Signal processing and machine learning techniques in DC microgrids:a review☆

0 Introduction

In this modern era,electricity has become the backbone of technological progress, economic growth, and societal well-being.With the global transition toward cleaner and more sustainable energy solutions, there is an increasing focus on integrating renewable energy sources, such as solar photovoltaics (PVs), wind energy, and battery storage systems, into power grids.The rising energy demand,coupled with concerns over uncertain climate conditions and fossil fuel depletion, has accelerated the shift toward decentralized power generation.However, integrating renewable energy sources into traditional alternating current (AC) grids presents significant challenges, including power conversion losses, stability concerns, and complex synchronization issues.To overcome these challenges,direct current (DC) microgrids (MGs) are emerging as a promising solution due to their ability to efficiently integrate renewable energy, storage systems, and modern power electronic converters [1].Further, researchers have turned their efforts toward hybrid MGs to benefit from both AC and DC MGs.A sample architecture of the hybrid MG is depicted in Fig.1.This MG consists of renewable and non-renewable energy sources, AC and DC loads, and a utility grid connection.

DC sources are directly connected to the DC bus,whereas AC sources are connected at a point of common coupling in the AC bus system.However, although this setup leverages the advantages of hybrid MGs, critical aspects such as grid control and isolating incipient faults remain challenging.Henceforth, most applications are built based on DC MGs rather than AC MGs owing to their significant advantages.Generally, DC MGs enable faster response times,better voltage regulation,and simplified power control strategies, making them highly suitable for applications such as smart grids, data centers, industrial facilities,electric vehicle(EV)charging stations,space missions, and remote electrification projects [2,3].Despite these advantages,DC MGs present significant operational challenges, particularly in terms of fault detection, protection, and islanding diagnosis.This review investigates the challenges of DC MG protection, covering critical aspects such as fault current characteristics, grounding systems,fault detection methods, protective devices, islanding detection, power quality disturbance (PQD) detection,and fault location techniques.Each area is reviewed comprehensively, and the future trends in DC MG protection are outlined.

A robust DC MG protection scheme must ensure proper grounding, fast and accurate fault detection, current limiting,and the use of advanced DC circuit breakers(DCCBs) [1].Grounding enhances grid reliability, minimizes leakage, and supports effective fault detection by meeting safety, fault ride-through, and detection requirements.Rapid fault progression in DC lines, coupled with low resistance and reactance, demands innovative fault location strategies balancing performance, complexity,and cost [4].Nonlinearities from power electronics challenge detection and coordination.Most studies have been addressed through current-profile-based protection.Although these methods are fast, they are unreliable for high-impedance faults (HIFs), while current derivative and directional methods suffer from noise sensitivity.For MG, DCCBs are critical, offering galvanic isolation, fast response,low loss,long life,and reliability,outperforming AC breakers for DC fault dynamics [2,3,5,6].PQDs, such as sags, swells, harmonics, and transients, often due to faults and switching events, degrade equipment reliability,necessitating real-time PQD detection and classification[3,6].Islanding detection is harder in DC grids due to the absence of reactive power, necessitating the use of voltage-based methods in place of traditional AC strategies.Standards like IEEE 1547 and UL 1741 require adaptations for DC [7].Distributed generation (DG)integration introduces protection challenges: altered fault current levels, blinding of protection, false tripping, unintentional islanding, and auto-reclosure issues.Protection devices must be selective, sensitive, fast, reliable, and cost-effective,with adaptive,intelligent protection schemes necessary for dynamic MGs [8].Conventional static schemes are inadequate for dynamic, islanded systems.Modern strategies, such as dual-setting directional overcurrent relays, rate of change of voltage relays, and I-V logic, enhance adaptability [9].Artificial intelligence, via neural networks, support vector machines, and evolutionary algorithms,is used to optimize protection settings and enable real-time adaptability.Communication-assisted relays further improve coordination but raise cybersecurity concerns, highlighting the need for intelligent, adaptive,and secure protection systems [10].

Fig.1.MG with renewables and distribution loads.

In response to these challenges, numerous researchers have conducted extensive studies on fault diagnosis and protection schemes for DC MGs.This growing research interest is highlighted by a bar graph in Fig.2 [5], which is based on a Scopus trend analysis, highlighting the increasing focus on DC MG protection strategies in recent years.This growing research interest reflects the critical need for robust fault diagnosis techniques to ensure the reliable operation of DC MGs.Various fault types,including pole-to-pole faults, pole-to-ground faults, and HIFs,can significantly impact MG stability if they are not detected and isolated promptly.Hence,effective protection schemes play a crucial role in minimizing fault isolation time, preventing equipment damage, and enhancing the overall resilience of MGs.

To address these crucial aspects, this paper provides a comprehensive review of various protection schemes employed in DC MGs, focusing on fault diagnosis and islanding diagnosis.In pursuit of these objectives, authors have extensively researched different DC MG articles.Based on this research, the articles have been categorized according to the algorithms and techniques employed, as follows:

1. Mathematical-based methods: These approaches utilize system equations and control algorithms to detect and isolate faults based on key parameters such as voltage,current, and impedance properties.

2. Signal processing-based techniques: These methods employ advanced signal analysis tools, including Fourier, wavelet, and Hilbert-Huang transforms, to extract fault features in the time and frequency domains for classifying various faults.

3. Machine learning (ML)-based approaches: In this approach, different ML techniques, such as artificial neural networks (ANNs) and support vector machines(SVMs), are utilized to attain superior and accurate results in terms of fault detection and classification.Additionally, they focused on islanding event detection and classification through ML.

4. Hybrid models: These techniques combine multiple approaches to increase fault detection reliability,reduce false positives, and adapt to varying MG conditions.

Fig.2.Year-wise publications on DC MG protection in the Scopus database.

Each of these protection schemes is crucial for ensuring the seamless and smooth operation, safety, and efficiency of DC MGs.However, significant challenges persist, particularly in real-time implementation and adaptability to diverse MG architectures.Despite numerous studies in this domain, there remains a need for a comprehensive review that consolidates existing protection strategies and critically evaluates their effectiveness in real-world scenarios.This paper addresses this gap by synthesizing advancements and emerging trends in DC MG protection,offering a structured analysis of fault detection, isolation, and islanding strategies.By evaluating their strengths and limitations,this study aims to guide future research and technological developments in this critical area.Ultimately,the collective efforts of researchers continue to drive progress toward sustainable and resilient energy solutions, unlocking the full potential of DC MGs in modern power systems.

1 Review of mathematical approach-based fault diagnosis schemes in DC MG

The mathematical approach-based fault diagnosis schemes have typically relied on mathematical equations,circuit analysis, and system modeling to detect abnormal conditions and fault occurrences within the DC MG.These approaches have garnered significant attention due to their low computational time,ability to handle fault scenarios, and cost-effectiveness in avoiding the need for extensive communication channels [11].In the quest for reliable operation of DC MGs,this section presents a comprehensive review of various mathematical approachbased fault diagnosis schemes employed in DC MGs.These schemes have encompassed fault detection, classification,and fault location estimation.To provide a systematic analysis, the mathematical-based approaches are categorized into four primary methods: impedancebased, overcurrent-based, derivative current-based, and communication-based schemes, which have been illustrated as follows.

1.1 Impedance-based fault diagnosis schemes

This section introduces impedance-based protection strategies.In impedance-based fault diagnosis techniques,impedance measurements are utilized to identify and pinpoint faults in DC MGs.These methods offer rapid fault detection,proving specifically beneficial in MGs with DG sources.Earlier studies[12-15]focused on fault detection and localization using cable parameters,such as resistance, inductance, and capacitance.The active impedance approach has been used for fault location and classification, providing reconfiguration information upon fault occurrence for isolating P-P-G and N-P-G faults [12,13].In [14], the impedance-based approach discriminated between faulty and non-faulty states via fixed threshold impedance, focusing on P-P-G and N-P-G fault types,which were classified by current outcome signs.A different approach introduced in [15] detected faults through resistance-inductance-capacitance (R-L-C) network parameters,monitored by protection devices through multiple thresholds.Although accurate, incorporating cable resistance and inductance has escalated computational demands.

Recent strategies have adopted individual parameters for fault detection, classification, and location estimation.In [16,17], cable resistance and equivalent circuit resistance-based methods were employed for precise fault detection and location.Zonal-wise DC MG configurations were introduced in[11]to enhance reliability and were validated in real-time for fault detection and isolation.Online fault location estimation was conducted through resistance to tackle faults occurring during supply from both ends and to estimate their fault distance.This approach was specifically useful while considering ring MGs [17].Furthermore, inductance-based fault location methods were introduced in[2,13],estimating line inductance using cable and known parameters from local measurements[12].The authors proposed a novel bidirectional power flow for multi-source DC MGs, surpassing unidirectional analysis.In [18], IEDs were used to capture voltage and current to estimate unknown parameters through least squares for faulty cables.The authors developed an accurate faultaddressing mathematical model based on line inductance[18].In [19], a non-iterative protection scheme using a power probe unit was developed for fault location, particularly during HIFs, using line impedance and attenuation coefficient.

Remarkably, these fault detection and isolation strategies obviate the need for a dedicated communication channel, exemplifying their efficiency and elegance.However,these methods have a few critical limitations,which include challenges in fault classification, the need for complex power units,reliance on predefined thresholds, a potential for false tripping due to improper coordination, and challenges in accurately estimating fault locations in HIF.

1.2 Overcurrent-based fault diagnosis in DC MG

In addition to impedance-based protection schemes,overcurrent-based fault diagnosis methods have garnered attention.These cutting-edge techniques capitalize on the meticulous analysis of current magnitudes to swiftly detect and localize faults within the DC MG.In the overcurrent protection scheme,threshold levels for current magnitudes are predefined.When the current exceeds the predefined threshold, it indicates a fault, and protective devices take appropriate actions to isolate the faulty section [20].In[21-25], the authors proposed overcurrent protection schemes for fault detection and isolation.A gridconnected PV DC MG was examined in [21], simulating faults and diagnosing them on the power systems computer aided design(PSCAD)or Electromagnetic Transient Dynamic Computer(EMTDC)platform,followed by realtime experiments on dSPACE 1104.Overcurrent-based protective relays were developed to interrupt and isolate the fault at the converter [22].The authors assumed the CBs are embedded with a converter to protect the connected equipment, particularly for handling P-P and P-G faults.In[23],a novel communication-assisted overcurrent protection scheme for PV-based DC MGs was presented,serving both islanded and grid-connected modes.The scheme introduced innovative relay operation strategies by setting relays through zones and addressing coordination challenges using a blocking scheme and communication infrastructure.A directional overcurrent (DOC)protection was introduced in[24],capturing current direction and isolating sections during faults.The DOC-based schemes minimized relay response time [25], while combined overcurrent and DOC relays coordinated for cable fault detection and isolation [23,24].

Although overcurrent-based relays have achieved faster fault clearance and detection times for P-P and P-G faults,they have limitations, including complexity in setting appropriate thresholds, unsuitability for low magnitude and HIF, inability to accurately estimate fault location,and inadequacy in fault classification.Additionally,achieving optimal coordination among multiple protective relays can be more complicated, with sensitivity to load variations, limited adaptability to system changes, false tripping due to inrush and noise content signals,and inadequate consideration of fault resistance.

1.3 Communication and differential-based protection schemes for DC MGs

This section explores the communication and differential-based protection schemes for DC MGs, utilizing communication infrastructure and differential measurements for efficient fault detection and accurate localization.Authors have reviewed various publications[26-30] to highlight the latest advancements and benefits of these methods.In [26], the modified squared poverty gap index difference was estimated from the fault current.Communication-based protection for P-P and P-G faults employs IEDs and solid-state circuit breakers (SSCBs),using predefined thresholds via high-bandwidth fiber optics.The authors restored fault zones through SSCB reclosing and classified faults based on current polarity:positive for external faults and negative for internal faults.Superimposed current-based and fault tree (FTR)-based protection schemes were developed using communication between IEDs [27,28].In the FTR method, the reliability index [27] is adopted through Boolean logic to protect DC ring microgrid (DCRM).To derive protection decisions, superimposed currents were presented on the Δiplane[29].In[30],an inverse-time power-based protection scheme for oscillations was presented to avoid communication failure by coordinating with relays to protect the DC MG against faults.Furthermore, the disturbance index was calculated by estimating the sample-to-sample difference to detect the fault.However,these methods exhibit vulnerabilities in communication, design complexity,sensitivity to fault resistance, dependence on disturbance index, coordination obstacles, and false tripping risks.These factors impact reliability, scalability, compatibility,costs, and expertise.Differential current-based protection schemes are employed to address these communication and resistance issues.

To reduce the fault detection time, the differential current is measured from both ends, and the sum of the currents is compared with the threshold to trip the circuit breaker [31].Thereafter,to enhance reliability,cumulative sum(CUSUM)-based differential fault protection schemes were introduced[32,33].In[32],modified CUSUM, which has utilized a window-by-window comparison, was applied along with unit protection to address P-P, P-G,and series arc faults.Inspired by [32], researchers performed backward and forward differential approaches via CUSUM for the IEEE-9 bus DC MG [33].Further,to address the time synchronization error, a multi-sample differential scheme was introduced in [34].This scheme performed multiple comparisons over a sampling window to ensure the stability of high-speed differential protection schemes.Consequently, these differential current-based approaches yielded fruitful results in terms of accurate detection, detection time, and coordination among relays.Moreover, these methods demonstrated the capability to address HIFs.

Despite their advantages, differential protection schemes demand high-bandwidth communication channels to transmit signals and need a backup protection scheme.Furthermore, these schemes are sensitive to variations in system parameters and require precise parameter calibration.Despite these limitations, authors,inspired by [33] and [34], considered difference current for fault detection and location estimation in various MG configurations.This motivated the adoption of differential current protection strategy for fault analysis in[33].

1.3.1 Derivative current-based fault diagnosis in DC MG

Line derivative current-based approaches were introduced to address the limitations of overcurrent-based relay protection.These schemes have involved monitoring the rate of change of current to achieve quick and accurate fault detection and isolation.Various researchers have significantly contributed to implementing these schemes across differently configured DC MGs, as evidenced by studies[34-39].In[34],a novel current rate-of-rise protection strategy for DCRM was introduced,addressing highand low-resistance faults.The derivative current handled high-resistance faults, while an adaptive threshold-based secondary protection scheme monitored low-resistance faults.Validation included a 40 dB noise content during P-P and P-G faults.In [35], small artificial inductances were positioned at line heads to derive the derivative current from captured voltages during DCRM faults.Parameters such as conductor length, measurement error,resistance,and inductance contribute to fault location estimation.A fixed threshold set at 95% of the cable length initiated fault isolation by tripping one circuit breaker(CB), followed by communication with the other CB for faulted section isolation.Notably, accurate fault location estimation is independent of fault resistance.A backup protection scheme was introduced based on local measurements to protect the DC MG in case of communication failure [36].This scheme included derivative and integral characteristics of current to detect and classify P-P and P-G faults.Here, the V-trace tool was used for backup protection, and fault location was estimated.To address these shortcomings,authors introduced adaptive threshold strategies for the first and second derivatives of the current[37,38].In [37], the current derivatives distinguished normal and faulty conditions, with thresholds improved by compensating gains based on line current magnitude.For fault detection within 250 μs, [37] introduced an artificial line inductance method to extend threshold validity.In [38], non-unit protection with first and second derivative currents sidestepped communication delays from local measurements, achieving faster fault detection.To overcome the limitations of derivative current schemes, [38]deployed fixed and adaptive threshold strategies,targeting HIF detection and isolation.

These cutting-edge methods, based on derivative current signals, offer significant advantages, such as quick responses to fault events with higher sensitivity and accuracy, reduced false tripping, and accurate fault location estimation.Moreover, these protection schemes are compatible with system changes and advanced protection strategies.However, despite their numerous advantages,derivative current-based methods have limitations, such as high sensitivity to noise, dependence on signal quality,and the complexity of setting adaptive thresholds.They require careful calibration and tuning, making the relay regulation more challenging due to their faster response characteristics.

Despite their limitations, researchers have drawn inspiration from [38] and [39], leading to the development of modified derivative-based approaches that show promise as effective protection schemes for fault detection and location estimation in diverse MG configurations.Among these, the second-order derivative current-based fault detection scheme demonstrated superiority in [40], which the authors have implemented.Specifically, this paper explores the significance of the second-order derivative of the difference current.This approach enhances the sensitivity to noise-containing signals and enables accurate fault detection and isolation.

2 Comprehensive review of signal processing techniques for fault diagnosis in LV DC distribution network

Signal processing methods have recently gained significant popularity due to their numerous advantages.Motivated by these benefits, researchers have dedicated their efforts to employing sophisticated mathematical algorithms for efficient fault detection and location in DC MGs.Authors have presented a concise review of selected articles in various publications, contributing to fault diagnosis in DC MGs using signal processing techniques.

Generally, SP techniques have been used to analyze,enhance, and extract useful information from timedomain signals, such as current signals.Initially, these techniques transformed time-domain signals into the frequency domain to extract valuable insights.To achieve this, various SP methods have been employed, with the Fourier transform (FT), developed by French mathematician and physicist Joseph Fourier, serving as an initial pioneering approach[41].In FT,the signal is decomposed into sinusoidal frequency components,displaying the magnitude and phases of these components through transformation [41,42].Fault protection based on FT has been proven challenging due to its inability to handle nonstationary signals,as it assumes that the signal’s frequency components remain constant over time [42].Fast FT(FFT),involving discrete data point calculations and computational complexity, was introduced to overcome the limitations of FT [43].FFT computes the discrete FT(DFT)of a signal in the frequency domain.It decomposes signals into their constituent frequency components, facilitating the analysis of various frequencies present in the signal along with their associated magnitudes and phases.Similar to FT,FFT is unsuitable for non-linear signals due to the superposition assumptions.Recently, FFT-based fault protection was introduced in [44,45] by analyzing waveforms and spectra.In [44], arc faults in PV-based DC MGs were diagnosed through the spectrum from FFT calculations, which were used to recognize current interference.Further, FFT with Blackman window interpolation was used for diagnosing faults in switched reluctance motors [45].Short-time FT (STFT) was introduced to mitigate FFT limitations.STFT accommodates linear and non-linear signals, offering superiority over FT,including expedited computation and adaptable sampling rates.Due to these merits, researchers widely employ STFT in real-time applications across diverse domains,encompassing SP, audio and speech analysis, vibration,and biomedical SP[46,47].Specifically,STFT-based methods have been proposed for DC MG fault protection in[48,49].However, STFT accuracy depends on the number of samples per window, FFT points, and sampling frequency.The wavelet transform (WT) was introduced to overcome the limitations of STFT.In [49], WT decomposed signals into scales and frequencies via convolution with wavelet functions, yielding coefficients that encapsulate signal particulars and approximations by capturing both high- and low-frequency data.This multi-resolution strategy has advanced the scrutiny of time-localized and frequency-varying attributes.DC fault detection was conducted in [50,51] using WT.Here, WT enabled P-P fault[50]and arc fault[51]detection,line identification,and discrimination without communication channels[51].Mother wavelet selection was determined by the degree of correlation, indicated by non-zero wavelet coefficients.Specifically, filters were employed through discrete wavelet transform (DWT) to address low- and high-frequency components [51].Continuous wavelet transform (CWT)was employed for fault location estimation using voltage signals,utilizing the Morlet mother wavelet[52].Although WT variants have yielded fruitful results compared to FT variants,they exhibit a volatile nature without selecting an appropriate mother wavelet.

The Stockwell transform (S transform) was introduced in [53] to address the shortcomings of FT and WT.The S transform builds upon STFT and CWT while addressing their respective limitations [53].It refined the time-frequency representation by Gaussian windowing before executing STFT.Further, STFT outcomes have been weighted by the window’s reciprocal standard deviation,enhancing frequency localization.This process optimized the time-frequency analysis for transient and frequencyvarying features.In [54-56], the authors introduced the S transform for detecting faults [54], estimating location[55], and monitoring wind turbine conditions [56].Faults were identified by the high-frequency components, while disturbances were neglected by the zero components.Further, the S transform was modified to address its shortcomings, especially regarding window selection, resolving closely spaced frequency components, and accurately representing rapidly varying signals[57].However,the limitations of the S transform led to the exploration of alternative techniques,such as empirical mode decomposition(EMD).In EMD,the input signal is decomposed into intrinsic mode functions (IMFs) that capture various scales and frequencies adaptively through iterative identification of extrema and mean envelopes [58].Unlike the S transform’s predefined windowing, EMD dynamically adjusts to signal complexities, rendering it effective for handling non-stationary and nonlinear signals.EMD’s self-adaptive nature addresses the limitations of the S transform and offers enhanced capabilities for fault detection and transient analysis.Because of its capability,researchers have adopted EMD for various applications,including vibration monitoring, financial data processing,speech recognition, biomedical signal analysis, and fault diagnosis [58,59].High-impedance arc fault detection [60]and P-P fault detection [61] were conducted in [60,61].However, EMD encounters mode-mixing susceptibility and inadequate frequency-domain representation, hindering its use in DC fault detection.Moreover, its effectiveness is further constrained when dealing with the lowfrequency dominant signals typical of DC faults [61].Ensemble EMD (EEMD) was introduced to overcome these limitations, addressing them through noise-assisted multiple iterations that reduce mode mixing and enhance frequency-domain representation [62].Though EEMD addresses the limitations of EMD,it alone has been unable to detect faults accurately due to potential residual noise interference and challenges in accurately capturing lowfrequency components of a fault current.Further,the Hilbert-Huang transform (HHT) was introduced to address the limitations of EMD and EEMD and attain accurate fault detection [63].HHT is a combination of EMD and the Hilbert transform (HT).Here, EMD decomposes the signal, and HT estimates the instantaneous frequency and Hilbert spectrum for accurate fault diagnosis [63].Because of these advantages, HHT is widely used for various applications,specifically for fault detection in permanent magnet synchronous machines [64], distance protection in high voltage DC(HVDC)lines[65],and forgery detection in digital images [66].Moreover, the marginal Hilbert spectrum is utilized for detecting faults in compound bearing-gear systems [67].As is clear from the aforementioned SP techniques, HHT is more accurate and capable of addressing both linear and non-linear signals.However,these traditional approaches often struggle with fault detection and generally exhibit lower accuracy.

Further,to address limitations of SP,advanced SP techniques, such as local mean decomposition (LMD), variational mode decomposition (VMD), detrended fractal analysis (DFA), and multifractal detrended fractal analysis (MFDFA), were introduced.In [68], LMD enhanced signal decomposition by segregating the local mean elements from oscillatory modes, thus optimizing intrinsic signal extraction.Its adaptability to local signal characteristics enhanced the analysis of non-stationary and nonlinear signals, thereby amplifying fault detection capabilities [68].In [69,70], LMD is utilized for islanding and non-islanding disturbance detection [69] and open circuit fault detection in permanent magnet synchronous generators [70].Though LMD has demonstrated remarkable merits over earlier techniques, it has a few limitations,including overfitting, difficulty in determining local mean components, and impacts on fault detection accuracy.To address these limitations, one of the most powerful and widely popular techniques, VMD, was introduced [71].VMD decomposes signals into modes by minimizing variation while maintaining the frequency attributes[71].Each mode’s frequency spectrum is shifted, segregating the signal into IMFs that adaptively extract frequency components and reduce mode-mixing issues.VMD’s iterative optimization yields dynamic frequency-amplitude modes,useful in fault diagnosis and transient analysis [72,73].Here, VMD was employed for DC MG fault detection[72] and arc fault detection [73] and achieved accurate results within 10 ms.However, challenges for VMD include mode selection for complex signals, sensitivity to noise,penalty factor determination,and the potential need for supplementary optimization algorithms.

Further,to enhance the robustness of SP techniques for fault detection,time-domain techniques such as the Teager energy operator (TEO) [74], CUSUM [75], and Pearson’s correlation coefficient(PCC)[76]have garnered significant attention.These methods have demonstrated the ability to integrate with various SP techniques to detect faults by analyzing the transient and fault characteristics of signals.Generally, PCC measures the relationship between two variables, with its range varying from 1 to 1 [76].In the case of DC faults, where quick and accurate detection is required to isolate the faulty section, TEO is proven superior to CUSUM due to its effectiveness in detecting abrupt changes [77,78].Additionally, PCC is utilized to detect faults in DC ring MGs [79].Advanced methods such as HHT and TEO were integrated in [80] to further enhance fault detection effectiveness, achieving accurate DC fault detection in a PV-wind-based DCRM.Similarly,advanced SP techniques, such as EEMD, LMD, VMD,and HHT, have necessitated incorporating optimization techniques to address their limitations and create robust detection approaches.Different signal processing techniques are represented in Fig.3.

In this review, the authors have introduced various optimization techniques that can seamlessly integrate with SP methods,resulting in advanced and efficient signal processing strategies.

Various optimization algorithms have recently been developed to address real-time challenges, especially by hybridizing them with SP and ML techniques.Among these, swarm intelligence and evolutionary-based metaheuristic optimization algorithms emerged as powerful paradigms, as demonstrated in [80,81].Swarm intelligence algorithms are inspired by the collective behaviors observed in nature, such as the interactions among decentralized agents, such as insects or birds, enabling collaborative problem solving [81-88].Prominent algorithms in this category include particle swarm optimization [81],ant colony optimization [82], salp swarm algorithm [83],and firefly algorithm (FA) [84].Conversely, evolutionary algorithms are grounded in the principles of natural selection and genetic inheritance,utilizing dynamic frameworks that iteratively improve solutions within a population through selection, crossover,and mutation.Unlike swarm intelligence, these algorithms leverage genetic mechanisms to optimize solutions dynamically.Their self-improvement mechanism enhances precision,robustness,and adaptability, making them effective across a wide range of applications.Notable evolutionary algorithms include genetic algorithms (GAs) [85], cuckoo search algorithm [86], sine cosine algorithm [87], and whale optimization algorithm(WOA) [88].

Fig.3.Different signal processing techniques.

Among evolutionary algorithms, WOA stands out by emulating the hunting strategies of whales, offering a well-balanced exploration and exploitation mechanism.This distinct approach enhances convergence speed and improves solution quality, surpassing traditional evolutionary and swarm intelligence algorithms [88].WOA’s ability to efficiently locate optimal solutions while avoiding local optima, combined with its adaptability across various domains, highlights its critical role in optimizing fault detection and signal processing applications, among other examples.However, despite its numerous advantages,WOA is sensitive to control parameter tuning, and,although it maintains a balance between exploration and exploitation, it might still encounter challenges with local optima.To address these limitations,an improved version of WOA (IWOA) was introduced, which minimizes the objective function of the L-kurtosis index (LKI).This enhancement strengthens parameter control, reduces susceptibility to local optima, and accelerates convergence.For fault detection in DC ring MGs (DCRM), the hybridization of VMD with IWOA offers a powerful solution.By integrating VMD’s signal decomposition capabilities with IWOA’s optimization strengths, this hybrid approach achieves high accuracy and reliability, ensuring swift fault and islanding detection in DCRM.This combination addresses the critical need for fast, accurate, and reliable detection in modern signal processing applications.

From the earlier discussion, it is clear that popular SP techniques, such as EMD, EEMD, HHT, and VMD,require support from other algorithms,which may include a simple decision tree or any classifier, to diagnose DC faults.Unlike traditional SP techniques, fractal analysisbased SP techniques have recently garnered more attention from researchers due to their significant advantages,notably their ability to achieve accurate classification without any supporting algorithm.Fractal analysis has been used to extract the hidden characteristics of fault current signals in terms of scaling exponents, such as fluctuation analysis and rescaled-range analysis,which are statistical tools[89].Among the various methods of fractal analysis, DFA has attained widespread popularity for detecting and classifying non-stationary signals.DFA serves as a transformation tool to compress large datasets into low-intensity fluctuation curves, and it is widely used in scientific research due to its long-range detection capability and ability to remove fluctuation trends from signals.Moreover, DFA has extracted intrinsic statistical features from time series by eliminating extrinsic polynomial trends in various differential orders.Because of these advantages,DFA has been applied across various fields,including neuroscience, geophysics, and biomedical signal analysis[89,90].Additionally, it has been employed in power systems to diagnose faults [91].However, DFA faces challenges during fault diagnosis, such as abrupt jumps in detrended profiles from disconnected fitting polynomials,non-monotonic behavior, sensitivity to large datasets and non-stationary signals,and a lack of insight into local fractal components.MFDFA was introduced [92,93] t o address these limitations.MFDFA offers substantial advantages by capturing varying scaling behaviors across different fractal components of a signal, allowing for a more nuanced exploration of complex data.It has provided insights into the multifractal nature of signals,enhancing the ability to differentiate between distinct fluctuation patterns and enabling a more comprehensive characterization of multifractal systems.

However, one concern with MFDFA is its application to fixed polynomial orders alone, making it unsuitable for higher-order polynomials in the detrending process.To overcome these limitations, modified MFDFA has been presented in [94] for accurate fault classification in PV-wind-based DCRM.The authors have addressed fault detection using MFDFA, along with a secondary approach in the form of a sliding mode window, which has strengthened fault detection capabilities even for low-and HIFs.This method has exhibited superior performance compared to traditional signal processing techniques by utilizing trigonometric functions to estimate the fluctuation function.

3 Comprehensive review of ML techniques for fault diagnosis in LV DC distribution networks

The substantial advantages inherent in ML techniques have recently spurred their widespread adoption.Driven by these merits,researchers have actively deployed various ML techniques, such as K-means, K-nearest neighbor(KNN), restricted Boltzmann machine (RBM), random forest, SVM, multi-layer perceptron, extreme learning machine, ANN, recurrent neural networks, convolutional neural networks, deep belief networks, random vector functional link networks, and broad learning systems, for efficient fault detection and location in DC MGs.The protection applications and ML methods are visualized in Fig.4.This section presents a comprehensive analysis of the ML techniques used for fault diagnosis within lowvoltage DC distribution networks, highlighting their significance, advantages, and limitations in fault detection and classification tasks.The discussion encompasses their applications across various fault types,data preprocessing methods, feature extraction techniques, and performance metrics.Nowadays, K-means clustering analysis has gained significant importance for classifying various signals.Grouping data points into clusters based on similarities has aided signal pattern recognition and data segmentation[95].The iterative process has refined cluster centers, improving accuracy in applications such as image processing, genomics, market segmentation, clustering documents for text analysis, and anomaly detection[95,96].Moreover,researchers have utilized K-means clustering to classify internal and external faults in HVDC transmission lines [97].Thereafter, KNN was introduced for classification by identifying the k nearest neighbors from the training dataset using a chosen distance metric when a new data point is presented.The algorithm assigns the class label by majority vote among these neighbors,making KNN suitable for various applications such as image recognition and collaborative filtering [96,98].RBM was introduced in [99] to address the earlier limitations.RBM is an unsupervised neural network model with visible and hidden layers.It has learned by sampling hidden nodes activated based on inputs, and in the reconstruction phase, visible nodes have reconstructed data.RBM has excelled in feature learning, collaborative filtering,and dimensionality reduction[99,100].However,it has struggled with capturing long-range dependencies in data and has suffered from overfitting on small datasets.Further, SVM has been introduced for classification [101].SVM has transformed data into a higher-dimensional space to achieve fault classification via a hyper plane that maximizes the margin between classes in feature space,minimizing misclassifications.The procedure for the general ML technique is visualized in Fig.5.

RF has been an ensemble algorithm for classification and regression.It has built multiple decision trees using bootstrapped subsets and random feature selection and has combined their predictions for more accurate results[102,103].This approach has provided significant advantages over existing algorithms, such as reducing overfitting, handling noisy data, and enhancing generalization[103].Researchers have applied RF to classify and detect DC faults in PV-based DC MGs using local and master detectors [104,105].Nonetheless, during Simulink and real-time validations,it has faced challenges like computational intensity, limited interpretability, and scalability,attributed to the PV arrays’ data length.Further, to address the limitations of RF and enhance classification accuracy, MLP has been developed.MLP has been a feed-forward neural network with input, hidden, and output layers interconnected by weighted connections[106,107].During training, the network has adjusted weights using algorithms like backpropagation (BP) to minimize error.Due to its ability to capture complex relationships, MLP has been widely used for tasks like classification, regression, and pattern recognition.Researchers have used MLP for estimating the fault location in LVDC ring MGs[108]and for arc fault classification and location estimation[108].However,MLP has suffered from overfitting with insufficient training data and has required careful tuning of hyperparameters for optimal performance.

Advancing the field, emerging techniques like ANN,CNN, RNN, and DBN have been introduced [109-112].ANN has utilized interconnected layers to learn complex patterns from data, enabling it to handle various tasks[109].Seamlessly interlinking, RNN has specialized in sequential data by retaining memory through recurrent connections, making it suitable for time series and sequence tasks [111].Transitioning to CNN, it has been tailored for image data by employing convolutional layers to automatically learn hierarchical features [110].Lastly,DBN has consisted of generative units for unsupervised feature learning, later fine-tuned for classification tasks[112].Due to their substantial advantages, these ML techniques have been adopted in various fields, including DC MG fault diagnosis [113-117].Researchers have used ANN for DC fault detection and location estimation,specifically focusing on P-P and P-G faults by varying fault resistance[113].For fault detection and classification in the DC MG, CNN has been utilized [114], achieving 99%accuracy for two fault types: P-P and P-G.A distinctive form of RNN,known as the nonlinear auto-regressive exogenous(NARX)neural network,has been employed to safeguard DC MGs from cyber attacks, particularly identifying false data injection aimed at manipulating the voltage profile and associated DERs [115].Further, DBN has been utilized for fault diagnosis in EV DC charging points,successfully identifying different fault types [116].However,these methods have suffered from various limitations.They have been unable to classify a broader range of faults, while RNN’s vanishing gradient issue has affected long sequences.CNN has been prone to overfitting small datasets, and though DBN has had the capability to handle multiple classes,its complex layer-wise pre-training has hindered practicality and increased computational time.

Fig.4.Representation of ML methods and power system applications.

Fig.5.General procedure for machine learning technique.

To enhance the performance of ML, researchers have introduced ELM and its variants, such as kernel ELM(KELM), multi KELM (MKELM), and online sequential ELM(OS-ELM)[117-120].ELM has exhibited a structure similar to MLP, but here, the input-to-hidden layer weights have been randomly assigned,and output weights have been analytically determined to minimize the training error [117].This approach has enabled ELM to achieve rapid training times and good generalization performance.Traditional ELM has been extended by incorporating kernel functions to capture complex nonlinear relationships in data, leading to KELM [118].By incorporating a kernel trick,KELM has enhanced model expressiveness and simplified weight calculation, enabling effective handling of nonlinear patterns.To further improve KELM’s performance, MKELM has been introduced by integrating multiple kernel functions into KELM to capture different aspects of complex data relationships,making it more versatile for real-world scenarios [119].While KELM and MKELM have garnered significant importance in various applications, they have possessed limitations such as kernel selection, and the careful combination of kernels has led to overfitting and increased computational complexity.To overcome these challenges, OS-ELM has been introduced in[120]to process data chunk by chunk and update the model as new data arrives without retraining the entire model.As a result, it has enhanced scalability, reduced memory usage, and suited real-time applications.Due to these advantages,OS-ELM has been adopted for diagnosing DC faults in the DC MG.However, its performance has primarily depended on the selection of chunk size and learning rate.

The variants of ELM, particularly OS-ELM, have demonstrated significant potential across various applications.However, these techniques have not been directly suitable for diagnosing faults in complex signals, such as those encountered in DC MGs.To overcome this limitation, hybridizing OS-ELM with signal processing (SP)techniques has been necessary, and they have been explored in Section 5.

To address the aforementioned limitations and enhance the performance of ELM and its variants,RVFLN variants and BLS variants such as kernel RVFLN (KRVFLN),multi KRVFLN (MKRVFLN), and online sequential(OS-RVFLN), incremental BLS (INBLS), and cascaded BLS (C-BLS) have been introduced.In RVFLN, weights have been randomly generated between the input layer and hidden layers, with no need for iterative training, and the hidden nodes have been connected through activation functions [121].Also, there has been a direct link between input and output [122].During training, the output layer weights have been calculated analytically using the Moore-Penrose pseudoinverse,minimizing error.However,in ELM,there have been no direct links between input and output.As a result, RVFLN has reduced the overfitting risk, efficient learning, and scalability to large datasets.However,RVFLN has also had limitations,including fixed random weights,challenges in generalizing to complex relationships,lack of adaptability to changing data,and potential oversight in fine-tuning [122].To mitigate those limitations, a kernel function has been introduced in KRVFLN[123].Here,the kernel function has been utilized to capture complex patterns and relationships in data and also avoids the selection of hidden nodes and mapping functions.Though it has enhanced the accuracy,it has suffered from the selection of the kernel, and to tackle that,MKRVFLN has been introduced in[124].In MKRVFLN,multiple kernels have been incorporated with RVLFN(MKRVFLN).As a result, it has reduced overfitting and enhanced classification accuracy.Further, to address the limitations of RVFLN variants, OS-KRVFLN has been introduced[125].As with OS-ELM,here also data has been transformed chunk by chunk, enabling continuous online learning and model updates for adapting to evolving patterns.This has resulted in superior nonlinear feature learning,enhanced generalization,and seamless dynamic model updates.However,it has also faced learning rate and chunk size problems.To tackle that,the forgetting factor has been introduced to OS-RVFLN.The advantages of RVFLN and its variants have led to their adoption in various applications.However, in the context of fault diagnosis in DC MGs,only OS-RVFL and OS-KRVFLN have been exclusively applied due to their inherent capabilities.Other RVFLN variants may not have been used due to limitations that necessitate hybridization with SP techniques, as discussed in Section 5,to ensure their suitability for DC fault diagnosis.

Further, BLS is introduced to overcome the shortcomings of ELM, ELM variants, RVFLN, RVFLN variants,etc.BLS is a single-layer incremental model and an alternative to deep learning, extending from RVFLN [126].In recent days, has emerged as a prominent algorithm in various fields for detecting and classifying various datasets, owing to its numerous advantages.However, BLS also has shortcomings like random generation of weights,feature mapping, and computation intensity.To enhance the performance of BLS, INBLS and C-BLS are introduced.In INBLS,an incremental approach is used for fast remodeling the network without retraining the process by increasing the enhancement and feature nodes [127,128].Further, C-BLS involves stacking multiple BLS layers to form a deep architecture.Each layer learns increasingly abstract features from the input data, enabling enhanced pattern recognition and complexity in a hierarchical manner [128].The various machine learning techniques are applied in power system protection is illustrated in Fig.6.

4 Review of hybrid models for DC faults diagnosis

Though machine learning algorithms have enough advantages and have been capable of accurate detection and classification, the solo classifier of ML has also encountered various limitations, including less accuracy,more computational time for large datasets, sensitivity to noisy data, lack of interpretability, vulnerability to adversarial attacks, and difficulty in handling and dealing with complex and imbalanced datasets.To address these limitations, various researchers and scientists have combined two or more algorithms to attain their advantages.Especially, various researchers have found that the hybridization of ML, SP, and OA techniques has provided remarkable results in various applications.

Fig.6.Various machine learning techniques are applied in power system protection.

For fault detection in DC MG,FT and ANN have been hybridized,especially for arc fault diagnosis[129].In[130],the least-square method (LS), PSO, and FFT have been hybridized for fault diagnosis.LS has been used for fault current envelope estimation,PSO for LS attenuation coefficient adjustment, and then FFT has been applied to the resultant.However, this has raised computational complexity,parameter tuning needs,and system efficiency concerns.In[131,132],arc fault detection has been carried out by hybridizing FFT and STFT [131], and introducing a variable window approach to STFT for diagnosing DC faults [132].For accurate fault detection in a DC MG,WT has been hybridized with ANN [133].Further, DWT and TEO have been hybridized for accurate fault detection[134].However,they have not been suitable for fault location estimation,which has increased the delay in detection due to ANN training.Recently, S Transform has been hybridized with the Gaussian function [135], and Kalman filter [136] for diagnosing the DC faults.However, they have demanded data quality, increased complexity,required careful parameter tuning, and have had limited generalization.In [137], empirical wavelet transform(EWT) and SVM have been hybridized for DC arc fault detection by decomposing the oscillatory frequency components,which have been generated during the fault.Further, to enhance the performance of EMD, improved EEMD has been introduced by hybridizing with adaptive noise [138], and fuzzy systems [139] for fault diagnosis.However, they have not been suitable for more fault classification and have increased the computational complexity due to fuzzy logics.To overcome the limitations in EMD and its variants, improved HHT-based approaches have been introduced [140,141].In [140], HHT has been hybridized with traveling wave theory to compensate for its limitations during fault location estimation.In HHT,instead of EMD decomposition, VMD has been utilized for fault classification [141].However, false tripping for noisy signals and inability to classify and estimate fault location has occurred.DC arc fault detection has been carried out by improved LMD in [142,143].LMD has been integrated with multi-scale fuzzy entropies [142], and SVM [143].Though integrating results have enhanced the accuracy, they have had a few limitations, such as mode mixing issues, sensitivity to noise, lack of more information about the rest of the faults, and an inability to trip CB for HIFs.Further,to enhance the performance,VMD has been hybridized with PSO and SVM for classifying DC faults [144].Here, VMD key parameters have been optimized by PSO, and then SVM has been used for DC fault classification.Moreover, VMD and S Transform have been combined for fault location estimation in HVDC transmission lines [145].For obtaining efficient classification accuracy, cross VMD has been hybridized with FA and reduced kernel ridge regression [146].They have not been suitable for islanding detection in DC MGs and have been unable to classify more faults.To address those limitations, DFA has been hybridized with adaptive VMD (AVMD) in [85] for islanding and nonislanding detection and classification.Here, adaptive VMD has been obtained by using PSO for modes optimization.Though AVDM-DFA has exhibited remarkable results in the MATLAB environment, it has been unable to perform fault location estimation and relay tripping operations due to a lack of information regarding local fractal components.Further, MFDFA has been introduced for highlighting the shortcomings of DFA in response to fault diagnosis [88].Along with MFDFA,TEO and CUSUM have been adopted for accurate fault detection based on the threshold.This approach has had shortcomings, like delayed detection time, TEO being unable to deal with a large number of samples, and an inability to perform fault location estimation.To compensate for those limitations, authors have introduced a new approach, i.e., SMW-PCC based MFDFA in [147].

In recent times, many popular ML techniques have required hybridization to demonstrate remarkable performance in the field of fault protection in DC MGs.Among them,a few major hybrid models have been demonstrated in this review with respect to fault diagnosis [148-155].Fault detection and classification have been carried out through modified KNN [148].Here, principal component analysis has been used to reduce the sizes and eliminate the singularity of the outcomes from linear discriminant analysis.Thereafter, to overcome those limitations, CNN has been hybridized with a transfer learning-based approach in [148] for online fault detection and location estimation in the DC MG.Here, transients of voltage and current signals have been transferred for pre-training and then subjected to CNN for online fault detection and location estimation.Further,ELM has been enhanced by increasing layers, i.e., ML-ELM, and its parameters have been optimized through the ant lion algorithm[149], Nelder-Mead simplex with KELM [150], and MKELM with VMD [151] for fault diagnosis.Here,MKELM with VMD has attained more accurate results,but it has also faced both critical optimal parameters in VMD and multi kernels.To overcome that, improved OS-ELM has been hybridized with hybrid FA (HFA)and introduced in [152].Here, the improved nature of OS-ELM has been obtained by initializing the adaptive chunk size.As a result of HFA-improved OS-ELM, most of the shortcomings of the ELM variants have been addressed.However,learning rate and computational time have been more with HFA-improved OS-ELM, and estimating fault location has been difficult.In[153],expanded exponential robust RVFLN (EE-RRVFLN) has been introduced to monitor the PV and battery in a MG.Here,MPPT control has been observed by distributed adaptive droop control.Further, to address those limitations,KRVFLN has been hybridized with adaptive empirical WT(AEWT)[154,155].Here,FA has been used for selecting the significant features, which have been extracted from the AEWT, and then fault detection and classification have been estimated in terms of sensitivity and overall accuracy.However, the selections of exponential and mother wavelet have still been challenging tasks.Moreover,to address those limitations,authors have introduced HHT-TEO-based OS-RVFLN with FF for DC fault detection and classification in [156].Here, the HHT-TEO approach has been utilized for triggering the circuit breaker based on the threshold value for isolating the faulty section from the healthy section.Thereafter, OSRVFLN with FF has been used for fault classification and location estimation by utilizing the extracted significant features.The outcomes of this approach have shown a significant way for exploring the hybridization techniques for fault analysis.

As discussed earlier, BLS has attained remarkable results in various fields compared to other ML techniques.Moreover, to the best of the authors knowledge,there have been no articles to date diagnosing DC faults based on BLS.From the above discussion, BLS has also had limitations that need to be addressed to attain enhanced accuracy in terms of detection and classification.For that purpose, authors have introduced AVMD-improved BLS in [157].Here, the AVMD-IBLS framework has effectively addressed the dual challenge of fault detection and classification, including the critical and intricate task of identifying distributed generation(DG) islanding events.These events have been particularly significant as they require the precise detection and timely isolation of the utility grid from the standalone MG to ensure system stability and safety.By accurately classifying faults and detecting DG islanding events, the proposed approach has not only enhanced diagnostic reliability but has also paved the way for more efficient management of MG operations.This work has underscored the potential of hybridized techniques in advancing the state-of-the-art in DC fault analysis and has opened new avenues for research in the integration of BLS with advanced signal processing methodologies.The pros and cons of the existing literature have been demonstrated in Tables 1-4.Also, how machine learning,optimization and other statical methods are classified is represented pictorially in Fig.7.

5 Discussion

From the Tables 1, 2,and 3,the reviewed strategies for DC microgrids highlighting their respective strengths,limitations,and application areas.These strategies are critical for maintaining system stability and reliability during faults, especially under dynamic and uncertain operating conditions.The wide range of methods from conventional threshold-based approaches to advanced data-driven techniques reflects the evolving complexity of DC MGs due to increasing integration of distributed energy resources and power electronic converters.This section discusses the key insights from the literature,offering an organized view of the protection methods.The ultimate objective of this review is to support researchers and engineers in selecting or designing effective, adaptive, and scalable protection schemes for modern DC microgrid environments.

5.1 Overview of protection methods

In this paper, DC MG protection methods are categorized into four types i.e., mathematical, signal processing,machine learning, and hybrid approaches.Each of these strategies is tailored to address the unique challenges associated with DC systems,such as the absence of natural current zero crossings and the dynamic behavior of power electronic converters during fault conditions and islanding conditions.Researchers and scientists have focused their efforts on improving the accuracy,reliability,and speed of fault detection,classification,location estimation,and the isolation of the healthy section from the faulted section in DC microgrids.The protection strategy applied in microgrid is illustrated in Fig.8.Grouping the major protection methods into these categories clarifies how each approach contributes to protection of DC microgrid, and the key insights and characteristics of these strategies are explored as follows:

5.1.1 Mathematical based approaches

Typically, these approaches rely on equations for fault detection and location estimation, and they exhibit high accuracy for simple faults.However, these approaches often face challenges with low-magnitude faults and noise interference.For example,the least-squares method struggles to identify multiple simultaneous faults, while active impedance estimation may give incorrect results when nearby cables are quickly de-energized.Similarly,although directional overcurrent and superimposed current-based methods improve the speed of fault detection,they remain sensitive to transient and high-resistance faults, highlighting the need for reliable backup protection strategies.

5.1.2 Signal processing techniques

In order to overcome the limitations of conventional mathematical methods under noisy and complex scenarios,signal processing-based techniques have seen a significant trend shift to overcome such limitations.Signalprocessing-based techniques provide increased sensitivity and accuracy by extracting useful features from fault current signal data,thus proving to be more suitable for fault detection in today’s power systems.Methods like FT,WT,and HHT have proven to be very promising in the analysis of high as well as low-frequency fault signal components.Though FT and its extension FFT are widely employed,their stationarity assumption restricts their potential in the identification of intricate, time-varying faults.Conversely,techniques such as STFT,WT,and HHT offer better adaptability for non-stationary signals STFT, for example, can identify high and low impedance faults within under 2 s.However, these techniques are also having their own limitations such as tend to have high computing requirements and are liable to be sensitive to noise.For instance, HHT performs well with nonlinearsignals but has limitations such as mode mixing and end effects that lower its credibility for faults with short durations.New methods such as VMD, LMD, and feature mode decomposition (FMD) [x] have better fault characterization by maximizing mode extraction and coping with signal complexity better.Yet, they still present some challenges, such as parameter tuning and sensitivity to distortions, especially in the case of high-impedance faults.

Table 1 Mathematical approaches for fault-diagnosis.

S.No.& Ref Methodology Pros Cons Mathematical approaches for fault diagnosis Cable of drive system during fault.No need of communication 1 [130] Least-square based parameter estimation Unable to classify more than two faults.Difficult to detect if fault current magnitude is less.Active impedance estimation 2 [12] No need of communication.Fault zones will be protected.Fault clearance and reconfiguration will be obtained instantly.Backup protection scheme is required.Mal-operation due to instant de-energization of adjacent cables.Overall cost will be increased.3 [13] Unified-Impedance based estimation 4 [14] Accurate detection and classification of both shortcircuit and ground faults.Unable to recognize the type of fault except internal and external faults.Unable to deal with noise signals.Impedance based estimation Improved fault detection time and localization accuracy.Able to classify both AC and DC faults.5 [15] Threshold settings are estimated for each CBs.It can applicable for noise related scenarios.Unable to recognize the type of fault except P-P-G and N-P-G.Unable to use for single polar system.R-L-C based estimation 6 [16] Very useful for real-time fault diagnosis.Able to clear quickly in case of P-P faults due to its zones.Multiple thresholds are required and difficult to analyze arc faults.Delay in fault detection.Resistance based estimation 7 [19] Probe power unit based protection scheme.Able to detect HIFs and their location in radial MG.Relative error is very less.Difficult to analyze HF and location estimation.Moreover, exhibited delay in detection.Not suitable for LIFs.Also not addressed short circuit faults location.Accuracy depends on network parameter sensitivity.8 [22] This method more suitable for P-G and P-P faults.Detection time is very less.More sensitive to load variations and low magnitude and HIFs.inadequacy in fault classification.9 [23] Communication based overcurrent scheme Overcurrent protection scheme 10 [24] It reduced the relay response time during fault.It improved the security of MG Complexity in setting appropriate threshold.Limited adaptability to system changes.Unable to protect when the fault is low magnitude.Directional overcurrent scheme It can applicable for both grid-connected and island modes.Blocking scheme overcomes the co-ordination problems.It is cost-effective.Coordination is a main challenge among relays.False tripping due to inrush and noise signals 11 [25] Modified squared poverty gap index based fault estimation Restore fault zones by SSCBs reclosing to clear the fault.Accurately classified internal and external faults.Unable to deal with arc faults and HIFs.Enabling IEDs during low fault currents is challenging to isolate the fault zone.12 [26] Superimposed currentbased FTR method External faults are isolated by using reliability index.IEDs allows to isolate the fault cable, at the instant of fault.It is more sensitive to fault resistance and load variations.Failure of line segments due to communication causes severe damage.13 [28] Sample-to-sample estimation through disturbance index is more accurate.No need of communication channel.Accurate fault detection Inverse-time power based protection scheme False local data causes false tripping.Depends on frequency oscillations.It necessitates efficient monitoring and enhances cost of operating.14 [31] Fast fault detection and requires very few samples.Doesn’t require any backup protection system.Capable of arc faults detection.Difficult to analyze low magnitude faults.It is sensitive to noise and window selection is crucial.Its applicability is limited.16 [33] Current rate-of-rise protection strategy Modified CUSUM based protection scheme 15 [32] The usage of CUSUM for differential current gives rapid fault detection.Only fault zone is isolated.Selection of threshold is difficult for various faults.If window length is more, the delay in detection time.Differential Current Method Effectively addressed high and low resistance faults with derivative current.Adaptive threshold is useful for accurate detection.More sensitive to noise signals.More complexity in setting adaptive thresholds.Careful calibration is required.Secondary protection enriches the cost.17 [36] Limited faults classification.Relying solely on locally measured current signals.Implementation of real-time is difficult.Multiple location faults are unable to address.Local Measurements-Based Backup Protection Provides continuous power supply to loads.More robust in noise conditions.Even communication failures the faults are detected through backup protection scheme.It is adaptable to variable conditions.18 [38] First and second order derivative used for fast fault detection.Communication delay is avoided.Incepted faults are cleared within milliseconds.Non-unit Protection scheme Scalability and adaptability of the scheme.Its accuracy of detection depends on signal quality.Regulation of relays more challenging due to faster response.

Table 2 Comparison among Advanced signal processing techniques.

S.No.&Ref Methodology Pros Cons 1 [41-45] FT and FFT FT is incapable to handle non-stationary signal due to its assumption of constant frequency.FFT is not suitable for non-linear signals due to superposition assumptions.These approaches leads to potential risks 2 [47,48] STFT Capable of detecting both HIFs and LIFs within 2 s.It utilizes high frequency components.Ability to differentiate between fault types and transient events.FT and FFT are capable to filter the noise for increasing the detection time.Easy to implement.More suitable for AC faults.The accuracy depends on the number of samples in a window.It requires accurate parameter tuning.Computational burden increases due to processing of frequency components.3 [50] WT It captures both high and low-frequency data.Offers multiresolution analysis.It can capture signal variations over time and frequency simultaneously.4 [51] DWT Arcing and non-arcing signals are accurately classified.Thresholds are fixed through signal-to-noise ratio.More suitable for diagnosis of instant changes.5 [53-57] S-transform Internal and external faults are classified accurately.Due to consideration of high frequency components for fault detection its capability is enriched for both AC and DC fault diagnosis.Reduces the cost and computational burden.6 [58-61] EMD It adjusts the complexities of a signal.Suitable for both nonstationary and nonlinear signals.Accurately detected P-P and arc faults.Enables early fault detection.Maintenance cost is reduced.Selection of appropriate wavelet is crucial.Potential sensitivity to noise and signal distortions.Its accuracy violates when the signal consist noise.Not applicable for all faults.Delay in detection time.Selection of mother wavelets requires additional approach.It enhances the computational burden.Selection of window impacts the detection accuracy.Relies on advanced mathematical techniques.Requiring expertise to analyze signals.Potential sensitivity to variations of parameters.Difficulty to implement in real-time applications.Encounters mode mixing susceptibility.Inadequate frequency-domain representation.Reliance on measured data.Unable to utilize for uncertain conditions.Neglecting the low frequency components causes loss of information.7 [62] EEMD Reducing mode mixing suitable for non-stationary and nonlinear signals.Frequency-domain representation is enriched by multiple iterations with white noise.Reduce the maintenance cost.Improves reliability.Challenge in capturing the low-frequency components.Increases the computational burden due to iterations and noise.It leads false tripping if it consist noise content.Careful tuning is required for selection of ensemble trials with added noise magnitude.8 [63-67] HHT It is more suitable for non-linear and non-stationary signals.The HHT spectrum can captures fault events in the signals.It is more compatible to integrate with other techniques.It shows mode mixing and end effect for short-term duration signals.HHT spectrum is sensitive to noise signals.For large data sets it is not suitable due to high computational burden.9 [68,69] LMD It is suitable to identify islanding and non-islanding phenomenon.It offers localized fault frequency analysis and facilitates the interpretation.Difficulty in determining the local mean components.Selection of window size and threshold values impact the detection accuracy.Over fitting also reduces accuracy.10 [70,72] VMD Selection of IMFs and penalty factor requires additional optimization approach.While dealing with HIFs, it exhibited computational burden.Selection of threshold is It is capable of diagnosing transients also dynamic frequency-amplitude modes.It optimizes the modes according to user defined.It is versatile and applicable to both AC and DC faults diagnosis.It is more robust against noise.difficult for HIFs and large fault datasets.11 [74] TEO It is more sensitive to load variations,uncertainties in source power and the noise content.Selection of improper threshold leads false tripping.It is not suitable for complex nature signals.12 [81] PSO Capable to deal with complex optimization problems.The Teager energy of the signal provides more flexible for fault analysis.It requires three consecutive samples for fault detection.Faster in response in both simulation and realtime.Simple to implement by tuning few parameters.It is compatible to integrate with other techniques.Converges quickly at near optimal solutions.Not suitable for high-dimensional or multimodal optimization.Not suitable for discontinuous or non-smooth fitness functions.Sensitive to the initial population and swarm size.Limited Exploration of search space.13 [83] SSA It balances exploration and exploitation effectively.It is more robust to noise in fault data.Its parallel processing enhances diagnosing accuracy.It is relatively simple to implement and allows versatility with other techniques.14 [85] FA Due to large search capabilities it is suitable for complex problems.More robust to noise and uncertainties.It can flexibly navigates non-convex optimizations based on landscape fitness.15 [86] GA It efficiently searches for global optimal values.It maintains population diversity to ensure various solutions.GA exhibits robustness to noise in fault data.It exhibits premature convergence and suboptimal results.Difficulty in handling constraints.Careful tuning is required to balance exploration and exploitation.Exhibits suboptimal results with multimodal fitness functions.Its capability is limited and is not suitable for complex and high-dimensional datasets.Careful parameter tuning is required for optimal results.It is difficult to identify a gene when it falls in premature convergence.Selection of initial population and encoding scheme impacts efficiency.Selection of mutation operator and crossover is essential.It exhibits high computational complexity.(continued on next page)

Table 2 (continued)

S.No.&Ref Methodology Pros Cons 16 [87] SCA It is more flexible to combine with other algorithms.Easy to implement.Reduced execution time.17 [88] WOA This approach effectively balances exploration and exploitation, possesses global optimization capability,demonstrates versatility in handling faults, and offers a simple implementation.18 [84] DFA This approach based on fractal analysis and it has the capability of detection and classification of non-stationary signals with long-range correlations.Also removes fluctuation trends and extracts intrinsic features in time series from the signal.19 [86] MFDFA The ability of capturing the scaling behaviour of different fractal signals including local and global to enhances the detection accuracy.Ability to differentiate the multiple fluctuation fractals.Also, applicable for complex datasets.20 [158] FMD Feature mode decomposition (FMD) method is very much useful in early fault feature detection.It separates mode effectively under uncertain conditions.21 [159] SVMD Successive variational Mode decomposition (SVMD) gives better decomposition outcomes as compared to existing techniques like empirical mode decomposition and variational mode decomposition.It doesn’t have theoretical convergence evidence.Difficulty in handling constraints.Requiring careful adjustment for optimal performance.It exhibits slow convergence speed in complex problems,and requires careful for coefficients of encircle prey and spiral updation.Also exhibits poor performance for highdimensional data sets This approach exhibits abrupt jumps in detrended profiles,non-monotonic behavior, sensitivity to large data and nonstationary signals, computational complexity, sensitivity to parameters, limited interpretability, and a lack of insight into local fractal components.Unable to address local fractals.It is not applicable for higher-order polynomials.Unequal division of scales results poor performance.It is only suitable for time-series data sets.It suffers from mode mixing under dynamic conditions.SVMD takes more computational time due to iterative procedure.Also, parameters for this method need to optimize otherwise may not get better outcomes.

5.1.3 Machine learning techniques

While signal processing techniques enhance fault detection, their performance often depends on feature extraction and expert defined thresholds for selecting IMFs.In order to resolve these shortcomings and strengthen adaptability towards handling complex fault data, different approaches of machine learning are being used today.Traditional methods such as SVM, KNN, and RF have already been used extensively for fault classification owing to their ability to learn patterns from large datasets.Recently, new approaches like Gradient Boosting,XGBoost, RVFLN, and BLS pretrained models optimize speed and accuracy, particularly when working with unbalanced or higher dimensional data.Notably,RVFLN and BLS are recognized for their rapid training and suitability for online learning, making them promising for real-time applications.Despite their advantages, these ML techniques require careful optimization and can pose challenges in real-time deployment due to computational demands and the risk of overfitting.Nonetheless, the integration of ML into power system protection continues to evolve, offering promising solutions for improving system reliability and fault management.

5.1.4 Hybrid approaches

Even though hybrid models have been introduced to address the limitations of standalone signal processing and machine learning methods, their real-time implementation remains a significant challenge.Techniques like DWT-TEO and MKELM with VMD enhance fault detection by combining precise signal decomposition with robust classification, demonstrating strong performance across AC/DC faults and islanding scenarios.Similarly,OS-ELM with HFA and EE-RRVFLN employ dynamic clustering and feature selection to achieve high accuracy in detecting complex and imbalanced fault conditions.However,these models often require substantial computational resources and careful parameter tuning, which hinders their applicability in real-time systems.Specifically,while MKELM with VMD and EE-RRVFLN, variants of BLS and hybridized BLS models perform well in noisy environments and power-sharing tasks,their accuracy may degrade when dealing with unbalanced data or during hardware integration.Although hybrid models offer significant potential by combining the strengths of multiple techniques, further refinement is needed to improve their adaptability and reduce computational complexity in real-world applications.This can be achieved through boosting strategies and optimization algorithms capable of efficiently tuning parameters for increasingly complex models.

Table 3 Comparison among machine learning techniques.

S.No.& Ref Methodology Pros Cons 1 [97] K-means It is more suitable for large datasets.Easily interpretable and facilitating efficient implementation.KNN It is efficient for normalized datasets and remains accurate with added data.It requires no training period, is easy to implement, and executes quickly.3 [100] RBM This classifier autonomously learns features from data,handles missing data, and complements supervised learning techniques.It operates without the need for labelled data during training.RBM reconstruction enables data denoising,and feature visualization.2 [98]4 [101] SVM It is more accurate due to its transformation to higher dimensional space.It is more suitable for small datasets and robust against overfitting.5 [102] RF It is more robust against overfitting and capable to handle the noise data.More suitable for feature selection and provides accurate and fast faults classification.6 [107] MLP Due to learning of complex non-linear relationships, it is used in various applications.It has the capable to learn and extract suitable features from the raw data.It has capable of approximating any continuous function.7 [109] ANN It can learn from data and adapt their behavior without changing programming.It is capable of protecting MG during dynamic changes.It is more suitable for intermittent faults.8 [114] CNN CNNs automatically learn relevant features from raw input data, capturing spatial relationships for accurate fault localization.They utilize pooling layers for higher-level feature extraction and demonstrate fast learning, particularly with small datasets.9 [115] RNN RNNs excel in handling sequential data, allowing for early fault detection.They can accommodate input sequences of variable lengths and maintain memory of previous inputs to enhance the capability.10 [116] DBN It can handle labeled and unlabeled data effectively.It can manage large datasets, and stack with other models for better fault detection.They also excel in managing missing data and reducing dimensionality, enhancing fault CA.This classifier is not suitable for discriminating multiple classes,its reliance on Euclidean distance,and the requirement for a predefined number of clusters.Additionally, it struggles with low data samples, which can lead to unreliable results.Its performance depends on neighbor selection and isn’t ideal for unscaled datasets with varying dimensions.Additionally,computing distances for large datasets can be computationally expensive.It struggles with capturing long-range dependencies in data and can suffer from overfitting on small datasets.Its performance depends on the hyperparameter choices.It has limited applicability.It is not suitable for both end measurements.Exponential features may lead to false fault location.Not suitable for large datasets and difficult to analyze noise related signals.Selection of kernels is a challenging task.It can be computationally intensive and memory-consuming,especially for large datasets.Additionally, tuning hyperparameters can be necessary, and interpreting results may be challenging due to its ensemble nature.It suffers from overfitting with insufficient training data and requires meticulous hyperparameter tuning.Additionally, it can be computationally intensive, particularly for large datasets, and may struggle with noisy or insufficient data.Requires BP for weights computation Due to the many parameters in ANNs, complexity and overfitting increase.It can be difficult to interpret and understand the decisions behind the fault classification.Its performance heavily relies on the quality and quantity of training data.ANNs can be computationally expensive and time-consuming to train.Complexity is more due to multiple layers.It is not suitable for less fault samples classification and isolation of the fault.Also,suffers from fragmentation problem.Its performance is heavily depends on the quality of the training data.Not suitable for noisy or insufficient data.Its black-box nature hinders interpretation.RNNs may encounter exploding gradient problems during training, potentially leading to difficulties in capturing longterm dependencies and resulting in false fault detection.RNNs may struggle with unsupervised training on large datasets,and their training can be computationally expensive compared to ANNs.Additionally,the finite memory of previous inputs may limit RNNs’ ability to capture extensive contextual information.DBNs may increase overfitting and are best suited for larger datasets,requiring extended training times.They may struggle to classify faults independently and need output layer tuning with supplementary algorithms for optimal performance.Additionally, their computational complexity and reliance on random weights can lead to lower accuracy.11 [117] ELM The hidden layer parameters in ELMs are randomly chosen and cannot be fine-tuned, which may lead to overfitting or underfitting, particularly with small datasets.The use of random weights can result in lower accuracy.Additionally,ELMs may struggle to capture complex hierarchical features effectively in fault diagnosis.(continued on next page)Simple in architecture, fast training, no need of hyperparameter tuning at hidden layers.It can perform well on unseen data after training.It is less susceptible to overfitting issues.

Table 3 (continued)

S.No.& Ref Methodology Pros Cons 12 [118] KELM Kernel function is used to capture the complex patterns,reduces the overfitting risk, it is suitable for large-scale applications.It has the capable of mapping non-linear data into a higher-dimensional space using kernel trick.The selection of kernels poses a significant challenge and demands more resources,particularly with complex kernels.It exhibits poor performance with the noisy data or real-time fault diagnosis.It has limited flexibility, and interpretability challenges.13 [119] MKELM Multiple kernels are used to capture the complex and nonlinear relationships in dataset.Enhances the CA as compared to other ELM variants.Capable to deal with large and complex datasets.Selecting and combining multiple kernels can be challenging,requiring careful tuning and potentially impacting computational resources.The training process of MKELMs may be computationally expensive due to the combination of multiple kernels,demanding more resources.MKELM can be difficult to interpret, further complicating model analysis.14 [120] OS-ELM It may suffer from instability issues related to matrix inversion and require regularization techniques to mitigate them.Continuous model updating can lead to overfitting, especially with MG fault data.It may not be suitable for all fault detection tasks and can exhibit computational burden with complex or very large datasets.Performance of OS-ELM depends on the selection of chunk size and learning rate.15 [122] RVFLN It offers rapid training and testing and can handle both noisy OS-ELM efficiently manages large datasets, ideal for big data applications.It demonstrates good generalization performance, even with limited training data, while reducing memory usage.With its sequential learning approach, OSELM trains data quickly, making it suitable for real-time applications.and noise-free datasets effectively.The use of direct links enhances performance and reduces overfitting.It is scalable to large datasets and captures nonlinear transformations for faults through functional links,expediting model development in fault detection.The performance of RVFLN is contingent on selecting appropriate regularization parameters.Nonlinear transformations introduced by functional links add complexity, aiding in generalization to complex relationships but potentially limiting adaptability to changing data.Additionally, the fixed random weights of RVFLNN may require boosting to optimize model performance.16 [123] KRVFLN It eliminates the need for selecting hidden nodes and mapping functions by utilizing kernel functions.These kernels capture complex and non-linear relationships from the fault signals,enhancing the adaptability of RVFLN.Kernel selection poses a challenge in KRVFLN, requiring careful consideration.Kernelization introduces redundancy,necessitating fine-tuning and influencing model performance.The chosen kernel significantly impacts KRVFLN’s sensitivity to dynamic conditions.17 [124] MKRVFLN It has the capable to capture wide range of complex and nonlinear relationships through multiple kernels.It reduces overfitting and enhances fault CA.It is more robust to noise and data variability.Able to deal with unseen fault dataset.Integrating multiple kernels complicates the model and makes selecting and tuning kernel parameters challenging.It is not suitable for small datasets and also struggle with classifying imbalanced fault datasets.OSKRVFLN 18 [125] It updates online with each data chunk,negating the necessity for retraining.It effectively addresses unseen faults and excels in nonlinear feature learning and generalization.OS-RVFLN’s memory efficiency allows it to handle large streaming fault data with high performance.Kernel function enhances OSKRVFLN’s ability to filter out noise and disturbances for diagnosing faults accurately.Its performance is depends upon the selection and tuning of kernel parameters.It struggles to adapt to significant changes in fault patterns, requiring expertise in dynamic fault environments.It updates based on recent data, potentially losing historical fault patterns,and its performance is sensitive to hyperparameter tuning.Sequential learning with small datasets leads to suboptimal results.19 [127] BLS It is simple in structure, easy to build, faster in training and testing, reduces the computational time.It can continuously learn and adapt to new data without retrain the entire network.It can efficiently handle high-dimensional data.It utilizes random mapping to create a vast pool of features.Exhibits more accuracy even compared to deep learning methods.It uses an iterative training procedure to adjust the weights in the network.Careful tuning is required for hyperparameters in BLS.BLS lacks control over specific features generated through random mapping.Potentially higher computational complexity.Integration of other learning techniques causes overfitting.Challenges in handling imbalanced datasets effectively.The redundancy and oversampling of weights.20 [160] The multiclass adaboost algorithm is chosen due to its ability to adaptively update weight values by iteratively calculating the error rates of weak classifiers.Multiclass adaboost Unlike traditional AdaBoost, this is limited to binary classification.The multi-class adaboost relies on inadequate classifier selection and is computationally demanding.21 [161] XG boost The extreme gradient boosting (XGBoost) algorithm has strong generalisation performance and noise tolerance.Utilising the concepts of feature column sampling and data sampling, XGBoost not only speeds up training but also successfully avoids over-fitting.If appropriate regularization and validation procedures are not used,XGBoost may overfit when trained on tiny datasets.

Table 4 Comparison among hybrid models.

S.No.& Ref Methodology Pros Cons 1 [134] DWT-TEO It enables early fault detection,especially DC shortcircuit faults.Capable to distinguish arc faults and cable faults.Enhances the detection accuracy.2 [137] EWT-SVM It is more suitable for HIFs detection and classification.More robust to noise related signals and ability to handle non-stationary signals.3 [138] Improved EEMD This approach eliminates the mode mixing problem.Effectively distinguishes the both LIFs and HIFs.Essential features are extracted from the singular points of mutation and cumulative slopes.It adopts noise to handle the noise signals.It suitable for various applications.4 [140] HHT based Travelling wave Accurately estimates the fault location.The characteristics of time-frequency and propagation provide clear information of fault.Adoptable to complex MGs.Reduces false alarms.5 [142] LMD-SVM LMD extracts crucial information in frequency bands.Effectively classify the DC arc faults through features.Fault detection accuracy increases.More robust to noise and non-stationary signals.Capable to deal with real-time applications.6 [144] VMD-PSO-SVM It is capable to classify both cable faults and ground faults.VMD outcomes are more reliable for fault detection.Capable of dealing with different types of signals contributes the various domains.It offers robust fault detection across various scenarios.7 [85] AVMD-DFA Enhances the detection and classification accuracy.Capable of detection and classification of nonstationary signals with long-range and local correlations.Adaptable to various faults in different MGs.More robust to noise and non-stationary signals.Capable to detect islanding and nonislanding events.8 [87] MFDFA-TEO Capable of detecting and classifying even low magnitude faults by extracting features from time series of fluctuations and trends.Enhances fault CA by effectively capturing the scaling behavior of different fractals from the signal.Suitable for realtime applications, reducing delay in detection time due to fewer samples.9 [148] Modified KNN Principal component analysis is used to reduce the size, eliminating singularity from the signal.It processes 5000 samples per second, reducing features.Suitable for large datasets, it doesn’t require accurate mathematical modeling, making it suitable for LIFs in MGs.ML-ELM with Ant-Lion algorithm 10 [120] Critical ML-ELM parameters optimized for accuracy.Effective for both internal and external fault classification.Improves fault diagnosis accuracy.Optimization enhances global search capability.Captures fault-specific information from vibration data.Handles complex relationships and large datasets effectively.Improper selection of mother wavelet for decomposition leads to suboptimal results.Not suitable for fault location estimation.Not suitable for HIFs detection and classification.It introduces computational complexity, careful tuning is required for hyperparameters.It can only identify fault and non-fault events.Enhances the computation time while dealing with real-time applications.It enhances the computational time while handling with large and real-time data.Improper selection of thresholds and hyperparameters lead to suboptimal results.Limited interpretability and transparency during decision process.Not suitable for fault location estimation.When samples sizes are low, it leads to overfitting.Not suitable for faults classification.Performance relies on about fault characteristics and it leads to suboptimal results.Careful tuning is required for hyperparameters.It requires comprehensive testing for validation.Sensitivity to noise, and not suitable for low resistance faults.Exhibits computational complexity while handling large datasets.Selection of thresholds for fault detection is a challenging task.It requires additional optimization for optimal hyperparameters.Not suitable when there are more features.It introduces more computational burden during training.Not suitable for fault location estimation.Accurate parameter tuning is required.selection of kernel is a challenging task.Difficulty to implement in real-time.Difficult to identify fractals for detection in real-time environment.For parameters tuning in AVMDDFA expert is required.Not suitable for fault location estimation and relay tripping.Unable to identify the unseen data.Not suitable for HIFs detection.Not suitable for polynomial functions.Difficulty to understand the fractal components and time-domain analysis.Not suitable for incident fault location estimation.Due to fewer samples processing for each time it is unable to adopt large datasets.Not suitable for noise or variations in the data.Enhances the complexity of the MG.Fine-tuning of this ML models require expertise.It demands computational resources for training and testing.It relies heavily on the closest data points for classification.Not suitable for real-time fault diagnosis.Expertise needed for hyperparameter fine-tuning.Suboptimal results with distant ants.Unable to extract sensitive information from low-magnitude faults.Not suitable for real-time implementation.Slow convergence impacts MLELM training.Requires meticulous optimization.(continued on next page)

Table 4 (continued)

S.No.& Ref Methodology Pros Cons 11 [152] MKELM with VMD Easy implementation, with performance depending on just two easily tunable parameters.Coordinated algorithms ensure convergence to optimal solutions.Can handle both AC and DC faults, detecting islanding and non-islanding events.Suitable for various MG configurations.Threshold selection challenging for low-magnitude faults.Susceptible to noise interference.Limited capacity to capture distribution system complexities.Kernel selection poses a significant challenge.Further research needed to tailor it for arc fault diagnosis.Not suitable for fault location estimation due to kernelization.Inapplicable for imbalanced systems.12 [153] OS-ELM with HFA Chunk-wise data processing ensures accurate fault classification.Significantly reduces PV tracking error.HFA minimizes randomness in OS-ELM,enhancing robustness.Improves power quality and fault detection.Dynamic input clustering for better adaptation to changing environmental conditions.It exhibits high learning rate and computational time.Real-world implementation may face additional challenges.Performance relies on hyperparameter selection and tuning.Stable voltage profile is not a concern.Unsuitable for fault location estimation.Unable to handle imbalanced datasets.13 [154] EE-RRVFLN Not suitable for detecting low-magnitude faults.Inapplicable for imbalanced datasets.Unsuitable for real-time implementation, increasing computational burden.Suboptimal results in dynamic conditions.More costly for real-time deployment.PV and wind MPPT control achieved via droop control, reducing tracking error.Efficient features selected through FA, enhancing classification accuracy.Suitable for various MGs, capable of classifying up to 6 fault classes.Extracts sensitive information from signals.Effective power sharing capability.14 [155] KRVFLN with AEWT Capable to detect DC cable and ground faults with high accuracy and sensitivity, even in noisy environments.Addresses coordination limitations in low-voltage DC MGs, preventing false tripping.Fast decision-making with Gaussian kernel functions for high-dimensional search spaces,handling noise and volatility.It is computationally sound, but exhibits delay in detection.Not suitable for real-time implementation due to complexities.Requires careful fine-tuning of hyperparameters.Additional optimization algorithms increase complexity.More sensitive to environmental changes.15 [162] ICEEMDAN with Multiclass adaboost Effective decomposition of nonlinear and nonstationary dc signals is made possible by improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN),which improves feature extraction for precise islanding detection in microgrid systems.The multi-class AdaBoost method facilitates the detection of different islanding and non-islanding occurrences and offers good classification accuracy.Because of its high processing cost, ICEEMDAN could not be practical for real-time applications without optimization.The quality and selection of characteristics from IMFs are crucial to the technique.AdaBoost needs a diversified and welllabeled dataset, which might be challenging to collect in real-world microgrid settings.15 [154] ACVMD with RKRR Its intricacy may restrict its interpretability and simplicity of use, though, as it necessitates meticulous parameter adjustment and an adequate amount of labelled data.By efficiently classifying and decomposing complicated signals, the combination of Reduced Kernel Ridge Regression (RKRR) and Adaptive Cross (ACVMD) provides reliable and accurate monitoring of islanding and power quality issues in microgrids.

5.2 Performance under different conditions

The effectiveness of fault protection strategies in DC microgrids is influenced by multiple interrelated factors.These include the nature of faults, system architecture,and dynamic operating conditions:

5.2.1 Influence of fault type and location

The detection and classification of faults significantly depends on their type of fault i.e., P-P, P-G, HIFs, etc.,and location within the network.Faults near the source tend to generate higher current magnitudes and are easier to detect, whereas faults at remote nodes or with high impedance may produce subtle signatures that complicate identification.The absence of natural current zero-crossing in DC systems further adds to the challenge, requiring more sensitive and accurate detection methods.

5.2.2 Impact of network topology and converter behavior

The structural configuration of the DC microgrid whether radial, ring, or mesh-combined with the presence of power electronic converters, influences fault current characteristics.Converter control strategies, switching speeds, and limited fault current contributions can alter system dynamics during disturbances.Protection methods must therefore consider both the static topology and the real-time behavior of converters to ensure accurate fault detection and coordination.

Fig.7.Representation of machine learning, optimization and other statistical methods for fault diagnosis.

Fig.8.The protection strategy applied in microgrid.

5.2.3 Effect of transient disturbances and islanding events

External transients,such as load fluctuations,switching operations, and measurement noise, can obscure fault signatures and lead to false positives or delayed responses.Additionally, unintentional islanding alters protection boundaries and affects relay coordination.Reliable fault protection must distinguish between genuine faults and non-fault disturbances while ensuring timely islanding detection to maintain system stability.

5.3 Emerging trends and gaps

Despite notable technical advancements,fault detection and protection strategies continue to face critical challenges.One key development is the incorporation of AI and ML techniques particularly models such as CNNs,RNNs, and RVLNs which enable adaptive learning,allowing systems to respond effectively to dynamic and evolving fault conditions.In parallel, hybrid methods are gaining attention by combining ML algorithms with signal processing techniques.This integration enhances protection accuracy by improving the extraction and classification of relevant features, especially in noisy electrical environments.The use of BLS frameworks within these hybrid approaches also reduces response time, enabling near real-time fault identification even in decentralized or remote DC microgrids.However,several critical gaps persist.Many existing methods depend on fixed thresholds or assume uniform fault characteristics, which compromises their reliability under varying load conditions, converter control modes,and during islanding events.Fault localization also remains a challenge while faults near the source are relatively easier to detect and locate, those occurring farther along the feeder or near distributed loads often produce weak fault signatures,complicating accurate identification.Furthermore, the absence of lightweight, lowlatency models that can be deployed on embedded edge devices limits the scalability of protection systems.To address these issues, future frameworks must be contextaware, dynamically adaptive, and capable of coordinated operation with converter control systems.

5.4 Future research directions

There are several areas for future research aimed at overcoming current limitations and enhancing the performance of fault detection and protection systems in DC microgrids:

5.4.1 Automatic selection of informative IMFs in VMD-based methods

Current VMD implementations rely on manually set parameters and fixed mode counts,often leading to redundant or misleading features.Future research should focus on developing adaptive optimization frameworks to automatically identify the most fault-relevant IMFs, potentially using entropy measures or sparsity constraints.

5.4.2 Design of ultra-lightweight hybrid models for DC microgrid protection

While hybrid models improve accuracy, their size and complexity often limit real-time deployment in DC microgrids.A promising direction is the development of simplified architectures that integrate shallow classifiers with low-order transforms,optimized for microcontrollers or FPGA-based systems,without sacrificing detection precision.

5.4.3 Threshold-free fault detection using dynamic feature learning

A major limitation of existing fault detection methods is their reliance on static thresholds for classifying faults,which often fail under dynamic conditions such as load transients, converter switching noise, or changing fault resistances.A promising direction is the development of unsupervised or self-organizing learning frameworks such as clustering-based fault detectors or adaptive entropy models—that automatically learn fault-relevant patterns in real time.These frameworks could eliminate the need for fixed thresholds by adapting to evolving statistical or structural changes in the signal, enhancing robustness in both HIF and LIF scenarios.

5.4.4 Dynamic islanding detection for DC microgrids

Fault signatures in DC microgrids change during islanding due to altered fault current paths, reduced magnitudes, and dynamic converter behavior, while most protection methods overlook converter control dynamics.Future research should focus on context-aware models that integrate real-time converter states (e.g., control loop outputs, reference current dynamics), load profiles, and islanding status to intelligently distinguish between actual faults and converter-induced transients.This could be achieved by combining signal-based features with control-layer variables using hybrid models or decisionlevel fusion frameworks improving both fault detection accuracy and selectivity during islanded operation.

By addressing these gaps and focusing on emerging trends, future research can significantly enhance the reliability, accuracy, and efficiency of fault detection and protection systems, paving the way for safer and more resilient DC microgrids.

6 Conclusion

This paper provides a comprehensive review of protection strategies for DC microgrids (DC MGs),highlighting their evolution through a year-wise analysis of key research publications.A structured overview of DC MG protection is presented,encompassing fault characteristics,detection methods, isolation mechanisms, classification techniques,and fault location estimation.The reviewed literature is organized into four primary categories: mathematical approaches, signal processing methods, machine learning techniques, and hybrid models each offering unique strengths and limitations in addressing challenges of DCMG protection.Through this organization of approaches, the paper identifies prominent trends such as the rise of intelligent, data-driven solutions and the integration of real-time signal processing for enhanced fault diagnosis in dynamic microgrid environments.Notably,the review underscores several research gaps, including the limitations of static threshold-based methods, challenges in locating faults, which are incipient away from the source,and the need for lightweight,low-latency models deployable on embedded systems.By projecting these developments and limitations,this review serves as a foundation for guiding future research toward more robust,adaptive, and context-aware protection schemes.It encourages the development and simulation of practical hybrid models capable of real-time deployment,ultimately contributing to the safe,efficient,and resilient operation of upcoming DC microgrids.

CRediT authorship contribution statement

Kanche Anjaiah:Writing-original draft,Methodology,Investigation. Jonnalagadda Divya: Writing - review &editing,Conceptualization.Eluri N.V.D.V.Prasad:Visualization, Formal analysis. Renu Sharma: Visualization,Supervision, Conceptualization.

Ethical approval

Accept the guideline of journals.

Availability of data and materials

No data available.

Funding

No funding available.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

This review was conducted as part of academic work and received no external funding.The authors thank their institution for academic support.