitzenfrei et al., 2020) and topological prop­ sponding to the eigenvalues. It was proved (Stanković et al., 2019) that
erties (Giustolisi et al., 2019; Meng et al., 2018) were proposed to an eigenvector associated with a smaller eigenvalue will have a lower
analyze the features of a WDN. Nevertheless, due to the complex spatial frequency, i.e., lower varying speed and better smoothness on the
relationship of WDN pressures, it remains to investigate appropriate defined graph. The idea of GFT is to represent the original signal H by a
hydraulic metrics and data analysis methodologies to enable a more linear combination of eigenvectors, thereby dividing it into components
efficient feature description and utilization. Thus these questions still with different frequencies. This is achieved by:
remain to be answered: (1) what kind of metrics are more suitable to
H = U− 1 H (4)
quantify the spatial relationship of pressures; (2) to what extent can we
better reveal the hydraulic features with limited information of a WDN;
H = UH (5)
(3) how can we use GSP to find our desired features, and further use
these features to reconstruct unmonitored pressures. where H = [H 1 , H 2 , …, H N ] is the Laplacian spectrum. Eqs. (4) and (5)
In this study, a new methodology to bridge WDN hydraulics with are called GFT and inverse graph Fourier transform (IGFT) of the graph
GSP is introduced. The WDN structure is represented by a graph, and the signal H, which employ U as the basis set. As ui which corresponds to a
nodal heads are represented by signals defined over the graph vertices. smaller eigenvalue has lower frequency, the former elements of H
Different metrics are proposed to characterize the WDN hydraulics and represent the spatially slow varying parts of H, while the latter elements
quantify the similarity of adjacent nodal heads, with consideration of correspond to the fast varying parts.
limited knowledge of the WDN. Edge weights of the constructed graph The variation of nodal heads in a WDN is determined by the WDN
are determined according to the proposed metrics, which enable the physical structure and pipe flow velocity. Generally, for a WDN in a
graph to better represent the hydraulic relationship of WDN heads. normal steady-state operating condition, the change of heads on con­
Based on the constructed graph and GSP technology, a graph-based head nected nodes is limited. Thus the adjacent heads vary slowly and the
reconstruction (GHR) method is proposed to estimate unmonitored distribution of heads is approximately (not strictly) “smooth” over the
pressures of WDNs. GHR is tested on networks of different sizes and graph. Based on this, we can assume that H is bandlimited and can be
compared with different pressure estimation methods to illustrate its approximately represented by its low-frequency parts, i.e.,
effectiveness and robustness.
H ≈ BH (6)
2. Methodology
H ≈ UBH (7)
2.1. Graph-based head reconstruction
where B is the N × N F -bandlimited matrix, B[i, i] = 1 for each i ≤ F
and other elements are zeros. F is the frequency limit, which decides
The basic idea of GHR is to use GSP to exploit the spatial features of
the number of low-frequency items used to approximate the original
WDN pressures and reconstruct unmeasured values. In this section, we
signal. BH extracts the first F low-frequency elements of H . Eq. (7) is
will demonstrate how to use graph and graph signals to represent a WDN
equivalent to
and its pressures, and provide an overview of GSP and the proposed
unknown pressure estimation method. A graph consists of a set of ver­ H ≈ UBF H F = UF H F (8)
tices V = {1, 2, …, N} and a set of weighted edges B = {wij }i,j∈V
connecting the vertices. Intuitively, the nodes (including tanks and where BF is constructed by the non-zero columns of B, UF is con­
reservoirs) of a WDN can be represented by the vertices of a graph, and structed by the first F rows of U, and H F = [H 1 , H 2 , …, H F ]. For
the pipes (including pumps) can be represented by the edges. If nodes i large WDNs, in order to speed up the calculation, we can only solve the
and j are linked by pipe ij, then wij > 0, otherwise wij = 0. The adjacency first F eigenvectors of L to construct UF , and the methods to solve the
matrix W of the weighted graph can be constructed as W[i, j] = wij , eigenvectors for large sparse symmetric matrices can be used (Lehoucq,
where [i, j] represents the element at the ith row and the jth column. The 1998).
heads of the WDN nodes can be regarded as signals on the graph vertices, Measuring nodal heads in the WDN by pressure meters can be
and can be arranged in a vector form: considered as sampling the original signal H. Assuming there are M
pressure meters, the measurements are
H = [H1 , H2 , …, HN ]T (1)
ys = Ds H (9)
where Hi is the head of node i, as well as the signal on vertex i of the
graph. where ys is the measured heads, Ds is an M × N sampling matrix,
In classical signal analysis, signals can be decomposed onto a set of Ds [i, j] = 1 if the ith pressure meter is installed at node j, and zeros
basis vectors with different frequencies by Fourier (spectral) analysis. otherwise. Substituting Eq. (8) into Eq. (9) and we will get
This approach can be extended to the graph signals using graph Fourier ys = Ds UF H (10)
F
transform (GFT), which allows a graph signal to be represented in the
graph frequency domain spanned by a set of basis vectors. Hence, based on Eq. (10), the low-frequency spectrum H F can be estimated by
different signal varying speeds with respect to the graph topology can be the measured nodal heads. let X = Ds UF , then
described (Mateos et al., 2019). GFT is based on the graph Laplacian ( )
matrix L: ̂ F = pinv(X)y = XT X − 1 XT y
H s s (11)
( )
L = diag 1T W − W (2) where H F is the estimation of low-frequency spectrum; pinv() returns
the pseudo inverse. X needs to have full column rank in order to solve
where diag() converts a vector into a diagonal matrix having the vector Eq. (11), and a necessary condition requires F ≤ M, i.e., the number of
as main diagonal; 1 denotes a vector with all elements equal to one. The low-frequency items should be smaller than the number of pressure
eigen decomposition of L is: meters. After H F is calculated, we can reconstruct the low-frequency
L = UΛUT (3) parts of H by
̂ = UF H
H (12)
where Λ is a diagonal matrix with the eigenvalues of L in ascending F

order; U = [u1 , u2 , …, uN ], which is composed of eigenvectors corre­

3
X. Zhou et al. Water Research 217 (2022) 118416

where Ĥ represents the reconstructed nodal heads. The elements in H


̂ uniform along all the water supplying pipes (except for the known big
are the estimation of unknown nodal heads, and pressures can be ob­ users), i.e., the nodal demands are assumed to be:
tained after subtracting the heights of the nodes. (Qt −

Qb )Lij
Rij = (15)
Lsum
2.2. Bridging WDN hydraulics and graph signal processing ∑
Qn = 0.5 Rij (16)
For a real-life WDN, the pressure changes between different nodes
∀i,j=n

are not uniform in most cases. Therefore, the smoothness assumption in


where Qt is the total demand of the WDN, Qb is the demand of big users,
Eqs. (6) and (7) could not be strictly satisfied, which will introduce er­
Lsum is the sum of the length of all water supplying pipes, Lij is the length
rors and lead to deviations in the reconstructed results. Thus, we can
of pipe ij, Rij is the demand of pipe ij, and Qn is the demand of node n. The
define appropriate weights to enhance the smoothness of pressure
edge weights can then be calculated by simulating head losses using the
change among nodes so that the original heads are better approximated
by the low-frequency parts. roughly constructed model, and are denoted as w(uQ) herein. The ele­
For an edge that connects vertex i and j, the weight wij measures the ments in w(uQ) will also be normalised similar to Eq. (14). Compared with
similarity of signals on the connected vertices. Larger wij corresponds to w, w(uQ) is a not-so-accurate metric to measure the WDN hydraulics, but
more similarity or mutual effects between the signals (Mateos et al., it still reveals the impact of WDN topology and the possible distribution
2019). In literature, wij is generally calculated by the inverse of ‘dis­ of user demands. The assumptions to calculate w(uQ) are especially useful
tance’ between the vertices, as farther distance usually corresponds to for WDNs with relatively uniformly distributed demands, such as WDNs
smaller similarity. For example, in spatial radio signal processing, the in residential areas.
weights are determined by the inverse distance between real-world lo­ The calculation of w and w(uQ) involves hydraulic simulation by
cations of the vertices (Redondi, 2018); and in electrical circuits, the solving continuity and energy equations. In some cases, an index that
weights are defined as reciprocal values to the electrical resistance does not need to perform hydraulic simulation but can measure the
(Stanković et al., 2019), as the resistance represents the ‘distance’ in a WDN hydraulics as accurately as possible can be useful, as it provides a
circuit system. Analogically, the ‘distance’ of nodes in a WDN with cheaper and more convenient way to estimate unknown pressures with a
respect to their heads can be measured by the head loss of pipes that compromise of acceptable decreasing of reconstruction accuracy.
connect the nodes. Thus, the ‘distance’ (i.e., the similarity) of two con­ Referring to the Hazen-Williams formula, hij can be represented by:
nected nodes i and j is given by:
10.67q1.852 Lij
(17)
ij
1 ⃒ ⃒ hij =
⃒ ⃒, ⃒hij ⃒ ≥ ε1 Cw1.852
ij D4.87
ij
⃒hij ⃒
wij = wji = { (13)
1 ⃒⃒ ⃒⃒ where qij , Dij , and Cwij are the flow rate, diameter and pipe roughness
, hij < ε1 coefficient of pipe ij, respectively. Substituting q = v⋅πD2 /4 into Eq. (17)
ε1
and we will get
where hij is the head loss of pipe ij; ε1 is a threshold value to avoid the
1 Lij
appearance of extremely large values, set at 0.01 in our study. Eq. (13) hij = k1 ⋅v1.852
ij ⋅ 1.852 ⋅ 1.166 (18)
Cwij Dij
uses the inverse of head loss to measure the similarity, as smaller head
loss indicates less difference between the adjacent nodal heads. For
where k1 is a constant and vij is the velocity of pipe ij. Assume that the
pumps and valves in WDN, as they will cause a significant head change
velocity and roughness of all pipes are uniform, then we will get
between the connected nodes, very small weights are assigned to the
corresponding edges. In order to make the metric unitless and easier for D1.166
(19)
ij
comparison, a min-max normalization is applied: w(uv)
ij = w(uv)
ji = k2
Lij
wij − min(w) + ε2
wij = (14) where k2 is a constant that can be neglected in the subsequent proced­
max(w) − min(w) + ε2
ures. wij is the weight of edge ij assuming uniform velocity (and
(uv)

where min() and max() return the minimum and maximum of the set, roughness). It is further min-max normalized. The calculation of w(uv)
respectively; w = {wij }, which is the set of wij corresponding to all the only requires the topology, pipe length and diameter of a WDN, which is
pipes in the WDN; ε2 is a small positive value used to avoid zero values, easy to be investigated and measured. w(uv) provides a very rough esti­
set as 0.001(max(w) − min(w)) in our study. mation of pipe head loss, and the similar metric also has been used to
w plays the role of bridging WDN hydraulic features and graph analyze WDN features in other studies (Ulusoy et al., 2018; Sitzenfrei
structures. It is determined by various WDN hydraulic parameters such et al., 2020).
as topology, roughness coefficients, and nodal demands. Calculating w Finally, as a very low-cost alternative, we can assume all weights as
involves an integrated hydraulic simulation with the requirement of a uniform. For all the edges, set
hydraulic model (Rossman, 1999), of which the parameters are difficult
to be accurately determined as discussed in Section 1. Nevertheless, the w(u) (u)
ij = wji = 1 (20)
case studies of our study will show that a hydraulic model with inac­
curate parameters can still provide enough information to calculate w where wij is the weight of edge ij assuming uniform weights.
(u)

for pressure reconstruction. The four metrics proposed in this section, i.e., w, w(uQ) , w(uv) , and w(u)
If a hydraulic model is not currently available, we can still establish a have different assumptions to measure WDN hydraulics as much as
rough model to estimate w under certain assumptions. Basic parameters possible, and will lead to different reconstruction accuracy. Choice of
such as topology, nodal elevations, and pipe lengths and diameters can which metric to use can be determined by how much WDN information
easily be obtained from design files and/or geographic information is available. These metrics will be further tested and discussed in the
systems (GIS) data (Abdelbaki et al., 2017). The parameters which are following sections.
difficult to determine are pipe roughness and nodal demands, as they are In summary, GHR combines WDN hydraulics to enhance the signal
not directly measurable (Zhou et al., 2018). Therefore, the assumptions smoothness on the graph domain, and reconstruct unknown nodal heads
are: (1) pipe roughness of all pipes are uniform; (2) user demand is

4
X. Zhou et al. Water Research 217 (2022) 118416

by calculating the low-frequency parts of the Laplacian spectrum (H F ). bursts and pump failure are not considered. (6) Mean absolute error
The mechanism and steps of GHR are shown in Fig. 2. By choosing a (MAE) is used to measure the performance of GHR, given as:
suitable metric, the weights corresponding to different pipes can be
calculated with available WDN information, and graph Laplacian matrix 1 ∑ Nu
MAE = ̃ i − Hi |
|H (22)
L can be constructed. The eigenvectors of L form the basis of GFT, as Nu i=1
shown in Fig. 2a and b. The low-frequency parts contain most of the
information about the nodal heads (Fig. 2c), therefore the low-frequency where H̃ i is the estimated head of node i, Nu is the number of unknown
spectrum is calculated by a limited number of measured (sampled) nodal nodes.
heads to approximately represent the original value, while the high-
frequency parts can be assumed as zeros (Fig. 2d and e). Finally, the 3.1. Overall performance and comparison with other methods
slow varying parts of the original signal are reconstructed and are used
to estimate the unknown nodal heads, as shown in Fig. 2f. A middle-sized real-life network is used to illustrate the construction
and performance of GHR. The network, as shown in Fig. 3a, consists of 4
3. Case studies and results reservoirs, 480 nodes, and 567 pipes with a total length of 147 km.
About 57,000 m3 of water are supplied from 4 reservoirs per day. The
GHR is applied to two WDNs of different sizes to demonstrate its nodal heads at 8:00 am are considered during the test, because the
applicability and effectiveness. To evaluate the performance of GHR largest nodal demands occur at this time and the varying of heads is the
under different working conditions and parameter settings, synthetic most significant. The nodal heads of the network range from 35 m to 45
data are generated using EPANET2 toolkit (Rossman, 1999) to represent m, as shown in Fig. 3a. A total of 30 pressure meters are used, and the
field data. Pressures at nodes where the meters are installed are set as pressures of reservoirs are also assumed to be known. The meter loca­
measurements, and other data are ‘ground truth’ to evaluate the per­ tions when using w as weights are marked as triangles in Fig. 3a. Thus,
formance of different pressure estimation methods. A number of as­ the objective is to estimate the unknown heads of the rest of 450 nodes
sumptions and presets are made: (1) when calculating w, the pipe with 34 known pressures. Fig. 3b shows the head loss of pipes. Many of
roughness coefficients (Cw) and nodal demands (Q) contain errors by the head losses are larger than 1 m, with the largest head loss reaching
default, as they are the most uncertain parameters in a WDN model and 6.12 m, which indicates the fast varying and less smoothness of heads
can hardly be directly measured (Zhou et al., 2018). Unless otherwise among different nodes, and makes it more difficult to infer the unknown
stated, Cw and Q are assigned with Gaussian white noise N (0, σ2 ) with a pressures from known measurements. Thus, the tested WDN can better
standard deviation of 10 (σ Cw = 10) and 20% of the corresponding true illustrate the performance of GHR under a less favourable condition.
demands (σQ = 20%), respectively. (2) When used, the topological To illustrate the novelty of GHR, Fig. 4 presents how the original
structure, node elevation, pipe length, and pipe diameter of the WDNs nodal heads are approximated by a set of low-frequency basis vectors.
are assumed to be known and certain. (3) Locations of pressure meters Some typical basis vectors and corresponding spectrums (see Eqs. (4)
are determined according to D-optimality (Winer, 1962; Tsitsvero et al., and (5)) are shown. The former values of H , such as H 1 and H 5 are
2016; Lorenzo et al., 2018), given by: much larger than the latter values. Therefore, ignoring the latter value
( ) will have little impact, as discussed in Eqs. (6) and (7). From another
S = arg max logdet XT X (21)
|S|=M perspective, the former basis vectors, such as u1 , u5 , and u15 change
more smoothly and globally, while the latter vectors such as u100 and
where S is the set of pressure meters. (4) For all the tests, the frequency u300 show local and rapid changes. At the same time, Fig. 3(a) shows that
limits are set as F = 0.9M to ensure X has full column rank and Eq. the real nodal heads mainly include global changes, which is consistent
(11) is solvable. (5) For illustration purposes, the tested WDNs operate with the former low-frequency basis vectors. The analysis shows that
under normal conditions, and abnormal conditions such as main pipe using a limited number of low-frequency basis vectors to approximate

Fig. 2. Schematic diagram of GHR.

5
X. Zhou et al. Water Research 217 (2022) 118416

Fig. 3. The basic information of the WDN. (a) The real nodal heads and pressure meter locations. (b) The head loss of pipes.

Fig. 4. Typical basis vectors and corresponding spectrums of the original heads.

6
X. Zhou et al. Water Research 217 (2022) 118416

Fig. 5. The nodal head estimation errors by different methods. (a–d) The estimation errors of GHR using w, w(uQ) , w(uv) , and w(u) as weights, respectively. (e–f) The
estimation errors of nearest neighbor interpolation and universal Kriging interpolation, respectively.

the original heads is feasible in practical WDNs. of the network) have significantly increased error. Using uniform
In Fig. 5, the performance of GHR with different weight metrics (w, weights assumes that all adjacent nodal heads have the same similarity,
w(uQ) , w(uv) , and w(u) , see Section 2.2) are tested, and are compared to two which will lead to low accuracy when a large head loss occurs. However,
spatial interpolation methods. Fig. 5a–d shows the head reconstruction compared with the interpolation methods, GHR with uniform weights
results using the four metrics. As shown in Fig. 5a, using w as weights, still achieves better results as the topology structure of the WDN is
the MAE of the estimated heads is only 0.25 m, and 83% of estimation considered.
errors are less than 0.5 m. Beside, comparing with Fig. 3, we can find For comparison purposes, nearest neighbor interpolation and uni­
that the head differences of many adjacent nodes are larger than 1 m, versal Kriging interpolation (Li and Heap, 2008) are tested and the re­
while GHR still obtains accurate estimation results for these nodes, sults are presented in Fig. 5e and f. These methods achieve lower
which also shows the effectiveness of GHR. Therefore, despite the accuracy compared with the results of GHR. As discussed in Fig. 1, the
parameter uncertainties (σ Cw = 10, σ Q = 20%), GHR can provide very interpolation methods cannot incorporate the topological structure and
good estimation results with the help of a hydraulic model. The calcu­ hydraulic features of WDN, therefore larger errors may occur. Universal
lation of w(uQ) actually builds a temporal hydraulic model with roughly Kriging interpolation is more accurate than nearest interpolation, as it
assigned hydraulic parameters to determine the weights, leading to a considers the spatial distance between nodes. However, it still shows
larger MAE of 0.29 m (Fig. 5b). Although the calculation of w and w(uQ) poor performance for nodes that are far from the meters or located at the
requires to perform hydraulic simulations, GHR shows good robustness end of the network, because the values of these nodes are less relevant to
to hydraulic parameter uncertainties, which will be further discussed in their distance to measured nodes.
Section 3.2.
Fig. 5c shows the results of GHR using w(uv) . Calculating w(uv) needs 3.2. Parameter uncertainties
less WDN knowledge and do not require performing hydraulic simula­
tion. Although the accuracy decreases to MAE=0.48 m, 64% of the er­ Fig. 5a shows that GHR can provide good estimation results with the
rors are still less than 0.5 m, and 99% are less than 2 m. Larger errors help of hydraulic models despite parameter uncertainties. It is mean­
occur when the nearby pipes have extremely large or small flow veloc­ ingful to investigate whether using GHR with w can achieve better ac­
ities. Using GHR with w(uv) provides a convenient and low-cost alter­ curacy than direct hydraulic simulation under different parameter
native to estimate unknown WDN pressures, as the required parameters uncertainties. In this section, the impact of parameter uncertainties,
(e.g., length and diameter of pipes) are easier to obtain. The results using including systematic errors and random errors, are tested, and the re­
uniform weights (w(u) ) is shown in Fig. 5d. The overall estimation error sults of GHR and hydraulic simulation under different degrees of un­
increases to MAE = 0.51 m, and some nodes (especially nodes at the end certainties are compared. Each test is repeatedly performed 50 times
with regenerated noises to Cw and Q. The noises follow N (μCw , σ2Cw ) and

7
X. Zhou et al. Water Research 217 (2022) 118416

Fig. 6. The results by GHR and hydraulic simulation with different levels of systematic and random errors.

N (μQ , σ2Q ), where μCw and μQ are the systematic error of Cw and Q, section, different numbers of meters and measurement errors are tested
respectively. For GHR, the pressure of 30 measured nodes and 4 reser­ with w, w(uQ) , w(uv) , and w(u) , respectively. The measurements are
voirs are used, same as in Section 3.1. The results are shown in Fig. 6, assigned with Gaussian white noise with standard deviation (σH )
and it can be found that GHR is less affected by both systematic and ranging from 0 m to 0.5 m. Each parameter set is performed 50 times
random errors. For example, the hydraulic simulation results can only with regenerated errors, and the results are shown in Fig. 7. Using more
have better accuracy than GHR when there are small parameter un­ meters will improve the accuracy of GHR, as more information about the
certainties (σQ < 20% and σCw < 10) with no systematic error (μCw = 0 spatial distribution of nodal heads can be provided. With 75 pressure
and μQ = 0). Focusing on the first abscissa (no random errors), with meters, the MAE of GHR using w as weights can decrease to only 0.10 m
different levels of systematic errors, the MAEs of hydraulic simulation when small measurement errors are assigned(Fig. 7a). We also identified
results increase 0.31 m, 0.36 m, and 0.70 m, respectively, while the MAE that using more meters can reduce the performance differences among
of GHR results only increase 0.02 m, 0.03 m and 0.05 m, respectively different weights. For example, the difference of MAE when using w and
(Fig. 6a to 6d). Under a specific systematic error, the increasing of w(u) is around 0.5 m with only 15 meters, whereas the difference de­
random errors (σQ and σ Cw ) also has more significant impacts on the creases to around 0.15 m with 75 meters. Therefore, when limited in­
error of hydraulic simulation results. The robustness of GHR to hydraulic formation about a WDN is known, we can still install a sufficient number
parameter uncertainties lies in that it combines both the information of meters, and use GHR with the more easy-to-obtain weights to estimate
from WDN hydraulics and the available measurements, and the hy­ the unknown pressures with high accuracy.
draulic model act only as a reference to calculate w. The introduced The MAE with different degrees of measurement errors are compared
errors in w can be tolerated as long as the smoothness hypothesis is in Fig. 7a-d. The accuracy decreases with larger measurement error, as
ensured. On the other hand, although a hydraulic modal can utilize field well as a wider distribution range of MAE for different tests. However,
measurements to achieve better accuracy by model calibration methods, GHR still shows good robustness with the increasing of the errors, and
it is not easy to achieve, as discussed in Section 1. the MAE of estimation results can be even smaller than σ H with sufficient
meters. For instance, when σH increases from 0.01 m to 0.5 m, the MAE
only increases by 0.2–0.3 m, and the MAE of GHR using w is only around
3.3. Number of meters and measurement errors 0.4 m when σH = 0.5m. The robustness of GHR to measurement errors
comes from the fact that GHR discards the spatially high-frequency parts
Other factors that have significant effects on the performance of GHR and emphasis the low-frequency parts, thus implicitly filtering out part
include the quantity and quality of available measurements. In this

8
X. Zhou et al. Water Research 217 (2022) 118416

Fig. 7. The performance of GHR with different numbers of meters and measurement errors.

of the measurement errors. w(uQ) has better performance especially for the nodes near the main
pipes, as these pipes have large head losses. However, using w(u) still
3.4. Performance on large network shows good performance with an MAE of 0.167 m, especially for the
nodes with slow pressure changes.
In this section, GHR is tested on a large looped network to further In order to further illustrate the performance of GHR under different
illustrate its performance on real-life WDNs. The network, as shown in operation conditions, it is applied to real-time estimate unknown nodal
Fig. 8a, contains 12,523 nodes, 2 sources, and 14,822 pipes. It was heads during 24 h. The operating data of the second day of the original
initially used in the Battle of the Water Sensor Networks (BWSN) (Ost­ BWSN network are used, because a large demand variation occurs on
feld et al., 2008). For simplicity and better illustration of the effect of this day, thus the impacts of daily demands variations can be better
customer demand variation, control rules in the original network are tested. The tests are performed 50 times with regenerated errors for
deleted. The nodal heads at 7:00 am, when the network has the largest model parameters and measurements. The graph weights w are calcu­
nodal demands, are used to test the performance of GHR. Model un­ lated based on only the WDN hydraulics at 7:00 am (same as in Fig. 8),
certainties (σ Cw = 10 and σQ = 20%) and measurement errors (σ H = and are used to estimate unknown heads at different operation condi­
0.2 m) are considered during the tests. The real nodal heads of the WDN tions throughout the day. The MAEs and the demand variation are
range from 66 m to 75 m, where the upper left parts show faster pressure shown in Fig. 9. It can be seen that GHR can have better accuracy with
change, and the bottom right parts change slowly, as shown in Fig. 8a. lower nodal demands, as low nodal demands result in slow variation of
The estimation error with 100 pressure meters and different weight nodal heads, which is more consistent with the smoothness hypothesis of
metrics are shown in Fig. 8b, c, and d, respectively. Despite the existence GHR. Comparing the MAE at 7:00 am and 9:00 pm, we can find that
of parameter uncertainty and measurement errors, GHR with w accu­ although the graph weights are calculated using the hydraulics of a
rately estimated the nodal head of more than 10,000 unknown nodes different operation, the increase in MAE is small. Therefore, it is feasible
with an MAE of 0.13 m. Although the existing pumps and valves cause to use the w obtained from a specific operating condition to estimate
dramatic nodal head change that cannot be assumed as ‘smooth’, the unknown nodal heads under other operating conditions.
head of these nodes can still be successfully estimated by assigning small
weights and setting appropriate meters. As shown in Fig. 8c, Using w(uQ) 4. Discussions
as weights reduces the overall estimation accuracy, as Cw and Q are
roughly determined and will introduce large uncertainties when calcu­ Compared with hydraulic simulation, estimating unknown nodal
lating weights. Compared with Fig. 8d that using w(u) as weights, using pressures from available measurements can be a more easy-to-

9
X. Zhou et al. Water Research 217 (2022) 118416

Fig. 8. The performance of GHR in a large looped WDN with different metrics.

implement alternative approach, but there are few successful methods in the construction and maintenance of hydraulic models under a lot of
that can achieve high accuracy results. Through bridging WDN hy­ cases, and using GHR provides a convenient and low cost approach to
draulics with GSP, GHR provides a novel method to reconstruct un­ achieve good estimation results without the requirement to accurately
known nodal heads with not only high accuracy, but also good determine WDN hydraulic parameters. For example, although some
robustness to parameter uncertainties and measurement errors. Based water companies have currently available hydraulic models, the absence
on the examples and results presented above, comprehensive discussions of long-term maintenance and calibration results in a significant
are summarized as follows. decrease in model accuracy and thus the simulated results cannot be
The main contribution of GHR is that it provides an effective but low- directly applied to infer WDN hydraulics. In that case, using the inac­
cost methodology to estimate the unknown WDN pressures. When an curate hydraulic model to calculate w and employ GHR to reconstruct
accurate hydraulic model is available or there are sufficient technicians unknown pressures could be a very convenient way to achieve good
and investments to perform elaborated model calibration, integrated results. Even with no hydraulic models, GHR can use w(uQ) , w(uv) and w(u)
hydraulic simulation is still the best way to calculate the unknown to exploit more readily WDN information such as pipe length and di­
pressures. However, water companies would have limited investments ameters, and reveal the WDN hydraulics as much as possible to improve

10
X. Zhou et al. Water Research 217 (2022) 118416

to precisely determine WDN physical parameters (comparing


with hydraulic models), and can provide accurate estimation
results (comparing with traditional data interpolation methods).
(2) GHR shows high estimation accuracy under large parameter
uncertainties and measurement errors, indicating its robustness
for real-life applications.
(3) The similarity metrics proposed in this study can significantly
improve the performance of GSP methods in analysing WDN
pressure data.

Although the main focus of this study is to estimate unknown WDN


data, the idea that combining WDN hydraulics with graph signal pro­
cessing methods also presents a new perspective to analyze the spatial
features of irregularly distributed data, and provides insights for solving
other problems in the research field of water resources and environ­
ments, such as estimating unknown water quality in WDNs or basins, or
Fig. 9. The MAE of GHR at different times of a day. detecting anomalies such as pipe bursts and incident contaminations.
Further multidisciplinary studies, such as combining machine learning
its accuracy. Compared with traditional data interpolation methods, and deep learning technologies are also recommended.
GHR has better performance because it considers the topology of a WDN
and reveals the relationship of heads at different nodes, thus effectively Declaration of Competing Interest
improving the estimation accuracy.
GHR shows better performance in WDNs with low pipe flow rates The authors declare that they have no known competing financial
and slow varying nodal heads, as these WDNs are more in accordance interests or personal relationships that could have appeared to influence
with the smoothness hypothesis of GHR. Therefore, for these WDNs, we the work reported in this paper.
can use fewer meters as well as a smaller frequency limit F to achieve
the required accuracy. For WDNs with high pipe flow rates, the head Acknowledgment
losses are larger and the spatial distribution of nodal heads changes
faster. Thus more meters are required to ensure the estimation accuracy. This work was financially supported by National Natural Science
During the test, F is set as 90% of the number of meters. However, Foundation of China (Grant Nos. 52100113, 51879139) and China
when there are large measurement errors or inappropriate setting of Postdoctoral Science Foundation (Grant No. 2021M691824). X.Z. ac­
meter locations, we would recommend smaller F to ensure the stability knowledges the support of the ShuiMu Tsinghua Scholar Program of
for the calculation of low frequency spectrum (Eq. (11)). Beside, if there Tsinghua University.
are a large number of meters, we recommend keeping F < 100 as too
large F will introduce fast varying basis vectors, which may also References
decrease the stability of GHR.
GHR is based on WDN operating status that already has occurred and Abdelbaki, C., Benchaib, M.M., Benziada, S., Mahmoudi, H., Goosen, M., 2017.
Management of a water distribution network by coupling GIS and hydraulic
been monitored. Thus, it is difficult to simulate a hypothetic operating
modeling: a case study of Chetouane in Algeria. Appl. Water Sci. 7 (3), 1561–1567.
status that has not occurred yet. For example, if we need to know the https://doi.org/10.1007/s13201-016-0416-1.
impact of closing a valve or adding a large demand, hydraulic simulation Alshattnawi, S.K., 2017. Smart water distribution management system architecture based
is the more appropriate way than GHR. Beside, GHR estimates WDN on internet of things and cloud computing. In: Proceedings of the 2017 International
Conference on New Trends in Computing Sciences (ICTCS).
pressures in a global way, and it cannot accurately consider the head Atasoy, S., Donnelly, I., Pearson, J., 2016. Human brain networks function in
change of a certain pipe, especially for valves and pumps that may cause connectome-specific harmonic waves. Nat. Commun. 7 (1), 10340. https://doi.org/
significant pressure changes. Although we can assign very small weights 10.1038/ncomms10340.
Boatwright, S., Romano, M., Mounce, S., Woodward, K., Boxall, J., 2018. Optimal sensor
to the edges that represent valves/pumps, additional meters are placement and leak/burst localisation in a water distribution system using spatially-
required, which will lead to increasing investments. Therefore, when constrained inverse-distance weighted interpolation. In: Proceedings of the 13th
GHR is applied to WDNs with a large number of pumps and valves, its International Conference on Hydroinformatics. Palermo, Italy.
Chu, S., Zhang, T., Yu, T., Wang, Q.J., Shao, Y., 2021. A noise adaptive approach for
cost and performance would be affected. nodal water demand estimation in water distribution systems. Water Res. 192,
116837 https://doi.org/10.1016/j.watres.2021.116837.
5. Conclusions Conejos Fuertes, P., Martínez Alzamora, F., Hervás Carot, M., Alonso Campos, J.C., 2020.
Building and exploiting a digital twin for the management of drinking water
distribution networks. Urban Water J. 17 (8), 704–713. https://doi.org/10.1080/
A novel GHR method is proposed to estimate unknown pressures of 1573062X.2020.1771382.
WDNs. GHR assumes the spatial distribution of nodal heads is smooth, Giustolisi, O., Ridolfi, L., Simone, A., 2019. Tailoring centrality metrics for water
distribution networks. Water Resour. Res. 55 (3), 2348–2369. https://doi.org/
and uses GSP theory to reconstruct the unknown heads by their slow
10.1029/2018WR023966.
varying parts. It analyses WDN system hydraulic features to quantitate Guo, G., Liu, S., Wu, Y., Li, J., Zhou, R., Zhu, X., 2018. Short-term water demand forecast
the similarity of different nodes with respect to their heads, and assigns based on deep learning method. J. Water Resour. Plan. Manag. 144 (12), 04018076
specific weights to the graph structure. On this basis, the smoothness of https://doi.org/10.1061/(ASCE)WR.1943-5452.0000992.
Hajebi, S., Roshani, E., Cardozo, N., Barrett, S., Clarke, A., Clarke, S., 2015. Water
nodal head distribution on the graph domain is enhanced, thus accurate distribution network sectorisation using graph theory and many-objective
estimation results can be achieved with a limited number of meters. optimisation. J. Hydroinform. 18 (1), 77–95. https://doi.org/10.2166/
GHR is tested on two different WDNs, and the key findings are sum­ hydro.2015.144.
Hajgat’o, G., Gyires-T’oth, B.A., & Pa’al, G.J.A., 2021. Reconstructing nodal pressures in
marized below. water distribution systems with graph neural networks. https://arxiv.org/abs/2
104.13619.
Herrera, M., Torgo, L., Izquierdo, J., Perez-Garcia, R., 2010. Predictive models for
forecasting hourly urban water demand. J. Hydrol. 387 (1–2), 141–150. https://doi.
(1) GHR provides an efficient and convenient way to estimate un­ org/10.1016/j.jhydrol.2010.04.005 (Amst).
known nodal pressures, which does not require high investment Hutton, C.J., Kapelan, Z., Vamvakeridou-Lyroudia, L., Savic, D.A., 2014. Dealing with
uncertainty in water distribution system models: a framework for real-time modeling

11
X. Zhou et al. Water Research 217 (2022) 118416

and data assimilation. J. Water Resour. Plan. Manag. 140 (2), 169–183 doi:10.1061/ Rossman, L.A., 1999. The EPANET programmer’s toolkit for analysis of water
%28ASCE%29WR.1943-5452.0000325. distribution systems. In: Proceedings of the 29th Annual Water.
Jain, R.K., Moura, J.M.F., Kontokosta, C.E., 2014. Big data + big cities: graph signals of Samir, N., Kansoh, R., Elbarki, W., Fleifle, A., 2017. Pressure control for minimizing
urban air pollution [exploratory SP]. IEEE Signal Process. Mag. 31 (5), 130–136. leakage in water distribution systems. Alex. Eng. J. 56 (4), 601–612. https://doi.org/
https://doi.org/10.1109/MSP.2014.2330357. 10.1016/j.aej.2017.07.008.
Kang, D., Lansey, K., 2009. Real-time demand estimation and confidence limit analysis Savic, D.A., Kapelan, Z.S., Jonkergouw, P.M.R., 2009. Quo vadis water distribution
for water distribution systems. J. Hydraul. Eng. 135 (10), 825–837 doi:10.1061/% model calibration? Urban Water J. 6 (1), 3–22. https://doi.org/10.1080/
28ASCE%29HY.1943-7900.0000086. 15730620802613380.
Kumar, S.M., Narasimhan, S., Murty Bhallamudi, S., 2010. Parameter estimation in water Sharoonizadeh, S., Mamizadeh, J., Sarvarian, J., 2016. Comparison of solution methods
distribution networks. Water Resour. Manag. 24 (6), 1251–1272. https://doi.org/ for analyzing water distribution networks under pressure-deficient conditions.
10.1007/s11269-009-9495-1. J. Water Supply Res. Technol. Aqua 65 (4), 330–341. https://doi.org/10.2166/
Li, J., Heap, A.D., 2008. A Review of Spatial Interpolation Methods for Environmental aqua.2016.084.
Scientists. Geoscience Australia, Australia. Sitzenfrei, R., Wang, Q., Kapelan, Z., Savić, D., 2020. Using complex network analysis for
Lima, G.M., Brentan, B.M., Manzi, D., Luvizotto Jr., E., 2017. Metamodel for nodal optimization of water distribution networks. Water Resour. Res. 56 (8) https://doi.
pressure estimation at near real-time in water distribution systems using artificial org/10.1029/2020WR027929 e2020WR027929.
neural networks. J. Hydroinform. 20 (2), 486–496. https://doi.org/10.2166/ Soldevila, A., Blesa, J., Fernández-Cantí, R.M., Tornil-Sin, S., Puig, V., 2019. Data-driven
hydro.2017.036. approach for leak localization in water distribution networks using pressure sensors
Lorenzo, P.D., Barbarossa, S., Banelli, P., 2018. Sampling and Recovery of Graph Signals. and spatial interpolation. Water 11, 1500. https://doi.org/10.3390/w11071500
Cooperative and Graph Signal Processing. Academic Press, pp. 261–282. (Basel).
Mallick, K.N., Ahmed, I., Tickle, K.S., & Lansey, K., 2002. Determining pipe groupings for Soldevila, A., Blesa, J., Jensen, T.N., Tornil-Sin, S., Fernández-Cantí, R.M., Puig, V.,
water distribution networks. 128(2), 130–139. 10.1061/(ASCE)0733-9496(2002) 2021. Leak localization method for water-distribution networks using a data-driven
128:2(130). model and dempster–shafer reasoning. IEEE Trans. Control Syst. Technol. 29 (3),
Mateos, G., Segarra, S., Marques, A.G., Ribeiro, A., 2019. Connecting the dots: 937–948. https://doi.org/10.1109/TCST.2020.2982349.
identifying network structure via graph signal processing. IEEE Signal Process. Mag. Stanković, L., Daković, M., Sejdić, E., 2019. Introduction to graph signal processing. In:
36 (3), 16–43. https://doi.org/10.1109/MSP.2018.2890143. Introduction to Graph Signal Processing, 1. Springer International Publishing,
Meng, F., Fu, G., Farmani, R., Sweetapple, C., Butler, D., 2018. Topological attributes of pp. 3–99.
network resilience: a study in water distribution systems. Water Res. 143, 376–386. Tremblay, N., Borgnat, P., 2014. Graph wavelets for multiscale community mining. IEEE
https://doi.org/10.1016/j.watres.2018.06.048. Trans. Signal Process. 62 (20), 5227–5239. https://doi.org/10.1109/
Mohammadi, N., Taylor, J.E., 2017. Smart city digital twins. In: Proceedings of the IEEE TSP.2014.2345355.
Symposium Series on Computational Intelligence (SSCI). Tshehla, K.S., Hamam, Y., Abu-Mahfouz, A.M., 2017. State estimation in water
Monsef, H., Naghashzadegan, M., Farmani, R., Jamali, A., 2018. Pressure management in distribution network: a review. In: Proceedings of the IEEE 15th International
water distribution systems in order to reduce energy consumption and background Conference on Industrial Informatics (INDIN).
leakage. J. Water Supply Res. Technol. Aqua 67 (4), 397–403. https://doi.org/ Tsitsvero, M., Barbarossa, S., Lorenzo, P.D., 2016. Signals on graphs: uncertainty
10.2166/aqua.2018.002. principle and sampling. IEEE Trans. Signal Process. 64 (18), 4845–4860. https://doi.
Mounce, S.R., Khan, A., Wood, A.S., Day, A.J., Widdop, P.D., Machell, J., 2003. Sensor- org/10.1109/TSP.2016.2573748.
fusion of hydraulic data for burst detection and location in a treated water Ulusoy, A.J., Stoianov, I., Chazerain, A., 2018. Hydraulically informed graph theoretic
distribution system. Inf. Fusion 4 (3), 217–229. https://doi.org/10.1016/S1566- measure of link criticality for the resilience analysis of water distribution networks.
2535(03)00034-4. Appl. Netw. Sci. 3 (1), 31. https://doi.org/10.1007/s41109-018-0079-y.
Olsson, G., 2020. Urban water supply automation – today and tomorrow. J. Water Supply Wei, Z., Pagani, A., Fu, G., Guymer, I., Chen, W., McCann, J., Guo, W., 2020. Optimal
Res. Technol. Aqua 70 (4), 420–437. https://doi.org/10.2166/aqua.2020.115. sampling of water distribution network dynamics using graph fourier transform.
Ostfeld, A., Uber, J.G., Salomons, E., et al., 2008. The battle of the water sensor networks IEEE Trans. Netw. Sci. Eng. 7 (3), 1570–1582. https://doi.org/10.1109/
(BWSN): a design challenge for engineers and algorithms. J. Water Resour. Plan. TNSE.2019.2941834.
Manag. 134 (6), 556–568, 10.1061/(ASCE)0733-9496(2008)134:6(556). Winer, B.J., 1962. Statistical Principles in Experimental Design. McGraw-Hill Book
Preis, A., Whittle, A.J., Ostfeld, A., & Perelman, L., 2011. Efficient hydraulic state Company, New York, US.
estimation technique using reduced models of urban water networks. 137(4), Xu, W., Zhou, X., Xin, K., Boxall, J., Yan, H., Tao, T., 2020. Disturbance extraction for
343–351. 10.1061/(ASCE)WR.1943-5452.0000113. burst detection in water distribution networks using pressure measurements. Water
Lehoucq, R.B., Sorensen, D.C., Yang, C., 1998. ARPACK USERS GUIDE: Solution of Large Resour. Res. 56 (5) https://doi.org/10.1029/2019WR025526 e2019WR025526.
Scale Eigenvalue Problems By Implicitly Restarted Arnoldi Methods. SIAM, Zhou, X., Xu, W., Xin, K., Yan, H., Tao, T., 2018. Self-adaptive calibration of real-time
Philadelphia, PA. demand and roughness of water distribution systems. Water Resour. Res. 54 (8),
Radhakrishnan, V., Wu, W., 2018. IoT technology for smart water system. In: 5536–5550. https://doi.org/10.1029/2017WR022147.
Proceedings of the IEEE 20th International Conference on High Performance Zhou, X., Tang, Z., Xu, W., et al., 2019. Deep learning identifies accurate burst locations
Computing and Communications (HPCC/SmartCity/DSS). in water distribution networks. Water Res. 166, 115058 https://doi.org/10.1016/j.
Redondi, A.E.C., 2018. Radio map interpolation using graph signal processing. IEEE watres.2019.115058.
Commun. Lett. 22 (1), 153–156. https://doi.org/10.1109/LCOMM.2017.2762318.

12

You might also like