energies

Article
Deep Learning-Based Algorithm for Internal Fault
Detection of Power Transformers during Inrush Current
at Distribution Substations
Sopheap Key, Gyu-Won Son and Soon-Ryul Nam *

Department of Electrical Engineering, Myongji University, Yongin 17058, Republic of Korea;

keysopheap123@gmail.com (S.K.); onyang01111@gmail.com (G.-W.S.)
* Correspondence: ptsouth@mju.ac.kr; Tel.: +82-31-330-6361

Abstract: The reliability and stability of differential protection in power transformers could be
threatened by several types of inferences, including magnetizing inrush currents, current transformer
saturation, and overexcitation from external faults. The robustness of deep learning applications
employed for power system protection in recent years has offered solutions to deal with several
disturbances. This paper presents a method for detecting internal faults in power transformers
occurring simultaneously with inrush currents. It involves utilizing a data window (DW) and
stacked denoising autoencoders. Unlike the conventional method, the proposed scheme requires
no thresholds to discriminate internal faults and inrush currents. The performance of the algorithm
was verified using fault data from a typical Korean 154 kV distribution substation. Inrush current
variation and internal faults were simulated and generated in PSCAD/EMTDC, considering various
parameters that affect an inrush current. The results indicate that the proposed scheme can detect the
appearance of internal faults occurring simultaneously with an inrush current. Moreover, it shows
promising results compared to the prevailing methods, ensuring the superiority of the proposed
method. From sample N–3, the proposed DNN demonstrates accurate discrimination between
internal faults and inrush currents, achieving accuracy, sensitivity, and precision values of 100%.

Keywords: inrush current; data window; deep learning; differential protection; internal fault
Citation: Key, S.; Son, G.-W.; Nam,
S.-R. Deep Learning-Based Algorithm
for Internal Fault Detection of Power
Transformers during Inrush Current 1. Introduction
at Distribution Substations. Energies A power transformer is an essential component used in power systems where voltage
2024, 17, 963. https://doi.org/ conversion is required. To ensure efficient operation in power systems, current differential
10.3390/en17040963 protection is conventionally adopted as the primary protection, which is based on Kirchoff’s
Academic Editor: Ahmed Abu-Siada current law. However, it is susceptible to unwanted abnormalities such as magnetizing
inrush currents during transformer energization and a parallel connection of transformers
Received: 29 January 2024 under normal operations, as well as CT saturation due to overexcitation. These abnor-
Revised: 14 February 2024
malities might negatively result in the mis-operation of the current differential protection.
Accepted: 15 February 2024
An inrush current is a non-sinusoidal and high-magnitude current generated due to flux
Published: 19 February 2024
saturation in the transformer during energization. The magnitude of an inrush current is
highly dependent on the switching angle, the amount of residual flux, and the sizes of the
transformers. The fundamental principles and derivation of magnetizing an inrush current
Copyright: © 2024 by the authors.
are presented in [1].
Licensee MDPI, Basel, Switzerland.
Since magnetizing inrush currents generally has a large ratio of the second-harmonic
This article is an open access article component compared to an internal fault and normal conditions, harmonic blocking and
distributed under the terms and restraint have been designed to avoid false operations due to inrush currents [2] and have
conditions of the Creative Commons been widely employed in commercial relays [3]. Moreover, with the newly improved
Attribution (CC BY) license (https:// material of modern transformers, second-harmonic restraint/blocking faces the downside
creativecommons.org/licenses/by/ of lower second-harmonic components during transformer energization [4]. Therefore,
4.0/). the conventional scheme in transformer protection can be blocked for several cycles due

Energies 2024, 17, 963. https://doi.org/10.3390/en17040963 https://www.mdpi.com/journal/energies

Energies 2024, 17, 963 2 of 18

to an indecisive threshold. In addition, the energization of a faulty transformer may

reduce the sensitivity of harmonic restraint due to the high ratio of the second harmonic in
healthy phases and leads to low reliability during the energization of a faulty transformer.
Thus, novel functionalities must be proposed or integrated with the existing differential
protection to enhance reliability and security in detecting internal faults during energization.
Recently, the Korean Electric Power Company (KEPCO) reported numerous failures of
the differential relay in the field when internal faults occurred during an inrush current,
leading to a malfunction in the differential relay, as conventional harmonic blocking could
not respond to them and continuously blocked the differential relay until the presence of
the second harmonic fell under a set value. Therefore, a new scheme for power transformer
protection is urgently required to secure the stability of power systems.

1.1. Literature Review and Related Works

Conventionally, the utilization of the second-harmonic principle is widely adopted in
power transformer protection against inrush currents, as described in the above section.
However, this method has been proven to be ineffective in several circumstances [5,6].
During internal faults, there is a large ratio of the second harmonic in a few cycles, which
blocks the differential relay from operating, resulting in damage to power transformers. An
extensive outage and a blackout were reported in [7] when the power transformer protection
mis-operated under inrush conditions. Moreover, as the power system expands, the second-
harmonic components increase on long transmission lines when the transformers are
connected to shunt reactors or series capacitors [8]; as a result, differential protection is
bypassed when this scenario occurs.
Several transformer-protection techniques have been actively proposed to identify
the inrush condition, such as artificial neural networks, fuzzy logic, wavelet transform,
and mathematical-based algorithms. A statistical approach based on Principle Component
Analysis (PCA) was described in [9] to differentiate inrush currents, internal faults, and
overexcitation conditions. It captures 2D feature space as a pattern recognition for each
abnormal condition. Methods based on fuzzy and artificial neural networks were proposed
in [10,11], and a correlation-based algorithm was developed for inrush current discrimi-
nation [12]. For a similar purpose, a method combining a support vector machine as the
classifier and a wavelet transform for feature extraction was also proposed in [13]. A deep
learning application was proposed in [14,15] to address current transformer saturation on
transmission lines, and another deep learning-based approach was also proposed in [16] to
remove the decaying DC offset in a power system.
As signal processing techniques based on wavelet transforms have proven to be effi-
cient tools for the analysis, detection, and classification of non-stationary signals at various
levels of time–frequency resolution in the literature, they could be applicable in real-time
devices. For instance, a wavelet transform has been utilized to address existing issues in
power systems such as fault detection, location, and classification [17,18], as well as in the
differential protection of power transformers [19–23]. Although it has good performance
without the need of harmonic information, there are some limitations for practical appli-
cations in power system protection, such as the strong influence of the mother wavelet
and time delay. However, it does not provide an answer for internal fault detection during
inrush conditions, which is a significant concern in transformer differential protection. An
improved wavelet transformation, namely the Real-Time Boundary Stationary Wavelet
Transform (RT-BSWT), was proposed in [24] to detect internal faults during inrush currents.
Despite the improvement made, a high sampling rate is required, and it may be susceptible
to noise. A process to identify an inrush current based on the enhanced GSA-BP approach
was proposed in [25] to discriminate inrush currents from fault currents in transformers.
A low-computation method based on a fault component network was developed
in [26] to enhance the accuracy of transformer protection, regardless of magnetizing inrush
conditions. A method based on the current and voltage ratio was demonstrated in [27],
where it deployed the absolute difference of the current and voltage to differentiate inrush
Energies 2024, 17, 963 3 of 18

currents from internal faults. A unidirectional index was utilized to detect the direction of
magnetizing inrush currents in power transformers [28]. The detection of inrush currents
based on the dead angle was introduced in [29]. If the waveform distortion is so severe
that the wave width is less than 140◦ , it will cause a delay in protection or even a wrong
judgment; therefore, the efficacy of this method presents a drawback. A new adaptive
coordination approach between generator and transformer was proposed to enhance the
abnormal operating conditions [30].

1.2. Key Contributions and Organization

Motivated by the above-mentioned problems with the conventional approach, this
paper presents a protection scheme to discriminate internal faults and inrush currents
by combining a data window with deep neural networks (DNNs). In recent years, new
techniques based on intelligent methods have demonstrated a robust distinction between
inrush currents and internal faults for power transformer protection, overcoming the
drawbacks of traditional differential protection. To detect inrush currents and internal
faults, the proposed scheme first utilizes the data window to obtain the distinctive feature
signal that separates the region of internal faults and inrush currents. The proposed
scheme can identify internal faults during inrush currents. It not only provides stability
when these two abnormal conditions occur simultaneously but also improves the response
time compared to conventional harmonic-blocking methods. Furthermore, the proposed
scheme is applicable to inrush currents and internal faults of various magnitudes due
to its normalization quantity during the preprocessing prior to deep-learning training.
Then, a DNN is employed to discriminate internal faults from inrush currents. The key
contributions of the proposed work can be highlighted as follows.
1. A wide range of applicability, regardless of inrush current magnitude, the residual
flux in power transformers, internal fault magnitude, and fault angles;
2. An improved discrimination of internal faults, considering winding-ground faults
during inrush currents;
3. A universal application for other power transformers with different characteristics;
4. A data window-based operation without the need for a threshold.
The rest of this paper is organized as follows. Section 2 highlights the literature review
of the behavior of inrush currents using a data window and addresses issues related to
the second-harmonic-blocking method. This section also includes information on data
acquisition and dataset preparation for training, along with a detailed description of inrush
current features. Section 3 presents the proposed deep neural network (DNN) method and
its structure. The simulation setup, implemented in both Python and PSCAD/EMTDC,
is detailed in Section 4. Section 5 addresses the results of the proposed method for in-
rush current and internal fault detection and provides a comparative analysis using the
conventional approach. A discussion of the performance evaluation based on statistical
percentages is demonstrated in Section 6. Lastly, Section 7 includes concluding remarks
and information regarding potential future works.

2. Problem Statement
This section presents the principles and approaches utilized for internal fault and
inrush-current detection based on a data window. To facilitate understanding in the
subsequent sections, a list of relevant acronyms and their definitions are provided in Table 1.
Energies 2024, 17, x FOR PEER REVIEW 4 of 18

Energies 2024, 17, 963 Table 1. Relevant acronyms, units, and their definitions. 4 of 18

Acronym Unit Definition

DW
Table 1. Relevant acronyms, units, and their definitions.
Data window
DNN Deep neural network
AE Acronym Unit Definition
Autoencoder
SAE DW Stacked autoencoder
Data window
HAR DNN Deep neural
Second-harmonic network
restraint
UNI AE Autoencoder
Unidirectional index
SAE Stacked autoencoder
EKF HAR Extended Kalman filter
Second-harmonic restraint
idiff UNI A Differential currentindex
Unidirectional
m EKF Size of data window
Extended Kalman filter
k i diff A Differential
Last index of differential current
current
m Size of data window
xnorm A Normalized differential current
k Last index of differential current
θk xnorm A Parameter
Normalized setdifferential
model current
yi θk Stochastic variable of theset
Parameter output
modelclass
S yi Stochastic variable
Activation of the output class
function
S Activation function
LAE Reconstruction loss for AE
LAE Reconstruction loss for AE
L L Softmax loss function
Softmax loss function
𝜎 σ Weight decaydecay
Weight

2.1.
2.1.Overview
OverviewofofMagnetizing
MagnetizingInrush
InrushCurrent
Currentand
andSecond
SecondHarmonic
HarmonicRatio
Ratio
An
Aninrush
inrushcurrent
currentisisthe
thehigh
high current
currentdrawn
drawnby by
a transformer
a transformerwhen it is itinitially
when en-
is initially
ergized. It is caused by an abrupt change in magnetic flux within
energized. It is caused by an abrupt change in magnetic flux within the transformerthe transformer core and
iscore
proportional to the current
and is proportional flowing
to the currentthrough
flowingthethrough
primarythe winding.
primaryFigure 1 illustrates
winding. Figurea1
differential current and the ratio
illustrates a differential current and of the second harmonic. As mentioned in the
ratio of the second harmonic. As mentioned in the Introduc-
tion, a common
Introduction, approachapproach
a common to differential protection
to differential in powerintransformers
protection involvesinvolves
power transformers second
harmonic-based
second harmonic-based blockingblocking
to prevent unnecessary
to prevent tripping.tripping.
unnecessary Due to theDuesubstantial ratio of
to the substantial
the second
ratio of theharmonic during transformer
second harmonic energization,
during transformer it can beiteffectively
energization, used to used
can be effectively distin-
to
guish an inrush
distinguish current
an inrush fromfrom
current internal faults.
internal However,
faults. However,as shown
as shown in in
Figure
Figure1,1,aasecond
second
harmonic
harmonicmay mayalso
alsobebegenerated
generatedduring duringinternal
internalfaults
faultsdue
duetotoaadecaying
decayingDC DCcomponent
component
from
fromfaults.
faults.At Atthe
themoment
momentofoftransformer
transformerenergization,
energization,thetheratio
ratioof
ofthe
thesecond
secondharmonic
harmonic
rapidly increases
rapidly increases to to approximately
approximately 60%. 60%. Consequently,
Consequently, the the harmonic-blocking
harmonic-blocking method method
blocks the
blocks the operation
operation of the differential
differential relay
relayin inthis
thisscenario,
scenario,leading
leadingtotopotential
potentialdamage
damage to
tothe power
the power transformer.
transformer.

Figure
Figure1.1.Waveform
Waveformof
ofdifferential
differentialcurrent
currentpertaining
pertainingto
toinrush
inrushcurrent
current and
and internal
internal faults.
faults.

DW is a technique applied in power system protection for fault detection, direction

estimation, time-series forecasting, and fault classification. It yields promising results at
every instant when there is a significant fluctuation in the waveform. The results based
on the data window from [14–16] are noticeable when dealing with abnormal conditions.
Inspired by this concept, we develop a DW which was originally proposed in [31] to
detect power swings on a transmission line. Considering the measured differential current
 DW is a technique applied in power system protection for fault detection, direction
estimation, time-series forecasting, and fault classification. It yields promising results at
every instant when there is a significant fluctuation in the waveform. The results based
on the data window from [14–16] are noticeable when dealing with abnormal conditions.
Energies 2024, 17, 963 5 of 18
Inspired by this concept, we develop a DW which was originally proposed in [31] to detect
power swings on a transmission line. Considering the measured differential current
𝑖 x1 , …, xk , where 𝑘 is the last index of the measured differential current, the
di f f = {x
iequation derived xk }, where
1 , . . . , from a set ofk DWs
is theon
last index
the of the differential
measured measured differential current, the
current is expressed in
equation derived
(1) as follows: from a set of DWs on the measured differential current is expressed in (1)
as follows:
x 1 x1 x2x2 .…. . xxmm-1−1 xm xm
 
 x 2 x2 x3x3 .…. . xxmm xm+1 x m +1 
Set of DWs (1)
Set of DWs =  . ⋮ ⋮ ⋱ ⋮ ⋮ (1)
 
. .. . . 
 .. x x.. … . x .. x .. 
k-m+1 k-m+2 k- 1 k
x k − m +1 x k − m +2 ... x k −1 xk
2.2. Data Window of Inrush Current and Internal Faults
2.2. Data Window of Inrush Current and Internal Faults
The waveform in the DW, as described in (1), forms an abundance of distinctive
The waveform in the DW, as described in (1), forms an abundance of distinctive
waveform
waveform characteristics
characteristics at ateach
eachsample
samplepoint. point.These
Thesecharacteristics
characteristics enable
enable DNNDNN to cap-
to capture
ture the unique features distinguishing inrush currents from internal
the unique features distinguishing inrush currents from internal faults. Figure 2 illustrates faults. Figure 2 illus-
trates
a regiona region of instantaneous
of instantaneous differential
differential currentscurrents
under a DW under a DW
with withofa one
a length length
cycle.of This
one
cycle. This figure clearly demonstrates the DW under conditions
figure clearly demonstrates the DW under conditions of an inrush current and an internal of an inrush current and
an internal
fault. Prior fault. Prior to
to a sudden a sudden
spike spike indue
in the current theto current due toenergization,
transformer transformer every energization,
value in
every value in the DW is zero in each sample. Upon closing
the DW is zero in each sample. Upon closing the circuit breaker, there is a sudden the circuit breaker, there is a
change
sudden change in
in the magnetic theleading
flux, magnetic to aflux, leadingincrease
significant to a significant increase in
in the differential the differential
current. Similarly,
current.
an internalSimilarly, an internal
fault also manifests fault also manifests
a sudden changeaat sudden change
the initial at the
point initial
in the point in
differential
the differential current, posing a challenge for conventional methods
current, posing a challenge for conventional methods to discriminate internal faults from to discriminate in-
ternal faults from inrush currents when both conditions coincide.
inrush currents when both conditions coincide. As illustrated in Figure 2, every value in As illustrated in Figure
2,
theevery
DW value
beforein the DW
point beforedesignating
A is zero, point A is this zero,region
designating this region
as the normal as the (state
condition normal 0).
condition
Upon reaching(state point
0). Upon reaching
A, the value of point A, the
the last value
index of the
of the DW last index of
becomes the DWindicating
positive, becomes
positive, indicating
the occurrence of athe occurrence
transient stateof(state
a transient
1). If state (state 1).
an internal If an
fault internal
and fault current
an inrush and an
inrush current occur simultaneously
occur simultaneously at this point,
at this point, it becomes it becomes
challenging challenging
to determine eachtodisturbance.
determine
each disturbance.
Therefore, no actionTherefore,
will beno action
taken will be
during taken
this during this
transition. transition.
At point B, anAt point B,
internal an
fault
internal fault exhibits different characteristics from an inrush
exhibits different characteristics from an inrush current. For internal faults, the valuecurrent. For internal faults,
the value zero
becomes becomes
for a zero
fault for a faultangle
inception inceptionof 0◦ angle of 0° or
or negative fornegative for a fault
a fault inception of 90◦ .
inception
angle
angle
Whenof 90°.
this When this
behavior behaviorthe
is detected, is algorithm
detected, the algorithm
promptly changespromptly
to state changes to state
3 (internal 3
fault);
(internal
otherwise, fault); otherwise,
it identifies it identifies
an inrush current an inrush current
(state 2). (state 2). delay
The sampling The sampling
betweendelay points be-A
tween
and B ispoints A and
less than oneB cycle,
is lessspecifically
than one cycle, specifically
58 samples, 58 samples,
considering considering
that one that one
cycle corresponds
to 64 corresponds
cycle samples. Theto reason for thisThe
64 samples. delay is to for
reason achieve a clearisdiscrimination
this delay to achieve a clear between inrush
discrimina-
currents and internal faults at a fault inception angle of 0 ◦.
tion between inrush currents and internal faults at a fault inception angle of 0°.

Figure 2. Illustration of a DW of a differential current under the conditions of an inrush current

Figure 2. Illustration of a DW of a differential current under the conditions of an inrush current
(upper) and an
(upper) and an internal
internal fault
fault (below).
(below).

Therefore, it is apparent that this difference can be effectively used as an important

feature to distinguish internal faults from inrush currents and to create learning labels for
DNNs, which will be explained later in the following section.

2.3. Dataset Acquisition for Training and Testing Procedure

A thorough analysis is necessary to achieve high accuracy and generalization in the
discrimination model for inrush currents and internal faults using a DNN-based method.
The generation of sufficient datasets for training DNNs is crucial for accurately discriminat-
Energies 2024, 17, 963 6 of 18

ing between the mentioned abnormalities. To obtain diverse datasets for inrush currents,
extensive simulations are required for subsequent analysis. The training dataset consid-
ers influencing parameters in inrush conditions, such as the residual flux in the power
transformer, the switching angle, and the polarity of the residual flux. The inrush current
magnitude is at its maximum when the transformer switches on at 0◦ . Moreover, the
polarity of the residual flux significantly impacts the magnitude of the inrush current.
The influencing parameters for inrush currents and internal faults are listed in Table 2.
The datasets for inrush currents accumulated 170 inrush conditions, corresponding to
228,140 datasets available for training and testing. The influencing parameters for the inter-
nal faults are given in Table 2. The datasets for the internal faults accumulated 90 cases of
a–g faults, corresponding to 111,870 datasets available for training and testing. The inrush
current and internal faults are randomly partitioned into training and testing datasets with
an 80% to 20% ratio, respectively.

Table 2. Dataset of the inrush current and internal faults for the DNN procedure.

Case Parameters Value

Switching angle (◦ ) 0, 10, 20, 30, 40, 50, 60, 70, 80, 90
Inrush −80, −70, −60, −50, −40, −30, −20,
Residual flux (%)
−10, 0, 10, 20, 30, 40, 50, 60, 70, 80

Internal Fault inception angle (◦ ) 0, 10, 20, 30, 40, 50, 60, 70, 80, 90
fault Winding location (%) 0, 10, 20, 30, 40, 50, 60, 70, 80

2.4. Dataset Preprocessing for Training

The preprocessing stage for training DNNs is the most crucial part, determining the
outcome of the trained model. It serves as a platform for DNNs to quickly comprehend the
problem statement and the approach to achieving the expected outcomes in the final stage.
As the magnitude of an inrush current varies depending on the influencing parameters
listed in Table 2, it is challenging to determine a specific threshold for the correct label
for DNNs. Therefore, normalization is introduced to address the problem of numerical
instability and uncertain thresholds caused by the varying magnitudes of an inrush current.
The derived equation for normalizing the training input is given in (2) as follows.
xi
xnorm = (2)
xmax

where xmax is the maximum value captured in the measured differential current. Normaliz-
ing the input dataset scales the training input within the range of [–1, 1]. Additionally, this
process enhances the robustness and capability of the proposed DNN, making it applicable
to datasets from different systems. Once a set of DWs are formed, as described in (1), and
the label for each condition is defined, as described in Section 2.2, we convert the multi-class
region into a binary form using one-hot encoding, as shown in Table 3.

Table 3. Binary form using one-hot encoding.

Class Label Binary Form

0 Normal condition 0001
1 Transient 0010
2 Inrush current 0100
3 Internal fault 1000

3. Deep Neural Network (DNN)-Based Discrimination

DNNs have undergone continuous evolution, demonstrating a strong capability to
address challenging problems in recent years, particularly in cases where conventional
Energies 2024, 17, 963 7 of 18

methods struggle with nonlinear issues. This section introduces the concepts and strategies
implemented to discriminate inrush currents and internal faults. To enhance the structure
of DNNs, the proposed discrimination scheme adopts unsupervised pre-training using
stacked autoencoders and supervised fine-tuning. The details of benchmark models are
demonstrated well in [32,33].

3.1. Principle of Autoencoders

An autoencoder is the basic component of an SAE learning in an unsupervised
way, typically containing an encoder and a decoder. In a simple autoencoder, the input
x ∈ Rn (x1 , x2 , . . ., xn ) is included in the training dataset. The input is then encoded to a
low dimension and restored to its original dimension in the decoding part. The training
uses backpropagation to minimize the reconstruction error of the input features. Once the
training converges, the transformed features (f 1 , f 2 , . . ., fn ) are saved and used to train other
autoencoders. The encoder employs a deterministic mapping function to map input x to
the hidden layer f. The encoding process is given as follows in (3), where W 1 and b1 are the
weight and bias of the encoding parts.

f = S(W1 x + b1 ) (3)

The decoder reconstructs the hidden layer representation (f ) to obtain the output
(x̂), as shown in (4), where W 2 and b2 are the weight and bias of the decoding parts. S
denotes the activation function for training the AE, and ReLU is used for both the encoder
and decoder.
^
x = S(W2 f + b2 ) (4)
The parameters of the AE are optimized to minimize the reconstruction error, as shown
in (5).
1 N
 
^
N n∑
LAE x, x = ( x − x̂θ ( f θ ( x̂ ))) (5)
=1

3.2. Framework of Stacked Autoencoder

An SAE is a neural network consisting of multiple layers of AEs, where the features
of each AE are stacked and fed as inputs to the successive AE. The first AE is trained in a
bottleneck fashion with the initial weight and bias (w1 and b1 ). The input is compressed
into a low-dimensional feature through the encoding function and then restored back to
its original dimension in the decoding layer. After removing the decoding layer (x̂) in the
first AE, a new hidden layer (h2 ) and an output (ĥ1 ) are stacked onto the first AE. Using
a similar process, many AEs are successively stacked together to form a deeper network
structure. This process is commonly known as pre-training because it adopts a greedy-layer
training method. Finally, an output layer is trained with the given label (binary form of
the abnormality) to discriminate between inrush currents and internal faults. All optimal
SAE weights and biases (wi and bi , where i = 1, 2, . . ., n), which are obtained during the pre-
training process, are fine-tuned using the backpropagation algorithm to achieve significant
improvements in discrimination ability. The construction process of a three-layer SAE is
depicted in Figure 3.
 pre-training process, are fine-tuned using the backpropagation algorithm to achieve sig
Energies 2024, 17, 963 nificant improvements in discrimination ability. The construction process of8aofthree-laye
18
SAE is depicted in Figure 3.

Figure 3. Construction process of a three-layer SAE used in simulations.

Figure 3. Construction process of a three-layer SAE used in simulations.
3.3. Fine-Tuning and SoftMax Classifier
3.3. Fine-Tuning
The pre-trainedandmodel
SoftMax fromClassifier
the SAE can be further optimized by using parameters
from all encoding layers during
The pre-trained model from the pre-training
the SAE can phasebewith the backpropagation
further optimized by usingalgorithm
parameters
to minimize
from errors. Using
all encoding layersthe weights
during the(Wpre-training
i ) and biases phase
(bi ) from an SAE
with as initial values
the backpropagation algo
for fine-tuning enables the deeper network to generalize more effectively
rithm to minimize errors. Using the weights (Wi) and biases (bi) from an SAE as initia to other inrush
variations produced by power transformers. We assign labels and extract features from the
values for fine-tuning enables the deeper network to generalize more effectively to othe
SAE to the classifier layer for the precise discrimination between normal conditions, inrush
inrush variations
currents, produced
and internal by power transformers.
faults. Consequently, We assign
the classification labels
outputs andminimal
exhibit extract features
from the SAE to the classifier
errors, resulting in high accuracy. layer for the precise discrimination between normal condi
tions,A inrush
SoftMaxcurrents,
classifier and internalinfaults.
is employed Consequently,
the classifier the classification
layer to discriminate among theoutputs
four exhibi
minimal
classes errors,
listed resulting
in Table in highthe
3. It estimates accuracy.
posterior probabilities of each class in the range of
[0,1], and the hypothesis
A SoftMax is calculated
classifier is employed as follows.
in the classifier layer to discriminate among the fou
classes listed in Table 3. It 
estimates the
posterior probabilities
 T i  of each class in the range
of [0–1], and the hypothesis is
i = 1| x i , θ 
p ycalculated as follows. e θ1 x
i i

 T i
 p y = 2| x , θ  1  e θ2 x 
hθ ( xi ) = .. p yi =1|x  =,θ  . θT xi (6)
  i
.⎡  ..⎡e T i ⎤
⎤kl=1 eθl xi 
T 1
  ∑
⎢ ii ⎥ 1 θ T⎢xei θ2 x ⎥
)= kp| x y, θ=2|x ,θ =
i

i
hpθy(xi=
⎢ ⎥ ek (6
θT x i ⎢ ⎥
⎢ i ⋮ i ⎥ ∑l=1 e l ⎢ ⋮T ⎥
k

where yi is the stochastic variable of ⎣the =k|x ,θclass

p youtput ⎦ corresponding
⎣eθk xi ⎦ to input dataset
xi , and j represents the output class, encompassing four conditions: normal, transient,
where yi is the stochastic variable of the
output T class corresponding to input dataset xi
inrush, and internal fault. θ = θ1T , θ2T , . . . , θkT denotes the parameter set of the model.
and j represents the output class, encompassing four conditions: normal, transient, inrush
Consequently, the output of the SoftMax classifier is given in a 4-dimensional vector
and internal fault. 𝜃 𝜃 ,𝜃 ,…,𝜃 denotes the parameter set of the model. Conse
containing four possible classes. The maximum probability of each class is determined
quently,
as follows. the output of the SoftMax classifier is given in a 4-dimensional vector containing
four possible classes. The maximum
i probability i of each
i class is determined as follows.
Classx = argmax j=1,...,k p ( y = j | x , θ ) (7)
Class(x
Likewise, the SoftMax classifier i) = argmax
converges to the global
j=1,…,k
i
=j xi ,θ) by iteratively opti-
p(yminimum (7
mizing the cost function in (8) using categorical cross entropy.
Likewise, the SoftMax
 classifier converges to the  global minimum by iteratively op
T xi
timizing the cost1function ink (8) using categorical e j cross
θ entropy.
 + σ ∑k ∑ N θij2
n o
m
L = − ∑i=1 ∑ j=1 S yi = j log T xi i =1 j =1
(8)
N ∑ k
e l θTxi 2
θ
1 l =1 e j σ
L= ∑m
i=1 ∑j=1 S y =j log
k i
T i + ∑ki=1 ∑nj=1 θ2ij (8
N ∑kl=1 eθl x 2
Energies 2024, 17, x FOR PEER REVIEW 9 of 18
Energies 2024, 17, 963 9 of 18

where𝑦yi is
where is the
the iith scalar
scalar value
valuefrom
from the
the SoftMax
SoftMaxoutput
outputin (7),𝑆Srepresents
in(7), representsthe
theindicator
th
indicator
function, 𝜎 is included in the cost function to penalize large values of the parameters,
function, σ is included in the cost function to penalize large values of the parameters, and andL
Lisisstrictly
strictlyconvex.
convex.
AAflowchart
flowchartofofthe theproposed
proposedDNN
DNNisisdepicted
depictedin
inFigure
Figure4.4.

Figure 4. Flowchart of the proposed DNN to discriminate between inrush currents and internal faults.
Figure 4. Flowchart of the proposed DNN to discriminate between inrush currents and internal
faults.
4. Simulation Model
This section will highlight the simulation setup used to generate datasets for verifying
4. Simulation Model
the proposed DNN.
This section will highlight the simulation setup used to generate datasets for verify-
4.1.
ing thePSCAD/EMTDC
proposed DNN. Model
The effectiveness of the proposed technique was verified using a typical Korean
4.1.
154PSCAD/EMTDC
kV distribution Model
substation. A simulation model of a 154/23 kV distribution system
withThe
40 MVA power transformer
effectiveness of the proposedand technique
a Y-Y configuration
was verified was builtain
using PSCAD/EMTDC,
typical Korean 154
kVas distribution
illustrated insubstation.
Figure 5. A The samplingmodel
simulation frequency was set
of a 154/23 kVtodistribution
3840 Hz orsystem 64 samples
with per
40
cyclepower
MVA in 60 Hz systems. The
transformer andsource
a Y-Ywas defined bywas
configuration the specific
built in parameters
PSCAD/EMTDC, listed in
asTable
illus-4.
In thisinstudy,
trated Figure only winding-ground
5. The faults were
sampling frequency considered
was set to 3840 Hzfor or
evaluation,
64 samples with
pervariations
cycle in
60 Hz systems. The source was defined by the specific parameters listed in Table 4. Infaults
in the fault inception angles and percentages of the winding faults. The winding this
were simulated
study, by varying the
only winding-ground faultwere
faults location in the transformer,
considered winding
for evaluation, between 10%
with variations and
in the
90%, in steps of 20%, from the winding terminal on the primary side
fault inception angles and percentages of the winding faults. The winding faults were of the transformer.
The fault by
simulated inception
varyingangle of the
the fault internal
location faults
in the varied from
transformer, 0◦ to 90
winding
◦ , in steps of 15◦ , with
between 10% and 90%,
inreference
steps of to the from
20%, phase-A
the current.
windingDuring
terminal theongeneration
the primary of magnetizing inrush currents,
side of the transformer. Thea
residual flux was considered in the range of − 80% to 80%, in steps of 10%,
fault inception angle of the internal faults varied from 0° to 90°, in steps of 15°, with refer- and different
switching
ence to theinstances
phase-A were considered
current. During between 0◦ and 90
the generation
◦.
of magnetizing inrush currents, a
Energies 2024, 17, x FOR PEER REVIEW 10 of 18

Energies 2024, 17, 963 residual flux was considered in the range of −80% to 80%, in steps of 10%, and different
10 of 18
switching instances were considered between 0° and 90°.

Figure 5. Single line of PSCAD modelling.

Figure 5. Single line of PSCAD modelling.
Table 4. Source and transformer parameters used in PSCAD modelling.
Table 4. Source and transformer parameters used in PSCAD modelling.
Specification Parameters Value
Specification Parameters
R1 , R2 Value0.0419
Positive and
L
R1, R2 1 , L2 0.8921
0.0419
Negative
Positive and C1 , C2 0.0128
Source L 1, L 2 0.8921
Negative R0 0.0293
Zero C1, C2L0 0.0128
2.6657
Source
R0 C0 0.0293
0.0042
Positive leakage reactance
Zero L0 %Z 2.6657
10.99
Transformer 154/23 kV Air core reactance %X 20
Magnetizing current C0 %Im 0.00421
Positive leakage reactance %Z 10.99
Transformer
Air core reactance
4.2. Deep Neural Network Model %X 20
154/23 kV
Magnetizing current %Im 1 Google.
Tensorflow is one of the most common deep learning platforms developed by
It offers a high-level API to optimize neural network models and the training procedure of
4.2.
the Deep Neural
proposed Network
DNN Model
model. Therefore, the Tensorflow library is adopted in this paper to
construct the network
Tensorflow is onemodel
of theand to train
most it to discriminate
common deep learning between inrush
platforms currents and
developed by
internal Itfaults.
Google. offers a high-level API to optimize neural network models and the training pro-
cedure In of
boththethe unsupervised
proposed and supervised
DNN model. Therefore, learning modes, alibrary
the Tensorflow categorical cross-entropy
is adopted in this
loss was
paper employedthe
to construct to network
quantifymodel
the error
andbetween
to train the
it tonetwork output
discriminate and the
between reference
inrush cur-
output. The Adam optimizer
rents and internal faults. was used to build the network for gradient backpropagation
and parameter
In both the updates in every
unsupervised andepoch. A decaying
supervised learning learning
modes,rate was applied
a categorical to enhance
cross-entropy
convergence performance and to expedite the training process, preventing
loss was employed to quantify the error between the network output and the reference issues related
to overfitting. It was initiallywas at 8 ×to10
set used 3 and then exponentially decreased with each
output. The Adam optimizer build the network for gradient backpropagation
and parameter updates in every epoch. A decaying learning rate was applied toare
iteration. The structure of the DNN and the training parameters for each AE given
enhance
in Table 5.
convergence performance and to expedite the training process, preventing issues related
3
to overfitting.
Table It was
5. Structure of theinitially
proposedsetDNN
at 8×10 and then
and training exponentially decreased with each it-
parameters.
eration. The structure of the DNN and the training parameters for each AE are given in
Table 5. AE1 AE2 AE3 SoftMax Layer
Neuron 30 18 9 4
TableBatch
5. Structure
size of the proposed
128 DNN and training
64 parameters. 64 64
Learning rate 0.001 0.0024 0.0019 0.0159
AE1 AE2 AE3 SoftMax Layer
Neuron 30 18 9 4
5. Simulation Results
Batch size 128 64 64 64
In this section, the efficiency of the proposed DNN is verified and compared to the
Learning rate 0.001 0.0024 0.0019 0.0159
unidirectional index method in [28], the conventional harmonic-blocking scheme [34],
and the Extended Kalman filter in [35]. Graphical illustration and evaluation metrics
5. Simulation Results
make it abundantly evident that the proposed method is effective against inrush currents
and In this section,
internal faults. the
In efficiency of the
Figures 6–11, proposed
DNN, UNI, DNNand HAR is verified
denoteand thecompared
proposed toDNNthe
unidirectional index method in [28], the conventional harmonic-blocking
method, the unidirectional index in [28], and the second-harmonic-blocking approach [34], scheme [34], and
the ExtendedThe
respectively. Kalman filterKalman
Extended in [35]. filter
Graphical
in [35]illustration and evaluation
is used for comparison when metrics make
internal it
faults
abundantly evident EKF
are present, because that only
the proposed
detects the method
instance is of
effective
internalagainst inrush
faults. It currents
is generally and
known
that protection relays in power system protections operate after one cycle. Therefore, the
evaluation of the proposed DNN and alternative methods will be discussed based on the
58th (=N–6) and 61st (=N–3) samples from each abnormality.
 monic sharply increased at the closing instance, the HAR was theoretically effective in
monic sharply increased at the closing instance, the HAR was theoretically effective in
quickly detecting the inrush current. The UNI detected the inrush current after a timing
quickly detecting the inrush current. The UNI detected the inrush current after a timing
delay due to the data window, while the proposed DNN detected it at 0.231 s, with a
delay due to the data window, while the proposed DNN detected it at 0.231 s, with a
slightly quicker response than the DNN reference. Based on Figure 6, it is evident that the
slightly quicker response than the DNN reference. Based on Figure 6, it is evident that the
Energies 2024, 17, 963
proposed DNN presented a promising output in noticing the inrush current after sample
proposed DNN presented a promising output in noticing the inrush current after 11 sample
of 18
N–6, which was comparable to the HAR and UNI.
N–6, which was comparable to the HAR and UNI.

Energies 2024, 17, x FOR PEER REVIEW 12 of 18

The performance of the proposed DNN was also evaluated considering transformer
energization with the maximum residual flux, which was approximately 80%. The
amount of residual flux heavily influenced the magnitude of the inrush current; as a result,
the magnitude of the inrush current nearly doubled in this case, as demonstrated in Figure
7. It can be seen that the HAR yielded the best output among the three approaches in this
case. Considering a time delay, the UNI responded to the inrush current at 0.234 s,
Figure
Figure 6.6. Results
Results of
of inrush-current
inrush-current detectionin
detection inaacase
casewith
with noresidual
residualflux
fluxand
andatata aswitching
switchingangle
angle
whereas
Figure 6. the DNN
Results demonstrated
inrush-current a quicker
detection case withno
in adetection instance
no residualthan
flux the
and UNI. For instance,
at a switching angle
of 0°.
of
the0◦DNN
of 0°.. detected inrush currents faster (one sample) and more accurately than the UNI.

5.2. Case Study 2: Inrush Current at a Switching Angle of 90°

Switching a power transformer at 90° with no residual flux does not impact the op-
eration of conventional differential relays and produces the least inrush currents. How-
ever, the maximum flux in the power transformer strongly influences the nonlinear nature
of the magnetizing inrush current, as depicted in Figure 8. The magnitude of the inrush
current in this case is similar to that depicted in Figure 6. Therefore, the detection of the
inrush current was examined at the maximum switching angle and with residual flux. As
displayed in Figure 8, the HAR showed the most promising outcome, as it reacted to the
first instance of an inrush current due to the presence of the second harmonic ratio. Due
to the data window used in the UNI and DNN, their detections showed a timing delay of
less than 1 cycle. In particularly, the DNN yielded a more promising outcome than the
UNI, as it was 8 samples quicker. That is, the DNN faultlessly detected the inrush current
Figure
Figure 7. Results of
of inrush-current
inrush-currentdetection
detectioninina case
a case with
with maximum
maximum residual
residual fluxflux
andand
at aat a switch-
switching
Figure
after 7. Results
the ◦61st
of inrush-current detection in a case with maximum residual flux and at a switch-
ing angle of
angle of 0 . 0°. (=N–3) sample from the switching instance.
ing angle of 0°.

8. Results
Figure 8. Resultsof ofinrush-current
inrush-currentdetection
detectioninin
a case with
a case maximum
with residual
maximum fluxflux
residual andand
at a at
switching
a switch-
ing of 90of◦ . 90°.
angle
angle

5.3. Case Study 3: Energization of a Power Transformer in the Presence of an Internal Fault
Energizing a power transformer in the presence of an internal fault is a challenging
task for conventional protections, as the ratio of the second harmonic may cause the dif-
ferential relay to be blocked, potentially leading to severe damage to the power trans-
former. In this case, we consider a–g faults for internal faults. Figure 9 shows the results
of internal-fault detection when a power transformer was energized in the presence of an
internal fault. The evaluation was conducted in two different scenarios at fault inception
angles of 0° and 90°.
As shown in Figure 9a, the conventional HAR method detected the inrush current
rather than the internal fault due to the presence of the second harmonic in the decaying
DC component generated during the internal fault. Consequently, it prevented the inter-
nal fault from being detected, resulting in the blocking of the differential relay operation.
In contrast, the UNI detected the differential current as an inrush current instead of an
 to internal faults, even though the HAR and UNI failed to detect them. As shown in Figure
9a, for the fault inception angle of 90°, the HAR failed to detect the internal fault for several
cycles, highlighting a drawback of using HARs in modern transformers. In contrast, the
proposed DNN successfully detected the internal fault, starting from just one sample later
than the DNN reference. Similarly, as illustrated in Figure 9b, the DNN exhibited a prom-
Energies 2024, 17, 963 12 of 18
ising output in discriminating between inrush currents and internal faults at a fault incep-
tion angle of 0°.

(a)

Energies 2024, 17, x FOR PEER REVIEW 14 of 18

methods, as it did not respond to the internal fault in this case. UNI is only applicable
when there is a direction of the waveform on the positive or negative side, as its bidirec-
tional index makes it vulnerable to internal faults. The proposed DNN could detect the
internal fault with a time delay of less than one cycle from the fault inception. The evalu-
ation was performed on internal faults at fault inception angles of 0° and 90°, as illustrated
in Figure 10a,b, respectively. The results show that the proposed DNN can detect internal
faults after a time delay of less than one cycle, regardless of the fault inception angle.
The influence of external faults on the proposed DNN can be ignored since the dif-
(b)
ferential current will be zero during an external fault. Therefore, the DNN bypasses exter-
Figure 9. Results of internal-fault detection when a power transformer is energized in the presence
nal faults and allows
9. Results relevant protection
of internal-fault detectionschemes
when outside
a power the protection iszone
transformer to operate
Figure
of an internal fault: (a) fault inception angle of 0° and (b) fault inception angle of 90°.energized in the presence of
based on disturbance criteria. ◦ ◦
an internal fault: (a) fault inception angle of 0 and (b) fault inception angle of 90 .
5.4. Case Study 4: Phase-A-to-Ground Internal Faults Occurring during the Energization of a
Power Transformer
The proposed DNN was validated during an internal fault occurring a few cycles
after the switching of a power transformer. The harmonic-blocking scheme blocked the
operation of the differential relay due to the large ratio of the second harmonic at the onset
of an internal fault. This could lead to damage to the power transformer and should be
avoided.
A power transformer was switched on for energization at 0.22 s, and the internal fault
occurred at 0.32 s, as demonstrated in Figure 10. With the interference of the internal fault,
the HAR showed unsatisfactory results as soon as the internal fault occurred. The HAR
blocked the differential relay from operating for around two cycles, which could nega-
tively affect the power transformer. The UNI showed the worst results among the three

(a)

(b)
Figure 10. Results of the detection of phase-A-to-ground internal faults occurring during the ener-
Figure 10. Results of the detection of phase-A-to-ground internal faults occurring during the en-
gization of a power transformer: (a) fault inception angle of 0° and (b) fault inception angle of 90°.
ergization of a power transformer: (a) fault inception angle of 0◦ and (b) fault inception angle
of ◦.
90Case
5.5. Study 5: Phase-B–C-to-Ground Internal Faults Occurring during the Energization of a
Power Transformer
To demonstrate the capability of the proposed DNN across different fault types,
phase-B–C-to-ground internal faults are considered in this case. Figure 11 presents a case
of a phase-B–C-to-ground internal fault at a different time node considering a fault incep-
tion angle of 0°. The internal fault depicted in Figure 11 occurs three cycles after the inrush
 fault for several cycles. The EKF exhibits low sensitivity to the internal fault because the
estimated current from the EKF produces noise. Unlike these three methods, the proposed
Energies 2024, 17, 963 DNN demonstrates an accurate and reliable output in discriminating internal faults 13 with
of 18
a given time delay.

Figure11.
Figure 11. Results
Results of
ofthe
thedetection
detectionof
ofaaphase-B–C-to-ground
phase-B–C-to-groundinternal
internalfaults
faultsoccurring
occurringduring
duringthe
the
energization
energizationof
ofaapower
powertransformer.
transformer.

5.1. Case Study 1: Inrush Current at a Switching Angle of 0◦

6. Discussion on the Performance Evaluation Metrics
Magnetizing inrush currents are generated due to the remanent magnetism and no-
To effectively evaluate the performance of the proposed DNN, three indicators were
load closing of a power transformer. The closing instance significantly influences the
selected as evaluation metrics: accuracy, sensitivity, and precision. Traditionally, accuracy
waveform characteristics of the inrush current, while its remanent magnetism mainly
alone is insufficient to determine whether the proposed DNN yields a promising outcome.
affects its amplitude.
To visualize the stability of the proposed DNN method, a confusion matrix was used,
Transformer energization cases without and with residual flux are studied in this
section. Figure the
summarizing classification
6 shows the results performance and providing
when a transformer a visual
without residual representation of the
flux was energized
actual
at and predicted
a switching instanceclasses. The evaluation to
of 0◦ , corresponding matrix
0.2 s.was assessed
As the using
ratio of the the following
second four
harmonic
performance indices: TP (true positive), TN (true negative), FP (false
sharply increased at the closing instance, the HAR was theoretically effective in quicklypositive), and FN
(false negative).
detecting the inrush current. The UNI detected the inrush current after a timing delay due
to the data window, while the proposed DNN detected TP +TF it at 0.231 s, with a slightly quicker
response than the DNN reference.ACC Based = on Figure 6, it is evident that the proposed DNN (9)
(TP + TF + FP + FN)
presented a promising output in noticing the inrush current after sample N–6, which was
comparable to the HAR and UNI. TP
SEN = (10)
The performance of the proposed DNN was (TP +also
FN)evaluated considering transformer
energization with the maximum residual flux, which was approximately 80%. The amount
of residual flux heavily influenced thePRE magnitude TPof the inrush current; as a result, the
= (11)
magnitude of the inrush current nearly doubled (TPin+ this
FP) case, as demonstrated in Figure 7.
It canConventionally,
be seen that theaccuracy
HAR yielded (ACC)the bestthe
shows output among the
authenticity of athree approaches
detection method,indefin-
this
case. Considering a time delay, the UNI responded to the inrush current at 0.234
ing the correct detections over the total numbers of detections, including correct and false s, whereas
the DNN
ones. demonstrated
Sensitivity (SEN) ameasures
quicker detection instance
the proportion ofthan theand
inrush UNI.internal
For instance,
faults the
thatDNN
were
detected
correctlyinrush currents
identified among faster
the (one
actualsample) and
labels. It is amore accurately
crucial metric inthan the UNI. because
discrimination,
it influences the decision to allow the differential relay to operate when an internal fault
5.2. Case Study 2: Inrush Current at a Switching Angle of 90◦
occurs during inrush currents. A high percentage of SENs is essential to determine the
Switching
stability of theaproposed
power transformer at 90◦ with
DNN. Precision (PRE) noisresidual
anotherflux does notmetric
important impactrequired
the opera-
to
tion of conventional differential relays and produces the least inrush currents.
affirm the correctness of the proposed DNN. For instance, it demonstrates the capability However,
the maximum
of the proposedflux in the
DNN topower
isolatetransformer strongly
internal faults influences
from inrush the nonlinear
currents when both nature of the
abnormal-
magnetizing inrush current, In
ities occur simultaneously. as depicted in Figure
other words, 8. The magnitude
it demonstrates of the
the ability of inrush currentde-
internal-fault in
this case is similar to that depicted in Figure 6. Therefore, the detection of the inrush current
tection without mistakenly identifying it as an inrush current. A comparative analysis was
was examined at the maximum switching angle and with residual flux. As displayed in
conducted, and the evaluation metrics are presented in Table 6. The effectiveness of these
Figure 8, the HAR showed the most promising outcome, as it reacted to the first instance
metrics was assessed at the 58th and 61st samples from the beginning of each abnormality.
of an inrush current due to the presence of the second harmonic ratio. Due to the data
In cases where a power transformer is energized in the presence of an internal fault,
window used in the UNI and DNN, their detections showed a timing delay of less than
the aim is to avoid a situation where the DNN mistakenly detects it as an inrush current
1 cycle. In particularly, the DNN yielded a more promising outcome than the UNI, as it
instead of an internal fault. Therefore, the DNN places emphasis on minimizing FNs;
was 8 samples quicker. That is, the DNN faultlessly detected the inrush current after the
61st (=N–3) sample from the switching instance.
Energies 2024, 17, 963 14 of 18

5.3. Case Study 3: Energization of a Power Transformer in the Presence of an Internal Fault
Energizing a power transformer in the presence of an internal fault is a challenging task
for conventional protections, as the ratio of the second harmonic may cause the differential
relay to be blocked, potentially leading to severe damage to the power transformer. In this
case, we consider a–g faults for internal faults. Figure 9 shows the results of internal-fault
detection when a power transformer was energized in the presence of an internal fault. The
evaluation was conducted in two different scenarios at fault inception angles of 0◦ and 90◦ .
As shown in Figure 9a, the conventional HAR method detected the inrush current
rather than the internal fault due to the presence of the second harmonic in the decaying
DC component generated during the internal fault. Consequently, it prevented the internal
fault from being detected, resulting in the blocking of the differential relay operation. In
contrast, the UNI detected the differential current as an inrush current instead of an internal
fault. The EKF could not discriminate the internal fault from the inrush current. Moreover,
the inaccuracy increased as the EKF estimated differential currents with noise. Unlike the
conventional HAR and UNI methods, the proposed DNN demonstrated an impressive
success rate in discriminating the internal fault from the inrush current after the 58th sample
from the abnormality. In this manner, the DNN exhibited high sensitivity to internal faults,
even though the HAR and UNI failed to detect them. As shown in Figure 9a, for the
fault inception angle of 90◦ , the HAR failed to detect the internal fault for several cycles,
highlighting a drawback of using HARs in modern transformers. In contrast, the proposed
DNN successfully detected the internal fault, starting from just one sample later than the
DNN reference. Similarly, as illustrated in Figure 9b, the DNN exhibited a promising
output in discriminating between inrush currents and internal faults at a fault inception
angle of 0◦ .

5.4. Case Study 4: Phase-A-to-Ground Internal Faults Occurring during the Energization of a
Power Transformer
The proposed DNN was validated during an internal fault occurring a few cycles
after the switching of a power transformer. The harmonic-blocking scheme blocked the
operation of the differential relay due to the large ratio of the second harmonic at the
onset of an internal fault. This could lead to damage to the power transformer and should
be avoided.
A power transformer was switched on for energization at 0.22 s, and the internal fault
occurred at 0.32 s, as demonstrated in Figure 10. With the interference of the internal fault,
the HAR showed unsatisfactory results as soon as the internal fault occurred. The HAR
blocked the differential relay from operating for around two cycles, which could negatively
affect the power transformer. The UNI showed the worst results among the three methods,
as it did not respond to the internal fault in this case. UNI is only applicable when there is a
direction of the waveform on the positive or negative side, as its bidirectional index makes
it vulnerable to internal faults. The proposed DNN could detect the internal fault with a
time delay of less than one cycle from the fault inception. The evaluation was performed
on internal faults at fault inception angles of 0◦ and 90◦ , as illustrated in Figure 10a,b,
respectively. The results show that the proposed DNN can detect internal faults after a time
delay of less than one cycle, regardless of the fault inception angle.
The influence of external faults on the proposed DNN can be ignored since the differ-
ential current will be zero during an external fault. Therefore, the DNN bypasses external
faults and allows relevant protection schemes outside the protection zone to operate based
on disturbance criteria.

5.5. Case Study 5: Phase-B–C-to-Ground Internal Faults Occurring during the Energization of a
Power Transformer
To demonstrate the capability of the proposed DNN across different fault types, phase-
B–C-to-ground internal faults are considered in this case. Figure 11 presents a case of a
phase-B–C-to-ground internal fault at a different time node considering a fault inception
Energies 2024, 17, 963 15 of 18

angle of 0◦ . The internal fault depicted in Figure 11 occurs three cycles after the inrush
current takes place. Similar to Case Study 4, the UNI successfully detects the instance
of the inrush current; however, the operation of the differential protection is continually
blocked for almost one cycle after an internal fault occurs. On the other hand, the UNI
proves to be effective in responding to the inrush current but fails to detect the internal
fault for several cycles. The EKF exhibits low sensitivity to the internal fault because the
estimated current from the EKF produces noise. Unlike these three methods, the proposed
DNN demonstrates an accurate and reliable output in discriminating internal faults with a
given time delay.

6. Discussion on the Performance Evaluation Metrics

To effectively evaluate the performance of the proposed DNN, three indicators were
selected as evaluation metrics: accuracy, sensitivity, and precision. Traditionally, accuracy
alone is insufficient to determine whether the proposed DNN yields a promising outcome.
To visualize the stability of the proposed DNN method, a confusion matrix was used,
summarizing the classification performance and providing a visual representation of the
actual and predicted classes. The evaluation matrix was assessed using the following four
performance indices: TP (true positive), TN (true negative), FP (false positive), and FN
(false negative).
TP+TF
ACC = (9)
(TP + TF + FP + FN)
TP
SEN = (10)
(TP + FN)
TP
PRE = (11)
(TP + FP)
Conventionally, accuracy (ACC) shows the authenticity of a detection method, defin-
ing the correct detections over the total numbers of detections, including correct and false
ones. Sensitivity (SEN) measures the proportion of inrush and internal faults that were
correctly identified among the actual labels. It is a crucial metric in discrimination, because
it influences the decision to allow the differential relay to operate when an internal fault
occurs during inrush currents. A high percentage of SENs is essential to determine the
stability of the proposed DNN. Precision (PRE) is another important metric required to
affirm the correctness of the proposed DNN. For instance, it demonstrates the capability of
the proposed DNN to isolate internal faults from inrush currents when both abnormalities
occur simultaneously. In other words, it demonstrates the ability of internal-fault detec-
tion without mistakenly identifying it as an inrush current. A comparative analysis was
conducted, and the evaluation metrics are presented in Table 6. The effectiveness of these
metrics was assessed at the 58th and 61st samples from the beginning of each abnormality.
In cases where a power transformer is energized in the presence of an internal fault,
the aim is to avoid a situation where the DNN mistakenly detects it as an inrush current
instead of an internal fault. Therefore, the DNN places emphasis on minimizing FNs;
otherwise, incorrect detections could lead to damage to the power transformer. The DNN
detects the internal fault at the 61st sample, which is three samples later than the DNN
reference; therefore, the DNN experienced three FNs in this case. The performance of the
proposed DNN and the other methods was evaluated at the 58th (=N–6) and 61st (=N–3)
samples from the beginning of each abnormality. It is noted that detection with a time
delay of 61 samples will be sufficient to protect the power transformer, as the protection
decision will be made after 64 samples.
According to the percentages presented in Table 6, it is evident that all four methods
correctly classified the normal condition from the other two abnormalities without any
defects. For inrush conditions, the HAR was undoubtedly proven to be effective, achieving
the highest metrics at the 58th and 61st samples. The UNI exhibited good performance
in detecting inrush currents, with ACC, SEN, and PRE values of 99.852%, 93.814%, and
Energies 2024, 17, 963 16 of 18

95.724%, respectively. The UNI is unable to achieve the highest metric at the 61st sample,
as inrush currents were detected at the 62nd sample in some cases. On the other hand,
the UNI performed poorly when experiencing internal faults, as it was more sensitive to
inrush currents. The DNN displayed a promising evaluation index in detecting the inrush
duration at the 58th sample, yielding the highest ACC, SEN, and PRE values of 99.526%,
100%, and 99.523%, respectively. At the 61st sample, the DNN could accurately classify
between inrush currents and internal faults, achieving 100% for all three metrics.

Table 6. Evaluation metrics at the 58th (=N–6) and 61st (=N–3) samples from the beginning of
each abnormality.

Accuracy (%) Sensitivity (%) Precision (%)

Method Case
N–6 N–3 N–6 N–3 N–6 N–3
HAR 100 100 100 100 100 100
UNI 100 100 100 100 100 100
Normal
EKF 100 100 100 100 100 100
DNN 100 100 100 100 100 100
HAR 100 100 100 100 100 100
UNI 99.852 99.932 93.814 95.613 95.724 97.741
Inrush
EKF - - - - - -
DNN 99.526 100 100 100 99.523 100
HAR - - - - - -
UNI - - - - - -
Internal fault
EKF 90.513 92.364 69.192 72.951 71.231 72.367
DNN 99.931 100 100 100 98.842 100
HAR - - - - - -
UNI Inrush and - - - - - -
EKF Internal fault 91.103 92.136 69.583 71.369 70.124 70.364
DNN 99.651 100 99.642 100 100 100

Furthermore, the DNN demonstrates excellent performance in detecting internal faults

during inrush currents. The evaluation index produced by the DNN outperformed the
other three methods at sample N–6, achieving ACC, SEN, and PRE values of 99.651%,
99.642%, and 100%, respectively. At sample N–3, the DNN achieved the best metrics (ACC,
SEN, and PRE), all at 100%. In contrast, the EKF showed worse performance compared to
the DNN in this study, as it mis-detected the internal faults due to the difference between
the measured and estimated currents. Moreover, EKF is inapplicable to other systems
and significantly relies on a threshold to detect internal faults, presenting a less favorable
discrimination between inrush currents and internal faults. At sample N–3, it yielded ACC,
SEN, and PRE values of 92.136%, 71.369%, and 70.364%, respectively.

7. Conclusions
This paper proposes a DNN-based method to discriminate between inrush currents
and internal faults utilizing a data window. The effectiveness of the proposed DNN
was assessed through numerical simulations, including inrush currents, internal faults,
and cases where the inrush current coincided with internal faults. Despite achieving
less accurate results during inrush currents, compared to HAR, DNN performs better in
detecting internal faults, even during inrush conditions. Based on graphical illustrations
and evaluation metrics, DNN successfully detects internal faults during inrush conditions,
enabling the differential relay to operate without delay, regardless of the fault inception
angle and residual flux. As DNN does not require a specific threshold to perform the
discrimination, it can be applied to different systems to discriminate inrush currents from
internal faults.
Energies 2024, 17, 963 17 of 18

HAR and UNI are insufficient to deal with both inrush currents and internal faults
occurring together. Although EKF can detect internal faults, the effectiveness of EKF is
reduced in other systems due to an indecisive threshold. The deficiencies of the prevailing
methods, such as reliance on physical parameters and indecisive predefined thresholds,
decrease their reliability and generality. In comparison to prevailing methods (HAR,
UNI, and EKF), the proposed DNN shows promising results from sample N–3, achieving
accuracy, sensitivity, and precision values of 100%. It is considered to be one of the
promising solutions for discriminating between inrush currents and internal faults. The
proposed DNN may produce errors in the presence of CT saturation. Our future work
involves developing a discrimination model for the main and backup protections that
considers CT saturation and implementing the proposed DNN to discriminate internal
faults from inrush currents in real time. The experiment will be based on hardware
implementation, which consists of RTDS and EVM boards.

Author Contributions: Conceptualization, S.K. and S.-R.N.; methodology, S.K., G.-W.S. and
S.-R.N.; supervision, S.-R.N.; validation, S.K. and G.-W.S.; writing—original draft, S.K. and S.-R.N.;
writing—review and editing, S.K., G.-W.S. and S.-R.N. All authors have read and agreed to the
published version of the manuscript.
Funding: This research was supported in part by Korea Electric Power Corporation (Grant number:
R23XO05-07). This research was also supported in part by the Korea Research Foundation, with
funding from the government (Ministry of Education) in 2021 (No. NRF-2021R1F1A1061798).
Data Availability Statement: Data are contained within the article.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Gopika, R.; Deepa, S. Study on Power Transformer Inrush Current. IOSR J. Electr. Electron. Eng. 2017, 2, 59–63.
2. Baoming, G.; Dealmeida, A.; Qionglin, Z.; Xiangheng, W. An Equivalent Instantaneous Inductance-Based Technique for Dis-
crimination between Inrush Current and Internal Faults in Power Transformers. IEEE Trans. Power Deliv. 2005, 20, 2473–2482.
[CrossRef]
3. Guzman, A.; Fischer, N.; Labuschagne, C. Improvements in transformer protection and control. In Proceedings of the 2009 62nd
Annual Conference for Protective Relay Engineers, College Station, TX, USA, 30 March–2 April 2009; pp. 563–579.
4. Hamilton, R. Analysis of transformer inrush current and comparison of harmonic restraint methods in transformer protection.
IEEE Trans. Ind. Appl. 2013, 49, 1890–1899. [CrossRef]
5. Zhang, L.; Wu, Q.; Ji, T.; Zhang, A. Identification of inrush currents in power transformers based on higher-order statistics. Electr.
Power Syst. Res. 2017, 146, 161–169. [CrossRef]
6. Mo, C.; Ji, T.; Zhang, L.; Wu, Q. Equivalent statistics based inrush identification method for differential protection of power
transformer. Electr. Power Syst. Res. 2022, 203, 107664. [CrossRef]
7. Hodder, S.; Kasztenny, B.; Fischer, N.; Xia, Y. Low second-harmonic content in transformer inrush currents—Analysis and
practical solutions for protection security. In Proceedings of the 2014 67th Annual Conference for Protective Relay Engineers,
College Station, TX, USA, 31 March–3 April 2014; pp. 705–722.
8. Lu, Z.; Tang, W.H.; Ji, T.Y.; Wu, Q.H. A Morphological Scheme for Inrush Identification in Transformer Protection. IEEE Trans.
Power Deliv. 2009, 24, 560–568. [CrossRef]
9. Vazquez, E.; Mijares, I.I.; Chacon, O.L.; Conde, A. Transformer differential protection using principal component analysis. IEEE
Trans. Power Deliv. 2008, 23, 67–72. [CrossRef]
10. Afrasiabi, S.; Afrasiabi, M.; Parang, B.; Mohammadi, M.; Samet, H.; Dragicevic, T. Fast GRNN-Based Method for Distinguishing
Inrush Currents in Power Transformers. IEEE Trans. Ind. Electron. 2021, 69, 8501–8512. [CrossRef]
11. Afrasiabi, S.; Afrasiabi, M.; Parang, B.; Mohammadi, M. Integration of Accelerated Deep Neural Network into Power Transformer
Differential Protection. IEEE Trans. Ind. Inform. 2019, 16, 865–876. [CrossRef]
12. Samet, H.; Ghanbari, T.; Ahmadi, M. An Auto-Correlation Function Based Technique for Discrimination of Internal Fault and
Magnetizing Inrush Current in Power Transformers. Electr. Power Compon. Syst. 2015, 43, 399–411. [CrossRef]
13. Mao, P.; Aggarwal, R. A novel approach to the classification of the transient phenomena in power transformers using combined
wavelet transform and neural network. IEEE Trans. Power Deliv. 2001, 16, 654–660. [CrossRef]
14. Key, S.; Ko, C.-S.; Song, K.-J.; Nam, S.-R. Fast Detection of Current Transformer Saturation Using Stacked Denoising Autoencoders.
Energies 2023, 16, 1528. [CrossRef]
15. Key, S.; Kang, S.-H.; Lee, N.-H.; Nam, S.-R. Bayesian Deep Neural Network to Compensate for Current Transformer Saturation.
IEEE Access 2021, 9, 154731–154739. [CrossRef]
Energies 2024, 17, 963 18 of 18

16. Sok, V.; Lee, S.-W.; Kang, S.-H.; Nam, S.-R. Deep Neural Network-Based Removal of a Decaying DC Offset in Less Than One
Cycle for Digital Relaying. Energies 2022, 15, 2644. [CrossRef]
17. Silva, K.; Souza, B.; Brito, N. Fault Detection and Classification in Transmission Lines Based on Wavelet Transform and ANN.
IEEE Trans. Power Deliv. 2006, 21, 2058–2063. [CrossRef]
18. Costa, F.B.; de Souza, B.A.; Brito, N.S.D. Detection and Classification of Transient Disturbances in Power Systems. IEEJ Trans.
Power Energy 2010, 130, 910–916. [CrossRef]
19. Gaouda, A.M.; Salama, M.M.A. DSP Wavelet-Based Tool for Monitoring Transformer Inrush Currents and Internal Faults. IEEE
Trans. Power Deliv. 2010, 25, 1258–1267. [CrossRef]
20. Saleh, S.A.; Scaplen, B.; Rahman, M.A. A new implementation method of wavelet-packet-transform differential protection for
power transformers. IEEE Trans. Ind. Appl. 2011, 47, 1003–1012. [CrossRef]
21. Costa, F.B. Fault-induced transient detection based on real-time analysis of the wavelet coefficient energy. IEEE Trans. Power Deliv.
2014, 29, 140–153. [CrossRef]
22. Medeiros, R.P.; Costa, F.B.; Silva, K.M. Power Transformer Differential Protection Using the Boundary Discrete Wavelet Transform.
IEEE Trans. Power Deliv. 2015, 31, 2083–2095. [CrossRef]
23. Medeiros, R.P.; Costa, F.B. A Wavelet-Based Transformer Differential Protection with Differential Current Transformer Saturation
and Cross-Country Fault Detection. IEEE Trans. Power Deliv. 2017, 33, 789–799. [CrossRef]
24. Medeiros, R.P.; Costa, F.B. A wavelet-based transformer differential protection: Internal fault detection during inrush conditions.
IEEE Trans. Power Deliv. 2018, 33, 2965–2977. [CrossRef]
25. Ruhan, Z.; Mansor, N.N.B.; Illias, H.A. Identification of Inrush Current Using a GSA-BP Network. Energies 2023, 16, 2340.
[CrossRef]
26. Peng, F.; Gao, H.; Huang, J.; Guo, Y.; Liu, Y.; Zhang, Y. Power Differential Protection for Transformer Based on Fault Component
Network. IEEE Trans. Power Deliv. 2023, 38, 2464–2477. [CrossRef]
27. Ali, E.; Helal, A.; Desouki, H.; Shebl, K.; Abdelkader, S.; Malik, O. Power transformer differential protection using current and
voltage ratios. Electr. Power Syst. Res. 2017, 154, 140–150. [CrossRef]
28. Liu, P.; Jiao, B.; Zhang, P.; Du, S.; Zhu, J.; Song, Y. Countermeasure to Prevent Transformer Differential Protection from False
Operations. IEEE Access 2023, 11, 45950–45960. [CrossRef]
29. He, B.; Zhang, X.; Bo, Z. A New Method to Identify Inrush Current Based on Error Estimation. IEEE Trans. Power Deliv. 2006, 21,
1163–1168. [CrossRef]
30. Elsadd, M.A.; Yousef, W.; Abdelaziz, A.Y. New adaptive coordination approach between generator-transformer unit overall
differential protection and generator capability curves. Int. J. Electr. Power Energy Syst. 2019, 118, 105788. [CrossRef]
31. Rao, J.G.; Pradhan, A.K. Power-Swing Detection Using Moving Window Averaging of Current Signals. IEEE Trans. Power Deliv.
2014, 30, 368–376. [CrossRef]
32. Hinton, G.E.; Osindero, S.; Teh, Y.-W. A Fast Learning Algorithm for Deep Belief Nets. Neural Comput. 2006, 18, 1527–1554.
[CrossRef]
33. Bengio, Y.; Lamblin, P.; Popovici, D.; Larochelle, H. Greedy Layer-wise Training of Deep Networks. In Proceedings of the NIPS’06
Proceedings of the 19th International Conference on Neural Information Processing Systems, Vancouver, BC, Canada, 4–7 December 2006;
pp. 153–160.
34. Ken, B.; Normann, F.; Casper, L. Considerations for Using Harmonic Blocking and Harmonic Restraint Techniques on Transformer
Differential Relays. SEL J. Reliab. Power 2011, 2, 1–17.
35. Gunda, S.K.; Dhanikonda, V.S.S.S.S. Discrimination of Transformer Inrush Currents and Internal Fault Currents Using Ex-tended
Kalman Filter Algorithm (EKF). Energies 2021, 14, 6020. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.