DATA SCIENCE AND APPLICATIONS, VOL. 4, NO.

2, 2021

Machine Learning Model Applications for Fault

Detection and Classification in Distributed
Power Networks
Jose Eduardo Urrea Cabus*, 1, İsmail Hakkı Altaş1
1
Department of Electrical and Electronics Engineering, Karadeniz Technical University,
61080, Trabzon, Turkey

1
Abstract—This paper compares various unsupervised finally cleared in the shortest amount of time to maintain
feature extraction techniques and supervised machine learning reliance and service continuity [4].
models for fault detection and classification in a power Methods for fault detection, classification, and location in
distributed generation system. The modified IEEE 34 bus test
power distribution systems in the presence of DGs and the
feeder was implemented for the study case simulated through
PowerFactory DigSILENT software. Data analysis results from intelligent agent’s incorporation have been studied and
three-phase voltages and currents collected were performed in published over the years [6, 10]. For instance, in [5] and [8],
Python. Simulation results confirm that by applying data mining has been incorporated into the protection
dimensionality reduction techniques such as feature extraction functions to extend performance limits by producing
and wavelet family selection adequately, high identification and enhanced conditions of states within the present protection
classification accuracy can be obtained, excluding the less
technology limitations over the distribution network. On the
essential characteristics and preventing the machine learning
models from overfitting or underfitting the datasets. other hand, in [11], they incorporate statistical machine
learning methods for preventive maintenance, using
Index Terms— Data mining, fault diagnosis, feature historical electrical data collected that has been turned into
extraction, machine learning. models that aim to forecast the risk of failures for
components and systems. Due to their versatility, high
I. INTRODUCTION capability, and high accuracy performances, machine
One of the most demanding high-quality services, which learning implementation for fault detection and classification
is experiencing vast and rapid development nowadays, is the in power systems has grown significantly in recent years
electricity supply [1]. Consequently, the electrical grid has [12]. In [1], [4], [6] and [18], authors have used Artificial
been defined as the most extensive engineered system Neural Networks (ANNs) for fault detection, classification,
worldwide due to its indispensability in our daily lives and and location, collecting three-phase voltages and currents as
importance to the economies and progress of countries [2, input data for performing the output predictions. In [7], two
3]. Likewise, continuity and reliability have become techniques for fault identification and classification have
essential requirements for customers that are particularly been presented based on TA-QSSVM and A-QSSVM,
susceptible to power blackouts [4]. Therefore, fault respectively, a modification of the Support Vector Machine
condition detection is critical for reliable services [5]. (SVM) algorithm; both methods are unsupervised and online
Furthermore, detecting short circuits in distribution networks with good performance during their accuracy calculation.
is much more difficult than in transmission networks because Besides, [9] introduces a protection scheme using statistical
they are typically unbalanced and asymmetrical due to the models like energy, entropy, and standard deviation for
increasing incorporation of renewable energy on the load microgrids using Wavelet Transforms (WTs) and Decision
side as distributed generation (DGs) [6]. This consequence Trees (DTs) as a discriminating function. [12] has presented
has triggered the modernization and development of the a semi-supervised machine learning approach based on co-
smart grids, incorporating modern measurement and training over a microgrid, where the harmony search
communication systems into the power systems’ real-time algorithm is implemented to identify optimal wavelet
monitoring [5, 7, 8]. Nowadays, the stakeholders’ main goal transform families during the data pre-processing step. In [3,
is to improve the power grids by making them more 14], the authors tested three machine learning algorithms and
intelligent, reliable, and sustainable. Fault detection systems used the third level of decomposition for the wavelet
yield a practical, fast, and reliable form of relaying transform in data pre-processing. The simulations showed
operations. Additionally, they should perform satisfactorily the high capability of the Random Forest algorithm (RF)
under multiple operating conditions and diverse electrical over the other algorithms. In [15], there are three families of
grid parameters. When it comes to fault detections, they are discrete wavelet transforms (DWT) for feature extraction
supposed to be detected first, then correctly classified, and over the input data (e.g., motor current). After their
extraction, classification results were performed by RF and
XGBoost machine learning algorithms. [16, 17], they
Manuscript received July 07, 2021; accepted June 05, 2022. implemented WT and DWT for feature extraction,
*Corresponding author: [email protected]

11
Cabus and Altaş: Machine Learning Model Applications for Fault Detection and Classification in Distributed Power Networks

respectively. Zero-sequence current has been collected to consisted of unbalanced loads distributed over the array of
perform fault detection in conjunction with Principal three-phases and one-phase (i.e., AN and BN) grid
Component Analysis, SVM, and the Adaboost+CART configurations; operational voltage levels are 24.9 kV and
algorithm as discriminant functions. 4.16 kV, with a total load of 1769 kW and 1044 MVAr
This research is based on the detection and classification distributed over the grid; two capacitors installed at buses
of faults in a power-distributed network. Despite the 844 and 848; and two regulators located in the line segments
previous contributions mentioned, a lack of interest was at 814-850 and 832-852, respectively. The test feeder is
observed in selecting the WT or DWT families for data pre- modelled and tested using the DIgSILENT PowerFactory
processing. Furthermore, through this research, it has been software. Additionally, a three-phase meshed topology
concluded that the appropriate selection of WFT or DWT arrangement has been incorporated between nodes 816 and
families and level of decomposition can improve the 832, with a length of 2.5km and 301 configurations [20].
performance and accuracy of the machine learning Distributed generations have been modelled and placed at
algorithms. Additionally, unsupervised dimensionality the weakest point over the feeder; two three-phase PV
reduction algorithms are applied to feature extraction systems and one generator model available in the software's
techniques. Simulation results validate that using feature static generator library of the DIgSILENT software are used
extraction techniques and wavelet packet transform selection and placed in nodes 840, 848, and 890, respectively [20-22].
effectively can achieve high identification accuracy by Fig. 1 shows the modelled IEEE 34 bus test feeder. Fault
removing the less relevant features from consideration, simulations have been carried out all over the electrical
preventing the machine learning algorithms from overfitting system and are stored in Excel files. The detection and
or underfitting the datasets. classification stages have been performed in Jupyter
The rest of the paper is organized as follows: Section II Notebook and Python 3 with the implementation of Numpy,
reviews the proposed methods. Simulation results are Skicit-learn, and Pandas libraries available in the software.
presented in Section III. Conclusions are presented in Computer information Intel (R) Core (TM) i5-6400 CPU
Section IV. @2.70Hz, RAM 8GB, X64 bits Windows 10 Enterprise.

II. PROPOSED METHOD

The proposed method is based on three-phase voltages
and currents collected from the IEEE 34 bus test feeder
simulations: The zero-sequence voltage signal is
implemented for wavelet family selection using the Fig. 1. Modified IEEE 34 bus test feeder
minimum entropy decomposition and the Support Vector C. Input Data
Machine algorithm. After its identification, the model is set
Fig. 1 shows the modelled test feeder in DIgSILENT
with the best wavelet mother and decomposition level results
software, where different fault cases were simulated. Fault
to perform the feature extractions. In the training and test
data was generated through the DPL code and different
subsets, the data is split and scaled. Six types of
conditions such as random fault resistance, different fault
unsupervised dimensionality reduction are applied in order
locations, distributed generation level penetration, and
to reduce a high-dimension to a low-dimension, increasing
various fault current inception angles. Table I shows a
the machine learning algorithm’s performance. The fault
summary of the fault conditions in this paper.
detection is performed by applying zero-sequence voltage
components. For ground fault detection, the zero-sequence TABLE I. FAULT DATA CONFIGURATIONS
current is used for its identification. The three-phase voltage Condition Values
and current features are applied for the exact fault
Fault type LLL, LL, LLG, LG
classification.
Fault resistance in () 0, 20,50,80,100
A. Experimental Electrical Model Fault location in (%) 10, 25, 50,75,95
A three-phase system with a balanced or unbalanced Fault angle inception in () 0, 45,90,120
electrical grid has been described as the power system's DGs’ level penetration in (%) 0, 25, 50, 75, 100
backbone [3, 12]. It is also crucial to foresee the absence of
results in terms of reliability and continuity in balanced or After simulating the model, three-phase voltages and
unbalanced three-phase systems. Similarly, distribution currents are recorded. The collected data is then processed
networks are seeing a surge in the penetration of distributed with the Clarke Transform (1) to obtain the voltage and
generation, which is being driven by financial considerations current zero sequence component [3, 23]. It is essential to
and cost savings. The advantages of adopting decentralized mention that the fault or non-fault presences during the
generation and grid topology designs as a long-term solution energy harvesting process have been omitted. Through the
for the rising quantities of load demand and interconnections signals’ analysis, it was concluded that different operating
have been proven in previous studies [20]. states exhibit different behaviors, which can be observed in
the harmonic spectrum content or magnitudes of dominant
B. Case of Study
frequencies. Therefore, a zero-sequence voltage component
The IEEE 34 bus test feeder, located in Arizona, USA, has been implemented to detect faults over the grid and
was selected as a test system in this paper. The system select the best wavelet transform for data processing and
12
 DATA SCIENCE AND APPLICATIONS, VOL. 4, NO. 2, 2021

feature extraction; additionally, the zero-sequence current TABLE II. STATISTICAL FEATURES
component has been used for ground fault detection and Feature Functions
classification. Energy (3)

Mean Absolute (4)

(1)

R.M.S (5)
D. Wavelet Transform
A wavelet transform is a powerful tool consisting of low-
Variance (6)
and high-pass filters used in signal processing, voice
recognition, and a wide variety of applications due to its
good performance in frequency components (i.e. high and Standard Deviation (7)
low-frequency) [13]. In this paper, the features contained in
the voltage and current waveforms and zero-sequences are
extracted with the discrete wavelet transform [3, 9, 12, 13,
Kurtosis (8)
15]. The discrete wavelet transform (DWT) coefficients are
employed and collected using equation (2) derived from the
discretization of CWT [15]. Where (j, k) are integer
positives, ψ* represents the mother wavelet's complex Skewness (9)
conjugate, and x(t) is the signal analyzed.

(10
(2) Shape Factor
)

The use of several wavelet families for signal analysis has (11
Impulse Factor
been documented in previous research; however, no reason )
or technique for picking the mother wavelet has been
(12
presented. Aside from that, the mother wavelet selection and Crest Factor
)
decomposition levels may have a significant impact on how
regardless of its size the feature vectors are, which can result (13
Clearance Factor
in significantly different outcomes from the study. As a )
consequence, in this paper, a combination of the grid
searching method and the support vector machine is used to (14
Shannon’s Entropy
optimize the selection of wavelet family and grades of )
decomposition.
One of the essential procedures when dealing with data is
E. Feature Extraction that the data must be on the same scale because they might
The data collected from the simulation is raw data contain characteristics that vary significantly in magnitudes,
presented in quite large dimensions and unscaled. Thus, to units, and range, affecting the performance of the algorithms
obtain high accuracy, robust results, and low computational implemented for their analysis. Previous research has shown
complexity, the data must be as small as possible in that rescaling the data increases the accuracy and
dimensions. In this paper, the proposed method preprocesses performance of the algorithms. Therefore, the
the voltage and current signals through the discrete wavelet standardization function has been implemented, consisting
transform and extracts the most useful statistical features by of rescaling the data to have a mean of 0 and a standard
applying feature extraction techniques that might contain deviation of 1. Equation (15) shows its representation.
essential data during transient events such as faults or
perturbations [5, 9, 15, 24]. Then features are implemented (15)
to build the data-mining model using dimensionality
reduction and machine learning algorithms for fault
detection and classification. Twelve statistical features were F. Dimensionality Reduction
used for feature extraction in this paper; the dataset has 7655 Data compression is an essential subject in machine
instances, which contain three-phase voltage and current learning because it can help improve data storage,
with a 2-second duration. The details of the statistical computational efficiency, and predictive performance.
features are shown in Table II. Unsupervised dimensionality reduction techniques have
been implemented in this paper for feature extraction to
identify the most relevant and not relevant patterns
effectively, summarizing the original feature dataset from a
high-dimensional space onto a low-dimensional feature

13
Cabus and Altaş: Machine Learning Model Applications for Fault Detection and Classification in Distributed Power Networks

subspace for their processing [10, 24-27]. A summary of the TABLE IV. METRICS AND SCORING LIST
unsupervised dimensionality reduction techniques Name Functions
implemented in this paper is shown in Table III. Confusion (16
Matrix )
TABLE III. UNSUPERVISED DIMENSIONALITY REDUCTION
TECHNIQUES Accuracy
(17
Symbol Name )
PCA Principal Component Analysis Precision
KPCA Kernel Principal Component Analysis (18
)
LLE Locally Linear Embedding
Recall (19
Isomap Isometric Mapping )
DL Mini-Batch Dictionary Learning F1 (20
)
ICA Independent Component Analysis
Log-loss (21
G. Machine Learning Algorithms for Decision Making )
Machine learning algorithms aim to infer comprehensible
correlations or discover patterns between system variables in III. SIMULATION RESULTS
datasets that can be used later to forecast or comprehend
A. Wavelet family selection
system behaviors. Moreover, choosing the correct algorithm
for those tasks often depends on the amount, quality, and After calculating the zero components from the 3-phase
correlation of their data features, which involves a process of voltages and currents using equation (1), a subset of 600
trial and error [9, 24-27]. The proposed study uses five instances was collected randomly to process the best wavelet
supervised machine learning algorithms (i.e. Logistic family transform and decomposition levels using a zero-
Regression (LR), Support Vector Machine (SVM), Decision sequence voltage signal. In this paper, the Haar, db3, db4,
Trees (DT), Random Forest (RF), and K-Nearest Neighbors db6, Sym4, and Coif2 wavelets were selected and compared
(KNN)) in contrast with six unsupervised dimensionality in order to find the best wavelet family and decomposition
reduction algorithms to perform the data analysis in DWT level for signal analysis [3, 9, 12, 13, 15]. Table II: Twelve
selection, ground-fault detection, detection, and features were calculated from the subset. A proportion of
classification faults over the grid. 70:30 subset was used for training and testing models using
The machine learning algorithms’ robustness depends on the Grid Search algorithm and Support Vector Machine as a
their optimal hyperparameters’ adjustment. Therefore, the discriminant algorithm. The accuracy of detailed
proper values of the hyperparameters may boost the components at different decomposition levels is presented in
efficiency of the training model and test dataset accuracy Fig. 2. Accuracy results show that the db3 wavelet family at
[24-26]. In this paper, the Grid Search Algorithm has been the seventh decomposition level provides the most
proposed as an optimal model for hyperparameter selection. distinctive accuracy value (i.e., 96.85%); thus, the wavelet
This algorithm is based on implementing a trial and error decomposition step has been set to those values in order to
method to select the best parameters that give the best determine the features of the three-phase voltages and
accuracy from a subset of hyperparameters. However, currents for the next steps of the model.
choosing the best hyperparameters sometimes leads the
model to underfit or overfit the dataset, providing an
unsatisfactory performance. Consequently, a K-Folds Cross-
Validation model has been implemented to trade off the bias
and variance to avoid underfitting or overfitting the training
model [24-26].
H. Model Evaluation
The extracted features using pipeline models are fed into
the five machine learning algorithms. Quantifying the quality
of algorithms’ predictions is assessed by applying the
metrics and scoring listed in Table IV [26]. TN, TP, FP, and
FN mean True Negative, True Positive, False Positive, and Fig. 2. Wavelet family results
False Negative. For multiclass labels, a weighted average
B. Fault Detection Results
has been implemented; refer to [24-26]. Besides, Receiver
Operating Characteristic (ROC) curves and the Area Under The proposed fault detection algorithm is based on the
the Curve (AUC) percent have been implemented to zero-sequence voltage signal. The training and test ratios
visualize the algorithm’s performance. were chosen as 65:35, consisting of an unbalanced dataset
(i.e., no-faults and faults); twelve features were calculated
and scaled before performing dimensionality reduction. The
14
 DATA SCIENCE AND APPLICATIONS, VOL. 4, NO. 2, 2021

Grid Search algorithm was used for the best hyperparameter

selection. A 5-fold cross-validation approach was utilized in
order to validate the trained model and trade-off the bias and
variance. The most remarkable results are presented in Figs.
3–8; a summary of errors is presented in Table V.

Fig. 3. Fault detection results using Logistic Regression

Fig. 8. Fault detection ROC-AUC curves results

Fig. 4. Fault detection results using Support Vector Machine
C. Ground Fault Detection Results
For future decision-making, it is essential to verify if the
fault that occurred is or is not a ground fault. Hence, the
zero-sequence current signal is implemented to identify
whether a ground fault has happened in a feeder. Feature
vectors are extracted based on Table II for their analysis. In
addition to verifying its presence or not through machine
Fig. 5. Fault detection results using Decision Tree learning algorithms, it is being investigated whether its
detection can be improved by applying dimensionality
reduction algorithms, as proposed in Table III. The results
are presented in Fig. 9 to Fig. 14. In addition, a summary is
presented in Table VI, where it presents the loss functions
used for its identification and predictions.

Fig. 6. Fault detection results using Random Forest

Fig. 9. Ground fault detection using Logistic Regression

Fig. 7. Fault detection results using K-Nearest Neighbors

The best-adjusted model using dimensionality reduction

techniques was analysed through the F1-score, where
99.61% of LR through K-PCA was reached. Furthermore, Fig. 10. Ground fault detection using Support Vector
99.92% of the SVM and RF and 99.73% of the DTs were Machine
reached using the ICA algorithm. Besides, 99.96% of the K-
NN algorithm was reached using PCA and K-PCA
algorithms.

Fig. 11. Ground fault detection using Decision Trees

15
Cabus and Altaş: Machine Learning Model Applications for Fault Detection and Classification in Distributed Power Networks

Fig. 12. Ground fault detection using Random Forest Fig. 15. Fault classification using Logistic Regression

Fig. 13. Ground fault detection using K-Nearest Neighbors Fig. 16. Fault classification using Support Vector Machine

Fig. 17. Fault classification using Decision Trees

Fig. 18. Fault classification using Random Forest

Fig. 14. Ground fault detection ROC and AUC curves results

Among the best results of F1-scores obtained with

dimensionality reduction algorithms, the following can be Fig. 19. Fault classification using K-Nearest Neighbors
highlighted: 96.62% obtained through the implementation of
the ICA algorithm in SVM, 97.69% in DT implementing From the results obtained, it can be highlighted that most
PCA, 97.81% in RF with ICA, and 99.23% in KNN with K- of the proposed models achieved significant increases in the
PCA. On the other hand, despite the improvements achieved discriminant algorithms, which led to better performance in
in LR with dimensionality reduction algorithms, its poor identifying faults on the electrical network. For instance, in
performance in ground fault identification can be seen in the the LR algorithm, 99.81% was reached by implementing the
results, reaching 90.92%. PCA algorithm; in SVM and RF, 99.85% and 99.53% were
D. Fault Classification Results reached using the ICA algorithm, respectively. Besides, in
The first step for decision-making in situations of DT and KNN, 98.84% and 99.66% were reached by
uncertainty is to know that something is happening, which applying PCA and K-PCA algorithms, respectively.
leads to the identification of the situation discussed in the
previous sections. After getting prior knowledge, it is
necessary to identify the puzzle’s pieces or situation
experienced. Therefore, following the same model presented
above, a fault classification model is proposed by
implementing dimensionality reduction and machine
learning algorithms. The results of the applications are
presented below, from Fig. 15 to Fig. 19. A summary of cost
functions and classification results is presented in Table VII.

16
 DATA SCIENCE AND APPLICATIONS, VOL. 4, NO. 2, 2021

RF+LLE 0.18 0.04 1646 63 925 46

TABLE V. FAULT DETECTION SUMMARY RESULTS RF+Isomap 0.13 0.04 1644 62 926 48
Log- Mean RF+DL 0.24 0.09 1561 100 888 131
TP FP TN FN
Loss Abs. RF+ICA 0.13 0.03 1650 32 956 42
LR 0.135 0.035 2424 79 162 15 KN 0.22 0.07 1613 105 883 79
LR+PCA 0.043 0.012 2415 7 234 24 KN+PCA 0.25 0.01 1670 17 971 22
LR+KPCA 0.030 0.007 2422 2 239 17 KN+KPCA 0.34 0.01 1678 12 976 14
LR+LLE 0.122 0.069 2366 111 130 73 KN+LLE 0.63 0.02 1662 29 959 30
LR+Isomap 0.089 0.043 2361 37 204 78 KN+Isomap 0.45 0.04 1634 54 934 58
LR+DL 0.132 0.060 2350 71 170 89 KN+DL 0.30 0.09 1552 110 878 140
LR+ICA 0.045 0.009 2415 0 241 24 KN+ICA 0.38 0.01 1681 24 964 11
SVM 0.068 0.010 2429 16 225 10
SVM+PCA 0.019 0.004 2427 0 241 12 TABLE VII. FAULT CLASSIFICATION SUMMARY RESULTS
SVM+KPCA 0.026 0.007 2424 4 237 15 Log- Mean Correct Mis-
SVM+LLE 0.027 0.004 2437 10 231 2 Loss Abs. Classified classified
SVM+Isomap 0.033 0.007 2425 6 235 14 LR 0.24 0.14 2571 109
SVM+DL 0.118 0.053 2334 38 203 105 LR+PCA 0.05 0.05 2675 5
SVM+ICA 0.008 0.001 2435 0 241 4 LR+KPCA 0.13 0.18 2628 52
DT 0.105 0.056 2321 32 209 118 LR+LLE 0.12 0.31 2598 82
DT+PCA 0.058 0.010 2421 9 232 18 LR+Isomap 0.16 0.39 2590 90
DT+KPCA 0.037 0.007 2426 7 234 13 LR+DL 0.15 0.40 2585 95
DT+LLE 0.138 0.016 2406 9 232 33 LR+ICA 0.07 0.08 2670 10
DT+Isomap 0.116 0.012 2413 6 235 26 SVM 0.09 0.09 2604 76
DT+DL 0.214 0.040 2385 52 189 54 SVM+PCA 0.09 0.09 2629 51
DT+ICA 0.037 0.005 2432 6 235 7 SVM+KPCA 0.75 0.06 2635 45
RF 0.027 0.011 2425 15 226 14 SVM+LLE 0.09 0.12 2583 97
RF+PCA 0.026 0.005 2434 9 232 5 SVM+Isomap 0.08 0.08 2612 68
RF+KPCA 0.033 0.005 2436 10 231 3 SVM+DL 0.03 0.02 2669 11
RF+LLE 0.020 0.003 2436 6 235 3 SVM+ICA 0.02 0.01 2676 4
RF+Isomap 0.029 0.004 2434 7 234 5 DT 0.32 0.07 2627 53
RF+DL 0.246 0.015 2406 6 235 33 DT+PCA 0.40 0.05 2649 31
RF+ICA 0.018 0.001 2435 0 241 4 DT+KPCA 0.40 0.05 2649 31
KN 0.055 0.026 2393 23 218 46 DT+LLE 0.45 0.04 2645 35
KN+PCA 0.026 0.007 2437 0 241 2 DT+Isomap 0.62 0.07 2632 48
KN+KPCA 0.026 0.007 2437 0 241 2 DT+DL 0.81 0.10 2617 63
KN+LLE 0.091 0.003 2436 4 237 3 DT+ICA 0.55 0.06 2637 43
KN+Isomap 0.041 0.002 2433 0 241 6 RF 0.28 0.12 2577 103
KN+DL 0.047 0.015 2404 5 236 35 RF+PCA 0.23 0.21 2665 15
KN+ICA 0.016 0.002 2434 0 241 5 RF+KPCA 0.20 0.07 2625 55
RF+LLE 0.12 0.07 2622 58
TABLE VI. GROUND FAULT DETECTION SUMMARY RESULTS RF+Isomap 0.10 0.02 2664 16
Log- Mean RF+DL 0.21 0.07 2624 56
TP FP TN FN
Loss Abs. RF+ICA 0.14 0.02 2667 13
LR 0.30 0.12 1454 76 912 238 KN 0.09 0.04 2646 34
LR+PCA 0.26 0.11 1450 48 940 248 KN+PCA 0.12 0.02 2671 9
LR+KPCA 0.26 0.11 1450 48 940 248 KN+KPCA 0.12 0.02 2671 9
LR+LLE 0.26 0.12 1506 135 853 186 KN+LLE 0.24 0.05 2654 26
LR+Isomap 0.28 0.13 1411 61 927 281 KN+Isomap 0.15 0.04 2650 30
LR+DL 0.34 0.14 1426 97 891 266 KN+DL 0.34 0.09 2617 63
LR+ICA 0.26 0.11 1450 48 940 248 KN+ICA 0.11 0.03 2660 20
SVM 0.19 0.09 1473 23 965 219
SVM+PCA 0.16 0.05 1606 39 949 86
IV. CONCLUSIONS
SVM+KPCA 0.12 0.04 1613 35 953 79
SVM+LLE 0.19 0.05 1629 59 929 63 This paper studied the importance of wavelet family
SVM+Isomap 0.14 0.04 1656 68 920 36 selection for feature extraction. Additionally, unsupervised
SVM+DL 0.33 0.09 1535 84 904 157 dimensionality reduction models have been applied to the
SVM+ICA 0.15 0.04 1629 51 937 63 machine learning algorithms over a distributed network to
DT 0.26 0.08 1554 81 907 138 improve predictions and classifications. The method
DT+PCA 0.48 0.03 1651 37 951 41
developed utilizes three-phase voltages and currents as
DT+KPCA 0.43 0.05 1614 52 936 78
DT+LLE 0.53 0.04 1638 56 932 54 inputs to the model. Then, the zero-sequence components
DT+Isomap 0.67 0.05 1635 76 912 57 are calculated and implemented for wavelet selection,
DT+DL 0.38 0.09 1557 104 884 135 ground fault identification, and fault detection. The
DT+ICA 0.79 0.04 1647 54 934 45 unsupervised dimensionality reduction algorithms extract the
RF 0.18 0.06 1632 114 874 60 most relevant features for training the machine learning
RF+PCA 0.48 0.09 1642 29 959 50 algorithms. The grid search algorithm and K-fold cross-
RF+KPCA 0.11 0.03 1643 28 960 49 validation were implemented to choose the appropriate

17
Cabus and Altaş: Machine Learning Model Applications for Fault Detection and Classification in Distributed Power Networks

hyperparameters and reduce overfitting or underfitting in the vol. 34, no. 2, pp. 328-345, Feb. 2012, doi:
10.1109/TPAMI.2011.108.
models. The simulation results prove that some of the [12] T. S. Abdelgayed, W. G. Morsi and T. S. Sidhu, Fault Detection and
proposed models have achieved satisfactory performance in Classification Based on Co-training of Semisupervised Machine
fault detection and classification over the network studied. Learning, in IEEE Transactions on Industrial Electronics, vol. 65, no.
2, pp. 1595-1605, Feb. 2018, doi: 10.1109/TIE.2017.2726961.
[13] C., Hwan Kim and R. Aggarwal, Wavelet transforms in power
ACKNOWLEDGMENT systems. II. Examples of application to actual power system
Part of this work was presented orally at the IV. transients, in Power Engineering Journal, vol. 15, no. 4, pp. 193-202,
Aug. 2001, doi: 10.1049/pe:20010404.
International Conference on Data Science and Applications [14] H. Okumuş And F. M. Nuroğlu, Power System Event Classification
2021 with the title “Machine Learning Application for Fault Based on Machine Learning, 3rd International Conference on
Detection in Power Distributed Network”. Computer Science and Engineering (UBMK), Sarajevo, Bosnia And
Herzegovina, pp.402-405, 2018.
[15] N. Toma, R. and Kim, J.-M. Bearing Fault Classification of Induction
REFERENCES Motors Using Discrete Wavelet Transform and Ensemble Machine
[1] Karić A., Konjić T. Jahić A. Power System Fault Detection, Learning Algorithms. Appl.Sci. 2020,
Classification and Location using Artificial Neural Networks, In: 10,5251.https://doi.org/10.3390/app10155251Qwqwq
Hadžikadić M., Avdaković S. (eds) Advanced Technologies, [16] R., Papia and M., Debani, Support vector machine-based fault
Systems, and Applications II. IAT 2017. Lecture Notes in Networks classification and location of a long transmission line. Engineering
and Systems, vol 28. Springer, Cham. (2018) Science and Technology, an International Journal. 2016, 19.
https://doi.org/10.1007/978-3-319-71321-2_8 10.1016/j.jestch.2016.04.001.
[2] Y. Mo et al., Cyber–Physical Security of a Smart Grid Infrastructure, [17] Guo, Mou-Fa, Z. Xiao-Dan, C. Duan-Yu and Y. Nien-Che. Deep-
in Proceedings of the IEEE, vol. 100, no. 1, pp. 195-209, Jan. 2012, Learning-Based Earth Fault Detection Using Continuous Wavelet
doi: 10.1109/JPROC.2011.2161428. Transform and Convolutional Neural Network in Resonant
[3] H. Okumuş and F.M. Nuroglu, Wavelet Based Fault Detection and Grounding Distribution Systems. IEEE Sensors Journal.2017, PP. 1-
Classification Algorithm for a Real Distribution Feeder. EMITTER 1. 10.1109/JSEN.2017.2776238.
International Journal of Engineering Technology. 2019. 7. [18] Ray, P., Panigrahi, B.K. and Senroy, N., Hybrid methodology for
10.24003/emitter.v7i1.382. fault distance estimation in series compensated transmission line. IET
[4] Jamil, M., Sharma, S.K. and Singh, R. Fault detection and Gener. Transm. Distrib.,7:431-439.2013,https://doi.org/10.1049 /iet-
classification in electrical power transmission system using artificial gtd.2012.0243
neural network. SpringerPlus 4, 334 (2015). [19] J. Jiang et al., A Hybrid Framework for Fault Detection,
https://doi.org/10.1186/s40064-015-1080-x Classification, and Location—Part I: Concept, Structure, and
[5] H. A. Tokel, R. A. Halaseh, G. Alirezaei and R. Mathar, A new Methodology, in IEEE Transactions on Power Delivery, vol. 26, no.
approach for machine learning-based fault detection and 3, pp. 1988-1998, July 2011, doi: 10.1109/TPWRD.2011.
classification in power systems, 2018 IEEE Power & Energy Society 2141157.Qwqw
Innovative Smart Grid Technologies Conference (ISGT), Washington, [20] M. Davoudi, V. Cecchi and J. R. Agüero, Effects of stiffness factor
DC, USA, 2018, pp. 1-5, doi: 10.1109/ISGT.2018.8403343. on bus voltage variations in the presence of intermittent distributed
[6] Dehghani, Farzad & Khodnia, Fereydoun & Dehghan, Esfandiar. generation, 2015 North American Power Symposium (NAPS),
Fault location of unbalanced power distribution feeder with Charlotte, NC, USA, 2015, pp. 1-6, doi:
distributed generation using neural networks. CIRED - Open Access 10.1109/NAPS.2015.7335187.
Proceedings Journal. 2017. 1134-1137. 10.1049/oap- [21] R. C. Dugan and W. H. Kersting, Induction machine test case for the
cired.2017.0007. 34-bus test feeder -description, 2006 IEEE Power Engineering
[7] N. Shahid, S. A. Aleem, I. H. Naqvi and N. Zaffar, Support Vector Society General Meeting, Montreal, QC, Canada, 2006, pp. 4 pp.-,
Machine based fault detection & classification in smart grids, 2012 doi: 10.1109/PES.2006.1709506.
IEEE Globecom Workshops, Anaheim, CA, USA, 2012, pp. 1526- [22] K. Balamurugan, D. Srinivasan, and T. Reindl, Impact of Distributed
1531, doi: 10.1109/GLOCOMW.2012.6477812. Generation on Power Distribution Systems, Energy Procedia, Volume
[8] A. V. Masa, S. Werben and J. C. Maun, Incorporation of data-mining 25, 2012, Pages 93-100,ISSN1876-6102, https://doi.org/10.1016/
in protection technology for high impedance fault detection, 2012 j.egypro.2012.07.013.Sfds
IEEE Power and Energy Society General Meeting, San Diego, CA, [23] Sasa Mujovic, Snezana Vujosevic and Luka Vujosevic (2018) Zero-
USA, 2012, pp. 1-8, doi: 10.1109/PESGM.2012.6344893. Sequence Voltage-based Method for Determination and
[9] S. Samantaray, D. Mishra and G. Joos, A Combined Wavelet and Classification of Unloaded Overhead Line Operating Conditions at
Data-Mining Based Intelligent Protection Scheme for Microgrid, the Moment of Energization, Electric Power Components and
2018 IEEE Power & Energy Society General Meeting (PESGM), Systems,46:2,162-176,DOI:10.1080/15325008.201 8.1433252
Portland, OR, USA, 2018, pp. 1-1, doi: [24] Raschka S. and Mirjalili V., Python Machine Learning: Machine
10.1109/PESGM.2018.8586480. Learning and Deep Learning with Python, Scikit-learn, and
[10] Chen, K., Huang, C. and He, J., Fault detection, Classification and TensorFlow, Packt
location for transmission lines and distribution systems: a review on Publishing,2017,ISBN1787125939/9781787125933
the methods. High [25] Ankur A. Patel, Hands-On Unsupervised Learning Using Python,
Voltage,1:25,33,2016.https://doi.org/10.1049/hve.2016.0005. O’Reilly Media, Inc., 2019, ISBN9781492035640
[11] C. Rudin et al., Machine Learning for the New York City Power Grid, [26] https://scikit-learn.org/stable/index.html (Accessing date: 9 Feb 2021)
in IEEE Transactions on Pattern Analysis and Machine Intelligence, [27] https://medium.com/ (Accessing date: 14 Feb 2021)
[28] https://www.mathworks.com/help/ (Accessing date: 20 Feb 2021)

18