Machine Learning Based Protection Scheme for Low Voltage AC Microgrids

Abstract

The microgrid (MG) is a popular concept to handle the high penetration of distributed energy resources, such as renewable and energy storage systems, into electric grids. However, the integration of inverter-interfaced distributed generation units (IIDGs) imposes control and protection challenges. Fault identification, classification and isolation are major concerns with IIDGs-based active MGs where IIDGs reveal arbitrary impedance and thus different fault characteristics. Moreover, bidirectional complex power flow creates extra difficulties for fault analysis. This makes the conventional methods inefficient, and a new paradigm in protection schemes is needed for IIDGs-dominated MGs. In this paper, a machine-learning (ML)-based protection technique is developed for IIDG-based AC MGs by extracting unique and novel features for detecting and classifying symmetrical and unsymmetrical faults. Different signals, namely, 400 samples, for wide variations in operating conditions of an MG are obtained through electromagnetic transient simulations in DIgSILENT PowerFactory. After retrieving and pre-processing the signals, 10 different feature extraction techniques, including new peaks metric and max factor, are applied to obtain 100 features. They are ranked using the Kruskal–Wallis H-Test to identify the best performing features, apart from estimating predictor importance for ensemble ML classification. The top 18 features are used as input to train 35 classification learners. Random Forest (RF) outperformed all other ML classifiers for fault detection and fault type classification with faulted phase identification. Compared to previous methods, the results show better performance of the proposed method.

1. Introduction

New DG (Distributed Generation) technologies are being developed, and more DGs are entering the distribution system as the conventional power grid approaches its maximum capacity and concerns about the glasshouse gas emissions by traditional power plants grow. This has allowed microgrids to emerge as an essential element of the modern distribution system. While it offers many advantages, there are a number of concerns to be resolved, among which fault protection is a major challenge. Protection devices in existing DN (Distribution network) consist of non-directional overcurrent (OC) protective relays, reclosers, fuses and sectionalisers [1] and are mostly radial. The protection schemes were initially designed for the unidirectional flow of power, but with the increase in bidirectional power flow, coordination between fault protection devices can be compromised [2], especially when the microgrid is in an autonomous (AUTO) mode of operation. Additionally, achieving correct selectivity and sensitivity poses a great challenge in the practical application of microgrids [3]. As a result, conventional protection techniques will not offer satisfactory protection in the future. Moreover, future intelligent grids will require precise single- and double-pole tripping to boost overall resilience and economic benefits. This requires correct fault type classification (FTC) and faulted phase (FP) identification to avoid tripping of healthy phases in fault events [4].

A common tool for analysing non-stationary, noisy, aperiodic, transient and intermittent signals is wavelet transform (WT). WT can examine the signal in time-frequency domains which give it superiority over Fourier and short-time Fourier transform [5]. WT unhides features within the original signal for detailed analysis. Additionally, applying machine learning (ML) to detect and classify faults offers a promising solution.

To detect microgrid islanding and fault disturbances, refs. [6,7] proposed methods based on WT. Both studies used Daubechies (dB) as the mother wavelet. The former considered only the negative-sequence component, dB5 and line-to-line (LL) faults, while the latter considered the positive- and zero-sequence component, dB10 and double-line-to-ground (LLG) faults. After extracting the signal, discrete wavelet transform (DWT) is applied to decompose it. By detecting variations in different parameters and comparing them to the pre-fault or threshold values, it is claimed that islanded mode and power quality issues can be identified. Furthermore, a DT-based protection scheme is proposed in [8] for FD and FTC. For FE, Discrete Fourier transform is used to extract features from voltage and current phasors.

On the other hand, ML-based microgrid fault protection methods are proposed in [9,10]. Wavelet Packet Transform and Multiresolution decomposition of WT are respectively used to pre-process the current and voltage signals to extract features based on negative sequence components and total harmonics distortion (THD) to train and evaluate the ML classifier. Likewise, a microgrid fault identification and classification method is proposed in [11], which processes line current using Haar wavelet to generate various coefficients. Detail coefficient d3 was selected to calculate different features to train the decision tree (DT) to detect and classify faults. Additionally, a Brownboost (BB) ensemble approach is used for FD and FTC in [12]. Hilbert–Huang transform is adopted for feature extraction from signals. Signals used are the current differential between the two ends of the line.

Similarly, a combination of Maximal Overlap Discrete Wavelet Transform and Extreme Gradient Boost is applied for FD and FTC in [13]. FejerKorovkin is used as the mother wavelet to extract features from the three-phase current and its zero-sequence current component. Further, ref. [14] proposed a technique for identifying various fault types in a microgrid by applying DWT to calculate the wavelet coefficients that are fed to different ML classifiers. Conversely, a high impedance fault detection approach using empirical wavelet transform is proposed in [15]. Different time-frequency components are first acquired using the WT to decompose the differential coefficient of wavelet energy. Feature components with the highest permutation entropy are then selected, which are used to identify high impedance faults.

In ref. [16], the authors used actual fault signals instead of normalising and extracting features. Although this gives high efficiency, a significant drawback of using the raw signals instead of transforming them into numerical features is that the ML model becomes prone to overfitting. Similarly, using a large number of features increases processing time. Hence, there is a need to extract unique features that result in high accuracy.

From the literature, it can be noticed that WT is commonly used for feature extraction. The main shortcoming of protection schemes using WT for feature extraction for ML is selecting an optimal mother wavelet basis function. Applying various mother wavelets to the signal may produce a variety of outcomes [17,18,19] that can cause protection system misoperation. Additionally, the type of chosen mother wavelet significantly impacts DWT, producing quite varied outcomes. Moreover, fault inception angle and sampling rate affect DWT response. As a result, most DWT-based protection techniques or derived features are efficient for specific parameters and cannot be generalised without using a different mother wavelet [20]. Figure 1 demonstrates the WT shortcoming by comparing the root-mean-square error (RMSE) for the reconstructed current signals using all the approximate and detail coefficients of a few wavelets for a low-resistance LG fault on phase A.

The RMSE seems to be negligible when all the approximate and detail coefficients are used to reconstruct the signal, but it is not feasible to use all the coefficients. When only one of the detail level coefficients corresponding to maximum relative energy is used to reconstruct the signal, there is a significant change in RMSE for different wavelets. Figure 2 shows a comparison of Level 9 detail coefficient of a current signal $Ip{h}_A$ and reconstructed signal for a low resistance LG fault on phase A using Symlets 5 ($Sym5$) as MW.

When only Level 9 detail coefficient is used to reconstruct the signal, comparison of RMSE for different wavelets is shown in Figure 3.

Unlike most researchers who have used WT to extract features, this research proposes new feature extraction methods and examines dimensionality reduction techniques, statistical tools, and impulsive and signal processing metrics to extract unique features for ML classifiers, to overcome the challenges associated with the selection of the optimal mother wavelet.

The most common type of asymmetrical faults is a line-to-ground fault (LG), accounting for 65 to 70% of all faults in an electric power system, followed by LLG faults that occur around 15 to 20% of the time, while LL faults make up 5 to 10% of all faults [21]. Symmetrical faults (LLL) are rare but the most severe of all electrical faults [22]. All these different types of faults are detected and classified in this paper, which is an extension of the research work published in [23].

Many improvements have been made. Previously, only LG faults were classified for faulted phase and variations in fault resistance. Nine signals, line-to-line voltage, phase voltage and short-circuit currents for phases A, B and C were used. Although these signals were enough to classify LG faults, they were not enough for FD. Since there was only LG fault classification, no data was collected to detect the fault and no-fault conditions. A small dataset was used with manual parameter tuning, which can not match the automated optimisation, using every possible variation in order to improve performance. Additionally, features were selected using an iterative process instead of using FS techniques. The main contributions of this paper are:

• A large amount of data is collected for varying fault conditions and no-fault cases. A new 400 × 500 × 10 microgrid fault dataset is built for 400 scenarios, each with 500 samples for 10 signals.
• Two new feature extraction (FE) techniques, Peaks Metric and Max Factor, are formulated and applied.
• Eight other FE methods, most of which have not been used for microgrid fault detection and classification, are investigated for suitability.
• A new 400 × 10 × 10 dataset with unique features for fault detection and classification in AC microgrids is built for 400 cases, 10 FE techniques and 10 signals.
• Various feature ranking techniques have been used to reduce the number of predictors. 35 ML algorithms with optimal hyperparameters have been trained to find the models with the highest possible accuracy for the fewest possible predictors.
• Validation of the trained models is carried out by using unseen data for making predictions.

After the introduction, Section 2 describes the test microgrid model used to record signals through electromagnetic transient (EMT) simulations for wide variations in operating conditions. Different techniques, including new factors proposed for extracting features, are presented in Section 3. Section 4 presents the methods used for feature selection. The proposed method for detecting and classifying faults is described in Section 5. The results and analysis are summarised in Section 6. Section 7 presents the conclusion and future work.

2. Test Microgrid and Simulations

The low voltage AC test microgrid simulated in DIgSILENT PowerFactory shown in Figure 4 is a part of a radial distribution system operating at 415 V, 50 Hz. It is connected to the main grid through an 11 kV/415 V transformer. There are three DG sources, two photovoltaic (PV) systems, which are inverter-interfaced distributed generators (IIDGs) and a synchronous generator-based microturbine to maintain microgrid stability in AUTO mode by providing sufficient damping component and rotational inertia. A commercial load is connected to Bus 1, and domestic loads are connected to Bus 2 and 3. Bus 2 is the point of common coupling (PCC). The circuit breaker after the transformer is used for switching between the two microgrid operational modes.

Data is recorded through EMT simulations for 400 cases. For every case, simulations are carried out for 0.05 s with a step size of 0.0001 s to obtain 500 observations for each of the 10 signals. The signals include three phase voltage $\left(Vp{h}_{ABC}\right)$ in kV, three phase current $\left(Ip{h}_{ABC}\right)$ and short-circuit current $\left(Is{h}_{ABC}\right)$ in kA and frequency $\left(Freq\right)$ in Hz.

Simulations are carried out for ten faults to collect data for fault detection (FD) and FTC with FP. Apart from variations in fault resistance, reactance, inception angles, number of cycles and locations, all faults are simulated for grid-connected (GC) and AUTO mode to identify the variations in fault current level and other signals. Three cases are used for the LG faults: bolted fault with 0 $\Omega$ resistance, low resistance ground fault with 5 $\Omega$ resistance, and high resistance fault with a value of 400 $\Omega$. For all other faults, the first case (C1) has no fault resistance or reactance. The second case (C2) has a resistance of 0.1 $\Omega$ and a 0.001 $\Omega$ reactance. The third and last case (C3) has a resistance of 0.1 $\Omega$ and a comparatively greater reactance of 1 $\Omega$.

Waveforms of 10 signals: $Vp{h}_{ABC}$, $Ip{h}_{ABC}$, $Is{h}_{ABC}$ and $Freq$ for a fault and three NF cases are shown in Figure 5. Waveforms for LL-AB fault for case C3 in GC mode is shown in Figure 5a. The NF case of loads switching on and off in AUTO mode is shown in Figure 5b; the NF case of load switching off in GC mode is shown in Figure 5c; and the NF case of MG switching from AUTO to GC mode is shown in Figure 5d.

For FD, equal sets for different faults are categorised as Fault, while various cases of normal operation, load switching and grig switching are classified as No Fault (NF). On the other hand, FTC data has been organised to classify the fault type and FP. Data for the NF conditions include simulations of connecting and disconnecting 5 kW, 50 kW and 200 kW load in both modes, switching from GC to AUTO and vice versa with and without load switching. Additionally, simulations without any fault, load or grid switching are also included to differentiate between a fault and NF conditions. Data collected for symmetrical faults include LLL, LLLN and LLLNG faults. Due to close similarity in collected signals, all cases of symmetrical faults are categorised as LLL faults.

3. Feature Extraction

Feature extraction (FE) involves transforming raw data into numerical features while retaining the information in the original data. This helps in preventing overfitting and gives better results instead of applying ML to the raw data. Two novel FE techniques, Peaks Metric and Max Factor are proposed and applied in this research. Additionally, the suitability of using Standard Deviation, First and Second Principal Components [23], Total Harmonic Distortion [10], Kurtosis, Crest Factor, Shape Factor and Skewness [24,25] to extract useful features is investigated. A total of 100 unique features are obtained. Kurtosis, Crest Factor, Shape Factor and Skewness are commonly used FE techniques for bearing fault diagnosis but have not been applied before to detect and classify faults in an AC microgrid to the best of the author’s knowledge.

Moreover, using Principal Component Analysis to detect and classify AC microgrid faults is also not common and was previously proposed by the authors of this paper.

3.1. Standard Deviation

The standard deviation $\left(STD\right)$, for a variable vector x composed of N scalar observations is defined as

$$STD\left(x\right)=\sqrt{\frac{{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^2}{(N-1)}}$$

(1)

where $\overline{x}$ is the mean of x:

$$\overline{x}=\frac{1}{N}\sum _{i=1}^N{x}_i$$

(2)

The variation in $STD$ of $Vp{h}_B$ for random cases of NF and LL-AB fault is shown in Figure 6. There is a notable difference between fault and NF features, which is desired for training ML classifiers. If the features also called predictors are very closely distributed, the probability of misclassification increases.

3.2. Peaks Metric

Peaks Metric $\left(PM\right)$ is a novel metric proposed in this research and is defined as the ratio of the mean of the peak values ${\overline{x}}_{peaks}$ in the signal to the mean $\overline{x}$ of the signal.

$$PM\left(x\right)=\frac{{\overline{x}}_{peaks}}{\overline{x}}$$

(3)

where

$${\overline{x}}_{peaks}=\frac{1}{N}\sum _{i=1}^N{x_{peaks}}_i$$

(4)

For C3, LL-AB fault in GC mode, the deviation in $freq$ is shown in Figure 7.

For the above observation, there are four peaks with values 50.8540, 50.5164, 50.7120 and 50.4684. The ${\overline{x}}_{peaks}$ is 50.6377, while $\overline{x}$ is 50.0130. For the above case, the value of $PM$ is 1.0125. The difference in $PM$ of $Freq$ for LL-AB fault and NF conditions is shown in Figure 8.

The proposed $PM$ considers all the peaks and takes their mean to represent the signal better, instead of just using the max value. When a fault occurs, the peaks of the waveform change before there is any significant change in the signal’s energy. Therefore the $PM$ can warn of faults when they first initiate.

3.3. Max Factor

Max Factor $\left(MF\right)$ is the second novel metric proposed in this research and is the ratio of maximum value $x_{max}$ to the absolute value of mean $\left|\overline{x}\right|$ of the signal.

$$MF\left(x\right)=\frac{x_{max}}{\left|\overline{x}\right|}$$

(5)

For a bolted LG fault on phase B in AUTO mode, three phase current signals $Ip{h}_{ABC}$ are shown in Figure 9, and the signal for $Ip{h}_B$ is shown separately in Figure 10 to demonstrate the application of proposed metric $MF$.

In Figure 10, the max value of the $Ip{h}_B$ is 1.2775, while $\left|\overline{x}\right|$ is 0.0108. For the above case, the value of $MF$ is 117.7036. For the no fault case, the max value of the $Ip{h}_B$ is 0.343, while $\left|\overline{x}\right|$ is 0.026, resulting in $MF$ of 13.273. The difference in $MF$ of $Ip{h}_B$ for NF and LG-B fault cases is shown in Figure 11.

3.4. Principal Component Analysis

Principal component analysis $\left(PCA\right)$ is mainly applied to reduce dimensionality in order to decrease the processing time and avoid overfitting the model [23]. The first step in $PCA$ is to calculate the covariance matrix. The covariance matrix $\left(C_M\right)$ of any two variables x and y, is the matrix of pairwise covariance $\left(cov\right)$ calculations between each variable.

$$C_M=\left[\begin{array}{cc}cov(x,x)& cov(x,y)\\ cov(y,x)& cov(y,y)\end{array}\right]$$

(6)

where

$$cov(x,y)=\frac{{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^{*}\left(y_i-\overline{y}\right)}{N-1}$$

(7)

$\overline{x}$ and $\overline{y}$ are the mean values of x and y, respectively and ∗ denote the complex conjugate. Eigenvalues are used to calculate the eigenvectors for the covariance matrix, which are then used to extract patterns. The first eigenvector represents the eigenvalue that has the highest variance. For the eigenvalue, which has the next highest variance, the second eigenvector corresponds to it, and so on. The matrix that results is as follows…

$$E=\left[\begin{array}{ccccc}e{\overrightarrow{v}}_1\hfill & e{\overrightarrow{v}}_2\hfill & .\hfill & .\hfill & e{\overrightarrow{v}}_p\hfill \end{array}\right]$$

(8)

Only the first and second eigenvectors are selected to obtain the first $\left(p{c}_1\right)$ and second principal components $\left(p{c}_2\right)$.

$$E^{\prime }=\left[\begin{array}{cc}e{\overrightarrow{v}}_1\hfill & e{\overrightarrow{v}}_2\hfill \end{array}\right]$$

(9)

The projection of the vectors onto the new base that is consistent with $p{c}_1$ and $p{c}_2$ is used to represent the new features.

$$p{c}_{1,2}=E^{\prime }·{\left[x_i-\overline{x}\right]}^T$$

(10)

where $x_i$ and $\overline{x}$ respectively represent the variable and the mean vector of the original data, whereas $p{c}_{1,2}$ represents new features. The difference in $p{c}_1$ for $Vp{h}_A$ for NF and LLG-AB fault scenarios is shown in Figure 12.

3.5. Kurtosis

The Kurtosis $\left(Kurt\right)$ of a signal x is defined in (12) [25].

$$Kurt\left(x\right)=\frac{\frac{1}{N}{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^4}{{\left[\frac{1}{N}{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^2\right]}^2}$$

(11)

The $Kurt$ of the normal distribution is 3. A fault in the system will change the value, greater than or less than 3. The difference in $Kurt$ for $Ip{h}_B$ for NF and LLL fault cases is shown in Figure 13.

3.6. Crest Factor

Crest Factor $\left(CRES\right)$ is the ratio of the maximum absolute value to the RMS [25].

$$CRES\left(x\right)=\frac{x_m}{x_{rms}}$$

(12)

where $x_m$ is the maximum absolute value of the signal:

$$x_m=\underset{i}{max}\left|x_i\right|$$

(13)

and $x_{rms}$ is:

$$x_{rms}=\sqrt{\frac{1}{N}\sum _{i=1}^N{\left|x_i\right|}^2}$$

(14)

The $CRES$ of a sinusoidal current waveform for purely resistive load is 1.414. Figure 14 show the difference in $CRES$ of $Ip{h}_A$ for NF and LLL fault condition.

3.7. Shape Factor

Shape Factor $\left(SF\right)$ is the ratio of RMS to the mean of the absolute value [24]. The $SF$ is independent of the signal dimensions, but it relies on the signal shape.

$$SF\left(x\right)=\frac{x_{rms}}{\frac{1}{N}{\sum }_{i=1}^N\left|x_i\right|}$$

(15)

Figure 15 show the difference in $SF$ of $Vp{h}_C$ for NF and LL-CA fault cases.

3.8. Total Harmonics Distortion

The $THD$ is the amount of distortion in the signal compared to the undistorted signal. It is defined as the ratio of the square root of the summation of all harmonics squared (from second harmonic) over the fundamental component [10]. $THD$ is an essential measure in power systems. A lower value gives lower peak currents, higher power factor and system efficiency.

$$THD\left(x\right)=\frac{\sqrt{{\sum }_{n=2}^{\infty }{x}_n^2}}{x_1}$$

(16)

where $x_n$ is the n-th harmonic of x and $x_1$ is the fundamental component. $THD$ difference of $Vp{h}_C$ for NF and LG(C) fault cases is shown in Figure 16.

3.9. Skewness

The skewness $\left(Skew\right)$ shows the irregularity of signal distribution [25]. Distribution symmetry can be impacted by faults resulting in an increased level of skewness.

$$Skew\left(x\right)=\frac{\frac{1}{N}{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^3}{{\left[\frac{1}{N}{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^2\right]}^{3/2}}$$

(17)

The difference in $Skew$ of phase C for NF and LLG(CA) fault conditions is shown in Figure 17.

4. Feature Selection

Feature selection (FS) is the process of reducing features or predictors to provide the best predictive power in modelling a set of data, as not all features are useful. The goal is to find the fewest possible features with the highest possible accuracy. Finding the best features essentially remains an iterative process and requires deep domain knowledge. Feature selection aids in improving the speed and accuracy of prediction as it:

• Prevents overfitting: modelling with many features can make the model more susceptible to specific observations in training data.
• Reduces model size: fewer features increase computational performance and require less memory for embedded deployment.

Features are ranked using the Kruskal–Wallis H-Test (KW) [26] for FD and FTC with FP. The top 18 features for FD are shown in Figure 18.

Additionally, estimates of predictor importance for the classification ensemble methods are also computed by summing the estimates over weak learners in the ensemble for each input predictor. A high value indicates that this predictor is important. Predictor importance for FD using Bagged Trees (BT) ensemble, where bagging is short for bootstrap aggregation [27], is shown in Figure 19.

5. Methodology

By applying domain knowledge and FS methods, the top 18 features are selected for FD, and the top 18 features are chosen for FTC with FP. Numerous ML algorithms are trained and tested, as will be discussed in Section VI. Random Forest (RF) [28,29] outperformed all other ML classifiers for FD and FTC with FP.

RF is an ensemble method similar to BT but differs in the growing phase. A subset of features is randomly selected for each decision split in RF as shown in Figure 20.

In BT, at each node, all the features are candidates for splitting, which can result in the same features giving the highest accuracy, being used at different nodes and for numerous DTs, as shown in Figure 21. This causes overfitting in the model. On the contrary, in RF, each tree is grown using a separate random subset of data; therefore, every decision tree that makes up a RF is unique. RF has superior accuracy compared to BT, as it minimises overfitting, while BT is better than a single decision tree.

For FD, the RF model with optimal hyperparameters grows 488 individual trees. On the other hand, for FTC with FP, the optimised RF model has 271 trees. The trained RF models are deployed to detect and classify faults. Figure 22 presents a schematic diagram of the proposed protection scheme.

The process starts with measuring local signals, followed by FE, fed to the trained RF models. A trip signal is issued when a fault is detected by converting the data type of label to the real-world numerical value. Furthermore, FTC with FP is displayed.

6. Results and Analysis

The top 18 features are used as input to train 35 classification learners for FD and FTC with FP. These include Classification Ensembles, Naive Bayes, Neural Networks, Discriminant Analysis, Support Vector Machine (SVM), Classification Trees and k-nearest neighbours (KNN) [30]. Ten-fold cross-validation (CV) is applied to the training dataset to protect against overfitting. Hyperparameter tuning of all models is performed to improve accuracy. Predictions are made using unseen data. The top 5 models with CV and test accuracy for FD are shown in

Table 1. Test accuracy of ML models with 18 features for FD.

Model	CV Accuracy	Test Accuracy
Optimized RF	100%	99.8%
Optimized GB	100%	99.1%
Gaussian Naive Bayes	100%	98.2%
Optimized KNN (Cosine)	97.9%	96.4%
Neural Network (Bilayered)	95.9%	93.5%

.

RF displayed the highest test accuracy. The accuracy of ML models with further reduction in the number of features for FD and FTC with FP is also investigated. Model accuracy dropped sharply for less than 18 predictors. The best combination of 18 predictors for FD includes 3 predictors obtained by $STD$ of $Vp{h}_{ABC}$, and 9 using $p{c}_1$ of $Is{h}_{ABC}$, $Ip{h}_{ABC}$ and $Vp{h}_{ABC}$. The remaining 6 include $CRES$ of $Is{h}_{ABC}$, $MF$, $PM$ and $CRES$ of $Freq$.

The Simulink Classification Ensemble Predict block with trained RF model is used to validate the model predictions with trip signal issued when fault is detected. For a set of 20 new observations, with alternating 5 NFs followed by fault cases, the model only misclassified once, where it predicted a fault as an NF. The alternating trip signals for the 20 observations are shown in Figure 23.

Likewise, for FTC with FP, the top 18 predictors are used to train 35 classification learners. The top 5 models with 10-fold cross-validation and test accuracy are shown in

Table 2. Test accuracy of ML models with 18 features for FTC with FP.

Model	CV Accuracy	Test Accuracy
Optimized RF	100%	99.4%
Decision Tree (Fine)	99.3%	98.1%
Optimized SVM (Gaussian)	97.1%	94.9%
Neural Network (Wide)	95.2%	91.4%
Linear Discriminant	93.8%	89.3%

.

The best combination of 18 predictors for FTC with FP includes 6 obtained by $STD$ of $Vp{h}_{ABC}$ and $Ip{h}_{ABC}$, and 9 using $p{c}_1$ of $Is{h}_{ABC}$, $Ip{h}_{ABC}$ and $Vp{h}_{ABC}$. The remaining 3 include $CRES$ of $Is{h}_{ABC}$.

The test accuracy comparison of the proposed method with other FD and FTC methods is shown in Figure 24 and Figure 25.

The results show excellent performance by the proposed protection scheme compared to previous methods. Apart from achieving very high accuracy for FD and FTC with FP, very high protection sensitivity is also attained for both modes of microgrid operation for various fault types and cases.

7. Conclusions and Future Work

The FD methods based on WT feature extraction can be affected by the type of selected mother wavelet, which may cause the protection system misoperation. Subsequently, most WT based protection techniques are efficient for specific parameters and cannot be generalised without using a different mother wavelet. To overcome this shortcoming of WT, a new protection scheme for AC microgrids is developed in the proposed research. Novel FE techniques, Peaks Metric and Max Factor are applied, apart from exploring other FE methods to examine the suitability for detecting and classifying faults. After the signals are pre-processed, the features are extracted and then the best performing features are selected using FS techniques. Various ML classifiers are trained and tested. For FD and FTC, with 18 predictors, RF outperformed all other ML classifiers for FD and FTC with FP. Simulink is used to validate the model predictions with a trip signal issued when a fault is detected. Accurate FD, FTC with FP identification, and high protection sensitivity for wide variations in operating conditions make the protection scheme superior to earlier methods.

Future work will integrate the proposed FD and FTC method into a multi-agent based protection scheme for meshed microgrid and testing it on a real-time digital simulator to evaluate the performance.