=Paper=
{{Paper
|id=Vol-2940/paper44
|storemode=property
|title=Revisiting Security Threat on Smart Grids: Accurate and Interpretable Fault Location Prediction and Type Classification
|pdfUrl=https://ceur-ws.org/Vol-2940/paper44.pdf
|volume=Vol-2940
|authors=Carmelo Ardito,Yashar Deldjoo,Eugenio Di Sciascio,Fatemeh Nazary,Vishnu Ramesh,Sara Abraham,Vinod P,Isham Mohamed,Corrado A. Visaggio,Sonia Laudanna
|dblpUrl=https://dblp.org/rec/conf/itasec/ArditoDSN21
}}
==Revisiting Security Threat on Smart Grids: Accurate and Interpretable Fault Location Prediction and Type Classification==
https://ceur-ws.org/Vol-2940/paper44.pdf
Revisiting Security Threat on Smart Grids: Accurate
and Interpretable Fault Location Prediction and Type
Classification
Carmelo Ardito1 , Yashar Deldjoo1 , Eugenio Di Sciascio1 and Fatemeh Nazary1
1
Politecnico di Bari, Italy
Abstract
This work revisits security threats on smart electrical grids and focuses on the dimensions of dependability,
serviceability, and accountability, which constitute the security requirements of an SG application. The
first two dimensions deal with fault diagnosis and location, while the last element tackles building the
system more transparent. We proposed a data-driven machine-learned fault prediction system that can
provide abrupt and accurate fault type classification and location prediction. Furthermore, we reported
a feature interaction visualization and elaborated on how this step can facilitate interpretation of the
results and assessment of the security threats in the SG. The evaluation of the system is carried out on a
large-scale dataset comprised of approximately 1.9K training samples. Results show the effectiveness of
the proposed system both in prediction and interpretability steps.1
Keywords
smart grid„ interpretability, security, fault prediction
1. Introduction and Context
Smart electrical grids – a.k.a smart grids (SGs) – are a complex infrastructure of distributed
energy resources, appliances, and facilities that enable efficient usage, and optimization of assets,
thereby reducing power consumption and investment costs. In simple terms, an SG can be
viewed as an electricity market with the capacity to generate and store energy and shift load for
customers [1]. Electrical grids however are susceptible to a variety of electrical abnormalities,
failures, and security threats, which if not tackled abruptly, in some cases they can leave a
devastating impact on lives and the country’s critical industries, leading to a national dilemma.
As an example, in August 2003, in the Northeast United States and Ontario, Canada, a fault
occurred n the 345KV line transmission line, which triggered a cascade effect, and caused a
wise-area blackout for several days. As the consequence, more than 50 million people were left
without power, and an estimated loss of 4 to 10 billion dollars was induced [2, 3].
Traditionally, protection devices and circuit breakers are used to monitor faulted lines and
locations [3]. However, the power outage examination report [2] outlined that undesired
functioning of protective relays and circuit breakers was the main reason for the cascading
1
Authors are listed in alphabetical order. Corresponding author: Fatemeh Nazary
ITASEC’21: Proceedings of the 15th Italian Conference on Cyber Security, April 07–09, 2021, Online event
Envelope-Open carmelo.ardito@poliba.it (C. Ardito); yashar.deldjoo@poliba.it (Y. Deldjoo); eugenio.disciascio@poliba.it
(E. D. Sciascio); fatemeh.nazary@poliba.it (F. Nazary)
© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
effects and subsequent blackout in North America’s disaster in 2003. These examples call for
intelligent, fast, and accurate power system security assessment and fault diagnosis systems.
One of the key requirements of SGs is the management of electrical flow to guarantee
the reliable, secure, and economical transmission of electricity. Gunduz et al. [1] outline
certain security objectives that have to be satisfied to ensure security in an SG, they include
confidentiality, integrity, and availability — or the CIA triad. Confidentiality refers to the
protection of data from unauthorized disclosure. Integrity refers to the prevention of data
from unauthorized alteration and destruction, and availability means accessibility of data to
authorized parties in the SG when needed without compromising security. Typical cyber-attacks
in SG applications target at least one of the CIA triad. In this work, we focus our attention on
availability threats defined in this work as electrical faults that can occur in the distribution
line, causing interruption or delay to the power delivery to consumers.
In addition to the security objectives of the CIA triad, there are certain security requirements
that need to be fulfilled to ensure cyber-security in SG applications. They include authentication,
accountability, privacy, dependability, survivability among others [1]. We focus our attention
on the dimensions dependability, survivability and accountability, and study them in a real-life
simulation of electrical grid and fault scenario. Dependability is the capacity of the system to
deliver its services in-time and accurate manner, and that services be delivered even during fault
periods. A requirement for ensuring dependability includes fault detection, fault forecasting,
and fault prevention. Survivability aims to deliver services both in the existence of malicious
activities and external faults. Fault location, maintainability, and security policies are among
prominent measures for survivability. Accountability operations allow understanding who
is responsible for a problem in the times they emerge, which is achieved by providing more
transparent evidence about the grid functioning.
Motivated by this observation, in this work we present a data-driven self-healing system
using machine learning (ML) techniques that can perform automatic and timely detection
technologies of fault types and locations. Furthermore, the proposed system can keep human
operators informed about why certain decisions were made by producing an interpretable
visualization of the outcome. We try to answer the following research questions throughout
this work.
• What is the fault type occurred in the electrical grid? known as the fault type classification
(FTC) problem.
• Where the fault has occurred within the electrical grid network? known as the fault
location prediction (FLP) problem.
• Why the ML system made certain FLP, or FTC decisions? known as the interpretability
problem.
FTC and FLP enable isolation of the faulty region from the rest of the grid, via providing
necessary information regarding fault location and its characteristics, which expedites the
required repair works (survivability and dependability). The interpretability step however aims
to keep human operators involved in the control loop, by designing interpretable ML models
that can replace the black-box prediction models and produce rules that can be understood with
little inspection (accountability).
Recent literature for developing self-healing features on SGs has concentrated on machine
learning (ML) and transform domain for feature extraction as an input of ML system to predict
the location of the fault. In this regard, the most popular approaches for feature extraction
apply transform domain such as Discrete Fourier transform (DFT) [4] whose goal is to identify
fault location in presence of fault or attack. Some other works [5, 6, 7] apply Discrete wavelet
transform (DWT) to address the FLP and FTC problems.
The contributions of this work are multi-fold:
1. Information processing and prediction: for given measurement of the three-phase
voltage signals (one of the phase), extracted features from time, frequency, and wavelet
domains. For feature representation, we investigate the impact of several statistical
aggregation function including compute the 𝑛-th moment of the probability distribution
functions (PDFs) [8] (𝑛 ∈ [1, 4]) together with the energy and maximum level of the
signals. Further, we explore the predictive power of large class of linear and non-linear
classifiers (5 classifiers in total) on the above features. Compared with the prior literature,
this work investigates the impact of a large suite of factors such as the feature domain,
aggregation function for feature representation and the core classifier, which makes the
insights obtained very useful.
2. Interpretability: To better facilitate the interpretability of the black-box ML model
(classifier), we utilize a model-dependent feature importance measurement and visualize
the impact of features and their pairwise relationship on the classification outcome. In
particular, the proposed interpretability apporach would allow us to answer fundamental
questions about which feature class (time, frequency, or wavelet), and which statistical
aggregation operator (e.g., mean, norm, skewness) are the key to building discriminative
feature descriptors for the classification task.
3. Large scale dataset: Unlike prior works, we make use of a large scale dataset (containing
1848 measurements) to report the final results obtained.
The proposed FTC and FLP system in this work moves several steps forward our previous effort
in [9] along the following directions:
• In [9], we addressed only the FTC problem. In this work, we address both FTC and FLP
tasks.
• In [9], we used time and frequency representation analysis. Here, we use time, fre-
quency, and wavelet representations.
• In [9], for the interpretability analysis, we only showed the possibility of using interaction
plots. Due to space limitation, we did not provide details about what these interpretrations
actually are. In this work, we provide a deeper analysis of the interpretability step and
mention exactly which insights this step provides us about the impact of feature classes,
and statistical aggregation operators, used for feature representation.
• Previously our dataset was comprised of 140 training instances. In this work, we use a
dataset including 1848 samples, that is approximately 12 times bigger in size. This would
increase the validity of the results. Furthermore, to make the work reproducible, we plan
to release the dataset online.
IEEE-13 Node test feeder
(distribution grid)
Fault injection Fault injection 7 Fault types
to 4 zones
(1) Line-to-line
(1) 632-671 (AB, AC, BC)
(2) 632-633 (2) Line-to-ground
Measurements of three-phase voltage (3) 692-675 (AG, BG, CG)
signals from faulty zones
(4) 671-680 (3) Three-phase
(ABC)
Extracting Extracting
Signal-level features Transform-domain features
DFT DWT
Concatenation
Multi-class single label Interpretability via pairwise
Classification feature importance
FTC FLP
Figure 1: The flowchart of the proposed system
2. Proposed method
The proposed system in this work takes as the input a voltage signal, representing a voltage
measurement of the IEEE-13 node test feeder taken from certain phase (A, B, or C) and zone
in the grid, and produce as output two score related to sub-tasks: FLP (zone 1, 2, 3, and 4) and
FTC, line-to-ground (AG, BG, CG), Line-to-Line (AB, AC, BC), and three-phase fault (ABC). It
further produces a visualization of pairwise interaction of features on the classification outcome,
providing a global understanding of the role of different feature types and representation tech-
niques. The pipeline of processing steps (feature extraction and representation), classification,
and interpretability is shown in Figure 1.
2.1. Fault simulation and Feature Extraction
For simulation, we made use of the IEEE-13 node test feeder, which is a distribution network
operating at 4.16 kV. It consists of a set of characteristics that are in common with actual
networks such as unbalanced loads, voltage regulators etc [10]. We divided the network into
four critical zones, i.e. (1) zone 1: 632-671; (2) zone 2: 632-633; (3) zone 3: 692-675; and (4)
zone 4: 671-680. Faults were injected to all of the identified zones. Then three-phase voltage
signals were measured from all that zones. To model FTC, we injected seven different short
circuit faults (i.e., AG, BG, CG, AB, BC, AC, ABC) into each zone and for each fault type and
zone, 22 measurements were collected corresponding to 22 resistance values fault resistance
𝑅𝑓 values in the range [0.001-2]. The entire simulation time was set to 𝑡 = [0 − 0.022] seconds
where faults have been injected at a certain start time 𝑡 = 0.01 and revoked at 𝑡 = 0.02 for all the
experimental case, therefore, 𝑡𝑓 = [0.01 − 0.02] represents the faulty period while 𝑡𝑛 = [0 − 0.01]
indicates the normal period. For feature extraction, for having a relative feature values, all the
features extracted from the faulty period 𝑡𝑓 were normalized by the same feature which was
extracted from the non-faulty (normal) period 𝑡𝑛 .
The key step toward successful ML-based prediction is (i) definition and (ii) extraction and
representation of features that are descriminative of the respective task at hand (e.g., FTC,
FLP). In this work, we chose three categories of features that were used in prior literature with
different attention level [5, 6, 11, 12, 13]:
• Time-domain: it refers to the original data measured in time domain. For the given
voltage signal 𝑥(𝑡), six aggregation functions were applied to produce a feature vector of
dimensionality six to represent the time domain feature vector.
• Discrete Fourier transform (DFT): to obtain richer information about frequency, volt-
age signals were also transformed to the frequency-domain by applying discrete DFT
according to 𝑋 (𝑓 ) = ℱ (𝑥(𝑡)), where ℱ denotes the DFT operation. Afterwards, the
same six aggregation functions used in the time domain were applied to the computed
spectrum, to produce a feature vector of dimensionality six for the frequency-domain
signal.
• Discrete Wavelet transform (DWT): DWT is a digital signal processing technique
that applies the concept of multi-resolution analysis to the signals [14]. In contrast
to DFT, the DWT uses short window at high frequencies and long window for low
frequencies, thus closely capturing characteristics of the (non-stationary) signals. In
multi-resolution analysis, at decomposition level 𝑖, there are approximation 𝐴𝑖 and detail
𝐷𝑖 , wavelet coefficients. Motivated by prior works [6, 5], we use a high number of
level-decomposition (five), and use 𝐴5 , 𝐷1∶5 .
After representing each feature in the respective domain (time, frequency or wavelet), we
calculate the following statistics to represent a feature in the specific domain. The statistics are
obtained by applying similar aggregation functions to all signals.
• The maximum value of 𝑥(𝑡)
• The energy of 𝑥(𝑡)
• The first moment (mean) value of 𝑥(𝑡)
• The second moment (standard deviation) value of 𝑥(𝑡)
• The third moment (skewness) value of 𝑥(𝑡)
• The forth moment (kurtosis) value of 𝑥(𝑡)
• The maximum value of 𝑋 (𝑓 )
• The energy of 𝑋 (𝑓 )
• The first moment (mean) value of 𝑋 (𝑓 )
• The second moment (standard deviation) value of 𝑋 (𝑓 )
• The third moment (skewness) value of 𝑋 (𝑓 )
• The forth moment (kurtosis) value of 𝑋 (𝑓 )
• The maximum value of the coefficients: 𝐴5 and 𝐷1∶5
• The energy of the coefficients: 𝐴5 and 𝐷1∶5
• The first moment (mean) value of the coefficients: 𝐴5 and 𝐷1∶5
• The second moment (standard deviation) value of coefficients: 𝐴5 and 𝐷1∶5
• The third moment (skewness) value coefficients: 𝐴5 and 𝐷1∶5
• The forth moment (kurtosis) value coefficients: 𝐴5 and 𝐷1∶5
Thus, the dimensionality of the feature vectors in the time and frequency domain is 6 of the
time and frequency domain. For the wavelet, we use 6 (coefficients) × 6 (aggregation operations)
producing a feature vector of dimensionality 36 for the wavelet domain. These feature would
constitute the main features used in the classification tasks, which we use them individually or
in combination.
2.2. Fault type classification and fault location prediction
In this work, we further extend our previous work [9], in terms of the core classification task, and
the classifiers’ choices to perform that specific task. As for the core task, we try both fault type
classification (FTC) and fault location prediction (FLP), which both are essentially multi-class
classification problems. We make use of a suite of classifiers ranging from Decision-Tree, SVM,
to KNN, and Ensemble methods (Bagged-Tree, subspace k-nearest neighbors).
2.3. Interpretability
Methods for ML interpretability aim to make the decision made by the ML model more transpar-
ent and provide insight into why certain decisions were made. The methods for interpretability
can be generally coupled or decoupled from the decision model (i.e., the classifier). In this work,
we made use of a decision-model informed approach for interpretability, involving visualizing
the impact of pairs of features that have the highest impact on the classification task. In other
words, we fixed one of the best winning classifiers in the previous task, and search for the
best pairs of features that produce the highest classification accuracy. We enumerated all
the possible pairs and performed the outcome classification and visualized the result using a
heatmap visualization.
3. Experimental setup
In this section, we explain the experimental setup in detail such as dataset (cf. Section 3.1), the
training setup and classifiers (cf. Section 3.2) that are used to validate the performance of the
proposed system.
3.1. Dataset
As mentioned in Section 2.1, for data collection and creating the training dataset, the distribution
system (IEEE-13) is divided into four critical zones. Afterwards, seven different faults were
injected into all zones. To augment the training dataset with further data, the data collection
was repeated for 22 different fault resistance values 𝑅𝑓 in the range of 0.001 to 2 for each type
of fault as listed in Table 1. Finally, three-phase voltage signals were measured independently
as individual signals. As the result of these steps, in total 4 (zones) × 7 (faults) × 3 (phases) ×
22 (resistance values) = 1848 training instances were created, which represent the size of the
dataset used in this work for the empirical evaluation.
3.2. Classifiers and training setup
Five different classifiers were applied for FTC and FLP tasks which include: (i) Decision tree
(DT), (ii) support vector machine (SVM), (iii) k-nearest neighbors, (iv) ensemble (bagged tree),
and (v) ensemble (subspace KNN). For the ensemble (bagged tree) classifier, the learner type was
the decision tree and the number of learners was equal to 30. Likewise, for ensemble (subspace
KNN), the number of learners was set to 30 and the subspace dimension was equal to 18. To
speed up the experiments, we used a hold-out validation (80%-20%) for training and test set.
The exact statistics of the training and test set are shown in Table 2. We used MATLAB for
feature extraction and classification experiments.
Table 1
characteristic of fault types, locations, and resistances.
Item Details
phase to ground AG, BG, CG
Fault type phase to phase AB, AC, BC
three phase ABC
zone 1 branch 632-671
zone 2 branch 632-633
Fault location
zone 3 branch 692-675
zone 4 branch 671-680
0.0010, 0.0273, 0.0535, 0.0798
0.1061, 0.1323 0.1586, 0.1848
Fault resistance 0.2111, 0.2374, 0.2636, 0.2899
0.3162, 0.3424, 0.3687, 0.3949
0.4212, 0.4475, 0.4737, 0.5, 1, 2
Table 2
IEEE-13 dataset: |𝒟|𝑇 — total number of data in dataset, |𝒟|𝑇 𝑟 — number of samples in training, |𝒟|𝑇 𝑒 —
number of samples in testing.
dataset |𝒟|𝑇 |𝒟|𝑇 𝑟 |𝒟|𝑇 𝑒
IEEE-13 1848 1478 370
4. Results and discussions
We discuss the results of empirical evaluation of the proposed system through two main sub-
tasks, the classification step, and the interpretability analysis, as presented in the following.
Classification: Table 3 summarizes the classification results using five classifiers for the
FLP and FTC tasks based on the three feature classes, i.e., 𝑇 𝑖𝑚𝑒, 𝐷𝐹 𝑇 , 𝐷𝑊 𝑇 studied in this work.
On average, DWT produces the highest classification accuracy, for all the experimental cases.
Time-based is ranked second and DFT produces the lowest quality. DFT produces a better result
than the time-domain signal only with SVM and for the FTC task. This can be explained by
the fact, DWT exploits a multi-resolution analysis, and thus it can be seen as a time-frequency
approach. Finally, it could be noted the highest quality of classification is achieved when all
the features are combined. Thus, on average the following relation holds about the quality of
different feature descriptors: 𝐴𝐿𝐿 > 𝐷𝑊 𝑇 > 𝑇 𝑖𝑚𝑒 > 𝐷𝐹 𝑇. As for the classifier type, we can
note that the Ensemble methods typically provide the highest quality of classification, followed
by SVM, i.e., 𝐸𝑛𝑠𝑒𝑚𝑏𝑙𝑒 > 𝑆𝑉 𝑀 > Others.
In summary, for FLP the best results are obtained for (ALL, Ensemble subspace k-nearest
neighbors) where the accuracy is 100% followed (DWT, Ensemble) with 99.7%. For FTC, the
best accuracy is obtained for (ALL, SVM) with 95.4%, and Ensemble (BT) with 93.5%.
Feature analysis and interpretability: The visualization of feature importance analysis
is shown in the heatmap of Figure 2. The heatmap shows the impact of both feature classes
and 48 features and their pairwise relationship on fault location prediction. We chose pairwise
comparison instead of a set comparison between three or more features in order to simplify
the study, because that it is easy to visualize two feature interactions on a 2D space. We show
the results of FLP with decision tree classifier for the interpretability analysis. We answer the
following experimental questions with the interpretability analysis:
Q1. Which feature classes (domains) have the most impact on the prediction task?
according to Figure 2, it can be noted that most of the orange and yellow regions which
correspond to highly accurate classification outcomes are related to DWT. In other words, it is
interesting to understand that the DWT feature class contains more discriminative information
(and in particular for 𝑑1 and 𝑑2) in comparison with the time and frequency domain which are
mostly blue-colored (regardless of the feature aggregation method). These results provide more
transparent information about details of the ML prediction, and that in fact, DWT contains
more useful information compared with DFT and time signal, thanks to its multi-resolution
analysis and the filter bank.
Q2. Which aggregation functions (norm, mean, skewness, etc.) used for the feature
representation improve the classification accuracy the most? if we look at the results
in DWT part we can note that interestingly majority of yellow regions belong to 𝐾 𝑢𝑟𝑡𝑜𝑖𝑠 >
𝑠𝑘𝑒𝑤𝑛𝑒𝑠𝑠 > 𝑀𝑒𝑎𝑛. These results are insightful and reflect the importance of 𝑛-th moment PDF
statistics.
Q3. Which is the best interaction between domains and extracted features? we can
note that the general observation is that the interaction of DWT with other classes (DWT, DFT,
and time) improve the classification accuracy. However, the interaction of DFT and time does
not provide much useful information to classification. Finally, we can note that the best results
are obtained for (DWT, skewness, d1, and d2) and (DWT, kurtosis, d1, and d2).
Table 3
Classification accuracy (%) using 48 (6+6+36) features and five classifiers. The first and second best
results are shown in Bold and Italic, respectively.
Domain Classifier FLP FTC
DT 67.5 88.1
SVM 59.1 90.5
Time KNN 58 86.2
Ensemble (BT) 72.1 88.3
Ensemble (K) 63.1 88.9
DT 59.1 82.4
SVM 59.3 91.6
Frequency
KNN 62.3 85.1
(DFT)
Ensemble (BT) 65 87.3
Ensemble (K) 59.9 83.2
DT 99.2 92.1
SVM 98.9 93
Wavelet
KNN 98.6 93
(DWT)
Ensemble (BT) 99.7 93.5
Ensemble (K) 99.7 84.8
DT 99.7 94.3
SVM 98.6 95.4
All KNN 97.3 92.4
Ensemble (BT) 99.5 94.9
Ensemble (K) 100 84.8
5. Conclusion and future work
In this paper, we have addressed the security threats on the electrical grid, representing one
of the self-healing features for smart grids. First, we created a large-scale dataset composed
of 1.8K samples by injecting faults to the IEEE-13 distribution network and collecting data;
second, we built a data-driven approach to perform fault location detection (FLD) and fault
type classification (FTC) automatically and accurately. Our proposed approach relies on a
suite of features extracted from time, frequency, and more importantly wavelet domain. It
also makes use of several state-of-the-art classification techniques to test and measure the
importance of different feature classes. We further explored the role of different aggression
functions for feature representation. Finally, a distinctive contribution of this paper is to provide
an interpretability analysis for the above classification task in which we shed light on how
interpretability can provide evidence about why certain decision makings were made by the
ML system. Results are promising and witness the merits of the the proposed system to tackle
security issues in SGs. In future, we plan to investigate on the efficacy of the proposed system
on larger-sized IEEE node test feeders. In addition, designing similar system for detection of
(deliberate) attacks such as the ones based on adversarial machine learning [15] — or adversarial
attacks — constitutes another interesting open research challenge for future.
a5
d1
d2
d3
Norm (DWT)
d4
d5
a5
d1
d2
Mean (DWT) d3
d4
d5
a5
d1
d2
std (DWT) d3
d4
d5
a5
d1
skewness d2
d3
(DWT) d4
d5
a5
d1
kurtosis d2
(DWT) d3
d4
d5
a5
d1
max (DWT) d2
d3
d4
d5
norm
mean
std
DFT skewness
kurtosis
max
norm
mean
Time std
skewness
kurtosis
max
Figure 2: Visualization of the interaction of different features on the FLP classification results measured
in terms of accuracy. In total, the impact of pairwise interaction of 48 features is shown in this figure.
Acknowledgments
This work has been partially funded by e-distribuzione S.p.A company, Italy, through a PhD
scholarship granted to Fatemeh Nazary.
The authors also acknowledge partial support of the projects: Servizi Locali 2.0, PON ARS01-
00876 Bio-D, PON ARS01-00821 FLET4.0, PON ARS01-00917 OK-INSAID, H2020 PASSPARTOUT.
References
[1] M. Z. Gunduz, R. Das, Cyber-security on smart grid: Threats and potential solutions,
Computer networks 169 (2020) 107094.
[2] D. Burpee, H. Dabaghi, L. Jackson, F. Kwamena, J. Richter, T. Rusnov, K. Friedman, L. Man-
sueti, D. Meyer, Us-canada power system outage task force: final report on the implemen-
tation of task force recommendations (2006).
[3] M. Majidi, M. Etezadi-Amoli, M. S. Fadali, A sparse-data-driven approach for fault location
in transmission networks, IEEE Trans. Smart Grid 8 (2017) 548–556. URL: https://doi.org/
10.1109/TSG.2015.2493545. doi:1 0 . 1 1 0 9 / T S G . 2 0 1 5 . 2 4 9 3 5 4 5 .
[4] S. Amini, F. Pasqualetti, M. Abbaszadeh, H. M. Rad, Hierarchical location identification of
destabilizing faults and attacks in power systems: A frequency-domain approach, IEEE
Trans. Smart Grid 10 (2019) 2036–2045. URL: https://doi.org/10.1109/TSG.2017.2787690.
doi:1 0 . 1 1 0 9 / T S G . 2 0 1 7 . 2 7 8 7 6 9 0 .
[5] J. J. Q. Yu, Y. Hou, A. Y. S. Lam, V. O. K. Li, Intelligent fault detection scheme for microgrids
with wavelet-based deep neural networks, IEEE Trans. Smart Grid 10 (2019) 1694–1703.
URL: https://doi.org/10.1109/TSG.2017.2776310. doi:1 0 . 1 1 0 9 / T S G . 2 0 1 7 . 2 7 7 6 3 1 0 .
[6] T. S. Abdelgayed, W. G. Morsi, T. S. Sidhu, A new harmony search approach for optimal
wavelets applied to fault classification, IEEE Trans. Smart Grid 9 (2018) 521–529. URL:
https://doi.org/10.1109/TSG.2016.2555141. doi:1 0 . 1 1 0 9 / T S G . 2 0 1 6 . 2 5 5 5 1 4 1 .
[7] V. Veerasamy, N. I. A. Wahab, R. Ramachandran, M. Thirumeni, C. Subramanian, M. L. Oth-
man, H. Hizam, High-impedance fault detection in medium-voltage distribution network
using computational intelligence-based classifiers, Neural Computing and Applications 31
(2019) 9127–9143.
[8] A. Spanos, Probability Theory and Statistical Inference: Empirical Modeling with Observa-
tional Data, Cambridge University Press, 2019.
[9] C. Ardito, Y. Deldjoo, E. D. Sciascio, F. Nazary, Interacting with features: Visual inspection
of black-box fault type classification systems in electrical grids, in: Proceedings of the
Italian Workshop on Explainable Artificial Intelligence co-located with 19th International
Conference of the Italian Association for Artificial Intelligence, XAI.it@AIxIA 2020, Online
Event, November 25-26, 2020, 2020, pp. 135–141. URL: http://ceur-ws.org/Vol-2742/short8.
pdf.
[10] F. E. Postigo Marcos, C. Mateo Domingo, T. Gomez San Roman, B. Palmintier, B.-M. Hodge,
V. Krishnan, F. de Cuadra García, B. Mather, A review of power distribution test feeders in
the united states and the need for synthetic representative networks, Energies 10 (2017)
1896.
[11] Q. Cui, Y. Weng, Enhance high impedance fault detection and location accuracy via -pmus,
IEEE Trans. Smart Grid 11 (2020) 797–809. URL: https://doi.org/10.1109/TSG.2019.2926668.
doi:1 0 . 1 1 0 9 / T S G . 2 0 1 9 . 2 9 2 6 6 6 8 .
[12] E. A. Reche, J. V. de 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 (2019) 3612–3619. URL: https://doi.org/10.1109/TSG.2018.2832840.
doi:1 0 . 1 1 0 9 / T S G . 2 0 1 8 . 2 8 3 2 8 4 0 .
[13] V. Veerasamy, N. I. A. Wahab, R. Ramachandran, M. Thirumeni, C. Subramanian,
M. L. Othman, H. Hizam, High-impedance fault detection in medium-voltage distri-
bution network using computational intelligence-based classifiers, Neural Comput.
Appl. 31 (2019) 9127–9143. URL: https://doi.org/10.1007/s00521-019-04445-w. doi:1 0 . 1 0 0 7 /
s00521- 019- 04445- w.
[14] S. Mallat, A theory for multiresolution signal decomposition: The wavelet representation,
IEEE Trans. Pattern Anal. Mach. Intell. 11 (1989) 674–693. URL: https://doi.org/10.1109/34.
192463. doi:1 0 . 1 1 0 9 / 3 4 . 1 9 2 4 6 3 .
[15] Y. Deldjoo, T. D. Noia, F. A. Merra, A survey on adversarial recommender systems: from
attack/defense strategies to generative adversarial networks, ACM Computing Surveys
(CSUR) 54 (2021) 1–38.