Scalable Pareto set generation for multiobjective codesign problems in water distribution networks: a continuous relaxation approach
 1.1k Downloads
 5 Citations
Abstract
In this paper, we study the multiobjective codesign problem of optimal valve placement and operation in water distribution networks, addressing the minimization of average pressure and pressure variability indices. The presented formulation considers nodal pressures, pipe flows and valve locations as decision variables, where binary variables are used to model the placement of control valves. The resulting optimization problem is a multiobjective mixed integer nonlinear optimization problem. As conflicting objectives, average zone pressure and pressure variability can not be simultaneously optimized. Therefore, we present the concept of Pareto optima sets to investigate the tradeoffs between the two conflicting objectives and evaluate the best compromise. We focus on the approximation of the Pareto front, the image of the Pareto optima set through the objective functions, using the weighted sum, normal boundary intersection and normalized normal constraint scalarization techniques. Each of the three methods relies on the solution of a series of singleobjective optimization problems, which are mixed integer nonlinear programs (MINLPs) in our case. For the solution of each singleobjective optimization problem, we implement a relaxation method that solves a sequence of nonlinear programs (NLPs) whose stationary points converge to a stationary point of the original MINLP. The relaxed NLPs have a sparse structure that come from the sparse water network graph constraints. In solving the large number of relaxed NLPs, sparsity is exploited by tailored techniques to improve the performance of the algorithms further and render the approaches scalable for large scale networks. The features of the proposed scalarization approaches are evaluated using a published benchmarking network model.
Keywords
Multiobjective optimization Scalarization strategies Mixed integer nonlinear programs Complementarity constraints Water distribution networks Pressure management1 Introduction
The optimal operation of water distribution networks (WDNs) requires the satisfaction of multiple criteria, some of which may be conflicting, in order to deliver increasing water demand costefficiently (Newman et al. 2014). Some objectives include the reduction of leakage, improvements in network resilience, optimization of water quality, and an efficient use of energy in pumping. The optimal management of pressure is one of the most effective methods to reduce leakage; a small decrease in average zone pressure has been shown to result in significant reductions in leakage (Lambert 2001). The sectorization of water distribution systems has also provided some benefit in this respect. Networks are subdivided into smaller sectors, called District Metered Areas (DMAs), so that flow into and out of each DMA is continuously monitored, improving the management of pressure and leakage (Farley and Trow 2003). On the other hand, this has severely reduced network redundancy, affecting resilience and water quality negatively. In particular, the action of valves to sectorize the network and reduce average zone pressure can generate high diurnal pressure variability across the nodes (Wright et al. 2014; Wright et al. 2015). Over the past few years, various studies have established the influence of pressure variability on pipe failures (Lambert and Thornton 2011). Moreover, the recent studies MartínezCodina et al. (2015); Hoskins and Stoianov (2015); Rezaei et al. (2015) have highlighted the influence of different pressure variability indicators on the probability of occurrence of pipes’ failures. As a consequence, an effective pressure management has to consider both the minimization of average zone pressure and pressure variability.
As standard in WDN management, here we consider pressure management that is actuated by controlling pressure reducing valves (PRVs), which regulate the pressure at their downstream node, and boundary valves, which can allow a range of pressure differentials across a pipe as their setting is varied between fully open, closed or anything in between the two. In standard approaches, these valves are optimally controlled once engineers have installed them at locations chosen using practical knowledge.
In this paper, we depart from the standard practice and consider the problem of optimizing the location of the actuator (i.e. boundary valves and PRVs), and the optimal pressure settings simultaneously – what is referred to as a codesign optimization problem. Although such an approach generally offers better performance, the mathematical formulation for the simultaneous optimal design and operation problem presents significant challenges; it requires the solution of a difficult nonlinear optimization problem with both continuous and discrete variables. In the present study we consider nodal pressures, pipe flows and valve locations as decision variables, where binary variables are needed to model the placement of valves. In addition, the inclusion of hydraulic conservation laws results in nonconvex, nonlinear constraints. In this work, we investigate the codesign problem with respect to two objectives, i.e. the minimization of average zone pressure and pressure variability via the optimal placement and control of valves. Therefore, a multiobjective optimization problem is formulated to establish the tradeoffs between these two criteria. The resulting problem is a multiobjective mixed integer nonlinear program (MINLP).
Due to their underlying complex structure, multiobjective codesign optimization problems for water distribution networks are commonly studied using evolutionary algorithms (EAs), which can be easy to implement by coupling them with simulation software (Maier et al. 2014). However, EAs have some disadvantages. Firstly, they do not guarantee optimality of the solutions, not even local optimality. Secondly, a large number of function evaluations and simulations of the hydraulic model are required in order to generate useful solutions, which do not scale well with the size of the search space. Unlike mathematical optimization approaches, evolutionary methods do not explicitly or accurately handle nonlinear constraints.
Since optimal control profiles for pressure management need to satisfy strict physical, quality and economic constraints, we seek a method that guarantees an accurate handling of constraints. Moreover, we aim to propose a scalable approach for large scale water systems, exploiting the particular sparse structure of water distribution networks. Therefore, in the present work we propose the application of stateoftheart optimization methods for the solution of multiobjective codesign problems for water distribution networks.
Although the natural aspiration is to find a feasible optimal configuration for all the objectives, typically no single solution that simultaneously optimises all conflicting objectives exists. Therefore, the mathematical notion of Pareto optimality is adopted to characterise the best compromises between conflicting objectives.
A feasible point for a multiobjective optimization problem is a global Pareto optimum if there is no other feasible point that improves one or more of the objectives without making another worse (Miettinen 1998). A decision vector is called a local Pareto optimum if it is a global Pareto optimum only when the optimization is restricted to its neighbourhood. However, for nonconvex problems all the deterministic gradientbased methods for the generation of Pareto optima guarantee convergence only to local optima. Since in the present work we consider a nonconvex multiobjective optimization problem, we will refer to local Pareto optima simply as Pareto optima, unless otherwise stated.
For a generic nonlinear optimization problem with m objectives, the set of Pareto optima is uncountable and very complex; it is a stratified set, a pairwise disjoint union of differentiable manifolds of dimension (m−1) with boundaries and corners of lower dimensions (Wan 1978; Smale 1976). For instance, the Pareto set is a pairwise disjoint union of segments of curves for the case of 2 objectives. Continuation methods (Hillermeier 2001) try to track the local structure of the Pareto optimal manifolds near some initial solutions. However, only few continuation methods can deal with the nondifferentiable breakpoints introduced by inequality constraints (Martin et al. 2014) and handle nonconvex and nonlinear problems, where the Pareto set is disconnected (Hartikainen and Lovison 2014). Moreover, in the case of multiobjective mixed integer nonlinear optimization problems, as considered in the present work, discrete decision variables introduce additional complexities – the Pareto set may be composed of disconnected segments of curves and isolated points (Das 2000). Since in general the differential composite structure of the Pareto set is not explicitly known, algorithmic mathematical methods focus on the space of the objective functions rather than the decision variable space; for an exhaustive survey see Miettinen (1998), Marler and Arora (2004) and Andersson (2000).
In the present work, we are interested in techniques which approximate the image of the Pareto set through the objective functions, called Pareto front. The aim is to provide a wide and uniform distribution of points of the front so that the decision maker can choose the Pareto optimal solution which better fits its needs. This is called an a posteriori approach since the articulation of the preferences by the decision maker is made after the generation of the approximated Pareto front (Marler and Arora 2004). Common a posteriori approaches include scalarization methods (Marler and Arora 2004; Das and Dennis 1998; Messac et al. 2003; Kim and De Weck 2005), which parametrize the multiobjective problem into a series of singleobjective optimization problems that can be solved using standard nonlinear programming techniques. A popular scalarization method is the weighted sum (WS) of the objectives (Marler and Arora 2004). Despite its simplicity, some disadvantages of the WS approach include its inability to generate Pareto points in the nonconvex parts of the Pareto front and the fact that an even spread of weights may not correspond an uniform distribution of points of the front (Das and Dennis 1997). In the past 20 years new methods have been proposed to deal with these drawbacks, in particular, the Normal Boundary Intersection (NBI) method (Das and Dennis 1998) and the Normalized Normal Constraint (NNC) method (Messac et al. 2003). These techniques can be combined with efficient gradientbased algorithms to find optimal solutions even for largescale and highly constrained optimization problems, see for example (Logist et al. 2010).
However, these methods have been developed for continuous problems and their effectiveness for mixed integer problems is not (well) studied. The nonconnectivity of the Pareto sets and the presence of isolated Pareto optimal points can cause severe problems for the NBI method. On the other hand, WS and NNC methods are less likely to suffer from the disconnected nature of the Pareto set. In fact, we will show this to be the case in Section 5, where disconnected Pareto branches in Fig. 6 are discussed.
The application of the scalarization methods to multiobjective MIPs considered in this article relies on the solution of a series of singleobjective nonconvex optimization problems belonging to the class of mixed integer nonlinear programming (MINLP). The solution of these problems requires handling nonconvex nonlinear constraints in a discrete framework and hence it is particularly challenging, for a general surveys on MINLP see Lee and Leyffer (2012). A possible approach for nonconvex MINLP is the application of a branch and bound algorithm to find at least local solutions (D’Ambrosio and Lodi 2013). However, in our case this would result in infeasible computational time for large scale water distribution networks, since it is necessary to solve a series of MINLPs to generate the Pareto front. Since the integer variables involved in our optimization problem are binaries, they can be reformulated as complementarity constraints, enforcing the variables to take only one of the two complementary values 0 and 1. Nonetheless, the complementarity constraints result is a special feasible set which violates standard requirements for the application of gradientbased solution algorithms and specialized approaches are needed to solve these problems (Scheel and Scholtes 2000; Leyffer 2006; Raghunathan and Biegler 2005; Ralph and Wright 2004; Hu and Ralph 2004). In particular we focus on relaxation methods: the MPPC is converted into a series of nonlinear programs with relaxed feasible sets which satisfy standard regularity assumptions for the application of stateoftheart nonlinear programming (NLP) solvers. The sequence of solutions of the relaxed NLP subproblems will converge to a solution for the original MPCC, see Herty and Steffensen (2012), Leyffer et al. (2006) and Scholtes (2001).
In this article, in order to solve the difficult multiobjective mixedinteger nonlinear problems, we propose coupling a relaxation approach for solving MIPs with stateoftheart scalarization methods for generating Pareto fronts. Moreover, sparse techniques are used within the NLP solvers to improve the performance of relaxation methods since the large scale mixed integer nonlinear programs that arise in the framework of optimal codesign for water distribution networks have sparse constraints. The Jacobians and Hessians of the relaxed NLPs retain this sparsity, which we take advantage of.
In the following section, we motivate and formulate the multiobjective optimization problem for the minimization of average zone pressure and pressure variability. In Section 3, three different scalarization schemes, weighted sum, normal boundary intersection and normalized normal constraint methods are studied. We then outline the relaxation approach for the solution of mixed integer nonlinear programs in Section 4. Finally, in the last section we apply the presented methods to approximate the Pareto front of our multiobjective optimization problem using a selected benchmarking network as case study.
2 Problem formulation
Consider a water distribution networks with n _{0} fixed head nodes (e.g. water sources), n _{ n } demand nodes and n _{ p } pipes. This can be modeled as a direct graph \(\mathcal {G}=(V,E)\) where V is the set of nodes (V = n _{ n } + n _{0}) and E the set of links. Since the direction of flow in each pipe is itself an unknown, we associate two directed flows with each pipe. The bidirectional flows are included by introducing, for each physical pipe j, the two links j and j + n _{ p } and so the graph representation has that E=2n _{ p }. In this work, the optimization is performed by considering the network operation over an extended time horizon – we consider n _{ l } different time steps in a diurnal operation and associated demand conditions.
In conclusion, the unknowns of the optimization problem can be represented by the vector x=[p ^{1},q ^{1},…,\( p^{n_{l}},q^{n_{l}},v]^{T} \in \mathbb {R}^{N}\) with N = n _{ l }(n _{ n }+2n _{ p })+2n _{ p }. Note that B= {n _{ l }(n _{ n }+2n _{ p })+1,…n _{ l }(n _{ n }+2n _{ p })+2n _{ p }} is the index set of the components of \(x \in \mathbb {R}^{N}\) which correspond to binary variables.
Note that μ _{ V } encodes the pressure variability indicators that in MartínezCodina et al. (2015) are linked with pipe failures. The function in (2) is a quadratic form and can be written as x ^{ T } A x, with \(A \in \mathbb {R}^{N \times N}\) symmetric positive semidefinite. Moreover, since the quadratic form only depends on pressure head values, we have that x ^{ T } A x = p ^{ T } A _{ p } p with p=[p ^{1},…,p ^{ k }]^{ T } and \(A_{p} \in \mathbb {R}^{n_{l}n_{n} \times n_{l}n_{n}}\) symmetric positive semidefinite. A preliminary numerical analysis on the considered optimization problem has shown that gradientbased nonlinear programming solvers perform better if at least matrix A _{ p } is positive defined, hence we add a small regularization term δ>0 to the diagonal elements of A _{ p } so that it becomes diagonal dominant and then positive defined. Let \(\tilde {A_{p}}:=A_{p}+\delta I_{n_{l}n_{n}}\) and \(\tilde {A}\) the matrix such that \(x^{T}\tilde {A}x=p^{T}\tilde {A_{p}}p\). If the perturbation term δ is sufficiently small it will not significantly affect the optimization process. This is a standard regularization scheme, see Nocedal and Wright (2006). Finally, we define the second objective function as \(\mu _{2}(x):=x^{T}\tilde {A}x\).
For each \(x \in \mathbb {R}^{N}\), we have set g(x):=(g _{ i }(x))_{ i∈I } and h(x):=(h _{ i }(x))_{ i∈E }, the vectors corresponding to the rows of inequality constraints (3b), (3c), (4a), (4c), (4d) and equality constraints (3a), (4b), respectively. Therefore, \(g:\mathbb {R}^{N} \rightarrow \mathbb {R}^{I}\) is a polynomial nonconvex function, while \(h: \mathbb {R}^{N} \rightarrow \mathbb {R}^{E}\) is affine.
3 Mathematical methods for the solution of multiobjective optimization problems
Therefore, it is necessary to search for optimal compromises between these two objectives. Let’s define the feasible set of Problem (5a) as \(\mathcal {T}:=\{x \in \mathbb {R}^{N} \,  \, g(x)\geq 0, \, h(x)=0, \, x_{j} \in \{0,1\} \, \forall j \in B\}.\) Then, we present some technical definitions that characterise these tradeoffs as follows:
Definition 1
(Miettinen 1998, Section 2.2) A point \(x \in \mathcal {T}\) is said dominated by \(y \in \mathcal {T}\) if μ _{ j }(y)≤μ _{ j }(x) ∀j∈{1,2} and μ(y)≠μ(x).
Definition 2
(Miettinen 1998, Section 2.2) A point \(x^{*} \in \mathcal {T}\) is called global Pareto optimum for Problem (5a) if it is not dominated by any point in \(\mathcal {T}\).

\(x_{i}^{*}\) is minimizer of objective μ _{ i }. The point \(\mu _{i}(x_{i}^{*})\) is called the ith Anchor Point

\(\mu ^{u}:=(\mu _{1}(x_{1}^{*}),\mu _{2}(x_{2}^{*}))^{T}\) is the utopia point, a point in the objective function space that is in general infeasible,

utopia line is the line joining the points \(\mu (x_{1}^{*})\) and \(\mu (x_{2}^{*})\),
 the payoff matrix$${\Phi}:= \left[\begin{array}{lll} 0 & \mu_{1}(x_{2}^{*})\mu_{1}(x_{1}^{*}) \\ \mu_{2}(x_{1}^{*})\mu_{2}(x_{2}^{*}) & 0 \end{array}\right], $$

\(\textbf {e}:=(1,1)^{T} \in \mathbb {R}^{2}\).
3.1 Weighted Sum Method
It is easy to prove that a minimum of Problem (6a) is a Pareto optimal point for Problem (5a), see Miettinen (1998).
The WS method solves MINLP_{WS}(ω) for a given distribution of weights in order to generate various Pareto optimal points and approximate the Pareto front. However, as pointed out before, this method has some drawbacks: it is not possible to generate points on the nonconvex part of the Pareto Front and in most cases the distribution of points is not uniform.
3.2 Normal boundary intersection method
Note that \(\mu ^{u} + {\Phi } \left (\begin {array}{ll} 1\omega \\ \omega \end {array}\right ) \) is a point belonging to the utopia line. Moreover, −Φe is the quasinormal direction to the utopia line, pointing towards μ ^{ u }.
3.3 Normalized normal constraint method
The additional equality constraints of NBI are particularly stringent since they restrict the search in \(\mu (\mathcal {T})\) to a line. As pointed out in Das (2000), this is expected to generate complications specifically in the case of a multiobjective mixed integer nonlinear optimization problem. In fact, the discrete decision variables result in a Pareto front with disconnected branches and isolated points, where the intersection between the quasinormal direction and the Pareto front can be empty. On the other hand, the weighted sum method does not suffer from the disconnected nature of the Pareto front. Das and Dennis (1997) presented the following geometric interpretation of the WS method: a weight ω∈[0,1] defines a particular line in the space of the objective functions and solving problem MINLP_{WS}(ω) corresponds to translating that line along its orthogonal direction and choosing the first intersection point with \(\mu (\mathcal {T})\). As a result, the WS method does not experience the difficulties associated with disconnected Pareto fronts. Likewise, the NNC method is not expected to preclude the convergence to Pareto points which are isolated or belong to disconnected branches because its additional inequality constraint enforces a minimization over halfspaces, which is less strict than the equality constraints used in NBI.
However, when the image of the feasible set is folded, both the NBI and NNC methods will generate different local Pareto optima (or even non Pareto points) when different initial conditions are used for the corresponding subproblems (Messac et al. 2003). In Fig. 4b, we illustrate this situation; the segments of curve \([\mu (x_{2}^{*}),A]\) and \([E,\mu (x_{1}^{*})]\) constitute the global Pareto front, while the branches [B,C] and [D,E] are only locally Pareto optimal. Note that although is some examples the Pareto front does not have such complex structures (Logist et al. 2010), many engineering problems have a highly nonlinear, nonconvex structures that result in a disconnected Pareto front. Furthermore, even when the global Pareto front is connected, it could be made up from different locally optimal branches that are cut and glued together, corresponding to separate regions in the space of the decision variables (Hartikainen and Lovison 2014) (see also Fig. 6d and the discussions on it). Despite some differences, NBI and NNC are intrinsically related. The work by Logist and Van Impe (2012) shows that, given a weight ω, any stationary point of a NBIsubproblem is also a stationary point for the NNCsubproblem, provided that the additional inequality constraint of NNC is active. However, as pointed out also by the same authors, the front generated by NBI and NNC could be different in practice, due to the nonconvexity of the scalarization optimization subproblems.
4 Solution of singleobjective MINLPs: a continuous relaxation
Theorem 1
Assume that \((x^{k})_{k \in \mathbb {N}}\) is sequence of stationary points for REL(t ^{ k } ). Let \(x^{k} \xrightarrow {k \rightarrow +\infty } \bar {x}\) as \(t^{k} \xrightarrow {k \rightarrow +\infty }0\) . Finally, assume that \(\bar {x}\) satisfies a suitable constraints qualification for MPCCs. Then \(\bar {x}\) is a stationary point for Problem ( 10 ).
Proof
The proof follows as in (Herty and Steffensen 2012, Theorem 2.3) after the introduction of a slack variable w:=1−x. □
As shown in Ralph and Wright (2004), if Problem (10) satisfies a linear independence constraints qualification tailored for programs with complementarity constraints, then the relaxed problems (11) satisfy the standard linear independence constraints qualification (LICQ) for nonlinear programs  see also (Nocedal and Wright 2006, Ch. 12) Such a tailored constraints qualification is denoted by MPCCLICQ.
Moreover, as discussed in Scholtes and Stȯhr (2001), MPCCLICQ is a generic condition for mathematical programs with complementarity constraints. This means that even if a particular MPCC violates the constraints qualification, an arbitrary small perturbation of the constraints will result in the satisfaction of MPCCLICQ. It is possible to prove that Problem (10) satisfies MPCCLICQ, once a small perturbation is applied  for the sake of brevity, we do not include the proof here.
Moreover, it can be shown that the gradients of the equality constraints of Problem (11) are linearly independent, while rank deficiencies occur for some of the inequality constraints. Nonetheless, as reported in Section 5, no numerical difficulty was encountered when the interior point solver IPOPT (Waechter and Biegler 2006) was applied. In fact, only the linear dependence of equality constraints has a negative impact on interior point methods, while the gradients of inequality constraints cannot cause any rank deficiency of the matrices involved into these algorithms  see (Nocedal and Wright 2006, pp. 573–574).
This approach offers a scalable method for the solution of design problems for WDNs, since the relaxed problems can take advantage of sparsity in the constraints using tailored NLP solvers. Therefore, the relaxation approach is particularly convenient for application to multiobjective problems, where in order to generate a uniform and detailed Pareto front, it may be necessary to solve a large number of singleobjective MINLPs. In this case, the application of standard branch and bound techniques for MINLPs would require infeasible computational time, particularly when dealing with largescale water networks (see for example the computational experience reported in D’Ambrosio and Lodi (2013)).
5 Case study
At each iteration of WS, NBI or NNC methods, the corresponding singleobjective subproblem was solved using the continuous smooth relaxation illustrated in Algorithm 1, generating a sequence of stationary points for relaxed subproblems; each of these nonlinear programs was solved using the interiorpoint solver for large scale optimization problems IPOPT (Waechter and Biegler 2006). An interiorpoint solver was preferred to an activeset method because of the highly constrained structure of the optimization problem: an activeset approach needs to keep track of changes in the activeset of constraints and bounds, requiring additional computational resources (Nocedal and Wright 2006). In our implementation, IPOPT options mu_strategy and mu_oracle were set to “adaptive” and “LOQO”, respectively  see also Kawajir et al. (2015). Moreover, a warmstart technique was applied in both the relaxation algorithm (inner iteration) and scalarization methods (outer iteration) using the solution of a subproblem (and the corresponding Lagrangian multipliers) as initial condition for the next iteration.
Since the optimization problem is sparse, we have exploited the corresponding sparse structure of the Jacobian of the constraints through IPOPT. We have tested the behaviour of the solver with exact sparse Hessians and approximations computed by a limitedmemory quasiNewton method (LBFGS), available within IPOPT (Kawajir et al. 2015). It was noted that better performances were obtained using LBFGS approximation, both in terms of number of iterations and quality of solutions; therefore, the LBFGS Hessian approximation was used in our implementation. The linear systems within IPOPT were solved using the package MA57 (Duff 2004), which is well suited for solving sparse linear systems. After preliminary simulations, we selected the perturbation value δ=10^{−4} for the regularization factor in \(\tilde {A}\) (see Section 2); in practice, this choice is shown not to significantly perturb the optimization problem. All computations were performed within MATLAB 2015a for Windows, installed on a 2.50GHz Intel^{®} Xeon(R) CPU E5−26400 with 18 Cores.
In Section 3, we argued that the Pareto set for a multiobjective mixed integer nonlinear optimization problem is generally disconnected and may be composed of many isolated points. As Figs. 6 and 7 show, in this particular case study, the efficient tradeoffs between the two objectives are distributed along continuous and possibly disconnected branches but they are not represented by isolated points.
As expected, the proposed methods have generated a number of local Pareto optima, some of which dominated, due to the nonconvexity of the singleobjective subproblems. Therefore, the Pareto filter proposed in Messac et al. (2003) was used to identify the nondominated Pareto configurations obtained in each instance. As anticipated in Section 3, the weighted sum method has not suffered from the disconnected nature of the Pareto set and has generated good approximations of the front. However, as expected, the Pareto points are not uniformly distributed along the front. On the other hand, the application of the NBI method resulted in uniform Pareto fronts for the case of 1 and 4 valves. Nonetheless, a number of nonPareto points were obtained in all instances and the fronts for 2 and 3 valves are not well captured by the NBI method. As stated in Section 3, this behaviour is predictable since NBI is known to generate non Pareto points when disconnected local Pareto branches are present– see also the folded case in Fig. 4b.
Figures 6 and 7 show that the NNC method is fairly robust: it has generated uniform and wellspread Pareto fronts with fewer dominated points, in all shown instances. Note that, in many practical situations, the global Pareto front can be very complex; for example, in Fig. 6f it is composed of the union of separate local branches. In this case, the visual analysis of the front is not sufficient to determine whether the two Pareto curves are disconnected or intersect each other. It is known that nonlinear maps can project a set of disconnected domains into a connected image, or a disconnected one. Both cases are frequently encountered in the solution of nonlinear multiobjective optimization problems. An example of a disconnected Pareto front is discussed in Section 3  see also Fig. 4. On the other hand, some multiobjective problems can give rise to connected Pareto fronts, composed of the nondominated set of the union of multiple local branches that intersect and correspond to separate regions of the domain, see also Hartikainen and Lovison (2014).
This situation is also particularly interesting from the hydraulic application point of view. In fact, for the considered optimization problem, the jump from one local branch to another represents a change in the locations of the control valves. On the other hand, when moving along the continuous local Pareto curves, the locations are fixed and only the valve settings are modified. As can be observed in Fig. 6f, if we fix the valve locations in order to individually minimize PV (i.e. we are on the Pareto front on the right) and change the operational settings trying to decrease AZP, we can not obtain significant reductions in average zone pressure. However, we note that smaller values in AZP, with similar PV, can be achieved on the Pareto curve on the left that corresponds to a different valve location, which is the location that minimizes the individual objective of average zone pressure. Therefore, we can conclude that the solution of the purely operational problem (where the locations of the control valves are fixed) can lead to suboptimal pressure management when considering the two objectives.
In the present work we have only considered tradeoffs between 2 conflicting criteria. Nonetheless, other objectives can be included in a straightforward fashion, the optimization of water quality or resilience for example. The generalization and application of the weighted sum and normal boundary intersection methods to optimization problems with more than 2 objectives is straightforward (Marler and Arora 2004; Das and Dennis 1998). On the other hand, the normalized normal constraint method has to be modified in order to properly handle more than 2 objectives and the enhanced normalized normal constraint method of Sanchis et al. (2008) can be used.
Average computational time and number of IPOPT iterations needed to compute one point of the Pareto front, with each scalarization method considered in the present study
NNC  NBI  WS  

Iter  CPU  Iter  CPU  Iter  CPU  
n _{ v }=1  580  25 s  322  16 s  304  15 s 
n _{ v }=2  1224  55 s  1369  64 s  536  24 s 
n _{ v }=3  684  28 s  890  37 s  519  24 s 
n _{ v }=4  722  28 s  1120  45 s  674  29 s 
Moreover, note that we are considering an offline codesign problem, aiming to compute optimal valves’ locations and pressure settings. Once the optimization process is complete, it is possible to define for each installed valve a feedback rule based on the optimization results, see Ulanicki et al. (2000). Therefore, valves can ultimately be operated through an online feedback control system. Nonetheless, the application of standard techniques for nonconvex mixed integer programs would be impractical (D’Ambrosio and Lodi 2013), especially when considering large scale water networks. On the other hand, the reported computational experience is promising and suggests that the proposed methods can be successfully applied to large scale problems related to optimal valves’ placement and operation in water distribution networks. Finally, all computations were executed in series on a single workstation; however, scalarization methods can be easily parallelized  i.e. different Pareto points can be computed in parallel.
6 Conclusions
We have presented a multiobjective codesign optimization problem for optimal valve placement and operation in water distribution networks, addressing the minimization of average zone pressure and pressure variability. By modelling boundary valves and pressure reducing valves within the same framework, the optimization allows to identify which physical actuator should be used and, at the same time, it provides optimal location and pressure settings. The considered formulation results in a multiobjective mixed integer nonlinear optimization problem. Because of the conflict among the two objectives we have adopted the notion of Pareto optimality to provide a mathematical characterization of the best compromises between conflicting criteria. We have focused on the objectives space and we have investigated the application of scalarization methods for the generation of the image through the objective functions of the set of Pareto optima, the so called Pareto front. We have described weighted sum, normal boundary intersection and normalized normal constraint methods, discussing their strengths and limitations. It was shown that each scalarization method relies on the solution of a series of singleobjective sparse mixed integer nonlinear programs (MINLPs).
Since the integer variables considered are binary, they allow a reformulation as a mathematical program with complementarity constraints (MPCC) whose solution can be obtained as limit point of a sequence of stationary points of sparse nonlinear programs  a relaxation method. The application of scalarization and relaxation methods for the solution of multiobjective mixed integer nonlinear problems is, to the authors knowledge, novel. Furthermore, the iterative solution of sparse NLPs provides a scalable approach for large scale water distribution networks, since sparse techniques can be coupled with stateoftheart NLP solvers to efficiently find (at least local) solutions.
Finally, we have applied the presented methods to solve a biobjective optimization problem for a published benchmark network model, considering a varying level of actuation (i.e. number of valves). Our numerical simulations have shown that the normal boundary intersection method is most affected by the disconnected nature of the Pareto fronts. On the other hand, the weighted sum and normalized normal constraint methods have generated good approximations of the Pareto fronts. In line with expectations, the fronts obtained by the weighted sum method are not as uniform as the ones produced by the normalized normal constraint approach. Moreover, the normalized normal constraint method has been shown to be robust, since it has avoided most of the nonPareto points found by the others approaches. The results from the case study are promising and suggest that the presented scalarization approaches for the Pareto front generation, coupled with the relaxation method, can be applied to other multiobjective codesign optimization problems for large scale water distribution networks, even for problems with more than 2 objectives.
Notes
Acknowledgments
This work was supported by the NECImperial “Big Data Technologies for Smart Water Networks” project.
Supplementary material
References
 Andersson J (2000) A survey of multiobjective optimization in engineering design. Tech rep. Department of Mechanical Engineering, Linkȯping UniversityGoogle Scholar
 Araujo LS, Ramos H, Coelho ST (2006) Pressure control for leakage minimisation in water distribution systems management. Water Resour Manag 20(1):133–149. doi: http://dx.doi.org/10.1007/s1126900646353 CrossRefGoogle Scholar
 D’Ambrosio C, Lodi A (2013) Mixed integer nonlinear programming tools: an updated practical overview. Ann Oper Res 204(1):301–320. doi: 10.1007/s1047901212725 MathSciNetCrossRefzbMATHGoogle Scholar
 Das I (2000) Applicability of existing continuous methods in determining the pareto set for nonlinear mixedinteger multicriteria optimization problems. American Institute of Aeronautics and Astronautics, Tech. rep.CrossRefGoogle Scholar
 Das I, Dennis JE (1997) A closer look at drawbacks of minimizing weighted sums of objectives for Pareto set generation in multicriteria optimization problems. Structural Optimization 14(1):63–69. doi: 10.1007/BF01197559 CrossRefGoogle Scholar
 Das I, Dennis JE (1998) NormalBoundary Intersection: a new method for generating the pareto surface in nonlinear multicriteria optimization problems. SIAM J Optim 8(3):631–657. doi: 10.1137/S1052623496307510 MathSciNetCrossRefzbMATHGoogle Scholar
 Duff IS (2004) MA57—A code for the solution of sparse symmetric definite and indefinite systems. ACM Trans Math Softw 30(2):118–144. doi: 10.1145/992200.992202 MathSciNetCrossRefzbMATHGoogle Scholar
 Farley M, Trow S (2003) Losses in water distribution networks. Monitoring and Control. IWA Publishing, A Practitioners’ Guide to AssessmentGoogle Scholar
 Hartikainen ME, Lovison A (2014) PAINTSIcon: constructing consistent parametric representations of Pareto sets in nonconvex multiobjective optimization. J Glob Optim 62(2):243–261. doi: 10.1007/s1089801402329 CrossRefzbMATHGoogle Scholar
 Herty M, Steffensen S (2012) MPCC Solution Approaches For a Class of MINLPs with Applications in Chemical Engineering. Aachen Institute for Advanced Study in Computational Engineering Science, Tech. rep.Google Scholar
 Hillermeier C (2001) Generalized homotopy approach to multiobjective optimization. J Optim Theory Appl 110(3):557–583. doi: 10.1023/A:1017536311488 MathSciNetCrossRefzbMATHGoogle Scholar
 Hoskins A, Stoianov I (2015) A device, method and system for monitoring a network of fluidcarrying conduits. US Patent Application 20150308627Google Scholar
 Hu XM, Ralph D (2004) Convergence of a penalty method for mathematical programming with complementarity constraints. J Optim Theory Appl 123(2):365–390. doi: 10.1007/s1095700451540 MathSciNetCrossRefGoogle Scholar
 Jowitt PW, Xu C (1990) Optimal valve control in water distribution networks. J Water Resour Plan Manag 116(4):455–472. doi: 10.1061/(ASCE)07339496(1990)116:4(455) 10.1061/(ASCE)07339496(1990)116:4(455) CrossRefGoogle Scholar
 Kawajir Y, Laird CD, Waechter A (2015) Introduction to IPOPT: A tutorial for downloading, installing, and using IPOPT. Package DocumentationGoogle Scholar
 Kim IY, De Weck OL (2005) Adaptive weightedsum method for biobjective optimization: Pareto front generation. Struct Multidiscip Optim 29(2):149–158. doi: 10.1007/s0015800404651 CrossRefGoogle Scholar
 Lambert A (2001) What do we know about pressure: leakage relationships in distribution systems?. In: WA Conference ’system approach to leakage control and water distribution systems managementGoogle Scholar
 Lambert A, Thornton J (2011) The relationships between pressure and bursts ‘stateoftheart’ update. Water21  Magazine of the International Water Association 21:37–38Google Scholar
 Lee J, Leyffer S (eds) (2012) Mixed integer nonlinear programming, 1st edn. SpringerVerlag, New York. doi: 10.1007/9781461419273
 Leyffer S (2006) Complementarity constraints as nonlinear equations: Theory and numerical experiences. In: Dempe S, Kalashnikov V (eds) Optimization with Multivalued Mappings, Springer US, chap 2, pp 169–208Google Scholar
 Leyffer S, LȯpezCalva G, Nocedal J (2006) Interior methods for mathematical programs with complementarity constraints. SIAM J Optim 17(1):52–77. doi: 10.1137/040621065 MathSciNetCrossRefzbMATHGoogle Scholar
 Logist F, Van Impe J (2012) Novel insights for multiobjective optimisation in engineering using Normal Boundary Intersection and (Enhanced) normalised Normal Constraint. Struct Multidiscip Optim 45(3):417–431. doi: 10.1007/s0015801106988 MathSciNetCrossRefzbMATHGoogle Scholar
 Logist F, Houska B, Diehl M, Van Impe J (2010) Fast Pareto set generation for nonlinear optimal control problems with multiple objectives. Struct Multidiscip Optim 42(4):591–603. doi: 10.1007/s001580100506x 10.1007/s001580100506x CrossRefzbMATHGoogle Scholar
 Maier H, Kapelan Z, Kasprzyk J, Kollat J, Matott L, Cunha M, Dandy G, Gibbs M, Keedwell E, Marchi A, Ostfeld A, Savic D, Solomatine D, Vrugt J, Zecchin A, Minsker B, Barbour E, Kuczera G, Pasha F, Castelletti A, Giuliani M, Reed P (2014) Evolutionary algorithms and other metaheuristics in water resources: current status, research challenges and future directions. Environ Model Softw 62:271–299. doi: 10.1016/j.envsoft.2014.09.013 10.1016/j.envsoft.2014.09.013 CrossRefGoogle Scholar
 Marler RT, Arora JS (2004) Survey of multiobjective optimization methods for engineering. Struct Multidiscip Optim 26(6):369–395. doi: 10.1007/s0015800303686 MathSciNetCrossRefzbMATHGoogle Scholar
 Martin B, Goldsztejn A, Granvilliers L, Jermann C (2014) On continuation methods for nonlinear biobjective optimization: towards a certified intervalbased approach. J Glob Optim 64(1):3–16. doi: 10.1007/s1089801402013 MathSciNetCrossRefzbMATHGoogle Scholar
 MartínezCodina Á, Castillo M, Gonzálezzeas D, Garrote L (2015) Pressure as a predictor of occurrence of pipe breaks in water distribution networks. Urban Water J. doi: 10.1080/1573062X.2015.1024687
 Messac A, IsmailYahaya A, Mattson CA (2003) The norMalized normal constraint method for generating the Pareto frontier. Struct Multidiscip Optim 25(2):86–98. doi: 10.1007/s0015800202761 MathSciNetCrossRefzbMATHGoogle Scholar
 Miettinen K (1998) Nonlinear multiobjective optimization. doi: 10.1007/9781461555636
 Newman JP, Dandy GC, Maier HR (2014) Multiobjective optimization of clusterscale urban water systems investigating alternative water sources and level of decentralization. Water Resour Res 50:7915–7938. doi: 10.1002/2013WR015233 CrossRefGoogle Scholar
 Nocedal J, Wright SJ (2006) Numerical optimization, 2nd edn. SpringerVerlag, New York. doi: 10.1007/BF01068601 zbMATHGoogle Scholar
 Raghunathan AU, Biegler LT (2005) An interior point method for mathematical programs with complementarity constraints (MPCCs). SIAM J Optim 15(3):720–750. doi: 10.1137/S1052623403429081 MathSciNetCrossRefzbMATHGoogle Scholar
 Ralph D, Wright SJ (2004) Some properties of regularization and penalization schemes for MPECs. Optimization Methods and Software 19(5):527–556. doi: 10.1080/10556780410001709439 MathSciNetCrossRefzbMATHGoogle Scholar
 Rezaei H, Ryan B, Stoianov I (2015) Pipe failure analysis and impact of dynamic hydraulic conditions in water supply networks. Procedia Engineering 119:253–262. doi: 10.1016/j.proeng.2015.08.883 CrossRefGoogle Scholar
 Sanchis J, Martínez M, Blasco X, Salcedo JV (2008) A new perspective on multiobjective optimization by enhanced normalized normal constraint method. Struct Multidiscip Optim 36 (5):537–546. doi: 10.1007/s0015800701854 CrossRefGoogle Scholar
 Scheel H, Scholtes S (2000) Mathematical Programs with Complementarity Constraints: stationarity, Optimality, and Sensitivity. Math Oper Res 25(1):1–22. doi: 10.2307/3690420 MathSciNetCrossRefzbMATHGoogle Scholar
 Scholtes S (2001) Convergence properties of a regularization scheme for mathematical programs with complementarity constraints. SIAM J Optim 11(4):918–936. doi: 10.1137/S1052623499361233 MathSciNetCrossRefzbMATHGoogle Scholar
 Scholtes S, Stȯhr M (2001) How stringent is the linear independence assumption for mathematical programs with complementarity constraints? Math Oper Res 26(4):851–863. doi: 10.1287/moor.26.4.851.10007 10.1287/moor.26.4.851.10007 MathSciNetCrossRefzbMATHGoogle Scholar
 Smale S (1976) Global analysis and economics V. J Math Econ 3(1):1–14. doi: 10.1016/03044068(76)900021 MathSciNetCrossRefzbMATHGoogle Scholar
 Ulanicki B, Bounds P, Rance J, Reynolds L (2000) Open and closed loop pressure control for leakage reduction. Urban Water 2(2):105–114. doi: 10.1016/S14620758(00)000480 CrossRefGoogle Scholar
 Waechter A, Biegler LT (2006) On the implementation of a Primal Dual interior point filter line search algorithm for LargeScale nonlinear programming. Math Program 106 (1):25–57. doi: 10.1007/s101070040559y 10.1007/s101070040559y MathSciNetCrossRefGoogle Scholar
 Wan YH (1978) On the algebraic criteria for local Pareto optima. II. Trans Am Math Soc 245:385–397. doi: 10.1090/S00029947197805114171 10.1090/S00029947197805114171 MathSciNetCrossRefzbMATHGoogle Scholar
 Wright R, Stoianov I, Parpas P, Henderson K, King J (2014) Adaptive water distribution networks with dynamically reconfigurable topology. J Hydroinf 16(6):1280–1301. doi: 10.2166/hydro.2014.086 10.2166/hydro.2014.086 CrossRefGoogle Scholar
 Wright R, Abraham E, Parpas P, Stoianov I (2015) Control of water distribution networks with dynamic DMA topology using strictly feasible sequential convex programming. Water Resour Res 51(12):9925–9941. doi: 10.1002/2015WR017466 CrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.