Water Resour Manage (2016) 30:2719–2733

DOI 10.1007/s11269-016-1316-8

Leak Prediction Model for Water Distribution

Networks Created Using a Bayesian Network
Learning Approach

Sou-Sen Leu 1 & Quang-Nha Bui 1

Received: 13 December 2014 / Accepted: 31 March 2016 /

Published online: 11 April 2016
# Springer Science+Business Media Dordrecht 2016

Abstract Water leakage in water distribution systems (WDSs) can bring various negative
economic, environmental, and safety effects. Therefore, predicting water leakage is one of the
most crucial tasks in water resource management; however, it is also one of the most
challenging ones. Previous leakage-related studies have only focused on detecting existing
leaks. This paper presents a novel model using expert structural expectation–maximisation, for
predicting water leakage in WDSs. The model can take into account the uncertainty of
leakage-related factors and balance the contribution of monitoring data and prior information
in a Bayesian learning process to maximise leakage prediction accuracy. Moreover, the
proposed method can indicate the most crucial factors affecting water leakage. The results
of this study could benefit water utilities by aiding them in establishing an effective leakage
control plan to minimise the risk of water leakage. A case study is presented to demonstrate the
robustness and effectiveness of the proposed method.

Keywords Water distribution system . Leak prediction . Water leakage . Bayesian network

1 Introduction

Water leakage in urban water distribution systems (WDSs) is a major challenge for water
utilities worldwide. Depending on factors such as location, water network systems, and
maintenance condition, the amount of water loss attributable to leaks in WDSs may be as
much as 50 % of the distribution input (Araujo et al. 2006; Puust et al. 2010). Water leakage

* Quang-Nha Bui
[email protected]

1
Department of Construction Engineering, National Taiwan University of Science and Technology, 43,
section 4, Keelung Road, Taipei 10672, Taiwan
2720 S.-S. Leu, Q.-N. Bui

wastes water resources and energy and threatens environmental and public health (Besner et al.
2011; Malm et al. 2015).
Although many studies have considered the problem of water leakage in WDSs, most
studies have focused on detecting leaks. Puust et al. (2010) presented a review of leak
detection methods. Recently, some of these methods, such as the inverse transient analysis
method (Haghighi and Ramos 2012), noise log correlation method (Jin et al. 2007), and
computer-aid simulation optimisation method (Wu et al. 2010), have been tested. The common
limitations of these methods are cost, complex equipment system, and a heavy dependence on
experts; in addition, applying these methods on a large scale is difficult, and they can only be
applied to existing leakages.
Few researchers have focused on water leak prediction, which warrants more attention.
Leakage prediction models are beneficial for active pipe leakage control methods such as leak
prevention and detection (Lijuan et al. 2012). Leak prediction methods are used to identify
areas and pipes with a high probability of leakage, thereby allowing water utilities to devise an
appropriate active leakage control plan, such as pipe replacement, to replace pipes with a high
leakage probability. (Delgado-Galván et al. 2010) reported that active leakage control is the
most effective alternative for water supply managers to handle water leakage problems when
economic, social, and environmental concerns are considered.
Only recently have some techniques been developed specifically to address the leak
prediction problem. Lijuan et al. (2012) presented a model that was based on the
radial basis function (RBF) neural network for pipe leakage time prediction in which
leakage influencing factors act as the input vector and pipe leakage time acts as the
objective vector of the RBF neural network. In this model, the hidden layers in the
RBF neural network are regarded as a black box; therefore, the impact of individual
input factors on output events is inestimable. Liang et al. (2012) presented an
approach for estimating the risk of third-party interference to a pipeline system by
using a fault tree and self-organising maps algorithm. For this approach to be
effective, all basic events in the fault tree must be considered mutually independent
events; however, this is rarely the case in real water pipe leakage problems. Analysing
factors affecting leaks and the relationship between these factors can facilitate
predicting water leakage more effectively. Because many factors that lead to water
leaks could be interrelated, there should be no assumption of mutual relationships
among them. Thus, water leak prediction is a challenging task. Another difficulty in
water leakage prediction is that such events are rare and any prediction has large
uncertainty bounds. Therefore, a method is required that both considers the relation-
ships between influencing factors and ensures the uncertainty bounds of the
prediction.
To address the aforementioned problems, this paper proposes a Bayesian framework
for water pipe leakage prediction. The Bayesian network (BN) derived using the
Bayesian network learning (BNL) model represents the water leakage forming system,
which is used to update the probability of each element in the network, and the leak
probability of a target pipe is updated subsequently according to the leak probabilities
of other elements in the network. To overcome the limitation of available data, a
novel BNL algorithm called expert structural expectation–maximisation (ExSEM) is
proposed to incorporate prior expert knowledge into the BNL process. In principle,
the ExSEM algorithm balances the contribution of monitoring data and prior infor-
mation. When data are limited, the model estimation can be made primarily according
Be Leak Prediction Model for Water Distribution Networks Created 2721

to the prior information; when data are sufficient to be exploited by the model, the
model estimation relies mainly on the data.
The remainder of this paper is organised as follows: Section 2 reviews the statistics of water
leaks in Taiwan and identifies leakage factors. Section 3 introduces the basic concepts of the
BN and BNL model, and Section 4 describes the Bayesian framework for water leakage
prediction. Section 5 details a case study to demonstrate how the model is applied to predict
leaks. In Section 6, conclusions are made on the effectiveness of the proposed model.

2 Leaks in Taiwan and Identifying Leakage Factors

Because of the limited land and uneven seasonal distribution of rainfall in Taiwan, concern for
the protection and effective use of water resources is increasing, particularly regarding water
leakage management. According to a 2011 Taiwan Water Corporation (TWC) report, the total
water leakage volume in Taiwan was equivalent to 630 million m3 per year and water leakage
rate was approximately 20.51 %, which is higher than the global average leakage rate of 18 %
(TWC 2011). Figure 1 shows the distribution of the leakage rate throughout all water network
branches in Taiwan; the lowest and highest leakage rates are 11.03 and 35.84 %, respectively.
Figure 1 also illustrates that the leakage rate is unevenly distributed across regions, depending
on water network conditions and other factors.
The TWC report named the three main direct factors leading to water leaks as pipe
dislocation, cracks, and corrosion (TWC 2011). According to the report, pipe dislocation
causes leaks by extending the joint gap between pipes; factors in pipe dislocation may be pipe
length and the number of connected pipes. Pipe cracks can be identified by monitoring for
excessively high or uneven internal or external pressures. Pipe corrosion may be affected by
water quality, soil condition, and pipe material and age. According to statistics, the pipeline

Fig. 1 Leakage rate in the water network branches in Taiwan

2722 S.-S. Leu, Q.-N. Bui

system in Taiwan has an overage rate of approximately 33 % and PVC pipes account for 59 %
of total pipe length. Data provided by Taipei Water Company revealed that some major factors
for water leaks are ground movement, construction activities, load vibration, and pipe material,
age, depth, and corrosion (Kuo 2014).
Some researchers have studied the causes of water leakage. The International Water
Service Association (IWSA) reported that some major impact factors of water leakage
are movement, corrosion, traffic loading, water pressure, excavation, water tempera-
ture, geological conditions, construction quality, and pipe age, fitting, and material
(IWSA 1991). However, the influence of each factor varies widely among countries
and regions, all of which have different leak impact factors. Ho et al. (2010)
summarised the major leakage impact factors in Taiwan as improper pipes, pipe
corrosion, overload and vibration, water hammers, ground collapse, and construction
damage. Some common leakage impact factors mentioned by Puust et al. (2010) are
ground movement, high water pressure, excavation, temperature, ground conditions,
poor workmanship, and pipe age and defects. Table 1 summarises leak impact factors
from related research.
Defining water leak factors is difficult because many factors influence leaks indirectly
through other factors. The following sections of this paper present a systematic method of
determining the relationships and influence of water leakage factors.

3 Bayesian Network

The Bayesian network (also called the belief network and Bayesian belief network), intro-
duced at Stanford University in the 1970s (McCabe et al. 1998), has since been widely applied
in various disciplines such as medicine (Antal et al. 2004; Liu et al. 2006), system reliability
(Ching and Leu 2009; Doguc and Ramirez-Marquez 2009), risk analysis (Bouejla et al. 2014;
Luu et al. 2009; Martín et al. 2009), and more recently environmental assessment (Cha and
Stow 2014). The BN is a powerful modelling tool for complex problems because it can
describe numerous relevant factors simultaneously and express their relationship effectively,
and provides a mechanism to incorporate many kinds of prior information and expert
knowledge into learning to solving problems with many uncertainties (Antal et al. 2004).

Table 1 Summary of leak impact factors by relevant researches

research/source Leak impact factors

TWC (2011) pipe cracked, pipe dislocation, pipe corrosion, pipe length, number of connected pipes
IWSA (1991) ground movement, pipe corrosion, traffic loading, water pressure, excavation,
pipe age, water temperature, pipe material, pipe fitting, geological condition,
construction quality
Ho et al. (2010) improper pipe, pipe corrosion, overload and vibration, water hammer, ground
collapse, construction damage
Puust et al. (2010) ground movement, high pressure, excavation, pipe age, temperature, pipe defects,
ground conditions, poor quality of workmanship
Kuo (2014) ground movement, pipe material, pipe age, construction activities, pipe depth, load
vibration, pipe corrosion
Be Leak Prediction Model for Water Distribution Networks Created 2723

3.1 Bayesian Network Representation

A BN can be represented as a directed acyclic graph (DAG) in which nodes represent factors
of concern and edges represent casual relationships between nodes. In addition to the graph, a
BN has a set of conditional probability tables (CPTs) Θ = {θijk} specifying the probability of
each possible state of the node Xi according to each combination θijk of the states of its parents
Pai. A node without parents is called a root node and a node without children is a leaf node.
Given the parents Pai of the node Xi, node Xi is conditionally independent of all its
nondescendant nodes.
When the structure and parameters are already known, a BN defines uniquely a joined
probability distribution over all nodes in the graph, which can be expressed as

n 



PðX 1 ; X 2 ; …; X n Þ ¼ ∏ P X i
Pai ð1Þ
i¼1

Once the joined probability distribution in Eq. 1 is known, many inference algorithms can
be used for reasoning probabilities of nodes that have not been observed conditionally to the
values of observed nodes. A review of these inference algorithms can be found in the study by
Murphy (2001).

3.2 Basic Structural Expectation–Maximisation Algorithm for Bayesian Network

Learning

A BN model can be constructed on two general bases: information from domain

experts or from historical data. The first method requires meticulous and exhaustive
research by experts and is susceptible to errors caused by subjective assessment. The
second approach generally requires a large amount of data to be effective (Chickering
et al. 2004). Many methods have been proposed to create BNs from data (Bui and
Jun 2012; Eerola et al. 2011; Schulte et al. 2009; Wong and Guo 2008). Among these
BNL algorithms, structural expectation–maximisation (SEM), proposed by Friedman
(1998), is one of the most popular and can manage incomplete data. The SEM
method is an extension of the expectation–maximisation (EM) algorithm proposed
by Dempster et al. (1977) for BN parameter learning. In contrast to EM, SEM is
employed to find both the structure M and parameters Θ of the BN. The parameter
maximisation in EM is replaced by maximisation in the joint space {M, Θ}, which
alternately searches for the optimal structure and the optimal parameters corresponding
to this structure. To compensate for missing values, the exact score of BN is
approximated according to the expected values of the statistics.

4 Bayesian Framework for Water Leakage Prediction

The ExSEM algorithm serves as the basis of the proposed leakage prediction system.
The most challenging task in executing the proposed model is combining expert
information and monitoring data. There are several means of combining expert
knowledge into the BNL process, in which each method requires experts to provide
different types of information (Feelders and van der Gaag 2006; Heckerman et al.
2724 S.-S. Leu, Q.-N. Bui

1995; Liao and Ji 2009). In this study, experts provided feedback on the structure of
the true BN. A predefined network BN0 was given to experts to obtain their opinions
on how it should be modified, and the final modified network was designated as
expert network BNEx. To integrate this expert network into the ExSEM, a score
function called Bexpert Bayesian information criterion^ (ExBIC) was developed for
integrating two information sources (i.e. limited data and an expert network).

4.1 Expert Bayesian Information Criterion Score

The Bayesian information criterion (BIC) is often adopted to score a candidate network in the
BNL problem and can be expressed as follows:
    1  
BIC ¼ logP B* þ logP B* jD − Dim B* logðnÞ ð2Þ
2
where

 X n
X
qi X
ri
 
*

log B
D ¼ N *ijk log θi jk ð3Þ
i¼1 j¼1 k¼1

is the log-likelihood of candidate network B* given the data D; N*ijk is the occurrence number
{Xi = xk, Pai = paj} in D obtained by inferring from the current optimal network in the case of
incomplete data or otherwise by mere counting (Leray and Francois 2005). The term log P(B*)
in Eq. 2 represents the prior probability of candidate network B*; 0.5Dim(B*)log(n) is the
penalty for the complex model in which Dim(B*) is the number of parameters used for BN
representation and calculated as

  X
n
Dim B *
¼ ðri −1Þqi ð4Þ
i¼1

To incorporate the prior information from expert network BNEx into the BNL process, a
new ExBIC score is proposed to replace the original BIC score. The uniform prior probability
log P(B*) in the BIC score is replaced by the prior probability of the candidate network given
the expert network log P(B*|BNEx); therefore, Eq. 2 becomes

   1  
ExBIC ¼ logP B*
BN Ex þ logP B* jD − Dim B* logðnÞ ð5Þ
2
The calculation of log P(B*|BNEx) is performed using the prior probability concept of a
candidate network. The conditional probability of a candidate network given the expert
network is estimated by calculated the similarity of its DAG with BNEx as follows:




1
P B*
BN Ex ¼ ð6Þ
½nðn−1Þdiff ðBN Þ
Ex

where diff(BNEx) is the number of different edges (including added, deleted, and reversed edges)
between B* and BNEx, and n(n-1) is the number of possible directed edges in n nodes BN.
Equation 6 shows that diff(BNEx) = 0, B* is the same as BNEx; therefore, P(B*|BNEx) = 1.
When diff(BNEx) increases, the probability assigned to B* decreases, signifying that the more
Be Leak Prediction Model for Water Distribution Networks Created 2725

or less similar to the BNEx the candidate network is, the higher or lower, respectively, the score
given to the candidate network will be.
The contribution of the prior network to the learning process is measured by including a
prior weight α in the ExBIC score function:

   1  
ExBIC ¼ αlogP B*
BN Ex þ logP B* jD − Dim B* logðnÞ ð7Þ
2
where α ranges from 0 to 1 and represents the prior weight of expert information; α = 0
corresponds to the uniform prior and α =1 corresponds to the full log prior probability given
by the prior network in the ExBIC score.

4.2 Expert Structural Expectation–Maximisation Algorithm for Predicting Water

Pipe Leakage

The proposed ExSEM algorithm is a modification of the original SEM, adapting the new ExBIC
score into the BNL process. Accordingly, the inputs for the ExSEM algorithm include initial network
BN0, expert network BNEx, and leakage monitoring data. To search for the structure and parameters
{M, Θ} of the BN, a greedy hill-climbing search is employed in the ExSEM, which is also applied in
the original SEM. This is a simple and effective heuristic search tool. Although the greedy hill-
climbing search does not guarantee an optimal result, many previous studies have shown that it
obtains satisfactory solutions (Gámez et al. 2011). To increase the efficiency of the hill-climbing
search in ExSEM, the search engine is provided with a Bgood^ initial network BN0 to reduce search
time and avoid local optima. Figure 2 presents the generalised framework of the ExSEM algorithm.

Fig. 2 ExSEM algorithm

flowchart
Initial BN: BN0

Current best BN: M,

monitoring
EM parameter learning
data

Current best BN: M, *

Search engine ExBIC score

yes
Improve score? Expert
network

no

Return best BN: M *, *

2726 S.-S. Leu, Q.-N. Bui

5 Case Study

5.1 Input Data and Experimental Setup

The proposed approach was applied to predict water leakages in the Taipei water distribution
system. Data were collected for two district meter areas (DMAs) for 2 months by field
engineers. Sampling points were distributed evenly among the two DMAs to ensure impar-
tiality. To optimise the learning process of the ExSEM algorithm, continuous data were
discretised into discrete states. Finally, a data set comprising 2633 cases was input into the
model. Table 2 shows all the input factors and their states after discretisation.
In addition to a data set, ExSEM requires an expert network BNEx for the learning process.
To fulfil this requirement, an initial BN structure BN0 was built according to the leakage-
related factors in the dataset and a survey of previous studies. Figure 3a shows the BN0
structure. This BN0 was then presented to three experts, who suggested deleting one factor and
three edges and adding six edges. Figure 3b presents the final, expert-approved network BNEx.
After sufficient data was inputted, the system was implemented using the Bayes Net Toolbox
developed by Murphy (2001). To evaluate the contribution of the expert network to the ExSEM
learning result and the predictive ability of the system, we applied SEM for the learning BN with
the same data source and compared the structure of the two learned BNs and their predictive ability.
The following comparison demonstrates the superiority of ExSEM over SEM.

5.2 Bayesian Learning Results

ExSEM was applied to learn BN with expert prior weight α = 1 and then compare the learned
BN with the result of the SEM algorithm (corresponding to α = 0 in ExSEM). Figure 4

Table 2 All leakage related factors and their states

Node Factors states

X1 Leakage 0 (no), 1 (yes)

X2 Pipe corrosion 0 (no), 1 (yes)
X3 Pipe cracked 0 (no), 1 (yes)
X4 Pipe dislocation 0 (no), 1 (yes)
X5 Pipe depth A(<0.4 m), B(0.4 m-0.7 m), C(0.7 m-1.2 m), D(>1.2 m)
X6 Pipe material PB, PVC, SSP
X7 Pipe diameter A(<50 mm), B(50 ~ 150 mm), (C > 150 mm)
X8 Pipe age A(0 ~ 10), B(10 ~ 20), C(20 ~ 30), D(>30)
X9 Pipe length A(80 m), B(80 ~ 150 m), C(150 ~ 300 m), D(>300 m)
X10 Construction activities A (small), B(medium), C (large)
X11 Number of connected pipes A(0 ~ 10), B(11 ~ 20), C(20 ~ 30), D(>30)
X12 Soil class N value A(0 ~ 2), B(2 ~ 4), C(4 ~ 8), D(8 ~ 15), E(>15)
X13 Pressure surge A(0), B(0 ~ 5), C(>5)
X14 Ground movement 0 (no), 1(yes)
X15 Road width A(0 ~ 6 m), B(6 ~ 8 m), C(8 ~ 12 m), D(>12 m)
X16 Pump station radius A(0 ~ 2 km), B(2 ~ 4 km), C(4 ~ 6 km), D(>6 km)
X17 District meter area A, B
Be Leak Prediction Model for Water Distribution Networks Created 2727

(a) Initial Bayesian network BN 0

(b) Initial Bayesian network BN Ex

Fig. 3 Input BNs for ExSEM model: a Initial network BN0 and b Expert network BNEx

presents the BNs BNExSEM and BNSEM learned from the ExSEM and SEM algorithms,
respectively.
Comparing structural differences between BNExSEM and BNSEM in the results in
Fig. 4 reveals that the expert network recovered some relationships in BNExSEM, such
as X8-X6, X4-X3, X15-X5, and X7-X3, and restored some factors that were missing in
BNSEM, such as X7 and X15. Other relationships were not supported by BNEx but were
recovered by monitoring data, such as X17-X8, X17-X12, and X12-X14. Some relation-
ships appeared in BNEx, such as X13-X11, X5-X7, and X6-X2, but were excluded in
BNExSEM because they were unsupported by the data. Table 3 shows a comparison of
structural difference between BNExSEM and BN0, BNEx, and BNSEM.
Although relationships in BNs do not constantly represent a real relationship in practice,
most of them are easily explained. The same is true of relationships in BNExSEM because they
can be interpreted when considering monitoring data, a literature survey, or opinions from
experts. For example, the relationship between pipe age (X8) and pipe material (X6) can be
explained using observations from data and comments from experts. Stainless steel pipes
(SSP) can last longer than 30 years, whereas pipes made with materials such as PVC or PB are
normally replaced within 30 years. Another example is the relationship between pipe diameter
2728 S.-S. Leu, Q.-N. Bui

(a) Bayesian network learned from ExSEM algorithm: BN ExSEM

(b) Bayesian network learned from SEM algorithm: BN SEM

Fig. 4 BNs learning result from ExSEM and SEM algorithms: a BNExSEM and b BNSEM

(X7) and pipe cracks (X3); most pipes in our database had diameters of at least 150 mm; larger
pipes such as these are more crucial and will be monitored more closely. However, this
overrepresentation of large pipes created difficulty for SEM to determine the relationship

Table 3 Comparison of structural differences between BNExSEM and other BNs

Compares with BN0 Compares with BNEx Compares with BNSEM

Added edge Removed edge Added edge Removed edge Added edge Removed edge

X4-X3 X17-X8 X13-X11 X15-X5 X11-X3

X14-X3 X17-X12 X5-X7 X8-X6
X12-X14 X12-X14 X6-X12 X4-X3
X11-X6 X6-X2 X7-X3
X8-X6 X6-X4
X17-X8
Be Leak Prediction Model for Water Distribution Networks Created 2729

between pipe diameter and cracks. By using the expert network, ExSEM could recover this
relationship more accurately and produce a more complete network. A possible explanation of
the two aforementioned phenomena is that strong relationships could not be recognised and
recovered by SEM because of insufficient data; however, these relationships were restored by
ExSEM with the assistance of the expert network. This reveals the importance of integrating
expert knowledge into the Bayesian learning process and the manner in which expert
knowledge and monitoring data are combined to optimise the learned network. This optimi-
sation affects the structure of BNExSEM and enhances the prediction accuracy. The next section
describes how BNExSEM exceeded BNSEM in predictive capability.

5.3 Leakage Prediction Results and Comparison

This subsection compares the prediction results of ExSEM and SEM to evaluate the
practical predictive ability of the proposed model. To reduce the bias of the resulting
estimator, we used the k-folds cross validation technique introduced by Diamantidis et
al. (2000). Accordingly, the dataset was partitioned randomly into five equal folds,
each containing 526 data records. In each prediction, four folds were used to train the
model. To compute the threshold level of the leakage, the resulted BN was used to
predict the data in the training folds and then the predicted results were compared
with the real leak condition. The threshold level was determined to give the highest
percentage of correct prediction. Finally, the resulting model was used to predict the
remaining fold.
Figure 5 shows a comparison of the prediction results of ExSEM with those of SEM. The
average prediction accuracy percentage of ExSEM was 0.846, which exceeds the SEM
average result of 0.726. ExSEM also exhibited stability in its predictions; the standard
deviation of its predication results was 0.026. These results are promising, suggesting that
the model has practical applications.

5.4 Analysis of the Effects of Leakage-Related Factors

Determining the main factors in pipe leakage will aid water utilities in water network design
and leakage response planning. An analysis of the effect of each factor on water leakage was
performed according to the method presented by Akhtar and Utne (2014). The effects on

Fig. 5 Comparison of leakage

prediction results ExSEM vs. SEM
2730 S.-S. Leu, Q.-N. Bui

leakage were validated by setting the evidence on the leakage-related factors to the highest
state, one by one, and observing the posterior probability change in the leakage node. This
posterior probability change represents the most severe effect of a single factor on water
leakage. Figure 6 shows the analysis results. For instance, setting the evidence on Bpipe age^
to state D will increase the marginal probability of water leakage from 0.11 to 0.58, and setting
the evidence on Bnumber of connected pipes^ to state D will increase the marginal probability
of water leakage from 0.11 to 0.23. The Bno evidence^ factor in Fig. 6 represents the case for
which no evidence was entered into the BN; its probability of 0.11 represents the normal
leakage probability of the network.
Figure 6 also shows that pipe cracks (X4), pipe dislocation (X3), and pipe corrosion (X2) had
the highest effects on water leakage, a phenomenon that was expected because these factors are
the main factors that cause leakage and the occurrence of each will almost definitely lead to
leakage. These factors can be considered to determine the root cause of leakage because their
states will be changed by altering the states of their descendant factors.
Changing the states of the remaining factors will improve leakage conditions. The mean
posterior probability values of these factors was 0.37, indicating that factors with values higher
than this threshold severely risks of leakage. As demonstrated, construction activity has the
highest effect to leakage, followed by ground movement, pipe age, and pressure surges. This
result suggests that one of the most effective methods of preventing leakage is soundly
managing construction activities. In addition, the results regarding the phenomenon of subsi-
dence and subsidence prevention measures suggest that the regular replacement of old pipes
and the effective design of pumping systems to minimise pressure surges will greatly prevent

Fig. 6 Analysis of effects of leakage-related factors to water leakage

Be Leak Prediction Model for Water Distribution Networks Created 2731

water leakage. By contrast, pipe diameter has the lowest effect to leakage, followed by the
number of connected pipes and road width. That pipe diameter has the lowest effect is
somewhat unexpected. As explained above, the diameter of most pipes in our data was the
large size (150 mm or 200 mm), possibly affecting the analysis results. Collecting more data
on pipes with different diameters will resolve this problem.

6 Conclusions

This study proposed a novel leakage prediction system for WDSs that was based on a new
Bayesian learning approach, the ExSEM algorithm. In the prediction system, a new ExBIC
score was proposed to combine monitoring data and expert knowledge into the BNL process
as a balancing solution to enhance prediction results. The system was applied to the Taipei
WDS, and the results demonstrated that the proposed approach effectively accounted for extra
leakage-related factors and the relationships among these factors. The results also revealed that
the system provided more accurate predictions by integrating an expert network into the
ExSEM algorithm. Moreover, the predictive accuracy of the system could be continuously
improved by updating new data from monitoring systems and experts. This advantage,
combined with the quick-response ability of the BN, renders the system suitable for an online
leakage prediction and prevention system. Furthermore, an analysis of effects of leakage-
related factors revealed critical factors in water pipe leakage, such as construction activities,
pipe age, ground movement, and pressure surges that can be considered to minimise damage
during the water network design process.
The model has some limitations that must be addressed. Currently, the system cannot
handle continuous data; therefore, all data must be discretised before being inputted, poten-
tially affecting the learned network and predictions. Furthermore, the integration of expert
knowledge is not necessarily limited to the BN structure, and the integration of other sources of
information will enable the BN learn more efficiently. Addressing these limitations can
improve our system and its usefulness.

References

Akhtar MJ, Utne IB (2014) Human fatigue’s effect on the risk of maritime groundings – A Bayesian Network
modeling approach. Saf Sci 62:427–440. doi:10.1016/j.ssci.2013.10.002
Antal P, Fannes G, Timmerman D, Moreau Y, De Moor B (2004) Using literature and data to learn Bayesian
networks as clinical models of ovarian tumors. Artif Intell Med 30:257–281
Araujo LS, Ramos H, Coelho ST (2006) Pressure control for leakage minimisation in water distribution systems.
Manag Water Res Manag 20:133–149. doi:10.1007/s11269-006-4635-3
Besner M-C, Prévost M, Regli S (2011) Assessing the public health risk of microbial intrusion events in
distribution systems: conceptual model, available data, and challenges. Water Res 45:961–979. doi:10.
1016/j.watres.2010.10.035
Bouejla A, Chaze X, Guarnieri F, Napoli A (2014) A Bayesian network to manage risks of maritime piracy
against offshore oil fields. Saf Sci 68:222–230. doi:10.1016/j.ssci.2014.04.010
Bui AT, Jun C-H (2012) Learning Bayesian network structure using Markov blanket decomposition. Pattern
Recogn Lett 33:2134–2140. doi:10.1016/j.patrec.2012.06.013
Cha Y, Stow CA (2014) A Bayesian network incorporating observation error to predict phosphorus and
chlorophyll a in Saginaw Bay. Environ Model Softw 57:90–100. doi:10.1016/j.envsoft.2014.02.010
Chickering DM, Heckerman D, Meek C (2004) Large-sample learning of Bayesian networks is NP-Hard. J Mach
Learn Res 5:1287–1330
2732 S.-S. Leu, Q.-N. Bui

Ching J, Leu S-S (2009) Bayesian updating of reliability of civil infrastructure facilities based on condition-state
data and fault-tree model. Reliab Eng Syst Saf 94:1962–1974
Delgado-Galván X, Pérez-García R, Izquierdo J, Mora-Rodríguez J (2010) An analytic hierarchy process for
assessing externalities in water leakage management. Math Comput Model 52:1194–1202. doi:10.1016/j.
mcm.2010.03.014
Dempster AP, Laird NM, Rubin DB (1977) Maximum likelihood from incomplete data via the EM algorithm. J
R Stat Soc Ser B Methodol 39(1):1–38
Diamantidis NA, Karlis D, Giakoumakis EA (2000) Unsupervised stratification of cross-validation for accuracy
estimation. Artif Intell 116:1–16. doi:10.1016/s0004-3702(99)00094-6
Doguc O, Ramirez-Marquez JE (2009) A generic method for estimating system reliability using Bayesian
networks. Reliab Eng Syst Saf 94:542–550
Eerola T, Lensu L, Kamarainen J-K, Leisti T, Ritala R, Nyman G, Kälviäinen H (2011) Bayesian network model
of overall print quality: construction and structural optimisation. Pattern Recogn Lett 32:1558–1566
Feelders A, van der Gaag LC (2006) Learning Bayesian network parameters under order constraints. Int J
Approx Reason 42:37–53
Friedman N (1998) The Bayesian structural EM algorithm. In: UAI'98 Proceedings of the Fourteenth conference
on Uncertainty in artificial intelligence. Morgan Kaufmann, San Francisco, pp 129–138
Gámez J, Mateo J, Puerta J (2011) Learning Bayesian networks by hill climbing: efficient methods based on
progressive restriction of the neighborhood. Data Min Knowl Disc 22:106–148. doi:10.1007/s10618-010-
0178-6
Haghighi A, Ramos HM (2012) Detection of leakage freshwater and friction factor calibration in drinking networks
using central force optimization. Water Resour Manag 26:2347–2363. doi:10.1007/s11269-012-0020-6
Heckerman D, Geiger D, Chickering DM (1995) Learning Bayesian networks: the combination of knowledge
and statistical data. Mach Learn 20:197–243
Ho C-I, Lin M-D, Lo S-L (2010) Use of a GIS-based hybrid artificial neural network to prioritize the order of
pipe replacement in a water distribution network. Environ Monit Assess 166:177–189. doi:10.1007/s10661-
009-0994-6
IWSA (1991) Statistics on impact factors of water leakage. International Water Service Association. IWSA,
London
Jin Y, Yumei W, Ping L (2007) The genetic-algorithm-enhanced blind system identification for water distribution
pipeline leak detection. Meas Sci Technol 18:2178
Kuo TY (2014) Taipei water leak statistics. Taipei Water Company, Taipei
Leray P, Francois O (2005) Bayesian Network Structural Learning and Incomplete Data. Paper presented at the
Proceedings of the International and Interdisciplinary Conference on Adaptive Knowledge Representation
and Reasoning (AKRR 2005), Espoo, Finland
Liang W, Hu J, Zhang L, Guo C, Lin W (2012) Assessing and classifying risk of pipeline third-party interference
based on fault tree and SOM. Eng Appl Artif Intell 25:594–608. doi:10.1016/j.engappai.2011.08.010
Liao W, Ji Q (2009) Learning Bayesian network parameters under incomplete data with domain knowledge.
Pattern Recogn 42:3046–3056
Lijuan W, Hongwei Z, Zhiguang N (2012) Leakage Prediction Model Based on RBF Neural Network. In: Wu Y
(ed) Software Engineering and Knowledge Engineering: Theory and Practice, vol 114. Advances in
Intelligent and Soft Computing. Springer Berlin Heidelberg, pp 451-458. doi:10.1007/978-3-642-03718-
4_56
Liu T-F, Sung W-K, Mittal A (2006) Model gene network by semi-fixed Bayesian network. Expert Syst Appl 30:
42–49
Luu VT, Kim S-Y, Tuan NV, Ogunlana SO (2009) Quantifying schedule risk in construction projects using
Bayesian belief networks. Int J Proj Manag 27:39–50
Malm A, Moberg F, Rosén L, Pettersson TJR (2015) Cost-benefit analysis and uncertainty analysis of water loss
reduction measures: case study of the Gothenburg drinking water distribution system. Water Resour Manag
29:5451–5468. doi:10.1007/s11269-015-1128-2
Martín JE, Rivas T, Matías JM, Taboada J, Argüelles A (2009) A Bayesian network analysis of workplace
accidents caused by falls from a height. Saf Sci 47:206–214
McCabe B, AbouRizk S, Goebel R (1998) Belief networks for construction performance diagnostics. J Comput
Civ Eng 12:93–100. doi:10.1061/(ASCE)0887-3801(1998)12:2(93)
Murphy K (2001) The Bayes net toolbox for MATLAB. Computing Science and Statistics 33(2):1024–1034
Puust R, Kapelan Z, Savic DA, Koppel T (2010) A review of methods for leakage management in pipe networks.
Urban Water J 7:25–45. doi:10.1080/15730621003610878
Schulte O, Frigo G, Greiner R, Wei L, Khosravi H (2009) A new hybrid method for Bayesian network learning
With dependency constraints. In: Computational Intelligence and Data Mining, 2009. CIDM ′09. IEEE
Symposium on, pp 53–60
Be Leak Prediction Model for Water Distribution Networks Created 2733

TWC (2011) Taiwan water corporation report

Wong ML, Guo YY (2008) Learning Bayesian networks from incomplete databases using a novel evolutionary
algorithm. Decis Support Syst 45:368–383
Wu Z, Sage P, Turtle D (2010) Pressure-dependent leak detection model and its application to a district water
system. J Water Resour Plan Manag 136:116–128. doi:10.1061/(ASCE)0733-9496(2010)136:1(116)