You are on page 1of 10

Water Research 233 (2023) 119747

Contents lists available at ScienceDirect

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

A convenient and stable graph-based pressure estimation methodology for


water distribution networks: Development and field validation
Xiao Zhou a, Juan Zhang b, d, Shuyi Guo c, Shuming Liu b, *, Kunlun Xin c
a
College of Civil Engineering, Hefei University of Technology, Hefei, China
b
School of Environment, Tsinghua University, Beijing, China
c
College of Environmental Science and Engineering, Tongji University, Shanghai, China
d
Hebei Construction and Investment Group Water Investment CO., Ltd, Shijiazhuang, China

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

Keywords: Accurate estimation of unknown nodal pressures (nodal heads) is necessary for efficient operation and man­
Data estimation agement of water distribution networks (WDNs), but existing methods such as hydraulic simulation and data
Graph signal processing interpolation can hardly reconcile estimation accuracy with model construction and maintenance costs. Recent
Hydraulic model
developments in graph signal processing (GSP) techniques provide us with new tools to utilize information in
Nodal pressure
Water distribution network
WDN hydraulics and available measurements. In a pilot study, a graph-based head reconstruction (GHR) method
was proposed, which used GSP to reconstruct the spatially slow-varying parts of nodal heads from a limited
number of field measurements to approximate original heads. GHR has illustrated the effectiveness and ease of
implementation of GSP-based methods. However, due to the ill-conditioning reconstruction process and inherent
uncertainties, GHR may show unstable results with large errors if pressure meters are not installed at specific
optimized locations, which limits its applicability. To solve this problem and discover a stable and convenient
method that can support a wider range of applications, a graph-based head reconstruction method with improved
stability (GHR-S) is proposed. GHR-S utilizes a rough estimation of unknown pressures as pseudo measurements,
which provide additional constraints and avoid the occurrence of unreasonable results during the reconstruction
process. A middle-sized network with synthetic data illustrates the stability, convenience, and accuracy of GHR-S
with arbitrary meter locations and uncalibrated model parameters. GHR-S is also applied to a large real-life
network with field measurements, and successfully estimates the unknown pressures of 83,000 nodes with
only 58 measurements, showing its effectiveness in practical engineering.

1. Introduction hydraulic relationships and available measurements can be a more


economically feasible way, and various methods have been proposed.
Accurate knowledge of nodal pressures provides decisive informa­ The most widely used approach is to build a mathematical model to
tion for the management and scheduling of water distribution networks simulate the hydraulic behaviors of a WDN. Although various methods
(WDNs), such as leakage control (Xu et al., 2014; Bozkurt et al., 2022), and toolkits (Rossman, 1999, 2010) have been proposed, the main
optimal scheduling (Dai et al., 2018; Zhang et al., 2022), and anomaly challenge lies in the difficulty of accurately determining the vast number
detection (Zhou et al., 2019; Xu et al., 2020), which further helps water of model parameters (Savic et al., 2009). For example, the accurate
companies to reduce costs and save energy and resources. Due to the allocation of demands to thousands of nodes in a WDN is costly and
immense size of WDNs and the high construction cost of pressure meters, laborious, and the randomness of customer behaviors adds to the diffi­
for most water supply companies, meters are installed at only a few key culties. Also, the roughness coefficients of pipes with various physical
nodes such as important users and end nodes, and the density of conditions are hard to be accurately determined, as they are not directly
monitored data is often insufficient for the requirement of intelligent measurable (Zhou et al., 2018). Other uncertainties, such as errors
and elaborative management and scheduling. In order to obtain more caused by model skeletonization, demand aggregation, and instrumen­
comprehensive information, estimating unknown pressures based on tation errors (Hutton et al., 2014) also aggregate the inaccuracy of a

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

https://doi.org/10.1016/j.watres.2023.119747
Received 3 October 2022; Received in revised form 19 January 2023; Accepted 13 February 2023
Available online 14 February 2023
0043-1354/© 2023 Published by Elsevier Ltd.
X. Zhou et al. Water Research 233 (2023) 119747

hydraulic model. Therefore, hydraulic simulations could provide un­ the graph vertices. The inhomogeneous hydraulic relationship of
certain, erroneous, or misleading results if the parameters are not finely different nodes can also be considered by setting specific weights to
adjusted. Model calibration (also referred to as state estimation) graph links. On this basis, Zhou et al. (2022) proposed a new method­
methods (Savic et al., 2009; Du et al., 2018; Zhou et al., 2018; Chu et al., ology named graph-based head reconstruction (GHR) by bridging WDN
2021) have been proposed to find suitable model parameters according hydraulics with GSP. GHR divides the spatial distribution of pressures
to the difference between model simulation results and field measure­ into components with different varying speeds (i.e., frequencies). With
ments. However, as summarized by Zhou et al. (2022), the an assumption that the slow-varying components make up the majority
under-determination of calibration problems and the requirement of of the original heads, the low-frequency parts can be reconstructed from
complex computation and experienced data pre-processing result in the a limited number of known measurements to approximate the original
absence of an accurate and easy-to-use calibration method. Inevitable heads. GHR achieved accurate results without the requirements of
measurement errors also add to the difficulties of accurate calibration high-precision hydraulic parameters, illustrating the effectiveness and
(Hutton et al., 2014; Zhou et al., 2018). Therefore, a majority of water ease of implementation of GSP-based methods. Nevertheless, the per­
utilities still do not have ready-to-use high-precision models, especially formance of GHR relies on its assumption that the locations of pressure
for WDNs with large sizes and complex structures. meters are determined according to D-optimality (Winer, 1962; Tsits­
Due to the difficulties in building an elaborate hydraulic model, some vero et al., 2016). With arbitrary locations, due to the absence of con­
studies have turned to explore easier-to-use methods to estimate un­ straints to some unmeasured nodes, the reconstruction process could be
known pressures. For example, data interpolation methods assume that ill-conditioned, which may lead to unstable results and very large esti­
spatially closed nodes will have similar nodal heads (i.e., the summation mation errors (see Section 2.2). In practical engineering, however,
of nodal pressure and elevation), thus unknown heads can be estimated pressure meters are generally installed in advance, and other measuring
according to their distance to known measurements. Mounce et al. purposes also need to be taken into account, thus it can be difficult to
(2003) used a cubic spatial interpolation method to build a pressure ensure the designated meter locations. Therefore, the applicability of
drop surface caused by pipe bursts; similarly, Soldevila et al. (2019) GHR is still limited.
employed Kriging spatial interpolation methods to estimate nodal To develop an accurate and easy-to-use pressure estimation method
pressures, and compared the estimated pressures with their reference that can support a wider range of applications, improving the stability of
values to localize bursts. Boatwright et al. (2018) used the outputs of a GSP-based methods with arbitrary meter locations can be a feasible and
spatially-constrained inverse-distance weighted interpolation method to promising technological approach. In this paper, a graph-based head
determine the optimal location of meters for leak/burst localization. reconstruction method with improved stability (GHR-S) is proposed.
Although the data interpolation methods are convenient and easy to use, Taking advantage of the slowly-varying nature of WDN nodal heads,
the irregular topologies and complex hydraulic relationships of a WDN GHR-S estimates the very low-frequency parts of heads as pseudo
can lead to inconsistencies with the assumption that closer nodes have measurements, which provides additional constraints to unmeasured
similar heads. Also, the variations of nodal heads with respect to their nodes to avoid the occurrence of unreasonable results and ensure the
distance are not always uniform (Zhou et al., 2022). Thus, the data stability of the reconstruction process. GHR-S is tested on two WDNs of
interpolation methods will have limited accuracy, especially for WDN different scales. The results show that it can provide accurate pressure
with irregular branches, unconnected intersections, and pipes with large estimation results without the requirement of precise hydraulic pa­
head loss. Machine learning (ML) is another promising technique that rameters nor specific locations of pressure meters, illustrating the con­
learns the hydraulic relationships to estimate unknown pressures. Lima venience and effectiveness of GHR-S in practical engineering.
et al. (2017) developed a multilayer perceptron (MLP) to learn from
numerical simulated data and successfully estimated unknown pressures 2. Method
of two small-scale WDNs. Limited by the learning capability of MLP for
complex relationships, its application to larger WDNs could be ineffi­ The method presented in this study builds on and improves the work
cient. Hajgató et al. (2021) investigated the performance of graph neural by Zhou et al. (2022). In the following sections, the basic concepts of
network for larger WDNs and achieved high accuracy. Despite the po­ original GHR will firstly be introduced, followed by the framework of
tential of ML methods, training an ML model needs to perform a large GHR-S, which includes a modified graph signal reconstruction algorithm
number of hydraulic simulations with different scenarios, and the un­ considering the presence of measurement noise, and a two-stage
certainties in the hydraulic parameters may still propagate to the trained reconstruction method to estimate unknown pressures with improved
model. Besides, experienced setting of model parameters (hyper-­ stability.
parameters) and pre-processing of training data are often required when
developing an effective ML model, which also limits its large-scale
2.1. Graph-based head reconstruction
application.
Efficient analysis of interrelated relationships and spatial distribu­
The topology of a WDN can be represented by a graph G = (V ,E ),
tion patterns of nodal heads could be the key to accurate and easy-to-
where V = {1, 2, …, N} is a set of N vertices, representing nodes, tanks
implement pressure estimation. The major challenge lies in the inho­
and reservoirs in the WDN, and E = {wij }i,j∈V is a set of edges con­
mogeneous hydraulic relationship and irregular topology of a WDN.
necting the nodes, representing pipes, pumps and valves. wij > 0 if there
Romero et al. (2021) introduced a graph-based nodal head interpolation
is a link between node i and j, and wij = 0 otherwise. The adjacency
method, which calculated the head of a certain node as a weighted linear
contribution of connected neighboring nodes. Although the topology matrix is defined as W(i, j) = wij , and the Laplacian matrix is L =
information is considered, this study simply assumed that the relation (i. diag(1T W) − W, where 1 is a vector of all ones, and diag converts a
e., weight) of two connected nodes is inversely proportional to the pipe vector to a diagonal matrix using the vector as its main diagonal. The
length between the nodes, which has limited ability to reveal the hy­ Eigen decomposition of L is L = UΛUT , where U = [u1 , u2 , …, uN ] is an
draulic relationships of a WDN. Besides, the method provides less in­ orthogonal matrix consisting of eigenvectors, and Λ is a diagonal matrix
formation to the global variation of pressures thus large estimation consisting of eigenvalues. The nodal heads of a WDN can be represented
errors may occur in nodes far away from known measurements. Recent by signals defined on the graph vertices, i.e., H = [H1 , H2 , …, HN ]T ,
developments in graph signal processing (GSP) (Lorenzo et al., 2018; where Hi is the head of node i.
Stanković et al., 2019) provide better tools to interpret the hydraulic GHR is based on graph Fourier transform (GFT), which transforms
relationship and topology of a WDN. GSP can represent WDN topologies the original graph signal into a set of graph signals with different varying
by vertices and links on a graph, and treats nodal pressures as signals on speeds (i.e., frequency). For H defined on the vertices, its varying speed

2
X. Zhou et al. Water Research 233 (2023) 119747

on the graph can be measured by the value difference (i.e., the change of heads given by GHR.
nodal heads) between any two connected nodes, given by A simple example of GHR is given in Fig. 1. The original heads of the
10 nodes of the example WDN are transformed into a combination of

N ∑
N
( )2
E(H) = wij Hi − Hj (1) Laplacian spectrums H and eigenvectors U. The former eigenvectors,
i=1 j=1 such as u1 and u2 , show slow varying speeds on the graph, while the
latter eigenvectors such as u9 and u10 vary fast. The latter values of H
where E(H) is the frequency of signal H, larger E(H) denotes faster are very small and neglecting them will have limited impacts. The
varying speed of the signal. retained former values can then be calculated from field measurements
For an eigenvector ui , it is proved (Stanković et al., 2019) that to reconstruct the original nodal heads using Eqs. (8) and (9). The
boundary dividing the retained and the neglected parts is determined by
E(ui ) = 2λi (2)
F .
where λi is the eigenvalue corresponding to ui . An eigenvector with The basic assumption of GHR is that the spatial distribution of nodal
larger eigenvalue will show higher frequency. If the eigenvalues are heads is ‘smooth’, with low-frequency components containing the ma­
arranged in ascending order, then the former eigenvectors such as u1 jority of information. However, in most of WDNs, the pressure changes
and u2 will have low frequency, while the latter eigenvectors such as among nodes are not uniform, and the assumption cannot be strictly
uN− 1 and uN vary fast. As U is orthogonal, it spans a graph frequency satisfied, which will introduce errors during the reconstruction process.
domain RN , of which the eigenvectors are the basis. The orthogonal GHR solves this problem by assigning specific link weights to enhance
mapping of the original signal to the graph frequency domain is the smoothness. As defined in Eq. (1), the varying speed of the original
signal H depends on both the nodal signal differences and the weights of
H = U− 1 H (3) the corresponding links. For adjacent nodes i and j that may have large
signal difference, assigning smaller wij can reduce the overall varying
∑ speed of H (i.e., E(H)) and improve its smoothness on the graph. Four
N
H = UH = ui H (4)
different strategies to determine the weights under different availability
i
i=1
of WDN information were proposed by Zhou et al. (2022). For a link that
where H = [H 1 , H 2 , …, H N ] is called Laplacian spectrum, and Eqs. (3) connects vertex i and j, the basic strategy to calculate its weight can be
and (4) are GFT and inverse graph Fourier transform (IGFT) of the written as
original signal, respectively. As shown in Eq. (4), the original signal H ⎧
⎪ 1 ⃒ ⃒
can be decomposed into a linear combination of the eigenvectors ui . ⎪ ⃒⃒ ⃒⃒, ⃒hij ⃒ ≥ ε1

⎨ hij
GHR assumes that the spatial distribution of WDN nodal heads is wij = wji = (10)
approximately ‘smooth’ with slow varying speed. Thus, the former ⎪
⎪ 1 ⃒⃒ ⃒⃒

⎩ , hij < ε1
values of H such as H 1 and H 2 corresponding to the low-frequency ε1
components will contain the majority of information, while the latter
values contain little information and can be neglected with acceptable where hij is the head loss of pipe ij, and ε1 is a constant used to avoid
accuracy loss. On this basis, the original signal can be approximated by unreasonable values, set as 10− 3 in our study. The calculated weights
its low-frequency parts Hl , i.e., will be further normalized with a min-max normalization:
wij − min(w)
∑ (11)
F
wij = + ε2
H ≈ Hl = UF H F = ui H i (5) max(w) − min(w)
i=1

where min() and max() are the minimum and maximum of the given set,
where F is an integer called frequency limit, representing the number respectively; w is a set containing all the calculated weights, ε2 is a small
of retained low-frequency components. UF ∈ RN×F contains the first F constant to avoid the occurrence of zero value, set as 10− 8 in this study.
columns of U, and H F contains the first F values of H . The choice of For valves and pumps, a very small value will be assigned to their cor­
F will be further discussed in the following sections. As the Laplacian responding weights, indicating the rapid change of nodal heads on the
matrix L is sparse and symmetric, for large WDNs, we only need to solve adjacent nodes. For simplicity, we also use w to represent the normalized
the first F eigenvectors to construct UF to speed up the calculation, link weights in the following discussions.
which is achieved by ARPACK (Lehoucq et al., 1998) in this study. The link weights act as a bridge between graph signal processing
The pressure measurements can be regarded as a sample of original techniques and WDN hydraulic relationships. The calculation of w in­
signal H, i.e., volves hydraulic simulation to calculate pipe head loss, which integrates
y s = Ds H (6) various hydraulic information of the WDN such as pipe length and
diameter, pipe roughness, nodal demands, etc. Despite the requirement
where ys is a vector of measurements, Ds ∈ RM×N is a sampling matrix, of hydraulic simulation to calculate w, the work by Zhou et al. (2022)
and M is the number of meters. If pressure meter i is installed at node j, and the results in our study will show that GHR and GHR-S have good
then Ds (i, j) = 1, and zeros for other elements. Combining Eqs. (5) and robustness to model parameter uncertainties, and accurate results can be
(6), we will get achieved without careful parameter calibration.
Even when there is no ready-to-use hydraulic model, with certain
y s = Ds UF H (7)
F
assumptions, we can still exploit available WDN information as much as
H F can then be estimated by possible to roughly estimate link weights. For example, in the absence of
( )− 1 T T a hydraulic model, if basic information such as pipe length, diameter
̂F
H = UTF DTS Ds UF UF DS y s (8) and WDN topology is available, with an assumption that all the pipe
Combining Eqs. (5) and (8), the reconstructed low-frequency parts of roughness coefficients are the same and the user demands are uniformly
the original signal are distributed along water supplying pipes, a rough temporary model can
be constructed to calculate the pipe head losses and weights in Eq. (10).
( )
̂ l = UF UT DT Ds UF − 1 UT DT ys
H F S F S (9) The weights calculated by this strategy are denoted as w(uQ) herein.
When the available data of a WDN is insufficient to support hydraulic
the corresponding elements in H
̂ l are the estimated unknown nodal simulations, by assuming that the velocity and roughness of all pipes are

3
X. Zhou et al. Water Research 233 (2023) 119747

Fig. 1. Schematic diagram of GHR.

uniform, with an approximation of the Hazen-Williams formula, we can reconstruction process. As shown in Fig. 2, in the first stage, a rough
also roughly estimate link weights without hydraulic simulation by estimation of unknown nodal heads will firstly be reconstructed with a
small enough number of low-frequency components which can ensure
D1.166
(uv)
wij
(uv)
= wji =k
ij
(12) reconstruction stability. The rough estimations are regarded as pseudo
Lij measurements for all the unknown nodes. After that, the pseudo mea­
surements and the field measurements will be combined, where the
where wij is the roughly estimated weight assuming uniform velocities,
(uv)
values on nodes with installed meters are field measurements, while the
and k is a constant that can be neglected in the normalization process. values on the unknown nodes are pseudo measurements. In the second
The set of normalized weights can be denoted as w(uv) . In the case that stage, the combined measurements are used to reconstruct unknown
very limited WDN information is available, a very low-cost way is to set nodal heads again, with a large enough number of low-frequency com­
all link weights as uniform, denoted as w(u) = [1, 1, …, 1]T . More detailed ponents to fully reveal the spatial variation of original nodal heads. The
explanations and comparisons of the weight calculation strategies can be using of pseudo measurements provides additional constraints to avoid
found in supplementary materials and Zhou et al. (2022). unreasonable results on unmeasured nodes, thus improving stability.
Pseudo measurements should be assigned with lower confidence (i.e.,
weights) than field measurements to avoid their inherent errors
2.2. Improving the stability of GHR
affecting the accuracy of final reconstruction results.
Specifically, in the first stage, the pseudo measurements Hp can be
Solving Eq. (8) requires the matrix Ds UF to be full column rank, i.e.,
calculated by:
the rank of matrix Ds UF should be equal to F Horn and Johnson,
2012). A necessary condition is M ≥ F , i.e., the number of meters ( )− 1
Hp = UF p UTF p DTS Ds UF p UTF p DTS ys (13)
should be greater than the number of retained low-frequency compo­
nents. Nevertheless, even with sufficient meters, Ds UF could still be
rank-deficient or ill-conditioned (Lorenzo et al., 2018) depending on the where F p < F is a small integer, which determines the number of low-
WDN hydraulics and the location of pressure meters, which could lead to frequency components used to calculate pseudo measurements. F p is
failure of the reconstruction process or very large estimation errors. Eqs. set to 6 by default in this study. Detailed discussions about the choosing
(8) and ((9) can be interpreted as solving H F to minimize the difference of F p are presented in Fig. 6.
between Ds Hl (the reconstructed heads at the metered nodes) and ys . Before the second stage, ys and Hp will be integrated as combined
With large F and non-optimized locations of meters, the reconstruction measurements Hc . Field measurements are assigned to measured nodes
process may tend to reduce the error at the measured nodes while in Hc , and pseudo measurements are assigned to unmeasured nodes:
ignoring the unconstrained nodes, thus an ‘overfitting’ problem will
Hc = DTs ys + Hp − DTs Ds Hp + v (14)
appear, and the estimation error of some unmeasured nodes could lose
control. Besides, the problem can be exacerbated by uncertainties where v is a noise vector, which is assumed to be zero-mean with
arising from the smoothness assumption and measurement errors. Using mutually independent elements. The covariance matrix of v is R =
more meters and small enough F helps to increase the stability, but this
will also lead to higher construction costs and lower estimation accuracy
as too many frequency components are neglected. Choosing appropriate
locations of meters according to experimental design theory such as
D-optimality (Winer, 1962; Tsitsvero et al., 2016; Zhou et al., 2022) is
another way to enhance the stability, which is adopted by Zhou et al.
(2022). However, for a majority of WDNs, it is difficult to ensure that all
the meters are installed in the desired locations.
The slow-varying nature of WDN heads provides us with a new
perspective to improve the stability of the signal reconstruction process.
Even with very few low-frequency components, the results by GHR can
still show the general trend and approximate range of the original heads,
which can be used as a reference to restrict the occurrence of unrea­
sonable results during the reconstruction process. On this basis, a two-
stage framework is proposed to enhance the stability of the Fig. 2. The two-stage framework of GHR-S.

4
X. Zhou et al. Water Research 233 (2023) 119747

E{vvT }, where R is a diagonal matrix. In this study, if a meter is installed systematic and random error of Cw, respectively, while the noises
at node i, then R(i,i) = 1, otherwise R(i,i) = σ2v , where larger σv denotes assigned to Q follow μQ = 5% and σ Cw = 20% of the original demands,
smaller confidence of pseudo measurements. The choosing of σv will be respectively. ‘Real nodal heads’ are generated by hydraulic simulation
further discussed in Fig. 6. In the second stage, Hc will be regarded as using the EPANET2 toolkit (Rossman, 1999) and WNTR package (Klise
measurements that contain different levels of noise. et al., 2018). Fig. 3b shows the model with assigned parameter un­
In order to solve H F in the presence of measurement noises, this certainties, where the pressures at metered nodes are set as measure­
study improves the calculation strategy of GHR using the best linear ments, and pressures at other nodes are ‘ground truth’ to evaluate the
unbiased estimator (BLUE) (Kay, 1993), given by: performance pressure estimation results. (2) Other basic information
( )− 1 such as pipe length, diameters, and network topology are assumed to be
( ) ( )− 1
̂ F = UT DT Ds RDT − 1 Ds UF
H F S S UTF DTS Ds RDTS Hc (15) known and accurate. (3) The original model rather than the ‘real-life
network’ is used to calculate graph weights (w) of GHR-S, thus the
The using of BLUE enables the reconstruction process to take into calculation of w will be affected by the assigned parameter uncertainties.
account the confidence of different measurements. Besides, as the The results of other weight calculation strategies (w(uQ) , w(uv) , and w(u) )
combined measurements have values on all nodes (i.e., Hc ∈ RN ), the will be discussed in Supplementary Materials. (4) the locations of
measurement matrix Ds of Hc is an identity matrix, thus Eq. (15) can be pressure meters are randomly determined by a roulette algorithm,
simplified as: where nodes adjacent to longer pipes will have a larger probability to be
( ) chosen. (5) Unless otherwise specified, the default parameters of GHR-S
̂ F = UT R− 1 UF − 1 UT R− 1 Hc
H (16)
F F are F = 0.7M, F p = 6, and σ v =10. (6) The accuracy of estimation re­
sults is measured by mean absolute error (MAE):
and the reconstructed heads are:
( ) 1 ∑ Nu
̂ l = UF UT R− 1 UF − 1 UT R− 1 Hc
H F F (17) MAE = ̂ i − Hi |
|H (18)
Nu i=1
As all the nodes are assumed to be ‘measured’ in the second stage,
Eq. (16) can be solved as long as F ≤ N, but too large F will introduce where H ̂ i is the estimated head of node i, Nu is the number of nodes to be
unnecessary computational demand and may still lead to unstable re­ estimated.
sults. Therefore, in this study, we set F = 0.7M (i.e., 70% of the number The performance of GHR-S is firstly tested with pressure measure­
of available meters) to balance computational demand, stability, and ments collected from 4 reservoirs and 30 randomly chosen meters (the
accuracy. The impact of different F on the results is also discussed in locations are shown in Fig. 3a as triangles). To illustrate the calculation
Fig. 7. process of GHR-S, Fig. 3c presents the pseudo measurements Hp calcu­
To sum up, GHR-S includes the following steps: (1) choose suitable lated in the first stage of GHR-S, and the differences between pseudo
F p and calculate the pseudo measurements Hp , (2) combine Hp with measurements and real nodal heads are shown in Fig. 4a. It can be
field measurements ys to obtain Hc , and (3) solve the reconstructed observed that the general trend of pseudo measurements is consistent
nodal heads H ̂ l with Eq. (17). with the original heads, but differences exist in local variations. Despite
the local errors, pseudo measurements still provide an approximate
3. Case studies and results range of unknown nodal heads, which avoids the occurrence of unrea­
sonable results in the second stage of GHR-S.
3.1. Case 1 Using the default parameter settings, GHR-S shows good estimation
results (MAE = 0.40 m) for most of the unknown nodes, as shown in
In this case, GHR-S is tested on a real-life middle-sized WDN with Fig. 4b. Only 32 nodes, which are mainly distributed at branches or ends
synthetic data to illustrate its effectiveness and make comparisons with of the WDN, have an estimation error larger than 1 m. For comparison
other pressure estimation methods. The WDN consists of 480 nodes, 4 purposes, head estimation results by GHR, Kriging interpolation (Li and
reservoirs, and 567 pipes, and supplies 57,000 m3 of water to a 62 km2 Heap, 2008), and hydraulic model with parameter uncertainties are also
area per day, as shown in Fig. 3a. During the tests, the following as­ given in Fig 4c-f. The performance of GHR is significantly affected by the
sumptions and default settings are made: (1) The parameters in the choosing of F as the locations of meters do not conform with D-opti­
original hydraulic model (Fig. 3a) are assigned with considerable un­ mality. It achieves the smallest MAE when F = 12, as shown in Fig. 4c.
certainties to reproduce a ‘real-life network’. Gaussian noises N (μ, σ 2 ) Larger F will lead to unstable results. For example, when setting F =
are assigned to pipe roughness coefficients (Hazen-Williams coefficients, 24 (same as GHR-S in Fig. 4b), GHR will give unreasonable results with
Cw) and nodal demands (Q), as these parameters are most difficult to be MAE = 351 m. Obviously, GHR needs to use a small enough F to ensure
precisely determined in a hydraulic model (Zhou et al., 2018). Both its stability with the randomly chosen meter locations, but this will also
systematic errors and random errors are considered. The default noises lead to insufficient frequency components to reveal detailed changes,
assigned to Cw follow μCw = − 5 and σCw = 10, representing the thus the accuracy is limited. A more detailed comparison of GHR-S and

Fig. 3. (a) The original hydraulic model. (b) The model with assigned parameter uncertainties, which are used to reproduce real-life WDN. (c) Pseudo measurements
in the first stage of GHR-S.

5
X. Zhou et al. Water Research 233 (2023) 119747

Fig. 4. The nodal head estimation errors by different methods. (a) The error of pseudo measurements in the first stage of GHR-S. (b-f) Estimation errors of GHR-S,
GHR with F = 12, GHR with F = 24, Kriging interpolation, and hydraulic model with parameter uncertainties, respectively.

GHR with respect to F will be given in Fig. 7. Kriging interpolation locations and measurement errors. It is obvious that using more meters
infers unknown nodal heads according to their spatial distance to will improve the performance of GHR-S, as more information will be
measured nodes, and the results are shown in Fig. 4e. As it ignores the given. Besides, as the locations of meters are randomly chosen during
irregular topologies and complex hydraulic features of a WDN, large the tests, with the limited number of meters, there will be no meter in
errors may occur, especially for nodes at branches of ends of the WDN. some parts of the WDN, and large estimation errors will occur. In
Therefore, Kriging interpolation methods show poorer accuracy even practical applications, adding some meters in a few key points can
compared with GHR using limited number of low-frequency compo­ effectively improve accuracy. Although larger measurement errors will
nents. Fig. 4f uses the original hydraulic model to simulate the ‘real-life decrease the estimation accuracy, as GHR-S only retains the low-
network’, which illustrates the error of hydraulic simulation with frequency components of the nodal heads, it can filter out part of the
parameter uncertainties (μCw = − 5, μQ = 5%, σCw = 10, and σ Q = random errors that lie in the measurements, showing a good ability to
20%). The pressures of fixed-head nodes (i.e., reservoirs) are set as resist the impact of errors. For example, increasing σ H from 0 m to 0.1 m
known. With the increasing water supply distance, the model simulation has no obvious effect on the accuracy; even when σH is increased to 0.5
error cumulates, and the accuracy of nodes at the ends of the WDN m, the overall estimation accuracy will decrease no more than 0.2 m.
significantly decrease. Comparatively, although GHR-S uses the model F p and σv are used to calculate pseudo measurements and determine
with the same uncertainty to calculate w, it integrates information from their weights in the first stage. In Fig. 6, the impacts of the two pa­
both hydraulic relationships and field measurements, thus it achieves rameters are analyzed using one-at-a-time sensitivity analysis, i.e., the
better accuracy and avoids complex and experience-intensive model parameter analyzed will be changed to different values in turn, and
calibration processes. other parameters will remain at their default values. Each setting of
As GHR-S solves low-frequency components according to field parameters will be repeatedly performed 50 times with random loca­
measurements, the data quality and quantity of measurements will show tions of 30 meters. From the results, we can find that GHR-S can
significant impacts. Different numbers of meters and different levels of maintain good performance over a wide range of parameters, such as
measurement errors are tested, and the results are shown in Fig. 5. The 3 ≤ F p ≤ 7 and 10 ≤ σ v ≤ 50. Too large F p will reduce the accuracy
measurement errors are assumed to be zero-mean Gaussian noises with because the stability during the calculation of pseudo measurements
standard deviations (σ H ) ranging from 0 m to 0.5 m. Each test is cannot be guaranteed. Although larger F p could be tolerated in large-
repeatedly performed 50 times with regenerated random meter size WDNs and/or when using more meters, smaller values such as
3~7 are still suggested as this ensures both stability and accuracy. As
pseudo measurements are only rough estimations with certain errors,
overemphasizing them with a small σ v will interfere with the informa­
tion of field measurements, thus the accuracy decreases. On the con­
trary, too large σ v will lead to insufficient weights of pseudo
measurements, thus less constraint can be provided to avoid unstable
results in the second stage of GHR-S. For larger WDNs with more nodes,
due to the higher ratio of pseudo measurements, a proper increase of σv
is recommended to balance the information between pseudo measure­
ments and field measurements.
Benefiting from the use of pseudo measurements, GHR-S can keep
high stability over a wide range of F . Fig. 7 presents a comparison of
GHR-S and GHR with F ranging from 8 to 30. Each test is repeatedly
performed 50 times with 30 randomly chosen meters. The overall results
are shown in Fig. 7a. Obvious estimation errors appear in the results of
GHR as long as F is larger than 10, and the estimation accuracy will
Fig. 5. The estimation accuracy of GHR-S under different number of meters become even worse with the increase of F . Comparatively, GHR-S can
and measurement errors. achieve accurate and stable results with F ranging from 13 to 30, and

6
X. Zhou et al. Water Research 233 (2023) 119747

Fig. 6. The estimation accuracy of GHR-S with different values of F p and σv .

Fig. 7. A comparison of GHR-S and GHR with different values of F . (a) The overall results of 50 tests. (b) The best results in 50 tests.

the MAEs for most of the tests are smaller than 1 m despite the optimality) are also shown. Although GHR can achieve very good ac­
randomness of meter locations, which shows better stability with arbi­ curacy (MAE=0.31 m, F = 29) with optimized meter locations, in
trary choosing of meter locations and F . The best results in the 50 tests practical engineering, it is often hard to install pressure meters in full
can represent the performance of the tested methods with relatively accordance with D-optimality. With random meter locations, even for
better meter locations, which are shown in Fig. 7b. For comparison the best results in 50 tests, GHR can only achieve MAE=0.49 m, F =
purpose, the performance of GHR with optimized meter locations (D- 18, and inappropriate choosing of F (e.g., F > 23) will still lead to

Fig. 8. Pressure estimation results by (a) GHR-S and (b) Hydraulic model simulation.

7
X. Zhou et al. Water Research 233 (2023) 119747

large estimation errors. The best accuracy achieved by GHR-S in the 50 requirements in the accuracy of hydraulic parameters such as nodal
tests is MAE=0.36 m, which is very close to the performance of GHR demands and pipe roughness coefficients, and enables engineers to
with optimized meter locations. Therefore, GHR-S shows a superior conveniently estimate unknown pressures with reduced investments in
convenience in practical engineering as it can achieve high estimation model construction and maintenance. GHR-S also has superiority over
accuracy with less requirement of meter locations and parameter traditional spatial interpolation methods, as it can better handle the
choosing. When GHR-S is applied to other WDNs, a larger proportion of irregular topology and complex hydraulic relationships of WDNs to
F is recommended when the number of meters is small (e.g., we can achieve higher accuracy.
set F = M when M<15), as this will ensure a sufficient number of low- Although the application of GHR-S does not rely on specific locations
frequency components. On the contrary, the proportion of F should be of meters, it is still recommended that the meters are distributed evenly
decreased with a large number of meters (e.g., F = 0.3M ∼ 0.6M when among the network with an appropriate focus on the end nodes and
M>100), as too many low-frequency components will benefit little to important users. Besides, if a majority of pressure meters can be installed
the reconstruction accuracy and may lead to unstable results. in designated locations, using GHR with optimized meter locations can
achieve better accuracy. GHR-S is especially suitable for WDNs with
lower flow velocity over the pipes, as the pressure change of these WDNs
3.2. Case 2 is relatively slow, and the smoothness assumption will result in less
accuracy loss. For these WDNs, using w(uv) and w(u) as weights would
In this case, a large real-life WDN with field measurements is used to able to achieve acceptable accuracy with very low cost. For WDNs with
validate the effectiveness of GHR-S in practical engineering. The WDN larger pipe flow rates and rapidly changing nodal pressures, using more
lies in eastern China and supplies 330,000 m3 of water per day to a 611 meters will help to ensure accuracy.
km2 area with a population of 1.09 million. It contains 13 reservoirs, A hydraulic model is still the best choice when there are sufficient
870 km pipes with diameters over 100 mm, and more than 83,000 investments to carefully calibrate model parameters and maintain model
nodes, as shown in Fig. 8a. Pressures were collected from 62 meters with accuracy, as hydraulic simulation can provide integrated results
pre-determined locations. As ‘ground truth’ is not known for unmetered including various hydraulic information with high interpretability.
nodes, the 58 pressure measurements are selected to reconstruct the However, for a number of water supply companies, ready-to-use accu­
nodal heads, and the remaining four meters are used to validate the rate hydraulic models are still absent, and investments in model con­
reconstructed results. The parameters of GHR-S in this case are: F p = 6, struction and maintenance could be unaffordable. Under the
F = 0.7M = 41, which are the same as in case 1; the ratio of unknown circumstances, GHR-S provides a low-cost, effective, and accurate
nodes to measured nodes is about 100 times higher than case 1, so σ2v is alternative way to estimate pressures. Besides, other weight calculation
correspondingly increased, set as σv = 100. A hydraulic model is strategies (w(uQ) , w(uv) , and w(u) ) can also be used as a more low-cost way,
available, but the model parameters are not carefully calibrated. The which will decrease the accuracy but still performs better than tradi­
pipe roughness coefficients of the model are all set to 120. The nodal tional data interpolation methods. Detailed discussions about different
demands of some large users are determined according to real-time strategies to determine graph weights can be found in Supplementary
remote water meters, and the remaining demands are assumed to be Materials and Zhou et al. (2022). It should be noted that GHR-S esti­
uniform along water supplying pipes. The model simulation results are mates unknown pressures of a WDN in a holistic manner, thus it can
also given as a comparison, and the results are shown in Fig. 8. hardly provide reliable information for subtle hydraulic states such as
As shown in Fig. 8a, GHR-S successfully reconstructs the unknown the head loss of a certain pipe. It is also difficult for GHR-S to simulate an
nodal heads for more than 83,000 nodes with the 58 measurements. The operating status with a hypothetic hydraulic condition, as its estimation
reconstruction errors at the 4 validate nodes are 0.26 m, 0.74 m, 0.21 m is based on measurements from existing operating status.
and − 0.17 m, respectively, which shows a better performance compared As an easy-to-implement way to estimate unknown pressures, GHR-S
with the results given by hydraulic model simulation results (0.14 m, can contribute to smarter and more efficient management of WDNs in
1.86 m, 1.08 m and 0.73 m, respectively, as shown in Fig. 8b). The better many ways. A simple example is the application in supervisory control
accuracy of GHR-S contributes to its combination of hydraulic re­ systems. Estimated pressures at unknown nodes can help decision-
lationships and field measurements, and the using of pseudo measure­ makers to obtain more comprehensive and visualized information
ments ensures the stability with arbitrary locations of meters. Besides, about WDN operating status (Tufa and Abate, 2022). GHR-S can also
GHR-S represents the complex WDN topologies and hydraulic relation­ provide key basic data for various intelligent technologies in the
ships by graph structure and specific graph weights, which enables it to research field of WDNs. For example, analyzing the spatial variation of
accurately reveal rapid regional pressure changes (for example, the re­ pressures during abnormal events is a commonly used methodology to
gions near the three reservoirs in the center left of the WDN, marked as detect and localize anomalies (Marzola et al., 2022; Wang et al., 2022),
District 1), and nodal heads with irregular topological structures (for which can be easily combined with GHR-S. Comprehensive knowledge
example, the nodes lie in the lower left of the WDN that are nearby but of pressures can also help in formulating pressure management schemes
not connected with each other, marked as District 2). Unknown pres­ (Olsson, 2020; Price et al., 2022), which will decrease energy con­
sures of nodes in these districts can hardly be accurately estimated by sumption and background leakage of WDNs. Besides, the estimated
traditional interpolation results. pressures will provide additional information on WDN operation status,
In this case, running GHR-S consumes 131.1 s on our testing com­ which may help to solve the under-determination of parameter cali­
puter (AMD Ryzen 9 5900HS CPU 3.30 GHz, 16GB memory), which is bration problems (Chen et al., 2022).
applicable for most online models with a 5 min to 1 h update frequency.
More details about the network and the performance of GHR-S with 5. Conclusions
synthetic data can be found in Supplementary Materials.
The proposed GHR-S provides a convenient, stable, and accurate
4. Discussions method to estimate unknown pressures in WDNs. GHR-S is based on the
GHR method presented by Zhou et al. (2022), which divides nodal
GHR-S provides a convenient and low-cost way to estimate unknown pressures into components with different spatially varying speeds (i.e.,
WDN pressures, as it can achieve good accuracy without careful cali­ frequencies), and uses the low-frequency parts to reconstruct the orig­
bration of hydraulic parameters. It improves GHR by allowing arbitrary inal nodal heads with an assumption that the original heads mainly
pressure meter locations with better stability. Compared with simulating consist of slowly varying components. To improve the stability of GHR
nodal pressures by hydraulic models, GHR-S has much lower and enable wider application, a two-stage nodal head reconstruction

8
X. Zhou et al. Water Research 233 (2023) 119747

framework is employed in GHR-S, which utilizes pseudo measurements Dai, P.D., Cuong, L.Q., Van Dai, B., 2018. Optimal pump scheduling to pressure
management for large-scale water distribution systems. In: AETA 2017 - Recent
to restrict the occurrence of unreasonable results during the head
Advances in Electrical Engineering and Related Sciences: Theory and Application.
reconstruction process. GHR-S is tested on two cases, and the key find­ https://doi.org/10.1007/978-3-319-69814-4_51.
ings are summarized below. Du, K., Ding, R., Wang, Z., et al., 2018. Direct inversion algorithm for pipe resistance
coefficient calibration of water distribution systems. J. Water Resour. Plann. Manag.
144 (7), 04018027 https://doi.org/10.1061/(ASCE)WR.1943-5452.0000948.
(1) GHR-S can achieve high accuracy without the requirement of Hajgató, G., Gyires-Tóth, B. and Paál, G., 2021. Reconstructing nodal pressures in water
precise hydraulic parameters or specific meter locations. It contrib­ distribution systems with graph neural networks. arXiv:2104.13619.
utes a new low-cost solution for pressure estimation problems, Horn, R.A., Johnson, C.R., 2012. Matrix Analysis. Cambridge University Press.
Hutton, C.J., Kapelan, Z., Vamvakeridou-Lyroudia, L., Savic, D.A., 2014. Dealing with
especially for large complex WDNs which would require high in­ uncertainty in water distribution system models: a framework for real-time modeling
vestments to build accurate hydraulic models. and data assimilation. J. Water Resour. Plann. Manag. 140 (2), 169–183. https://doi.
(2) GHR-S is a significant improvement over GHR. The using of org/10.1061/(ASCE)WR.1943-5452.0000325.
Kay, S.M., 1993. Fundamentals of Statistical Signal Processing. Prentice Hall PTR, Upper
pseudo measurements enables the graph-based pressure estimation Saddle River, New Jersey.
methods to be applied to most WDNs with arbitrary locations of Klise, K.A., Murray, R., Haxton, T., 2018. An overview of the water network tool for
meters, which significantly extends the applicability of these resilience (WNTR). In: 1st International WDSA /CCWI 2018 Joint Conference,
Kingston, Ontario, Canada.
methods. Lehoucq, R., Sorensen, D., Yang, C., 1998. ARPACK USERS GUIDE: Solution of Large
(3) The successful application in a large complex real-life network Scale Eigenvalue Problems by Implicitly Restarted Arnoldi Methods. SIAM,
with field measurements illustrates the effectiveness of GHR in Philadelphia, PA. https://doi.org/10.1137/1.9780898719628.
Li, J., Heap, A.D., 2008. A Review of Spatial Interpolation Methods for Environmental
practical engineering.
Scientists. Geoscience Australia, Australia.
Lima, G.M., Brentan, B.M., Manzi, D., Luvizotto Jr., E., 2017. Metamodel for nodal
The pressure estimation results of GHR-S can further support various pressure estimation at near real-time in water distribution systems using artificial
intelligent technologies for WDN management such as pressure man­ neural networks. J. Hydroinform. 20 (2), 486–496. https://doi.org/10.2166/
hydro.2017.036.
agement and leakage detection, and combinatory studies with these Lorenzo, P.D., Barbarossa, S., Banelli, P., 2018. Sampling and recovery of graph signals.
technologies are recommended. Besides, the using of graph signal pro­ In: Djurić, P.M., Richard, C. (Eds.), Cooperative and Graph Signal Processing.
cessing methods in WDNs also provides a new perspective for solving Academic Press, pp. 261–282.
Marzola, I., Mazzoni, F., Alvisi, S., Franchini, M., 2022. Leakage detection and
similar problems in the field of water resources and environments, such localization in a water distribution network through comparison of observed and
as water quality estimation, analysis of pollutant transmission, and simulated pressure data. J. Water Resour. Plann. Manag. 148 (1), 04021096 https://
precipitation analysis. Although pseudo measurements are calculated by doi.org/10.1061/(ASCE)WR.1943-5452.0001503.
Mounce, S.R., Khan, A., Wood, A.S., Day, A.J., Widdop, P.D., Machell, J., 2003. Sensor-
reconstructing original heads from a limited number of low-frequency fusion of hydraulic data for burst detection and location in a treated water
components in this study, other strategies such as forecasting technol­ distribution system. Inf. Fusion 4 (3), 217–229. https://doi.org/10.1016/S1566-
ogies can be investigated to improve the performance of GHR-S. 2535(03)00034-4.
Olsson, G., 2020. Urban water supply automation – today and tomorrow. J. Water Supply
70 (4), 420–437. https://doi.org/10.2166/aqua.2020.115.
Price, E., Abhijith, G.R., Ostfeld, A., 2022. Pressure management in water distribution
Declaration of Competing Interest systems through PRVs optimal placement and settings. Water Res. 226 https://doi.
org/10.1016/j.watres.2022.119236.
The authors declare that they have no known competing financial Romero, L., Puig, V., Cembrano, G., Blesa, J., Meseguer, J., 2021. A fully data-driven
approach for leak localization in water distribution networks. In: 2021 European
interests or personal relationships that could have appeared to influence Control Conference (ECC), Rotterdam, Netherlands. https://doi.org/10.23919/
the work reported in this paper. ECC54610.2021.9654928.
Rossman, L.A., 1999. The EPANET programmer’s toolkit for analysis of water
distribution systems. In: Proceedings of the 29th Annual Water. https://doi.org/
Data availability 10.1061/40430(1999)39.
Rossman, L.A., 2010. An overview of EPANET Version 3.0. Water Distribution Systems
Data will be made available on request. Analysis 2010, Tucson, Arizona, USA. https://doi.org/10.1061/41203(425)3.
Savic, Kapelan, Z.S., Jonkergouw, P.M.R., 2009. Quo vadis water distribution model
calibration? Urban Water J. 6 (1), 3–22. https://doi.org/10.1080/
15730620802613380.
Acknowledgements Soldevila, A., Blesa, J., Fernández-Cantí, R.M., Tornil-Sin, S., Puig, V., 2019. Data-driven
approach for leak localization in water distribution networks using pressure sensors
and spatial interpolation. Water 11, 1500. https://doi.org/10.3390/w11071500.
This work was financially supported by National Natural Science Stanković, L., Daković, M., Sejdić, E., 2019. Introduction to graph signal processing. In:
Foundation of China (Grant No. 52100113, 51879139) and China Stanković, L., Sejdic, E. (Eds.), Vertex-Frequency Analysis of Graph Signals. Springer
International Publishing, pp. 3–99. https://doi.org/10.1007/978-3-030-03574-7_1.
Postdoctoral Science Foundation (Grant No. 2021M691824). Vol. 1.
Tsitsvero, M., Barbarossa, S., Lorenzo, P.D., 2016. Signals on graphs: uncertainty
principle and sampling. IEEE Trans. Signal Process. 64 (18), 4845–4860. https://doi.
Supplementary materials
org/10.1109/TSP.2016.2573748.
Tufa, G., Abate, B., 2022. Assessment of accessibility and hydraulic performance of the
Supplementary material associated with this article can be found, in water distribution system of Ejere Town. J. Water Supply 71 (4), 577–592. https://
the online version, at doi:10.1016/j.watres.2023.119747. doi.org/10.2166/aqua.2022.012.
Wang, C., Xu, Q., Qiang, Z., Zhou, Y., 2022. Research on pipe burst in water distribution
systems: knowledge structure and emerging trends. J. Water Supply 71 (12),
References 1408–1424. https://doi.org/10.2166/aqua.2022.150.
Winer, B.J., 1962. Statistical Principles in Experimental Design. McGraw-Hill Book
Company, New York, NY, US. https://doi.org/10.1037/11774-000.
Boatwright, S., Romano, M., Mounce, S., Woodward, K., Boxall, J., 2018. Optimal Sensor
Xu, Q., Liu, R., Chen, Q., Li, R., 2014. Review on water leakage control in distribution
Placement and leak/burst localisation in a water distribution system using spatially-
networks and the associated environmental benefits. J. Environ. Sci. 26 (5),
constrained inverse-distance weighted interpolation. In: 13th International
955–961. https://doi.org/10.1016/S1001-0742(13)60569-0.
Conference on Hydroinformatics, Palermo, Italy.
Xu, W., Zhou, X., Xin, K., Boxall, J., Yan, H., Tao, T., 2020. Disturbance extraction for
Bozkurt, C., Firat, M., Ateş, A., 2022. Development of a new comprehensive framework
burst detection in water distribution networks using pressure measurements. Water
for the evaluation of leak management components and practices. J. Water Supply
Resour. Res. 56 (5) https://doi.org/10.1029/2019WR025526 e2019WR025526.
71 (5), 642–663. https://doi.org/10.2166/aqua.2022.031.
Zhang, C., Liu, H., Pei, S., Zhao, M., Zhou, H., 2022. Multi-objective operational
Chen, X., Zhou, X., Xin, K., et al., 2022. Sensitivity-oriented clustering method for
optimization toward improved resilience in water distribution systems. J. Water
parameter grouping in water network model calibration. Water Resour. Res. 58
Supply 71 (5), 593–607. https://doi.org/10.2166/aqua.2022.136.
https://doi.org/10.1029/2021WR031206 e2021WR031206.
Chu, S., Zhang, T., Yu, T., Wang, Q.J., Shao, Y., 2021. A noise adaptive approach for
nodal water demand estimation in water distribution systems. Water Res. 192,
116837 https://doi.org/10.1016/j.watres.2021.116837.

9
X. Zhou et al. Water Research 233 (2023) 119747

Zhou, X., Liu, S., Xu, W., Xin, K., Wu, Y., Meng, F., 2022. Bridging hydraulics and graph Zhou, X., Xu, W., Xin, K., Yan, H., Tao, T., 2018. Self-adaptive calibration of real-time
signal processing: a new perspective to estimate water distribution network demand and roughness of water distribution systems. Water Resour. Res. 54 (8),
pressures. Water Res. 217, 118416 https://doi.org/10.1016/j.watres.2022.118416. 5536–5550. https://doi.org/10.1029/2017WR022147.
Zhou, X., Tang, Z., Xu, W., et al., 2019. Deep learning identifies accurate burst locations
in water distribution networks. Water Res. 166, 115058 https://doi.org/10.1016/j.
watres.2019.115058.

10

You might also like