Journal of Autonomous Intelligence (2024) Volume 7 Issue 1

doi: 10.32629/jai.v7i1.1124

Original Research Article

Power grid monitoring based on Machine Learning and Deep
Learning techniques
Marco Bindi*, Carlos Iturrino-García, Maria Cristina Piccirilli, Francesco Grasso, Antonio Luchetta,
Libero Paolucci
Department of Information Engineering, University of Florence, 50139 Florence, Italy
* Corresponding author: Marco Bindi, m.bindi@unifi.it

ABSTRACT
Background: In this work, some application examples of machine learning and deep learning techniques in the
monitoring of electricity distribution services and infrastructures are proposed. Three different fields of application are
considered to highlight the use of techniques based on neural networks: detection and classification of power quality
disturbances, monitoring of underground cables in medium voltage lines and diagnosis of joints in high voltage overhead
power lines. Methods: In the field of power grid monitoring, this work proposes a classification method based on a
complex valued neural network to assess working conditions of junction regions in high-voltage overhead lines and
insulating materials in medium voltage underground networks. The purpose of this method is to prevent the rupture of
joint structures and the abnormal degradation of underground cables via frequency response measurements. This approach
allows the direct processing of complex measurements and to reduce the computational effort compared to other methods
available in the literature. Results: The results obtained in the monitoring of underground cables and joints of high voltage
lines guarantee an overall classification rate higher than 90%. In the field of power quality, several deep learning and
machine learning methods are proposed to detect the most common voltage disturbances. Conclusions: In this paper, an
innovative use of widespread algorithms such as convolutional neural networks is proposed with excellent results.
Furthermore, the use of a complex-valued neural network in electrical infrastructure monitoring is presented, introducing
a minimally invasive classification method that could be instrumental in the transition from corrective to predictive
maintenance in the near future.
Keywords: power quality disturbances; machine learning; neural networks; medium voltage cables; fault diagnosis; high
voltage systems

ARTICLE INFO 1. Introduction

Received: 8 August 2023 This paper proposes new approaches to monitor electricity
Accepted: 7 September 2023
Available online: 27 October 2023 distribution services, as this is one of the most important activities to
COPYRIGHT
ensure the continuity and quality of electricity grids. Electrical
Copyright © 2023 by author(s).
continuity is compromised by fault situations, while power quality is a
Journal of Autonomous Intelligence is problem that, in recent years, is becoming increasingly important [1].
published by Frontier Scientific Publishing.
This work is licensed under the Creative
One of the main causes is the growing use of electronic devices, which
Commons Attribution-NonCommercial 4.0 can be connected to the network both as loads and as energy injectors[2–4].
International License (CC BY-NC 4.0).
https://creativecommons.org/licenses/by-
The integration of renewable energy generators significantly increases
nc/4.0/ the complexity of electrical transmission and distribution systems, the
maintenance of which requires the introduction of new techniques from
other fields to detect and locate malfunctions[5–7]. Most of these new
techniques use artificial intelligence algorithms to prevent catastrophic
consequences; for example, in the research of Bindi et al.[8] and in the
work of Hayder and Saidi[9], possible approaches to avoid failures in

1
photovoltaic applications are presented. On the other hand, the widespread diffusion of electronic devices in
any sector, from industry to the domestic, from commerce to the public sector, has improved the efficiency and
accuracy of numerous services and the automation of many different processes. Unfortunately electronic
devices have a nonlinear behavior that can produce distortions in voltage and current waveforms of the power
grid with a consequent degradation of its quality. Detection and classification of these disturbances is important
to avoid degradation of electrical components and power losses that lead to losses in revenue.
Another fundamental aspect to ensure the continuity of the electricity distribution service is the
monitoring of underground cables. Medium voltage lines near urban centers generally consist of underground
cables. Given the great difficulty of accessing these cables, the possibility of checking their state of health
without removing them plays a fundamental role in organizing maintenance operations [10].
In the case of high voltage overhead lines are concerned, the prevention of catastrophic events and the
rapid localization of the damage are essential to reduce recovery times and organize maintenance interventions.
This is not easy due to the length of this kind of lines[11].
Depending on the problem to be treated, it is necessary to apply the most appropriate computational
technique. For example, in order to detect and classify disturbances affecting grid power quality, various
techniques are currently exploited such as Kalman filters[12] and wavelet transforms[13].
There are several fault location techniques in underground lines, but many of them cannot be used online
during normal network operation[14,15]. One of these methods[16] is based on the measurement of dielectric
losses due to the degradation of the insulating material, but requires measurements to be performed in the
laboratory. Other methods are based on the reflection of the measurement signals and are able to locate the
point of the conductor with an impedance value different from the nominal one[17,18]. This group of techniques,
called Time Domain Reflectometry (TDR), does not require pulling out cables, but it does not prevent
insulation problems.
Regarding the transmission lines, there are various methods that can be used to monitor the operational
state of overhead conductors, such as Load Flow analysis and state estimators[19–21]. Although these methods
can identify incorrect behavior, they may not be able to determine the type and location of the malfunction.
Given that joints are one of the most critical components, it is essential to have knowledge of their level of
deterioration. In order to achieve this, TDR techniques can be employed to remotely detect the impedance of
the joint by analyzing signals that are reflected from the areas of the conductor where there are
discontinuities[22,23]. However, the accuracy of this technique can be affected by measurement errors,
environmental noise, line length, and any changes to the environmental conditions along the line.
One effective approach to addressing the issues mentioned earlier is to utilize machine learning methods.
In addition to automating grid monitoring, these techniques can help resolve problems associated with
conventional procedures, such as those for underground cables and overhead transmission lines. They employ
an adaptive pattern classification mechanism, which enables them to perform a robust classification even in
the presence of unclear system models and noisy environments. Following an initial training phase, a neural
network can extrapolate the acquired information to make generalizations.
The main innovative contributions of this paper can be summarized as follows:
 propose the use of Long Short-Term Memory (LSTM) algorithms, Convolutional Neural Networks (CNN) and
the mixed architectures LSTM-CNN to detect the most common electrical disturbances in low voltage
electrical grids;
 present a possible configuration of a complex neural network to obtain the detection and localization of
insulation problems in medium voltage underground cables;
 improve a minimally invasive monitoring method for assessing the health of joints in high voltage overhead

2
 lines.
The paper is organized in the following way. The problem of power quality disturbance detection and
classification is faced in Section 2. The monitoring of underground cables in medium voltage lines is
considered in Section 3, and the diagnosis of joints in high voltage electrical lines in Section 4. Finally, Section
5 contains the conclusions and some ideas for future work.

2. Power quality analysis

Power quality is a critical aspect of modern electrical systems, with implications for the safe and reliable
operation of all types of electrical equipment. The IEEE defines Power Quality (PQ) as: “The concept of
powering and grounding sensitive equipment in a matter that is suitable to the operation of that equipment" In
other words, power quality refers to the consistency and reliability of the voltage, frequency, and waveform of
the electrical power delivered to a facility or device. Any variations or disturbances in power quality can result
in equipment damage, data loss, or even system failure, leading to significant economic costs and potential
safety risks. As our reliance on electronic devices and sensitive equipment continues to grow, the need to
maintain high power quality has become increasingly important[24,25]. Optimal power quality implies standard
voltage and current values and a non-deviation from frequency of voltage and current signals in the grid. Power
quality disturbances can affect the normal functioning of electrical components such as transformers, capacitor
banks, power lines, electric motors, household electronics, etc. This in turn affects everyday life as well as
industrial processes. These disturbances include frequency component injection, amplitude deviation,
waveform and symmetry distortion of the three-phase voltages and currents. Their identification and
classification are important in order to counteract and find the origin of the distortion. This can be achieved by
using specialized instrumentation for data acquisition of voltage and current signals on the power grid. In the
context of power quality, machine learning algorithms can be used to identify patterns and anomalies in
electrical power systems. They have proved to be accurate and precise when dealing with detection and
classification of disturbances. Several of these algorithms, using different architectures and feature extraction
in the time domain and/or the frequency domain, are currently exploited to carry out this task[26,27].
The most frequent power quality disturbances in low voltage electrical networks can be divided into three
main groups: amplitude fluctuations, sudden transients and steady state harmonic pollution. A more detailed
classification of disturbances includes voltage sag, voltage swell, harmonic disturbances, interruption, notch
and transient. For each kind of these disturbances a specific machine learning algorithm can be more efficiently
used, so a comparison among different kinds of techniques is very useful. For example, in the study of Iturrino
et al.[28] the effectiveness of three deep learning architectures for power quality disturbance detection and
classification is investigated. They are the Long Short-Term Memory (LSTM), the Convolutional Neural
Networks (CNN) and the mixed architecture LSTM-CNN. To compare the different techniques, the work
proposed by Iturrino et al.[28] focuses on the different feature extraction techniques of the different architectures
and their effectiveness on classifying power quality disturbances.
Also, deep learning architectures trained with simulated data have proven to be successful in detecting
and classifying different types of disturbances in an experimental setup. For this reason, a Matlab®/Simulink
model has been built to simulate the voltage and current time signals on a micro grid including several different
industrial loads (a three-phase dynamic load, connected to an electrical engine with variable load, a linear load,
a nonlinear load injecting disturbances). Through the simulation, a current and voltage dataset was created for
training and validating the classification algorithms. These disturbances include: under voltage (sag), over
voltage (swell), harmonics distortions, transient, notching and interruption. To create a balance dataset and
optimize the performance of deep learning algorithms it is necessary to reproduce different disturbances by
varying amplitude, duration, and intensity of the signal window. To improve the performance of classification

3
and converge to a generalized result, the data in the simulated dataset was augmented. After training with the
augmented dataset, the algorithms have been tested using experimental voltage measurements containing the
disturbances previously mentioned. For testing to be successful, an experimental testbench has been created in
order to reproduce and measure several real-time disturbances. In Figures 1–4, the results of the classification
using experimental data of the different classifiers previously mentioned are shown.

Figure 1. Classification results. LSTM real time classification.

Figure 2. Classification results. CNN real time classification.

4
 Figure 3. Classification results. LSTM-CNN real time classification.

Figure 4. Classification results. LSTM-CNN with optimized hyperparameters real time classification.

This experiment was executed using a sliding window method and classifying sections of the measured
signal in an orderly fashion. In Figures 5 and 6, a bar chart containing the recall and precision of the different
deep learning architectures classifying different power quality disturbances is shown. In this test, the different
classifiers are compared using the measured recall and precision.
In the research of Iturrino et al.[29] a new type of algorithm was developed called the Single Shot Power
Quality Disturbance Detection (SSPQDD) algorithm. This algorithm is based on an object detection algorithm
for image detection and classification. Voltage signals can contain a sequence of disturbances in a single
window frame of a certain duration. The SSPQDD is a pretrained VGG-16 architecture modified with two
outputs in order to accurately detect and classify multiple disturbances in a single window frame. In Figure 7,
a comparison of the SSPQDD with different classical deep learning architectures is shown. It shows the
SSPQDD was able to correctly detect and classify the sequence of disturbances in a single window frame
where the other architectures misclassified.

5
 Figure 5. Comparison. Precision chart of the classifiers.

Figure 6. Comparison. Recall chart of the classifiers.

Figure 7. Test results. Classification results of the SSPQDD of a signal with three disturbances.

6
3. Localization of malfunctions in underground cables
Monitoring medium voltage lines can prevent power outages for thousands of users. Given the
underground position of the cables, it is necessary to adopt specific techniques, usually based on artificial
intelligence techniques, to prevent catastrophic failures. This can be achieved by monitoring the working
temperature of successive cable sections, as temperature increases can indicate insulation problems [30].
Malfunctions can be localized by identifying the section with a higher temperature. If current overload and/or
an unusual increase in ambient temperature occur, their duration must be monitored to prevent degradation of
the cable insulation[31]. A complex neural network can be used to identify the working temperature of several
cable sections by measuring the frequency response of the line[32]. In this paper a possible application of the
complex classifier is proposed, processing measurements of the line voltage gain for different frequencies. To
achieve the purpose of monitoring and locating the part of the line in the worst conditions, the network branch
is divided into successive sections. Each of these sections is represented by a π-model shown in Figure 8.

Figure 8. Equivalent circuit. Lumped circuit of a single cable section.

Since the resistance of the conductor increases with temperature and the change in resistance introduces
a corresponding change in the frequency response, its measurements contain all the information about the
working temperature. Thus, the neural classifier processes the magnitude and phase of the transfer function
measurements.
The dataset matrix required for the classifier training phase contains numerous samples of each fault class
obtained by randomly varying the temperature of the cable sections. These changes are converted to resistance
values using Equation (1) and Equation (2). Subsequently, the values of the resistors are used in the model of
the line.
𝜌𝑇 = 𝜌20 [1 + 𝛼(𝑇 − 20)] (1)
where the resistivity 𝜌20 refers to the conductor material at 20°C, 𝑇 is the working temperature and 𝜌𝑇 is
the corresponding resistivity. Finally, α is the thermal coefficient of the conductor material.
′
𝜌𝑇
𝑅𝑙𝑇 = (2)
𝑆
where S represents the cross-section of the cable. The dataset used during the training phase contains several
samples of each possible fault class. Assuming a maximum level of working temperature equal to 105 ℃[33]
and considering the phase current approximately equal to the rated current of the cable, with the ambient
temperature equal to 25 ℃, the operating temperature of the cable is approximately 70 °C [34]. Two possible
working conditions are considered for each cable section: an operating temperature below 70 ℃ represents the
nominal condition, while a cable temperature above 70 ℃ indicates the presence of a thermal malfunction.
The minimum cable temperature considered in this paper is 30 ℃ and describes the minimum load condition.
The overheat situation is 135 ℃ and indicates that there is a problem.

3.1. Case study

A 900-meter medium voltage (20 kV) underground line consisting of three RG7H1M1 single core cables
with a 35mm2 section represents the case study analyzed here. This cable has a rated current of 205 A and the

7
whole line is theoretically divided into three successive sections. Furthermore, the hypothesis of a single fault
is assumed and this means that overheating can only affect a single section at a time. This means that there are
four error classes: the first (class 0) describes the nominal operating condition of each part of the cable, the
other 3 classes describe the thermal malfunction of each section of the line (n = 1, 2, 3). Table 1 summarizes
the characteristics of a single piece of cable and Figure 9 shows the case study.

Table 1. Characteristics of a single cable part.

Cable length [km] Conductor radius [mm] Sheath radius [mm] Outer radius [mm]
0.3 7 17 23.2
Line-line spacing [mm] Line formation Insulation relative permittivity Cable section [mm2]
46.4 Flat 2.4 35

Figure 9. Underground line. Medium voltage underground line divided into three sections.

As previously mentioned, the main physical phenomenon considered in this application is the relation
between the working temperature and cable resistance (only the longitudinal resistance is the variable element):
as the operating temperature increases, the resistance of the conductor also increases. When a large change in
conductor resistance occurs, there is also a change in frequency response. This means that the transfer function
contains all the information on the working temperature of the cable. To obtain the frequency response
measurements it is necessary to inject high-frequency signals into the line using the typical tools of Power Line
Communication. Magnitude and phase of the voltage transfer function are processed by the neural classifier to
detect any thermal malfunctions.
In this application, dataset creation is completely done on Matlab/Simlinks. SapWin software is used to
verify Testability of the overall circuit[35], and the result obtained indicates the possibility of distinguishing the
four fault classes through the selected measure.

3.2. Simulation procedure

Once all the possible working conditions have been defined, it is necessary to create the dataset matrix.
As previously mentioned, a Simulink model is used, where the transmission of signals is simulated by
introducing a PLC system. The turn ratio of the matching transformers is fixed to adapt the transmitter/receiver
resistance with the characteristic line impedance. The coupling circuit used in this case is capacitive and
presents a fourth order high pass filter with a minimum bandwidth frequency of 3 kHz. The injection of signals
in the (3 ÷ 50) kHz band is considered. To simulate all the possible fault classes, the Simulink/Simscape scheme
is managed by a Matlab script, which varies the resistance values. The corresponding voltage transfer function
is measured in magnitude and phase for 𝑛𝑓 test frequencies belonging to the band taken into consideration.
In this procedure, two different approaches are compared: the first involves the selection of 4 test
frequencies while the second is based on a Principal Component Analysis (PCA). In the first case, 4 frequencies
obtained by dividing the total bandwidth into three equal parts are used in the Simulink/Simscape diagram to

8
extract the transfer function measurements and create a dataset containing 300 random samples for each fault
class. Next, 100 frequencies are considered and a PCA is applied to reduce the number of inputs. In this way
it is not necessary to choose the test frequencies beforehand. Again 1200 examples are generated from the four
failure classes at the test frequencies. A linear PCA is applied to the dataset matrix containing the measurements
relating to 100 frequencies, maintaining the 98% of the informative content. This produces the reduction of
the dimensionality of each input from 100 complex numbers to 4. This means that two dataset are available to
classify the state of health of the line using four complex numbers for each input. In the first case, these
numbers correspond to magnitudes and phases of the measured transfer function. In the second case they are
obtained following the projections of the measurements obtained for 100 frequencies in a space with reduced
dimensionality. In the new coordinate system, the physical meaning of the measurements is not maintained,
but a certain percentage of the information content is preserved (98% in this case).

3.3. Neural network setup and results

In this case, the output configuration of the complex neural network is based on the single failure
hypothesis. Therefore, three binary neurons are used in the output layer, one for each cable section taken into
consideration. Figure 10 illustrates the structure of the output layer of the neural network.

Figure 10. Neural network structure. Configuration of the output layer of the complex neural network: use of a binary neuron for
each section of cable.

These neurons are used to distinguish the nominal condition of the corresponding cable section from the
over temperature situation. In particular, the value 0 corresponding to the beginning of the upper half plane is
used to indicate the nominal working temperature, while the lower border of the second sector indicates the
overheating. As previously mentioned, two datasets are used in this case to evaluate the performance of the
classifier. In the first case, the transfer function measurements obtained at 4 test frequencies are used. The
results obtained in terms of the overall classification rate are shown in Figure 11a, while Figure 11b presents
the performance of the complex neural network for each class separately. Initially, hold out validation is used
to evaluate the classification results: this means that 80% of the data is used to modify the weights of the neural
network (learning) while the remaining 20% is used to verify performance (testing). Note that the results
presented are obtained by setting 20 neurons in the hidden layer. In Figure 11a, the red line is used to describe
the classification rate obtained during the learning phase, while the blue one refers to the testing phase. Finally,
the training procedure is repeated by applying a linear PCA to the simulated measurements obtained at 100
frequencies. By maintaining 98% of the information content, the results shown in Figure 12a,b are obtained.
The performance obtained by keeping 20 neurons in the hidden layer is slightly better than that obtained in the
previous case. Therefore, the use of a linear PCA can be considered a valid alternative to the frequency
selection. Instead of implementing a rather complex mathematical method, it is possible to perform

9
measurements at multiple frequencies and then reduce the dimensionality of the inputs to avoid a heavy
structure of the neural classifier. To obtain a comparable result using all the measurements obtained at 100
frequencies, it is necessary to use a number of neurons in the hidden layer 3 times higher. Table 2 summarizes
these considerations, where No PCA indicates the use of four test frequencies selected through the
mathematical method, Linear PCA indicates the use of a linear PCA to the dataset containing simulated
measurements at 100 frequencies and 100 Frequencies indicates direct use of this dataset.

Figure 11. Classification results. Results obtained using 4 specific frequencies: (a) global classification rate; (b) classification rate
for each output neuron.

Figure 12. Classification results. Results obtained by PCA: (a) global classification rate; (b) classification rate for each output
neuron.

Table 2. Classification results obtained by selecting 4 frequencies and using a linear PCA.
Configuration Classification rate (CR%) N. hidden neurons N. complex inputs
No PCA 92.92 20 4
Linear PCA 98.75 20 4
100 Frequencies 96.70 60 100

4. Prognosis of joints in high voltage electrical lines

To identify the operating state of high-voltage overhead transmission lines, it is important to focus on the

10
joints (as shown in Figure 13), which are responsible for maintaining electrical continuity along the phase
conductor. Joints are among the most vulnerable components of the line, and problems such as voltage drops,
power losses, and decreased power quality may be caused by their deterioration. The physical characteristics
of the joints are described through three terms: Δ𝜆 is the length, 𝑑 is the thickness and 𝐻 is the height. Then,
when a partial break occurs, the term 𝑥 is used to indicate the thickness of the break and the ℎ term for its
height.
For high voltage electrical lines, predictive maintenance requires assessing the health of the joints and
taking preventive measures to avoid any potential failures. A multi-layer feedforward neural network based on
multi-valued neurons[32] can be employed[36] for this purpose. The voltage transfer function is still used as input
for the neural network. Once nominal ranges have been established for the electrical parameters of the line
pattern, any deviation can be used to indicate a possible fault.
The monitoring action consists in the comparison between the theoretical frequency response and the
simulated measurements at multiple frequencies. The variation obtained from the comparison describes the
change in the parameters of the joints and therefore the departure from the nominal condition.

Figure 13. Joint structure. Joint modelling in nominal condition and in the case of a partial breakage.

The neural network is responsible for performing the task of classifying the severity of joint degradation.
The procedure for assessing the health of the joints can be summarized as follows.
First, the model of the elementary section of the line, consisting of the cascade connection of the conductor
model and the joint model (Figure 14), must be introduced. Indeed, the infrastructure can be seen as a sequence
of joints and sections of conductor. The conductor equivalent circuit typically used in the literature is the
canonical π model[21], whose parameters depend on the mechanical characteristics of the conductor and are
considered constant per unit length[37]. The variation of interest is that of the junction regions. Among the three
parameters of the joint model (𝑅𝑠𝑗 , 𝐿𝑠𝑗 , and 𝐶𝑠𝑗 ), the resistance parameter is most affected by the formation
of oxide, as this modifies the resistivity, while 𝐿𝑠𝑗 and 𝐶𝑠𝑗 account for any breakage in the joint[38]. Once the
model is established, the neural network can perform the task of evaluating the health of the joints.

Figure 14. Equivalent model. Elementary section of an overhead line.

11
 The second step involves determining the potential health conditions of each joint region, which can be
categorized into three possible classes: nominal, structure oxidation, and partial breakage. The ranges for each
joint parameter can be calculated based on the information extracted from the research of De Paulis et al.[22],
as shown in Table 3.

Table 3. Fault classes.

Nominal conditions
𝑅𝑠𝑗 [Ω] 60 ∙ 10−6 ÷ (4.78 ∙ 10−5 √𝑓 + 7.77 ⋅ 10−4 )
𝐿𝑠𝑗 [𝐻] (1.3 ÷ 1.7) ∙ 10−6
𝐶𝑠𝑗 [𝐹] (0.0099 ± 0.011) ∙ 10−12
Oxidation conditions
𝑅𝑠𝑗 [Ω] (4.78 ∙ 10−5 √𝑓 + 7.77 ⋅ 10−4 ) ÷ 2
𝐿𝑠𝑗 [𝐻] (1.3 ÷ 1.7) ∙ 10−6
𝐶𝑠𝑗 [𝐹] (0.0099 ± 0.011) ∙ 10−12
Partial breaking conditions
𝑅𝑠𝑗 [Ω] >2
𝐿𝑠𝑗 [𝐻] (1 ÷ 1.3) ∙ 10−6
𝐶𝑠𝑗 [𝐹] (0.011 ± 0.16) ∙ 10−12

The third step is focused on selecting an appropriate set of frequencies based on the procedure outlined
in the study of Grasso et al.[39]. The voltage transfer function of the model can be calculated using SAPWIN[40],
which provides its symbolic formulation and allows the calculation of the response at specific frequencies
within reasonable time limits. Next, a training set is created. The symbolic network function can be handled in
MATLAB to generate the training samples. In fact, a specific MATLAB script calculates the analytical
expressions of magnitude and phase and evaluates their values for each test frequency and for each line health
state. This is achieved by replacing symbolic parameters with their numeric values. Therefore, all training
samples are obtained by randomly varying the parameters of each joint in the respective classes. During this
operation, the parameters of the conductor sections are changed randomly within a range of 10% of their
nominal value. In this way, the effects of ambient conditions and load current can be taken into account.

Case study and results

The complex neural network used to obtain the classification results has two layers and uses a pair of
binary output neurons for each junction region. The first neuron of each pair detects the oxidation process
while the second the partial breakdown mechanism. Figure 15 shows the global structure of the neural
network-based classifier.

12
 Figure 15. Complex-valued neural network. Overall structure of the complex neural network.

Since these neurons have two possible levels of output, the low value is used to describe nominal
conditions. When the output of the first neuron is high and that of the second is low, the combination represents
joint oxidation. If, on the other hand, both outputs have a high level, the corresponding junction region has
partial damage. Since the hypothesis of the consequentiality between the oxidation mechanism and partial
failure is assumed, the remaining combination is not taken into consideration. In general, there are 𝑁𝑐 = 3𝑁𝑔
possible output combinations, since each junction region can be oxidized, broken or fully functional (where
𝑁𝑔 is the total number of joints,). Figure 16 illustrates the case study considered.

Figure 16. Case study. Overhead line with three junction regions.

The results reported in Table 4 are obtained for a line with three junction regions and show that the
classification method identifies the operating conditions of each joint with an accuracy level higher than 90%.

Table 4. Characteristics of a single cable part.

Fault classes Neuron CR% for each neuron Joint CR% for each joint
Oxidation 1 95.11 1 94.07
Partial breakage 2 99.37
Oxidation 1 92.85 2 91.30
Partial breakage 2 99.22
Oxidation 1 97.89 3 96.70
Partial breakage 2 99.81

13
5. Conclusion
The use of machine learning algorithms has shown significant potential for detecting power quality
disturbances in electrical systems. The research presented in this article has explored different machine
learning techniques for power quality analysis. They can accurately classify different types of power quality
disturbances, including voltage sags, swells, interruptions, harmonics, and transients. Moreover, the
localization and classification of power quality disturbances using the SSPQDD have been also considered.
The results of our study suggest that machine learning-based approaches can provide an efficient and
reliable way to monitor power quality and detect disturbances in real-time. By analyzing the electrical signals
from power systems, these algorithms can identify patterns and anomalies that may indicate power quality
issues. Moreover, the ability to detect and classify power quality events can provide valuable information for
power system operators to take preventive measures and improve system performance.
Despite the promising results, there are still challenges to overcome in implementing machine learning-
based power quality analysis in practice. These challenges include data availability and quality, and
interpretability. Further research is needed to address these challenges and optimize machine learning models
for power quality analysis.
In summary, the use of machine learning for power quality analysis has significant potential to improve
the safety, reliability, and efficiency of electrical systems. As the demand for high-quality power continues to
grow, machine learning can provide an effective solution for power quality monitoring and detection, enabling
early detection and mitigation of disturbances and reducing the risk of equipment damage and system failures.
Furthermore, the theoretical results presented in this paper offer a possible approach for the detection and
localization of malfunctions in transmission and distribution lines. The goal of the simulations performed was
to demonstrate that the MLMVN-based classifier allows the identification of the state of the line using
measurements of its frequency response. In the case of the prognosis of joints on overhead lines, the hypothesis
of multiple failure with multiple severities was adopted. In this situation, the method allows the correct
monitoring (classification rate higher than 90%) of the joints contained in a branch of 800 meters. Exceeding
this limit, the performance in recognizing the operating conditions of all joints decreases. It is still possible to
study a single joint with good accuracy, but the overall performance is not sufficient. As regards the monitoring
of underground lines, the single failure hypothesis was assumed and the prognostic method was focused on
the detection of overheating. In this case, the classifier allows the correct monitoring (accuracy greater than
95%) of approximately 1 km of line divided into three successive sections. The increase in the length of the
line branch considered and therefore in the number of successive sections determines the reduction of the
classification rate and requires further developments.

Author contributions
Conceptualization, MB; methodology, CIG; software, MB and CIG; validation, AL, FG and LP; formal
analysis, MCP; investigation, LP; data curation, LP and CIG; writing—original draft preparation, MB;
writing—review and editing, MCP; supervision, AL and FG. All authors have read and agreed to the published
version of the manuscript

Acknowledgments
We would like to express our sincere gratitude to the Smart Energy Lab for providing us with the
necessary tools and resources for conducting our tests. Their support and collaboration have been invaluable
in enabling our research on energy efficiency and smart grid technologies (https://www.smartenergylab.eu/).

14
Conflict of interest
The authors declare no conflict of interest.

Abbreviations
TDR Time Domain Reflectometry
LSTM Long Short-Term Memory
CNN Convolutional Neural Networks
SSPQDD Single Shot Power Quality Disturbance Detection
PQ Power Quality
PCA Principal Component Analysis
MLMVN Multi-Layer neural network with Multi-Valued Neurons

References
1. Bindi M, Luchetta A, Paolucci L, et al. Applications of machine learning techniques for the monitoring of
Electrical transmission and distribution lines. In: Proceedings of the 2022 18th International Conference on
Synthesis, Modeling, Analysis and Simulation Methods and Applications to Circuit Design (SMACD); 12–15 June
2022; Villasimius, Italy. pp. 1–4.
2. Wang L, Qin Z, Slangen T, et al. Grid impact of electric vehicle fast charging stations: trends, standards, issues and
mitigation measures—An overview. IEEE Open Journal of Power Electronics 2021; 2: 56–74, doi:
10.1109/OJPEL.2021.3054601
3. Rafiqi IS, Bhat AH. Role of UPQC in addressing power quality issues in the power grid connected with renewable
energy sources. In: Proceedings of the 2022 1st International Conference on Sustainable Technology for Power
and Energy Systems (STPES); 4–6 July 2022; Srinagar, India. pp. 1–4.
4. Kanjiya P, Khadkikar V. Enhancing power quality and stability of future smart grid with intermittent renewable
energy sources using electric springs. In: Proceedings of the 2013 International Conference on Renewable Energy
Research and Applications (ICRERA); 20–23 October 2013; Madrid, Spain. pp. 918–922.
5. Auler LF, d'Amore R. Power quality monitoring controlled through low-cost modules. IEEE Transactions on
Instrumentation and Measurement 2009; 58(3): 557–562. doi: 10.1109/TIM.2008.2005264
6. Cappelli I, Carli F, Fort A, et al. Enhanced visible light localization based on machine learning and optimized
fingerprinting in wireless sensor networks. IEEE Transactions on Instrumentation and Measurement 2023; 72: 1–
10. doi: 10.1109/TIM.2023.3240220
7. Farhoodnea M, Mohamed A, Shareef H, Zayandehroodi H. Power quality impacts of high-penetration electric
vehicle stations and renewable energy-based generators on power distribution systems. Procedia Technology 2013;
11: 11–17. doi: 10.1016/j.protcy.2013.12.156
8. Bindi M, Corti F, Aizenberg I, et al. Machine learning-based monitoring of DC-DC converters in photovoltaic
applications. Algorithms 2022; 15(3): 74. doi: 10.3390/a15030074
9. Hayder AN, Saidi L. Applications of artificial neural networks with input and output degradation data for
renewable energy systems fault prognosis. In: Proceedings of the 2021 12th International Renewable Energy
Congress (IREC); 26–28 October 2021; Hammamet, Tunisia. pp. 1–6.
10. Ozbuber S, Bagriyanik M. A smart grid integration platform developed for monitoring and management of energy
systems. In: Proceedings of the 3rd International Istanbul Smart Grid Congress and Fair (ICSG); 29–30 April
2015; Istanbul, Turkey. pp. 1–5.
11. Belardi R, Bindi M, Grasso F, et al. A complex neural classifier for the fault prognosis and diagnosis of overhead
electrical lines. In: Proceedings of the 2020 International Conference on Advanced Electrical and Energy Systems;
18–21 August 2020; Osaka, Japan.
12. Kwan KH, So PL, Chu YC. An output regulation-based unified power quality conditioner with kalman filters.
IEEE Transactions on Industrial Electronics 2012; 59(11): 4248–4262. doi: 10.1109/TIE.2012.2193852
13. Kumawat PN, Verma DK, Zaveri N. Comparison between wavelet packet transform and M-band wavelet packet
transform for identification of power quality disturbances. Power Research 2018; 14(1): 37–45. doi:
10.33686/pwj.v14i1.142183
14. Pugach VN, Polyakov DA, Nikitin KI, Yurchuk DA. XLPE-insulated cables temperature monitoring for the
determination of their residual life. In: Proceedings of the Dynamics of Systems, Mechanisms and Machines
(Dynamics); 14–16 November 2017; Omsk, Russia. pp. 1–6.
15. Shafiq M, Kiitam I, Taklaja P, et al. Identification and location of PD defects in medium voltage underground
power cables using high frequency current transformer. IEEE Access 2019; 7: 103608–103618.

15
16. Heider MZ, Rahman MM, Al-Arainy AA. Study of frequency variant tan delta diagnosis for MV cables insulation
status assessment. In: Proceedings of the 2019 5th International Conference on Advances in Electrical Engineering
(ICAEE); 26–28 September 2019; Dhaka, Bangladesh. pp. 260–264.
17. Rahim NH, Chairul IS; Ghani SA, et al. Simulation of TDR circuit for the analysis of wave propagation in XLPE
cable model. In: Proceedings of the IEEE International Conference on Power and Energy (PECon); 2–5 December
2012; Kota Kinabalu, Malaysia. pp. 796–801.
18. Xun L, Shungui L, Ronghui H, et al. Study on accuracy traveling wave fault location method of overhead line—
Cable hybrid line and its influencing factors. In: Proceedings of the Chinese Automation Congress (CAC), 20–22
October 2017; Jinan, China. pp. 4593–4597.
19. Andersson G. Modelling and Analysis of Electric Power System [PhD thesis]. ITET ETH Zurich 2008.
20. Davoudi M, D’Antona G, Berizzi A. Sensitivity Analysis of Power System State Estimation Regarding to Network
Parameter Uncertainties [PhD thesis]. Politecnico di Milano-Electrical Engeneering Department-Doctoral
Program In Electrical Engineering; 2009–2012.
21. Monteiro JS, Neves WLA, Fernandes D, et al. Computation of transmission line parameters for power flow
studies. In: Proceedings of the Canadian Conference on Electrical and Computer Engineering 2005; 1–4 May
2005; Saskatoon, SK, Canada. pp. 607–610.
22. de Paulis F, Olivieri C, Orlandi A, et al. Exploring remote monitoring of degraded compression and bolted joints in
HV power transmission lines. IEEE Transactions On Power Delivery 2016; 31(5): 2179–2187. doi:
10.1109/TPWRD.2016.2562579
23. Jaffrey NA, Hettiewatte S. Corrosion detection in steel reinforced alluminium conductor cables. In: Proceedings of
the 2014 Australasian Universities Power Engineering Conference (AUPEC); 28 September–1 October 2014;
Perth, WA, Australia. pp. 1–6.
24. Shi X, Yang H, Xu Z, et al. An independent component analysis classification for complex power quality
disturbances with sparse auto encoder features. IEEE Access 2019; 7: 20961–20966. doi:
10.1109/ACCESS.2019.2898211
25. Liu Y, Jin T, Mohamed MA, Wang Q. A novel three-step classification approach based on time-dependent spectral
features for complex power quality disturbances. IEEE Transactions on Instrumentation and Measurement 2021;
70: 1–14. doi: 10.1109/TIM.2021.3050187
26. Mohan N, Soman KP, Vinayakumar R. Deep power: Deep learning architectures for power quality disturbances
classification. In: Proceedings of the 2017 International Conference on Technological Advancements in Power and
Energy (TAP Energy); 21–23 December 2017; Kollam, India. pp. 1–6.
27. Khokhar S, Mohd Zin AA, Memon AP, Mokhtar AS. A new optimal feature selection algorithm for classification
of power quality disturbances using discrete wavelet transform and probabilistic neural network. Measurement
2017; 95: 246–259. doi: 10.1016/j.measurement.2016.10.013
28. Iturrino GC, Grasso F, Luchetta A, et al. A comparison of power quality disturbance detection and classification
methods using CNN, LSTM and CNN-LSTM. Applied Sciences 2020; 10(19): 6755. doi: 10.3390/app10196755
29. Iturrino-García C, Patrizi G, Bartolini A, et al. An innovative single shot power quality disturbance detector
algorithm. IEEE Transactions on Instrumentation and Measurement 2022; 71: 1–10. doi:
10.1109/TIM.2022.3201927
30. Sturchio A, Fioriti G, Salusest V, et al. Thermal behavior of distribution MV underground cables. In: Proceedings
of the 2015 AEIT International Annual Conference (AEIT); 14–16 October 2015; Naples, Italy. pp. 1–5.
31. Sturchio A, Fioriti G, Pompili M, Cauzillo B. Failure rates reduction in SmartGrid MV underground distribution
cables: Influence of temperature. In: Proceedings of the 2014 AEIT Annual Conference—From Research to
Industry: The Need for a More Effective Technology Transfer (AEIT); 18–19 September 2014; Trieste, Italy. pp.
1–6.
32. Aizenberg I, Belardi R, Bindi M, et al. Failure prevention and malfunction localization in underground medium
voltage cables. Energies 2020; 14(1): 85. doi: 10.3390/en14010085
33. Available online: https://www.latrivenetacavi.com/it/cavi_categorie.aspx (accessed on 20 October 2023).
34. Gatta FM. Electrical Installations [Italian]. Esculapio; 2019.
35. Bindi M, Piccirilli MC, Luchetta A, et al. Testability evaluation in time-variant circuits: A new graphical method.
Electronics 2022; 11(10): 1589. doi: 10.3390/electronics11101589M
36. Bindi M, Aizenberg I, Belardi R, et al. Neural network-based fault diagnosis of joints in high voltage electrical
lines. Advances in Science, Technology and Engineering Systems Journal 2020; 5(4): 488–498. doi:
10.25046/aj050458
37. Hashmi GM, Lehtonen M. Covered-conductor overhead distribution line modeling and experimental verification
for determining its line characteristics. In: Proceedings of the 2007 IEEE Power Engineering Society Conference
and Exposition in Africa—PowerAfrica; 16–20 July 2007; Johannesburg, South Africa. pp. 1–7.
38. Bindi M, Grasso F, Luchetta A, et al. Smart monitoring and fault diagnosis of joints in high voltage electrical
transmission lines. In: Proceedings of the 2019 6th International Conference on Soft Computing & Machine
Intelligence (ISCMI); 19–20 November 2019; Johannesburg, South Africa. pp. 40–44.
39. Grasso F, Luchetta A, Manetti S, Piccirilli MC. A method for the automatic selection of test frequencies in analog

16
 fault diagnosis. IEEE Transactions on Instrumentation and Measurement 2007; 56(6): 2322–2329. doi:
10.1109/tim.2007.907947
40. Grasso F, Luchetta A, Manetti S, et al. SapWin 4.0—A new simulation program for electrical engineering
education using symbolic analysis. Computer Applications in Engineering Education 2015; 24(1): 44–57. doi:
10.1002/cae.21671

17