Received 27 July 2023, accepted 9 August 2023, date of publication 17 August 2023, date of current version 25 August 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3305958

High Impedance Fault Detection on Microgrids

Considering the Impact of VSC Based Generation
VITOR FERNANDO COUTO AND MIGUEL MORETO
Department of Electrical and Electronics Engineering, Federal University of Santa Catarina, Florianópolis 88040-900, Brazil
Corresponding author: Vitor Fernando Couto ([email protected])
This work was supported in part by Coordenação de AperfeiÇoamento de Pessoal de Nível Superior-Brasil (CAPES) under Grant 001.

ABSTRACT This paper presents an evaluation of the impacts of simulation models for high impedance
faults detection in microgrids with distributed generators interfaced by voltage source converters. Detailed
simulation models are used to consider the high frequency switching of power electronics and, consequently,
the associated harmonic distortions. Thus, for systems with a higher degradation of the current sine
waveform, the high impedance fault characteristics such as nonlinearity, asymmetry, and randomness may
remain hidden, compromising the performance of some existing detection techniques. Such models are not
commonly used in the high impedance fault detection literature. To evaluate the impact of the simulation
model’s complexity, a pretrained deep convolutional neural networks on the imageNet database is adapted
to classify current waveform based on a time-frequency representation image. The proposed method is
designed and tested using a microgrid model based on the CIGRE European medium voltage distribution
network benchmark. The simulation data set considers microgrid operating in islanded and connected
manner, radial and ring topology, with transients at different locations of the network. The results indicate that
high impedance fault detection methods for microgrid with voltage source converters generators need to be
designed using detailed simulation models to achieve better accuracy, and the proposed classifier presented
accurate and robust high impedance fault detection.

INDEX TERMS High impedance fault, microgrid, simulation models, transfer learning, voltage sourced
converter.

I. INTRODUCTION flows and different short-circuit levels depending on the

The concern with climate change, the greenhouse effect, operation mode. In addition, the variations associated with
and the pursuit for independence from fossil fuels has generation, loads, and configuration lead to challenges to be
led to a major transformation of electrical power systems. overcome [2].
In recent years, renewable distributed generation (DG) has Traditional protection schemes need to be adapted to these
been inserted into the distribution network to meet the new characteristics of MG. Among the problems to be over-
ever-increasing demand for electricity locally [1]. This trend come, high impedance fault (HIF) provides various complex
is expected to be stronger in future grids with a higher degree issues for the utility industry, as conventional overcurrent
of renewable energy sources that are usually interfaced using relays fail to detect HIFs due to the low current levels.
voltage source converters (VSCs). This type of fault can occur when an energized conductor
In this environment, microgrid (MG) have emerged, touches the surface with a high impedance, such as vegetation
facilitating the incorporation of DG power sources. Such near overhead distribution lines or when the conductor falls
a concept can result in greater reliability to the electricity to the ground. These situations expose electrical system
supply since they can operate connected or islanded from equipment, private properties, and especially human lives to
the main grid. Some of the concerns with protection risk [3].
systems in MG include the possibility of reversible power For this reason, more efficient detection methods were
proposed, most of them based on tracking voltage and current
The associate editor coordinating the review of this manuscript and distortions due to the HIF. This protection issue in distribution
approving it for publication was Hazlie Mokhlis . systems networks is not an outdated problem [4] and becomes
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
89550 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

even more challenging in a MG with a high penetration of The authors of [21] and [22] proposed HIF protection
generators interfaced by VSC [5]. strategies for MG while taking VSC DG into account, but they
An extensive review of HIF detection methods was discovered performance deterioration for the island operation
covered in [4]. Distortions due to nonlinearity of voltage mode. Detection and classification of MG faults in [23]
and current, asymmetry between positive and negative semi- take into account radial and meshed topology and also the
cycles, nonstationary current and, randomness arise from mode of operation. Additionally, in [24], a protection strategy
HIF. Detection methods typically use signal processing tech- for MG was presented, considering HIF and thresholds
niques to identify these features within voltage and current were established for residual voltage and current. However,
signals in a harmonic-free environment. Nevertheless, in MG to achieve selectivity in the HIF detection logic, a delay of at
characterized by a substantial deployment of VSC generators, least 0.5 seconds had to be included.
a notable level of distortion arises, potentially impeding the Based on the aforementioned studies, a small number of
effectiveness of detection algorithms. In addition, switching works took DG into account using detailed models. Even
transients have HIF-like characteristics, such as capacitor among those that did, they did not examine its impact on
bank connection and large single-phase load switches, which the classification problem. These findings highlight the need
can lead to false detections. for further investigation into this subject. Thus, this paper
There are some studies in the technical literature that intends to discuss the impact of detailed simulation models on
propose solutions to detect HIF in systems with DG, as in [6] HIF detection in MG systems and also proposes an effective
and [7]. Despite the consideration of DG and HIFs, the test classification method. The results were compared with a
systems utilized were free of harmonic distortions. Such recent method from the literature.
circumstances are improbable at present, but the prognosis The proposed detection method uses as input current sig-
suggests a worsening situation in the future for distribution nals converted into RGB images for a Deep Neural Network
systems, particularly in MGs. In the existing literature, (DNN) classifier. This approach has the advantage of being
the considered distortions [5], [8], [9], [10], [11] primarily independent of prior knowledge and human intervention in
stemmed from the utilization of detailed simulation models feature extraction. The goal is to show detailed simulation
for VSC generators. Nonetheless, the literature fails to models result in more realistic data, providing a challenging
encompass MG scenarios involving island operation and ring condition for HIF detection methods. In summary, the
topology. contributions of this paper can be stated as follows:
Certain papers have previously addressed the HIF problem • Proposition of a novel HIF detection method using
in MGs, but without providing specific details regarding image classification techniques. The approach uses VSC
the representation of distributed generation, as observed simulation models to accurately capture the effects
in [12], [13], [14], and [15]. of switching frequency and dynamic VSC generator
There are methods for detecting HIF in MG that take into characteristics.
account the application of comprehensive simulation models • The development of a methodology to identify HIF
for VSC-based generation. In [16], a protection scheme for using transfer learning from 2-dimensional DNN fea-
microgrids capable of detecting HIF was presented. The test tures trained on ImageNet [25].
system comprised an IEC model test microgrid incorporating • The transformation of a time series signal into an RGB
a DFIG-based wind farm, an inverter-based DG wind farm, image by means of a Continuous Wavelet Transform
and two synchronous DGs. However, the authors did not (CWT).
examine this aspect in their results. • Variations in the MG’s modes of operation (islanded or
A long short-term memory (LSTM) based protection connected) and topology (ring of radial) are taken into
scheme was proposed to detect HIF in hybrid MG [17]. The account during HIF detection.
PV DG consider a statistical modelling for solar irradiance • Comparative evaluation of numerical results with other
intermittency, combining Gaussian distribution function and existing approaches [15], [16], [17], [18], [20].
meteorological data. Although the DG was not a source of • Evaluation with five performance indexes and compare
harmonic intrusion in the simulation data, the approach’s the results with similar method [20] for the same
immunity to it was evaluated separately for nonlinear loads database.
and Gaussian noise. The real-world applicability of this • Noise resilience test adding current signals white
approach may be affected since Gaussian noise cannot Gaussian noise of varying levels (0 dB < SNR < 30 dB).
accurately recreate the dynamics of PV DG in transient states, • Tests are conducted with three different HIF models
particularly in high penetration VSC DG and MG islanded to evaluate the method’s ability to recognize HIF with
operation. This was also the case for the papers [18], [19]. different characteristics for which it has not been trained.
Moreover, in [20] the LSTM was used to classify HIF with The rest of this paper is organized as follows. Section II
three-phase current signal. The method achieve overall clas- explains MG features and VSC generators simulation models.
sification accuracy of 91.21% and outperforms well-known HIF features and simulation models used are described in
classifiers in tests performed. This approach consider PV DG Section III. Section IV contains the proposed methodology
with detailed model in a distribution network. and data set description. Section V provides the classification

VOLUME 11, 2023 89551

 V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

test results and analysis. At the final Section VI presents the

conclusions and further development intentions.

II. MICROGRID TOPOLOGY, LOADS AND GENERATORS

In MGs, it is possible to have reversible power flows, different
levels of short-circuit for the islanded and connected modes,
and the presence of VSC generators behave differently from
synchronous generators. All these features require different
protection than traditional distribution systems [26].
In this paper, the MG test system has simulation models
that consider the VSC switching circuits that reproduce
the associated harmonic distortions. These distortions can
impact the effectiveness of HIF detection algorithms, causing
misclassification or lack of sensibility [22], [27], [28].

A. MICROGRID
The system used is an adaptation of one of the test systems
presented in [29] by CIGRE’s commission C6 (Distribution
and Distributed Generation Systems). This system is called
Topology of European MV Distribution Network Benchmark, FIGURE 1. Single line diagram of the microgrid based on CIGRE’s
European MV distribution network benchmark.
and was chosen among the two that operate at medium
voltage (European and North American). The MG structure
has two feeders, but only feeder 1 is considered in this work.
System operation settings were changed from 20kV-50Hz to
25kV-60Hz. The loads are balanced, all the lines are overhead
and three-phase with non-insulated aluminum conductors.
The MG single line diagram is shown in the Fig. 1.
In [29] it is stated that grounding is a matter of regional
FIGURE 2. Simulation model of the PV generator.
preferences, being the European distribution networks typi-
cally isolated or grounded by impedance. In this work it is
considered a MG with a grounding zig-zag transformer (GT), (273 Vdc at maximum power) to 500 Vdc . The VSC converts
connected between bus 1 and T1, to create a low-impedance this continuous voltage to 260Vac , operating with a unit power
path for zero sequence currents, while being high impedance factor. It is also considered an RC filter to mitigate the
for positive and negative sequence currents. There are three harmonic distortions.
switches, S1 is the Common Connection Point (PCC) that The plant has an MPPT (Maximum Power Point Tracking)
allows the MG to operate in an isolated or connected mode. control that adjusts the operating point for the highest energy
When tie switches S2 and S3 are closed, the MG has a output. The simulations are performed with a temperature
loop configuration, and when both are opened, the system of 45◦ C and an irradiance of 1000 W /m2 . The remaining
becomes radial. parameters of this PV farm are the same as the one in [30].
The distributed generation sources are included using
detailed models, that represent static VSC power converters. C. WIND GENERATOR
The DG consists of three solar power plants, one wind power The wind power plant is based on 1.5 MW Double-Fed
plant and one synchronous generator. This synchronous Induction Generators (DFIG). The plant is composed of six
generator acts as a synchronous reference in islanded generators, with an output voltage of 575Vac , connected to
connection mode, responsible for power balance, modeled as Bus 3 of the MG through a 25kV/575V (D11yg) 9MW
a three-phase source with an internal R-L impedance. The transformer.
other generators controls are set to extract maximum power Fig. 3 shows the scheme for one of these generators,
from the available primary energy. where the implemented model includes a representation of
the IGBT-based converters (1620 Hz and 2700 Hz switching
B. PHOTOVOLTAIC GENERATOR frequencies). The armature winding (or stator) is connected
The MG contains three photovoltaic plants of 100kWp each, to the network, and the field winding (rotor) is connected
connected to the network by 100kVA, 25kV/600V (D11Yg) to VSC converters. The converters are called RSC (Rotor
transformers. It can be seen in Fig. 2, the plants have two Side Converter) and LSC (Line Side Converter) and allow
converters, the DC-DC and the VSC. The DC-DC boost the frequency of the field winding to be adjustable. This way,
converter regulates the output voltage of photovoltaic panels, the generator can operate at variable speeds with increased
which varies according to the intensity of solar irradiance efficiency.

89552 VOLUME 11, 2023

V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

asymmetry in voltage and current waveforms. The electrical

current resulting from this type of fault has a frequency
spectrum that varies over time. In addition, current magnitude
and conduction intervals are random events. This relationship
between voltage and current is non-linear and generates
harmonic components [33].
The majority of HIF detection techniques employ electrical
FIGURE 3. Simulation model of the DFIG wind generator and its current measurements, as illustrated in [3] and [4]. There
associated power converters. are also some methods using a combination of voltage
and current or magnetic fields. To classify the event,
characteristics are taken from the collected voltage and/or
current data. These characteristics may be in the time domain,
frequency domain, time-scale domain, or time-frequency
domain. The physical phenomenon of HIF very often involves
electrical arcing, making the electrical current random and
unpredictable. A variety of current waveforms are produced
by environmental factors including humidity and the kind of
contact surface.
The development of HIF detection techniques with satis-
factory performance is closely related to the quality of the
database used. This data can be obtained through field tests.
It is necessary to perform these field tests, special equipment
and skilled labor are required. The characteristics of these
data tend to be more specific to the distribution system
FIGURE 4. Simulation model of the non-linear load with two thyristor
rectifiers.
tested, making them difficult to generalize and use in other
systems [34]. Another option is the use of simulation models
to obtain this data. With a few adaptations, it is possible to
In the simulations, the plant operates with a constant wind apply HIF to different systems and test them for a wide variety
at 15 m/s and the reactive power is set to 0 MVAr. The same of events. The HIF waveforms obtained from simulation
model was used in [21] and [23]. models were equivalent to those from field tests, confirming
their validity [7].
D. NON-LINEAR LOAD This paper uses three different HIF simulation models
It was not something new that certain loads share characteris- to create the database for the proposed detection method.
tics with HIF, which can result in false triggers. These loads, These models were variations of Emanuel’s HIF model [33]
such as arc welders, furnaces, computers, and other non- that use a combination of electrical elements like voltage
linear applications, produce harmonic distortions [31]. Thus, sources, diodes, inductances, and resistances to reproduce
it is necessary to test the ability of HIF detection methods to the HIF characteristics. The Emanuel model was based on
distinguish between the fault and switching of this type of laboratory experiments, field measurements, and theoretical
load. In [27] and [32], Soheili et al had proposed methods calculations. Since then, a number of further articles have
for HIF detection, taking into account the presence of non- proposed refining and adapting this model to better represent
linear loads. In this paper, a detailed Matlab/Simulink model this physical phenomenon [4].
of a three-phase six-pulse bridge thyristor rectifier is used as The source-diode-resistance HIF model, shown in Fig. 5
a building block for a non-linear load as shown in Fig. 4. (a), has tuned parameters to suit the system voltage and is
Load 1 is supplied by two six-pulse bridges connected in connected between phase (Vph ) and ground. In this model, the
series, with their AC connections in a transformer with two positive branch (RP ) and (VP ) have smaller values relative to
secondary windings (one wye, one delta). Load 2 is supplied the negative branch (RN ) and (VN ) to reproduce the half-cycle
by a single six-pulse bridge. The combination of the two asymmetry and intermediate arc extinction observed in the
loads results in a consumption of 500 kW 800kVAr with 17% HIF waveforms. In this circuit, current flows to ground when
THDi. the instantaneous value Vph > VP , reverse when Vph > VN
and if VN < Vph < VP there is no fault current.
III. HIGH IMPEDANCE FAULTS In this paper, three different HIF models (named Model I,
The main detectable characteristics of HIFs are due to the Model II and Model III) are employed in Matlab/Simulink
formation of an electric arc. This arc generates a distorted for multiple simulations. The inclusion of three different
electrical current, alternating between conduction and no models helps build a database that includes all of the
conduction cycles. The voltage threshold to dielectric rupture HIFs’ features while preventing the proposed method from
is different for positive and negative voltages, resulting in becoming overfit.

VOLUME 11, 2023 89553

 V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

FIGURE 7. V-I characteristics of HIF (a) model I (b) model II (c) model III.
FIGURE 5. Emanuel type of HIF simulation circuits: (a) model I [35] and
(b) model III [7].
the characteristics of HIF current, such as: intermittence,
asymmetry, buildup, shoulder, randomness, and nonlinearity.
In the actual settings of the benchmark MG employed in
this paper, VP and VN vary randomly pm 10% around
8 kV and 5 kV independently every 0.1 ms. Also, the two
independently and randomly changing resistors RP and RN
are varying between 1000  and 1500 , at a rate of 0.1 ms
as well. The circuit of HIF model III is presented in Fig. 5 (b).
Fig. 7 depicts the V-I characteristics of the HIF models. The
first graphic, Fig. 7 (a), represents HIF model I with settings
FIGURE 6. Circuit of HIF simulation model II [37].
VN = 6155V , VP = 6180V , RN = 245 and VN = 245.
In Fig. 7 (b) the characteristics are to the HIF model II state 2.
A. HIF MODEL I The third graphic, Fig. 7 (c), show HIF model III. The total
The HIF I model was based on Emanuel’s model [33], where harmonic distortion for these examples of HIF is 20%, 8%
inductance and resistance have been replaced by resistors of and 22% respectively.
unequal values to reproduce the asymmetry in the current
IV. THE PROPOSED CLASSIFIER AND DATA SETS
waveform. This model reproduces low frequency phenomena
HIF detection is not a simple task, and as can be seen in the
typical of an arcing fault involving sandy soil. The circuit is
literature, there are several papers that propose to solve this
shown in Fig. 5 (a), and ten sets of typical values for DC
problem with different approaches [4]. In this work, some
sources and resistors were shown in [35].
features extracted from measurements of the test system have
This model, although being frequently used in the devel-
been evaluated, and the simply setting thresholds was not
opment of detection techniques, is unable to recreate the
able to determine the occurrence of a HIF. Therefore, a state-
build-up and shoulder characteristics of an HIF [36].
of-the-art pattern recognition technique is used to evaluate
the impact that simulation model complexity has on HIF
B. HIF MODEL II
detection method’s performance. This technique is described
The HIF model II was based on the Emanuel arc model but below.
used six arc models in parallel, as shown in Fig. 6, to improve
HIF simulation. Each one of the six arc circuit refers to the A. CLASSIFICATION FEATURES AND TRAINING METHOD
occurrence of an electric arc, and their values of resistances, In [38] the detection of HIF in the distribution network used
DC-sources, and switching time for five different fault types a time-frequency-based method. The algorithm performs
were detailed in [37]. These values were adjusted to fit the five steps to extract a set of minimally relevant features
waveforms obtained through real-world testing. The authors from the measured current to detect HIF. These tasks are
proposed one HIF model for each of five different surfaces, complex and involve the creation of 2D matrix with time-
with a variation in the maximum amplitude of the fault current frequency distributions, feature dimensionality reduction
between 9 and 90 A. (PCA - Principal Component Analysis), and classification by
This model reproduces the first eight cycles of HIF current support vector machine (SVM).
with some characteristics, such as asymmetry and nonlinear- The detection method proposed here also uses the
ity of the HIF, and all frequency components up to 12 kHz. time-frequency characteristics of the current but takes
fewer steps. From the three-phase current measurements
C. HIF MODEL III the instantaneous residual current is calculated (IR (t) =
The HIF model III is similar to model I, but the values for Ia (t) + Ic (t) + Ic (t)) [39]. Applying Continuous Wavelet
resistance and DC sources vary randomly and independently Transform (CWT) results in a RGB image which is the input
every 0.1 ms [7]. The model is also simple and gets all data of a pretrained DNN.

89554 VOLUME 11, 2023

V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

FIGURE 8. Examples of scalograms. The top images correspond to 3 HIF

images while the bottom ones are obtained from other switching events.

The DNN has an architecture that learns from data,

eliminating the need for feature extraction [40]. Thus,
a technique called transfer learning was used reducing the
need and effort for training data collection. Also, transfer
learning is applied to take advantage of the existing neural
networks, it has been trained on large data sets, such as
ImageNet [25] for image classification. The selection of
DNN represents a trade-off between three qualities: accuracy,
speed, and size. The DNNs VGG19, Squeezenet, GoogleNet,
and AlexNet were examined and tested. SqueezeNet [41]
was chosen due to its improved performance and compact
size, which facilitates implementation. The transfer learning
technique for HIF detection was used in [42], [43], and [44] FIGURE 9. Flowchart of the Proposed method.

but not with image classification.

The first and second steps is the calculation and conversion
of residual current IR into a scalogram image, which is the frequency, for HIF and nonHIF cases. The simulation
absolute value of the CWT coefficients of a signal. Each of model is discretized at 5 microseconds time step, due to
the simulations performed has a corresponding RGB image detailed representation of the VSC converters with switching
of size 227-by-227-by-3, where 3 is the number of color frequencies of 5kHz (PV) and 2.7kHz (DFIG). The same time
channels. These images form the data sets to train and test the step and sampling frequency was used in the simulations with
deep CNN. These images are time-frequency representations simplified models.
of IR current, and some examples are shown in Fig. 8. All HIF simulations models were presented in section III.
The third step consists of training the DNN to recognize Others power system switching transients were simulated.
HIF. During the transfer learning process, some of the These transients share some characteristic with HIF, such
18 layers of the SqueezeNet architecture are replaced as the high-frequency components of load and capacitor
with new ones. The first layers identify common image switching, and the low frequency components and asymmetry
characteristics (spots, edges, and colors), the last layers focus present in transformer inrush currents [3].
on specific characteristics to distinguish classes. The ‘drop9’ All simulations last for 2 seconds. During the first
layer, which prevents overfitting, is replaced with a dropout 0.5 seconds the system stabilizes in the steady state condition.
layer of probability 0.6. Also, the ‘conv10’ last learnable layer Thus, only the final 1.5 seconds were considered to obtain
is replaced by a new convolutional layer with two filters that the residual current (IR ) scalogram images. Simulations of
are the number of classes in the data set (HIF and nonHIF). training data set are listed below. All simulated transients con-
To finish the adjustment, the classification layer is replaced sidered 100%, 80% and 60% MG loading in all combinations
by a new one without class labels. The labels will come of topology (ring or radial) and operating mode (connected or
from the new training data set. The classification process is islanded).
presented in Fig. 9. 1) High impedance faults: intermittent, with and without
interruption;
B. DATABASE 2) Capacitor switching: 600 kVar, 400 kVar and 300 kVar;
A Simulink/MatLab model of the MG test system was used 3) Transformer switching: switching on and off with 10%
to collect the signal IR in bus 3, with 15.36kHz sampling and 30% loads;

VOLUME 11, 2023 89555

 V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

defined as:
 
DF ND
Cm = (1)
MF HC
where in the case of detection algorithms, number DF shows
the number of right detection of the HIFs, HC demonstrates
number of right decision of healthy conditions, MF is
FIGURE 10. Structure of the dataset formed through MatLab/Simulink
number of healthy condition that mistakenly classified as
simulation. HIFs and finally ND shows the number of HIFs that were not
detected [38]. The aim is to achieve the highest values of the
4) Linear load switching: single-phase, two-phase and five reliability indexes listed below.
three-phase connected in delta and wye; • Accuracy: it is the amount of correct detections in
5) Non-linear load switching: as shown in section II-D; relation to the overall number of cases. It emphasizes
6) Three-phase line switching: combination of S2 and S3 the Overall precision of the method.
opening and closing, shown in Fig 1; DF + HC
7) No transient: All MG operation mode and topology AC = (2)
DF + ND + HC + MF
combinations with no transient.
• Dependability: this index can be expressed as the ratio
These combination results in 1800 (900 HIF and 900 non- of the number of identified faults to the actual number of
HIF) different simulation conditions, as shown in Fig. 10, faults. Dependability is defined as faulty state detection
and two training data set are assembled. The first training set precision.
have 1260 images considering DGs with detailed model of
the VSC generators. The classifier trained with this data set DF
DP = (3)
is referred to as the Detailed Classifier. In the second data set, DF + ND
the detailed models for the DG’s are replaced by three-phase • Security: it demonstrates the technique’s capability to
source blocks available in MATLAB/Simulink. They are recognize the HIF only when it is present and not trip
specified to delivering the same active and reactive power for nonHIF conditions. It represents the healthy state
as DG’s detailed models. The Simple Classifier is derived detection precision.
from this last training data set. SqueezeNet training data set HC
is divided into 70% for training and 30% for validation. SC = (4)
HC + MF
Another data set, with different transient parameters from
training, is created to test the performance of the Detailed • Safety: this index takes into account the method’s
and Simple classifiers. This test data set has 540 simulation capacity to isolate faults that pose a risk to the general
(270 HIF and 270 nonHIF) only with DG’s detailed models. public, such as HIF in a congested area, assuring that
The data sets utilized for testing and training are different. faults are not confused with nonHIF. The number of
To prevent bias due to the small number of examples, the data correctly predicted nonHIF divided by the total number
for training and testing are manually divided in 70% and 30%, of nonHIF determine this criterion.
respectively. For the training process, this data is randomly HC
SF = (5)
split into training and validation, in a 70/30 ratio, as seen in ND + HC
Fig. 10. • Sensibility: this index can be defined as the percentage
The data sets for training and testing are organized to of correctly predicted faults over the total number of
perform five different tests. To ensure a proper evaluation of predicted faults. Emphasizing the danger of tripping
the proposed classifier, performance indexes are selected for sensitive loads.
use. DF
SN = (6)
DF + MF
C. PERFORMANCE EVALUATION
Due to the complexity of the HIF phenomenon and its D. COMPARATIVE EVALUATION
detection techniques, a complete set of criteria is needed for In addition to the evaluation indexes, the results are compared
their evaluation. While tripping the power system due to the with state-of-the-art HIF detection approaches, shown in
HIF guarantees public safety, it would also disconnect vital Table 1. The rates of Proposed approach were slightly lower
loads such as traffic lights, elevators, and hospitals. Thus, than those in [17], although this method requires seven
two criteria, safety and sensibility, are suggested for utilities currents measurement points. The [15], [18], [20] approaches
to use when performing a risk analysis on this tradeoff. The only employ one point with currents measurement, similar
complexity of HIF detection methods necessitates the use of to Proposed method, but for [16], ten voltages and currents
additional indices [3] to assess their efficacy. measurement points were needed, as presented in Table 2.
Five evaluation indexes derived from the confusion matrix The islanded operation mode of the test system was not
will be utilized. Confusion matrix is a two-by-two matrix taken into account in [15], [16], [17], [18], and [20] does

89556 VOLUME 11, 2023

V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

TABLE 1. Comparative evaluation of performance indexes.

TABLE 2. Data set information.

FIGURE 11. Classification performance for the complete data set test.

Simple SqueezeNet and Method [20] classifiers. All tests

were performed on simulation data with detailed DG models,
thus reproducing a more realistic simulation scenario.
not take noise into account. The test data sets also contain
a significant class imbalance, with samples in [15] consisting
of 120 HIF and 10 nonHIF, in [18] consisting of 928 HIF A. CLASSIFICATION TESTS FOR COMPLETE DATA SET
and 54 nonHIF, and in [20] consisting of 66 HIF and The complete data set comprises a total of 630 cases from
128 nonHIF, shown in Table 2. Since the evaluation index for each of the three HIF models and 630 non-HIF cases. The
the compared approaches have fairly similar values, all the latter corresponds to 245 cases of the other switching types
factors mentioned above should be taken into account. described in IV-B, 76 cases of switching non-linear loads,
The details regarding the input data type and sample sizes 133 cases of switching lines, and 76 cases of simulations
for the training and testing data sets can be found in Table 2. without any transients.
In [16] the number of samples for testing is not presented. The results of the test using the five performance indexes
The HIF detection technique presented in [20] are select mentioned previously are shown in Fig. 11. When trained on
to classify the same test data set used in the Proposed detailed data, the Proposed method utilizing the pretrained
method. This work was based on simulations with nonfault Squeezenet DNN outperformed the Method [20] employing
(normal operation, capacitor switching, load switching, LSTM in terms of all five parameters. When compared to
transformer inrush current) and fault (HIF, symmetrical and the simplified ones, the two Detailed classifiers produced
unsymmetrical fault) in IEEE 13-bus distribution system a significantly superior outcome. The results of these tests
with 300 kW PV DGs modeled in MATLAB/Simulink. The reveal that advanced pattern recognition techniques, in the
detection technique implemented a deep learning method absence of a robust database, yield inaccurate HIF detection
of RNN based LSTM with feature extraction using DWT methods. This is particularly evident when considering
analysis. The three-phase current signal was processed using simulations with detailed DG models, which closely emulate
db4 mother wavelet to extract each phase’s energy value real-world application scenarios. The methodology underper-
characteristics for training and testing the classifiers. formed solely based on the security parameter when using
The authors achieved classification accuracy of 91.21% simplified training data.
with a success rate of 92.41% in identifying HIF in the sim-
ulations performed. This method [20], using LSTM, follows B. TESTS FOR HIF MODEL II
the same rule for the training data as the SqueezeNet, result- The ground surface material, surface humidity, feeder archi-
ing in the Detailed Method [20] and Simple Method [20] tecture, voltage levels, weather, and load type are just a few
classifiers. of the variables that have an impact on HIFs [3]. Therefore,
creating a database from simulations that can accommodate
V. CLASSIFICATION TESTS, RESULTS AND ANALYSIS all conceivable configurations of HIF occurrence in an MG
Five test scenarios were established in order to determine is a highly challenging task. The tests with the HIF II and
the efficacy of the proposed HIF detection approach. The III models assess the resilience of the developed detection
complete database is used in the initial test. The capacity method. The idea is to subject the classifier to data from a
to recognize model II and III HIF is assessed in the second HIF model not used during the training process and evaluate
and third tests, respectively. The performance regarding its performance.
differentiating between non-linear load and HIF is analyzed The data set for training consists of 1050 cases (420 HIF,
in the fourth test. The ability to not identify switching on MG 630 nonHIF) and for testing are 570 cases (300 HIF,
transmission lines as HIF is tested in the fifth and final test. 270 nonHIF). The results are in Fig. 12. In order to create
The results are shown in the same figure for the two the test data set, all of the model II HIF data were removed
Detailed SqueezeNet and Method [20] and for the two from the training set and added to it.

VOLUME 11, 2023 89557

 V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

FIGURE 12. Classification performance for HIF model II. FIGURE 14. Classification performance for non-linear load.

FIGURE 13. Classification performance for HIF model III. FIGURE 15. Classification performance for line switching.

Compared to the base case, the Detailed Squeezenet were not trained with them. There were 76 examples for
classifier maintained the safety and sensitivity indices, training and 32 cases for testing in the first test, but all of
a small degradation in accuracy and a slightly lower them were only included in the test set here.
performance than 90% in dependability and safety. The The two Detailed classifiers performed very similar to the
Detailed Method [20] classifier, on the other hand, showed first test, shown in Fig. 14. Thus, it is possible to confirm that
a greater degradation in its performance indices, dropping non-linear load switching did not lead to misclassification
from 93% to 78% in accuracy and from 89% to 65% despite sharing features with HIF. The simple classifiers
in dependability. When compared to the base test using obtained slightly better results compared to the first test, but
the entire database, the performance difference for Simple due to their low performance, it is not feasible to use them in
classifiers was not significant. real applications.

C. TESTS FOR HIF MODEL III E. TESTS FOR TRANSMISSION LINE SWITCHING
The same procedure used in the previous test was applied The purpose of this test is to verify whether changing MG
to evaluate HIF model III. The training data set consists arrangements after the closing or opening of the tie switch
of 1050 cases (210 HIF model I and 210 HIF model II, causes misclassification. Many HIF detection techniques
630 nonHIF) and for testing 570 cases (300 HIF model III, consider radial distribution systems, which do not take into
270 nonHIF). In this test, the detailed classifiers produced account this kind of scenario. The data sets for this test
the best results, as seen in Fig. 13. The best accuracy for the are 1027 cases (630 HIF 397 nonHIF) and for testing are
Simple SqueezeNet classifier was achieved in this test. 570 cases (270 HIF 403 nonHIF). The transmission line
As shown in Fig. 7, the resistors and DC sources in switching simulation data was all included in the test set and
this HIF III model change randomly and independently removed from the training set.
every 0.1 ms. Despite this disparity from the other two HIF The evaluated indices produced the lowest scores for the
models, both Detailed classifiers were able to distinguish the Detailed Squeezenet classifier shown in Fig. 15. These rates,
characteristics of the HIF current signal in the scalogram however, surpass the highest performance the Method [20]
images. classifier could produce.

D. TESTS FOR NON-LINEAR LOAD CONNECTION F. TESTS FOR NOISE RESILIENCE

Non-linear loads have been previously mentioned in To evaluate the performance of the Proposed Method in
section II-D as having HIF-like features. Since these noisy environment, a white Gaussian noise of varying levels,
transients only arise in tests, the classifiers used in this test (0 dB < SNR< 30 dB) was added to current signals in

89558 VOLUME 11, 2023

V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

results demonstrated the robustness of the proposed method-

ology, indicating that neither overfitting nor underfitting
occurred while detecting HIF from models for which it was
not trained.
The noise resilience test was also performed. The results
revealed adequate performance for systems with up to
20 dB SNR, demonstrating reliability for realistic fault
scenarios. The increase in processing power and, with it, the
development of new classification techniques are motivating
factors for the proposition of more sophisticated and robust
HIF detection methods, such as the one proposed in this
FIGURE 16. Detailed Squeezenet classification performance for the paper.
complete data set with noise.
The next stage of HIF detection technique development
should consider more measurement points, a higher sample
test database. In the test, Detailed Squeezenet trained with rate, and harmonic synchrophasor measurements. The detec-
complete database perform the classification on complete tion method has the residual current as its input quantity.
database for test, with performance shown in Fig. 16. It has been verified that the approach is suitable for the
The classification method shows little degradation in cases considered. However, its applicability in other system
performance for a signal-to-noise ratio (SNR) of 30 dB topologies should be evaluated in future research, especially
and 20 dB. For this level of noise the method presents regarding the grounding scheme.
performance, robustness and reliability consistent with real-
world applications. REFERENCES
[1] A. A. Bayod-Rújula, ‘‘Future development of the electricity systems with
VI. CONCLUSION distributed generation,’’ Energy, vol. 34, no. 3, pp. 377–383, Mar. 2009.
In this paper, a new classifier is proposed for the problem of [2] S. Parhizi, H. Lotfi, A. Khodaei, and S. Bahramirad, ‘‘State of the art
in research on microgrids: A review,’’ IEEE Access, vol. 3, pp. 890–925,
HIF detection in MGs. An emphasis is given to the impact of 2015.
HIF modeling and DGs on classification performance when [3] A. Ghaderi, H. L. Ginn, and H. A. Mohammadpour, ‘‘High impedance fault
considering harmonic distortions caused by VSC based DG. detection: A review,’’ Electric Power Syst. Res., vol. 143, pp. 376–388,
The simulations consider detailed models of DG and Feb. 2017.
[4] M. Mishra and R. R. Panigrahi, ‘‘Taxonomy of high impedance fault
three different HIF models. The current measurements for detection algorithm,’’ Measurement, vol. 148, Dec. 2019, Art. no. 106955.
all three phases are aggregated into a single time series [5] B. K. Chaitanya, A. Yadav, and M. Pazoki, ‘‘An intelligent detection of
by residual current calculation. The tested transient events high-impedance faults for distribution lines integrated with distributed
generators,’’ IEEE Syst. J., vol. 14, no. 1, pp. 870–879, Mar. 2020.
were classified into HIF or nonHIF by a computationally
[6] P. K. Nayak, K. Sarwagya, and T. Biswal, ‘‘A novel high impedance fault
sophisticated pattern recognition algorithm employing RGB detection technique in distribution systems with distributed generators,’’ in
images, which represent the time-frequency properties of Proc. Nat. Power Syst. Conf. (NPSC), Dec. 2016, pp. 1–6.
the measured current signal. Five evaluation indexes, which [7] S. Gautam and S. M. Brahma, ‘‘Detection of high impedance fault in power
distribution systems using mathematical morphology,’’ IEEE Trans. Power
represent a more comprehensive set of evaluation criteria, are Syst., vol. 28, no. 2, pp. 1226–1234, May 2013.
employed to show the classification output. [8] D. A. Gadanayak and R. K. Mallick, ‘‘Interharmonics based high
To enhance the performance of HIF detection methods in impedance fault detection in distribution systems using maximum overlap
MG, it is necessary to consider detailed simulation models, wavelet packet transform and a modified empirical mode decomposition,’’
Int. J. Electr. Power Energy Syst., vol. 112, pp. 282–293, Nov. 2019.
as suggested by the results. A harmonic distortion-free [9] M. Biswal, M. Mishra, V. K. Sood, R. C. Bansal, and A. Y. Abdelaziz,
environment is not a realistic assumption for MG with VSC ‘‘Savitzky-Golay filter integrated matrix pencil method to identify high
based DGs. Also, the HIF model should reproduce this impedance fault in a renewable penetrated distribution system,’’ Electric
Power Syst. Res., vol. 210, Sep. 2022, Art. no. 108056.
phenomenon as accurately as possible in order to achieve
[10] M. Bhatnagar, A. Yadav, and A. Swetapadma, ‘‘A resilient protection
good performance in real applications. scheme for common shunt fault and high impedance fault in distri-
The results of the present study emphasize the need bution lines using wavelet transform,’’ IEEE Syst. J., vol. 16, no. 4,
for using more accurate simulation models to improve the pp. 5281–5292, Dec. 2022.
[11] M. Biswal, S. Ghore, O. P. Malik, and R. C. Bansal, ‘‘Development of time-
effectiveness of HIF identification. The accuracy of machine frequency based approach to detect high impedance fault in an inverter
learning algorithms, in particular, is strongly influenced by interfaced distribution system,’’ IEEE Trans. Power Del., vol. 36, no. 6,
the integrity of the training data, highlighting the critical role pp. 3825–3833, Dec. 2021.
[12] M. Manohar, E. Koley, and S. Ghosh, ‘‘Microgrid protection against
that high-quality data plays in achieving reliable and effective high impedance faults with robustness to harmonic intrusion and weather
outcomes. intermittency,’’ IET Renew. Power Gener., vol. 15, no. 11, pp. 2325–2339,
Two different HIF models and non-HIF transients were Aug. 2021.
used in a series of five tests. For all of the evaluated [13] D. R. R. Penido, L. R. de Araujo, V. T. S. Rodrigues, and
K. B. do Nascimento, ‘‘An analytical zero sequence method to locate fault
conditions, the proposed classifier outperformed a recent in distribution systems rich in DG,’’ IEEE Trans. Smart Grid, vol. 13,
HIF detection methodology, demonstrating its efficiency. The no. 3, pp. 1849–1859, May 2022.

VOLUME 11, 2023 89559

 V. F. Couto, M. Moreto: HIF Detection on Microgrids Considering the Impact of VSC Based Generation

[14] M. Wei, W. Liu, H. Zhang, F. Shi, and W. Chen, ‘‘Distortion-based [33] A. E. Emanuel, D. Cyganski, J. A. Orr, S. Shiller, and E. M. Gulachenski,
detection of high impedance fault in distribution systems,’’ IEEE Trans. ‘‘High impedance fault arcing on sandy soil in 15 kV distribution feeders:
Power Del., vol. 36, no. 3, pp. 1603–1618, Jun. 2021. Contributions to the evaluation of the low frequency spectrum,’’ IEEE
[15] A. Ahmadi, E. Aghajari, and M. Zangeneh, ‘‘High-impedance fault Trans. Power Del., vol. 5, no. 2, pp. 676–686, Apr. 1990.
detection in power distribution grid systems based on support vector [34] Q. Cui, K. El-Arroudi, and G. Joos, ‘‘An effective feature extraction method
machine approach,’’ Electr. Eng., vol. 104, no. 5, pp. 3659–3672, in pattern recognition based high impedance fault detection,’’ in Proc. 19th
Oct. 2022. Int. Conf. Intell. Syst. Appl. to Power Syst. (ISAP), Sep. 2017, pp. 1–6.
[16] F. Mumtaz, K. Imran, S. B. A. Bukhari, K. K. Mehmood, A. Abusorrah, [35] T. M. Lai, L. A. Snider, E. Lo, and D. Sutanto, ‘‘High-impedance
M. A. Shah, and S. A. A. Kazmi, ‘‘A Kalman filter-based protection fault detection using discrete wavelet transform and frequency range and
strategy for microgrids,’’ IEEE Access, vol. 10, pp. 73243–73256, RMS conversion,’’ IEEE Trans. Power Del., vol. 20, no. 1, pp. 397–407,
2022. Jan. 2005.
[17] A. G. Rameshrao, E. Koley, and S. Ghosh, ‘‘A LSTM-based approach for [36] A. Aljohani and I. Habiballah, ‘‘High-impedance fault diagnosis:
detection of high impedance faults in hybrid microgrid with immunity A review,’’ Energies, vol. 13, no. 23, p. 6447, Dec. 2020. [Online].
against weather intermittency and N-1 contingency,’’ Renew. Energy, Available: https://www.mdpi.com/1996-1073/13/23/6447
vol. 198, pp. 75–90, Oct. 2022. [37] A. R. Sedighi and M. R. Haghifam, ‘‘Simulation of high impedance ground
[18] É. M. Lima, R. D. A. Coelho, N. S. D. Brito, and B. A. D. Souza, ‘‘High fault in electrical power distribution systems,’’ in Proc. Int. Conf. Power
impedance fault detection method for distribution networks under non- Syst. Technol., Oct. 2010, pp. 1–7.
linear conditions,’’ Int. J. Electr. Power Energy Syst., vol. 131, Oct. 2021, [38] A. Ghaderi, H. A. Mohammadpour, H. L. Ginn, and Y.-J. Shin, ‘‘High-
Art. no. 107041. impedance fault detection in the distribution network using the time-
frequency-based algorithm,’’ IEEE Trans. Power Del., vol. 30, no. 3,
[19] M. Wei, F. Shi, H. Zhang, and W. Chen, ‘‘Wideband synchronous
pp. 1260–1268, Jun. 2015.
measurement-based detection and location of high impedance fault for
[39] M. Eslami, M. Jannati, and S. S. Tabatabaei, ‘‘An improved protection strat-
resonant distribution systems with integration of DERs,’’ IEEE Trans.
egy based on PCC-SVM algorithm for identification of high impedance
Smart Grid, vol. 14, no. 2, pp. 1117–1134, Mar. 2023.
arcing fault in smart microgrids in the presence of distributed generation,’’
[20] V. Veerasamy, N. I. A. Wahab, M. L. Othman, S. Padmanaban, K. Sekar, Measurement, vol. 175, Apr. 2021, Art. no. 109149. [Online]. Available:
R. Ramachandran, H. Hizam, A. Vinayagam, and M. Z. Islam, ‘‘LSTM https://www.sciencedirect.com/science/article/pii/S0263224121001731
recurrent neural network classifier for high impedance fault detection in
[40] S. J. Pan and Q. Yang, ‘‘A survey on transfer learning,’’ IEEE Trans. Knowl.
solar PV integrated power system,’’ IEEE Access, vol. 9, pp. 32672–32687,
Data Eng., vol. 22, no. 10, pp. 1345–1359, Jan. 2009.
2021.
[41] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally,
[21] S. Kar and S. R. Samantaray, ‘‘Time-frequency transform-based and K. Keutzer, ‘‘SqueezeNet: AlexNet-level accuracy with 50x fewer
differential scheme for microgrid protection,’’ IET Gener., Transmiss. parameters and <0.5 MB model size,’’ 2016, arXiv:1602.07360.
Distrib., vol. 8, no. 2, pp. 310–320, Feb. 2014. [Online]. Available: [42] Y. Zhang, X. Wang, J. He, Y. Xu, F. Zhang, and Y. Luo, ‘‘A transfer
https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/iet- learning-based high impedance fault detection method under a cloud-edge
gtd.2013.0180 collaboration framework,’’ IEEE Access, vol. 8, pp. 165099–165110, 2020.
[22] S. Kar and S. R. Samantaray, ‘‘High impedance fault detection in microgrid [43] Y. Zhang, X. Wang, Y. Luo, Y. Xu, J. He, and G. Wu, ‘‘A CNN based
using maximal overlapping discrete wavelet transform and decision tree,’’ transfer learning method for high impedance fault detection,’’ in Proc.
in Proc. Int. Conf. Electr. Power Energy Syst. (ICEPES), Dec. 2016, IEEE Power Energy Soc. Gen. Meeting (PESGM), Aug. 2020, pp. 1–5.
pp. 258–263. [44] R. Fan and T. Yin, ‘‘Convolutional neural network and transfer learning for
[23] M. Mishra and P. K. Rout, ‘‘Detection and classification of micro-grid high impedance fault detection,’’ 2019, arXiv:1904.08863.
faults based on HHT and machine learning techniques,’’ IET Gener.,
Transmiss. Distrib., vol. 12, no. 2, pp. 388–397, Jan. 2018.
[24] H. Laaksonen and P. Hovila, ‘‘Enhanced MV microgrid protection
scheme for detecting high-impedance faults,’’ in Proc. IEEE Manchester
PowerTech, Jun. 2017, pp. 1–6.
[25] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, ‘‘ImageNet: VITOR FERNANDO COUTO received the B.E.
A large-scale hierarchical image database,’’ in Proc. IEEE Conf. Comput. and M.E. degrees in electrical engineering (power
Vis. Pattern Recognit., Jun. 2009, pp. 248–255. systems) from the Regional University of Blu-
[26] A. A. Memon and K. Kauhaniemi, ‘‘A critical review of AC microgrid menau, Brazil, in 2011 and 2013, respectively.
protection issues and available solutions,’’ Electric Power Syst. Res., He is currently pursuing the Ph.D. degree in
vol. 129, pp. 23–31, Dec. 2015. electrical engineering with the Federal University
[27] A. Soheili, J. Sadeh, and R. Bakhshi, ‘‘Modified FFT based high of Santa Catarina, Florianópolis, Brazil. He is
impedance fault detection technique considering distribution non-linear an Innovation Researcher with the WEG Group.
loads: Simulation and experimental data analysis,’’ Int. J. Electr. Power His research interests include pattern recognition,
Energy Syst., vol. 94, pp. 124–140, Jan. 2018. [Online]. Available: signal processing, and artificial intelligence for
https://www.sciencedirect.com/science/article/pii/S0142061517306555 fault identification in electrical power systems.
[28] Q.-M. Xiao, M.-F. Guo, and D.-Y. Chen, ‘‘High-impedance fault detection
method based on one-dimensional variational prototyping-encoder for
distribution networks,’’ IEEE Syst. J., vol. 16, no. 1, pp. 966–976,
Mar. 2022.
[29] CIGRE Task Force C6.04.01. (2014). Benchmark Systems for Network
Integration of Renewable and Distributed Energy Resources. CIGRE. MIGUEL MORETO received the B.Sc. and M.Sc.
[Online]. Available: http://www.e-cigre.org/Order/select.asp?ID=16639 degrees in electrical engineering from the Fed-
[30] R. Azim, F. Li, Y. Xue, M. Starke, and H. Wang, ‘‘An islanding eral University of Rio Grande do Sul, Porto
detection methodology combining decision trees and Sandia frequency Alegre, Brazil, in 2003 and 2005, respectively,
shift for inverter-based distributed generations,’’ IET Gener., Trans- and the D.Sc. degree in electrical engineering
miss. Distrib., vol. 11, no. 16, pp. 4104–4113, Nov. 2017. [Online]. from the Federal University of Santa Catarina,
Available: https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/iet- Florianópolis, Brazil, in 2011. He is currently
gtd.2016.1617 an Adjunct Professor with the Federal University
[31] A. Sultan and G. Swift, ‘‘Security testing of high impedance fault of Santa Catarina. His research interests include
detectors,’’ in Proc. WESCANEX, 1991, pp. 191–197. computational intelligence applications to power
[32] A. Soheili, J. Sadeh, H. Lomei, and K. Muttaqi, ‘‘A new high impedance systems, disturbance diagnosis, protection systems, and power systems
fault detection scheme: Fourier based approach,’’ in Proc. IEEE Int. Conf. signal processing and synchrophasor measurements in distribution systems.
Power Syst. Technol. (POWERCON), Sep. 2016, pp. 1–6.

89560 VOLUME 11, 2023