A Novel Fault Detection and Classification Strategy for Photovoltaic Distribution Network Using Improved Hilbert–Huang Transform and Ensemble Learning Technique

Abstract

Due to the increased integration of distributed generations in distributed networks, their development and operation are facing protection challenges that traditional protection systems are incapable of addressing. These problems include variations in the fault current during various operation modes, diverse distributed network topology, and high impedance faults. Therefore, appropriate and reasonable fault detection is highly encouraged to improve the protection and dependability of the distributed network. This paper proposes a novel technique that employs an improved Hilbert–Huang Transform (HHT) and ensemble learning techniques to resolve these challenges in a photovoltaic distributed network. First, improved HHT is utilized to extract energy features from the current signal. Second, variational mode decomposition (VMD) is applied to extract the intrinsic mode function from the zero component of the current signal. Then, the acquired energy feature and intrinsic mode function are input to the ensemble learning technique for fault detection and classification. The proposed technique is implemented using MATLAB software environment, including a classification learner app and SIMULINK. An evaluation of the results is conducted under normal connected mode (NCM) and island mode (ISM) for radial and mesh-soft normally open point (SNOP) configurations. The accuracy of the ensemble bagged trees technique is higher when compared to the narrow-neural network, fine tree, quadratic SVM, fine-gaussian SVM, and wide-neural network. The presented technique depends only on local variables and has no requirements for connection latency. Consequently, the detection and classification of faults using the proposed technology are reasonable. The simulation results demonstrate that the proposed technique is superior to the neural network and support vector machine, achieving 100%, 99.2% and 99.7% accurate symmetrical and unsymmetrical fault detection and classification throughout NCM, ISM, and dynamic operation mode, respectively. Moreover, the developed technique protects DN effectively in radial and mesh-SNOP topologies. The suggested strategy can detect and classify faults accurately in DN with/without DGs. Additionally, this technique can precisely detect high and low impedance faults within 4.8 ms.

1. Introduction

Distribution generations (DGs) are becoming widely integrated into the distribution network (DN) because it enhances the efficiency, dependability, and stability of the DN. DG supplies energy by integrating small-scale renewable energy sources [1]. DG provides consumers to produce their own electricity, with the option of injecting any extra generated into the DN [2]. By the year 2050, photovoltaic panels are expected to supply about 20 percent of the electricity in the world. Photovoltaic-based production is the third most significant resource of renewable energy in terms of world installed capacity [3]. Injecting photovoltaic-based DG into the power grid is considered an extremely promising solution related to the usage of clean energy [4]. However, DG yields a fault current depending on the generator type, size, placement, and DN configuration. In order to maintain fault current within the permissible level and reduce the strain on the circuit breaker, the installation of fault current limiters is recommended [5]. The operational condition of DNs becomes more complex due to their adaptable mode of operation and quick response to the demand of consumers [6]. Both NCM and ISM are available with DN. ISM affects protection malfunction in DN [7]; because of the operation mode change in DNs, variations in current direction and magnitude cause enormous challenges in the fault identification process [8]. Consequently, the detection of faults in DNs in various operation modes is a significant challenge in DN development.

The diagnosis of high impedance fault (HIF) is a significant concern to distribution network protection developers. HIF happens when broken electrical cables contact high impedance conductors, i.e., concrete-base, sand, grassland, or gravel [9]. Owing to the HIF possessing a very small fault current compared with low impedance fault (LIF) [10], Conventional protection technique such as fuse, overcurrent relay, and distance relay is unreliable for detecting faults [11]. Despite the fact that the DN devices will not be damaged by the low HIF current, energized conductors are preceded by an arc when HIF is happening. However, the arc can initiate fire and electric shock, posing a crucial hazard to human life and DN equipment [12]. To prevent the harmful effects on the network, diagnosis of DN faults is vital. For these reasons, HIF needs to identify efficiently and accurately in DNs.

DN reliability could be improved by implementing SNOP in cases of network failure and transient disturbance [13]. In addition, SNOP can maintain active power flow management, voltage regulation, and reactive power compensation during normal operation [14]. Moreover, SNOP provides robustness by transmitting electrical power across nearby feeders and providing mesh network benefits [15]. The application of SNOP may affect the fault detection procedure by changing the DN topology from radial to mesh-SNOP.

In recent years, differential protection, signal processing, and machine learning technique (MLT) have been developed for fault detection in DN. In such an endeavor, Sharma and Samantaray in 2018 utilized differential phase angles to detect the fault in DN [16]. Nevertheless, the communication link is the backbone of this technology. Meanwhile, Mishra and Rout [17] developed a differential protection technique based on conventional HHT and MLTs. Conventional HHT extract features utilizing current signals and zero sequence components of current signals. Mishra and Rout observed that the Extreme learning technique outperforms other MLT such as support vector machines (SVM) and Naive Bayes classifier. Chaitanya et al. utilized VMD and Hilbert transform to develop a differential protection technique [18]. This technique requires a set of threshold value to detect faults. Nevertheless, both techniques in [17,18] need communication links. Dubey and Jena in 2020 presented differential protection based on the negative sequence component of impedance angle [19]. This technique does not verify for detecting symmetrical fault and ISM. However, besides the cost of differential protection techniques, the accuracy of the differential protection-based techniques is sensitive to loads in DN. Moreover, differential protection techniques fail when communication link malfunction happens. Therefore, only local information uses the proposed technique.

Substantial research works were conducted to achieve a feasible fault detection solution over the years. In [20], a technique based on wavelet-singular entropy and a fuzzy inference system was presented to detect and classify faults, where the wavelet singular entropy was employed to extract the detailed parameter of the positive component and three-phase current signal. These signals were employed as the inputs for the fuzzy inference system. Afterward, Fuzzy-sets and -rules were used to obtain the indexes of the fuzzy inference system. The indexes were then changed to perceptual variables in order to detect and categorize faults in DN-NCM. Although the proposed technique detected the faults accurately, further studies are necessary to determine the appropriate membership function of the fuzzy-set. Heidary et al. [21] presented a protection device based on the fault current limiter in combination with a circuit breaker. This protection device did not affect the DN in NCM but, it disconnected the faulty line during the fault event. Nevertheless, the influence of DG, ISM, mesh-SNOP, and HIF was not considered. In [3], Ahmadipour et al. presented fault detection techniques depending on wavelet transform and SVM. Wavelet transforms were employed to extract prominent features from voltage signals. Afterward, these features were utilized to train and test the SVM to detect and classify faults. Whereas, Patcharoen and Ngaopitakkul proposed a fault classification technique based on wavelet transform and decision tree technique [11]. Wavelet transform was utilized to extract significant features from the current signal. Afterwards, the decision tree technique was employed to classify faults. Again, both approaches [3,11] did not take into account the impact of ISM, and HIF. In addition, the accuracy of the developed technique can be enhanced with the selection of appropriate hyperparameters, activation functions and training algorithms. Srinivasa Rao et al. [22] proposed a neural network with an adaptive evolutionary training technique and wavelet decompositions with cascade SVM to detect and classify faults in DN. Srinivasa Rao et al. found that Cascade SVM is powerful and quicker than conventional SVMs. In 2021, Moloi and Yusuff presented a technique based on discrete wavelet transform, neural network, and genetic algorithm to fault diagnostic [23] where the discrete-wavelet transform was employed to extract statistical features from the current signal. Then, a neural network was used as a classifier technique to detect and classify fault. Whereas genetic algorithms were utilized to optimize the neural network performance. A few strategies for fault identification were developed by optimizing the MLT algorithm and tuning parameters. Nevertheless, both approaches [22,23] did not take into consideration the effect of DGs and ISM.

Several techniques for detecting HIFs in DN have been developed in recent years. Vyshnavi and Prasad [10] employed the fuzzy logic technique to detect HIF in DN while this technique did not consider the effect of ISM and DGs. Zheng et al. proposed communication-based protection of converter-interfaced islanded DNs [4]. This method used Park transformed grid voltage and wavelet parameters of modal current transients to identify the presence of solid and HIFs, respectively. In [24], a wavelet transform with an extreme learning machine is used to protect smart grids from HIF. As three-phase current signals from both sides of the power line are used for extraction of the high-frequency components used in this protection technique, an extremely reliable communication link is required. While Roy and Debnath [25] proposed an HIF detection technique by using wavelet transform to determine power spectral density. Nevertheless, this technique depends on the threshold value. The detection time depends on appropriate threshold values. Meanwhile, Manohar et al. [26] proposed the least squares-Adaline algorithm and improved SVM to detect and classify HIF in medium voltage DN. This technique did not consider the effect of SNOP application. Whereas, Xiao et al. suggested neural network and decision tree technique to detect HIF using transient zero sequence component of the current signal [27]. The effect of ISM and SNOP applications is not determined by this technique. Accordingly, it is observed that HIF can be diagnosed effectively and precisely using some sophisticated algorithms.

In this article, a combination of signal processing, and MLT was introduced to detect and classify faults in low voltage DN. Signal processing is used to extract hidden features from local measurements of the current signal. MLT is utilized to avoid inadequate threshold values which have a significant impact on the accuracy and the detection time. It is a known fact that in nonstationary signal analysis the VMD method outperforms empirical mode decomposition (EMD). As a result, VMD is adapted in HHT to improve the feature extraction method rather than the EMD. The extracted features are then trained with the MLT, namely the ensembled bagged trees method (EBTM) to improve the classification accuracy. Such a novel strategy adopted in this work of combining VMD in HHT for feature extractions and EBTM as a classifier has contributed to the following enhancements:

• The proposed strategy can detect and classify all types of faults accurately, which involve both symmetrical and asymmetrical faults in low voltage DN using only local input data. As such, it can be concluded that the proposed strategy is reasonable for DN protection compared with the other techniques that use communication links mentioned earlier.
• The effect of operation mode change between NCM and ISM has also been considered. However, the findings suggest that the proposed strategy works well on NCM but not likewise on ISM. This is mainly due to the limited size of the fault current data. In addition, an interesting point to note is that the proposed strategy provides a high degree of accuracy throughout a variety of operational modes.
• The proposed strategy can identify both the LIF and the HIF accurately. Hence, the HIF has a very low-fault current in contrast to the LIF. Therefore, the proposed strategy outperforms conventional protection techniques.
• The effect of mesh-SNOP on the proposed technique is considered while other existing techniques do not consider it. Despite the deployment of SNOP having the ability to affect the fault detection process by adapting the DN topology, the suggested method protects DN sufficiently in both radial and mesh-SNOP topologies.
• A novel fault detection and classification strategy using improved HHT, and EBTM is proposed. Where improved HHT consists of VMD and HHT because non-stationary signals can be addressed using VMD. The proposed EBTM-based technique is compared to five other machine learning techniques, which are trained and tested utilizing the same dataset that was used for EBTM. The proposed strategy has the capability to detect and classify faults precisely in DN with/without DGs.

This article is divided into five main sections. Section 2 provides a description of the proposed methodology. Section 3 presents a brief description of the case study. Section 4 illustrates results and discussions. Lastly, conclusions are illustrated in Section 5.

2. Methodology

In this section, the details of the proposed methodology are presented, where the improved HHT is used to extract current signal features. Additionally, Fault classification is achieved by the employment of EBTM.

2.1. Improved Hilbert–Huang Transform

The HHT is a time frequency-analysis-based approach that commonly combines EMD and the Hilbert transform [18]. HHT is a powerful, flexible, and accurate mechanism for extracting signal features. Furthermore, the non-stationary power signals can be analyzed by the HHT [1]. In addition, it is capable of performing high-precision analyses in the time- and frequency domains. Due to HHT’s superb outcomes, it has been commonly employed in low-frequency vibration analysis, electrical fault diagnosis, mechanical structural fault detection, and other fields. The VMD outperforms the EMD in multi-component signal decomposition, side-band identification, intra-wave feature extraction, and noise robustness [28]. Therefore, VMD was used instead of EMD to improve HHT performance in this article.

VMD is a unique, non-recursive, completely intrinsic, and adaptive signal processing technique that decomposes a signal into sub-signals or intrinsic mode functions (IMFs) or modes [29]. The VMD algorithm separates a signal $Y(t)$ into a finite number of IMFs $Y_k(t)$, and each signal has a band-limited bandwidth. The modes are identified by differences in sparsity properties, although the original input signal is reproduced in each case. The IMFs are formed as a sinusoidal waveform function.

$$Y_k(t)={\displaystyle {\sum }_{k=1}^K{M}_k(t)\cos {\theta }_k(t)}$$

(1)

where K represents the number of modes, $Y_k(t)$ are IMFs, $M_k(t)$ are the positive envelope of IMFs, and ${\theta }_k(t)$ are the non-decreasing phase. Reproducing an input signal by deconstructed modes has a particular sparsity feature. Specifically, a majority of each mode revolves around the center frequency (CF) [28]. The IMFs are indeed the reproduction of varying amplitudes of substantial disturbances contained in the input signal owing to the fault. All IMFs can be combined to reconstruct the original signal. To determine the bandwidth and CF of each IMF, the optimal solution to a variational problem is evolved continuously while maintaining a steady optimization strategy. For any signal, the decomposition problem can be formulated as follows:

$$\begin{array}{l}L\left(\left\{M_k\right\},\left\{w_k\right\},\lambda \right)=\alpha {\displaystyle \sum _k{\Vert \frac{d}{dt}\left[\left(\delta (t)+\frac{j}{\pi t}\right)\ast M_k(t)\right]e^{-j{w}_{k}t}\Vert }_2^2}\\ +{\Vert f(t)-{\displaystyle \sum _k{M}_k(t)}\Vert }_2^2+\langle \lambda (t),f(t)-{\displaystyle \sum _k{M}_k(t)}\rangle \end{array}$$

(2)

where $w_k$ represents the center angular frequency, $\delta$ denotes the Dirac distribution, $\ast$ is the convolution, $j$ is the imaginary unit with a value $j^2=1$. $\alpha$ is the quadratic penalty term, and $\lambda$ represents the Lagrangian multiplier (LM). The original signal $f(t)$ can be reconstructed by adding all the modes. Using the Alternate direction method of multipliers, Equation (2) can be effectively minimized. The corresponding updated CF and the evaluated modes in the frequency response are shown as follows:

$${\hat{M}}_k^{n+1}(w)=\frac{\hat{f}(w)-{\displaystyle \sum _{i<k}{\hat{M}}_i^{n+1}(w)-{\displaystyle \sum _{i>k}{\hat{M}}_i^n(w)+\frac{\hat{\lambda }(w)}{2}}}}{1+2\alpha {(w-w_k)}^2}$$

(3)

$${\displaystyle \sum _k\frac{{\Vert {\hat{M}}_k^{n+1}-{\hat{M}}_k^n\Vert }_2^2}{{\Vert {\hat{M}}_k^n\Vert }_2^2}<\epsilon }$$

(4)

Equations (3) and (4) provide the mathematical method for updating the modes and their CFs for the number of modes selected (n). Where ${\hat{M}}_k^{n+1}(w)$ and $w_k^{n+1}$ represent the amplitude and CF of the next mode (n + 1) of the kth level of decomposition. $\hat{f}$ indicates the Fourier transform of $f$ the original signal. $\epsilon$ is an appropriate tolerance rate. The LM is updated as

$${\hat{\lambda }}_{n+1}(w)={\hat{\lambda }}_n(w)+\tau \left(\hat{f}(w)-{\displaystyle \sum _k{\hat{M}}_k^{n+1}(w)}\right)$$

(5)

where $\tau$ denotes the LM update rate. The number of iterations is raised, and the algorithm continues to update the modes, CFs, and LM until convergence is achieved. The next procedure computes the instantaneous energy by applying HHT for each IMF as illustrated below [17]:

• Determine the analytic signal using Equation (6).

$$Z_k(t)=M_k(t)+jH\left\{M_k(t)\right\}=a_k(t)e^{j{\theta }_k(t)}$$

(6)

where $Z_k(t)$ represents the analytic signal and $H\left\{M_k(t)\right\}$ denotes Hilbert transform of $M_k$. In addition, $a_k(t)$ and ${\theta }_k(t)$ represent instantaneous amplitude and phase, respectively.

2.

The instantaneous energy and frequency can be calculated by using:

$$\mathrm{instantaneous} \mathrm{energy}= {\left|a_k\left(t\right)\right|}^2$$

(7)

$$\mathrm{instantaneous} \mathrm{frequency}=\frac{d{\theta }_k(t)}{dt}$$

(8)

In this article, only the instantaneous energy feature is utilized because it met our design requirements.

2.2. Sliding-Window Mechanism

Real-time features need to be preserved during the process of extracting fault features. Nevertheless, VMD in HHT requires a specific amount of signal length to assure a proper IMF output. This causes issues with the time-consuming fault feature extraction technique utilizing HHT.

The sliding window mechanism is employed in the proposed technique in order to minimize the impact that is induced by the data processing. With this mechanism, a fixed-length sample can be generated by sliding a fixed-length window throughout a specific duration. The sample is continually updated by sliding it simultaneously with the changing time. New data occur and old data are discarded after a fixed interval. An updated sample of a fixed length is generated [6]. As a result, the data for VMD can be obtained quickly. Furthermore, the performance of the mechanism will be improved by selecting suitable sliding window dimensions.

2.3. Ensemble Learning Technique

An MLT can be generally divided into three categories: supervised, semi-supervised, and unsupervised learning. One of the most common methods that are used for classifying is supervised machine learning. Classification ability is achieved by the use of training errors in the training function. This closed-loop feedback can improve classification accuracy in MLTs [30]. Therefore, only supervised MLTs were applied as the base learner for the Ensemble learning technique in this research.

There are three main steps to follow in order to implement the ensemble learning technique. The first phase involves manipulating the training dataset and building models using different learning algorithms. Member selection is the second phase of the process. This phase includes only selecting models that could make predictions. The output from multiple classifiers is combined into one final prediction in a third phase known as the member combination phase. In addition, there are three stages that need to be contributed to the task, and each step requires multiple classifiers. The existing procedures are as follows. Firstly, considering different perspectives to combine classifiers. Secondly, co-operating classifiers employing one or more perspectives. Thirdly, choosing classifiers based on a variety of criteria, including the use of basic ensemble techniques. To produce a final accurate prediction by combining the outcomes of several classifiers, designers employ several fundamental ensemble approaches including average, majority voting, weighted average, and weighted majority voting. Boosting, Bagging, and Random Subspace are the main categories of ensemble learning techniques for implementing machine learning classifiers [31]. In order to tackle the classification challenge, the EBTM was selected.

Bagging is alternatively called bootstrap aggregation. There are two key benefits of using bagging: it eliminates variance and reduces model overfitting by creating multiple classifiers using fixed bias and combining their findings by averaging. This method is quite effective when the features of the input have a large variation and minimal bias. Bagging generates numerous bootstrap sets of data from the training data, trains the data using a classifier, and then combines the results of each model using a convenient technique such as majority voting [32]. Figure 1 illustrates the overall process and structure of the EBTM, in which subset (n) is a group of datasets randomly selected from the original training database. Each model is developed with a subset of the training data sets, where every sample data could be selected multiple times during the training procedure. This procedure of selecting data is also identified as sampling with replacement. The last step consists of aggregating all predictions by the average [30].

2.4. Proposed Technique for Fault Detection

The proposed fault detection and classification technique using improved HHT and EBTM are illustrated in Figure 2. The proposed method is divided into two main processes: detection and classification. In order to detect the fault, VMD is used to extract IMFs (mode-3) from three-phase current signals. Afterward, HHT was employed to determine energy features. The sequence analysis was utilized to compute the zero-sequence component (ZSC) of the ninth harmonic from the three-phase current in the second process. The transmitting ZSC signals are boosted by using the gain. Subsequently, the VMD is employed to extract IMF (mode 5) from the ZSC of the current signal. Hence, IMF mode has been selected manually for both ZSC and current signal by using the signal processing toolbox\signal analyzer in MATLAB. Finally, both IMF (mode 5) and energy features are used for fault detection and classification using EBTM.

The overall flow chart of the proposed technique for detection and classification of fault in DN is shown in Figure 3.

3. Case Studies

A Low voltage DN with 11/0.4 KV voltage level is modeled and simulated in MATLAB (R2021a, Natick, MA, USA) Simulink environment as illustrated in Figure 4. The distribution grid is fed by two inverters interfaced DGs. Additionally, the inverters interfaced DGs are connected to the network through two step-up transformers 1 and 2.

The DN consists of two Photo-Voltaic farms, two inverters, two battery banks, two transformers, and four loads. The specification of grid components is demonstrated in

Table 1. The Specification of DN Components.

No.	Component	Specification
1	Photo-Voltaic farm	2 Parallel strings, 28 series-connected modules per string, irradiance 1000 W/m2
2	Inverter 1 and 2	12.5 kVA, 0.4 kV, 50 Hz, switching frequency 5 kHz Filter: Series Inductance 4.6 mH Series Resistance 0.4596 Ω, Shunt Capacitance 0.1102 µF
3	Two battery banks	Lead-Acid, 980 V, 2.7 Ah
4	Transformers	11/0.4 kV, 24 kVA, 50 Hz, D11/Yn
5	load 1 and 2	5 kW, 0.4 kV, 50 Hz
6	Load 3 and 4	10 kW, 11 kV, 50 Hz, 16 kW, 11 kV, 50 Hz

.

4. Results and Discussions

As previously stated, the feature is extracted from ZSC and current signals through VMD. VMD is capable of deconstructing the signal into several modes. At the same time, a high level of deconstruction is recommended in order to collect the maximum number of accessible signal features. In addition, there are several downsides to a higher level of decomposition. Increasing the number of modes will further influence the computational burden, which may lead to a rise in the relay’s response time. An increase in the relay’s response time may cause a significant danger to the power grid. Therefore, the decomposition modes are limited to a maximum of 5. Unlike another implementation of VMD, it is not required to capture the entire signal information because VMD was implemented within MATLAB/Simulink environment in this article. However, each signal should be converted to a moving window-based signal to successfully operate VMD in the MATLAB Simulink environment. the sample size is an essential component to address in addition to the sampling frequency. The samples’ size must first be carefully specified in order to make the moving window and proposed technique operate efficiently. Thus, 2048 samples per moving window are being used to achieve a balance between minimal computation time and comprehensive feature extraction based on extensive analysis. In this study, the VMD algorithm decomposed Five IMFs, as stated in Figure 5. Whereas the sampling time interval for the discrete simulation type is specified as 2.5 × 10^-6 second.

After extensive analysis considering all types of faults in DN, IMF (mode-3) was utilized to extract features from the current signal and IMF (mode-5) was employed to extract features from ZSC current signal. Afterwards, the training data are collected after applying HHT on IMF (mode-3) to acquire energy features. Figure 6 depicts the energy feature extracted from the current signal of phase A. Energy features extracted by improved HHT technique during double line to ground (DLG) on phases B and C in ISM are demonstrated in Figure 7. As illustrated in Figure 7a,b, the energy feature of the phase B variant from 581.33 to 574.72 during LIF while activating the SNOP, respectively. Similarly, the phase C energy feature changes from 505.88 to 499.52. Figure 7c,d depicts how the energy feature of phase B during HIF changes from 556.07 to 548.49 when SNOP is activated. Correspondingly, the phase C energy feature decreases to 495.52. Thus, the energy feature of the current signal during LIF surpasses HIF. Additionally, initiating the SNOP in DN reduces the energy feature of the current signal during both LIF and HIF.

4.1. Performance of MLT for Detection and Classification Fault

This section demonstrates the performance of artificial classifiers for detecting and classifying faults utilizing improved HHT, and VMD based on energy features, and IMF-mode 5, respectively. In this subsection, six artificial classifier techniques are evaluated and contrasted: Narrow-Neural Network, Fine Tree, Quadratic SVM, Fine-Gaussian SVM, Wide-Neural network, and EBTM.

Accuracy is indeed a statistical metric that is used to evaluate the effectiveness of MLTs. Accuracy measures the dependability between predicted and normal circumstances for both non-fault and fault events simultaneously. It can be calculated by using Equation (9).

$$\frac{\mathit{Total}(\tilde{F}+\tilde{\overline{F}})}{\mathit{Total}(F+\overline{F})},$$

(9)

where $\tilde{F}$ and $\tilde{\overline{F}}$ are the predicted fault and healthy cases. Additionally, $F$ and $\overline{F}$ represent the actual fault and healthy cases. Moreover,

Table 2. The obtained accuracy using numerous MLTs.

MLTs	Accuracy %
	With DG					Without DG
	NCM		ISM		Dynamic
	Radial	Mesh-SNOP	Radial	Mesh-SNOP	Radial and Mesh-SNOP	Radial
Narrow Neural Network	93.3	94.2	88.3	88.3	86.1	97.8%
Fine Tree	90	89.2	88.3	88.3	94.9	100.0%
Quadratic SVM	90	90	89.2	90.8	80.8	93.5%
Fine Gaussian SVM	90	90	94.2	93.3	72.2	93.5%
Wide Neural Network	98.3	98.3	95.8	96.7	85.9	97.8%
EBTM	100	100	99.2	99.2	99.7	100.0%

illustrates the obtained accuracy of different MLTs.

In this article, the re-substitution validation process is used for training and testing all MLTs in order to guarantee full benefit from all data. Furthermore, the training data of 396 cases in DN integrated with DGs during dynamic operation mode was trained using supervised machine learning classification learner application in MATLAB. The training data covers both operation modes of DN, including NCM and ISM. In addition, all types of faults were considered, for instance, single-line to ground, double-line, DLG, three lines, and three lines to ground (3LG) faults. Furthermore, HIFs and LIFs were studied under both operation modes for all fault types. To carry out LIFs and HIFs, the fault impedance is set as 0.01, 10, 80, 100, 500, and 1000 Ω correspondingly. The fault is established at 0.2 s and the three-phase current signal from bus-4. Moreover, the existence of SNOP on the DN and its impact on the proposed protection technique is taken into consideration. The circuit breaker is used to make a DN topology transition from a radial to a mesh-SNOP, as depicted in Figure 4. The MLT’s performance was trained and tested in DN with/without DGs. The training scenario in DN integrated with DGs is conducted based on radial topology with NCM, mesh-SNOP topology with NCM, radial topology with ISM, and mesh-SNOP topology with ISM. All types of faults, including symmetrical and asymmetrical faults, are conducted under these scenarios.

4.1.1. MLTs Performance in DN Integrated with DGs

• NCM operation mode

NCM and ISM fault currents are different in magnitudes, making it difficult to achieve using a protection mechanism with a pre-set setting threshold value. In Table 2, the performances of six MLTs in NCM are given for radial topology. The DN protection capabilities of the MLTs have been evaluated using 120 cases in a radial topology. In addition, HIF and LIF have been conducted in radial topology with NCM under various fault locations. It has been found that the lowest rates of accuracy can be achieved by the fine tree, quadratic SVM, and Fine Gaussian SVM machine learning with a 90 percent accuracy rate; Narrow Neural Network poses a 93.3% accuracy rate. Furthermore, the Wide Neural Network is observed to be 98.3% accurate. It has been observed that EBTM is able to achieve the highest possible accuracy with one hundred percent.

Whereas the performance of six MLTs can also be affected by changing the topology of NCM-DN between radial to mesh-SNOP as a result of unbalanced current distribution in the corresponding feeder. Figure 8 demonstrates a comparative analysis of the accuracy of all MLTs applied in this study under both radial and mesh-SNOP in NCM-DN. The MLTs have been tested for DN protection with 120 cases in a mesh-SNOP topology. Moreover, HIF and LIF are handled in mesh-SNOP topology with NCM at diverse fault locations. It has been found that the lowest rates of accuracy can be achieved by fine tree machine learning with 89.2%. While the Quadratic SVM and Fine Gaussian SVM accuracy rates are 90%. The Narrow Neural Network and Wide Neural Network accuracy rates are observed to be 94.2 and 98.3%, respectively. In contrast, it has been found that the optimal accuracy of 100% can be acquired by EBTM.

2.

ISM operation mode

Owing to the dynamic behavior of the operating conditions, traditional over-current relays might have considerable relaying issues. The performances of six MLTs in ISM are given for radial topology in Table 2. The DN protection abilities of the MLTs have been tested utilizing 120 cases in a radial topology. HIF and LIF are tested in radial topology with ISM in different fault locations. It has been discovered that the fine tree and Narrow Neural Network machine learning method can attain an accuracy rate of 88.3 percent, which is the lowest rate of accuracy achievable as compared with the other four MLTs. Conversely, the accuracy rates of the Quadratic SVM and Fine Gaussian SVM are found to be 89.2 and 94.2%, respectively. Whilst the accuracy rate of the Wide Neural Network has been decreased to 95.8 percent as compared to the NCM condition. On the contrary, EBTM is capable of reaching the accuracy of 99.2%, according to investigative work.

Consequently, the six MLTs’ performance can also be affected by changing the ISM-DN topology from radial to mesh-SNOP. The statistics of six MLTs in ISM are shown in Table 2 for mesh-SNOP topology. Figure 9 depicts a comparative analysis of the accuracy of six MLTs utilized in this study under radial and mesh-SNOP in ISM-DN. In a mesh-SNOP topology, 120 cases were utilized to assess the DN protection capabilities of the MLTs. HIF and LIF are investigated with ISM in mesh-SNOP topology with diverse fault locations. It has been found that the fine tree and Narrow Neural Network machine learning method have an accuracy rate of 88.3%, which is the minimum accuracy rate that can be attained among the other four MLTs. The accuracy rate of the Quadratic SVM, Fine Gaussian SVM, and Wide Neural Network are found to be 90.8, 93.3, and 96.7%, respectively. Nevertheless, EBTM is capable of obtaining 99.2 percent accuracy, which is the maximum conceivable accuracy in ISM, according to the outcomes of the investigation.

3.

Dynamic operation mode

In this section, the performance of six MLTs are tested extensively under NCM, and ISM. Furthermore, both radial and mesh-SNOP of DN topologies are investigated in all MLTs in order to obtain the overall accuracy. Based on 396 cases, the MLTs’ DN protection capabilities have been validated. The HIF and LIF are inspected in both radial and mesh-SNOP DN topology with a variety of fault positions. Figure 10 illustrates an overall accuracy of six MLTs employed in this study under dynamic operation mode in DN. It has been observed that the Fine Gaussian SVM machine learning method has an accuracy rate of 72.2 percent, which is the lowest accuracy rate that can be achieved with the other five MLTs. Whereas the accuracy rate of the Quadratic SVM machine learning method is found to be 80.8%. However, Wide Neural Network Narrow Neural Network, and Fine Tree machine learning methods have the capability to acquire 85.9, 86.1, and 94.9% accuracy rates. According to the findings, EBTM has been found to have the best accuracy, obtaining a perfect rating of 99.7 percent.

4.1.2. MLTs Performance in DN without DGs

In this section, the performance of six MLTs are tested in DN without DGs integration. Moreover, the redial DN without integration of DGs are investigated in MLTs in order to obtain the accuracy. The DN protection capabilities with MLTs have been validated based on 93 cases. Furthermore, the HIF and LIF are examined in a variety of fault locations. Figure 11 demonstrates the overall accuracy of six MLTs employed in this study under DN without DGs integration. It has been found that both the Quadratic SVM and Fine Gaussian SVM machine learning method have an accuracy rate of 93.5%, which is the minimum accuracy rate that can be attained among the other four MLTs. Additionally, the accuracy rate of both the Narrow Neural Network and Wide Neural Network are found to be 97.8%. Nevertheless, both fine tree and EBTM are capable of obtaining the optimal accuracy of 100 percent accuracy, which is the maximum conceivable accuracy.

As denoted in Table 2, the EBTM provided excellent accuracy when compared with other methods, i.e., Narrow-Neural Network, Fine Tree, Quadratic SVM, Fine-Gaussian SVM, and Wide-Neural network. Therefore, the fault classification challenge was addressed using the EBTM.

Table 3. The confusion matrix of proposed technique under dynamic operation mode.

		Predicted Class
		Number of faults	AG	BG	CG	AB	AC	BC	ABG	ACG	BCG	ABC	ABCG
True class	No fault	28	0	0	0	0	0	0	0	0	0	0	0
	AG	0	48	0	0	0	0	0	0	0	0	0	0
	BG	0	0	48	0	0	0	0	0	0	0	0	0
	CG	0	0	0	48	0	0	0	0	0	0	0	0
	AB	0	0	0	0	8	0	0	0	0	0	0	0
	AC	0	0	0	0	0	8	0	0	0	0	0	0
	BC	0	0	0	0	0	0	8	0	0	0	0	0
	ABG	0	0	0	0	0	0	0	48	0	0	0	0
	ACG	0	0	0	0	0	0	0	0	48	0	0	0
	BCG	0	0	0	0	0	0	0	0	0	48	0	0
	ABC	0	0	0	0	0	0	0	0	0	0	7	1
	ABCG	0	0	0	0	0	0	0	0	0	0	0	48

depicts the confusion matrix of the proposed technique under dynamic operation mode. Where 28 cases of no-fault conditions are considered in the training data. The no-fault data compose of both operation modes, mesh-SNOP active, loads variation (between +10% and −10%), and zero active power mismatch. While HIF is included for single-line to ground, DLG, and 3LG faults. Furthermore, different fault locations (between 0.6 and 0.8 km near bus 4) are conducted.

Afterward, the EBTM can be implemented in MATLAB Simulink environment using the classification ensemble predict block. Finally, the improved HHT, VMD, and EBTM were tested in a MATLAB Simulink environment to determine the detection time. As a result, the proposed method required only 4.8 ms to detect and classify all types of faults, including LIF and HIF. In addition, the change in operation mode between NCM and ISM has slight effects on the performance of this technique. Furthermore, this technique’s efficiency is unaffected by changing DN topology between radial and mesh-SNOP in both operation modes NCM and ISM. Furthermore, the observations indicate that the proposed technique works effectively on DN with/without DGs.

A comparison of the suggested system with various existing strategies is presented in

Table 4. Comparison of the suggested system with various existing strategies.

Ref.	Method	HIF	Mesh-SNOP	Operation Mode	Considered All Fault Types	Communication Link Required	Accuracy
[33]	Statistical morphology, recursive least square approaches, Butter-worth filter	✕	✕	Both	Missing 3LG fault	No	98%
[3]	Wavelet transform and SVM	✕	✕	NCM	Missing three lines fault	No	100%
[17]	HHT, differential protection, MLT	●	✕	Both	●	Yes	In NCM 96.99%, and ISM 96.75%
[34]	Hilbert transform, and Ada boost classifier	●	✕	Both	Missing three lines fault	No	98.75%
[25]	Wavelet transform, power spectral density, threshold value	●	✕	NCM	Missing 3LG fault	No	Not specified
[6]	Current phase difference, ZSC of current, HHT, adaptive, and threshold value	●	✕	Both	Missing 3LG fault	Yes	Not specified
[26]	least square and Adaline algorithm, Sine Cosine algorithm, and SVM	●	✕	Both	●	No	In NCM 99.56% and ISM 99.17%
Proposed Technique	Improved HHT, and EBTM	●	●	Both	●	No	In DN with DGs (dynamic mode) 99.7%, and without DGs 100%

. The performance of the majority of stated techniques is insufficient for detecting and classifying DN faults. In addition, A few of them detected all types of faults. Some detection strategies can be deployed during both the NCM and ISM, as shown in Table 4. While the proposed detection technique is effective for both NCM and ISM. Nevertheless, the proposed technique is capable of detecting and classifying all fault types in DN. HIF and LIF can be accurately diagnosed and classified by using the proposed method. Mesh-SNOP is not considered by other existing techniques. Furthermore, the effect of Mesh-SNOP has no effect on the proposed technique. Both techniques in [6,17] are required communication links which is not reasonable to implement in low voltage DN. Whereas the suggested technique depends only on local input and does not need a communications link. Therefore, the proposed method achieves superior results than those of the existing methods stated in Table 4.

5. Conclusions

Improving DN dependability is achieved by highly accurate fault detection. Conventional protective systems are inadequate in addressing DN issues because of their specific characteristics and operations. These issues include changing fault currents throughout various modes of operation, diverse DN topologies, and high impedance faults. The proposed technique provides an effective solution to operation mode changes from NCM to ISM, miscellaneous DN topologies radial and mesh-SNOP, and high impedance fault. In this paper, a novel protection technique based on improved HHT, and EBTM is proposed to detect and classify fault precisely in DN. VMD is utilized to extract features from ZSC of the current signal obtained by the sequence analyzer. After that, the energy features are acquired by applying improved HHT on three-phase currents. The IMF and energy feature act as input data of EBTM to detect and classify faults. The results demonstrate that the proposed technique can detect and classify symmetrical and asymmetrical faults without employing a communication link.

In addition, the proposed technique was tested under different operation conditions, i.e., NCM with radial, NCM with mesh-SNOP, ISM with radial, and ISM with mesh-SNOP. The investigations observed that the proposed strategy operates well on DN with/without DGs. Furthermore, both HIF and LIF can be detected and classified by using this technique. The accuracy of the ensemble bagged trees technique is higher when compared with the narrow-neural network, fine tree, quadratic SVM, fine-gaussian SVM, and wide-neural network. The proposed technique’s classification and detection accuracy in DN with DGs (dynamic operation mode) is 99.7% and without DGs is 100% and its detection time is 4.8 ms.

The presented technique is simple to implement due to its dependence on local variables and has no required connection latency. Therefore, the proposed technique for low voltage DN is a cost-effective fault detection and classification method. Finally, further investigation is still needed in the future for real-time authentication of the proposed technique using a hardware-in-the-loop environment and to improve the fault detection time.