Documente Academic
Documente Profesional
Documente Cultură
Payal Gupta, Thaw Tar Thein Zan, Mengmeng Wang, Justin Dauwels, Abhisek Ukil
PII: S1875-5100(18)30323-8
DOI: 10.1016/j.jngse.2018.07.012
Reference: JNGSE 2657
Please cite this article as: Gupta, P., Thein Zan, T.T., Wang, M., Dauwels, J., Ukil, A., Leak Detection
in Low-Pressure Gas Distribution Networks by Probabilistic Methods, Journal of Natural Gas Science &
Engineering (2018), doi: 10.1016/j.jngse.2018.07.012.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Leak Detection in Low-Pressure Gas Distribution Networks by Probabilistic Methods I
Payal Guptaa,∗, Thaw Tar Thein Zana , Mengmeng Wanga , Justin Dauwelsb , Abhisek Ukilc ,
Abstract
PT
The presence of leaks is a prevalent issue for aging gas distribution systems across the globe. These events, if not detected in time,
may bring about environmental and health hazards, besides economic losses. Therefore, the development of efficient detection,
quantification, and localization methods is crucial to all gas companies worldwide. In this paper, we present a leak monitoring
system, called Leak Analytics System (LAS) using a probabilistic approach to determine the location and the rate (severity) of
RI
leakage in low-pressure gas distribution networks. This work aims to develop a robust, cost-effective, and real-time online moni-
toring system for low-pressure gas distribution networks. The leakage events are estimated using pressure and flow data obtained
from steady-state modeling of the gas network. The robustness of the methodology is illustrated by analyzing gas networks in
the presence of measurement errors, which account for unavoidable sensor noise in flow and pressure data. The feasibility of the
SC
proposed method is demonstrated on a small artificial gas network. Moreover, the method is applied to a section of the Singapore
gas distribution network for a single as well as multiple leak scenarios. It is also experimentally shown that the severity of the leak
and the location for a single leak scenario can be determined within an accuracy of 95% and 80% respectively, even in the presence
of strong noise.
U
Keywords: Leak detection, Localization, Severity, Gas distribution network, Pressure, Flow.
AN
Nomenclature distribution networks. Transporting networks operate at very
high pressure (>1.6MPag) which transport gas from coastal
n̂` Number of nodes in leak zone supplied to regional areas. On the other hand, distribution net-
A Demand junction link incidence matrix works operate at medium (50kPag-0.3MPag) to low-pressure
M
B Link loop incidence matrix (0-50kPag) and distribute gas from regional areas to end cus-
D Inner diameter of the pipe (mm) tomers.
g Specific gravity of gas
L Pipe length (m) The failure of any of the pipeline network system leads to
D
m Number of branches (pipes) not only the loss of natural resources but also causes serious
n Number of nodes environmental impact and threat to public safety. Therefore,
n` Number of dummy nodes (zero demand nodes) it is essential to manage pipeline risk, prevent and mitigate
TE
P∗ Pressure at pressure regulator node the consequences of any failures (Muhlbauer, 2004) such as
Pi Pressure at the node i leaks. Several approaches have been applied to identify pos-
q Flow demand vector at the outlet nodes sible pipeline failure initiators and subsequently perform risk
Qj Gas flow inside the pipe j assessment (Veritas, 2010; Hopkins et al., 2009). Risk assess-
ment analysis for pipeline failures are based on qualitative and
EP
Pipeline networks are one of the vital and economically ef- pipeline networks across the globe are between 30 and 100
fective modes of transportation of natural gas. These networks years old. Hence, aging and lack of maintenance have increased
the risk of failures such as leaks in the pipes (especially at the
AC
(EIPO) through the national research foundation and Singapore energy market Due to maturity of transporting networks, several leak detec-
authority. Project LA/Contract No.: NRF2014EWT-EIRP003-002. tion techniques have been developed for its continuous moni-
∗ Corresponding author
toring (Billmann and Isermann, 1987; Zhang, 1997). (Ma et al.,
Email addresses: pgupta@ntu.edu.sg (Payal Gupta),
tttzan@ntu.edu.sg (Thaw Tar Thein Zan), mengmeng@ntu.edu.sg
2010) applied the negative pressure wave method that exploits
(Mengmeng Wang), jdauwles@ntu.edu.sg (Justin Dauwels), the pressure waves generated by leaks while (Rocha, 1989) ex-
a.ukil@auckland.ac.nz (Abhisek Ukil) plored the appearance of acoustic pressure waves caused by
Preprint submitted to Journal of Natural Gas Science and Engineering August 7, 2018
ACCEPTED MANUSCRIPT
PT
tion and localization implement as no mathe-
matical model is required
Mass/Volume Mass balance equation Transporting networks, easy principle Sensitive to arbitrary disturbances and
balance presents systematic devia- medium pressure net- dynamics of the pipeline leading to high
tion when leak occurs works false alarms
Transient based Conservation of mass, momen- Transporting networks, Can detect small leaks Sensor cost is high and model is com-
RI
methods tum and energy are utilized for medium pressure net- plex and require trained users
leak detection works
Distributed Presence of leaks causes a tem- Transporting networks, Fast response and high ac- Cost of implementation very high, diffi-
temperature perature change which is de- distribution networks curacy for leak location cult of retrofitting to existing pipes
SC
sensing tected by fibre optic cables
Probabilistic Bayesian probabilistic ap- distribution networks High accuracy and small High computation cost, not able to iden-
based method proach is used for leak sample data is required. tify multiple leaks
detection and localization Cost-effective
U
leaks. Pressure point analysis (PPA) (Bin Akib et al., 2011), et al., 2008). However, such methods require intensive human
another popular and cost-effective approach, detects leaks by AN involvement and lack real-time monitoring. Therefore, recent
comparing the current pipeline pressure values to earlier val- works on the low-pressure networks focus on real-time moni-
ues. (Wan et al., 2011) performed the detection as well as the toring utilizing specialized sensors to assist the detection and
localization of leaks via wavelet transform technology for sig- the localization of leaks. Typical sensors may include dis-
nal pre-processing and average-weighted localization scheme tributed temperature sensors (Tanimola and Hill, 2009; Cam-
for determining the leak location. However, the focus of the im- panella et al., 2016; Ukil et al., 2012), acoustic sensors (Karku-
M
plementation is limited to linear pipelines, without considering lali et al., 2016), and pressure/flow sensors (Reddy et al., 2016).
more complex pipeline networks. Other common approaches However, direct utilization of the sensor data for leak detec-
include frequency response based methods (Mpesha et al., tion and localization may result in false alarm problems and
2001; Askari et al., 2016), mass/volume balance (Dos Santos faulty diagnostics. This is because the sensors are affected
D
et al., 2011), steady-state or transient models (Hauge et al., by background noise and other environmental interference.
2007). In (Ebrahimi-Moghadam et al., 2016), a two-dimensional, tur-
bulent and compressible gas flow model is developed to inves-
TE
with small sample size (as limited leak data is available from
probability of distribution gas pipelines compared to transport-
network operators). Additionally, uncertainties such as sen-
ing pipelines, and makes distribution systems more vulnerable
sor noise, demand variations, etc., are easily handled by proba-
to accidents. Hence, there is a need to develop robust and cost-
bilistic approaches (Uusitalo, 2007). However, the localization
effective schemes for detecting and localizing gas leaks in low-
method is also limited to a single leak, and the author assumes
pressure gas distribution networks.
that the sensors are available at all nodes. The presence of a
Currently, most distribution networks operators follow the large number of exit points at domestic consumers makes it too
conventional methods for leak inspection. Conventional costly to install and deploy sensors at all exit points. Table 1
leak detection methods involve personnel patrolling along the summarizes all the above-mentioned approaches.
pipeline looking for specific sounds, visual effects or odors that
may be released during a leak (Murvay and Silea, 2012; Liu
2
ACCEPTED MANUSCRIPT
1.1. Our Contribution The demand junction link incidence matrix Ai is defined as:
In the view of the issues above and the lack of robust monitor-
+1, if gas in branch j enters node i
ing techniques for localizing leaks particularly in low-pressure
ai j =
−1, if gas in branch j leaves node i
distribution networks, we propose a leak monitoring system,
0, if branch j is not connected to node i .
referred to as Leak Analytics System (LAS). The methodology
involves steady-state modeling of the gas distribution network, The pressure values need to fulfil the energy conservation
combined with a probabilistic approach for leak location and law. The latter can be expressed as the following matrix equal-
severity estimation. Our work can be seen as an extension of ity:
the study of (Gupta et al., 2016) from single to multiple leaks. B∆P = 0, (2)
Our proposed method is able to handle multiple leaks. It does where the link loop incidence matrix B is defined as:
so by a divide-and-conquer approach: the distribution network
PT
is divided in leak zones, which are analyzed separately. In this
+1, if branch j is in loop i and the direction is same
bi j =
−1, if branch j is in loop i and the direction is opposite
way, the computational complexity can drastically be reduced.
To divide the network into leak zones, we introduce leak iso-
0, if branch j is not in loop i ,
lation step for cluster (leak zone) formation, comprising nodes
RI
and ∆P is the unknown node pressure vector in the network:
with high probability of having leaks. In our work, we assume
h iT
that only a few sensors are available, which is more realistic ∆P = ∆P1 ∆P2 ... ∆Pm . (3)
and practical. Also, to illustrate the reliability of our method-
ology, we test the approach in a real gas distribution network. For low-pressure (0-75kPag) gas networks, high-pressure flow
SC
We experimentally show that for a single-leak scenario, the leak equations (Coelho and Pinho, 2007) have been mathematically
locations and the severity can be estimated with an accuracy of reworked in order to ease up calculations with following as-
80% and 95% respectively, while for a multi-leak scenario the sumptions:
localization accuracy remains above 75% in the presence of low 1. steady and isothermal flow of gas,
U
levels of sensor noise. 2. compressibility factor of gas to be one,
The rest of the paper is organized as follows. Section 2 3. neglecting potential energy term,
presents the mathematical description of gas distribution net- 4. constant friction coefficient over the length of pipe which
AN
works at steady state. Section 3 and Section 4 illustrate the leak is given by:
localization and leak isolation stages respectively in the mathe- !
12
matical model for LAS. We present the experimental results and f = 0.0044 1 + . (4)
discussion based on a synthetic network and Singapore gas dis- 0.276D
M
tribution network in Section 5, followed by concluding remarks With all the above stated assumptions, Lacey’s flow equa-
in Section 6. tion (Segeler, 1968; Abeysekera et al., 2014) is derived for low-
pressure (0-75kPag) regime from high-pressure flow equation
and is given by:
2. Modeling Gas Distribution Networks
D
s
(P1 − P2 )D5
The gas distribution networks can be viewed as a combi- Q = 7.1 × 10 −3
, (5)
nation of nodes (pressure regulators, valves, customer outlets, gL
TE
etc.) and link components (gas pipes) with varying lengths and where the specific gravity of the gas g is 0.589. The nodal pres-
diameters, aimed at delivering gas from the supply stations to sure P is in mbar and the flow rate Q is in m3 /h. Eq. (5) can
the customers. Pressure regulators or alternatively known as also be re-written as:
district regulators reduce the pressure of the gas from high-
∆P = P1 − P2 = KQ2 , (6)
EP
In order to analyze the state of gas distribution networks, we station or the pressure regulator node, which converts high-
conduct steady-state simulations on the corresponding graphs, pressure gas to low-pressure. The sole purpose of the pressure
AC
yielding the flow and pressure information. The flow and pres- reduction station is to keep the output pressure near a certain
sure values within the network can be determined from the pressure level P set , which in our case is 5kPag for artificial
principle of mass and energy conservation (Osiadacz, 1988). network while 2kPag for Singapore network. The gas flow-
The flow values need to fulfil mass conservation at the junction ing through the pressure regulator nodes (denoted by P∗ ) can
nodes, which can be expressed in matrix form as follows: be modeled by the following relation:
Ai Q = q, (1) P∗ = P set , P∗ > P set , (8)
where Q is the standard flowrate values for the m pipes and can The above network formulation given by eq. (1), (2), and (6)
be written as: is solved iteratively until convergence with (8) as the constraint
h iT set to obtain the flow values inside pipes and pressure values at
Q = Q1 Q2 ... Qm . the nodes (Krope et al., 2011).
3
ACCEPTED MANUSCRIPT
PT
parameter set θ. In our system, the parameter set θ is defined by
the location θL of the leaks, and gas leakage rate (severity) θF Minimizing f (θ, σ) w.r.t. σ for a given θ leads to:
of the leak. Hence, by definition θ = (θL , θF ). T
1 X
Considering an existing monitoring system installed in the σ̂ = J(θ̂) = min
2
k(yi j (θ) − yi j )k2 . (14)
RI
θ TN
network at N monitoring locations, pressure and flow data are i=1
measured for T test cases. Let y = [yi j ]N×T denote the data Finally, invoking the total probability theorem, the uncer-
matrix measure at N monitoring locations and T test cases tainty in the model parameters (given the measured data y) is
where yi j is the vector comprising of pressure and flow val-
SC
given by (Katafygiotis et al., 1998):
ues at ith monitoring location during jth instance. Similarly,
flow and pressure quantities, denoted by ỹ(θ) = [ỹi j (θ)], can be p(θ|y) = c[J(θ)]−(T N−1)/2 . (15)
obtained from solving the model M(θ) corresponding to a par-
ticular value of parameter θ. As a consequence, the deviation 4. Leak Isolation Problem
U
between the model values ỹi j (θ) and actual data yi j is given by:
In this work, we assume that the potential leak points ac-
ei j (θ) = yi j − ỹi j (θ). (9) count for the pipe junctions, faulty flanges, etc., which con-
AN stitute the nodes of the network with zero demand. We refer
This deviation may be caused by modeling errors or the these nodes as dummy nodes. This is a standard assumption
unavoidable sensor noise while taking measurements. We in model-based leak detection and location literature in large
model these uncertainties by means of a Gaussian distribution scale networks (Ponce et al., 2014; Preis et al., 2010). This as-
parametrized by its mean and standard deviation σ, where we sumption reduces the complexity of the leakage model without
M
assume the mean to be zero. By Bayes theorem, the probability affecting the final results. Further, the leaks are simulated in the
distribution of the parameters θ and σ is obtained as: network by adding an additional consumption/ demand value at
assumed leak nodes which account for leakage rate (severity).
p(y|θ, σ)π(θ, σ) To demonstrate the reliability of the approach for all leak sizes,
p(θ, σ|y) =
D
, (10)
p(y) we generate the leakage rates at random between 0.1%-10% of
the total gas flowing inside the network.
where the term 1/p(y) is a constant for given data. Quantifi-
TE
ceptable values of [θ, σ]. points. This number is significantly increased in the case of
multiple leaks (k > 1) compared to single leak case (k = 1).
The posterior pdf (10) provides the relative plausibility of n
k = 2 .
P̀ n` n`
the models based on the observed data y. The prediction er- Hence, the complexity of this problem becomes
k=1
ror ei j (σ) is assumed to be Gaussian white-noise with standard
C
p(y|θ, σ) in eq. (10) can be written as follows: leak monitoring system, called Leak Analytics System (LAS);
T Y
N
the system can be separated into three stages: leak detection,
(yi j (θ) − yi j )2
" #
Y 1 leak isolation, and leak localization.
p(y|θ, σ) = √ exp − . (11)
i=1 j=1 2πσ 2σ2 The fault detection and isolation stages in LAS are achieved
through the evaluation of residuals between measurements and
From eq. (10) and eq. (11), the probability distribution on leak model outputs (Nejjari et al., 2015). For a given set of M pres-
parameters given sensor data y is obtained by: sure monitoring locations, the residual ri is the difference be-
tween the actual pressure measurement vector p and the esti-
T
"
P
k(yi j (θ) − yi j )k2 # mated pressure values p̂ in leak free case:
1 i=1
p(θ, σ|y) ∝ √ exp − . (12) ri = pi − p̂i . (16)
( 2πσ)T N 2σ2
4
ACCEPTED MANUSCRIPT
PT
Fig. 1: System architecture for the three stage Leak analytics system for leak localization in gas distribution networks.
RI
Algorithm 1 Leak analytics system for leak localization the correlation index. To form this cluster, the node with the
Input: Current sensor measurements from the network (y) and leak-free measurements ŷ; highest correlation value is chosen as the centroid node, and all
Network model.
Output: A near-optimal leakage rates and locations its neighboring nodes within a threshold distance form the leak
prone zone (cluster). The threshold distance depends on the size
SC
1. Residual ← y - ŷ //cf. eq. (16)
2. while Residual > Threshold do of the network or the number of nodes in the network. For our
centroid node ← max (corr(S i , y))//cf. eq. (18) analysis, we consider this distance to be twice the average pipe
1≤i≤NL
Leak zone ← all dummy nodes within threshold distance from centroid node. distance of the network.
for θL = 1, 2, . . . nodes in leak zone do Pressure sensor placement
102
restrict ← SetRestrictions (0 < θF < 10% total flow) Leak location
U
100 P1 101 P2
Golden section/ Nelder-Mead based search. Flow sensor placement
P9
while An optimization criterion is not reached do 120 P3
P8 P6
θF ← getConfig() 123 P10
105 P4
J(θ) ← eval J(θ) //cf. eq. (14)
end while
AN P13 122 P12 121
103
P32
P7 104
P5
106
P11 P31
Find (θkL , θkF ) such that J(θk ) = minθF J(θ) P15 P30 135 136 P51 P36
end for 124 134 P33
P14 108 P35 109
Find (θ̂L , θ̂kF ) such that J(θ̂) = minθL J(θk ) 107 P34
Correct leaks in the network and take measurement data y again. P16 P20
126 P37
3. end while P17 P18 P43
M
125 127 P46 118
P29 115
110 P44 113 P42
P21 P19 P25 P38
P22
The aim of these steps is to detect the presence of leaks and iso- 128
P24 130 P45
P41
129 117
P39
late the leak-free nodes of the network respectively and form a P47
D
on every available pressure sensor measurement by means of method (discussed in Section 3) to obtain the fault locations in
sensitivity analysis. For a network with n` possible faults (leaks the leak zone of the network. The leak localization stage in-
in nodes), we define the sensitivity matrix S as: volves inferring two parameters: leak location θL and leakage
rate θF . By means of the probabilistic framework, we estimate
p̂i j − p̂i the most probable leakage rate θ̂(i)
L of the parameter set θ for
(i)
si j = , i = 1, . . . , M; j = 1, . . . , n` , (17)
C
occurring at node j and p̂i is the pressure estimate at node i in system of pipe network equations at steady-state for all possi-
the leak-free scenario. Leak isolation is performed by analyz- ble leak positions in the network. We solve the above system in
ing the correlation between residual vector (16) and sensitivity MATLAB by golden section search and parabolic interpolation
matrix (17) (Pérez et al., 2011). The same analysis may also based optimization methods (Brent, 2013; Forsythe et al., 1977)
be preformed by means of different metrics. For each set of for single leak scenario. For the case of multiple leaks, we use
possible leaks, the correlation is obtained as: Nelder-Mead based optimization approach. Next, we calculate
the probability of occurrence pi for each possible leakage event
s jT r
Γj = , j = 1, . . . , n` . (18) i with the corresponding estimated leak severity θˆiL :
ks j kkrk
pi = Kc p(θ(i) , θ̂(i)
L |y), (19)
The objective of this analysis is to form a leak prone zone
by clustering the highly probable leak nodes n̂` ∈ n` based on where Kc is a normalization constant. The leak combination
5
ACCEPTED MANUSCRIPT
with the highest probability pi constitutes the most probable respectively. We assume a total of 12 sensors (6 pressure and 6
leak locations, and their corresponding severity values are the flow meters) to be placed for monitoring the network (shown in
rate of gas outflow from the leaks. Algorithm 1 and Fig. 1 il- Fig. 2). We selected the location of the sensors randomly. We
lustrate all the stages of the proposed methodology. We repeat provide the pressure and flow values at the monitoring loca-
these steps until the residual values (R) become below threshold tions in case of no leak in the network in Table Appendix A.1.
and no more leaks are detected in the network. As mentioned, we assume that leak points occur at the pipe
junctions, faulty flanges, etc., which constitute the nodes of the
network with zero demand. We indicate these nodes by aster-
5. Experimental Results and Discussion
isk in Table Appendix A.2. We assume the remaining nodes as
To test the leak localization method, we first apply it to a syn- customer outlet points with gas demand 30m3 /h.
PT
thetic network (see Fig. 2), and next to a section of the gas dis-
tribution network in Singapore (see Fig. 3a). First, we consider 5.1.2. Singapore Gas Distribution Network
a single leakage scenario followed by the occurrence of multi- Next, we consider an actual gas pipeline network, currently
ple leaks. Leakage rates vary between 0.1% - 10% of the total operating in Singapore to distribute town gas from supply sta-
gas supply. We generate synthetic measurements yi j according tions to consumers. The gas distribution network in Singapore
RI
to the model described in Section 3. Specifically, the measure- provides town gas to residential and commercial consumers,
ment vector yi j is a vector set of pressure and flow values at the and constitutes more than 3400 km of pipelines. The en-
sensor locations obtained from the network monitoring system. tire network is mainly underground and operates at three pres-
sure regimes: high-pressure (28–40bar), medium-pressure (3–
SC
5.5bar) and low-pressure (50 / 20 / 2kPag) (Singapore Power,
2016). Here we investigate a 3 km long section of low-pressure
(2kPag) pipeline network that covers the residential area in the
central region of Singapore. For this network, the average dis-
U
tance and the minimum distance between two adjacent nodes
AN are approximately 20m and 1m respectively. For security rea-
1E-10
1E-20
1E-30
Normalized Probability
1E-40
M
1E-50
(a)
(a) 1E-10
1E-20
D
1E-30
1E-40
1E-50
TE
1E-60
77 78 79 80 81 82 83 84 85 86 87 90
Node ID
(b)
Fig. 4: Results for probable leak location w.r.t. node IDs for (a) artificial net-
work and (b) Singapore network with zero noise. X-axis represents the node
IDs of the nodes inside the leak zone.
EP
(b)
1.0
Fig. 3: Gas distribution network in Singapore; (a) extracted network overlaid 0.8
on an aerial view (b) section of Singapore gas distribution network where pipes 0.6
0.4
Normalized Probability
0.2
inlet source (see Fig. 2). (Osiadacz et al., 1995) and (Djebed- 0.4
jian et al., 2008) considered a similar network for testing op- 0.2
77 78 79 80 81 82 83 84 85 86 87
25.4 mm). We provide the network specifications in Table Ap- Fig. 5: Results for probable leak location w.r.t. node IDs for (a) artificial and
(b) Singapore network in the presence of noise. X-axis represents the node IDs
pendix A.2. The average distance and the minimum distance
of the nodes inside the leak zone.
between two adjacent nodes are approximately 100m and 50m
6
ACCEPTED MANUSCRIPT
100
Average Distance
25
accuracy (%)
Leak location
90 100
Leak Location
Accuracy (%)
20
Error (m)
80 90
15
70 80
60 70 10
50 60 5
4
16
12
Leak Rate
8 2
4 1
0
0
0.0 0.5 1.0 1.5 2.0
0.0 0.5 1.0 1.5 2.0
PT
Sensor Noise (%)
Sensor Noise (%)
(a) (b)
Fig. 6: Accuracy of estimated leak location and mean absolute error of estimated leak severity w.r.t. noise value with 95% confidence interval for single leak scenario
in (a) artificial network and (b) Singapore network.
RI
sons, GIS data for the network was not provided. Therefore, in the gas networks. We add zero-mean Gaussian noise to
we extract the network skeleton from the available RGB im- the sensor measurements, independently at different locations.
age of the layout by applying a color based image segmentation For low-pressure distribution networks, the range of the pres-
SC
method based on K-means clustering algorithm. Finally, we im- sure readings is usually between 0-10kPag with an accuracy
pose an XY-coordinate system onto the network (see Fig. 3a). of ±0.1% (ABB, 2016). To illustrate the performance of the
We discretize the network into 150 nodes among which 50 scheme, we select the noise standard deviation (denoted by ε)
nodes are the gas outlets carrying a demand value, and the re- within and above this range, i.e., 0.5%, 1% and 2% in magni-
maining 100 nodes are the dummy nodes (without demand) tude of the pressure and flow data at the sensor locations. While
U
which may account for valves points, pipes junctions, etc. Note generating the observed data y for both networks, the leak loca-
that in Singapore’s gas network, approximately 70% of the AN tions are the same as in Section 5.2.
pipes are made of ductile iron (see Fig. 3b). Therefore, due The leakage events are predicted using the aforementioned
to corrosion, faulty flanges and pipe fitting, and other factors, steps and the normalized probability distribution is obtained
these dummy nodes have the highest risk for leaks. We assume (see Fig. 5) at different values of sensor noise standard devi-
a total of 16 sensors (8 pressure and 8 flow meters) to be placed ation. At the lower level of noise values, the normalized proba-
in the network for monitoring. We arbitrarily select the loca- bility of the most probable leak node is clearly larger compared
M
tions and the number of sensors. to the probability values of other nodes. However, we can see
that when the sensor noise in the measurement data increases
5.2. Detection of single leak without sensor noise from 0.5% to 2%, the probability of the most probable leak lo-
cation is reduced while the peak values of the nearby nodes are
First, we assume that a single leak is present in each net-
D
the leak location at node 122 (see Fig. 2) with a severity of 3.5
curacy. Concretely, the error remains below 12% for all 100
m3 /h, which corresponds to approx. 0.5% of the total gas sup-
test cases. Finally, to test the reliability of the leak localization
ply. For the Singapore network, we consider a single leak of 5
m3 /h at node 80 (see Fig. 3a). Fig. 4(a) and Fig. 4(b) show the
results for the normalized probability obtained for each possi-
EP
ble leak position inside leak zone in the artificial and Singapore
network respectively. We also provide the pressure and flow
values at the monitoring locations for the artificial network in
case of a single leak at node 80 in Table Appendix A.1. With
C
uniformly at random in the range 0.1% - 10% of the total gas monitoring locations. However, due to budget constraints, only
flow in the network. In Fig. 6, we report the accuracy in leak a limited number of sensors can be installed in the network and
detection along with mean absolute percentage error (MAPE) therefore, it is important to optimize the location of these sen-
of the leak amount and its 95% confidence level. In addition, sors,
we report the average distance error of the estimated leak loca- Placing sensors in a strategic manner is important, as we il-
tion from the actual leak location. We compute the accuracy in lustrate in Fig. 9. The first configuration (see Fig. 9(a)) where
leakage location to estimate the number of times the leak loca- the sensors are distributed throughout the network, identifies the
tion is correctly identified out of 100 random simulations, while leak location correctly, while the identification results are worse
we compute MAPE to assess the estimates of the leakage rate. in the second configuration (see Fig. 9(b)) where the sensors are
It can be seen from the figures that sensor noise mostly affects concentrated in one sub-region of the network. Consequently,
the detection of the leak location, and less the estimation of the it is important to determine the optimal location of sensors in
PT
leak strength. Even for strong noise, the leakage rate can still order to achieve the best possible leak localization for a given
be accurately estimated. budget.
Fig. 7 shows the test cases for Singapore network where the
estimated leak node does not coincide with the actual leak posi- 5.5. Detection of multiple leaks with sensor noise
RI
tion. We plot the shortest pipe distance from the actual leak po-
sition to the estimated leak location w.r.t. the normalized prob- After the successful localization of a single leak, we now in-
ability values. Localization errors below 5m, between 5m and vestigate the scenario with more than one leak. The implemen-
15m, and more than 15m are regarded as good, medium, and tation of the overall localization methodology follows the steps
SC
low localization accuracy. We can see that almost all detections in Algorithm 1. For each network, we evaluate two test cases at
are within the good and reasonable performance range. different noise values (see Table 2). Threshold based leak detec-
tion stage confirms the presence of leaks in the networks. For
multi-leak scenario, we apply the leak isolation stage to iden-
100
tify the leaks zones for each network. For the first test case, we
U
identification accuracy (%)
80 show the leak zones for both the networks in Fig. 10. Finally,
we apply the leak localization step on the leak zone nodes to
Leak location
60
AN estimate the leak locations and the leakage rates. The estimated
40
leak locations are corrected, and the above-mentioned steps are
repeated until residuals become below threshold. Table 2 lists
20
the estimated leak locations and the corresponding estimated
leakage rates along with the normalized probability value at the
M
0
0 2 6 10 14 18 22 26
Number of Sensors
estimated leak nodes combination (15). It can be seen from the
Fig. 8: Results for number of correct identification of leak location w.r.t. the
results that large leaks (>1% total flow) are always correctly
number of monitoring sensors in the network. identified and the severity of the leakages are also estimated
correctly.
D
TE
C EP
Fig. 9: Effect of monitoring locations on the localization of leaks. Fig. 10: Zone that contains potential leaks in (a) artificial network and (b) Sin-
gapore network. Blue circles indicate the potential leak points and red markers
indicate the actual leak locations.
AC
Table 2: Results for multiple leakage locations by varying noise values in Singapore network.
Artificial network Singapore Network
Actual Leak Nodes Actual Leak Rate (m3 /h) Actual Leak Nodes Actual Leak Rate (m3 /h)
(i) 122, 130 (ii) 105, 128 (i) 5.0, 4.0 (ii) 7.0, 3.0 (i) 78, 84 (ii) 62, 143 (i) 5.6, 9.0 (ii) 9.5, 3.7
Noise Estimated Leak Location Estimated Leakage Rate (m3 /h) Normalized Estimated Leak Location Estimated Leakage Rate (m3 /h) Normalized
Value Probability Probability
(%) Location 1 Location 2 Severity 1 Severity 2 Location 1 Location 2 Severity 1 Severity 2
122 130 5.0 4.0 1 78 84 5.3 8.8 0.89
ε=0
105 128 6.89 3.15 0.99 62 143 9.45 3.3 0.99
PT
122 130 5.18 4.05 0.8 52 78 3.7 6.9 0.56
ε=2
105 129 6.66 3.15 0.66 62 143 9.4 3.93 0.65
RI
uncertainties and noisy measurements. Here, our probabilis- We proposed a leak localization method LAS for localizing
tic methodology is modified accordingly, which allows for the leakage events in gas distribution networks. We successfully
explicit treatment of the uncertainties arising from the measure- tested this approach on a small artificial network as well as on a
ment noise. We ran all the experiments on a 4-core Intel Core section of the low-pressure (2kPag) Singapore gas distribution
SC
network. We identified the location of leakages with 80% accu-
racy and estimated the leakage rate with 95% accuracy, even in
the presence of noise. Additionally, we applied the proposed
100
LAS to the scenario of multiple leaks. For detecting multi-
U
ple leaks, the computational cost of the naive probabilistic ap-
Accuracy (%)
Leak Location
80
proach is significantly higher. We proposed to reduce this cost
by carrying out leak isolation, before applying the localization
60
AN algorithm. The localization accuracy remains above 75% for
40 multiple leak case scenarios. We demonstrated the effective-
ness of the developed system by considering noisy sensor data.
20
0 0.5 1 1.5 2
Certainly, the performance of our localization methodology
M
Sensor Noise (%) is linked to the measurement locations in the network and the
threshold values of the data for leak detection. For measuring
Fig. 11: Accuracy of estimated leak locations w.r.t. noise value with 95% con- the threshold levels, we assume that there are no leaks present
fidence interval in Singapore network.
in the network while taking measurements from the sensors.
D
gas distribution network with n` possible leaking nodes and k to obtain more reliable leak detection. We will address these
leakage locations is: topics in the future. We generated the pressure and flow data
n` utilized in this study through a gas network model and will re-
Nk = . (20) place by real sensor data from the Singapore gas network in
k!(n` − k)!
future work. Currently, our proposed system assumes a con-
EP
stant gas demand (30 m3 /h) which is a required input into the
The solution scheme to solve the optimization problem for network flow model. We will replace this demand value either
leak localization with k leakages involves an exhaustive search by gas consumption model (Van Westering, 2013) or by actual
over the discrete parameter subspace. Specifically, the most measured gas demand at the customer outlets in future work.
C
P
probable leakage event requires the solution of Nk optimiza-
k
tion problems, considering that each function evaluation of y
Refrences
AC
9
ACCEPTED MANUSCRIPT
Billmann, L., Isermann, R., 1987. Leak detection methods for pipelines. Auto- distribution networks. Control Eng. Practice 19 (10), 1157–1167.
matica 23 (3), 381–385. Ponce, M. V. C., Castañón, L. E. G., Cayuela, V. P., 2014. Model-based
Bin Akib, A., bin Saad, N., Asirvadam, V., 2011. Pressure point analysis for leak detection and location in water distribution networks considering an
early detection system. In: Signal Processing and its Applications (CSPA), extended-horizon analysis of pressure sensitivities. Journal of Hydroinfor-
2011 IEEE 7th International Colloquium on. IEEE, pp. 103–107. matics 16 (3), 649–670.
Brent, R. P., 2013. Algorithms for minimization without derivatives. Courier Preis, A., Whittle, A. J., Ostfeld, A., Perelman, L., 2010. Efficient hydraulic
Corporation. state estimation technique using reduced models of urban water networks.
Campanella, C. E., Ai, G., Ukil, A., Mar 2016. Distributed fiber optics tech- Journal of Water Resources Planning and Management 137 (4), 343–351.
niques for gas network monitoring. In: IEEE Int. Conf. on Ind. Technology Reddy, R. S., Payal, G., Karkulali, P., Himanshu, M., Ukil, A., Dauwels, J.,
(ICIT). Taipei, Taiwan, pp. 646–651. March 2016. Pressure and flow variation in gas distribution pipeline for leak
Coelho, P. M., Pinho, C., 2007. Considerations about equations for steady state detection. In: IEEE Int. Conf. on Ind. Technology (ICIT). Taipei, Taiwan,
flow in natural gas pipelines. J. of the Brazilian Soc. of Mech. Sciences and pp. 679–683.
Eng. 29 (3), 262–273. Rocha, M., 1989. Acoustic monitoring of pipeline leaks. ISA, Paper (89-0333).
PT
Djebedjian, B., Shahin, I., El-Naggar, M., 2008. Gas distribution network op- Rui, Z., Han, G., Zhang, H., Wang, S., Pu, H., Ling, K., 2017. A new model
timization by genetic algorithm. In: Ninth Int. Congr. of Fluid Dynamics & to evaluate two leak points in a gas pipeline. Journal of Natural Gas Science
Propulsion, Alexandria. and Engineering 46, 491–497.
Dos Santos, P. L., Azevedo-Perdicoulis, T.-P., Ramos, J. A., de Carvalho, J. M., Segeler, G. C., 1968. Gas Engineers Handbook (E-Book), Segeler. Ind. Press.
Jank, G., Milhinhos, J., 2011. An lpv modeling and identification approach Singapore Power, 2016. Singapore power. http://www.singaporepower.
to leakage detection in high pressure natural gas transportation networks. com.sg, [Online; accessed 18-September-2016].
RI
IEEE Trans. on Control Syst. Technology 19 (1), 77–92. Tanimola, F., Hill, D., 2009. Distributed fibre optic sensors for pipeline protec-
Ebrahimi-Moghadam, A., Farzaneh-Gord, M., Deymi-Dashtebayaz, M., 2016. tion. Journal of Natural Gas Science and Engineering 1 (4-5), 134–143.
Correlations for estimating natural gas leakage from above-ground and Ukil, A., Braendle, H., Krippner, P., 2012. Distributed temperature sensing:
buried urban distribution pipelines. Journal of Natural Gas Science and En- review of technology and applicat. Sensors J., IEEE 12 (5), 885–892.
SC
gineering 34, 185–196. Uusitalo, L., 2007. Advantages and challenges of bayesian networks in envi-
Forsythe, G. E., Moler, C. B., Malcolm, M. A., 1977. Computer methods for ronmental modelling. Ecological modelling 203 (3), 312–318.
math. computations. Van Westering, W., 2013. Gas distribution network modelling and optimization.
Gupta, P., Goyal, A., Dauwels, J., Ukil, A., 2016. Bayesian detection of leaks Veritas, D. N., 2010. Risk assessment of pipeline protection. RECOM-
in gas distribution networks. In: Ind. Electronics Soc., IECON 2016-42nd MENDED PRACTICE DNVRPF107.
Annual Conf. of the IEEE. IEEE, pp. 855–860. Wan, J., Yu, Y., Wu, Y., Feng, R., Yu, N., 2011. Hierarchical leak detection
U
Han, Z., Weng, W., 2011. Comparison study on qualitative and quantitative and localization method in natural gas pipeline monitoring sensor networks.
risk assessment methods for urban natural gas pipeline network. Journal of Sensors 12 (1), 189–214.
hazardous materials 189 (1-2), 509–518. Zhang, J., 1997. Designing a cost-effective and reliable pipeline leak-detection
Hauge, E., Aamo, O. M., Godhavn, J.-M., 2007. Model based pipeline moni-
AN
toring with leak detection. In: Proc. 7th IFAC Symp., On Nonlinear Control
system. Pipes and Pipelines Int. 42 (1), 20–26.
Soc., IECON 2016-42nd Annual Conf. of the IEEE. IEEE, pp. 412–416.
115 42.11 41.93 P29 35.26 35.95
Katafygiotis, L. S., Papadimitriou, C., Lam, H.-F., 1998. A probabilistic ap-
129 35.61 33.17 P35 70.01 70.70
proach to structural model updating. Soil Dynamics and Earthquake Eng.
134 36.65 35.26 P45 30.0 30
17 (7), 495–507.
TE
Krope, J., Trop, P., Goričanec, D., 2011. Flow-pressure anal. of loop gas net-
works. Int. J. of Syst. Applicat., Eng. & development 5 (4), 477–484.
Liu, J., Yao, J., Gallaher, M., Coburn, J., Fernandez, R., 2008. Study on
methane emission reduction potential in chinas oil and natural gas ind. Tech.
rep., Tech. rep., April.
EP
Ma, C., Yu, S., Huo, J., 2010. Negative pressure wave-flow testing gas pipeline
leak based on wavelet transform. In: Computer, Mechatronics, Control and
Electronic Engineering (CMCE), 2010 International Conference on. Vol. 5.
IEEE, pp. 306–308.
Mpesha, W., Gassman, S. L., Chaudhry, M. H., 2001. Leak detection in pipes
by frequency response method. J. of Hydraulic Eng. 127 (2), 134–147.
C
10
ACCEPTED MANUSCRIPT
PT
RI
SC
Table Appendix A.2: Synthetic gas network specifications
Branch Start End Length Diameter
ID Node Node (m) (inch)
1 101 100 216.1 6
2 102 101 199.9 6
U
3 106 102 98.1 8
4 106 105* 78.9 3
5 105* 104 139.9 3
6
7
104
104
101
103
92.0
199.9
4
3
AN
8 103 100 81.1 8
9 100 120 135.9 6
10 121 120 20.1 4
11 134* 121 93.9 4
M
12 121 122* 75.9 4
13 122* 123* 78.0 6
14 126 122* 153.0 4
15 124* 123* 70.1 6
16 125 124* 63.1 6
D
11
ACCEPTED MANUSCRIPT
PT
37 118* 109 15.8 6
38 117 118* 111.9 6
39 117 116 88.1 6
40 119 117 84.1 6
41 116 115* 70.1 6
RI
42 115* 113 150.0 6
43 113 108 103.9 4
44 113 110* 252.1 6
45 114 113 89.9 4
46 110* 107 118.9 8
SC
47 111 110* 152.1 12
48 111 112 56.1 12
U
AN
M
D
TE
C EP
AC
12
ACCEPTED MANUSCRIPT
HIGHLIGHTS
PT
RI
• Leak Analytics System based on probabilistic and sensitivity approach.
SC
• Three stages of the system: leak detection, leak isolation, and leak localization.
• Robust and accurate model for detecting single and multiple leaks.
U
AN
M
D
TE
C EP
AC