Electric Power Systems Research 223 (2023) 109646

Contents lists available at ScienceDirect

Electric Power Systems Research

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

A travelling wave-based fault locator for radial distribution systems using

decision trees to mitigate multiple estimations✰
L.S. Lessa a, C.V.C. Grilo a, *, A.L. Moraes a, D.V. Coury a, R.A.S. Fernandes b
a
Department of Electrical and Computer Engineering, São Carlos School of Engineering, University of São Paulo, São Carlos, Brazil
b
Department of Electrical Engineering, Federal University of São Carlos, São Carlos, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: Electrical systems have been facing transformations, such as distributed generation insertion, system expansion
Distribution systems and regulatory standards in order to increase reliability and quality of the power supply. Thus, fault location
Fault location methods must be updated to ensure accuracy in estimating the location of electrical faults. The delay in restoring
Multiple estimation
the system causes damage to utilities and consumers. Considering this, the current work presents an approach
Decision trees
Travelling waves
capable of locating faults accurately in radial distribution systems. At first, the distance is estimated using the
travelling wave theory with data acquired from two terminals. Next, due to the radial characteristic of the
system, the proposal aims to mitigate the problem of multiple estimation of faults. Thus, features are extracted
from the voltage and current signals, which are used as inputs of decision trees to identify the fault region. The
proposed approach was validated in a medium voltage distribution system, in which the results presented an
average error of 0.79% (with a standard deviation of 0.4%) in estimating the fault distances and an average
accuracy above 88.7% in identifying the region under fault. Thus, it was demonstrated that the proposed
methodology is efficient to locate faults, mitigating the problem of multiple estimation.

1. Introduction self-healing mode [5]. In the past, many fault location methods have
been developed, some of them based on the apparent impedance
DISTRIBUTION systems have been constantly modified according to calculation. Although their accuracy is affected by problems such as
technological advances. The standards that regulate the power sector high impedances and fault inception angles, these methods are still in
are increasingly rigorous and they prioritize the quality of the service of use [4].
the power supply. The electrical system reliability is improved when On the other hand, travelling wave-based location techniques have
faults are detected and isolated quickly and accurately by protective proven to be efficient in the current context of distributed systems [6],
devices. To decrease the interruption time of the power supply after a which is represented by constant modification of the system load profile
fault, several studies have been carried out to develop reliable and ac­ and diversification of the network topology [7]. Considering the ad­
curate fault locators [1–3]. vances in technology, modern travelling wave recorders are capable of
Over the last decade, electrical systems have grown due to the operating with high sampling rates (between 1.25 MHz and 20 MHz),
increasing number of distributed generations, mainly represented by digital signal processing, synchronization and data exchange. Thus,
wind and photovoltaic power plants. These sources are connected these meters allow fault locators based on the travelling wave theory to
completely or partially in large distribution systems via inverters. This overcome barriers, aiming to consolidate this kind of meter in the
fact increased the topology complexity of electrical systems and the market [8].
diversity in system parameters. As a consequence, the difficulties and However, it is worth mentioning that distribution systems have
challenges for locating faults were increased [4]. branches and shunt loads that poses an even greater challenge for fault
Looking to the future of protection in electrical systems, detection, locators due to the problem of multiple estimation [9]. Recently, many
location and fault isolation leading to the restoration of the service are authors have sought new methods to reduce this problem, as proposed in
pillars for distribution systems, open space for devices that operate in a [10] and [11]. The methodology proposed in both studies made use of

✰
Paper submitted to the International Conference on Power Systems Transients (IPST2023) in Thessaloniki, Greece, June 12–15, 2023.
* Corresponding author.
E-mail address: [email protected] (C.V.C. Grilo).

https://doi.org/10.1016/j.epsr.2023.109646
Received 11 December 2022; Received in revised form 27 March 2023; Accepted 21 June 2023
Available online 7 July 2023
0378-7796/© 2023 Elsevier B.V. All rights reserved.
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

the concept of low voltage zones, which is based on voltage measure­

ments and fault location estimation using the system impedance matrix.
The proposed approach used several meters in the system and an
adaptive threshold, based on the calculation of the apparent impedance,
associated with the fault distance to indicate regions of voltage sag, and
thus reduce the problem of multiple locations of the fault.
Other studies proposed to reduce multiple estimation regions based
on data mining [12,13], which received the voltage and current signals
recorded by smart meters. Based on the extracted signals, a feature
vector was created that helped to define the relationship between the
missing events and the regions.
As an alternative to the methods presented for multiple estimation
reduction, the approach proposed by [9] made use of only one phase
involved in the fault. This approach used positive sequence components
and an adaptive threshold to reduce the multiple estimation of faulty
regions. Furthermore, the authors analyzed the behavior of the meth­
odology as a function of the reduction in the number of meters.
As a solution to the difficulties encountered in the related literature,
this article proposes an application of a fault location technique based on
travelling waves together with decision trees to determine the distance
and region of the system under fault, respectively. The location tech­
nique adopted made use of signals recorded through meters, allocated at
the ends of the system. On the other hand, multiple estimation mitiga­
tion made use of data from a virtual meter positioned in the substation Fig. 1. One-line diagram showing the first test system and division of regions
(topology I).
and connected meters at the ends of the system. These data were
delivered to a set of specialized classifiers in each type of fault. It is also
worth mentioning that each classifier is responsible for a region of the which can be visualized in Fig. 2. It is worth mentioning that the stages
system. for fault detection and classification were not implemented in this work,
In short, the contributions of this work are as follows: as some approaches in the literature present high precision for these
tasks [14,15].
• applicability of the technique of locating faults based on travelling
waves for distribution systems; 3.1. Data acquisition
• proposal of the necessary structure for fault location in distribution
systems, from the stage of data acquisition, precise calculation of the In this stage, in accordance with [3], the signals were sampled at a
fault location, as well as the mitigation of the multiple estimation frequency of 12 MHz and each meter recorded voltages and currents of
problem; all phases of the system. Following this, the recorded signals were stored
• mitigation of the multiple estimation problem using machine in a database accessed for the offline step execution. It is important to
learning algorithms. mention that due to the fact that the fault locator required measure­
ments from two terminals, it was considered that the meters were syn­
The work is organized as follows. Section II presents the test system chronized by a GPS (Global Positioning System). Furthermore, in this
and details about the simulations performed. Section III presents the measurement strategy, the meters must be allocated at the substation
complete structure and steps for online and offline execution of the and at the end of each branch.
proposed methodology. The results with discussions are presented in
Section IV. Finally, the conclusions are presented in Section V.

2. Modelled and simulated radial distribution system

Fig. 1 presents the first test system based on the CIGRE electrical
distribution system (topology I), which is inspired by a real distribution
system located in southern Germany. This system consists of a balanced
and symmetrical medium voltage network, composed by two 20 kV
feeders (T1 and T2) and 14 buses. For the execution of this study, this
system was modeled in the PSCAD™/EMTP™ software and adapted to
operate at a frequency of 60 Hz to meet the Brazilian context. Thus, all
parameters of the system lines, represented for the frequency of 50 Hz,
were recalculated to adjust the frequency of 60 Hz. Meters were allo­
cated at the ends of the feeders, called M1 to M6, to collect the system
operation data.
In addition, the CIGRE system was segmented into 5 non-overlapping
regions (S1 to S5). These areas are important for the multiple estimation
mitigation process.

3. Proposed methodology

In this section, details of the proposed methodology for locating the

fault and mitigation of the multiple estimation of faults are presented, Fig. 2. Flowchart representative of the proposed methodology.

2
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

3.2. Fault detection and classification electrical systems is adequate, because it can measure the time in­
stants of the reflections of the travelling waves generated during a
The latest methodologies adopted for fault detection and classifica­ fault [1,3].
tion were based on intelligent algorithms. These steps can be performed
directly on the relay and implemented according to [14] and [15]. In practice, the Wavelet transform entails grouping pairs of filters
Considering that the fault detection rate reached in these papers is (low pass and high pass) at each scale step. These steps can be consid­
99.9%, the models implemented in this paper did not use data repre­ ered as successive approximations of the same function, where each
senting non-fault situations. approximation details information from a given frequency range. This
successive filtering process is called Multi-Resolution Analysis.
To detect the arrival time of a travelling wave, each instant of the
3.3. Fault distance estimation signal was analyzed with a threshold, established by means of the
comparison with the maximum value of the detail coefficients, for each
This subsection details the steps taken to estimate the distance from meter, from a signal in a steady-state situation. To detect the correct
the fault using two terminals. Fig. 3 describes the steps performed in this time, the value established for the threshold was increased by 10%.
process. Depending on the network noise level, the time obtained was the
maximum value of the signal in up to 24 samples after considering the
1) Modal Transformation: in distribution systems, signals are best threshold. In this technique, only the time instant of arrival of the first
observed by decoupling phases into modal components, adopting the wave, reflected in each of the two terminals, was detected.
modal transformation technique. Modal transformation allows the
three-phase system to be treated as a system with three single-phase 3) Fault distance estimation: after detecting the highest value using the
circuits using the three decoupled modes - ground mode (mode 0) threshold established to the terminal meter A, the same procedure
and two aerial modes (α and β modes) - each with its own charac­ was repeated with the signal recorded by the other meter (B). These
teristics [6]. instants were denoted as Ta and Tb, as presented in Fig. 4 for a fault
case.
The modes were obtained by (1), which were a transformation ma­
trix applied to three-phase systems: After detecting the time instants, the delay was calculated using (3):
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
V0 1 1 1 Va Tdelay = Tb − Ta . (3)
⎣ Vα ⎦ = 1⋅⎣ 2 − 1 − 1 ⎦ ⋅ ⎣ Vb ⎦, (1)
3 √ ̅̅
̅ √ ̅̅
̅
Vβ 0 3 − 3 Vc Thus, the distance by the travelling wave was calculated and sub­
tracted from the total length of the line [1,2], based on the formulation
where V0 , Vα and Vβ are modal vectors of voltage and Va , Vb and Vc the of the distance given by (4):
voltage phase vectors.
The propagation speed for each mode was calculated by (2), as ̂ = l − Vp × Tdelay ,
D (4)
demonstrated in [1]: 2

1 in which D̂ is the estimated distance from the fault (in meters); Vp is the
v = √̅̅̅̅̅̅, (2)
LC wave propagation speed in the stretch for α mode (in m/s); l is the length
of the section covered by the meter (in meters); Tdelay is the time dif­
in which L and C are obtained from the series impedance and shunt ference between the measuring terminals (in seconds).
capacitance of the sequence components used to represent the electrical
system.
3.4. Determining possible fault regions
2) Time instant detection: the Wavelet transform is widely accepted for a
From the distance of the fault, estimated by the locator, it was
great variety of signals that are not periodic and that may contain
possible to determine in advance in which possible regions of the system
sine signals and impulsive signals, a common feature in electrical
the fault could have occurred. Considering the one-line diagram of
power systems. Thus, the application of Wavelet to locate faults in
Fig. 1, if the fault occurred at 7.5 km from the M1 meter, the possible
regions of the fault would be S2 or S3.
As presented in the next subsection, each system region had a specific
algorithm for analyzing the occurrence or not of a fault. Therefore, based
on the distance estimated by the fault locator, a pre-selection of possible
faulted regions was performed. Thus, only the expert algorithms of the
sensitized regions were activated. Furthermore, if no fault occurred, the
algorithms were kept on standby.

Fig. 3. Flowchart representing the steps performed to estimate the distance of

the faults. Fig. 4. Detection of the time instants of the reflected waves.

3
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

3.5. Mitigation of multiple estimations • Based on signal amplitude: Root mean square (F8), Peak value
(F9), and Difference between maximum and minimum window
This subsection presents the methodology used to deal with the (F10).
problem of multiple estimation. As seen before, the presence of branches
in distribution systems leads to a greater difficulty in dealing with the Finally, the ratio of the voltage and phasor current of each phase was
exact fault location. In order to address this problem, this article pre­ performed to represent the apparent post-fault impedance (F11); and the
sents an approach based on decision trees that complements the fault ratio between the superimposed phase voltage and current to represent
locator response by identifying the faulted region, as illustrated in Fig. 5. the apparent overlapping impedance of the system (F12). Each of these
features was separated into real (R) and imaginary (X) to represent the
1) Pre-processing: to perform this subroutine, the voltage and current data delivered to the intelligent algorithm.
signals presented by a virtual meter were considered, which emu­
lates the behavior of a real meter positioned at the substation. By 2) Identifying the fault region: in this study the machine learning algo­
applying Kirchhoff’s laws, the voltages and currents were obtained at rithm known as Extra-Tree-Classifier (or known as Extremely Ran­
a point of interest of the system. In addition to the virtual meter data, domized Trees Classifier) was used [18], which is capable of dealing
voltage and current data originating from real meters located at the with classification problems. This algorithm created Decision Trees
ends of the system, downstream of the substation, were considered. at random, combining the results of each tree to find the final deci­
Moreover, the application of a down-sampling step in the acquired sion. A Decision Tree has a structure similar to a flowchart, hierar­
signals was considered to adjust the sampling to 256 samples/cycle chically structured and comprising a set of interconnected nodes. For
(15,360 Hz), and thus contribute to reducing the computational each data sample presented for the tree, each internal node performs
burden of this subroutine. a conditional test of the type "if <condition>, then ...; else ...", to
determine an output response [19].
From the voltage and current signals obtained by the meters, the zero
sequence components were obtained, represented by V0 and I0 , As presented later, the reference distribution system of this study was
respectively. Afterwards, the superimposed components were calculated divided into regions. The proposed algorithm had the task of identifying
involving the voltages and instantaneous currents of each of the phases the region of the system in which the fault possibly occurred, thus
(Va sup , Vb sup , Vc sup , Ia sup , Ib sup and Ic sup ) and zero sequence voltages mitigating multiple estimations. To do this, the problem was modeled as
and currents (V0 sup , I0 sup ), only for the virtual meter. These calcula­ a binary case, in which class 1 represents the occurrence of the fault in
tions, represented by the difference between the post- and pre-fault each region and class 0 to non-occurrence. For each type of fault an
signals, reduce the impact on the system against the loading variations ensemble of Extra-Trees was established, in which each of them was
[16]. responsible for a region of the system.
In addition, to reduce the influence of high frequency transients on The classifiers were implemented in Python language using the
the calculation of superimposed components, the second signal cycle ExtraTreesClassifier function of the Scikit-Learn library. The algorithms
prior to the beginning of the fault was considered as a pre-fault and the were configured to operate with unbalanced database and with a
third signal cycle after the fault as a post-fault. number of estimators equal to 100.
Afterwards, features of the previously presented signals were For training and validation of the proposal, 2,484 fault cases were
extracted. Thus, the features proposed in [17] were extracted, as simulated in the PSCAD™/EMTDC™. Based on these cases, a 10-fold
follows: cross-validation [20] was performed to train and validate the models,
while 30% of the data were used to test them. More details about the
• Based on statistics: Harmonic Mean (F1), Standard Deviation (F2), simulated cases are presented next.
Mean Deviation (F3) and Kurtosis (F4);
• Based on the amount of signal information: Entropy (F5), Shan­ 3.6. Simulation of fault conditions for topology I
non Entropy (F6) and Rényi Entropy (F7);
To validate the present work, the study was divided into two parts,
namely: (i) estimation of the fault distance and (ii) mitigation of mul­
tiple estimations. To analyze the performance of the fault locator, LG,
LL, LLG and LLL faults were considered, distributed into 10, 25, 40, 50,
75 and 90% of the length of each section of the system. The fault re­
sistances ranged from 0, 10 and 25 Ω and the fault inception angles were
0, 45 and 90◦
On the other hand, in order to analyze the mitigation of multiple
estimation of faults, the same types of faults were considered, which
were applied in the regions S1 to S5 in distances of 100 in 100 meters for
each section, with fault resistances randomly varying between 0.001
and 30 Ω and with fault inception angles ranging randomly between
0◦ and 100◦ These parameters were adapted from [21].

3.7. Performance evaluation metrics

In this study, only the estimated distance from the fault was adopted
as a metric to evaluate the behavior of the fault locator. Therefore, the
relative error (5) was calculated by:

D− D̂
Errorrel [%] = × 100, (5)
l

Fig. 5. Flowchart of multiple estimation mitigation routine. where D is the actual distance from the fault. However, in order to obtain

4
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

an overview, the average percentage error was still considered. occurrence and severity provided by the fault of the electrical network.
To evaluate the performance results of each classifier used to miti­ The result of the algorithm’s performance in relation to the type of fault
gate the multiple estimation, individually for each region, the use of is observed in Fig. 7.
confusion matrices was considered [22]. Through a confusion matrix, The greatest influences occurred for LG and LLG faults. The biggest
the number of false positives (FP), false negatives (FN), true positives errors were 1.09% and 1.03%, for LG and LLG faults, located by the M5-
(TP), and true negatives (TN) resulting from the validation process for M6 meter pair. The smallest errors were 0.52% and 0.53% in LL and LLL
the test data were reported and compared. Given this information, the faults. Although the largest error obtained was 2.7%, the results validate
accuracy (Acc) of the model was calculated, according to (6): the fault locator when used in distribution systems with radial topology,
that is, for a scenario that presents greater complexity when detecting
TP + TN
Acc|[%] = × 100. (6) time instants of the travelling waves.
TP + TN + FP + FN
Additionally, the percentage of multiple estimation reduction (R%),
4.2. Mitigation of multiple estimations
proposed by [12], was calculated as follows:
∑
Dr (n) Unlike the estimation of the fault distance, the subroutine for miti­
R% = ∑ × 100, (7)
Dsme (n) gating multiple fault estimation is a stochastic process that needs to go
through a learning stage. This subroutine was performed to identify in
where n is the total number of simulated faults; Dr (n) represents the which region of the system the fault occurred. Figs. 8–11 show the in­
distance between the beginning of the region under multiple estimation dividual performance of each classifier for a given region and type of
to the fault location, taking into account only the cases in which the fault. "F" indicates the presence of a fault in the region, while "NF" in­
faults were correctly located by the algorithm; and Dsme (n) is similar to dicates the absence of a fault in the region. The first element of the
Dr (n), however, considering all the faults correctly located or not. matrix (TP) denotes that there was a fault in the region and the estimator
correctly identified it as such. The second element (FP) denotes that
4. Results there was no fault in the region, but the estimator incorrectly identified
it as faulty. The third element (FN) denotes that there was a fault in the
The results obtained from the proposed methodology considering region, but the estimator failed to identify it. The fourth element (TN)
topology I are presented in this section. First, the results of the fault denotes the absence of a fault in the region and the estimator correctly
locator based on travelling waves are analyzed, and then the results of identified it as such.
multiple estimation mitigation are presented. Observing the results for single-phase faults, represented in Fig. 8, it
can be observed that the classifiers individually achieved an average
accuracy of 95.1% to indicate whether a fault occurred (class 1) in a
4.1. Location of faults given region of the system. For regions S2, S3 and S4, for example, the
accuracies found for TP were 92.7%, 87.3% and 90.7%, respectively.
In all the results presented below, the average relative errors for all Moreover, in relation to TN (no occurrence of fault in the region), the
cases of the fault were considered. Only average errors were considered classifiers individually reached accuracies between 99.3% and 100% for
to evaluate the performance, as the location techniques based on trav­ single-phase faults, highlighting the results attributed to regions S1 and
elling waves are robust and have virtually no considerable influences for S5, with performances equal to 100%.
inception angles and the resistances of the fault. Concerning biphasic faults (Figs. 9 and 10), the performance was
Fig. 6 shows the performance of the locator in relation to the distance quite similar to the scenario of single-phase faults. In general, the ac­
from the fault occurrence to the meter. Errors were higher for faults very curacies varied within the range of 99.3% and 100% for the identifica­
close to the meter due to the high frequencies resulting from the fault. tion of TN and 80.6% and 100% for TP regarding non-grounded biphasic
Considering the pairs of meters corresponding to the feeder derived from faults; and accuracies between 99.2% and 100% for TN and between
the T1 transformer, the largest errors did not exceed 1.5%. The largest 87.8% and 100% for TP regarding grounded biphasic faults. Therefore,
error of 2.7% was obtained for the pair of feeder meters derived from the it can be concluded that the absence of ground in biphasic faults implied
T2 transformer. The major errors occurred due to the difficulty of in a subtle reduction in the performance when identifying the faulty
detecting the instant time of the first wavefront reflected at the location region, especially concerning regions S3 and S4.
of the fault. Faults applied exactly at half the length of the line between The individual results of the classifiers for three-phase faults are
the pair of meters obtained the smallest errors ranging between 0.04% presented in Fig. 11. For TN identification, accuracies between 99.4%
and 0.36%, almost representing an exact location of the fault. and 100% were achieved; while for TP accuracies, they were between
Another performance analysis of the proposed algorithm was carried 91.3% and 100%.
out for 10 different types of faults, which were applied in all pre- Although the aforementioned results indicate a high accuracy, they
established distances, prioritizing the relationship of the probability of only show the individual performance of each classifier for a given faulty
region. Above all, the main scope of this work is to indicate in which of

Fig. 6. Locator accuracy for different distances from the electrical system (to­
pology I). Fig. 7. Locator accuracy in relation to the type of fault (topology I).

5
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

Fig. 8. Results for fault line-ground.

Fig. 9. Results for fault line-line.

Fig. 10. Results for fault line-line-ground.

the regions of the system the fault occurred. The performance obtained accuracy for LLL faults (92.6%) and a lower accuracy for LG faults
specifically for each type of fault is shown in Fig. 12. (85.6%).
Based on the mitigation results (Fig. 12), an average accuracy of It is important to note that the unmitigated faults, indicated in Fig. 12
88.7% was obtained in the correct identification of the fault region. represent the cases in which the algorithm either indicated more than
Among the four types of faults tested, there was a higher identification one region as a fault location or did not indicate any region even if there

6
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

Fig. 11. Results for fault line-line-line.

Table 2
Mitigation result in terms of R% for topology I.
Fault type LG LLG LL LLL

R% 89.7 85.9 90.0 93.5

technique was applied to verify the robustness of the proposed

approach. In [11], in turn, the R% calculated by [9] showed that the
average percentage of reduction was less than 80%, which was much
lower than the performance attained in this study.

5. Tests for a different topology

In order to validate the reliability of the proposed framework, the

methodology was implemented in a second topology for the distribution
network shown in Fig. 13 (topology II).
Fig. 12. Multiple estimation mitigation results for topology I. This system differs from the one in Fig. 1 due to the presence of an
extended branch belonging to zone S4 and the meters were relocated at
was a fault in any of them. For LG type faults, for example, 11.4% of the ends of each branch.
these were not identified in the regions subject to multiple estimation Tests were carried out considering faults to the new topology,
and 1.0% returned with more than one indicated region. In the case of varying fault resistances and angles. The location methodology using
LLG faults, 13.9% were not identified and 0.5% with more than one two terminals achieved better performance in fault location for topology
region indicated. For LL faults, 9.8% were not identified and 1.2% ob­ II. This can be observed in Fig. 14, where faults applied in a branch with
tained more than one indicated region. And for LLL faults, 6.5% of faults 20.1 km between meters M1-M4 in the system of Fig. 2 (topology II)
were not identified and 0.9% showed more than one identified region as presented more accurate calculations of the fault distance if compared to
output (Table 1). the pair of M1-M4 in topology I (Fig. 1).
Furthermore, Table 2 presents the evaluation of the results using the The performance in topology II reached a fault location 11% better
R% metric. LLL faults continue to represent the cases in which there was than in topology I, and for faults located near the middle of the line, the
the greatest reduction in the multiple estimation (93.5%), while the LLG error was almost zero. It should be noted that for the methodology
had the lowest reduction (85.9%). adopted, the greatest difficulty lies in locating faults near the meters due
Finally, it is important to comment that, despite different approaches to the difference in propagation time between the meters. Furthermore,
and test distribution systems, the results are in line with those obtained other combinations of meter pairs (different from M1-M4 mentioned)
in [9,12] and [11]. While the present study reached an average per­ did not show significant variations from the values depicted in Fig. 6 for
centage reduction value (R%) of 89.8%, the work by [9] obtained an different fault locations.
average reduction value of 91.9%. Regarding [12], the average per­ Regarding the problem of multiple estimation, it is worth noting that
centage of reduction achieved was 94.2%. However, no cross-validation changing the system topology did not have any adverse effect on the
algorithm’s performance, as shown in Fig. 15. The results of the con­
ducted tests indicate that an accuracy rate of over 92.3% was achieved
Table 1 in correctly detecting the fault zone, which is higher than the value
Results in terms of non-identification of region under fault or unmitigated
obtained for topology I. Furthermore, it can be observed that the best
multiple estimation for topology I.
and worst detection performances were still found for LLL and LLG
Type of problem Fault type faults, respectively. Detailed results of the mitigation are shown in
LG LLG LL LLL Table 3.
No region identified 11.4% 13.9% 9.8% 6.5% The results based on the R% metric are presented in Table 4. It is
Multiple regions identified 1.0% 0.5% 1.2% 0.9% worth noting that the values obtained are similar to those found for

7
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

Fig. 13. The second test system and division of regions (topology II).

Table 3
Results in terms of non-identification of region under fault or unmitigated
multiple estimation for topology II.
Type of problem Fault type

LG LLG LL LLL

No region identified 8.0% 11.4% 6.8% 2.6%

Multiple regions identified 0.2% 1.2% 0.3% 0.2%

Table 4
Mitigation result in terms of R% for topology II.
Fig. 14. Fault location errors for topologies I and II considering meters M1-M4. Fault type LG LLG LL LLL

R% 92.2 87.0 92.9 96.9

topology I. Only a slight increase of approximately 3% was noted for LG,

LL, and LLL faults, and a 1.1% increase was observed for LG faults.

6. Conclusions

This article proposed an application of a fault location technique

based on travelling waves for radial distribution systems together with
decision trees to determine the faulty region of the system, respectively.
To do this, a methodology was proposed that integrates the accuracy of a
fault locator based on travelling waves with an ensemble of decision
trees responsible for identifying the faulty region with high precision
and, consequently, mitigate the problem of multiple estimation. Unlike
the approaches found in the literature, a virtual meter was considered
and allocated in the substation to obtain the voltage and current signals
used as inputs (followed by a feature extraction stage) of the multiple
Fig. 15. Multiple estimation mitigation results for topology II. estimation mitigation technique. Considering the results, it was
observed that the proposed approach was able to locate the faults pre­
cisely, mitigating the problem of multiple estimations. It is important to

8
L.S. Lessa et al. Electric Power Systems Research 223 (2023) 109646

mention that the method presented in this work can be adapted to [2] E.C.M. Maritz, J.M. Maritz, M. Salehi, A travelling wave-based fault location
strategy using the concepts of metric dimension and vertex covers in a graph, IEEE
different topologies. For each analyzed system, it is essential to divide
Access 9 (2021), 155 815–155 825.
the previous segmentation into non-overlapping regions. Then, for the [3] L. Xie, L. Luo, Y. Li, Y. Zhang, Y. Cao, A traveling wave-based fault location method
multiple estimation method, the decision tree model for each region employing vmd-teo for distribution network, IEEE Trans. Power Deliv. 35 (4)
should be retrained, considering the information collected by the meters (2020) 1987–1998.
[4] IEEE, Ieee guide for determining fault location on ac transmission and distribution
installed at the ends of each branch. lines, in: IEEE Std C37.114-2014 (Revision of IEEE Std C37.114-2004), 2015,
pp. 1–76.
[5] G. Liu, T. Jiang, T.B. Ollis, X. Li, F. Li, K. Tomsovic, Resilient distribution system
Funding
leveraging distributed generation and microgrids: a review, IET Energy Syst.
Integr. 2 (4) (2020) 289–304.
This work was supported by the São Paulo Research Foundation [6] R. Kumar, D. Saxena, A traveling wave based method for fault location in multi-
(FAPESP), Brazil [Grant Number 2021/04872-9] and Coordenação de lateral distribution network with dg. 2018 IEEE Innovative Smart Grid Technologies -
Asia (ISGT Asia), 2018, p. 712.
Aperfeiçoamento de Pessoal de Nível Superior (CAPES), Brazil [Finance [7] K. Prabakar, A. Singh, M. Reynolds, M. Lunacek, L. Monzon, Y.N. Velaga, J. Maack,
Code 001]. S. Tiwari, J. Roy, C. Tombari, I.Mendoza Carrillo, X. Yang, S. Iyengar, W. John, D.
Anand, A. Gopstein, and D. Vaidhynathan, “Use of traveling wave signatures in
medium-voltage distribution systems for fault detection and location,” Technical
CRediT authorship contribution statement Report, 2021.
[8] F.V. Lopes, R.L.A. Reis, K.M. Silva, A. Martins-Britto, E.P.A. Ribeiro, C.M. Moraes,
L.S. Lessa: Conceptualization, Methodology, Software, Formal M.A.M. Rodrigues, Past, present, and future trends of traveling wave-based fault
location solutions, in: 2021 Workshop on Communication Networks and Power
analysis, Investigation, Data curation, Visualization, Writing – original Systems (WCNPS), 2021, pp. 1–6.
draft, Writing – review & editing. C.V.C. Grilo: Conceptualization, [9] Y.V. Tresso, R.A. Fernandes, D.V. Coury, Reducing multiple estimation for fault
Methodology, Software, Formal analysis, Investigation, Data curation, location in medium voltage distribution networks, Electr. Power Syst. Res. 199
(2021).
Visualization, Writing – original draft, Writing – review & editing. A.L. [10] F.C.L. Trindade, W. Freitas, J.C.M. Vieira, Fault location in distribution systems
Moraes: Conceptualization, Methodology, Software, Formal analysis, based on smart feeder meters, IEEE Trans. Power Deliv. 29 (1) (2014) 251–260.
Investigation, Data curation, Visualization, Writing – original draft, [11] F.C. Trindade, W. Freitas, Low voltage zones to support fault location in
distribution systems with smart meters, IEEE Trans. Smart Grid 8 (2017)
Writing – review & editing. D.V. Coury: Supervision, Resources, 2765–2774.
Writing – review & editing, Project administration, Funding acquisition. [12] J.V. Sousa, E.A. Reche, D.V. Coury, R.A. Fernandes, Cloud computing in the smart
R.A.S. Fernandes: Supervision, Writing – review & editing. grid context: an application to aid fault location in distribution systems concerning
the multiple estimation problem, IET Gener. Transm. Distrib. 13 (18) (Sep. 2019)
4222–4232.
Declaration of Competing Interest [13] E.A. Reche, J.V. d. Sousa, D.V. Coury, R.A.S. Fernandes, Data mining-based method
to reduce multiple estimation for fault location in radial distribution systems, IEEE
Trans. Smart Grid 10 (4) (2019) 3612–3619.
The authors declare that they have no known competing financial [14] M. Shafiullah, M.A. Abido, S-transform based ffnn approach for distribution grids
interests or personal relationships that could have appeared to influence fault detection and classification, IEEE Access 6 (2018) 8080–8088.
[15] M. Mishra, P.K. Rout, Detection and classification of micro-grid faults based on hht
the work reported in this paper. and machine learning techniques, IET Gener. Transm. Distrib. 12 (2) (2018)
388–397.
Data availability [16] R. Aggarwal, D. Coury, A. Johns, A. Kalam, A practical approach to accurate fault
location on extra high voltage teed feeders, IEEE Trans. Power Deliv. 8 (3) (1993)
874–883.
Data will be made available on request. [17] F.A.S. Borges, R.A.S. Fernandes, I.N. Silva, C.B.S. Silva, Feature extraction and
power quality disturbances classification using smart meters signals, IEEE Trans.
Indus. Inform. 12 (2) (2016) 824–833.
[18] P. Geurts, D. Ernst, L. Wehenkel, Extremely randomized trees, Mach. Learn. 63 (1)
Acknowledgments (2006) 3–42.
[19] I.H. Witten, E. Frank, M.A. Hall, Data Mining: Practical Machine Learning Tools
and Techniques, 3a ed., Morgan Kaufmann, Amsterdam, 2011. Amsterdam.
The authors would like to thank the São Carlos School of Engineer­ [20] G. James, D. Witten, T. Hastie, R. Tibshirani, An Introduction to Statistical
ing, University of São Paulo, Brazil for the facilities provided. Learning: With Applications in R, Springer Publishing Company, 2014.
Incorporated.
[21] E.A. Reche, J.V. d. Souza, D.V. Coury, R.A.S. Fernandes, Data mining-based
References method to reduce multiple estimation for fault location in radial distribution
systems, IEEE Trans. Smart Grid 10 (4) (2019) 3612–3619.
[1] F.V. Lopes, K.M. Dantas, K.M. Silva, F.B. Costa, Accurate two-terminal transmission [22] R.O. Duda and D.G. Stork, “Pattern classification,” 2a ed., New York: Wiley, vol. 10,
line fault location using traveling waves, IEEE Trans. Power Deliv. 33 (2) (2018) no. 4, pp. 3612–3619, 2001.
873–880.