A Comprehensive Review of Intelligent Islanding Schemes and Feature Selection Techniques for Distributed Generation System

Detection of unintentional islanding, defined as inadvertently separation of distributed generators (DGs) from the utility grid, is a major challenging issue for modern distribution networks. Islanding detection becomes problematic especially when the local generation matches or closely matches the local load. Therefore, there are strict requirements for accurate, fast, and reliable islanding detection of renewables and DG-based systems. Various islanding schemes have been proposed in the literature, which can be categorized as remote, local, and intelligent-classifier-based schemes. Recently, intelligent schemes have gained attention due to their superior properties and advantages relative to traditional approaches. This paper overviews the shift in research from traditional schemes to intelligent islanding schemes. It also highlights the major obstacles, challenges, advantages and disadvantages, and future research directions of intelligent schemes. In this study, the intelligent-classifier-based islanding detection schemes presented over the last decade are analyzed objectively and comprehensively from all aspects of islanding detection. This research further highlights feature selection schemes and the most common parameters used for islanding detection. Finally, based on a detailed and critical analysis, the findings and potential recommendations are presented.


I. INTRODUCTION
Interconnection of distributed generators (DGs) has gained attention due to the electricity market deregulation, capital investments, requirement of reliable and better power quality (PQ), and environmental concerns. DG integration can reduce transmission and distribution losses, generate revenue from excess power, low or zero emissions (for renewables such as wind and solar), and capability of handling power operation during the absence of the main utility [1]. In the conventional electrical power system (EPS), the production of power is centrally operated, and power is delivered to customers through transmission and distribution networks. The primary disadvantages of conventional networks are their high cost and transmission losses, environmental issues, and the unidirectional flow in the network [2], [3]. However, DG interconnection also poses challenges such as elongated payback times, the intermittent nature of renewables, and glitches in the power system [4]. In addition, DG integration can also result in unbalanced voltage and frequency along with power quality problems [5].
With enhanced penetration of DGs, detection of unintentional islanding in the power system becomes a non-trivial task. Unintentional islanding occurs when the DG gets separated from the main utility without any planned intention, i.e. due to the tripping of the circuit breaker (CB) [6]. The tripping of circuit breaker could occur due to system failure, unbalanced power, line outage, generator tripping, human error, natural disasters, and other disturbances [7]. Failure to detect this issue can result in severe consequences for both the system (damage to DG and related equipment) and human life (maintenance workers and consumers). The islanding problem becomes more severe when the local generation matches or closely matches the local load. As per the IEEE 1547-2018 standard, islanding should be detected within a period of 0.16 to 2 s [8]. Therefore, islanding detection should justify these requirements of dependability, security, and fast response time. The islanding scenario in DG integrated EPS is shown in Fig. 1, once the CB1 is opened, the DG is isolated from the rest of the system and becomes the only available source for local loads. The islanding issue has been under study for several years, and different studies have been conducted to overcome this major issue of power systems with DGs. Various islanding detection schemes (IDSs) have been introduced in the literature. Each type has its pros and cons based on the non-detection zone (NDZ), detection time, and PQ. IDSs can be classified into three main categories, as shown in Fig. 2. Remote IDS employs communication infrastructure for the detection of islanding. A communication link is deployed between the DG and main utility, this communication link requires additional instruments [9]- [11]. These instruments are generally high-cost sensors, telecommunication tools, and control systems. The remote IDS have comparatively higher system and running costs, compared to active and passive techniques. Therefore, for small-scale systems, remote IDSs are not suitable; however, these IDSs are commonly used for high-scale projects. The upside of remote IDS is that they have zero NDZ, no degradation of PQ, and can handle complex DG integrated EPS [12]. Some examples of remote IDSs are the power line carrier communication (PLCC) [13]- [15], supervisory control and data acquisition method (SCADA) [10], [16], [17], signals produced disconnect (SPD) [18]- [20], transfer trip schemes [21]- [23], impedance insertion method [21], and phasor measurement unit (PMU) [24]- [29].
Local schemes are based on monitoring various electrical parameters such as voltage, current, frequency, and power in addition to the injection of disturbances in DG-EPS for islanding detection. IDS are further categorized as active, passive, and hybrid islanding detection schemes. Active IDSs utilize external disturbances by injecting a troubling signal into DG output, this external signal injection introduces variation in system parameters [30], [31]. By calculating the variation in system parameters with thresholds, active methods detect islanding. The merits of active IDSs are low NDZ and lesser detection time, while they require additional setup for disturbance injection and may harm the PQ of DG-EPS. Various active IDSs are available in the literature, for example impedance measurement (IM) [32]- [34], active frequency drift (AFD) [35]- [38], Sandia frequency shift (SFS) [39]- [42], Sandia voltage shift (SVS) [43]- [46], and sliding mode frequency shift (SMFS) [31], [47]- [50].
Passive islanding is another commonly used IDS for DG-EPS, where system parameters are monitored at the point of common coupling (PCC), which shows variations when the utility is isolated from the DG-EPS [51]- [53]. Based on these variations, the protective relay operates to detect islanding by comparing it with the defined threshold values [54], [55]. Passive IDSs are economical and uncomplicated schemes that pose no harm for PQ, thus considered as realistic solutions for DG-EPS. The downside of these IDSs is their large NDZ and need for thresholds [56]- [58]. The passive IDSs used in literature are over/under voltage and frequency methods (O/U V&F) [59]- [62], the rate of change of frequency/power (ROCOF/P) [63]- [67], total harmonic distortion method (THD) [18], [68]- [71], and phase jump detection (PJD) methods [68], [72]- [75].
Hybrid IDSs are a combination of active and passive schemes. The PQ problem in active IDSs and large NDZ in passive IDSs can be resolved using the hybrid IDSs [76]- [78]. The downside of hybrid IDSs is complexity and higher detection time. Some hybrid schemes are presented in the literature which are based on positive feedback and voltage unbalance, SFS and ROCOF, voltage unbalance and frequency set point, voltage, and real power shift, RACOV and ROCOP, and hybrid SFS and Q-f Curve IDSs [70], [79]- [83].
Islanding detection based on remote and local schemes has its advantages and disadvantages [84]- [87]. Remote schemes require a communication interface and are reliable for large systems but impractical for small systems because of their complexity and higher cost [10], [88], [89]. In contrast, local schemes are simple, easily applicable, and are low-cost but they have some disadvantages, such as active methods have noise and PQ issues while passive islanding schemes have a large NDZ and low speed [36], [71], [90]- [92]. With bugs and pitfalls in local and remote schemes, intelligent classifier-based schemes are gaining more attention. The key reasons for shifting towards intelligent classifier-based IDS are the exemption from threshold settings, no noise and PQ problems, low NDZ, high speed, and no communication channel intervention, which make intelligent schemes more reliable and acceptable.
A. PREVIOUS STUDIES Some researchers have conducted review studies on islanding detection strategies while concentrating primarily on traditional islanding systems such as remote, local (active and passive), and signal-processing-based schemes. A few researchers have reviewed intelligence-based islanding schemes. In [5], islanding schemes were categorized into four different groups: remote, local, signal-processing-based, and intelligent classifiers. A total of 85 research publications were reviewed and classified in this research. The comparison of different islanding schemes was also illustrated in [93]- [96].
Another review research for islanding detection of microgrids was conducted in [19], where the major focus was only on local and remote IDS. In this study, various performance indices like (NDZ, detection time, PQ) and other technical problems were also reviewed and analyzed. A study on the computational intelligence-based islanding detection of DGs was presented in [97]. This review summarises islanding strategies focused on conventional and computational intelligence-based schemes. It also presented a comparison of the performance of intelligence-based and traditional schemes [98]- [100].
The shifting trend in islanding detection from classical methods to machine learning-based methods was introduced in a comprehensive survey [101]. This study reviewed the basics of the islanding issue, types, test standards for islanding detection, and the reasons for the trend shifting from conventional to machine learning-based IDSs. The reasons highlighted in this research are the consequent threshold selection, fast detection, NDZ, PQ, and robustness in dealing with complex conditions, which make machine learning-based methods a rational choice for islanding detection [102].
A comprehensive analysis of modern islanding schemes for a DG network in terms of merits, efficiency, efficacy, and feasibility was presented in [103]. It also presented an investigation of different schemes by comparing their time of detection and the computational burden [20], [51], [104], [105]. A detailed review of various islanding schemes with their strengths and weaknesses was presented in [106]. In this research, the analysis was classified into three major groups: classical methods (local and remote), signal-processing methods, and intelligence-based methods. Various schemes were compared and assessed for future recommendations based on different performance metrics such as detection time, accuracy, and efficiency [107]- [109]. A review of the local and remote islanding detection techniques, with their advantages and disadvantages, was reviewed for the DG-based system in [110]. The technical issues of islanding detection proposed by different researchers were reviewed and compared in a detailed way in [1]. This study compared all the IDSs considering their advantages and disadvantages compared to each other and recommended future trends in the field of IDSs.

B. CONTRIBUTION
The main inspiration for this work is the availability of extensive literature and the shift in interest from traditional islanding schemes to intelligent classifier-based schemes over the last decade. The studies discussed in the previous paragraphs have stressed mainly traditional islanding schemes, their benefits, and drawbacks. The above review papers have not covered specific research focused only on intelligent islanding schemes. The research trend and recent advancements in state-of-the-art intelligent islanding schemes for the power system have shown a great deal of interest. It is also important to carry out a critical and thorough analysis of intelligent islanding schemes based on this void and the need for time. VOLUME 9, 2021 In intelligent islanding schemes, feature selection is a key factor affecting accurate and reliable classification, and this vital aspect of islanding schemes has not been studied previously. Furthermore, there exists no study on the electrical parameters that have the greatest influence on islanding detection, as reported in the literature. Therefore, the contribution of this study in comparison with existing review studies is outlined as follows.
• A detailed overview of islanding issues, various types of IDSs, test systems, and standards for islanding detection in DG-EPS is presented.
• A systematic and comprehensive analysis of islanding detection schemes based on various intelligent classifiers is performed along with their pros and cons.
• Different feature-selection schemes used for islanding detection and those that may be used in the future are outlined.
• The most important electrical parameters used for islanding detection are highlighted, based on different methods available in the literature.
• Comparison of various intelligent-classifier-based schemes is carried out based on performance indicators such as accuracy, NDZ, and detection time.
• Finally, potential guidance and recommendations have also been provided after the review of the IDSs.
The rest of this paper is organized as follows. Section II presents a brief introduction of the islanding types and their effects. In section III, the islanding test standards and test system topologies are presented. A complete survey and assessment of various state-of-the-art intelligent schemes studied from the last decade are presented in section IV. Section V presents the feature selection schemes for various IDSs. The most useful and dominating features for islanding detection are presented in section VI. In section VII performance indicators are discussed. Comparison, discussion and future recommendations are presented in section VIII, and conclusions of this study are presented in section IX.

II. INTENTIONAL AND UNINTENTIONAL ISLANDING
The word islanding refers to the separation of an operating power system having both loads and generation from the central utility grid. In an EPS, the islanding phenomenon is categorized into two main types: 1) intentional islanding and 2) unintentional islanding. Intentional islanding is a planned operation used primarily for system maintenances and operational issues, whereas unintentional islanding events are caused by sudden faults, load switching, and CB tripping because of a main utility power outage.

A. INTENTIONAL ISLANDING
Intentional islanding is the anticipated and systematic separation of the main grid from the power network to avoid a major breakdown [111]. The main reason for this separation is to overcome the blackouts and cascading problems of a power system. Cascading failure can be described as a fault process that results in the tripping of another element of the grid successively [112]. This separation is also employed for system maintenance, voltage enhancement, power-loss scenarios, temporary faults, and improving the efficiency of the network. Intentional islanding is viable and worthwhile for the power system. Several researchers have worked on intentional islanding and control over several years to make it practically applicable and to shield the power system from the harsh outcomes of unintentional islanding [92]. It makes the power grid stable in a manageable island area to support the rest of the system. Effective and stable operation in islanding mode requires a balanced load and generation.
Numerous algorithms have been established for the detection and division of power systems to stable islands. Some of these include the wide-area measurement systems (WAMS) algorithm for a wide-area blackout [113], security-based method [114], effects of degraded communication, and load variability on-grid splitting [115], improved spectral clustering for a multi-objective controlled islanding system [116]. The algorithms used in different studies include ant search mechanism [117], fast greedy and bloom algorithms [118], comprehensive learning particle swarm optimization (CLPSO) algorithm [119], two-stage stochastic optimal islanding method [120], mixed-integer linear programming (MILP) [121], algebraic graph method [122], and ordered binary decision diagram (OBDD) method [123]. In addition to these methods, several countries have their own intentional islanding microgrid systems for realizing safe and reliable distribution networks [124].

B. UNINTENTIONAL ISLANDING
Unintentional islanding occurs without any planned intention or prior information from the main utility grid or independent power producers. There could be several reasons for this unknown tripping, but the dominating drivers are the occurrence of faults at the main utility grid, system failure on the utility side, human error due to negligence, and natural disasters [125]. The harsh outcome of this unintentional islanding is the separation of the DGs from the main power system. In power systems, several islanding standards have been prescribed to overcome this critical issue [126]. A detailed survey of international standards on islanding is presented in the next section. The undesirable consequences of unintentional islanding include the inability of DGs to handle the abrupt change, synchronization issues after re-connection, and the uncontrollable behavior of the DGs during islanding because of the load and generation mismatch [127]. Unintentional islanding poses a threat to the security of power systems, which can harm utilities, equipment, and maintenance personnel. The record of mega outages and power system failures due to unintentional islanding shows that a lack of power supply has contributed to major lapses in security and significant economic downturns [128], [129]. In this article, IDSs are critically analyzed to resolve these challenges and highlight the sensitivity of the security against anti-islanding.

III. TEST STANDARDS AND SYSTEMS FOR ISLANDING DETECTION A. TEST STANDARDS FOR ISLANDING DETECTION
Due to these islanding issues in the power system, several islanding detection standards have been set by different researchers and organizations. Several standards are being used globally for islanding detection considering specific quality factors (QF), detection time, acceptable frequency, and voltage ranges. These standards are used to direct the installation and construction of the power systems and distribution networks during islanding scenarios [130]- [135]. Table 1. summarises the aforementioned international standards based on the thresholds of the QF, acceptable frequency range (AFR), acceptable voltage range (AVR), and detection time. Several countries have their standards for islanding detection, as presented in Table 2. The most widely used standards and benchmarks are IEEE 1547-2018, IEEE 929-2000, UL 1741, and IEC 62116. The variance in the thresholds of the QF, AFR, AVR, and detection time among the various global standards are because of differences in power network and distribution frequency criteria in different countries/regions.

B. TEST SYSTEM FOR ISLANDING DETECTION
Similar to islanding detection standards, there are several test systems for islanding tests. There is no global standardized and definite test system; different countries and organizations use their test systems for islanding detection. These systems include single DG, multiple DGs, same type DGs, different DG types, and hybrid-type test systems. Every country has its distinct test system for analyzing the consistency and practicality of its networks under islanding conditions.
One of the most commonly used and generic test systems is shown in Fig. 3, which is based on IEEE 1547-2018, IEEE 929-2000, UL 1741, and IEC 62116. This is the recommended test system for analyzing islanding conditions by opening the CB to isolate the DG from the main utility. A parallel resistor-inductor-capacitor (RLC) load is connected as the load for the test system because the RLC load is considered as the worst-case scenario for islanding detection.

IV. INTELLIGENT-CLASSIFIER-BASED ISLANDING SCHEMES
Local and remote IDSs have their benefits, but they also have some critical issues. In local methods, the large NDZ of the passive IDS and PQ issues in active IDS make them unsuitable for islanding detection. Furthermore, remote IDSs are costly and infeasible for small distribution networks. Because of the above concerns, researchers and engineers have shifted their focus towards intelligent-classifier-based IDSs. In this research, we summarize and critically analyze the shift in research from traditional IDSs to intelligentclassifier-based IDSs.
Advancements in artificial intelligence have improved life using machine-learning theories in every field of life such as medicine, material sciences, and engineering fields. Nowadays, the applications of state-of-the-art intelligent classifiers in electrical engineering, and primarily in EPSs, are growing rapidly. In EPSs, islanding detection is the most challenging issue that is required to be addressed around the world. The most commonly used intelligent-classifier-based schemes for islanding detection are decision trees (DTs), artificial neural networks (ANNs), support vector machines (SVMs), fuzzy logic (FL), adaptive neuro-fuzzy inference systems (ANFISs), convolution neural networks (CNNs), and deep neural networks (DNNs). Intelligent classifier-based IDSs first derive features from the obtained signal, which are then used as an input to the intelligent classifier in the form of a feature vector, and the classifier makes decisions based on input features as shown in Fig. 4. The use of these methods for islanding detection has increased significantly. Fig.5 depicts the implementation of each technique for islanding detection discussed in this paper. As depicted by Fig. 5, several different forms of intelligent strategies have been used to detect islanding.

A. DECISION-TREE-BASED IDS
DT is the most widely used tool as an intelligent classifier. The recursive partitioning process is applied to each attribute for its value authentication. The DT classifies the given data into pre-defined classes; it can be either binary or multiple classifications. To begin segmentation, a prediction is used to convert the root node into child nodes. The resolution can be done from a child node for additional divisions. The architecture of the DT is presented in Fig. 6, while Table 2 shows the summary of DT-based IDS presented.
The feature vectors are produced from the transient current and voltage data using discrete wavelet transformation (DWT) in [136] for islanding detection. Only the four most important characteristics were validated using classification and regression trees (CART) out of the twelve features. The expanded test system of [136] with voltage-source-converter (VSC)-based DC source was provided in [137], and the performance of the suggested approach was evaluated and compared to that of other passive islanding approaches in terms of noise impact, NDZ, and response time.
DWT and DT were utilized in [138], however, for multiple frequency bands with varied characteristics and features, only one transient signal (voltage) and low DWT levels, from D1-D9, were used. To determine the superiority the accuracy of the proposed scheme was compared to that of many passive IDSs and DT-based methods. In [85], the discrete Fourier transform (DFT) is used to extract the voltage and current signals, as well as their related characteristics, utilizing a data mining technique. The DFT was used to obtain a total of 27 characteristics that were influenced during islanding, and the DT was trained in real-time for islanding and non-islanding detection in the proposed method.
A novel hybrid IDS based on DT and inverter-based positive feedback was presented in [139]. An intelligent islanding relay based on multivariate analysis and data mining techniques is presented in [140]. Then, to manage the protection and thresholds of each DG, DT was utilized for the tripping logic. Before feeding them into the DT classifier, a total of eleven time-dependent characteristics were collected and pre-processed to eliminate noise and inconsistencies. The suggested approach was validated using the offline test results and a hardware-in-loop (HIL) experiment.
A hybrid IDS technique that combines the DT and Sandia frequency shift (SFS) methods for multiple inverters-based DGs was proposed in [141]. Two test systems were used to verify the resilience of the proposed IDS under varied operating circumstances and changing load configurations. A hybrid approach for both inverter-based and synchronous DGs was proposed in [142]. A real-time simulator was utilized to determine the NDZ border, and a DT was used to classify the islanding and non-islanding occurrences. The suggested IIR provides great reliability and security in addition to a large decrease in NDZ.
The selection of features is one of the most technical and critical measures in islanding detection. A novel feature selection technique was presented by [143] based on modified multi-objective differential evolution (MMODEA) and extreme learning machine (ELM). In the off-line mode, a total of 16 features were extracted using MMODEA-ELM, and subsequently, the optimum features were selected by objective function formulation. In the online mode, these optimal features were used to identify islanding and non-islanding scenarios using the DT classifier. A new scheme was presented in [144] to generate the DT logic for the categorization of islanding and non-islanding events based on the active and reactive power imbalance, which directly corresponds to the NDZ. A sequential feature selection method was applied to choose the 12 best features from a total of 30 electrical features. The proposed scheme reduces the NDZ by over 54% than the standard relay function and was verified on the HIL system.

B. ARTIFICIAL-NEURAL-NETWORK-BASED IDS
ANNs are one of the main methods employed in machine learning. These are brain-inspired devices, as the ''neural'' part of its name implies, that are built to mimic how humans learn. Neural networks consist of input and output layers, as well as (in most cases) a hidden unit layer, which converts information into something that the output layer can use, as shown in Fig. 7 while Table 3 summarizes the ANN-based IDS that have been presented.
In [60], an IDS was introduced for inverter-based DGs using signal parameter estimation. The rotational invariance technique (ESPRIT) is used for the extraction of features at the point of common coupling (PCC) and the naïve Bayes classifier is used for the classification. The validity of the proposed scheme was confirmed via three-fold cross-validation. A framework was implemented based on an ANN for islanding detection of the distributed synchronous generators in [145]. The proposed method can detect the islanding situation by calculating the voltage waveform at the distributed generator terminals. In addition, a method for selecting the data was suggested to enhance the training of the ANN. To evaluate the system efficiency as well as non-detection areas, the concept of the time-performance region was implemented.
The optimization approach and minimum-feature-based IDS were presented in [146] with the use of an ANN. Evolutionary programming (EP) and particle swarm optimization (PSO) were introduced for improving the accuracy of ANNs. In this research, the behavior of 16 different features was analyzed, and only three features were selected for the training of the ANN. After a thorough assessment of the online and offline testing results, the ANN-PSO classifier exhibited the highest accuracy as compared with the stand-alone ANN and ANN-EP. In [147], an IDS for the reduction of the NDZ was presented, wherein the classification of the islanding and non-islanding events was conducted using DWT for feature extraction and ANN as a classifier. A modified ELM technique presented as weighted bidirectional ELM (WB-ELM) and Hilbert-Huang transform (HHT) was introduced as a novel IDS in [148]. The proposed IDS was implemented and assessed (dependability and security) on multiple DGs based on test systems and the IEEE 13-bus system. The HHT was used for the feature extraction from the non-stationary voltage signal at the DG end, and the extracted features were then fed to the WB-ELM classifier for islanding and non-islanding detection.
Two feature-extraction methods, the multiplier-based method (MBM) and Andrews's plot-based method (APBM) were introduced in [27] for feature extraction and dimension reduction. The phasor measurement units provide time-synchronized measurements to the phasor data concentrator (PDC). These raw data were then input into the APBM and MBM for pre-possessing before sending them to the ANN classifier. In islanding, optimal feature selection is the key moderator for appropriate and accurate detection, which is not emphasized in the literature. In [149], the author introduced the wrapper method for the selection of sensitive features and coupled the modified multi-objective differential evolution algorithm with a kernel-based ELM classifier. Only 3 out of 45 features were selected and used based on the proposed schemes by evaluating their performance indexes with and without noise. The proposed scheme VOLUME 9, 2021 performed well and accurately with acceptable detection time as compared to the existing intelligent-classifier-based and traditional schemes.
A unique and novel local-data-mining anti-islanding system for synchronous generator (LDMAIS-SG) was presented in [150], based on a powerful data-mining tool, which is known as a code repository (DAMICORE). A total of 10 different features were selected and pre-processed using low-pass Butterworth filters to eliminate the noise. The proposed scheme performed well in differentiating between islanding and other disturbances. In [151], an adaptive ensemble classifier called an ELM, and a phase-space feature extraction technique was introduced. The proposed IDS comprises two steps: in the first step, a unique feature extraction method (phase-space technique) is used to extract features from the three-phase voltage. Then, classification is carried out with ensemble ELM in the second step. Multiple events were tested on two test systems for the performance evaluation of the proposed IDS and compared with the random forest (RF) scheme. The proposed scheme performed well in all the evaluations.
A universal islanding detection scheme is introduced for both inverter-and synchronous DGs in [152], which performed well with high accuracy, zero NDZ for both DG types, and fast detection times. The proposed method comprised three parts: 1) feature extraction (twenty-one features were extracted, which can be influenced by islanding), 2) feature selection using forward feature selection and backward feature selection (four features are selected based on their accuracy in the shortest time), and 3) classification to differentiate islanding and non-islanding occurrences using RF algorithms. A modified DWT, known as slantlet transform (SLT) with a superior lead of two vanishing moments and a better time localization than DWT, has been implemented in [153] for the selection of features. The extracted features have been used to detect islanding and other disturbances while employing a ridgelet probabilistic neural network (RPNN). The combination of SLT and RPNN exhibited better performance with a 100% accuracy compared with the combination of DWT and RPNN and that of DWT and the probabilistic neural network.
An intelligent islanding detection system based on an intrinsic mode function feature-based grey wolf integrated artificial neural network is proposed in [6]. The nodal voltage is pre-processed to extract vital features by the Hilbert transform. Fourier transform (FT) and machine learning algorithm, K-nearest neighbor technique (KNN) is proposed for islanding detection of microgrids in [154]. In this study, nine features from the voltage and current signal were extracted using the discrete Fourier transform. An extreme machine learning and wavelet transform has been used as a classification to discriminate against islanding events from non-islanding events [155]. To exploit the various useful features of the DG bus, the negative sequence voltage and the current signal were used to obtain the 2 basic mathematical morphology operators, erosion dilation difference filter  and opening-closing difference operator. Tunable Q-factor wavelet transforms (TQWT) and ANN-based schemes are proposed in [156], the feature extraction step is conducted using TQWT, and classification of islanding is done by ANN based on conjugate gradient algorithm.

C. SUPPORT-VECTOR-MACHINE-BASED IDS
An SVM is a discriminative classifier explicitly described by a hyperplane separator and first introduced by Vapnik [157]. An SVM is based on the structural risk minimization theory which minimizes the upper limit on the expected risk. The tuning parameters of an SVM are the kernel, regularization, gamma, and margin. Fig. 8 presents the architecture of the SVM and the SVM-based IDS that have been presented are summarized in Table 4.
In [84], in order to overcome the confines of conventional islanding schemes, a scheme is proposed for islanding detection using multiple features and an SVM. Five features were used, which were obtained from five network parameters, namely, frequency, voltage, rotor angle, ROCOV, and ROCOF at PCC. A total of 2760 events were generated for offline SVM training, and the trained SVM is then employed for islanding detection in real-time. Based on the outcome of the above research, a multi-feature-based IDS was presented for the NDZ in the subcritical region of the vector surge relay [158]. From five network variables, the process of feature extraction was completed, and a total of 2022 incidents were produced for the SVM training and testing. Linear kernel, polynomial kernel, and Gaussian radial basis function kernel were used for the performance evaluation of the proposed scheme.
A new IDM for single-phase inverter-based DG was introduced in [90]. In the initial step, autoregressive (AR) signal modeling was used at the PCC for the extraction of the voltage and current signal features. In the next step, the SVM predicts the islanding state from the determined characteristics. The IEEE 13-bus system is used to conduct the study, and the trained SVM is tested under several islanding and non-islanding conditions. This scheme detects the islanding correctly within 50 ms after the event starts. The IDS for hybrid DGs comprising both photovoltaic and wind generation connected to the IEEE 30-bus system was presented in [159]. The negative sequence component of the voltage signal was pre-processed through three different advanced signal-processing schemes, namely, the hyperbolic S-transform, time-time transform (TTT), and mathematical morphology methods for feature extraction. The extracted features were then fed to the SVM for the classification of islanding and other PQ disturbances.
Wavelet transform (WT) is the widely applied signal processing technique for islanding feature extraction but has some shortcomings with non-linear loads and harmonics. To overcome this lag of the WT, in [160], Renyi entropy was employed with WT to identify and categorize seven PQ disturbances. The features extracted using Renyi and WT were then trained in SVM classifiers and tested in real-time scenarios. SVM is used as a dual-functional classifier for the detection of islanding and grid-connected modes and  the classification of various faults during normal grid operation [161]. This scheme was implemented on a MATLAB model based on the parameters of a real-time PV power plant.

D. FUZZY-LOGIC-BASED IDS
Fuzzy logic is a problem-solving approach influenced by human decision-making, which benefits from the human ability to reason with ambiguous or provisional data and was presented by Zadeh [162]. Fuzzy logic is a basic mathematical logic that uses partial truth as a continuous value between 0 and 1 instead of a discrete value, as shown in Fig. 9. The FL-based IDS that have been presented are summarized in Table 5.
Islanding detection for DGs using the concept of a fuzzy membership function (MF) was demonstrated in [163] with an accuracy of 100% islanding detection. In the initial step of the proposed method, 11 features affected by islanding were extracted. The second step consists of two parts: the extracted features were fed to the DT for the initial classification boundaries, and the fuzzy MFs were then used for rule-based classification. In the case of large noise, the WT fails to detect islanding, and thus, a discrete fast S-transform (DFST) with a fuzzy expert system (FES) was proposed in [164] to resolve this problem. Negative voltage and current sequences were used and pre-processed by the DFST for extracting 24 different features, and only four significant features were then fed into the FES for the classification of islanding and non-islanding events. The validation of the proposed IDS was obtained by testing it on two different test systems with different islanding and non-islanding events. In addition, the performance is compared with previously implemented schemes and the proposed method has demonstrated superiority in terms of reliability, detection time, and sensitivity.
A novel hybrid scheme based on the SFS and FL was presented in [41] for the elimination of the NDZ. Initially, the fuzzy load parameter elimination is trained for the selection of the load parameters (R, L, and C) for the appropriate load quality factor and then the SFS for detection. A novel IDS was proposed in [165] using hybrid fuzzy positive feedback (PF) to reduce the interference injection. In the proposed scheme, the PF was continuously varied, while the PF gain was fixed in the traditional methods. The presented scheme VOLUME 9, 2021  shows a 77.3% decrease in detection time and low injection disturbances. The majority of the IDS measures exhibit a change in electrical parameters at a single point called the PCC, but in [76], a multi-connection point model has been implemented for smart grids focused on the probability of islanding (PoI). The PoI at various points was measured using active, passive, and communication-based islanding schemes and sent to the central microgrid control (CMGC). The parameters were obtained from the voltage and current signals using the DWT and then fuzzy neural networks are used for islanding detection.

E. ADAPTIVE-NEURO-FUZZY-INFERENCE-SYSTEM-BASED IDS
ANFIS, a mixture of ANN and FL, is a robust computational system for a non-linear and complicated system with limited training data. It was first introduced in the early 1990s, with two parts called premise and consequence and five layers (fuzzification, rule, normalization, defuzzification, and summation layers) [166]. Fig. 10 presents the complete structure of the ANFIS while the ANFIS-based IDS demonstrated thus far are summarized in Table 6.
In [167], ANFIS-based IDS for inverter-based DGs was presented. During islanding, the proposed system checks the sensitivity of specified parameters at particular DG locations. It determines the pattern vector Xi by analyzing the behavior of the current and voltage data and then feeds it to the ANFIS for islanding detection. In [168], a two-step ANFIS-based IDS was presented. The distribution system was first simulated in PSCAD at the PCC for the extraction of five chosen indices. The retrieved data is then sent into the ANFIS toolbox in MATLAB for the categorization of islanding and non-islanding occurrences in the second phase. The energy analysis of wavelet coefficients and the ANFIS algorithm for the classification of islanding and non-islanding events were used to develop the new IDS presented in [169]. A total of eight distinct electrical signals that were impacted by any islanding or non-islanding disturbance were monitored; these parameters were then put into a wavelet energy calculator to extract features, which were then fed into ANFIS for islanding and non-islanding detection. The suggested IDS had an NDZ of almost zero, no threshold settings, and no PQ issues.
The ROCOF is the most widely used passive islanding detection approach; it was utilized as an input parameter for the ANFIS in [170] to identify islanding and non-islanding scenarios. Various non-islanding situations, including load switching, capacitor switching, and motor starting, were used to validate the suggested system, and the given IDS worked effectively with almost zero NDZ while overcoming threshold setting challenges. In [171] a unique ANFIS algorithm approach for low-voltage inverter-interfaced microgrid islanding detection. The relevant data for the ANFIS classifier was gathered from seven electrical characteristics utilizing relay metering sensors at the PCC. The suggested  hybrid method may be customized to solve several problems. An ANN was utilized to provide flexible learning capabilities, while FL was used to discover nonlinear connections. The proposed IDS did not affect the system PQ and significantly reduced NDZ.

F. DEEP-NEURAL-NETWORK-BASED IDS
DNNs are a popular group of machine-learning algorithms that are implemented in the depths and widths of smaller architectures by stacking layers of neural networks [172]. DNN has a special structure as it has a relatively large and complex hidden layer in the middle of the input and output layers, as shown in Fig. 11. To be called a DNN, this hidden segment should contain at least two layers. Owing to their nature, DNNs can distinguish patterns better than shallow networks [173]. Machine-learning researchers extend the horizons of profound thinking by searching for potential applications of DNNs in other fields such as EPSs. Lately, the DNN has been used in EPSs for islanding detection and has shown excellent results in eliminating the islanding problem. Table 7 summarizes the DNN-based IDS that have been presented.
The concept of deep neural learning was first implemented and demonstrated for the classification of islanding and other grid disturbances in [174]. A novel feature-extraction technique using wavelet decomposition and multi-resolution singular spectrum entropy was introduced. Initially, the PCC voltage signal during islanding and other grid disturbances were decomposed by the wavelet decomposition, and the multi-resolution singular spectrum entropy was then calculated. The extracted features were then used for the training and testing of the DNN. The training process of the deep architecture comprised two steps: 1) initialization of weights utilizing greedy layer-wise unsupervised learning, and 2) fine-tuning of the earlier initialized weights using supervised data. The proposed approach exhibits a better performance in terms of both accuracy (98.3%) and detection time (0.18 s) than other methods of classification.
CNN is also a type of DNN; however, similar to the DNN, this technique is not used much in EPSs for the detection and classification of islanding. CNN was first implemented in EPS applications in 2019, and IDS was proposed based on image classification with CNN [175]. The novelty of the proposed IDS is that it converts the time series data having distinct information about islanding and non-islanding events into images. These scalogram images obtained using continuous WT with Morse wavelet were then fed into the CNN classifier for the classification of islanding and non-islanding events. A total of 205 islanding and non-islanding events were generated for the validation of the proposed IDS, and 60% of the data was used for training purposes, while the remainder of the data was used for testing. The scheme demonstrates an accuracy of 98.78% and a preeminent performance for noisy data. Long short-term memory (LSTM) for the first time used for islanding detection in [176], the scheme proposed a two-step approach for islanding detection. In the first  step, useful features were extracted from voltage and current signals using DFT, and then the extracted six features were fed to LSTM for event detection.
Intelligent-classifier-based IDSs are discussed critically in detail and a comparative investigation of intelligent schemes in islanding detection is summarized. Table 8 presents the pros and cons of intelligent classifiers discussed and used for islanding detection. While Table 9 gives a comparison between intelligent IDSs and conventional IDSs based on various performance indices. Reliability refers to the accuracy of anything being evaluated by a process and DNN has the highest reliability. DT has the lowest reliability among intelligent classifiers while conventional IDSs (remote techniques) have the highest reliability. Complexity provides the means to conceptualize the research challenge as a complex adaptive process, focusing on the patterns of interactions between various system components at different points and times. ANN, FL, active, and passive IDS are the most desirable methods in terms of complexity. While DNN, CNN, ANN, and SVM score comparatively well in terms of accuracy. Finally, DNN, CNN, and ANFIS are influential in terms of detection speed. Intelligent classifiers have zero impact on the PQ of DG-EPS, while active IDSs have the highest impact on PQ. Implementation cost is an important aspect of the DG-EPS system modeling and construction. Intelligentclassifier-based IDSs are economical compared to remote and hybrid techniques.

G. OTHER ISLANDING SCHEMES
Signal processing techniques are used in the modified passive IDSs to increase detection performance, minimize VOLUME 9, 2021  detection time, and reduce NDZ. The researchers improved existing islanding detection schemes and developed new methods by using techniques like the Fourier transform (FT), WT, S-Transform (ST), and TTT. These techniques facilitate the analysis and extraction of key features from a measured signal, allowing for more efficient power system operations. Identification of islanding and non-islanding occurrences is feasible with the knowledge of these retrieved features.

1) FT-BASED IDS
FT is a frequency-domain analysis technique for extracting signal characteristics at certain frequencies. Because the FT is incapable of incorporating time-domain analysis, a short-time Fourier transform (STFT) is used to evolve several frames of the signal in the moving window. Other popular approaches include DFT and fast FT, which convert a discrete-time series of finite duration into a discrete-frequency sequence. FT is used for feature extraction from voltage, current, and frequency signals in intelligence-based IDS, and many IDS have been suggested, which are addressed in section IV [85], [154], and [176].

2) WT-BASED IDS
WT is a useful technique for extracting important characteristics from distorted voltage, current, or frequency signals. The signals are converted into several temporary scales, such as the mother wavelet, which creates small waves called wavelets. The wavelet coefficients of the observed signal are compared to a pre-defined threshold value in WT-based IDSs. The islanding situation will be identified if these coefficients reach a value greater than the pre-defined threshold value. The drawbacks of such techniques are the effects of mother wavelet selection, threshold settings, and various sampling frequencies. Intelligent IDSs made extensive use of various forms of WT for islanding detection, as described in section IV [136], [138], [153], [155], and [160].

3) ST-BASED IDS
The ST idea is a development of the WT concept. It transforms a two-dimensional frequency-domain function into a time-domain function. The ST technique, like other time-domain approaches, is used to extract key characteristics from a recorded signal at PCC, allowing the islanding state to be detected. ST produces the S-matrix and the corresponding time-frequency contours from the recorded voltage or current signals at the DG terminals. The ST approach takes more computing memory to process a signal than other related techniques. Furthermore, such techniques have a long processing time. In [159] and [164] ST is used for feature extraction for intelligent IDSs discussed in section IV.

4) TTT-BASED IDS
By providing a time-time distribution on a specific window, the TTT method analyzes and transforms a one-dimensional time-domain signal into a two-dimensional time-domain signal. Low-frequency components are concentrated at various locations in the TTT technique, but high-frequency components are focused on the localization point with the highest energy concentration. TTT's time-local view usefulness through the scaled window is one of its characteristics, making it a good approach for change detection in signals and systems. In [159] employs TTT-based feature extraction of the negative sequence of the PCC voltage signal to detect islanding in 25 milliseconds. TTT is used in [180], where the TTT pattern of the three-phase disturbances clearly shows distinct signatures. Individual events are discovered to have a distinct pattern that can be utilized to detect islanding. The graphical result analysis in [181] illustrates the TTT capacity of detection and the localization of islanding disturbances in a hybrid DG system over WT, ST, and HST. The energy content and standard deviation of the converted signal are computed to assess performance.

V. FEATURE SELECTION SCHEMES
The most intelligent IDSs usually follow three basic steps, which consist of feature extraction, feature selection, and classification. Generally, more features mean more information, but it is practically difficult to implement because the extraction of multiple features requires more computational power. A significant problem, therefore, emerges in choosing the best and minimum features. Moreover, the collection of suitable features decreases the time needed for classification training and testing of a dataset. From the literature, we have found that not all features are valuable, and some of them are less sensitive to islanding detection and classification.
To increase the efficiency and speed of rating classification, the use of various feature selection schemes omits redundant features. Intelligent-classifier-based learning schemes that use data for any output have been developed for different problems in the last decade. To train any intelligentclassifier-based scheme, a large amount of data is required in terms of several instances and features. This is called feature selection for managing high-dimensional data by identifying responsive and redundant features [182].
Machine-learning algorithms allow computers to create a method based on the input data. Thus, if relevant data is provided, the output of the learning algorithm will be improved. The quest for schemes that improve the quality of input data thus helps to improve the agent's output performance and thus improved data quality can be achieved. For example, by removing the noisy instances and by discriminating between relevant, irrelevant, and redundant data items, an optimal learning algorithm can be developed [183], [184].
The technical reasons for the requirement of feature selection schemes are as follows: a) curse of dimensionality (overfitting), b) Occam's razor (simplicity of model), and c) garbage in, garbage out (poor-quality input will produce poor-quality output). Various schemes for the selection of features have been presented and implemented by researchers, but the main approaches for the selection of features are of three types as summarized in Fig. 12 and the working principles are presented in Fig. 13.
The Wrapper method was presented in 1997 by two researchers in [185]. The Wrapper approaches work with a sub-set of features that test each subset based on the performance quality of an algorithm, using a search technique to observe the range of potential feature sub-set [186]. In any machine-learning algorithm, filter methods independently select features from a data set depending on the feature characteristics as the characteristics are extracted from the data before learning [187]. Embedded methods complete the selection process inside the machine's algorithm during the model training, which is why they are called embedded methods [188].
The feature selection is the most significant and critical step, but several researchers have not emphasized this part in the islanding detection. The majority of the proposed IDSs select features explicitly based on the analysis of the literature without mentioning the selection criteria and selection reasons. For their proposed IDS, only a few researchers used feature selection schemes. The feature selection wrapper method has been used in [149] to select the three best features out of 45. In [136], CART was used to choose features. The sequential feature-selection process was employed to choose 12 of the best features out of a total of 30 electrical features in [144]. The forward sequential feature selection  and the backward sequential feature selection were adopted for feature selection in [153]. As discussed above, considering the importance and need for optimal selection of features, IDSs should use any feature selection schemes based on their data and the requirements of the proposed scheme.

VI. MOST DOMINATING FEATURES FOR ISLANDING DETECTION
In islanding and non-islanding situations, the electrical features of EPS show various fluctuations as compared to those in normal operation. Some of these parameters are more sensitive to disturbances from islanding and non-islanding whereas, some are less sensitive, while some of them have no impact [189], [190]. Based on the sensitivity of these parameters under irregular circumstances, researchers have selected specific parameters for their proposed IDS. The number of parameters for each IDS depends on their method and the feature selection technique used. As mentioned in section 5, several researchers have not used any feature selection methodology and have selected parameters based on previous and evolving studies. In [136], 12 features were extracted, whereas in [85], 27 features were extracted for islanding detection. Similarly, a total of 16 features were extracted using MMODEA-ELM and by objective function formulation, the most optimum features were selected in [143]. In [149], initially, 45 features affected by islanding conditions were extracted, while in [71], 10 features were selected; a total of 21 features were selected in [152] during the feature-extraction step, and 11 features were extracted in [80]. A wide variety of features for islanding detection have been derived ranging from three to 45 features in the above IDSs. Some specific and significant features are impaired and very prone to islanding and non-islanding disturbances, which were utilized by the majority of researchers in their proposed schemes. Such features provide a lot of information concerning behavioral changes and are very useful for machine-learning classifiers to distinguish between these disturbances. ROCOF, ROCOV, total harmonic distortion (THD), power factor, ROCOP, and ROCOQ are widely used owing to their traits that are useful for classifying islanding and non-islanding situations. Table 10 provides a full overview of the most dominant features and their uses in the various IDSs.

VII. PERFORMANCE INDICATORS FOR IDSs
The timely and precise operation of the appropriate technique is critical to the IDSs' performance. The NDZ, parallel RLC load, and quality factor are the three primary performance indicators that describe an IDS's operating capabilities. A successful operation under these severe conditions confirms the use of IDSs. If a technique effectively identifies the islanding situation in such circumstances, the approach's superiority is highlighted, and international requirements are satisfied.

A. NDZ
NDZ is an area that is not easily identified by traditional protective relays in islanding detection. The NDZ is thought to be a summative assessment for IDSs. An NDZ is typically assessed based on a range of active and reactive power mismatches in which the voltage and frequency relays are unable to identify the islanding condition in a reasonable timeframe. The NDZ boundary limits may be established using (1) and (2), and the region of critical and non-critical operating conditions can be identified [103].
where V max , V min , f max , and f min are the maximum and minimum voltage/frequency threshold limits of the relay. P and Q represent the power mismatches prior to the main grid disconnection while Q f is the load quality factor.

B. PARALLEL RLC LOADS
Most loads in power networks are inductive, but inverterbased DG units yield maximum kilowatt-hours by operating at unity power factor. When combined with a parallel RLC load, this is considered the worst-case situation for detecting islanding. Similarly, when the connected load and DG power are perfectly matched, islanding identification becomes more difficult [23]. The relationship between the PCC voltage and frequency in a parallel RLC circuit is as follows: V PCC is reliant on the active power of the islanded system, as shown in (3) and (4).

C. QUALITY FACTOR
The load quality factor is a critical element for determining the IDS's dependability and robustness [103]. Because the VOLUME 9, 2021 size of the NDZ and detection accuracy is affected by the load quality factor, the value of the load quality factor has a major impact on the performance analysis of the IDSs. The quality factor can be expressed mathematically as (5).
where R, L, and C are the effective load resistance, inductance, and capacitance, respectively. Table 11 shows a comparison of several existing intelligent classifier-based IDSs based on many indices like the used test system, detection time (s), achieved accuracy, and the NDZ. Table 9 compares intelligent classifiers with traditional IDS based on different criteria such as reliability, complexity accuracy, detection speed, impact on PQ, and implementation cost. In the recent research on islanding detection conducted by our team, various intelligent classifiers were compared based on accuracy, precision, recall, and F_1 score in [3]. Ada-boost performs very accurately with the highest accuracy, precision, recall, and F_1 score, while DT performance is worst among all models. Fig. 14 illustrates a comparison of the ensemble learning models (Ada-boost and RF) and canonical methods (MLP, DT, and SVM) with the same data and test system. The change in research trends from traditional to intelligent-classifier-based IDS over the past decade has been outlined and objectively evaluated in this paper. This work was focused on intelligent-classifier-based IDSs such as DT, ANN, SVM, FL, ANFIS, CNN, and DNN. Different aspects of feature selection were also studied in this research, as this is the most critical step in islanding detection algorithms. The review also outlined the most widely used and influential features for islanding detection.

VIII. COMPARISON, DISCUSSIONS, AND FUTURE RECOMMENDATION
This study concludes that machine-learning-based schemes show a high degree of robustness in all performance indices that are important for any IDS, such as negligible NDZ, high PQ, multiple DGs, and fast detection time, and are capable of managing complex and large EPSs. Following are the recommendations for future studies, which are based on the evaluation and review of all the performance indices for islanding detection.
• Intelligent-classifier-based methods are strongly recommended based on performance indices that are significant for islanding system design.
• Based on this study, feature selection should be regarded as an essential step in any proposed islanding scheme.
• Several features offering a high accuracy is a fallacy in machine learning, as a larger number of features burdens the system and reduces the speed and accuracy of the algorithm. The selection of features should, therefore, be based on precision and prevalence and be as few as possible. • DNN-and CNN-based approaches are desirable for large and complex systems, but conventional schemes offer the best efficiency and are still very convenient for implementation in small and basic systems.
• Signal processing techniques schemes such as DWT, DFT, TTT, and HHT can be used to reduce noise, maximize device performance, and minimize dimensionality during the extraction of features.
• In the future, advanced signal processing technologies combined with a machine-learning algorithm might be used to develop an accurate island detection method.
• In future research, hybrid strategies based on intelligent classifiers and conventional techniques are suggested.
• Several recent intelligent classifiers, such as LSTM, RNN, Encoder-Decoder, and so on, are not used in power system research, although they may be in future studies.

IX. CONCLUSION
A detailed and in-depth analysis of intelligent-classifier-based IDSs has been presented in this review. Systematic and detailed research is carried out based on intelligent/machine learning theories for islanding detection. This research outlines the feature-selection schemes used in the literature and those that could be used in the future for islanding detection. In this article, based on a literature review, the most important electrical parameters for islanding detection are also highlighted. The IDS were categorized into three main classes remote, local, and intelligent classifier-based scheme. Classic IDS approaches, such as active, passive, and remote approaches, have their own merits and demerits, and because of their shortcomings, researchers have recently shifted towards intelligent-classifiers-based schemes. The intelligent-classifier-based IDSs have major advantages in terms of performance indices, such as NDZ, detection time, precision, PQ, noise, and accuracy as compared to conventional schemes. It can be concluded that the implementation of intelligent classifier-based IDS can play a major role in the efficient and viable detection of DG islanding. Implementation of these techniques in islanding detection will also increase the stability of the power system and power supply efficiency. Therefore, artificial intelligence-based approaches are also favored and can be used in real-time applications to efficiently execute DG islanding operations.