Data_Mining_for_Classification_of_Power_Quality_Problems_Using_WEKA_and_the_Effect_of_Attributes_on_Classification_Accuracy
Data_Mining_for_Classification_of_Power_Quality_Problems_Using_WEKA_and_the_Effect_of_Attributes_on_Classification_Accuracy
(2018) 3:29
Protection and Control of
https://doi.org/10.1186/s41601-018-0103-3 Modern Power Systems
Abstract
There is growing interest in power quality issues due to wider developments in power delivery engineering. In
order to maintain good power quality, it is necessary to detect and monitor power quality problems. The power
quality monitoring requires storing large amount of data for analysis. This rapid increase in the size of databases has
demanded new technique such as data mining to assist in the analysis and understanding of the data. This paper
presents the classification of power quality problems such as voltage sag, swell, interruption and unbalance using
data mining algorithms: J48, Random Tree and Random Forest decision trees. These algorithms are implemented
on two sets of voltage data using WEKA software. The numeric attributes in first data set include 3-phase RMS
voltages at the point of common coupling. In second data set, three more numeric attributes such as minimum,
maximum and average voltages, are added along with 3-phase RMS voltages. The performance of the algorithms is
evaluated in both the cases to determine the best classification algorithm, and the effect of addition of the three
attributes in the second case is studied, which depicts the advantages in terms of classification accuracy and
training time of the decision trees.
Keywords: Power quality problems, Classification, Data mining, Decision trees, J48, Random tree, Random forest, WEKA
© The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to
the Creative Commons license, and indicate if changes were made.
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 2 of 12
online or offline analysis is needed to be carried out to is faster and provides better classification accuracy at every
classify the disturbances [4–7]. case with and without noise. It is also easier to implement
Vast and increasing volumes of data obtained from than SVM. Moreover, the decision tree worked satisfactor-
power quality monitoring system, requires the use of ily with both synthesized and real signals.
data mining technique for analyzing the data. Data min- Decision trees such as J48, Logistic Model Tree
ing technology is an effective tool to deal with massive (LMT), Reduced Error Pruning (REP) Tree, Random
data, and to detect the useful patterns in those data. In Tree, Simple Cart, Random Forest are used for the clas-
power systems, data can be raw waveforms (voltages and sification purpose [15–17]. Random Forest is used for
currents) sampled at relatively high sampling frequen- the classification of PQ disturbances [18] and fault rec-
cies, pre-processed waveforms (e.g., RMS values) or sta- ord detection in data center of large power grid [19]. J48
tus variables (e.g., if a relay is opened or closed) which is compared with Random Forest in the classification of
are typically sampled at low sampling frequencies [8]. power quality disturbances and found that Random For-
Classification of data is an important task in the data est is more accurate than J48 [20]. The performance of
mining process that extracts models for describing clas- Random Tree is observed to be better than REP Tree,
ses and predicts target class for data instances. Simple Cart [21], Logical Analysis of Data (LAD) Tree
Today, several standard classifiers are available, among and Random Forest [22] for the classification purpose. It
which the decision trees are most powerful and popular has been found that whenever correct attributes are se-
for both classification and prediction. Decision trees are lected before classification, accuracy of data mining algo-
flexible enough to handle items with a mixture of rithms is improved significantly [23, 24]. This paper
real-valued and categorical features, as well as items with focuses on how data mining techniques of J48, Random
some missing features. These are more interpretable Tree and Random Forest decision trees are applied to
than other classifiers such as Artificial Neural Networks classify power quality problems of voltage sag, swell,
(ANN) and Support Vector Machines (SVM) because interruption and unbalance. The effect of data attributes
they combine simple questions about the data in an on the classification accuracy and time taken for training
understandable way [9]. They are expressive enough to the decision trees is also discussed.
model many partitions of the data that are not as easily The paper is organized as follows: Section 2 gives defi-
achieved with classifiers that rely on a single decision nitions and causes of power quality problems like volt-
boundary such as logistic regression or SVM. Decision age sag, swell, interruption and unbalance along with
trees naturally support classification problems with more their typical figures. Section 3 deals with the basics of
than two classes and can be modified to handle regres- data mining and explains about J48, Random Tree and
sion problems. Finally, once constructed, they classify Random Forest algorithms. This Section also briefs
new items quickly [10]. about WEKA software used for implementing data min-
In [11], SVM, ANN, logistic regression, Naïve Bayes, ing for the classification purpose. In Section 4, the
classification and regression trees, C5.0 algorithm, Quick, MATLAB simulation circuit is given which is used for
Unbiased and Efficient Statistical Tree (QUEST), generating the data for various power quality problems.
CHi-square Automatic Interaction Detector (CHAID) and The testing and results of the data mining algorithms
discriminant analysis have been implemented for classifica- obtained from WEKA are discussed in Section 5. Finally,
tion on nine datasets. According to the experimental re- Section 6 gives conclusions of the work from the ob-
sults, C5.0 model proved to have the best performance. served results.
The performances of J48 decision tree, Multi-Layer Per-
ceptron (MLP) and Naïve Bayes classification algorithms
2 Power quality problems
were studied with respect to training time and accuracy of
Power quality problem is defined as any power problem
prediction [12]. It is observed that MLP takes highest train-
manifested in voltage, current, or frequency deviations
ing time for each of the data instances than J48 decision
that results in failure or misoperation of customer
tree and Naïve Bayes classifiers. The accuracy of Naïve
equipment. Some of the commonly occurring power
Bayes reduces as the data size increases. J48 and MLP
quality problems in a power system are voltage sag,
showed high accuracies with low as well as higher data
swell, interruption and unbalance [25].
sizes. The performance of ANN and SVM is evaluated for
the classification of sag, swell, interruption, harmonics and
flicker [13]. It is found that SVM outperform ANN in 2.1 Voltage sag
terms of classification accuracy and computation time. Ten Voltage sag is defined as a decrease in RMS voltage be-
different types of disturbances such as sag, swell, interrup- tween 0.1 p.u. to 0.9 p.u. at the power frequency for du-
tion with and without harmonics, are classified using SVM rations from 0.5 cycles to 1 min, reported as the
and decision tree [14]. It is observed that the decision tree remaining voltage. Voltage sags can occur due to short
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 3 of 12
2.3 Interruption
An interruption occurs when the supply voltage or load and (v) the rapid increase in computer technology [26].
current decreases to less than 0.1 p.u. for a period of time The ultimate goal of data mining is to discover useful in-
not exceeding 1 min. Interruptions can be the result of formation from large amounts of data in many different
power system faults, lightning, equipment failures and ways using rules, patterns and classification [27]. Data
control malfunctions. Interruption is illustrated in Fig. 3. mining can be used to identify anomalies that occur as a
result of network or load operation, which may not be
2.4 Voltage unbalance acknowledged by standard reporting techniques. It is
In a 3-phase system, voltage unbalance takes place when proposed that data mining can provide answers to the
the magnitudes of phase or line voltages are different, or end-users about PQ problems by converting raw data
the phase angles differ from the balanced conditions, or into useful knowledge [28, 29].
both. The sources of voltage unbalance are unbalanced Many people treat data mining as a synonym for an-
faults, single-phase loads on a three-phase circuit and other popularly used term, Knowledge Discovery from
blown fuses in one phase of a 3-phase capacitor bank. Data (KDD), while others view data mining as merely an
The three phase voltages during an unbalanced fault are essential step in the process of knowledge discovery. The
as shown in Fig. 4. knowledge discovery process is an iterative sequence of
the following steps: (i) Data cleaning, (ii) Data integra-
3 Data mining tion, (iii) Data selection, (iv) Data transformation, (v)
Data mining is a process that uses a variety of data ana- Data mining, (vi) Pattern evaluation and (vii) Knowledge
lysis tools to identify hidden patterns and relationships presentation. Steps (i) through (iv) are different forms of
within the data. These tools are a mixture of machine data pre-processing, where data are prepared for mining.
learning, statistics and database utilities. Data mining The data mining step may interact with the user or a
has recently obtained popularity within many research knowledge base. The interesting patterns are presented
fields over classical techniques for the purpose of analyz- to the user and may be stored as new knowledge in the
ing data due to (i) a vast increase in the size and number knowledge base. The preceding view shows data mining
of databases, (ii) the decrease in storage device costs, as one step in the knowledge discovery process, albeit an
(iii) an ability to handle data which contains distortion essential one because it uncovers hidden patterns for
(noise, missing values, etc.), (iv) continuous progress in evaluation. However, in industry, in media, and in the
the implementation of automatic learning techniques research milieu, the term data mining is often used to
Fig. 1 Voltage waveform during Sag Fig. 3 Voltage waveform during Interruption
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 4 of 12
3.1.1 J48
J48 is an open source Java implementation of the C4.5
algorithm in the WEKA data mining tool. It creates a
binary tree. It is one of the most useful decision tree ap-
proach for classification problems. It employs top-down
and greedy search through all possible branches to con-
struct a decision tree to model the classification process.
In order to classify a new item, it first needs to create a
decision tree based on the attribute values of the avail-
able training data. So, whenever it encounters a set of
Fig. 4 Three phase voltages during Unbalance condition
items (training set), it identifies the attribute that dis-
criminates the various instances more clearly. This fea-
ture, which is able to tell us more about the data
refer to the entire knowledge discovery process [30].The instances, so that we can classify them the best, is said
data mining process differs from classical statistical to have the highest information gain. Now, among the
methods in the way that statistical methods focus only possible values of this feature, if there is any value for
on model estimation, while data mining techniques which there is no ambiguity, i.e., for which the data in-
focus on both model formation and its performance. stances falling within its category have the same value
Another significant difference is that statistical methods for the target variable, then that branch is terminated
fail to analyze data with missing values, or data that con- and the target value is assigned to it. For the other cases,
tains a mixture of numeric and qualitative forms. Data another attribute is selected which gives the highest in-
mining techniques, instead, can analyze and cope intelli- formation gain. The process is continued in this manner
gently with records containing missing values, as well as until a clear decision is obtained about what combin-
a mixture of qualitative and quantitative data, without ation of attributes gives a particular target value, or till
tedious manual manipulation [31, 32]. all the attributes are completed. In the event that all the
Data mining starts with the real data, collected from the attributes are finished, or if the unambiguous result can-
real equipment. In fact, more the diversified data, more not be obtained from the available information, we as-
accurate and better result is obtained. So, if hundreds of sign this branch a target value that the majority of the
parameters are recorded and available for analysis, data items under this branch possesses. Now that we have
mining can consider and use all the data which is col- the decision tree, we follow the order of attribute selec-
lected. Data mining methods are well equipped to handle tion as we have obtained for the tree. By checking all the
large amount of data and to detect the useful patterns in respective attributes and their values with those seen in
these data that allow us to improve the performance. Data the decision tree model, the target value of the new in-
mining methodologies and algorithms have their origins stance can be predicted. J48 classification is based on
in many different disciplines. For example, researchers on the decision trees or rules generated from them [34].
artificial intelligence have proposed various methods and The simple tree structure of J48 is as shown in Fig. 5.
techniques that can efficiently “mimic” how real people
(“experts”) can detect difficult hidden patterns in large
amounts of complex data. 3.1.2 Random tree
A Random Tree is a decision tree that is formed by a
stochastic process. In standard tree, each node is split
3.1 Methods: Data mining algorithms
using the best split among all attributes. In a Random
There are many data mining algorithms available, among
Tree, each node is split using the best among the subset
which the most widely used algorithms for classification
of randomly chosen attributes at that node. Random
are J48, Random Tree and Random Forest. These are de-
Tree algorithm has an option to estimate the class prob-
cision trees which use divide-and-conquer strategies as a
abilities for classification. Random Trees have been in-
form of learning by induction. Thus, these algorithms
troduced by Leo Breiman and Adele Cutler. This
use a tree representation, which helps in pattern classifi-
algorithm can deal with both classification and regres-
cation in data sets, being hierarchically structured in a
sion problems [21, 35]. The structure for a Random Tree
set of interconnected nodes. The internal nodes test an
is shown in Fig. 6.e
input attribute/feature in relation to a decision constant
and, this way, determines what will be the next descend-
ing node. Therefore, the nodes considered as leaves clas- 3.1.3 Random forest
sify the instances that reach them according to the This algorithm uses a set of classifiers based on deci-
associated label [33]. sion trees. Random Forest fits many classification
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 5 of 12
trees to a data set and then combines the prediction 3.2 Data mining tool: WEKA software
from all the correlated trees. Each tree depends on WEKA, formally called Waikato Environment for Know-
the value of a separately sampled random vector. ledge Analysis, is a computer program that was developed
Random Forest corresponds to a collection of com- at the University of Waikato in New Zealand for the pur-
bined decision trees {hk(x, Tk)}, for k = 1, 2,..., n, pose of identifying information from raw data gathered
where n is the number of trees and Tk is the training from agricultural domains. WEKA is a state-of-the-art fa-
set built at random and identically distributed, hk rep- cility for developing machine learning techniques and
resents the tree created from the vector Tk and is re- their application to real-world data mining problems. It is
sponsible for producing an output x. a collection of machine learning algorithms for data min-
The trees that make up the Random Forest are built ing tasks. The algorithms are applied directly to a dataset.
randomly selecting ‘m’ (value fixed for all nodes) attri- WEKA supports many different standard data mining
butes in each node of the tree; where the best attribute tasks such as data pre-processing, classification, clustering,
is chosen to divide the node. The vector used for train- regression, visualization and feature selection. The basic
ing each tree is obtained using a random selection of the premise of the application is to utilize a computer applica-
instances. Thus, to determine the class of an instance, all tion that can be trained to perform machine learning cap-
the trees indicate an output and the most voted is se- abilities and derive useful information in the form of
lected as the final result. So, the classification error de- trends and patterns. WEKA is an open source application
pends on the strength of individual trees of the forest that is freely available under the GNU general public li-
and the correlation between any two trees in the forest cense agreement. It is user friendly with a graphical inter-
[20]. Figure 7 shows the tree diagram of a Random For- face that allows for quick set up and operation. WEKA
est. The various differences between the three data min- operates on the predication that the user data is available
ing algorithms are presented in Table 1. as a flat file or relation, this means that each data object is
Class B
Class A Class A
Majority voting
described by a fixed number of attributes that usually are disconnecting the supply. The voltage unbalance is created
of a specific type, normal alpha-numeric or numeric by a 3-phase unbalance fault. The 3-phase RMS voltages
values. The WEKA application allows novice users a tool calculated at the Point of Common Coupling (PCC) are
to identify hidden information from database and file sys- used as the main data for classification of the power quality
tems with simple to use options and visual interfaces [36]. problems. The data is sampled at a frequency of 2 kHz.
From the simulation, 400,001 data samples are obtained,
4 Circuit for generating data for classification among which 31,438 samples contain sag, 22,506 samples
The circuit shown in Fig. 8 is modelled in MATLAB Simu- contain swell, 5441 samples contain interruption, 14,268
link. The circuit consists of a 33/11 kV distribution sub- samples contain unbalance problem and remaining 326,348
station connected to a 2 km distribution line having a 11/ samples have no power quality problems. This data is used
0.433 kV distribution transformer supplying to a load of for classification by data mining algorithms.
190 kW, 140 kVAr [37]. It is simulated to get the data for
various voltage sags, swells, interruptions and unbalance 5 Results and discussion
problems. Voltage sags are created by balanced 3-phase to The data samples obtained from simulations carried out
ground faults with varied fault impedance and duration, for on the system shown in Fig. 8, are stored in a datasheet.
different categories of sags. Voltage swells are created by Using this data, a class attribute is formulated which is
switching capacitors of different capacitances connecting to used to differentiate sag, swell, interruption and unbal-
the line, for varied durations to get different categories of ance. With this information, an ARFF (Attribute-Rela-
swells. Interruptions are introduced by opening circuit tion File Format) file is written. ARFF file is an ASCII
breaker 1 (CB 1) for different time durations, thereby text file that describes a list of instances sharing a set of
attributes. ARFF files were developed by the machine mining in WEKA indicating total number of instances,
learning project at the Department of Computer Science the number of attributes and number of samples under
of the University of Waikato for use with the WEKA each class of power quality problems along with a bar
machine learning software. The file has a header section graph. The attributes used in this case are the numeric
followed by data section. The header section contains re- values of three phase RMS voltages, namely Va, Vb and
lation declarations mentioning the name of the relation Vc along with the class attribute. The class attribute
and attribute declarations listing the attributes (the col- value is “NoProb” for samples containing no power qual-
umns in the data) with their types [38]. ity problem, “Sag” for samples with voltage sag, “Swell”
The ARFF file is used to load the data into WEKA when samples contain voltage swell, “Intr” for samples
software for the classification of the power quality prob- containing interruption and “Unbal” for samples with
lems. Figure 9 shows the pre-processing stage of data voltage unbalance condition.
Table 2 Comparison of Data Mining Algorithms with 4 attributes for evaluation on training set
S. No. Algorithm Cases Tested Correct Classification Incorrect Classification Training
Time (s)
No. of Samples Accuracy (%) No. of Samples Inaccuracy (%)
1 J48 Sags 31,438 100 0 0 3.59
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,257 99.9229 11 0.0771
No PQ problems 326,348 100 0 0
Overall 399,990 99.9973 11 0.0027
2 Random Tree Sags 31,438 100 0 0 1.91
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,268 100 0 0
No PQ problems 326,348 100 0 0
Overall 400,001 100 0 0
3 Random Forest Sags 31,438 100 0 0 25.51
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,268 100 0 0
No PQ problems 326,348 100 0 0
Overall 400,001 100 0 0
The data loaded into WEKA is used to train the data in WEKA. It indicates the total number of instances, the
mining algorithms: J48, Random Tree and Random For- number of attributes and number of samples under each
est for the classification purpose. After training, the al- class of power quality problems along with a bar graph.
gorithms are tested based on the given training set and The information is same as that shown in Fig. 9, except for
as well as using stratified 10-fold cross validation [39]. the number of attributes taken.
The results obtained after testing the algorithms using Using the data of seven attributes, loaded into WEKA,
training set are indicated in Table 2. It is observed that the data mining algorithms are trained and tested. The
the overall accuracy of J48 algorithm is 99.9973%, results obtained after testing the algorithms using train-
whereas Random Tree and Random Forest algorithms ing set are indicated in Table 4. It is observed that the
have an accuracy of 100% in the classification of the overall accuracy of J48 algorithm is 99.9983%, whereas
power quality problems. It is also clear that the training Random Tree and Random Forest algorithms have an
time taken by the Random Tree is only 1.91 s, which is accuracy of 100% in the classification of the power qual-
very less as compared to J48 and Random Forest. ity problems. It is again clear that the training time
The results obtained after testing the algorithms using taken by the Random Tree (1.88 s) is very less as com-
stratified 10-fold cross validation are shown in Table 3. pared to J48 and Random Forest algorithms. Comparing
From the results, it is seen that the Random Tree has a the results of Tables 2 and 4, it is observed that the clas-
more overall accuracy (99.9943%) and takes less training sification accuracy of J48 algorithm is improved in seven
time (1.86 s) as compared to J48 and Random Forest al- attributes case. Random Tree and Random Forest has
gorithms. From Tables 2 and 3, it is clear that with only 100% accuracy in both the cases. It is also observed that
four attributes in the data, Random Tree is best of the the training time taken by all the algorithms is reduced
three algorithms as it has more accuracy and takes very in seven attributes case.
less time for training. Table 5 shows the results obtained after testing the algo-
In the next case, along with Va, Vb, Vc and class attri- rithms using stratified 10-fold cross validation. From the
bute, three more extra numeric attributes are included. results, it is seen that the Random Forest has highest over-
They are average (Vavg), minimum (Vmin) and maximum all accuracy (99.9973%) whereas Random Tree has lowest
(Vmax) values of the three phase voltages. Figure 10 shows training time (1.75 s) as compared to other algorithms.
the pre-processing stage of data mining for seven attributes Comparing the results of Tables 3 and 5, it is clear that for
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 9 of 12
Table 3 Comparison of Data Mining Algorithms with 4 attributes for stratified 10-fold cross-validation
S. No. Algorithm Cases Tested Correct Classification Incorrect Classification Training
Time (s)
No. of Samples Accuracy (%) No. of Samples Inaccuracy (%)
1 J48 Sags 31,418 99.9364 20 0.0636 4.58
Swells 22,506 100 0 0
Interruptions 5438 99.9448 3 0.0552
Unbalances 14,242 99.8178 26 0.1822
No PQ problems 326,340 99.9975 8 0.0025
Overall 399,944 99.9858 57 0.0142
2 Random Tree Sags 31,431 99.9777 7 0.0223 1.86
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,255 99.9089 13 0.0911
No PQ problems 326,345 99.9991 3 0.0009
Overall 399,978 99.9943 23 0.0057
3 Random Forest Sags 31,429 99.9714 9 0.0286 24.5
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,255 99.9089 13 0.0911
No PQ problems 326,345 99.9991 3 0.0009
Overall 399,976 99.9938 25 0.0062
all the algorithms, the classification accuracy is improved From all the results obtained by testing the algorithms
and the training time is reduced using seven attributes. for classification of power quality problems, comparison
Thus, it indicates that the generalization capabilities of the of overall performance of the algorithms is indicated
algorithms are enhanced by including the extra attributes briefly in Table 6. It is observed that Random Forest
in the second case. gives most accurate results, but takes more time for
Table 4 Comparison of Data Mining Algorithms with 7 attributes for evaluation on training set
S. No. Algorithm Cases Tested Correct Classification Incorrect Classification Training
Time (s)
No. of Samples Accuracy (%) No. of Samples Inaccuracy (%)
1 J48 Sags 31,436 99.9936 2 0.0064 3.26
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,263 99.9650 5 0.0350
No PQ problems 326,348 100 0 0
Overall 399,994 99.9983 7 0.0017
2 Random Tree Sags 31,438 100 0 0 1.88
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,268 100 0 0
No PQ problems 326,348 100 0 0
Overall 400,001 100 0 0
3 Random Forest Sags 31,438 100 0 0 22.58
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,268 100 0 0
No PQ problems 326,348 100 0 0
Overall 400,001 100 0 0
Table 5 Comparison of Data Mining Algorithms with 7 attributes for stratified 10-fold cross-validation
S. No. Algorithm Cases Tested Correct Classification Incorrect Classification Training
Time (s)
No. of Samples Accuracy (%) No. of Samples Inaccuracy (%)
1 J48 Sags 31,424 99.9554 14 0.0445 3.68
Swells 22,506 100 0 0
Interruptions 5440 99.9816 1 0.0184
Unbalances 14,252 99.8878 16 0.1122
No PQ problems 326,348 100 0 0
Overall 399,970 99.9923 31 0.0077
2 Random Tree Sags 31,429 99.9713 9 0.0287 1.75
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,258 99.9299 10 0.0701
No PQ problems 326,347 99.9997 1 0.0003
Overall 399,981 99.995 20 0.005
3 Random Forest Sags 31,434 99.9872 4 0.0128 21.25
Swells 22,506 100 0 0
Interruptions 5441 100 0 0
Unbalances 14,262 99.9579 6 0.0421
No PQ problems 326,347 99.9997 1 0.0003
Overall 399,990 99.9973 11 0.0027
Asha Kiranmai and Jaya Laxmi Protection and Control of Modern Power Systems (2018) 3:29 Page 11 of 12
10. Kingsford, C., & Salzberg, S. L. (2008). What are decision trees? Nat 37. Suresh, K., & Chandrashekhar, T. (2012). Characteristics analysis of voltage
Biotechnol, 26(9), 1011–1013. sag in distribution system using RMS voltage method. ACEEE Int J on
11. Sampson, A. (2012). Comparing classification algorithms in data mining. A Electrical and Power Engineering, 3(1), 55–61.
Thesis, Central Connecticut State University, New Britain, Connecticut. 38. Soman, K. P., Diwakar, S., & Ajay, V. (2006). Data mining: Theory and practice.
12. Akinola, S., & Oyabugbe, O. (2015). Accuracies and training times of data Delhi: PHI Learning Pvt. Ltd..
mining classification algorithms: An empirical comparative study. Journal of 39. Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy
Software Engineering and Applications, 8, 470–477. estimation and model selection. Montreal, Quebec, Canada: In Proc.
13. Manimala, K., Selvi, K., & Ahila, R. (2008). Artificial intelligence techniques 14th International Joint Conference on Artificial Intelligence (IJCAI); 2.
applications for power disturbances classification. World Academy of Science, p. 1137–1143.
Engineering and Technology, 22, 833–840.
14. Upadhyaya, S., & Mohanty, S. (2016). Localization and classification of power
quality disturbances using maximal overlap discrete wavelet transform and
data mining based classifiers. IFAC-Papers OnLine, 49(1), 437–442.
15. Jeya Sheela, Y., & Krishnaveni, S. H. (2017). A comparative analysis of various
classification trees (pp. 1346–1352). Kollam: IEEE International Conference on
Circuit, Power and Computing Technologies (ICCPCT).
16. Pandey, P., & Prabhakar, R. (2016). An analysis of machine learning techniques
(J48 & AdaBoost)-for classification (pp. 1–6). Delhi: IEEE 1st India International
Conference on Information Processing (IICIP).
17. Sewaiwar, P., & Verma, K. K. (2015). Comparative study of various decision
tree classification algorithm using WEKA. International Journal of Emerging
Research in Management & Technology, 4(10), 87–91.
18. Anton Domini Sta. Cruz & Jordan Rel C. Orillaza (2017). Classification of single
and multiple PQ disturbances based on DWT and RF classifiers. Penang: IEEE
Region 10 Conference, TENCON 2017.
19. Zhou, J., Ge, Z., Gao, S., & Yanli, X. (2016). Fault record detection with random
forests in data center of large power grid (pp. 25–28). Xi’an: IEEE PES Asia-
Pacific Power and Energy Engineering Conference (APPEEC).
20. Borges, Fábbio A. S., Fernandes, Ricardo A. S., Lucas, A. M. & Silva, Ivan N.
(2015). Comparison between random Forest algorithm and J48 decision trees
applied to the classification of power quality disturbances. Las Vegas, Nevada,
USA: Int'l Conf. Data Mining (DMIN’15), 146–147.
21. Kalmegh, S. (2015). Analysis of WEKA data mining algorithm REP tree, simple
cart and random tree for classification of Indian news. International Journal
of Innovative Science, Engineering & Technology, 2(2), 438–446.
22. Kalmegh, S. R. (2015). Comparative analysis of WEKA data mining algorithm
RandomForest, RandomTree and LADTree for classification of indigenous
news data. International Journal of Emerging Technology and Advanced
Engineering, 5(1), 507–517.
23. Choudhary, N. K., Shinde, Y., Kannan, R., & Venkatraman, V. (2014). Impact of
attribute selection on the accuracy of multilayer perceptron. International
Journal of IT & Knowledge Management, 7(2), 32–36.
24. Dinakaran, S., & Thangaiah, D. R. J. (2013). Role of attribute selection in
classification algorithms. International Journal of Scientific & Engineering
Research, 4(6), 67–71.
25. Dugan, R. C., McGranaghan, M. F., Santoso, S., & Wayne Beaty, H. (2004).
Electrical power systems quality (2nd ed.). New York: Mc Graw-Hill.
26. Groth, R. (2000). Data mining: Building competitive advantage. USA: Prentice
Hall.
27. Mannila, H. (1996). Data mining: Machine learning, statistics and databases.
Stockholm, Sweden: In Proc. 8th Inter. Conf. on Scientific and Statistical
Database Systems.
28. Olaru, C., & Wehenkel, L. (1999). Data mining. IEEE Computer Applications in
Power, 12, 19–25.
29. Santoso, S. & Lamoree, J. D. (2000). Power quality data analysis: From raw
data to knowledge using knowledge discovery approach. IEEE PES Summer
Meeting.
30. Han, J., Kamber, M., & Pei, J. (2011). Data Mining: Concepts and Techniques
(3rd ed.). Burlington, Massachusetts, United States: Morgan Kaufmann.
31. Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., & Uthurusamy, R. (1996).
Advances in knowledge discovery and data mining. Menlo Park: AAAI Press.
32. Westphal, C., & Balxton, T. (1998). Data mining solutions: Method and tools
for solving real-world problems. USA: Wily.
33. Witten, I. H., & Frank, E. (2005). Data mining: Pratical machine learning tools
and techniques (2nd ed.). San Francisco: Morgan Kaufmann Publishers.
34. Classification methods. www.d.umn.edu/~padhy005/Chapter5.html .
35. Hamsagayathri, P., & Sampath, P. (2017). Performance analysis of breast
cancer classification using decision tree classifiers. Int J Curr Pharm Res, 9(2),
19–25.
36. WEKA manual. www.nilc.icmc.usp.br/elc-ebralc2012/minicursos/WekaManual-3-
6-8.pdf.