Professional Documents
Culture Documents
Multiobjective Optimization of Control Valve Installation and DMA Creation For Reducing Leakage in Water Distribution Networks
Multiobjective Optimization of Control Valve Installation and DMA Creation For Reducing Leakage in Water Distribution Networks
Abstract: This paper presents a novel methodology for optimizing simultaneously the installation of control valves and the creation
Downloaded from ascelibrary.org by Zhejiang University on 11/27/22. Copyright ASCE. For personal use only; all rights reserved.
of district metered areas (DMAs) in water distribution networks (WDNs). This methodology was developed through the multiobjective
approach, by considering, as decisional variables, the sites for control valve installation and isolation valve closure. The proposed algorithm
is based on the hybrid combination of three algorithms, a multiobjective genetic algorithm, which is entrusted with valve site search, and two
embedded algorithms, the first based on iterated linear programming (LP) and the second based on graph theory, aimed at searching for the
optimal settings of control valves and at partitioning the WDN into DMAs, respectively. The hybrid algorithm attempts to find optimal
solutions in the trade-off between the following objective functions to be optimized simultaneously: total installation cost, daily leakage
volume, and demand uniformity across DMAs. The applications to a small Lebanese WDN proved that the methodology can find, especially
for high values of the total installation cost, effective control valve installations, and isolation valve closures in terms of leakage abatement
while obtaining a uniform distribution of demands across DMAs. DOI: 10.1061/(ASCE)WR.1943-5452.0001114. © 2019 American Society
of Civil Engineers.
Author keywords: Multiobjective optimization; Genetic algorithms; Linear programming; Water distribution; Leakage; Pressure;
Control valve; District metered area (DMA); District; Partitioning.
Introduction locations and settings using a genetic algorithm (GA) and LP,
respectively. The algorithms proposed by Araujo et al. (2006),
When service pressure is higher than the minimum desired values Ali (2015), and Covelli et al. (2016) adopt GAs for both issues.
required for full demand satisfaction, water utilities often choose to However, Ali (2015) incorporated physical knowledge of the
install control valves. One of the main benefits of service pressure WDN to improve convergence toward optimal solutions. Liberatore
regulation is the abatement of leakage (Farley and Trow 2003). and Sechi (2009) proposed a two-step procedure for determining
Other benefits include reduction in the yearly number of pipe optimal valve location and control. In the first step, candidate sets
breaks and extension of the infrastructure life. for the location of valves are restricted to pipes defined based on
The optimal location of control valves for service pressure and hydraulic analysis. Metaheuristic scatter search routines were used
leakage reduction has been one of the most explored topics in the in the second step to identify the best solution in the location and
framework of water distribution network (WDN) management, control problems by optimizing a weighted multiobjective function
with numerous algorithms proposed over the years starting in the that considers the cost of inserting valves and the penalty for node
1990s. The algorithm proposed by Jowitt and Xu (1990) makes use pressures that do not meet the requirements. A harmony search
of iterated linear programming (LP) to search for the optimal con- approach was used by De Paola et al. (2017) to optimize both
trol valve settings that minimize leakage at each time slot of the control valve locations and settings. The algorithms of Pezzinga
day. Vairavamoorthy and Lumbers (1998) suggested replacing LP and Gueli (1999), Nicolini and Zovatto (2009), and Creaco and
with sequential quadratic programming while formulating an ob- Pezzinga (2015a, b) differ from those just described in that they
jective function that allows minor violations in the targeted pressure were developed using a multiobjective approach. In fact, they en-
requirements. Reis et al. (1997) tackled the optimization of valve able constructing Pareto fronts of optimal solutions in the trade-off
between number of control valves, as a surrogate for the installation
1
Associate Professor, Dipartimento di Ingegneria Civile e Architettura, cost, and daily leakage. Nicolini and Zovatto (2009) used a multi-
Univ. of Pavia, Via Ferrata 3, Pavia 27100, Italy; Honorary Senior Research objective GA to optimize both valve locations and control settings.
Fellow, College of Engineering, Physical and Mathematical Sciences, Univ. Pezzinga and Gueli (1999) and Creaco and Pezzinga (2015a, b) used
of Exeter, Exeter EX4 4QF, UK; Adjunct Senior Lecturer, School of Civil, LP for optimal valve control. However, the former authors tackled
Environmental and Mining Engineering, Univ. of Adelaide, Adelaide 5005, optimal valve location through a fully deterministic procedure, based
Australia (corresponding author). ORCID: https://orcid.org/0000-0003 on the sequential addition (SA) of beneficial valves up to a maximum
-4422-2417. Email: creaco@unipv.it number of valves installable in the WDN. The latter, on the other
2
Associate Professor, Dept. of Civil Engineering, Lebanese Univ., hand, used a multiobjective GA. The comparison of SA and multi-
Rafic Hariri Campus, Hadath B.P. 2, Baabda, Lebanon. Email: hatem
objective GA for control valve location was presented by Creaco and
_mhaidar@hotmail.com
Note. This manuscript was submitted on July 24, 2018; approved on Pezzinga (2018), who showed that the former has the advantage of
February 28, 2019; published online on August 14, 2019. Discussion per- lower computational burden, whereas the latter performs better at
iod open until January 14, 2020; separate discussions must be submitted for finding beneficial valve locations in terms of leakage abatement.
individual papers. This paper is part of the Journal of Water Resources Compared to the other works in the scientific literature, a further
Planning and Management, © ASCE, ISSN 0733-9496. novelty of the work of Creaco and Pezzinga (2015a, b) lies in the
et al. 2016; Galdiero et al. 2016; Laucelli et al. 2017; Di Nardo et al. isolation valve.
2018a, b; Liu and Han 2018; Liu et al. 2018) to separate DMAs, the The gene values 2 and 3, both of which encode isolation valve
option of closing an isolation valve is usually evaluated as a cost- closure in the generic pipe, have the same hydraulic effect, i.e., in-
effective alternative to the installation of a flow meter, as long as finite pipe resistance. Nevertheless, the two gene values are consid-
it does not cause any pressure deficits in the WDN. Therefore, the ered to have different effects in terms of the creation of DMAs.
question arises as to whether isolation valve closures could be op- Specifically, in the case of a gene value equal to 2, the pipe is con-
timized to obtain simultaneously a large leakage abatement and sidered a present link during WDN partitioning. The generic pipe
an effective WDN partitioning inside an optimization framework associated with a gene value of 3, on the other hand, is removed
that also includes installation of control valves. A multiobjective from the layout before the application of the WDN partitioning
methodology is presented in what follows to explore this issue. algorithm. The distinction between gene values of 2 and 3 was
In the following sections, first the methodology proposed is de- made to enable NSGAII to select which closed isolation valves will
scribed, followed by applications to a real WDN. The paper ends affect the formation of DMAs.
with conclusions. After the initial population’s generation, the individuals evolve
in generations through processes of crossover and mutation while
being organized in fronts of various levels of dominance. In the
Methodology algorithm, the number of population individuals and the maximum
number of generations must be fixed to obtain accurate results
The algorithm proposed in this work is based on the coupling of within an acceptable computation time. For each individual of
three algorithms: NSGAII, the sequence of instructions shown in Fig. 1 is carried
Algorithm 1: Multiobjective GA for the optimal location of out, which includes the main feasibility check and the execution
control valves and for the identification of isolation valves to be of Algorithms 2 and 3 in fitness evaluation.
installed and closed in the WDN;
Algorithm 2: WDN partitioning algorithm; Main Feasibility Check
Algorithm 3: To search for the optimal settings of the control After valve location according to the genes, the generic individual
valves at each time slot of WDN operation. of NSGAII undergoes the main feasibility check. In this context,
The three component algorithms are described in detail in the the individual is considered infeasible when isolation valve closures
following subsections. cause the disconnection of some nodes from the sources or nodal
Compared to the work of Creaco and Pezzinga (2015a, b), the pressure deficits below the desired pressure head hdes (m) under
main novelty of the approach described here lies in the introduction conditions of closed isolation valves (for both gene values 2 and 3)
of Algorithm 2 for WDN partitioning. This results in a methodol- and open control valves. The degree of infeasibility is expressed as
ogy that optimizes simultaneously control valve installations and a function of the number of disconnected nodes and of the maxi-
DMA creation while searching for optimal solutions in the trade- mum pressure head violation. It results in a penalization of the ob-
off between installation costs, leakage, and demand uniformity jective functions, which are assessed as explained in what follows.
across DMAs. Furthermore, Algorithm 3 was improved in leakage
estimation. In fact, a more general relationship to express leakage Assessment of Demand Uniformity Cv across DMAs
as a function of service pressure was implemented, as will be shown (Objective Function)
in what follows. The WDN partitioning algorithm (i.e., Algorithm 2) is then run to
According to the Talbi (2002) classification, the algorithm is a subdivide the WDN into a certain number N dis of DMAs. To this
low-level hybrid algorithm, since Algorithms 2 and 3 are embedded end, the WDN structure resulting from the removal of the pipes
in Algorithm 1 as functional parts. The algorithm aims to obtain associated with gene values equal to 3 is considered. The removal
optimal solutions in the trade-off between installation costs and daily of various pipes, as proposed by the generic individuals of NSGAII,
leakage volume, which fit in with WDN partitioning into DMAs. enables different DMA configurations to be obtained through the
partitioning algorithm. This algorithm enables identification of the
total number N bp of boundary pipes between DMAs and the cal-
Algorithm 1: Multiobjective Genetic Algorithm NSGAII culation of how many closed isolation valves are placed inside or
outside the boundary pipes (numbers equal to N ivb or N ivnb, respec-
Individual Encoding tively, with N iv ¼ N ivb þ N ivnb ). A flow meter must be installed at
The multiobjective GA NSGAII (Deb et al. 2002) is used as each of the boundary pipes without a closed isolation valve to mon-
Algorithm 1 for the optimal location of control valves and for itor the exchange of flow. These boundary pipes are as numerous
the identification of the isolation valves to be closed in the as N flowmeter ¼ N bp − N ivb . After the application of the WDN
WDN, with the objective of simultaneously optimizing demand partitioning algorithm, a secondary feasibility check is performed.
where Cval;i is the valve cost in the ith of the npv pipes that enable
control valve installation or isolation valve closure. In Eq. (1), if the
ith of the npv valve-related genes is equal to 0, the corresponding
value Cval;i is set at 0; if the ith gene is equal to 1, Cval;i is equal to
the cost of a control valve; finally, if the ith gene is equal to 2 or 3,
Cval;i is equal to 0 if an isolation valve is already present in the ith
pipe (because it has already been installed for other purposes, such
as segment isolation) or to the cost of an isolation valve if the latter
device is absent. Furthermore, Cflowmeter;i is equal to the cost of the
flowmeter if the generic ith of the np WDN pipes was identified as
a flowmeter-fitted boundary pipe by the WDN partitioning algo-
Downloaded from ascelibrary.org by Zhejiang University on 11/27/22. Copyright ASCE. For personal use only; all rights reserved.
X
nΔt
WL ¼ W L;j ð2Þ
j¼1
where nΔt represents the number of time steps associated with the
daily operation of the WDN.
In the presence of infeasibility, a penalized daily leakage volume
is estimated without running Algorithm 3, starting from the daily
leakage volume in the no-control scenario. Algorithm 3 cannot be
the source of any infeasibility for an initially feasible individual
because it preserves the meeting of nodal pressure constraints.
Improvements of NSGAII
The various individuals of NSGAII are compared based on their
fitness, composed of the three objective functions described earlier,
to be simultaneously optimized. When the population is old
Fig. 1. Logic flux of instructions to carry out for each individual of
enough, the entire population belongs to a single Pareto band of
genetic algorithm.
undominated individuals. Compared to the NSGAII proposed by
Deb et al. (2002), the upgrades proposed by Creaco et al. (2016)
to improve the algorithm convergence in the presence of more than
In this context, the isolation valves associated with gene values two objective functions were implemented. Another variation was
equal to 2, which were retained as active links during WDN par- implemented to prevent the population’s aging. If all of the popu-
titioning, are now also removed from the WDN, to determine lation’s individuals belong to a single band of undominated indi-
whether any of these isolation valves causes disconnections inside viduals before the maximum number of generations is reached, a
the generic DMA. Failure to comply with this secondary feasibility check is made about the presence of duplicate individuals. Whether
check results in an additional increase in the degree of infeasibility present, they are regenerated to insert new genetic characters into
of the individual. the population.
The first objective function is then assessed as the demand uni-
formity across DMAs. It is calculated as the coefficient of variation Algorithm 2: Fast-Greedy Partitioning Algorithm
Cv of DMA demands, the ratio of the standard deviation to the aver-
age value, to be minimized. The lower Cv, the more uniform is de- The fast-greedy partitioning algorithm (FGPA), developed through
mand across DMAs. graph theory by Clauset et al. (2004) and based on the concept of
modularity as an indicator of the strength of partitioning, is used as
Assessment of Total Installation Cost Cinst Algorithm 2 to obtain WDN partitioning into N dis DMAs. Various
(Objective Function) authors have used FGPA (e.g., Diao et al. 2013; Ciaponi et al. 2016;
After N val control valves have been installed and N iv isolation Liu et al. 2018). Though Liu et al. (2018) recently proved that this
valves have been closed based on the NSGAII individual’s encod- algorithm has both advantages and disadvantages, it was chosen as
ing, the installation cost Cinst of the intervention is assessed (second Algorithm 2 in the present work due to its simple implementation.
objective function to be minimized) as FGPA is applied to the WDN configuration associated with the
sponding to gene values of 3 have been removed from the WDN W L;j ¼ QLi Δt ð6Þ
layout. This is done to drive WDN partitioning toward DMAs with i¼1
some closed isolation valves placed at the boundary pipes. Different
distributions of closed isolation valves associated with gene values where QLi (m3 =s) is the leakage outflow from the ith of the np
equal to 3 produce different configurations of DMAs, which may network pipes during the jth time step, and Δt (s) is the temporal
have a different performance in terms of leakage abatement and length of the time step itself.
demand uniformity. Within the optimization problem, control valve settings can be
For the application of FGPA, modularity can be expressed defined as coefficients V that take on real positive values lower than
starting from the incidence topological Matrix A, with size np xnn 1, by means of which the resistance of the generic pipe fitted with
(np and nn are the total numbers of pipes and nodes, respectively). control valve is modified in such a way as to take account of the
In the generic row of A, associated with the generic network pipe, presence of the valve itself. Vector V ¼ ðV 1 ; V 2 ; : : : ; V k ; : : : ;
the generic element can take on the values 0, −1, or 1, whether the V N val Þ of control valve settings can then be defined, where N val
node corresponding to the matrix element is not at the end of the is the number of pipes fitted with a control valve.
Downloaded from ascelibrary.org by Zhejiang University on 11/27/22. Copyright ASCE. For personal use only; all rights reserved.
pipe, is the initial node of the pipe, or is the final node of the pipe, Within the objective function in Eq. (6), the leakage outflow in
respectively. Starting from A, Vector K (nnx1) and Matrix B cubic meters per second from the generic pipe can be assessed
(nn × nn) can be calculated using the following expressions: through the following relationship derived from May (1994):
where ci and cj are the DMAs to which the ith and jth nodes where Hi;1 (m), H i;2 (m) and zi;1 (m), zi;2 (m) are the heads and
belong, respectively. Finally, δðci ; cj ) is equal to 1, whether the elevations, respectively, for the end nodes of the pipe.
ith and jth nodes belong to the same DMA (that is, ci ¼ cj ). Simple calculations transform Eq. (7) into the following Eq. (9):
Otherwise, δðci ; cj Þ ¼ 0.
The objective of the algorithm lies in obtaining a WDN partition
QLi ¼ Li ðC1L;i h0.5 1.5
i þ C2L;i hi Þ ð9Þ
featuring a high value of M. After the target number N dis of DMAs
has been set, the algorithm starts by considering each WDN node
constituting a DMA of its own. Therefore, at the initial step, the where C1L;i (m3=2 =s) and C2L;i (m1=2 =s) are coefficients that have to
number of DMAs in the WDN is equal to the number of nodes, be calibrated based on measurements on the WDN.
which is nn. At the second step, two DMAs are joined together, Leakage outflows QLi from pipes can be regarded as the ele-
and the number of DMAs becomes nn − 1. The aggregation pro- ments of a vector QL (np × 1), evaluated through the following
cess is repeated in the following steps until the network merges to vector equation:
N dis DMAs. At the generic step, the choice of the two DMAs to
merge is made to obtain the highest positive variation in M. jA10 jðH0 − z0 Þ þ jA12 jðH − zÞ 0.5
QL ¼ diagðCL1 Þ
2
Algorithm 3: Linear Programming Algorithm jA10 jðH0 − z0 Þ þ jA12 jðH − zÞ 1.5
þ diagðCL2 Þ ð10Þ
The iterated LP, an upgraded version of that proposed by Creaco 2
and Pezzinga (2015a, b), is used as Algorithm 3 for the optimiza-
tion of control valve settings located following the generic individ- where CL1 and CL2 are vectors with size np × 1, the elements of
ual of NSGAII. The main upgrade lies in leakage estimation by the which are Li C1L;i and Li C2L;i , respectively.
fixed and variable area discharge (FAVAD) equation (May 1994; In Eq. (10) Matrixes A12 and A10 are derived from A. Matrix
Van Zyl and Cassa 2014), which guarantees better estimation of A12 (np × n) is obtained by considering the columns corresponding
leakage over a large range of service pressures than the conven- to the n network nodes with unknown head. Matrix A10 (np × n0 )
tional emitter equation (Rossman 2000). is obtained by considering the columns corresponding to the n0
Algorithm 3 is used to search for the optimal setting of the con- nodes with fixed head. H (n × 1) and H0 (n0 × 1) are the vectors
trol valves in the jth of the nΔt time steps that characterize the net- of nodal heads for the unknown head nodes and for the fixed head
work operation (in terms of hourly variations in source head and nodes, respectively. Finally, z (n × 1) and z0 (n0 × 1) are the vec-
demand coefficient). The objective of the optimization is to min- tors of nodal elevations for the unknown head nodes and for the
imize the leakage volume W L;j at each time step, calculated by fixed head nodes, respectively. In Eq. (10), the division by scalar 2
means of the following relationship: and the powers 0.5 and 1.5 are applied to each element of the vector.
Results
The results of the optimizations are shown in the graph in Fig. 3,
which reports the Pareto band of W L as a function of Cinst and Cv .
The graph has Cinst and W L on the x- and y-axes, respectively. The
third objective function, Cv , is reported in the graph through the
tones of the dots: the darker the dot, the higher the value of Cv .
Globally, the graph shows decreasing values of W L as Cinst in-
creases from about €50,000 to about €400,000. Furthermore, for
a prefixed value of Cinst , the highest values of Cv , associated with
solutions with nonuniform distribution of demand across DMAs,
lead to lower values of W L . These solutions are more tailored to
leakage minimization than to WDN partitioning. Indeed, the lower
envelope of the Pareto band, including the solutions with the high-
est value of Cv for each value of Cinst , coincides with the Pareto
front obtainable through the biobjective (Cinst − W L ) optimization
performed by Creaco and Pezzinga (2015a, b). When moving from
low to high values of Cinst , the range of values obtained for W L as a
function of Cv becomes smaller and smaller. This means that, for
Fig. 5. Solutions with (a) ID = 3; and (b) ID = 4. See text and Tables 1 Fig. 6. Solutions with (a) ID = 5; and (b) ID = 6. See text and Tables 1
and 2. and 2.
whereas a larger total cost of control valves (Cval ) is obtained in smaller and smaller. As a result, the methodology proposed in this
Solution 2. The number N vault and cost Cvault of vaults is the same work can yield solutions enabling uniform demand distributions
for the two solutions even if Solution 2 features one additional across DMAs without compromising too much the performance
device to be inserted in a vault (N val þ N flowmeter ¼ 7 for in terms of leakage abatement. This concept holds true also when
Solution 2 versus N val þ N flowmeter ¼ 6 for Solution 1). This Solutions 5 and 6, which feature Cinst values around €166,000
happens because one of the vaults in Solution 2 hosts both a flow- (Fig. 6 and Tables 1 and 2), are considered. In fact, the results for
meter and a control valve. Incidentally, it must be remarked that this high value of Cinst point out that the Cv reduction of 0.25 pro-
the possibility to consider cost reductions by inserting a control duces an increment in the leakage reduction of 105 m3 , which is
valve and a flowmeter into the same vault represents an additional smaller than those obtained above for lower values of Cinst . This
benefit of the methodology presented in this work. happens even if the lower leakage volume is obtained in Solution 6
Solutions 3 and 4 (shown in Fig. 5 and in Tables 1 and 2) have a through the installation of five additional control valves, in compari-
similar value of Cinst , around €101,000. While the former solution son with Solution 5.
has more numerous flowmeters (6 versus 5), the latter features more
numerous control valves (4 versus 3). Compared to Solutions 2 Discussion
and 1, Solutions 4 and 3 have a similar difference in terms of Cv
(i.e., a difference of about 0.25). However, the difference in terms To test the overall effectiveness of the methodology, an additional
of W L is smaller in the case of Solutions 3 and 4. In fact, this differ- optimization was performed under conditions of more complex
ence is around 195 m3 , while the W L difference between Solution 1 topography. Specifically, the ground elevation was increased to
and Solution 2 is about 276 m3 . This happens because, as the num- 920 m asl in the nodes in the core of the WDN layout (Fig. 2).
ber of control valves installed in the network grows, the benefit Considering a fixed head of the source equal to 962 m, these nodes
derived from the installation of additional control valves becomes feature now only a minor pressure surplus, in comparison with the
on the graph theory for DMA creation, and (3) iterated linear pro- 2
gramming for the optimization of control valve settings. The ap-
plications to the main skeleton of a real WDN showed that the þ d − 2A21 B11 NA12 H þ 2A2nv Bnvnv ðAnv0 H0 þ Anv2 HÞ
methodology yields control valve installations and isolation valve − A21 B11 NA10 H0 ð23Þ
closures associated with various levels of daily leakage volume and
of fit to WDN partitioning, evaluated as demand uniformity across where Matrixes Bpvpv and Bnvnv are derived from Matrix B11 by
DMAs. For a prefixed installation cost, including installation of considering the rows and the columns associated with the pipes
vaults, flowmeters, and control valves, the methodology provides fitted with a control valve and lacking a control valve, respectively.
solutions with smaller leakage abatement and higher fit to WDN Matrixes Apv0 and Apv2 are derived from Matrixes A10 and A12 ,
partitioning, as well as solutions with larger leakage abatement and respectively, by considering the rows associated with the pipes fit-
lower fit to WDN partitioning. However, the variation in daily leak- ted with a control valve. Matrixes Anv0 and Anv2 are derived from
age volumes as a function of the fit to WDN partitioning is small in Matrixes A10 and A12 , respectively, by considering the rows asso-
some cases, especially for high installation cost values, associated ciated with the pipes lacking a control valve. In Eq. (22), Vectors V
with many installed control valves. Therefore, the good fit of the and H and Matrix N are relative to iteration iter and Vector P
solutions provided by the methodology to WDN partitioning does (np × 1) is yielded by the following relationship:
not always compromise the high performance in terms of leakage
reductions compared to the no-control scenario. In fact, compared jA10 jðH0 -z0 Þ þ jA12 jðHiter -zÞ
P¼ ð24Þ
to an optimization aiming only at minimizing leakage with no ac- 2
count for WDN partitioning, solutions can be found with similar
leakage abatement performance but better fit to WDN partitioning.
A future development of the work will concern the implemen- Data Availability Statement
tation of different WDN partitioning algorithms from that used
in this work (Clauset et al. 2004), including spectral algorithms The data used for the calculations are made available upon request
(e.g., Di Nardo et al. 2018a, b) and algorithms based on modularity by the authors of the paper.
functions specifically oriented toward WDNs (Giustolisi and
Ridolfi 2014). Furthermore, case studies with different character-
istics from that considered in this paper will be analyzed. This will Acknowledgments
enable identification of the best WDN partitioning algorithm and
of its most suitable settings to be used when WDN partitioning is This work was done within the project Energy and Leakage
optimized simultaneously with installation of control valves. Optimization in Lebanese Water Distribution Networks, funded by
the Lebanese University. The authors wish to thank the anonymous
reviewers, who provided very useful comments for the improve-
ment of the paper.
Appendix. Linearization for the Iterated LP Algorithm
Applying linear theory concepts to Eq. (11), the following linear
objective function is obtained: References