Journal of Environmental Management 226 (2018) 308–319

Contents lists available at ScienceDirect

Journal of Environmental Management

journal homepage: www.elsevier.com/locate/jenvman

Research article

Many-objective optimization model for the flexible design of water T

distribution networks
João Marquesa,∗, Maria Cunhaa, Dragan Savićb
a
INESC Coimbra – Institute for Systems Engineering and Computers at Coimbra, Department of Civil Engineering, University of Coimbra, Coimbra, Portugal
b
Centre for Water Systems, School of Engineering, Computing and Mathematics, University of Exeter, United Kingdom

A R T I C LE I N FO A B S T R A C T

Keywords: This paper proposes a many-objective optimization model for the flexible design of water distribution networks
Water distribution networks (WDNs), including four objectives. Two objectives are related to the WDNs' hydraulic capacity, the minimization
Many-objective optimization of the pressure deficit and the undelivered demand. The third objective is the traditional cost minimization while
Uncertainty the fourth minimizes carbon emissions. These objectives concern network reliability, and financial and en-
Multi-stage design
vironmental concerns. They can give rise to solutions embedding new trade-off in design perspectives. There is a
Flexibility
gap in the literature when it comes to dealing with many-objective problems for designing and constructing a
Simulated annealing
WDN over a long-term planning horizon and using a staged design scheme that includes the consideration of
uncertainty. A solution obtained through this process can be implemented in the first stage and the WDN is
prepared for the possible occurrence of various future scenarios. These scenarios can consider expansions of
WDNs to different development areas, in different time stages. Furthermore, defining a multi-staged design
allows implementing the design of the first stage and reassessing the whole process in the end of each stage when
more plausible future scenarios can be investigated. The solution of complex problems such as these needs
improved algorithms to produce the Pareto front and so enable the trade-off between the objectives to be ex-
amined. An enhanced algorithm, based on the simulated annealing concept and capable of handling the critical
scalability issues encountered in previous algorithms with respect to drawing the Pareto front for many-objective
problems where a high-dimensional space is involves, is presented. The results obtained allow a thorough
analysis of trade-offs between objectives and confirm the importance of considering the minimization of all those
four objectives and the advantages of using a flexible approach to design WDNs to better inform decision makers.

1. Introduction Different approaches have been followed, ranging from the initial
attempts in which single objective models were solved by the linear
Water distribution network (WDN) design is one of the most com- programming gradient method (Alperovits and Shamir, 1977), through
plex problems in the management of urban water systems (Mala- nonlinear programming methods (Lansey and Mays, 1989), to the more
Jetmarova et al., 2018). The complexity of the design problem stems recent heuristics methods used to solve more complex problems such as
mainly from its discrete and nonlinear nature, multiple criteria for real WDNs. Examples of heuristics that dealt with the single objective
evaluation, as well as uncertainties inherent in long-term planning. This (cost minimization) problem optimization of WDNs include genetic
work presents a many-objective model for the flexible design of WDNs, algorithms (Savic and Walters, 1997), simulated annealing (Cunha and
based on a multi-stage scheme. The aim is to account for multiple Sousa, 1999), ant colony optimization (Maier et al., 2003) and harmony
benefits of the design by minimizing four objectives, namely, pressure search (Geem, 2006). The simplest approach to minimizing the cost via
deficits, undelivered demand, construction costs, and carbon emissions. a single-objective optimization model was most likely developed be-
Many studies of the optimal design of WDNs have been published over cause of the high complexity of these systems, and because financial
the last three decades. Mala-Jetmarova et al. (2018), present a detailed resources to construct infrastructure are always limited. However, ap-
literature review where all the concepts, main contributions, develop- proaches involving more objectives have subsequently been devised
ments, trends and limitations on this subject are analysed and high- because WDN design often exhibits multiple, conflicting objectives
lighted. (Savic, 2002).

∗
Corresponding author.
E-mail addresses: [email protected] (J. Marques), [email protected] (M. Cunha), [email protected] (D. Savić).

https://doi.org/10.1016/j.jenvman.2018.08.054
Received 20 July 2017; Received in revised form 10 August 2018; Accepted 12 August 2018
0301-4797/ © 2018 Elsevier Ltd. All rights reserved.
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

In addition to cost minimization, a considerable body of literature Pareto dominance relation in high-dimensional space. In fact, the per-
has considered other objectives, mostly using two-objective models. For formance of the Pareto dominance-based algorithms developed for
example, the pressure deficit reduction at nodes as a measure of net- solving problems involving two or three objectives decreases con-
work reliability was included by Keedwell and Khu (2004), Kapelan siderably as the number of objectives increases. Various Multiobjective
et al. (2005) and Fu and Kapelan (2011) for the optimal rehabilitation Evolutionary Algorithms - MOEAs (NSGA-II, PAES, MOEA/D-PBI, etc.,
of an existing WDN (New York Tunnels case study). Similarly, as presented in Jiang and Yang, 2017), and Multiobjective Simulated
Atiquzzaman et al. (2006) applied the same objective for the Hanoi and Annealing Algorithms - MOSAs (SMOSA, PDMOSA, WDMOSA, etc., as
Alperovits and Shamir (1977) networks. The minimization of the un- presented in Suman and Kumar, 2006, Bandyopadhyay et al., 2008 and
delivered demand was used by Tanyimboh and Seyoum (2016) for a Sengupta and Saha, 2018), developed in the past face severe scalability
real UK case study. The maximization of a network resilience index was issues (Jiang and Yang, 2017; Sengupta and Saha, 2018). Many-objec-
included by Wang et al. (2015), while the minimization of greenhouse tive optimization poses new challenges, such as the increasing pro-
gas emissions was studied by Wu et al. (2008), Herstein et al. (2009), portion of non-dominated solutions and the inadequacy of Pareto
Wu et al. (2010) and Stokes et al. (2015) for the optimal design and dominance to create enough selection pressure in problems with a
operation of new WDNs. Shokoohi et al. (2017) added the maximiza- higher number of objectives.
tion of water quality as another objective. Although the two-objective To emphasize the difference between the needs for solving problems
WDN design problem is sufficiently complex to warrant the use of so- involving more than three objectives, a new term “many-objective op-
phisticated heuristic methods, going beyond two objectives brings ad- timization problems” (MaOPs) has appeared in the literature (Farina
ditional challenges for optimization methods (Deb, 1999). and Amato, 2002; Purshouse and Fleming, 2003, 2007). Since Chand
Wu et al. (2013) proposed a three-objective model, including and Wagner (2015) established that at least three objectives are needed
minimizing costs, improving reliability and reducing greenhouse gas for a problem to be considered an MaOP the concept has gained
emissions. Khu and Keedwell (2005) analyse the additional design widespread popularity (Li et al., 2015; Jiang and Yang, 2017; Sengupta
choices given by an optimization model with six objectives compared and Saha, 2018).
with design solutions obtained with a two-objective NSGA-II model. This paper proposes a many-objective optimization model for the
Giustolisi and Berardi (2009) considered four objectives (investment design of WDNs that can be expanded over the planning horizon. The
costs, cost of pipe breaks, preferential selection of only one pipe ma- literature shows that the four objectives considered in this work can
terial and reliability) for the rehabilitation of water distribution net- significantly influence WDN design. However, no research has yet ex-
works (case study, real UK water distribution network), using the amined the trade-offs between these four objectives (minimizing pres-
multiobjective genetic algorithm, OPTIMOGA. Fu et al. (2013) solved sure deficits, undelivered demand, costs and carbon emissions) while
the rehabilitation and operation of water distribution networks (case also considering a multi-stage design for WDNs to tackle uncertainty
study: Anytown) with the following objectives: investment costs, op- issues and, as such, define flexible solutions. Therefore, it is important
erating costs, hydraulic failure, leakage, water age, and fire-fighting to address this research gap by understanding the compromises that can
capacity, using the ε-NSGA-II algorithm. However, none of these studies be achieved between these objectives to better inform utility decision
deals with the much more complex problem of the design and con- makers. A scenario-based approach is proposed to deal with uncertain
struction of a WDN over a long-term planning horizon through a staged future circumstances by considering a planning horizon that consists of
design scheme including consideration of uncertainty. a number of construction stages during which new urban areas can be
Staged design is important for considering uncertainty issues (e.g., established over time. Only the design for the first stage of the planning
uncertain nodal demands or network deterioration), as it allows inter- horizon has to be implemented at the time of decision making.
vention at different stages of the planning horizon by defining flexible However, the scenarios for the future staged development are essential
solutions in the process of WDN development and permitting planners to provide flexibility (de Neufville and Scholtes, 2011) so that the first
to act (and change decisions) as new information becomes available stage solution can cope with a range of possible future conditions. As an
(Spiller et al., 2015). This has an advantage over the traditional single- adaptive scheme is implemented, the whole process can be repeated at
stage design where an intervention is fixed for the entire planning the end of each stage when more plausible future scenarios can be in-
horizon, which can result in an under- or over-designed WDN. Huang vestigated.
et al. (2010) address the problem of the optimal design of WDNs in Solving a four-objective and multi-stage design problem is a major
stages that takes uncertainties of future water demand into account. challenge for optimization algorithms as it involves exploring a large
They developed flexible solutions but used a single-objective model for solution space and comparing many solutions during the optimization
cost minimization. Creaco et al. (2015) investigate three distinct ap- process. This complexity increases significantly when a staged design
proaches to phasing (or staging) the design of WDNs while also con- (one that is carried out in stages using short time steps at each stage) is
sidering a minimum cost objective. Creaco et al. (2014) also propose an considered under uncertain futures (Mala-Jetmarova et al., 2018).
approach for phasing the construction of WDNs considering two ob- Another challenge is extracting information from a large number of
jectives, the minimization of costs and the pressure surplus. Their so- non-dominated solutions found on the approximation to the Pareto
lution uses a multiobjective genetic algorithm and draws some com- front.
parisons with the fixed design solutions, which shows that the staged The model developed in this paper is solved by an enhanced si-
design is better. Basupi and Kapelan (2015a) and Basupi and Kapelan mulated annealing algorithm, based on the concepts presented in
(2015b) propose a different approach to develop flexible solutions, Kirkpatrick et al. (1983), and exploring the amount of domination
taking demand uncertainty into account by applying staged interven- concept (Bandyopadhyay et al., 2008) capable of overcoming the major
tions in the network and considering a two-objective model. The model drawbacks of scalability previously mentioned.
that minimizes cost and maximizes resilience was solved by an NSGA-II The remainder of the paper is organized as follows. In section 2, the
algorithm. Marques et al. (2015a) include the flexible WDN design optimization model is formulated, and section 3 details the optimiza-
problem with a two-objective optimization formulation involving the tion tool used to solve the many-objective model. The case study, the
minimization of costs and pressure deficits. Marques et al. (2017) results and discussion are presented in section 4 and, finally, the last
solved a three-objective optimization problem by minimizing costs, section sets out the conclusions.
carbon emissions and pressure deficits. Both studies used a simulated
annealing algorithm to solve the optimization problem. 2. Multi-stage and many-objective model
The analysis of literature on WDN and on several other fields, when
there are many objectives at stake, shows the ineffectiveness of the The proposed many-objective model can solve problems in urban

309
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

areas that are expanding through population growth, by exploiting hydraulic analysis by summing, over all scenarios, all stages after t = 1
adaptable WDN design over a planning horizon. Efficient interventions and all network nodes, the maximum undelivered demand for all de-
for a number of scenarios (represented by different decision paths that mand conditions. In (4), Drqn,d,t,s = demand required at node n for
can be followed over a planning horizon) need to be defined. A flexible demand condition d for time stage t and in scenario s and
design through a staged scheme is implemented. The staged design Ddln,d,t,s = supply delivered at node n for demand condition d for time
means that the WDN constructed can be adapted in future stages as new stage t and in scenario s.
information becomes available. Therefore, the chosen solution must The objectives of minimizing pressure deficits and undelivered de-
satisfy design requirements for the first stage and to be prepared to cope mand are correlated as the supply of a node is driven by the pressure
with increasingly uncertain future scenarios. The model considers four under deficient pressure conditions. However, network designs that
different minimization objectives, which are, pressure deficit, un- only consider the pressure deficit objective will disregard the flow de-
delivered demand, total costs and carbon emissions. The objective livered at each node. For example, it is possible for different un-
function (1) is used to minimize the pressure deficit that can occur in delivered demand values to be experienced for the same pressure def-
future stages: icits distributed at a number of nodes. This can be due to the network
NSC NTI NNS nodes having different demands even when the pressure deficit is the
PD = ∑ ∑ ∑ d=Max {Max [0; (Prqmin, n, d − Pn, d, t , s )]}. same, e.g. a pressure deficit of 10 m at a node with demand of 60 L/s
1,...NDC (1) has an undelivered demand value that is different from that at a node
s=1 t=2 n=1
with demand of 10 L/s but with the same pressure deficit of 10 m. This
This pressure deficit (PD) is the sum of the pressure deficits, con-
also means that the undelivered demand minimization objective could
sidering all scenarios (NSC), all time stages (NTI) after t = 1, all net-
be redundant (relative to the pressure deficit minimization objective)
work nodes (NNS) and the maximum deficit for all demand conditions
when dealing with hypothetical networks with the same demand for all
(NDC). For t = 1, nodal pressures (Pn,d,t,s = pressure at node n for de-
the network nodes, but it is of critical importance to consider it if the
mand condition d for time stage t and in scenario s) have to be fully
network under analysis includes nodes with very different demand
satisfied, i.e., nodal pressure must be equal to or higher than the
values (the typical real world situation). In this case, the minimization
minimum required pressure. In future stages, it is possible for pressure
of undelivered demand is handled in the optimization process by se-
to be lower than this pressure (Prqmin,n,d = minimum required pressure
lecting solutions that tend to increase the network hydraulic capacity
at node n for demand condition d) but they should always be higher
(e.g. by using large pipe sizes) for supplying the network areas with the
than zero. This is achieved by including the minimum pressure con-
highest nodal demands. This is very important because most WDNs
straint (2) to limit Pn,d,t,s:
have critical nodes that normally experience high demand, such as
⎧ Pn, d, t , s ≥ Prqmin, n, d t = 1 ∀ n ∈ NNS ;∀ d ∈ NDC; ∀ s ∈ NSC . hospitals and schools. These are often held to be vulnerable users with
priority (with low undelivered demand) over less important areas with
⎩ Pn, d, t , s ≥ 0
⎨ t>1 (2)
very low demand. This cannot be done by simply minimizing the
Under pressure deficient conditions, demand cannot be totally met pressure deficits, and so this work also includes undelivered demand
and therefore the supply delivered to nodes is lower than the required minimization as an objective of the optimization model.
demand. A pressure driven hydraulic simulator EPANETpdd (Morley The other two objectives of the model are the minimization of the
and Tricarico, 2008) is used to simulate hydraulic conditions in the total cost (CT) and the minimization of carbon emissions (CE). The total
network by using the head-flow relationship proposed by Wagner et al. cost is given as the sum of the initial solution cost to be implemented in
(1988). The supply delivered (Ddl) is obtained for each node by using the first stage and the future costs computed by summing all possible
the expressions in (3) (Wagner et al., 1988), which considers three options (different scenarios are established as options that can be
different conditions: adopted in future stages), starting with the second stage of the planning
horizon. The CE are also given by summing the emissions of the first
⎧ Ddl = Drq if Pavl ≥ Prq
stage with a sum of the future carbon emissions of all the possible future
⎪ 1/2

⎨ (
P −P
Ddl = Pavl − P min
rq min
⎪ Ddl = 0 if P ≤ P
) × Drq if Prq > Pavl > Pmin . options that can be adopted. The CT and CE objectives are described in
detail in the case study section of this work. The many-objective model
⎩ avl min (3) also includes the usual hydraulic constraints of WDN design optimiza-
tion models to check for nodal mass balance, to compute the head loss
If the available nodal pressure (Pavl) is equal to or higher than the
in the pipe.
required pressure (Prq), all the demand required (Drq) is satisfied. For
As mentioned above, this model aims to find flexible development
nodes with heads between required pressure (Prq) and a minimum nodal
solutions by considering possible alternative decision paths that can be
pressure (Pmin) the supply is partially delivered and is computed ac-
followed in future. These alternatives should be identified in line with
cording to (3). No supply will be delivered to nodes with pressures
future developments and through the involvement of stakeholders.
below or equal to minimum nodal pressure (Pmin). The main purpose of
Therefore, the process of selecting future adaptations for the network
WDNs is to deliver water in sufficient quantity, of required quality and
and the corresponding probabilities is not the aim of this work; its main
with enough pressure. Expression (1) is used to minimize pressures
aim is to show how this approach can be used as a decision support tool
lower than a required level. But this is not the situation that best
to decision makers.
guarantees the minimum undelivered water in the WDN. For pressures
lower than required, the demand delivered is given by the expression
3. Many-objective optimization method
proposed by Wagner et al. (1988), which uses a nonlinear relationship
between available pressure (Pavl) and supply delivered (Ddl). In real
A heuristic based on the concept of simulated annealing
systems, due to spatial variability of demand, undelivered will also vary
(Kirkpatrick et al., 1983) is used to solve the four-objective optimiza-
even for the same pressure deficit. Therefore, a second objective func-
tion model. The enhanced simulated annealing algorithm appropriate
tion, given by (4), is used to minimize the undelivered demand (UD) at
to deal with MaOPs results from embedding the concept of amount of
nodes where the pressure is lower than the required pressure:
domination as developed by Bandyopadhyay et al. (2008) in the algo-
NSC NTI NNS rithm of simulated annealing presented in Marques et al. (2015b)
UD = ∑ ∑ ∑ d=1,Max
..., NDC
{(Drqn, d, t , s − Ddln, d, t , s )} . (updated version of Cunha and Sousa, 1999). The amount of domina-
s=1 t=2 n=1 (4)
tion concept (Δdom) between two solutions Sa and Sb is defined in (5) by
The undelivered demand is computed from the pressure-driven multiplying the differences in values of the objectives (Oi) divided by

310
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

the objective range (Ri), if this difference is other than zero: domination.
NOB The second situation (C2) occurs when the current solution and
Oi (Sa) − Oi (Sb) candidate solution are non-dominated with respect to each other. In
ΔdomSa, Sb = ∏ Ri
.
i = 1, Oi (Sa) ≠ Oi (Sb) (5) this case is not possible to improve one of the objectives without wor-
sening another one. Three situations can occur in (C2), depending on
Its use contributes to overcome the previously mentioned draw- the domination status between the candidate solution and solutions
backs of scalability due to the restricted use of the Pareto dominance in from the archive. First, at least one of the solutions from the archive
MaOPs problems. The amount of domination concept is used in this dominates candidate solution (C2.1) and the Pacp is computed with
algorithm to compute the acceptance probability in (6): Δdom given by the mean value of the amount of domination of the
−Δdom ⎞ solutions Sarch that dominate Scand. Then the candidate solution can be
Pacp = exp ⎛ . accepted as the new current solution if the Metropolis criterion is met.
⎝ T ⎠ (6)
Second, the candidate solution is non-dominated by any solution of the
The acceptance probability (6) is also function of the temperature archive (C2.2), then the candidate solution becomes the new current
parameter (T). Simulated annealing is based on an analogy with the solution and is added to the archive. In the third situation, the candi-
physical process of cooling a material in a heat bath; it represents the date solution dominates at least one of the solutions from the archive
physical process of heating a material up to its melting temperature and (C2.3), and then the candidate solution becomes the new current so-
then slowly reducing the temperature to obtain a new material with a lution, is added to the archive and solutions in the archive dominated
reinforced structure and the lowest internal energy state. Accepting by the candidate solution are removed.
worse solutions through the Metropolis criterion (Metropolis et al., Finally, in the third condition (C3), the candidate solution dom-
1953) is an essential property of simulated annealing that enables a inates the current solution and the current solution is removed from the
more extensive search by allowing the algorithm to escape from local archive, if this solution is in the archive. Again, three situations can
optima. Therefore, the parameter (T) in (6), plays an essential role in occur. First, at least one of the solutions from the archive dominate the
the method as it controls the optimization search by slowly reducing the candidate solution (C3.1) and the Pacp is computed with Δdom equal to
likelihood of accepting worse solutions (with worse values for the ob- the minimum value obtained by comparing the amount of domination
jectives) during the exploration of the solution space. Furthermore, the of each of the Sarch solutions that dominate Scand with Scand. Then the
algorithms stops when (T) reaches a low level and when there is a low candidate solution is accepted as the new current solution if the
percentage of accepted solutions. Metropolis criterion is met. If not accepted, the current solution be-
A pseudo-code describing the most important steps of the algorithm comes equal to the solution from the archive that served to determine
is shown in Fig. 1. Δdom. Second, the candidate solution is non-dominated by any solution
The many-objective simulated annealing initializes by loading the of the archive (C3.2), then the candidate solution is converted to the
WDN data from the EPANET file and by defining the current solution new current solution and is added to the archive. And in the third si-
(Scurr). Then the iterative process starts until stop criterion is met. This tuation, the candidate solution dominates at least one of solutions from
iterative procedure generates candidate solutions (Scand) that have to archive (C3.3) then the candidate solution becomes the new current
satisfy all constraints of the optimization model; a new candidate so- solution, is added to the archive, and the solutions in the archive
lution is generated and the four objectives (O) of the model are eval- dominated by the candidate solution are removed.
uated (O1 = PD, O2 = UD, O3 = CT and O4 = CE). Then, a process The algorithm is repeated for a number of iterations at each tem-
inspired on the on the work of Bandyopadhyay et al. (2008) to deal perature until equilibrium is reached (Fig. 1). The iterative process ends
with many-objectives, is followed by comparing the dominance status when a stopping criterion is met. At the end of the process, the archive
between candidate solution and the current solution (the last accepted is composed of a set of the found non-dominated solutions of the pro-
solution), and, if needed, the solutions stored in the archive. The ar- blem. These are usually called Pareto solutions (Pareto, 1896).
chive stores the non-dominated solutions found during the iteration When a large number of objectives are considered in the optimi-
search. To determine the Pacp, the amount of domination (Δdom) is zation model, a high number of non-dominated solutions can be found
computed in different forms (using (5)), also according to the dom- and therefore a large archive has to be used. In fact, it saves not only the
inance status between candidate and current solution and the archive objective function values of the solutions, but also the values for all
solutions (Sarch) that dominate candidate solution (Bandyopadhyay decision variables (such as all the pipe diameter sizes) that are asso-
et al., 2008). Three different situations (C), as shown in Fig. 1, can ciated with the solutions. This requires allocating a considerable
occur after comparing the dominance status. If current solution Scurr amount of computational memory to the optimization process
dominates candidate solution Scand than the situation is (C1) and the (Bandyopadhyay et al., 2008), which is why the size of the archive has
Pacp is computed with Δdom given by the mean value of the amount of to be carefully defined. Initially, it has to be set large enough to ac-
domination of the solutions Scurr and Sarch that dominate Scand. commodate as much information as possible, bearing in mind the
In case C1, all objective values determined for Scurr are better possible future number of non-dominated solutions that can be gener-
(lower) than or equal to those for Scand and at least one of the objectives ated, but simultaneously avoiding dimensionality problems. In general,
of Scurr has a value lower than the value of an objective in Scand (7): for the same number of iterations, the time required to solve a many-
∀ i ∈ 1,2,3,4: Oi (Scurr ) ≤ Oi (Scand ) objective problem is greater than the time required to solve the same
and ∃ i ∈ 1,2,3,4: Oi (Scurr ) < Oi (Scand ). problem but with fewer objectives. This is because, during the opti-
(7)
mization process of the simulated annealing algorithm, new solutions
In other words, the candidate solution is worse than the current are generated and compared with the non-dominated solutions found so
solution. In this case, the candidate solution can only be accepted as the far and saved in the archive. Since a large number of non-dominated
new current solution if the Metropolis criterion (Metropolis et al., 1953) solutions are found in many-objective problems, more time is needed to
is met. The Metropolis criterion is met when the acceptance probability perform these comparisons and thus more computational time is used
(Pacp), given by (6) is higher than a reference probability (Pref = random than is needed for problems with fewer objectives. The simulated an-
(0,1)) randomly generated. Therefore, if Pacp > Pref the candidate so- nealing algorithm makes use of a set of parameters that influence the
lution is accepted as the new current solution. The purpose of using the number of interactions of the algorithm and the corresponding com-
amount of domination concept (5) to determine the Pacp used in the putational time required to run the optimization. This means that to
Metropolis criterion, is to obtain high Pacp to solutions with small obtain solutions for the many-objective problems in a reasonable
amounts of domination and low Pacp for those with a high amount of amount of time these parameters must also be calibrated, considering

311
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

START PROGRAM
Load the WDN file (EPANET file)
Define current solution (Scurr) and evaluate objective functions
DO until stop criterion met
Generate a random candidate solution (Scand) that meets all constraints
Evaluate objective functions corresponding to candidate solution
IF current solution dominates candidate solution THEN (C1)
IF Metropolis criterion met THEN
Current solution equal to candidate solution
END IF
ELSE IF current solution and candidate solution are non-dominated with respect to each other THEN (C2)
IF at least one of the solutions from archive dominate candidate solution THEN (C2.1)
IF Metropolis criterion met THEN
Current solution equal to candidate solution
END IF
ELSE IF candidate solution is non-dominated by any solution of the archive THEN (C2.2)
Current solution equal to candidate solution
Copy candidate solution to the archive
ELSE candidate solution dominates at least one of the solutions from archive THEN (C2.3)
Current solution equal to candidate solution
Copy candidate solution to the archive
Remove from the archive solutions dominated by candidate solution
END IF
ELSE IF candidate solution dominates current solution THEN (C3)
Remove current solution from the archive if this solution is in the archive
IF at least one of the solutions from archive dominates candidate solution THEN (C3.1)
IF Metropolis criterion met THEN
Current solution equal to candidate solution
ELSE
Current solution equal to archive solution with minimum amount of domination to Scand
END IF
ELSE IF candidate solution is non-dominated by any solution of the archive THEN (C3.2)
Current solution equal to candidate solution
Copy candidate solution to the archive
ELSE candidate solution dominates at least one of the solutions from archive THEN (C3.3)
Current solution equal to candidate solution
Copy candidate solution to the archive
Remove from the archive solutions dominated by candidate solution
END IF
END IF
IF equilibrium at each annealing temperature is met THEN
Reduce the annealing temperature
END IF
END DO
END PROGRAM

Fig. 1. Many-objective simulated annealing pseudo-code (C1-C3 indicate the three main situations in the process).

the additional complexity of the many-objective problems relative to remaining nodes and pipes are included in the development areas with
problems with fewer objectives. the characteristics defined for this specific analysis as follows: nodes 18,
19, 20, 21 and 24 with elevation 15.3 m; node 22 with elevation
24.4 m; nodes 23, 25 and 26 with elevation of 36.6 m; nodes 18, 19, 20,
4. Results and discussion 21, 22, 23, 25 and 26 with the demand of 12.6 l/s; node 24 with the
demand of 31.6 l/s; pipes 35, 37, 38, 39, 41, 43, 45, 48, 49, 51, 52, 53
4.1. Case study and 55 with the length of 1,830 m; pipes 36, 42, 44, 46 and 47 with the
length of 3,660 m and pipes 40, 50 and 54 with the length of 2,745 m.
The application of the four objective model is demonstrated on a The commercial diameters shown in Table 1, which are taken from
WDN based on the one proposed by Lansey and Mays (1989), that has Lansey and Mays (1989), are used for the WDN pipe design. This table
been changed to include expansion areas. This is a hypothetical WDN shows the unit cost and the carbon emissions for each diameter per unit
but has similarities to real networks. This network has 55 pipes, 2 of length. The carbon emissions are computed by a process proposed by
identical pumps to transfer energy to the flow from a reservoir with a Marques et al. (2015c). The Hazen-Williams coefficient is equal to 120
fixed water level and 27 nodes. The layout of the network is presented for all commercial diameters.
in Fig. 2. The capital cost and the carbon emissions are calculated based on
To meet urban expansion, three development areas are considered the values in Table 1 and the capital cost of installing pumps in links 56
and represented in Fig. 2 by the network areas with orange (DA1), and 57 (Fig. 2), is computed according to Lansey and Mays (1989). To
green (DA2) and blue (DA3) hatching. These development areas can be compute the operating costs (calculated from the pumping energy
constructed in future stages of the planning horizon, in the wake of the costs) an energy unit cost of 0.12(USD)/kWh is used and an interest rate
scenarios that are described below. The characteristics of the nodes 1 to of 12% is assumed for the determination of the current value of the
17 and pipes 1 to 34 can be found in Lansey and Mays (1989). The

312
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Fig. 2. The layout of the water distribution network with three possible development areas.

Table 1 the city, if such development has occurred in the second stage. If DA1 is
Diameter, unit cost and carbon emissions. installed in the second stage, there is the possibility in the third stage to
Diameter Unit Carbon Diameter Unit Carbon expand to DA2 or DA3, or Maintain the system. If DA2 is installed in the
(mm) cost emissions (mm) cost emissions second stage, expansion to DA1, to DA3 or Maintain are the options. If
(USD/ (TonCO2/m) (USD/ (TonCO2/m) DA3 is selected in the second stage, expansion to DA1, to DA2 or
m) m) Maintain are the options. Finally, if Maintain is selected in the second
152 68 0.48 406 192 0.96
stage, it is admitted that the WDN will again Maintain the system “as is”
203 91 0.59 457 219 1.05 in the third stage. Each option has associated probabilities to indicate
254 113 0.71 508 248 1.14 the most probable alternatives that can be taken. For example, the
305 138 0.81 610 305 1.32 probability to develop areas DA1 and DA2 (0.3) in stage t = 2 is higher
356 164 0.87
than that to develop area DA3 (0.2) or Maintain (0.2) the system. Also,
if DA1 is chosen in t = 2, the most probable alternative in t = 3 is to
costs. The daily water use is divided into four demand conditions of 6 h develop area DA2 (0.7), which is close to DA1, and Maintain the system
has the lowest probability (0.1). The decisions that are taken under
each, whose value is given by multiplying the nodal demand by 0.67 for
the first demand condition from 0:00 to 6:00; by 1.28 for the second different decision paths establish possible future intervention scenarios
in the network, with probabilities given by multiplying the probabilities
demand condition from 6:00 to 12:00; by 1.08 for the third demand
condition from 12:00 to 18:00, and by 0.97 for the fourth demand in that decision path (those that are presented on the right-hand nodes
in Fig. 3). These values show scenarios 1 and 4 (probability of 0.21), as
condition from 18:00 to 24:00 (Lansey and Mays, 1989).
As stated, the hatched areas in Fig. 2 represent parts of the city that the most probable and scenarios 3 and 6 as the least probable (prob-
can be developed in different stages of the planning horizon. To design ability of 0.03). In real world applications, these probabilities can be
a network that can be expanded in the future it is advantageous to given by expert judgement.
embed the flexibility perspective from the beginning (or flexibility in The probabilities of these options are used in the many-objective
systems as defined by de Neufville and Scholtes, 2011). This work takes model to compute the cost objective (8) and the carbon emissions ob-
a planning horizon of 60 years divided into three 20-year stages, when jective (9):
new decisions can be taken. These are organized according to decision NSC NTI t

paths of options (or scenarios) that can occur over the planning horizon CT = Min Cph1 + ∑ ∑ ⎛⎜Cpht,s · ∏ prbkt,s⎞⎟
and are detailed in Fig. 3. s= 1 t= 2 ⎝ kt = 2 ⎠ (8)
In the first stage, a solution must be chosen to be implemented NSC NTI t
“now”, and therefore pipes 1 to 34 (Fig. 2) have to be designed and CE = Min CO2 ph1 + ∑ ∑ ⎛⎜CO2 pht,s · ∏ prbkt,s⎞⎟
installed in the initial stage. In the second stage, the authorities are s=1 t=2 ⎝ kt = 2 ⎠ (9)
planning to allow urban development in one of the three areas (DA1,
In (8), Cph1 = cost of the solution to be implemented in the first
DA2 and DA3) of the city shown in Fig. 2, or, if there is no need,
stage; Cpht,s = cost of future design solutions in scenario s for stage t;
maintain the same system (Maintain). In the third stage, the authorities
kt = time stage from t = 2 until the stage under analysis and
are also planning to allow urban development in an additional area of
Prbkt,s = probability of scenario s in time stage kt. Expression (8) is used

313
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

t=1 (0-19 years) t=2 (20-39 years) t=3 (40-59 years)

DA2 (0.7) Scenario 1 (0.21)

DA1 (0.3) DA3 (0.2) Scenario 2 (0.06)

Maintain (0.1) Scenario 3 (0.03)

DA1 (0.7) Scenario 4 (0.21)

DA2 (0.3) DA3 (0.2) Scenario 5 (0.06)

Maintain (0.1) Scenario 6 (0.03)

Initial design (1)
DA1 (0.3) Scenario 7 (0.06)

DA3 (0.2) DA2 (0.3) Scenario 8 (0.06)

Maintain (0.4) Scenario 9 (0.08)

Maintain (0.2) Maintain (1) Scenario 10 (0.2)

Fig. 3. Decision tree for the planning horizon and probabilities of occurrence (in brackets).

to minimize the total cost (CT) objective and is given by summing the This is because achieving a small improvement in the undelivered de-
cost of the initial solution (the first term of (8)) to be implemented in mand and pressure deficit objectives in this part of the solution space
the first stage and the future costs (second term of (8)), computed by a requires a large increase in the cost objective and a rise in the carbon
summing of all possible options, starting from the second time stage, emissions objective. Therefore, obtaining a solution with a very low or
given for future stages of the planning horizon. For each option, the zero undelivered demand, means accepting a costly solution with a very
costs are computed by multiplying the future cost of each design option high cost. This is important because it gives decision makers informa-
by the probability of taking that option. The objective to minimize the tion about a cost threshold level, (in this case, for CT lower than 0.2)
carbon emissions (CE) is given in (9) where CO2ph1 = carbon emissions above which the undelivered demand and pressure deficit values start
in the first stage; CO2pht,s = carbon emissions in scenario s for time to fall at the very low rate. The same conclusion can be drawn for the
stage t. The carbon emissions are evaluated by the methodology set out pressure deficit objective, as in solution set S1 all solution points are the
in Marques et al. (2015c), which considers the whole life cycle in- same colour (blue). The other part of the Pareto front, solution set S2,
cluding the extraction of raw materials, transport, manufacture, as- consists of a large number of solutions with low cost objective values
sembly, installation, disassembly, demolition and/or decomposition. and creates a surface in a three-dimensional space.
The carbon emissions are given by summing the emissions in the first In addition to the representation of the four-dimensional view (3D
time stage and by a weighted sum of the carbon emissions in all the axes and the colour scheme), the different combinations of two (out of
possible future options available. the four) objectives is explored to identify the benefits of using all four
objectives. The two-dimensional trade-offs are shown in the next figures
5. Results by means of dark blue points and lines representing the trade-offs be-
tween each combination of two objectives and grey points representing
In a four objective optimization model, it is challenging to extract the remaining non-dominated solutions obtained for the four-objective
information from a large number of non-dominated solutions found and problem.
from the configuration of the Pareto front. Fig. 4 includes the non- Fig. 5 shows a clear conflict between cost and undelivered demand
dominated solutions obtained by the many-objective simulated an- (5a). Practically the same format can be seen in trade-offs between cost
nealing method and were built in the Ahrens et al. (2005) visualization and pressure deficit in (5b).
application. The pressure deficit objective is represented using a colour Fig. 5 shows how the cost reduction can be achieved with increased
scheme ranging from blue to red, where the pressure deficit ranges from undelivered demand and pressure deficit. The trade-off between cost
0 (blue) to 1 (red) and the undelivered demand, total cost and carbon and undelivered demand can be explained as follows. The high cost
emission are plotted on the 3D axes of the figure. Furthermore, all four solutions contain large diameter pipes, which results in networks with
objectives are normalized between (0, 1). high pressure at nodes and thus with low undelivered demand. The
Two distinct parts can clearly be seen in the Pareto front of solutions differences in the design (decision space) of these solutions can be seen
(Fig. 4). Fronts of this type can occur when the trade-offs between the in Fig. 6. This figure details the solutions d1 and d2 marked in Fig. 5a).
objectives are dominated by one of them in some region of the solution These two particular solutions have been selected to point out the dif-
space. In Fig. 4b) these two parts are highlighted by sets of solutions ferences between designs in opposed areas of the solution space. They
identified as set S1 in purple, and set S2 in grey. Set S1 consists of a correspond to scenario 1 (a scenario with a high probability of occur-
small number of high cost solutions that create a Pareto front that is rence) that includes the urban development of DA1 in t = 2 and the
closer to a curve than a surface in a three-dimensional space. This in- urban development of DA2 in t = 3. However, it should be noted that
dicates that in this part of the solution space the trade-off between the the solutions include all the paths considered in Fig. 3.
objectives (represented in the axes) is controlled by the cost objective. In Fig. 6, the pipe diameters (displayed for each pipe) used in d1 are

314
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Fig. 4. a) Pareto set of solutions b) Pareto set of solutions divided into sets S1 and S2.

larger than in d2, which predominantly contains the minimum possible mainly the minimum possible diameter (152 mm), whereas in d4 the
diameter (152 mm). This increases the total cost objective of (CT) from pipes have larger diameters which enables a reduction in undelivered
32 × 107 (USD) in d2 to 40 × 107 (USD) in d1. However, the un- demand from 2414 L/s in d3 to 1355 L/s in d4 while the level of carbon
delivered demand objective (UD) is reduced from 1649 L/s in d2 to emissions is practically unchanged, at 5.5 × 106 TonCO2 for d3 and
202 L/s in d1. Furthermore, the pressure deficit objective (PD) is also 5.6 × 106 TonCO2 for d4. The same conclusions can be drawn when
reduced from 2661 m in d2 to 403 m in d1 by increasing the cost relating pressure deficits and carbon emissions.
(Fig. 5b). In Fig. 9a) we can see that solutions with the lowest cost also have
It is interesting to see in Fig. 7a) that undelivered demand can be the lowest values of carbon emissions (solutions near to the graph
reduced from 1 to 0.55 while maintaining low levels of carbon emis- origin). However, the reduction in carbon emissions is slower for costs
sions of 0–0.05. The same trend can be seen in (7b), where pressure above 0.2 and is faster for solutions with costs below 0.2.
deficits are related to carbon emissions. Finally Fig. 9b) shows the relation between undelivered demand
For high undelivered demand solutions, the network pipe sizes are and pressure deficits. So far, Figs. 5 and 7 do not show any advantage of
smaller (with high head losses) and pumps have to guarantee minimum including both objectives related to the undelivered demand and the
pressure requirements and thus have high energy consumption. As pressure deficit. However, Fig. 9b) identifies solutions with similar
network pipe sizes start to increase, the undelivered demand is reduced, pressure deficits but with different undelivered demand values. The
however, the carbon emissions remain at the same level because the main advantage of considering minimization of both the pressure def-
reduction of carbon emissions related to energy expenditure is offset by icit and the undelivered demand as objectives is that solutions with
the increase in carbon emissions due to the use of high diameter pipes. lower undelivered demand for practically the same level of pressure
This can be seen in Fig. 8, which compares the design of solutions d3 deficit can be identified. This can be seen from solutions d5 and d6,
and d4 marked in Fig. 7a). indicated in (9b) by diamonds, and which are presented in Fig. 10.
Comparing solutions d3 and d4, the pipe diameters used in d3 are Furthermore, to evidence the advantages of considering the undelivered

Fig. 5. Trade-offs between pairs of objectives. a) CT vs. UD, b) CT vs PD.

315
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Fig. 6. Solutions d1 and d2 for scenario 1 that compare designs and values of CT, UD and PD.

demand as an objective in the model, the same optimization problem from 26 L/s to 16 L/s in t = 3. In node 16 undelivered demand falls
was again solved, without considering undelivered demand as an ob- from 36 L/s to 31 L/s in t = 2 and from 42 L/s to 37 L/s in t = 3, and
jective but retaining all the other objectives and conditions. One of the this is related to the increased hydraulic capacity of the pipes that
solutions obtained without the undelivered demand objective but with supply these nodes. From the SNUD solution obtained without con-
practically the same level of pressure deficit as in solutions d5 and d6 is sidering the minimization of UD as objective in the optimization model,
also presented in Fig. 10, as SNUD (Solution obtained with No Un- it can be seen that the UD is high even for the same level of pressure
delivered Demand objective). To understand the impact in terms of deficit in d6, (UD = 128 L/s in SNUD) in t = 2 and (UD = 182 L/s in
nodal conditions, these values are presented for the particular condi- SNUD) in t = 3. The differences in the design of SNUD and d6 (the
tions of scenario 1. solution with the lowest UD values) are mainly related to the lower
The designs in Fig. 10 include the pipe diameters of each network network capacity in the central part of the city considered in solution
link and also the values rounded to units of pressure deficit and the SNUD, compared with solution d6, which increases the value un-
undelivered demand at each node for t = 2 and for t = 3. From d5 and delivered demand value at nodes 6, 7, 9 and 10 in this part of the city.
d6 it can be seen that the pressure deficits are practically the same Comparing the solutions and objectives shows that all four objec-
(PD = 165 m in d5 and PD = 167 m in d6) in t = 2 and (PD = 253 m in tives have different trade-offs with one another and consequently for-
d5 and PD = 255 m in d6) in t = 3, but the undelivered demand is mulations without one of these objectives would predispose decision
higher for d5 than for d6 (UD = 114 L/s in d5 and UD = 104 L/s in d6) makers towards neglecting potential improvements in one of the ob-
in t = 2 and (UD = 173 L/s in d5 and UD = 146 L/s in d6) in t = 3. jectives relative to the other three. Based on the results from the case
Analysis of the network designs for these solutions makes it possible to study, it is possible to conclude that the solution space has two parts, a
conclude that the decrease in undelivered demand offered by solution curve that indicates the control of the cost objective relative to the
d6 is attained by the higher capacity introduced in the pipes that supply others for high cost solutions and a surface part that identifies com-
nodes 9 and 16 with large pipe diameters. These are the network nodes promises between all objectives. This work considers a looped network,
with high demand and, therefore, solutions that cut undelivered de- which is the most common type of real world network. Furthermore,
mand have to increase the network capacity to supply these nodes. given the energy required to deliver water in this WDN, increasing the
Therefore, solution d6 uses a pipe diameter of 305 mm from node 10 to pipe capacity not only increases the network reliability (cutting un-
9 whereas solution d5 uses the minimum pipe diameter of 152 mm. delivered demand and pressure deficit), it also reduces the energy
Also in solution d6, a pipe diameter 508 mm is used from node 11 to consumption of pumps by reducing head losses. Therefore, for real
node 16, while solution d5 uses 305 mm. With these diameters, the water distribution networks with energy expenditure and looped
undelivered demand at node 9 falls from 21 L/s to 11 L/s in t = 2 and structures, trade-offs between objectives are expected.

Fig. 7. Trade-offs between pairs of objectives. a) UD vs. CE, b) PD vs. CE.

316
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Fig. 8. Solutions d3 and d4 for scenario 1 that compare design and values of CE, UD and PD.

6. Conclusions growing awareness of the implications of climate change.

Analysis of the results leads to some interesting conclusions and
The research set out in this paper was motivated by the interest in demonstrates the relevance of using this approach. The Pareto front is
solving problems not yet dealt with in the literature: those cast as many- composed of two main parts, one with a low density of solutions which
objective problems to develop flexible solutions for the design of a WDN are high in cost, and a second high-density part with low cost solutions
using a multi-stage scheme and including demand uncertainty issues. represented by a surface. This gives decision makers information about
An enhanced simulated annealing algorithm was developed that is a cost threshold level beyond which the undelivered demand and
capable of dealing with challenges encountered in MaOPs. Such chal- pressure deficit start to decrease at a high rate. Also, an increase in
lenges include handling the ineffectiveness of the Pareto dominance carbon emissions means that the objectives of undelivered demand and
relation in high-dimensional space and the need to create enough se- pressure deficit can be reduced. For undelivered demand and pressure
lection pressure in problems with a higher number of objectives. The deficit between 1 and 0.55, it is possible to keep a very low level of
methodology was fully analysed by applying it to a WDN case study for carbon emissions mainly because these solutions, with small diameter
a planning horizon of 60 years, divided into three stages of 20 years. In pipes, have large operating costs related to the high head losses. As
these stages, different decisions with specific probabilities can be taken network pipe sizes increase, the undelivered demand is reduced and the
according to the different decision paths. The results represent the non- carbon emissions remain at the same level, because the increase in
dominated solutions found by the algorithm, including a large number carbon emissions due to the use of pipes with high diameter is offset by
of staged solutions for the design of WDNs. This means that decision the reduction of carbon emissions related to energy expenditure. This is
makers can select a solution or a small set of solutions to evaluate in demonstrated by showing two solutions with practically the same level
more detail. Additionally, they can implement the first stage design of carbon emissions (5.5 × 106 TonCO2 in d3 and 5.6 × 106 TonCO2 in
without compromising future developments during the planning hor- d4) but with different undelivered demand values (2414 L/s in d3 and
izon. Furthermore, various combinations of two of the four objectives 1355 L/s in d4).
are explored to ascertain the benefits of using all four objectives. The Finally, this work also identifies the trade-offs between undelivered
objectives considered were the pressure deficit, undelivered demand, demand and pressure deficits. The analysis indicates there is the pos-
cost and carbon emissions. The pressure deficit and undelivered de- sibility to select more appropriate solutions in terms of undelivered
mand objectives were used in the optimization model to incorporate demand or pressure deficits, and therefore the consideration of these
network reliability by evaluating the conditions of insufficient pressure two objectives makes it possible to differentiate among these solutions.
and undelivered demand. The cost objective was used due to the high Comparisons made it possible to identify solutions with different levels
cost of these systems and carbon emissions were used due to the of undelivered demand (UD = 114 L/s in d5 and UD = 104 L/s in d6) in

Fig. 9. Trade-offs between pairs of objectives. a) CT vs CE and b) UD vs PD.

317
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Fig. 10. Solutions d5, d6 and SNUD for scenario 1 that compare designs and nodal values of PD and UD for stage t = 2 and t = 3.

t = 2 and with very similar pressure deficit values (UD = 173 L/s in d5 Acknowledgements
and UD = 146 L/s in d6) in t = 3. These differences are mainly related
to the reduction of the undelivered demand at the network nodes with This work is partially supported by the Portuguese Foundation for
high demand. This is, in fact, an important result as it proves the im- Science and Technology under project grant UID/Multi/00308/2013.
portance of taking the undelivered demand minimization objective into
account at nodes with high demand, which are usually the critical References
nodes of the network that supply important facilities. For the case study
in particular, the analysis of the network designs of these solutions Ahrens, J., Geveci, B., Law, C., 2005. ParaView: an End-user Tool for Large-data
shows that the fall in undelivered demand offered by solution d6 was Visualization. Visualization Handbook, vol. 836. pp. 717–731. https://doi.org/10.
1016/B978-012387582-2/50038-1.
attained by assigning higher capacity to the network pipes that supply Alperovits, E., Shamir, U., 1977. Design of optimal water distribution systems. Water
the nodes with high demand. Resour. Res. 13 (6), 885–900. https://doi.org/10.1029/WR013i006p00885.
Additionally, the advantages of considering undelivered demand as Atiquzzaman, M., Liong, S.-Y., Yu, X., 2006. Alternative decision making in water dis-
tribution network with NSGA-II. J. Water Resour. Plann. Manag. 132 (2), 122–126.
an objective in the model were further explored by solving the same https://doi.org/10.1061/(ASCE)0733-9496(2006)132:2(122).
optimization problem without considering it. The results clearly show Bandyopadhyay, S., Saha, S., Maulik, U., Deb, K., 2008. A simulated annealing-based
that undelivered demand decreases for the same level of pressure deficit multiobjective optimization algorithm: AMOSA. IEEE Trans. Evol. Comput. 12 (3),
269–283. https://doi.org/10.1109/TEVC.2007.900837.
if the minimization of the undelivered demand objective is considered
Basupi, I., Kapelan, Z., 2015a. Evaluating flexibility in water distribution system design
in the model. From the solutions, it is possible to confirm that the un- under future demand uncertainty. J. Infrastruct. Syst. 21 (2), 04014034. https://doi.
delivered demand increase (for solution SNUD with practically the org/10.1061/(ASCE)IS.1943-555X.0000199.
Basupi, I., Kapelan, Z., 2015b. Flexible water distribution system design under future
same level of pressure deficit as in d6) is related to the decrease in
demand uncertainty. J. Water Resour. Plann. Manag. 141 (4), 04014067. https://doi.
network pipe capacity in the areas of the city with the highest con- org/10.1061/(ASCE)WR.1943-5452.0000416.
sumption of water. Chand, S., Wagner, M., 2015. Evolutionary many-objective optimization: a quick-start
Flexible solutions make it possible to adapt the infrastructure as a guide. Surv. Oper. Res. Manag. Sci. 20 (2), 35–42. https://doi.org/10.1016/j.sorms.
2015.08.001.
function of different scenarios predefined in a decision tree. Each of the Creaco, E., Franchini, M., Walski, T., 2014. Accounting for phasing of construction within
solutions obtained for the four-objective optimization model, represents the design of water distribution networks. J. Water Resour. Plann. Manag. 140 (5),
a particular (single) design considering all scenarios of the decision 598–606. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000358.
Creaco, E., Franchini, M., Walski, T.M., 2015. Comparison of various phased approaches
tree. This means that the solution can cope with all pre-established for the constrained minimum-cost design of water distribution networks. Urban
scenarios and thus the idea is not to select a solution for a single optimal Water J. 13 (3), 270–283. https://doi.org/10.1080/1573062X.2014.991330.
scenario but to select solutions that could function as best as possible Cunha, M.C., Sousa, J., 1999. Water distribution network design optimization: simulated
annealing approach. J. Water Resour. Plann. Manag. 125 (4), 215–221. https://doi.
for a set of different scenarios. A selection of the one (optimal) solution org/10.1061/(ASCE)0733-9496(1999)125:4(215).
would be difficult as by the very definition of the Pareto set, each so- de Neufville, R., Scholtes, S., 2011. Flexibility in Engineering Design. Engineering sys-
lution in it is better in at least one objective than any other solution. In tems. pp. 293.
Deb, K., 1999. Multi-objective genetic algorithms: problem difficulties and construction of
the latter case, a single solution has to be selected by a decision maker
test problems. Evol. Comput. 7 (3), 205–230. https://doi.org/10.1162/evco.1999.7.
in order to be able to implement it. 3.205.
Farina, M., Amato, P., 2002. On the optimal solution definition for many-criteria opti-
mization problems. In: Annual Conference of the North American Fuzzy Information

318
J. Marques et al. Journal of Environmental Management 226 (2018) 308–319

Processing Society, 2002, pp. 233–238. https://doi.org/10.1109/NAFIPS.2002. and wastewater engineering – definitions, literature and decision guide. J. Environ.
1018061. Manag. 149, 271–281. https://doi.org/10.1016/j.jenvman.2014.09.031.
Fu, G., Kapelan, Z., 2011. Fuzzy probabilistic design of water distribution networks. Stokes, C.S., Simpson, A.R., Maier, H.R., 2015. A computational software tool for the
Water Resour. Res. 47 (5), W05538. https://doi.org/10.1029/2010WR009739. minimization of costs and greenhouse gas emissions associated with water distribu-
Fu, G., Kapelan, Z., Kasprzyk, J.R., Reed, P., 2013. Optimal design of water distribution tion systems. Environ. Model. Software 69, 452–467. https://doi.org/10.1016/J.
systems using many-objective visual analytics. J. Water Resour. Plann. Manag. 139 ENVSOFT.2014.11.004.
(6), 624–633. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000311. Suman, B., Kumar, P., 2006. A survey of simulated annealing as a tool for single and
Geem, Z.W., 2006. Optimal cost design of water distribution networks using harmony multiobjective optimization. J. Oper. Res. Soc. 57 (10), 1143–1160. https://doi.org/
search. Eng. Optim. 38 (3), 259–277. https://doi.org/10.1080/ 10.1057/palgrave.jors.2602068.
03052150500467430. Tanyimboh, T.T., Seyoum, A.G., 2016. Multiobjective evolutionary optimization of water
Giustolisi, O., Berardi, L., 2009. Prioritizing pipe replacement: from multiobjective ge- distribution systems: exploiting diversity with infeasible solutions. J. Environ.
netic algorithms to operational decision support. J. Water Resour. Plann. Manag. 135 Manag. 183, 133–141. https://doi.org/10.1016/j.jenvman.2016.08.048.
(6), 484–492. https://doi.org/10.1061/(ASCE)0733-9496(2009)135:6(484). Wagner, J.M., Shamir, U., Marks, D.H., 1988. Water distribution reliability: simulation
Herstein, L.M., Filion, Y.R., Hall, K.R., 2009. Evaluating environmental impact in water methods. J. Water Resour. Plann. Manag. 114 (3), 276–294. https://doi.org/10.
distribution system design. J. Infrastruct. Syst. 15 (3), 241–250. https://doi.org/10. 1061/(ASCE)0733-9496(1988)114:3(276).
1061/(ASCE)1076-0342(2009)15:3(241). Wang, Q., Guidolin, M., Savic, D., Kapelan, Z., 2015. Two-objective design of benchmark
Huang, D., Vairavamoorthy, K., Tsegaye, S., 2010. Flexible design of urban water dis- problems of a water distribution system via MOEAs: towards the best-known ap-
tribution networks. In: World Environmental and Water Resources Congress 2010, proximation of the true Pareto front. J. Water Resour. Plann. Manag. 141 (3),
pp. 4225–4236. https://doi.org/10.1061/41114(371)430. 04014060. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000460.
Jiang, S., Yang, S., 2017. A strength Pareto evolutionary algorithm based on reference Wu, W., Maier, H.R., Simpson, A.R., 2013. Multiobjective optimization of water dis-
direction for multiobjective and many-objective optimization. IEEE Trans. Evol. tribution systems accounting for economic cost, hydraulic reliability, and greenhouse
Comput. 21 (3), 329–346. https://doi.org/10.1109/TEVC.2016.2592479. gas emissions. Water Resour. Res. 49 (3), 1211–1225. https://doi.org/10.1002/wrcr.
Kapelan, Z.S., Savic, D.A., Walters, G.A., 2005. Multiobjective design of water distribution 20120.
systems under uncertainty. Water Resour. Res. 41 (11), W11407. https://doi.org/10. Wu, W., Simpson, A.R., Maier, H.R., 2008. Multi-objective genetic algorithm optimisation
1029/2004WR003787. of water distribution systems accounting for sustainability. In: Proceedings of Water
Keedwell, E., Khu, S.T., 2004. Hybrid Genetic Algorithms for Multi-objective Down under 2008, pp. 1750–1761.
Optimisation of Water Distribution Networks. Springer, Berlin, Heidelberg, pp. Wu, W., Simpson, A.R., Maier, H.R., 2010. Accounting for greenhouse gas emissions in
1042–1053. https://doi.org/10.1007/978-3-540-24855-2_115. multiobjective genetic algorithm optimization of water distribution systems. J. Water
Khu, S.T., Keedwell, E., 2005. Introducing more choices (flexibility) in the upgrading of Resour. Plann. Manag. 136 (5), 146–155. https://doi.org/10.1061/(ASCE)WR.1943-
water distribution networks: the New York city tunnel network example. Eng. Optim. 5452.0000072.
37 (3), 291–305. https://doi.org/10.1080/03052150512331303445.
Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P., 1983. Optimization by simulated annealing.
Science 220 (4598), 671–680. https://doi.org/10.1126/science.220.4598.671. Nomenclature
Lansey, K.E., Mays, L.W., 1989. Optimization model for water distribution system design.
J. Hydraul. Eng. 115 (10), 1401–1418. https://doi.org/10.1061/(ASCE)0733- CE: carbon emissions
9429(1989)115:10(1401). CO2ph1: carbon emissions in the first stage; CO2pht,s carbon emissions in scenario s for
Li, M., Yang, S., Liu, X., 2015. Bi-goal evolution for many-objective optimization pro- time stage t
blems. Artif. Intell. 228, 45–65. https://doi.org/10.1016/j.artint.2015.06.007. Cph1: cost of the solution to be implemented in the first stage
Maier, H.R., Simpson, A.R., Zecchin, A.C., Foong, W.K., Phang, K.Y., Seah, H.Y., Tan, C.L., Cpht,s: cost of future design solutions in scenario s for stage t
2003. Ant colony optimization for design of water distribution systems. J. Water CT: total costs
Resour. Plann. Manag. 129 (3), 200–209. https://doi.org/10.1061/(ASCE)0733- d: demand condition
9496(2003)129:3(200). DA: development area
Mala-Jetmarova, H., Sultanova, N., Savic, D., 2018. Lost in optimisation of water dis- Ddl: supply delivered
tribution systems? A literature review of system design. Water 10 (3), 307. https:// Ddln,d,t,s: supply delivered at node n for demand condition d for time stage t and in sce-
doi.org/10.3390/w10030307. nario s
Marques, J., Cunha, M., Savić, D.A., 2015a. Multi-objective optimization of water dis- Δdom: domination
tribution systems based on a real options approach. Environ. Model. Software 63 (1), Drq: demand required
1–13. https://doi.org/10.1016/j.envsoft.2014.09.014. Drqn,d,t,s: demand required at node n for demand condition d for time stage t and in
Marques, J., Cunha, M., Savić, D.A., 2015b. Using real options in the optimal design of scenario s
water distribution networks. J. Water Resour. Plann. Manag. 141 (2), 04014052. kt: time stage from t = 2 until the stage under analysis
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000448. n: network node
Marques, J., Cunha, M., Savić, D.A., 2015c. Using real options for an eco-friendly design NDC: number of demand conditions
of water distribution systems. J. Hydroinf. 17 (1), 20–35. https://doi.org/10.2166/ NNS: number of network nodes
hydro.2014.122. NOB: number of objectives
Marques, J., Cunha, M., Savić, D., Giustolisi, O., 2017. Water network design using a NSC: number of scenarios
multiobjective real options framework. J. Optim. 2017, 1–13. https://doi.org/10. NTI: number of time stages
1155/2017/4373952. O: objective
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953. Pacp: acceptance probability
Equation of state calculations by fast computing machines. J. Chem. Phys. 21 (6), Pavl: available nodal pressure
1087–1092. https://doi.org/10.1063/1.1699114. PD: pressure deficits
Morley, M.S., Tricarico, C., 2008. Pressure Driven Demand Extension for EPANET Pmin: minimum nodal pressure
(EPANETpdd). Centre for Water Systems, pp. 10 Technical Report. Pn,d,t,s: pressure at node n for demand condition d for time stage t and in scenario s
Pareto, V., 1896. Cours D'Economie Politique. F. Rouge, pp. 430. Prbkt,s: probability of scenario s in time stage kt
Purshouse, R.C., Fleming, P.J., 2003. Evolutionary many-objective optimisation: an ex- Pref: reference probability
ploratory analysis. In: 2003 Congress on Evolutionary Computation, vol. 3. pp. Prq: required nodal pressure
2066–2073. https://doi.org/10.1109/CEC.2003.1299927. Prqmin,n,d: minimum required pressure at node n for demand condition d
Purshouse, R.C., Fleming, P.J., 2007. On the evolutionary optimization of many con- Ri: range of objective i
flicting objectives. IEEE Trans. Evol. Comput. 11 (6), 770–784. https://doi.org/10. s: scenario
1109/TEVC.2007.910138. Sa: solution “a”
Savic, D., 2002. Single-objective vs. Multiobjective optimisation for integrated decision Sarch: solution from archive
support. In: Integrated Assessment and Decision, pp. 7–12. Sb: solution “b”
Savic, D.A., Walters, G.A., 1997. Genetic algorithms for least-cost design of water dis- Scand: candidate solution
tribution networks. J. Water Resour. Plann. Manag. 123 (2), 67–77. https://doi.org/ Scurr: current solution
10.1061/(ASCE)0733-9496(1997)123:2(67). SNUD: solution obtained with No Undelivered Demand objective
Sengupta, R., Saha, S., 2018. Reference point based archived many objective simulated t: time stage
annealing. Inf. Sci. https://doi.org/10.1016/J.INS.2018.05.013. (in press). T: temperature parameter
Shokoohi, M., Tabesh, M., Nazif, S., Dini, M., 2017. Water quality based multi-objective UD: undelivered demand
optimal design of water distribution systems. Water Resour. Manag. 31 (1), 93–108. WDNs: water distribution networks
https://doi.org/10.1007/s11269-016-1512-6.
Spiller, M., Vreeburg, J.H.G., Leusbrock, I., Zeeman, G., 2015. Flexible design in water

319