Drink. Water Eng. Sci.

, 15, 1–12, 2022

https://doi.org/10.5194/dwes-15-1-2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Technical note: Graph-theory-based heuristics

to aid in the implementation of optimized
drinking water network sectorization
Marius Møller Rokstad1 and Karel van Laarhoven2
1 Departmentof Civil and Environmental Engineering, Norwegian University of Science
and Technology (NTNU), Trondheim, 7491, Norway
2 KWR Water Research Institute, Nieuwegein, 3433, the Netherlands

Correspondence: Marius Møller Rokstad ([email protected])

Received: 30 April 2021 – Discussion started: 14 January 2022

Revised: 5 May 2022 – Accepted: 23 May 2022 – Published: 17 June 2022

Abstract. Drinking water distribution networks form an essential part of modern-day critical infrastructure.
Sectorizing a network into district metered areas is a key technique for pressure management and water loss
reduction. Sectorizing an existing network from scratch is, however, an exceedingly complex design task that
designs in a well-studied general mathematical problem. Numerical optimization techniques such as evolutionary
algorithms can be used to search for near-optimal solutions to such problems, but doing so within a reasonable
time frame remains an ongoing challenge. In this work, we introduce two heuristic tricks that use information of
the network structure and information of the operational requirements of the drinking water distribution network
to modify the basic evolutionary algorithm used to solve the general problem. These techniques not only reduce
the time required to find good solutions but also ensure that these solutions better match the requirements of
drinking water practice. Both techniques were demonstrated by applying them in the sectorization of the actual
distribution network of a large city.

1 Introduction In many countries, sectorization into DMAs has been a key

strategy for leak detection and water loss reduction through
Drinking water distribution systems (DWDSs) form an es- pressure management (UK Water Authorities Association,
sential part of modern-day critical infrastructure. As these 2008; Morrison et al., 2007; Farley et al., 2001). In other
large pipe networks are often hard to reach and interwoven countries, such as Norway and the Netherlands, however, wa-
with other urban infrastructure, modifying them is an ardu- ter utilities have only recently started to consider the advan-
ous and expensive task. Regardless, drinking water utilities tages of sectorization. Not having integrated sectorization in
need to consider applying new designs to their networks in their network design process at the start, these utilities now
order to, for instance, improve service, reduce water losses face the challenge of designing an efficient division of their
and to prepare in a resilient way for an uncertain future. complete infrastructure at once. This will involve balanc-
Here, we focus on one specific aspect of network redesign: ing many different criteria, prime among them the costs of
sectorization. Sectorization entails dividing a DWDS into placing or removing network components and the reduction
separate subnetworks, either by closing the boundary pipes in hydraulic performance that follows from decreased con-
between them or by outfitting the boundaries with pumps, nectivity in the network. Moreover, as DWDSs typically are
pressure reduction valves or flow meters. As a result, the wa- huge meshed systems, there often are an overwhelming num-
ter balance of individual subnetworks or District Metered Ar- ber of different approaches to dividing up the network.
eas (DMAs) can be obtained, and the pressures in the DMAs The key challenge of network sectorization lies in finding
can be managed separately. ways to efficiently divide the network in as many DMAs as

Published by Copernicus Publications on behalf of the Delft University of Technology.

2 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

possible with as few changes (which are costly) to the net- 2 Methods
work as possible. This essentially is a version of the (np-
hard) minimal k-cut problem (Kim et al., 2011). In the past 2.1 Basics of EAs and their application to sectorization
decades, problems such as these have inspired extensive re-
EAs are a type of optimization algorithm inspired by con-
search on the application of numerical optimization tech-
cepts from genetics. The general principle behind this type of
niques to aid in various aspects of DWDS design (Mala-
algorithm is shown schematically by the blue boxes in Fig. 1.
Jetmarova et al., 2018; Maier et al., 2014). The literature con-
First, a collection of possible solutions is created (a popula-
tains a multitude of examples of various methods applied to
tion of individuals). The solutions are tested for their per-
achieve effective sectorization of water distribution networks
formance according to the user’s performance criteria. The
(Alvisi, 2015; Brentan et al., 2018; Ciaponi et al., 2016; Diao
least successful solutions are discarded (natural selection),
et al., 2013, 2016; Laucelli et al., 2017; Liu and Han, 2018;
and the collection is supplemented with new solutions. The
Hajebi et al., 2015; di Nardo et al., 2014; Vasilic et al., 2020;
new solutions are generated by creating small variations in
Zhang et al., 2016, 2017). Such techniques allow drinking
well-performing solutions (mutation) or by combining ele-
water experts to explore their options in a systematic, auto-
ments from two well-performing solutions (reproduction or
mated way and to subsequently substantiate their choice for
crossover). The process is then repeated several times, grad-
specific, optimal solutions. One particularly versatile tech-
ually improving the quality of solutions (evolution).
nique that has received thorough attention in this context is
If individual solutions are judged based on only one per-
that of genetic algorithms (GAs) (Holland, 1975; Goldberg,
formance criterion, selecting the final candidate is a matter of
1989) and other members of the overarching family of evo-
selecting the solution with the highest performance. If multi-
lutionary algorithms (EAs).
ple performance criteria are used, however, it may occur that
While EAs are known to be suitable for sectorization-type
these criteria are at odds, so a choice must be made that ac-
problems, one of their limitations is the unreasonable amount
cepts a trade-off between the two. This trade-off is typically
of function evaluations required to converge to an optimal so-
represented by a Pareto front, a graph that scatters individual
lution properly (Kim et al., 2011). Moreover, when it comes
solutions on two or more axes that correspond to their scores
to practical application of EAs, improving their searching
according to the different criteria.
behaviour in terms of computational speed and quality of
The white boxes in Fig. 1 illustrate a basic way in which an
results remains an ongoing challenge (Maier et al., 2014;
EA can be applied to find solutions to the sectorization prob-
van Thienen et al., 2018). To this end, the various mech-
lem, i.e. to find ways to divide the network into subnetworks
anisms of the classic genetic algorithm are commonly ex-
with as few boundaries between them as possible:
panded, replaced or combined with heuristic tricks or com-
plete heuristic algorithms to improve performance (Maier – Initialization. Individual solutions are defined by
et al., 2014; Krasnogor and Smith, 2005; El-Mihoub et al., assigning every demand node in a DWDS to a
2006; van Laarhoven et al., 2018). The algorithms which in- given DMA.
clude these are commonly referred to as hybrid genetic algo-
rithms (HGAs) or memetic algorithms (MAs). – DMA performance criteria. Typical aspects that are im-
In this paper, we report two HGA techniques that have portant for the performance of a solution are, for in-
aided in the successful application of EAs for sectorizing stance, the sizes of the individual DMAs and the num-
real-life DWDSs of large towns in Norway and the Nether- ber of boundaries between them. To determine these as-
lands. Both techniques first use graph theory algorithms to pects, a specific representation of solutions is needed
extract aspects of the network structure in a formal way. This that includes the graph topology of the network. In other
information on network structure is used to guide the search- words, the solution must contain not only information
ing behaviour of the EA towards structures that are prefer- on the demand nodes of the DWDS but also on the pipes
able in a DWDS according to the criteria of water utility ex- between them. A specific representation may also in-
perts. Making the design criteria of utility experts explicit clude more detailed information on the functional prop-
in this way enhances trust in the technique and thereby en- erties of the DWDS – effectively forming a complete
hances the chances of practical implementation of numerical hydraulic model – so that the hydraulic performance
optimization. Secondly, computational time required to find of a solution might be assessed (for instance in terms
suitable solutions is reduced so that practical application be- of reduced supply capacity once a certain pipe between
comes feasible. DMAs is closed).

– DMA performance constraints. Rather than being used

as performance criteria that drive the direction of opti-
mization, solution properties may also be made subject
to constraints or boundaries. This ensures that networks
keep meeting practical requirements while their config-

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022

M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization 3

Figure 1. Schematic representation of the evolutionary optimization algorithm used. The blue boxes summarize the general operation of the
algorithm. The white boxes illustrate a basic implementation of the algorithm used to optimize a DMA configuration.

uration is changed to optimize the performance crite- ferred. As such, the individual results should be ordered
ria. This could be as simple as putting a minimum or in a Pareto front from which a solution must then be
maximum on the value of a performance criterion for chosen based on criteria outside the optimization pro-
a solution to be considered viable. Far more complex cess.
constraints may be useful or necessary, however. The
topic of Sect. 2, for instance, is ensuring that connec- 2.2 Shortest independent paths identification algorithm
tivity, redundancy and pressure requirements through- for reliability conservation and search space
out the network are met when every DMA boundary is reduction
outfitted with a pressure reduction valve of a particu-
lar setting (a constraint that requires multiple hydraulic One specific reason to sectorize a network may be to divide
simulations to evaluate for a single possible solution). it into pressure management zones. These are subnetworks,
separated by pressure reduction valves, aimed at reducing the
– Mechanisms to create candidate variation. Basic muta- pressure throughout the network. In this case, every bound-
tion can be achieved by splitting DMAs in two or by ary between zones has an implication for the hydraulic per-
merging adjacent DMAs. Crossover can be achieved by formance of the network. The hydraulics of the network,
taking specific DMAs from two solutions to construct however, are subject to many practical constraints and, in-
a new solution (paying attention to smoothing possible deed, regulations. It must for instance be ensured that, despite
gaps between solutions). the introduction of pressure reduction valves, the network is
able to supply sufficient water not only in the nominal situ-
– Results and Pareto front. The two criteria used here are ation, but also during a pipe failure calamity or while a fire
mutually exclusive, i.e. smaller DMAs are beneficial but hydrant is in use.
require more boundaries to realize, which is not pre-

https://doi.org/10.5194/dwes-15-1-2022 Drink. Water Eng. Sci., 15, 1–12, 2022

4 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

Figure 2. Flow chart with suggested algorithm for identifying principal paths.

Checking whether a particular solution will meet these ity and system robustness. This is achieved by representing
constraints will require a multitude of hydraulic simulations the network as a bi-directional weighted graph, identifying
under different scenarios. The computational time required the shortest path between source and demand node using Di-
to verify this could easily be of the order of tens of hours for jkstra’s algorithm, consecutively changing the weights of the
a single solution when a realistically large distribution net- graph, and rerunning the shortest-path algorithm so that the
work is considered. As such, the time required to check the paths which are maximally independent are identified.
performance of individuals is prohibitively large for an EA
to be used (in which easily millions of individuals need to be 2.2.1 Shortest-path algorithm outline
evaluated throughout the search). The need to check whether
a particular solution for sectorization of the drinking wa- The suggested procedure for identifying P independent paths
ter network will be in violation of the performance require- in a network is outlined in the flow chart in Fig. 2, and it
ments set by the utility or legislation has been limiting Trond- will be explained further in the following subsections. The
heim municipality’s capacity to optimize their sectorization method starts with loading the hydraulic model representa-
with respect to pressure management, as the number of hy- tion M of the network (see step 0 in Fig. 2; the hydraulic
draulic simulations and computational time would be im- model is assumed to contain the following:
practically high, thus making it virtually impossible for the
– a set of np links (L) of which P is the subset of links
utility to identify a globally effective solution for pressure
which are pipes (P ⊂ L), and B is the subset of links
management. The objective of the algorithm suggested here
which allow bi-directional flow (B ⊂ L) (as opposed to,
is to provide an alternative way of ensuring the hydraulic re-
for example, check valves). Each link i has defined a
quirements of solutions while optimizing a sectorization with (1) (2)
first and second node (Ni and Ni , respectively), a
pressure control zones in mind.
diameter (Di ) and a measure of its hydraulic resistance
The core of the approach is to evaluate to which extent in-
fi = f (ki ), where ki is the pipe’s absolute roughness);
dividual nodes in a water distribution system are served by
pipes do in addition have a defined length (Li ).
multiple, independent paths, as this is a measure of reliabil-

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022

M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization 5

(j )
– a set of nodes N, of which S is the subset of nodes which directions, thus I  Lp ) are changed to the value ri ,

are considered sources that can provide water into the Li
where ri > min(L) w̃k , thus ensuring that a path go-
system, i.e. reservoirs or water tanks (S ⊂ N). ∀k∈L
(pt)
ing through any one of the elements in Lj will have a
Then each link i in the network is assigned a weight w̃i ac- higher weight than any non-looped path that does not go
cording to its hydraulic conductance (step 1): through any of the elements, and thereby prompting the
– For the set of pipes (P ⊂ L), the weight is calculated shortest-path algorithm to minimize the number of ele-
(p)
to be proportional to the length (Li ) and some measure ments it has in common with the elements in Lj . The
of the hydraulic resistance (fi ), and it is inversely pro- algorithm then returns to identify the next (p + 1) short-
γ
portional to some exponent of the diameter (Di ) of the est path for node j . If the number of paths to be identi-
pipe. In this way, each pipe is assigned a weight accord- fied for node j has reached P , the algorithm moves to
ing to its hydraulic resistance w̃i = fDi Lγ i . the next node (j + 1).
i

– The weight for non-pipe links (valves and pumps; L r Thus, by identifying the shortest path between two nodes,
P) is calculated in the same way, with the assumption changing the weights of the edges in this path to a value
that its length is twice its diameter, as is also assumed that is
larger than the longest possible path in the graph
for open valves in EPANET (Rossman, 2000), i.e. w̃i = (r > w̃i ), and then running the algorithm to find the
2fi ∀i∈Lp
γ −1 . shortest path again with these changed weights, the shortest-
Di
path algorithm will identify a path that is as independent as
Based on these weights, a weighted bi-directional graph, G, possible from the paths that have already been identified.
representation of the network is constructed (step 2) in the
following way:
2.2.2 Performance metrics for independent paths
– each hydraulic node is represented as a graph vertex, i.e.
When the P independent paths for a set of nodes in a network
V ← N;
have been identified, one can utilize these paths to assess as-
– all bi-directional links (i.e. all links which allow flow in pects of topological reliability and resilience in the network
both directions; B ⊂ L) are represented as two edges in either for individual nodes or on a subsystem or system level.
(1) When P paths have been identified for any given node j ,
the graph, with one for each direction (one from Ni
(2) (2) (1) one can assess the degree of independence the paths to this
to Ni and one from Ni to Ni if i ∈ B);
node have from each other, by comparing which elements the
– all unidirectional links (i.e. all links which allow flow in pth path has to its preceding paths. This can be calculated as
only one direction; L r B) are represented as one edge. follows through the p-path independence proportion Ip,j :

After the weighted graph has been constructed, one can start 
( {
L̃
}) i
identifying paths for each node in the network. For each ( ) (j ) (j ) (j )
∀i∈ Lp  L1 L2  ... Lp−1
(j )
L̃
node j which is not considered a source (∀j ∈ N r S), the Ip,j = 1 −  ∀p > 1, (1)
following steps are undertaken. L̃i
(j )
∀i∈Lp
3. A copy of the graph is made, i.e. Ĝ ← G.
4. For each path p = 1, 2 . . . P , the shortest paths from where L̃i is some measure of distance for link i and may,
all sources S to nodes j are identified. The function for instance, pipe length Li or the probability of failure for
fSP (Ĝ, s, j ) represents applies Dijkstra’s algorithm for each link i. As shown in Eq. (1), this index calculates the
(j )
finding the shortest path between node j and s in total length of elements in path Lp that are also present in
(j ) (j ) (j )
the graph Ĝ. The function fSP returns N(s,j ) , L(s,j ) any of the preceding paths {L1  L2  . . .  Lp−1 } (i.e. el-
and W (s,j ) , which is the set of nodes, set of links and ements that are not independent), and it divides it with the
the sum of weights (total distance) in the path between s total length of the path.
and j , respectively. Hence, a value Ip,j = 1 means that the pth path to node j
shares none of the links from any of the preceding paths, and
5. Then, the shortest of the paths between S it is therefore a supply to node j which is completely inde-
and
{ j is chosen as the} p shortest path, i.e. pendent of any of the other paths. Conversely, if Ip,j = 0,
Np , Lp ← N(sp ,j ) L(sp ,j ) .
(j ) (j )
it means that all of the links in the pth path to j are al-
ready present in one (or more) of the preceding paths, and
6. If the number of paths to be identified for node j has that path p does not provide any supply redundancy to j .
not been reached (p < P ), all the weights ŵi of the
links that are in the current shortest path (going in both

https://doi.org/10.5194/dwes-15-1-2022 Drink. Water Eng. Sci., 15, 1–12, 2022

6 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

Figure 5. First shortest path from s to j (solid line shows links in

path, while dashed lines are not in path).
Figure 3. Fictitious example network, with pipe diameter labels.

Figure 6. Graph with updated weights after first shortest path is

identified (thick grey lines indicate edges with updated weights).
Figure 4. Network represented as weighted graph with edge
weights wi = Li /Di2 .

The independence of the pth path can also be expressed in

absolute terms:
( )
L̃ ∑
Ap,j = L̃i , ∀p > 1, (2)
( { })
(j ) (j ) (j ) (j )
∀i∈ Lp L1 L2  ... Lp−1
Figure 7. First and second shortest paths from s to j (illustrated
( )
L̃ with thin and thick solid lines, respectively).
where Ap,j is the length the p shortest path to node j shares
with a preceding path.
– Step 4–5. The shortest path between s and j is identified
2.2.3 Simple network example using the shortest-path algorithm, yielding a result as
shown in Fig. 5.
To illustrate how the suggested method in Fig. 2 works, a
fictitious example network is used (Fig. 3), and the P shortest – Step 6. The weights of the links that are in this path are
paths between a single source s and one node in this network updated and given a high weight, as illustrated in Fig. 6,
is illustrated, identifying only the three shortest paths. This and the algorithm returns to step 4.
example network has pipe diameters according to Fig. 3, and
all pipes are assumed to have the same length (L = 1000 m) – Step 4(2). The shortest-path algorithm is used again,
and friction factor (f = 1). The steps undertaken to identify with the altered weights, identifying the shortest path
the P paths between two nodes (s and j ) are as follows: between s and j , avoiding the links that have already
been identified in previous paths, yielding a result as
– Step 1–2. Based on the network properties (pipe diame- shown in Fig. 7.
ters and lengths), the weight of each edge is calculated,
and the weighted bi-graph is constructed, as illustrated – If further paths to node j are to be identified, the weights
in Fig. 4. (Although a bi-graph is constructed, only one are updated again (see Fig. 8), and the shortest-path
link between each node pair is visible in the example algorithm is used again to find the next shortest path
figures for simplicity of illustration.) (Fig. 9). This process is continued until the P shortest
paths have been identified.
– Step 3. When starting the process of identifying the
shortest paths to a new node j , a copy of the weighted Table 1 shows the results from the analysis of the three short-
(L) (L)
graph is made (Ĝ ← G). est paths from s to j . The indicators Ip,j and Ap,j show that

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022

M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization 7

Table 1. Summary of results for the three shortest-path analyses

from s to j , Fig. 4.

  (L) (L)
Path (p) wi Li Ip,j Ap,j
∀i∈Lp ∀i∈Lp

1 0.0464 4000 – –
2 0.4000 4000 1.00 0
3 0.1859 5000 0.80 1000
Figure 8. Graph with updated weights after first and second short-
est path are identified (thick grey lines indicate edges with updated
weights). A. The algorithm chooses a specific DMA (light grey node
cluster) and takes note of its boundaries (red links) with
the rest of the network (white nodes).

B. Every non-DMA node within a certain distance of the

t DMA is found (dark grey nodes). The links that con-
nect two nodes from this selection are collected as well.
The subgraph that is formed by these nodes and links is
used in the next steps.

Figure 9. First, second and third shortest paths from s to j (sorted

C. The isolated subgraph is expanded with the links that
from thinnest to thickest solid lines, respectively). originally connected the subgraph to the DMA (blue
lines), which are instead connected to a virtual source
node. Moreover, the isolated subgraph is expanded with
the algorithm has been able to identify two paths which are the links that originally connected the subgraph to the
completely independent of each other (paths 1 and 2), since rest of the network (red lines), which are instead con-
(L) nected to a virtual source node.
I2,j = 1.00, and one path that is partially independent from
(L)
the two preceding paths (path 3; I2,j = 0.80 < 1.00). D. The Edmonds–Karp max flow algorithm (Dinic, 1970)
is used to find the maximum number of independent
2.3 Max flow algorithm-based variator that guides paths (green lines) from the virtual source to the virtual
towards nearby minimal costs sink through the isolated subgraph.
The objective of the technique proposed here is to enhance E. The nodes in the isolated subgraph that can be reached
the basic EA approach for sectorization with an additional from the virtual source without using any link in the
variator. When applied to a given solution during the process, independent paths found in step D are found with a
this variator picks one DMA from the solution and searches breadth first search from the source (blue nodes).
for nodes in its vicinity that may be optimal to include in it,
based on the local network topology. On its own, the variator F. In the original DMA configuration, the nodes found in
creates a greedy searching behaviour around a solution, ac- step E are assimilated into the original DMA.
cepting the macroscopic structure of the DMAs and specif-
ically aiming to reduce the number of boundaries between In effect, the algorithm searches for local imperfections in
them. It is assumed that by carefully managing the rate at the interface between the DMA and the rest of the network.
which this hybrid variator is applied to parts of individual so- Firstly, boundaries that lead to the same “choke point” a
lutions, solutions can be seeded with optimal substructures, few nodes away can be reduced by including the nodes up
while the other, basic variators provide enough variation to to the choke point (such as is the case for the assimilated
avoid early convergence to local optima. node cluster on the right in the example). Achieving the same
through random splitting and merging of the DMA and its
2.3.1 Algorithm outline and simple network example neighbouring DMAs would take many generations in the EA.
Secondly, tiny DMAs that fall within the isolated subgraph
The individual steps of the variator are outlined below and entirely are joined to the DMA (such as the assimilated node
illustrated in Fig. 10. The lettered list matches the labels in cluster on the left in the example). This deliberately erad-
the figure: icates small node clusters from the solution, which would
eventually occur through random merging otherwise.
Combined, these effects optimize the number of bound-
aries with minimal changes to the original DMA. The exact

https://doi.org/10.5194/dwes-15-1-2022 Drink. Water Eng. Sci., 15, 1–12, 2022

8 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

Figure 10. Conceptual steps made by the hybrid variator to reduce the number of boundaries of a specific DMA by assimilation certain
nearby nodes. The individual steps are explained in more detail in the text.

extent of the changes allowed can be controlled by choosing Table 2. Algorithm settings during the experiments performed.
the depth of the subgraph around the DMA, which can be
used as a parameter of the optimization algorithm. General algorithm settings
Population 100
2.3.2 Case study Elitism 20 %
Generations 100
The performance of this approach was tested within the con-
text of a case study involving the optimization of a real Experiments
DWDS: the network of the city of The Hague, in the sup- Name Merge/ Crossover Hybrid
ply area of the Dutch water utility Dunea. At the moment split rate variator
of writing, The Hague’s network is strongly meshed and has rate rate
no DMAs implemented, other than one pilot DMA that sep-
EA1 0.03 0.1 –
arates ∼ 2000 customers from the other ∼ 48 000. Dunea
EA2 0.03 0.3 –
seeks to implement a DMA structure in The Hague as a part EA3 0.03 0.9 –
of their effort to better monitor the flow of their water supply. EA4 0.09 0.1 –
The subdivision of this network into DMAs was originally EA5 0.09 0.3 –
optimized with a basic NSGA-II EA for two objectives: mini- EA6 0.09 0.9 –
mizing the total number of DMA boundaries and minimizing EA7 0.27 0.1 –
the maximum DMA size (in terms of daily peak demand). EA8 0.27 0.3 –
Further details about the background and definition of the EA9 0.27 0.9 –
original optimization problem were previously described in HGA1 0.27 0.1 0.03
van Laarhoven and Gardien (2019). HGA2 0.27 0.1 0.09
Here, an optimization over a limited number of genera- HGA3 0.27 0.1 0.27
HGA4 0.27 0.1 0.81
tions (100) was repeated for different combinations of rates
of occurrence assigned to the variators (mutation through
merging/splitting of a single DMA, crossover by combin-
ing DMAs from two solutions and application of the hy- were performed with varying rates of the hybrid variator in-
brid variator described above to a single DMA). In the first cluded. The performed tests are summarized in Table 2.
tests (EA1 to EA9), only mutation and crossover were ap- The results of each test were abstracted in terms of the
plied to roughly scan for the most advantageous basic set- quality of the Pareto front (hypervolume; Cao et al., 2015)
tings. Then, for the most advantageous combination of basic with respect to the network’s combined base demand and the
mutation and crossover, additional tests (HGA1 to HGA4) maximum number of boundaries between DMAs allowed by

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022

M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization 9

Figure 12. Relation between MF and the length of path limiting the
MF value for each node in Trondheim’s DWDS.

ditional completely independent paths to node j remaining,

and it therefore seeks to find the path which shares the min-
imum path length with the previously identified paths. The
Figure 11. Example of the three shortest paths to a node in Trond- third shortest path (green line) shares a subset of its path with
heim’s DWDS. the first shortest identified, but the algorithm seeks to find
the solution where the first and third path has the minimum
the optimization: 4413 m3 h−1 and 500 boundaries, respec- length of elements in common. The results derived from the
tively. This allows us to investigate the influence of the hybrid independent path search can be used as a constraint when
variator on searching behaviour. The computational time of optimizing the allocation of flow-altering valves in a DWDS.
each test was registered to gain insight into the computational The results obtained for the independent paths in a DWDS
cost associated with the use of the hybrid variator. Each test can be compared to the more traditional “maxflow” value
was performed in triplicate to account for reproducibility. measure. When the maxflow between all sources and a node
in DWDS is calculated for an unweighted graph, the result-
ing maxflow value is equal to the number of completely in-
3 Results dependent paths between the sources and the node in ques-
tion. Summarized for Trondheim’s DWDS, Fig. 12 demon-
3.1 Result of shortest-path analyses strates the relation between the MF value between sources
An example result from the shortest independent path anal- and each node in the system, as well as the length of the
yses performed on Trondheim’s DWDS is demonstrated in path limiting each node from achieving a higher max flow
Fig. 11. In this example, the three shortest independent paths value. The system contains roughly 4700 nodes which only
from available sources to one example node, j , have been have one completely independent path from a source (red
identified. Node j is situated centrally in a small pressure area in Fig. 12). However, as the black line demonstrates,
zone in the periphery of Trondheim’s DWDS. The anal- there are less than 2000 of these nodes where the first and
ysis shows that the hydraulically closest source supplying second shortest path share more than 1 km of the path. Con-
node j is the reservoir situated in the west of the pressure versely, one can see from Fig. 12 that there are a few hundred
zone; the path from this source to node j is indicated with a nodes for which the single flow paths (where the first and
thick red line. Although there are multiple paths from the second shortest paths are shared) are longer than 5 km. Thus,
local reservoir to node j which are shorter than the sec- the results in Fig. 12 demonstrate how the proposed indepen-
ond shortest path identified (cyan line), the second shortest dent path measures can be used as a supplement to metrics
path is the only path which is completely independent of the describing one-sided or multi-sided supply in DWDSs.
first shortest path identified, as there are no paths which are
completely independent of each other from the reservoir to
node j . After the first and second shortest independent paths
have been identified, the algorithm finds that there are no ad-

https://doi.org/10.5194/dwes-15-1-2022 Drink. Water Eng. Sci., 15, 1–12, 2022

10 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

Figure 13. Solutions of the Pareto fronts that followed from the Figure 14. The results of the experiments, which are summarized
experiments summarized in Table 2. Blue corresponds to the results as the hypervolumes produced and plotted against the computation
from EA1 to EA9; orange corresponds to the results from HGA1 to time (each experiment was performed in triplicate).
HGA4.

larity indices, meshedness, centrality, algebraic connectivity,

3.2 Result of the max flow variator case study nodal degree, network density and so on – have been used
as metrics to assess the efficiency and resilience of sectoriza-
Figure 13 summarizes the combined results of the experi-
tion solutions, either during the optimization process or dur-
ments. The performances of all solutions in the Pareto fronts
ing a pre-processing step in which assumed optimal “build-
of all experiments are plotted, grouped only by colour in
ing blocks” of DMAs are identified (Khoa Bui et al., 2020).
terms of whether the hybrid variator was used or not. From
These approaches rely on a common assumption: that these
this, it becomes apparent that, indeed, the algorithm is able
theoretical, graph-theory-based heuristics are in some way
to produce a set of solutions of significantly higher quality
congruous with practical DWDS performance and therefore
in the same number of iterations when the hybrid variator is
beneficial as a driver for optimization. The techniques de-
used. At smaller DMA sizes, the number of DMA bound-
scribed in this chapter were designed to exploit graph the-
aries is reduced by 50–100. Especially from a practical point
ory concepts without the need for such an assumption, keep-
of view, this is a substantial increase in quality if one consid-
ing the objectives and constraints explicitly defined in terms
ers that the realization of only a single DMA boundary may
of the strategic goals of the water utility expert: a mini-
already cost a water utility thousands of euros.
mum number of required interventions (boundaries) while
The hypervolumes of the individual Pareto fronts are plot-
hydraulic function is ensured.
ted against the computation times that were required to pro-
The results presented in this paper represent computation-
duce them in Fig. 14. Experiment EA7 produced the lowest
ally quick ways of solving sectorization problems, while at
average Pareto volume while also having one of the lowest
the same time considering specific practical constraints. The
computational times of the basic experiments. Hence, the set-
shortest-path algorithm presented in Sect. 2.2 can be used as
tings for the merge, split and crossover rates of EA7 were
a pre-processing step that ultimately excludes pipes as viable
used in the HGAx experiments as well.
locations for pressure control zone boundaries, with practical
The data demonstrate that using the hybrid variator with
requirements and regulations in mind. This provides an ap-
a rate of occurrence of only a few percent already leads to
proach to use EAs for optimizing the design of pressure re-
a substantial decrease in hypervolume (increase in quality).
duction zones – while guaranteeing acceptable performance
The quality can be increased with higher rates of occurrence,
under a multitude of possible failure scenarios – in a way that
but this comes at a significant cost of computational time:
is computationally feasible.
HGA4 takes around 5 times longer than HGA1.
The hybrid variator presented in Sect. 2.3 can be used
(sparingly) in addition to other variators to add a local search
4 Conclusions component to the search behaviour that contributes to find-
ing stronger solutions more quickly with EAs (as shown in
For the past decade, graph theory has become an increasingly Fig. 13). The variator can be used more rigorously to find
popular tool for analysing and optimizing the performance even stronger solutions at the cost of substantial computa-
of DWDSs. Among other things, graph theory has been used tional time (as shown in Fig. 14). As a result, the variator
as a computationally efficient means for optimally sectoriz- can be a valuable additional asset when applying EAs in the
ing DWDSs into DMAs or other subnetworks. In most of water utility practice to optimize the design of DMAs.
the cases, graph-theory-derived concepts – such as modu-

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022

M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization 11

The shortest-path algorithm presented in Sect. 2.2 essen- Review statement. This paper was edited by Luuk Rietveld and
tially constitutes a search space reduction with a specific reviewed by two anonymous referees.
DMA functionality in mind, and the hybrid variator pre-
sented in Sect. 2.3 essentially constitutes a greedy optimiza-
tion step towards a specific DMA property. Although the ini-
tial results look promising and are able to provide results that References
are suitable for the water utility practice, both approaches
increase the risk for the optimization to get stuck in local op- Alvisi, S.: A New Procedure for Optimal Design of District
tima (i.e. to arrive at solutions that are “very good” but not Metered Areas Based on the Multilevel Balancing and Re-
“the best achievable”). Future work should focus on further finement Algorithm, Water Resour. Manage., 29, 4397–4409,
elucidating this potential trade-off. https://doi.org/10.1007/s11269-015-1066-z, 2015.
Brentan, B., Campbell, E., Goulart, T., Manzi, D., Meirelles,
G., Herrera, M., Izquierdo, J., and Luvizotto, E.: So-
Code availability. The code used for the The Hague case study cial Network Community Detection and Hybrid Opti-
was created under the general terms and conditions of the BTO joint mization for Dividing Water Supply into District Me-
research programme of the Dutch water utilities, which does not tered Areas, J. Water Resour. Pl. Manage., 144, 4018020,
allow for source code to be made publicly available. Similarly, the https://doi.org/10.1061/(ASCE)WR.1943-5452.0000924, 2018.
code used for the Trondheim case study is restricted from public Cao, Y., Smucker, B. J., and Robinson, T. J.: On using the hypervol-
access due to the terms and conditions applied from the funding ume indicator to compare Pareto fronts: Applications to multi-
source. criteria optimal experimental design, J. Stat. Plan. Infer., 160,
60–74, https://doi.org/10.1016/j.jspi.2014.12.004, 2015.
Ciaponi, C., Murari, E., and Todeschini, S.: Modularity-Based
Data availability. The data used to support the findings presented Procedure for Partitioning Water Distribution Systems into In-
in this paper are not available for sharing, due to the fact that the dependent Districts, Water Resour. Manage., 30, 2021–2036,
data describe details about actual water distribution systems, which https://doi.org/10.1007/s11269-016-1266-1, 2016.
are considered a critical infrastructures. Diao, K., Zhou, Y., and Rauch, W.: Automated Creation
of District Metered Area Boundaries in Water Distribu-
tion Systems, J. Water Resour. Pl. Manage., 139, 184–190,
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000247, 2013.
Author contributions. MMR designed the approach described in
Diao, K., Fu, G., Farmani, R., Guidolin, M., and Butler, D.: Twin-
Sect. 2.2 and executed the case study described in Sect. 3.1. KvL de-
Hierarchy Decomposition for Optimal Design of Water Distri-
signed the approach described in Sect. 2.3 and executed the case
bution Systems, J. Water Resour. Pl. Manage., 142, C4015008,
study described in Sect. 3.2. MMR and KvL prepared the article.
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000597, 2016.
di Nardo, A., Di Natale, M., Santonastaso, G. F., Tzatchkov,
V. G., and Alcocer-Yamanaka, V. H.: Water Network Sec-
Competing interests. The contact author has declared that nei- torization Based on Graph Theory and Energy Perfor-
ther they nor their co-author has any competing interests. mance Indices, J. Water Resour. Pl. Manage., 140, 620–629,
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000364, 2014.
Dinic, E. A.: Algorithm for Solution of a Problem of Maximum
Disclaimer. Publisher’s note: Copernicus Publications remains Flow in a Network with Power Estimation, Sov. Math Dokl., 11,
neutral with regard to jurisdictional claims in published maps and 1277–1280, 1970.
institutional affiliations. El-Mihoub, T. A., Hopgood, A. A., Nolle, L., and Battersby, A.:
Hybrid Genetic Algorithms: A Review, Eng. Lett., 13, 124–137,
2006.
Acknowledgements. The authors thank Dennis Gardien, Ina Ver- Farley, M., Water, S., Supply, W., Council, S. C., and
tommen and Peter van Thienen for their roles in the Dutch case Organization, W. H.: Leakage management and control:
study and Odd Atle Tveit, Birgitte G. Johannessen and Ran- a best practice training manual, World Health Organiza-
nveig Høseggen for their roles in the Norwegian case study. tion, Geneva, Switzerland, https://apps.who.int/iris/bitstream/
handle/10665/66893/WHO_SDE_WSH_01.1_eng.pdf (last ac-
cess: 10 June 2022), 2001.
Financial support. Part of this research was funded by the BTO Goldberg, D. E.: Genetic algorithms in search Optimization,
joint research programme of the Dutch water utilities. The research and MachineLearning, 1st Edn., Addison-Wesley Professional,
was also partly funded through the Universitetskommune collabora- ISBN 978-0201157673, 1989.
tion between the Norwegian University of Science and Technology Hajebi, S., Roshani, E., Cardozo, N., Barrett, S., Clarke, A., and
and Trondheim municipality. Clarke, S.: Water distribution network sectorisation using graph
theory and many-objective optimisation, J. Hydroinform., 18,
77–95, https://doi.org/10.2166/hydro.2015.144, 2015.
Holland, J.: Adaptation in Natural and Artificial Systems, U Michi-
gan Press., Michigan, USA, ISBN 978-0262581110, 1975.

https://doi.org/10.5194/dwes-15-1-2022 Drink. Water Eng. Sci., 15, 1–12, 2022

12 M. M. Rokstad and K. van Laarhoven: Two tricks to aid in WDN sectorization

Khoa Bui, X., S. Marlim, M., and Kang, D.: Water Network Parti- Rossman, L.: Epanet 2.0 User’s Manual, Water Supply and
tioning into District Metered Areas: A State-Of-The-Art Review, Water Resources Division, US Environmental Protection
Water, 12, 1002, https://doi.org/10.3390/w12041002, 2020. Agency, US EPA, Cincinnati, https://nepis.epa.gov/Adobe/PDF/
Kim, J., Hwang, I., Kim, Y.-H., and Moon, B.-R.: Genetic ap- P1007WWU.pdf (last access: 10 June 2022), 2000.
proaches for graph partitioning: a survey, Association for Com- UK Water Authorities Association: Report 26
puting Machinery, New York, NY, USA, ISBN 9781450305570, Leakage Control Policy & Practice, Ofwat,
https://doi.org/10.1145/2001576.2001642, 2011. London, UK, https://www.dwi.gov.uk/research/
Krasnogor, N. and Smith, J.: A tutorial for competent memetic algo- completed-research/drinking-water-distribution-networks/
rithms: model, taxonomy, and design issues, IEEE T. Evol. Com- leakage-control-policy-and-practice/ (last access: 10 Jan-
put., 9, 474–488, https://doi.org/10.1109/TEVC.2005.850260, uary 2017), 2008.
2005. van Laarhoven, K., Vertommen, I., and van Thienen, P.: Technical
Laucelli, D. B., Simone, A., Berardi, L., and Giustolisi, O.: Op- note: Problem-specific variators in a genetic algorithm for the
timal Design of District Metering Areas for the Reduction optimization of drinking water networks, Drink. Water Eng. Sci.,
of Leakages, J. Water Resour. Pl. Manage., 143, 4017017, 11, 101–105, https://doi.org/10.5194/dwes-11-101-2018, 2018.
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000768, 2017. van Laarhoven, K. A. and Gardien, D.: How to detect as many leaks
Liu, J. and Han, R.: Spectral Clustering and Multicriteria Decision as possible with as few flow meters as possible?, 12–15, 2019.
for Design of District Metered Areas, J. Water Resour. Pl. Man- van Thienen, P., Vertommen, I., and van Laarhoven, K.: Practical
age., 144, 4018013, https://doi.org/10.1061/(ASCE)WR.1943- application of optimization techniques to drinking water distri-
5452.0000916, 2018. bution problems, in: HIC 2018, vol. 3, 13th International Confer-
Maier, H. R., Kapelan, Z., Kasprzyk, J., Kollat, J., Matott, L. ence on Hydroinformatics, 1–6 July 2018, Palermo, Italy, 2136–
S., Cunha, M. C., Dandy, G. C., Gibbs, M. S., Keedwell, 2143, https://doi.org/10.29007/b54r, 2018.
E., Marchi, A., Ostfeld, A., Savic, D., Solomatine, D. P., Vasilic, Ž., Stanic, M., Kapelan, Z., Prodanovic, D., and Babic,
Vrugt, J. A., Zecchin, A. C., Minsker, B. S., Barbour, E. B.: Uniformity and Heuristics-Based DeNSE Method for Sec-
J., Kuczera, G., Pasha, F., Castelletti, A., Giuliani, M., and torization of Water Distribution Networks, J. Water Resour. Pl.
Reed, P. M.: Evolutionary algorithms and other metaheuris- Manage., 146, 04019079, https://doi.org/10.1061/(asce)wr.1943-
tics in water resources: Current status, research challenges 5452.0001163, 2020.
and future directions, Environ. Model. Softw., 62, 271–299, Zhang, Q., Wu Zheng, Y., Zhao, M., Qi, J., Huang, Y.,
https://doi.org/10.1016/j.envsoft.2014.09.013, 2014. and Zhao, H.: Leakage Zone Identification in Large-Scale
Mala-Jetmarova, H., Sultanova, N., and Savic, D.: Lost in optimi- Water Distribution Systems Using Multiclass Support Vec-
sation of water distribution systems? A literature review of sys- tor Machines, J. Water Resour. Pl. Manage., 142, 4016042,
tem design, Water, 10, 307, https://doi.org/10.3390/w10030307, https://doi.org/10.1061/(ASCE)WR.1943-5452.0000661, 2016.
2018. Zhang, Q., Wu Zheng, Y., Zhao, M., Qi, J., Huang, Y., and
Morrison, J., Tooms, S., and Rogers, D.: DMA Management Guid- Zhao, H.: Automatic Partitioning of Water Distribution Net-
ance Notes, International Water Association (IWA), Specialist works Using Multiscale Community Detection and Multiobjec-
Group Efficient operation and Management Water Loss Task tive Optimization, J. Water Resour. Pl. Manage., 143, 4017057,
Force, London, UK, https://iwa-network.org/learn_resources/ https://doi.org/10.1061/(ASCE)WR.1943-5452.0000819, 2017.
district-metered-areas-guidance-notes/ (last access: 10 Jan-
uary 2017), 2007.

Drink. Water Eng. Sci., 15, 1–12, 2022 https://doi.org/10.5194/dwes-15-1-2022