SciELO - Scientific Electronic Library Online

 
vol.35 número2MRP OPTIMIZATION MODEL FOR A PRODUCTION SYSTEM WITH REMANUFACTURINGANALYTIC NETWORK PROCESS AND BALANCED SCORECARD APPLIED TO THE PERFORMANCE EVALUATION OF PUBLIC HEALTH SYSTEMS índice de autoresíndice de assuntospesquisa de artigos
Home Pagelista alfabética de periódicos  

Serviços Personalizados

Journal

Artigo

Indicadores

Links relacionados

Compartilhar


Pesquisa Operacional

versão impressa ISSN 0101-7438versão On-line ISSN 1678-5142

Pesqui. Oper. vol.35 no.2 Rio de Janeiro maio/ago. 2015

http://dx.doi.org/10.1590/0101-7438.2015.035.02.0329 

Articles

CLASSIFICATION OF POWER QUALITY CONSIDERING VOLTAGE SAGS IN DISTRIBUTION SYSTEMS USING KDD PROCESS

Anderson Roges Teixeira Góes 1  

Maria Teresinha Arns Steiner 2   *

Rodrigo Antonio Peniche 3  

1Federal University of Paraná; Graduate Program in Numerical Methods in Engineering; Centro Politécnico, CP. 19081; 81531-990 Curitiba, Paraná, Brazil. E-mail: artgoes@ufpr.br

2Pontifical Catholic University of Paraná; Graduate Program in Industrial and Systems Engineering; Rua Imaculada Conceição, 1155, Prado Velho, 80215-901 Curitiba, Paraná, Brazil. E-mail: maria.steiner@pucpr.br

3Companhia Paranaense de Energia Elétrica (COPEL) Curitiba, Paraná, Brazil. E-mail: rodrigo.peniche@copel.com


ABSTRACT

In this paper, we propose a methodology to classify Power Quality (PQ) in distribution systems based on voltage sags. The methodology uses the KDD process (Knowledge Discovery in Databases) in order to establish a quality level to be printed in labels. The methodology was applied to feeders on a substation located in Curitiba, Paraná, Brazil, considering attributes such as sag length (remnant voltage), duration and frequency (number of occurrences on a given period of time). On the Data Mining Stage (the main stage on KDD Process), three different techniques were used, in a comparative way, for pattern recognition, in order to achieve the quality classification for the feeders: Artificial Neural Networks (ANN); Support Vector Machines (SVM) and Genetic Algorithms (GA). By printing a label with quality level information, utilities companies (power concessionaires) can get better organized for mitigation procedures by establishing clear targets. Moreover, the same way costumers already receive information regarding PQ based on interruptions, they will also be able to receive information based on voltage sags.

Key words: PQ classification; KDD Process; Artificial Neural Networks; Support Vector Machines; Genetic Algorithms; Quality Label

1 INTRODUCTION

For better service improvement, companies from different sectors of the economy have been utilizing high-sensibility computerized equipments in order to improve their services. These equipments require good Power Quality (PQ) in order to achieve perfect function. As a result, many studies related to PQ have been developed.

Costumers demand for both products and service quality, especially when they have support from the Federal or State governments Regulatory Agencies, have led utilities companies (power concessionaries) to look for solutions that may be able to satisfy their clients in what concerns PQ improvement and at the same time fulfill the requirements of the Brazilian Electricity Regulatory Agency (Agência Nacional de Energia Elétrica, ANEEL), created in Brazil in 1997.

Many are the disturbances that occur in electric system. These disturbances, usually called "events", can be either accidental (for example, tree branch fall; atmospheric discharges; and so on) or programmed (preventive maintenance) and they have direct influence on PQ. These events generate some PQ indicators or, also called, continuity indicators (both individual and collective), currently indicated by Brazilian utilities companies, which are related to PQ. The most outstanding indicators are (ANEEL, 2000): DIC index (Interruption Duration per Consumer Unit); FIC index (Interruption Frequency per Consumer Unit); DMIC (Maximum Interruption Duration per Consumer Unit); DEC index (Equivalent Interruption Duration per Consumer Unit) and FEC index (Equivalent Interruption Frequency per Consumer Unit). Among these indicators, the first three are individual indicators (per transformer) and the last two are regional indicators.

In the light of these considerations, the aim of this paper is to present a methodology that can be considered as an alternative to the fulfillment of what is proposed by ANEEL (2008). This agency has not yet defined performance standards regarding "voltage sag" events, but indicates that "concessionaires should follow up the performance of monitored bus bars and make it available on an annual basis". This information can be a reference for the bar performance on consumer units serviced by Medium and High Voltage Distribution System with sensitive loads and short-duration voltage variations.

This study proposes a PQ classification using only three attributes: voltage sag length, duration, and frequency (number of events during a certain period). As a response to the implement of this classification, a PQ label (or quality level to be printed in labels, or simply, quality label) is generated, classifying feeders according to a six-color scale, in which each color stands for a quality level (from "A" to "F", where "A" indicates the highest quality and "F" the lowest quality). This paper introduces an illustration of the methodology applied to feeders in a substation in the city of Curitiba, Parana, Brazil, which could be adopted and applied to other issues (Góes, 2012).

The inspiration to create a printed label for voltage sags, in order to indicate quality level, is a result of a literature review on studies performed by Casteren et al. (2005) and Cobben & Casteren (2006). These researchers outlined a PQ classification, without, however, presenting a methodology or technique to make the PQ effective for voltage sags. Moreover, they have not worked on applications using real data, instead, they use only dummy data.

Some questions can be listed in order to direct the present research: How to use real data in order to create a quality label? How to define what "regular quality" is, based on real data? How to classify an element/pattern that does not fit any of the classification levels in the quality label? In this study we introduce a methodology, as well as its application, that answer those questions. The methodology brings in its context the Knowledge Discovery in Databases (KDD).

In order to answer the first question, we used a historical database of a utility company from a substation stored within a period of four months (February to May 2008). The second question is answered by achieving the upper limit of the "C" range, which will be showed throughout this paper. Finally, in order to answer the third question we used three classic pattern recognition techniques: Artificial Neural Networks (ANN), Support Vector Machines (SVM) and Genetic Algorithm (GA) at the Data Mining stage (main stage of the KDD process).

This paper is organized in five sections, including this introduction. Section 2 discuss the real problem; section 3 presents a brief literature review, which indicates some studies that are related to this subject; the KDD process and the Data Mining techniques used. Section 4 shows the methodology applied to the real problem and some results and finally, in Section 5, the conclusions are presented showing the main contribution of this paper.

2 PROBLEM DESCRIPTION

Data were collected from a utility company in order to apply the developed methodology fulfilling the first aim of this research which consists in using real data to indicate the PQ. This company supplies 399 cities and 1,114 localities (districts, small towns and villages) in Parana State, Brazil and by the time of the survey it had 378 substations in order to supply around four million consumers (households, industries and others). There were 30 substations with around 300 feeders (approximately 10 feeders per substation) in the state capital city, specifically. A device is installed to detect PQ-events also measuring voltage sags in 14 of the 378 substations.Of these 14 devices, six devices are installed in substations in the capital city and its metropolitan area.

The methodology proposed in this study is applied to one of these substations composed of 12 feeders. However, it should be noted that this methodology can be applied to any substation as long as it has a data collector to capture the required information.

The historical records of events (voltage sags) required for the development of the proposal of this study are stored by the concessionaire in two data bases. In the first one, BD01, data are captured by the device installed in the bus bar of the substation. Each of these records contains 17 data (attributes) - "oscillographic identification", which consists in numbering the records using a software from the concessionaire; "date and timing of the event start", which indicates the initial time record of the PQ-related event; "type of event", which indicates the PQ-related phenomenon: voltage sag or voltage swell, among others; and "remnant voltage or root mean square (RMS)", which indicates the remnant voltage, in other words, the voltage "left" after the event occurrence at each of the voltage phases (Phase A, Phase B and Phase C).

The records from the second data base, BD02, were captured by the Distribution OperationSystem of the concessionaire and they are related to the interruptions. These records obtained by a software from the concessionaire provide 29 attributes among which there is "feeder identification" (name of the power grid feeder where the interruption was generated); "data and timing of event on set" (moment when the interruption occurred); "duration" (interruption duration); "type" (description of interruption type, which can be accidental, programmed or voluntary) and "affected component" (description of the affected electrical component).

The data used in the study were collected during a period of four months, between February and May 2008; BD01 was formed by 352 records (each one with 17 attributes) and BD02 was formed by 422 records (each one with 29 attributes). It is necessary to adopt a procedure in order to analyze and explore and transform this information in knowledge, thus, KDD process is used in order to explore the data that will ultimately produce the PQ label.

3 LITERATURE REVIEW

In this section, we discuss papers related to the subject here presented as well as the KDD process and an overview of techniques that were used in techniques for PQ label creation considering voltage sags.

3.1 Related Studies

There are many studies related to electrical grid disturbances (voltage sags, overvoltage, Total Harmonic Distortion, frequency, unbalanced circuits, among others) which involve many researches that have used Operations Research techniques in order to find the identification, location, classification and forecasting of these disturbances. Some of these researches are developed by Oleskovicz et al. (2006); Adepoju et al. (2007); Kaewarsa et al. (2008); Caciotta et al. (2009); Gencer et al. (2010); Kappor & Saini (2011) and Dash et al. (2012).

In a more detailed way, we can mention Saini & Kapoor (2012) who have conducted a review of 27 papers involving the study of signal processing (as Fourier and Wavelet transform based methods); intelligent techniques (as fuzzy logic, neural network, genetic algorithm and their fusion) and optimization techniques (Particle Swarm Optimization, Ant Colony Optimization) in PQ analysis. Tabular presentation of these 27 works has also been provided by the authors showing, mainly, the techniques applied and the data feature (synthetic or practical). Biswal et al. (2012) presented an approach for processing various non-stationary PQ waveforms through a Fast S-Transform with modified Gaussian window to generate time-frequency contours in order to extract relevant feature vectors for automatic disturbance pattern classification. The extracted features were clustered using Bacterial Foraging Optimization Algorithm (BFOA) based Fuzzy decision tree to give improved classification accuracy in comparison to the Fuzzy decision tree alone. The proposed technique was used to analyze the power signal disturbance in a realistic power network simulated by power system block set.

Morajev et al. (2009) presented a method of PQ classification using SVMs. The authors have got optimal features for the classifier two stage of feature selection. In first stage "Mutual Information Feature Selection" (MIFS; it reduce the dimensionality of inputs) and in the second stage "Correlation Feature Selection" (CFS; it can get optimal features) techniques were used for feature extraction from signals to build distinguished patterns for classifiers. In order to create training and testing vectors, different disturbance classes were simulated using parametric equations, i.e., pure sinusoid, sag, swell, harmonic, outage, sag and harmonics and swell and harmonics. The results showed that the classifier has an accuracy of 98.40%. Ozgonenel et al. (2013) presented the performance evaluation of SVM with "One against All" and different classification methods for PQ monitoring. The first aim was to conduct feature vector extraction and selection of PQ disturbances in a comparative way using EEMD ("Ensemble Empirical Mode Decomposition") and EMD ("Empirical Mode Decomposition") techniques. Feature vectors are extracted from the sampled power signals with the Hilbert Huang Transform (HHT) technique. The outputs of HHT are intrinsic mode functions (IMFs), instantaneous frequency (IF) and instantaneous amplitude (IA). Characteristic features are obtained from first IMFs, IF, and IA. These features were normalized along with the inputs of SVM and other classifiers.

All these papers deal with "Classification of Events for Power Quality". There is only one paper found at the literature (Casteren et al., 2005), which deals "Construction of a label for Classification of Power Quality". Casteren et al. (2005) were the first to introduce Figure 1, which was used latter by Cobben & Casteren (2006), in order to attempt to classify the voltage sags in order to indicate responsibility (consumer, equipment manufacturer or concessionaire) over the event and its mitigation measures examining the duration and the remaining value of such sags. With this data in hand Casteren et al. (2005) outlined a quality label according to frequency (number of occurrences) in order to classify voltage sags according to a table divided into nine different levels, as shown in Figure 2, grouped into three regions, where each region represents a responsibility area.

Figure 1 PQ label. Source: Casteren et al. (2005)

Figure 2 Responsibility over voltage sags (duration × remnant voltage). Source: Casteren et al. (2005)

The upper region (K0, M0, L0), with duration between 500 ms and 5 min and remnant voltage between 80% and 100%, is the manufacturer's responsibility. The intermediate region (K1, M1, L1), with analogous interpretation, is the consumer's responsibility area. Finally, the lower region (K2, M2, L2) is the concessionaire's responsibility. The authors do not report any detailed data from the sags, either measured or simulated. Therefore, the numbers reported in the present standard criteria are dummy data. Figure 3, as an example, indicates that a costumer may prove annually a maximum of five K1 sags, three M1 sags and two L1 sags; any value higher than these would result in penalties to the concessionaires. M2 sags are allowed only once every two years.

Figure 3 Example of criterion of characterization of voltage sags. Source: Casteren et al. (2005)

In order to facilitate the communication between consumers and concessionaires, the authors planned a classification label (quality label) of PQ based on the criteria of voltage sag characterization, as shown in Figure 4. According to this classification "A" stands for high PQ, and "E", for low PQ.

Figure 4 PQ label. Source: Casteren et al. (2005)

The classification of PQ shown in this Figure 4 must be linked to the criteria of characterization of voltage sags (Figure 3). Therefore, the authors used the upper "C" level limit criterion as shown in Figure 5. Additional criteria tables can be created analogously in order to define the upper A, B, C and D bounds.

Figure 5 PQ classification method (linking Figures 3 and 4). Source: Casteren et al. (2005)

This type of classification method is simple and consistent. It requires just some multiplication factors that are defined according to the criteria from the concessionaires. However some of the considerations made by the authors are not so obvious and it seems that there are no other studies in the literature that indicates the answers to the questions previously mentioned in Section 1: how to use real data to create a quality label? How to define what "regular quality" is, based on real data? How to classify an element that does not fit any classification range in the quality label such as K1, K2, M1 and L1, which belong to the values in the "B" range classification, although M2 and L2 have values in "D" range (Figure 6)?

Figure 6 Example of voltage sag events that do not fit in the framework of Figure 5. Source: Authors. 

Through Figures 5 and 6, one can observe that for the values of K1 = 2; M1 = 1.1; K2 = 0.25; M2 = 0.19, the classification of the element would be "B", but for the values of L1 = 2.3 and L2 = 0.29, the classification would be "E". In this way, we can observe that there is not a way, at this moment, to classify this element. So, in order to classify an element of this type, we are proposing the use of SVM classifier (presented the best performance among the three techniques researched: ANN, SVM and GA), as shown throughout the paper.

3.2 KDD Process

The methodology used in this study can be framed in the context of KDD process, which aim is to find information in databases through an automated search, thus establishing interest patterns that experts may fail to observe.

KDD process emerged in the beginning of the 1990s and it is strongly operating till nowadays. This statement can be confirmed through the studies of Han & Kamber (2006), Steiner et al. (2006), Atmami & Beldjilali (2007), Li & Kuo (2008), Bang et al. (2009), Liu et al. (2010), Bradwaj & Pal (2011), Mao et al. (2011), Arora & Bhatia (2012), Liu et al. (2012), Sassi (2012), among others.

The definition for KDD was first mentioned in a study performed by Frawley, Piatetsky-Shapiro & Matheus (1991), to which the authors define as "nontrivial extraction of implicit, previously unknown, and potentially useful information from data". Three years latter Brachman & Anand (1994) defined KDD as "knowledge-intensive task consisting of complex interactions, protracted over time, between a human and a (large) database, possibly supported by a heterogeneous suite of tools". However, the most common definition found in the literature is the one from Fayyad et al. (1995), in which KDD is "the non-trivial process of identifying valid, novel, potentially useful and ultimately understandable patterns in data".

Before these concepts KDD process was often mistaken with the concept of Data Mining, which, in fact, is, within the five stages of the KDD process, the main one. These stages shown in Figure 7 are: data selection; data preprocessing; data transformation; data mining and, finally, interpretation of the knowledge generated (Fayyad et al., 1995).

Figure 7 Stages in the KDD process. Source: Fayyad et al. (1995)

It is possible to find throughout this process both explicit and non-explicit knowledge, in other words, the information can be either known information or unexpected information which was not perceived before with database analysis. There can also be information with no significant relation, due to the lack of attributes or due to the inexistence of new knowledge to be discovered. Usually the information obtained is the ones that are not detected when traditional methods of data analysis are applied for decision making. That is because most of the traditional methods are capable of verifying only the explicit information in datasets.

Once the DM stage is the most important in the KDD process, the pursue of efficient methods is necessary as much as datasets grow in relation to the number of information that they may store, sometimes not making the relations between the various attributes explicit. Therefore, three techniques, that are standard in the literature, were used in this paper: Artificial Neural Networks or ANN (Mitchell, 1997; Haykin, 1999; Rodrigues et al., 2008; Setiono et al., 2009), Support Vector Machines or SVM (Vanipk, 1995, 1998; Burges, 1998; Ding & Dubchak, 2001; Sung & Mukkamala, 2003; Khemchandani et al., 2007) and Genetic Algorithms or GA (Goldberg, 1989; Holland, 1992; Angélico et al., 2013; Oliveira et al., 2013).

4 METHODOLOGY AND ITS APPLICATION

KDD process was used as the foundation of the methodology developed in this study in order to create the PQ that is basically composed of the following stages (adapted from Figure 7):

  • • Data Preprocessing (data cleaning and transformation);

  • • Association between databases (BD01 and BD02);

  • • Creation of the label itself. At this stage, Data Mining techniques (ANN; AG and SVM) were used in order to achieve pattern recognition, as already mentioned;

  • • Result Analysis.

4.1 Data Preprocessing

At this stage of the KDD process, the attributes described in section 2 were analyzed. From BD01, where each record has 17 attributes, eight were removed and nine, which were considered relevant, left; the number of attributes of BD02 was reduced from 29 to six.

During a meeting held with the concessionaire engineers, they informed that only the records with attributes in "Type" were "Accidental" should be considered once the other two "Types" ("Programmed" and "Voluntary") can have the PQ disturbances monitored. Once this information is exclusively for BD02, this database was filtered again and only five attributes were left, that way reducing the numbers of records from 422 to 181.

The transformation of attributes that indicate remnant voltage at each of the voltage phases was then performed in BD01. This transformation was called "aggregation of parameters", that is, the remnant voltage of the event was defined as the lowest value among the values achieved by the three voltage phases, which was an alternative indicated by ANEEL (2008). The event duration is defined as the maximum duration between the three phase/neutral events. These values were recorded in the new "Remnant voltage" attribute, and the "RMS voltage phase A"; "RMS voltage phase B" and "RMS voltage phase C" attributes were excluded from BD01.

The methodology proposed for the creation of a feeder PQ label considered only three attributes: remnant voltage sag; sag duration, and number of occurrences. The first two attributes are in BD01 (Table 1); the third attribute is the result of a simple occurrence counting. However, BD01 does not indicate the feeder that was affected by the event once the data concerning feeders are in BD02. Thus, it is necessary to associate BD01 records with BD02 records according to a procedure introduced in the next section.

Table 1 Some BD01 records after data transformation. Source: Authors. 

Id. Osc. Start Date Start Time Final Date Final Time Duration (s) Circuit RMS (%)
9 2008-02-06 07:28:35.034 2008-02-06 07:28:35.252 218 0 60.1
10 2008-02-06 20:04:14.805 2008-02-06 20:04:14.990 185 1 35.9
... ... ... ... ... ... ... ...

4.2 Data Association (BD01 and BD02)

Attributes related to "Time" were used in order to associate the date from BD01 and BD02. More specifically, "Start Date" and "Start Time" attributes in BD01 and "Start Date" and "Start Time" attributes in BD02 were used.

The engineers at the utility company defined the association or matching intervals for records in these databases in the following way: if a record in BD02 has the information "AR" (automatic restart) in the "Affected Component" attribute, a maximum time interval of 5 minutes in relation to a record in BD01 must be accepted, once databases record events (almost) immediately for this type of switch. In case the information is not an "AR", but instead, other types of switches, the interval could to be up to 2 hours. In these conditions, a matching was accepted and a record in BD03 with data was created.

This association generated a new data base called BD03, which contains 169 records. In other words, from the 352 records in BD01 and the 181 records in BD02 there are 169 associated records matching parameter "time" accordingly to the criterion above Table 2 shows in 10 columns some examples/records of this association. Information in columns 1 to 5 is data from BD01 while columns 6 to 10 are their respective associations found in BD02. In addition, in order to identify the 12 feeders in this substation, they will be generically called AA, AB, AC,..., AK, and AL.

Table 2 Examples of BD03 records (association between BD01and BD02). Source: Authors. 

Id. Osc. Start Date Start Time Duration (s) RMS (%) Feeder Component Affected Start Date Start Time Duration (s)
117 28/04/2008 14:28:43 185 46.3 AC Fly tap 28/04/2008 14:30 135
117 28/04/2008 14:28:43 185 46.3 AF AR performance 28/04/2008 14:29 1
117 28/04/2008 14:28:43 185 46.3 AF Fusible link performance 28/04/2008 14:42 41
121 28/04/2008 18:07:50 202 12.6 AC Conductor - AT 28/04/2008 18:32 344
121 28/04/2008 18:07:50 202 45.6 AI Conductor - BT 28/04/2008 18:16 403
136 02/05/2008 7:13:10 705 28.7 AF Pole 02/05/2008 07:15 36
139 08/05/2008 11:13:04 168 42.4 AI AR performance 08/05/2008 11:14 0

Table 2 shows that one record in BD01 may have more than one association with BD02 as it happens in the first three lines of the table where the "Oscillography Identification" attribute is 117. This indicates that the event captured in the substation was also "captured" or was originated in two feeders, "AC" and "AF"; where "AF" has two records for different affected components: "Fly Tap", "AR performance", or simply "AR" and "Fusible link performance".

4.3 Creating the PQ Label for the Feeders in a Substation

The classification for each BD03 record started with the construction of a classification table (Table 3) inspired by the proposal made by Casteren et al. (2005), as shown in Section 3.1, with the following attributes: remnant voltage sag; sag duration, and number of events. The division proposed for the Table 3 was as follows: two duration ranges were considered for the event: ≤ 500 and >500 milliseconds and five remnant voltage intervals: 10% to 19%; 20% to 39%; 40% to 59%; 60% to 79% and 80% to 90%.

Table 3 Classification considering sag duration and remnant voltage sag in the records. Source: Authors. 

RMS (%) Duration
≤ 500 miliseconds > 500 miliseconds
80 to 90% C1 C2
60 to 79% C3 C4
40 to 59% C5 C6
20 to 39% C7 C8
10 to 19% C9 C10

The connection between sag duration and remnant voltage sag can be better understood by observing Table 3, where 10 possible classes, C1, C2, ... to C10, are shown. It becomes evident that the shorter the duration, the higher the remnant voltage of the event, and the better weill bethe PQ classification for that event. Thus, the PQ of events has the following hierarchy: C1 ≥ C2 ≥ ... ≥ C10. In order to typify such classification, records in Table 2 are properly classified according to Table 3 and Table 4.

Duration (miliseconds) RMS (%) Feeder Record Classification
185 46.3 AC C5
185 46.3 AF C5
185 46.3 AF C5
202 42.6 AC C5
202 42.6 AI C5
705 28.7 AF C8

When performing this classification with the 169 records of BD03, it is assumed that the numbers of records from the "AA" feeder, for example, are the ones shown in Table 5. Record classification is achieved similarly for the other feeders in the substation. Table 5 shows that the "AA" feeder has two C5 events, one C7 and one C8. Table 6 considers all 169 records of all 12 feeders of the substation. It can be noted that only three of those ranges have records: C5, C7 and C8.

Table 5 Classification of voltage sags of the "AA" feeder. Source: Authors. 

RMS (%) Duration
≤ 500 miliseconds > 500 miliseconds
80 to 90% 0 0
60 to 79% 0 0
40 to 59% 2 0
20 to 39% 1 1
10 to 19% 0 0

Table 6 Classification of voltage sags in the analyzedsubstation, considering all the records. Source: Authors. 

RMS (%) Duration
≤ 500 miliseconds > 500 miliseconds
80 to 90% 0 0
60 to 79% 0 0
40 to 59% 149 0
20 to 39% 15 5
10 to 19% 0 0

In order to achieve the "average quality" of the substation under analysis, the number of events in Table 6 was divided by 12 (total feeders), obtaining the data of Table 7, already properly rounded.

Table 7 Average classification of voltage sags in the analyzed substation. Source: Authors. 

RMS (%) Duration
≤ 500 miliseconds > 500 miliseconds
80 to 90% 0 0
60 to 79% 0 0
40 to 59% 13 0
20 to 39% 2 1
10 to 19% 0 0

In order to create the PQ label, the values of six ranges were established, where "Range A" is the best PQ and "Range F" is the worst one. For each range, some factors (defined in conjunction with the concessionaire engineers) were multiplied in order to determine the upper limit of each range. Thus, Table 7 above represents the feeder average, that is, the upper limit of "Range C". The upper bound of "Range A" (Table 8) was obtained by multiplying values in Table 7 by 0.25.

Table 8 Upper bound of "Range A" in the PQ classification label of a feeder in a specific substation. Source: Authors. 

RMS (%) Duration
≤ 500 miliseconds > 500 miliseconds
80 to 90% 0 0
60 to 79% 0 0
40 to 59% 4 0
20 to 39% 1 1
10 to 19% 0 0

The upper bound of "Range B" was obtained by multiplying the values in Table 7 by 0.50. The upper bound of "Range D" was obtained by multiplying the values in Table 7 by 1.5. The upper bound of "Range E" was obtained by multiplying the values in Table 7 by a 2.0 factor. Finally, the upper bound of "Range F" was obtained by verifying the highest value set for the feeders under analysis. These values are shown in the PQ classification label (Figure 8).

Figure 8 PQ classification label of feeders of a specific substation. Source: Authors. 

Once the parameters for the label have been established, defining the lower and upper bound of occurrences for RMS intervals on each category, the classification of each feeder only requires verifying which range interval shown in Figure 8 the feeder fits in. However, it becomes evident that such task is not so simple once only five of the 12 feeders fit exactly in these ranges of values, all of them with "A" quality classification. The feeders are: AA, AG, AJ, AK and AL. The five feeders show values for C5 belonging to the [0,4] interval and for C7 and C8, in the [0,1] interval, as illustrated in Figure 9.

Figure 9 Feeders classified directly from the PQ label. Source: Authors. 

Other feeders could not be directly classified. In feeder AH, for example, the value of C5 equals 16, which indicates that its classification would be D. However, in this feeder, C7 and C8 are outside D class intervals. For that reason, in order to classify the other feeders, we used three Data Mining techniques, already briefly described in Sections 3.2.1 to 3.2.3.

4.3.1 Data Mining Techniques Application

The DM stage is the most important stage in the KDD process once it is the moment when pattern recognition techniques are applied either through heuristic or through metaheuristic procedures. In this study, such procedures are applied aiming the PQ classification of feeders in a substation.

The techniques, used in order to classify the feeders that could not be directly classified, are the ANN, SVM, and GA according to their performance in solving the problem discussed in this paper. The specific features of these techniques are approached in A, B and C items below. Next their general features are described.

Cross-validation procedure was used in order to evaluate all the techniques, more precisely the stratified three-fold. The dataset was divided into two subsets: 2/3 for the training phase and 1/3 for testing phase. Once there is six classification ranges ("A" to "F"), each one with three RMS levels, 30 dummy records were created per range obeying the limits of each range (Figure 8), thus obtaining a total of 180 records from which 120 were used for training and 60 for testing. Since it is stratified, each set (training and testing) was formed by classes ("A" to "F") with the same number of elements.

In addition, the training of each of the techniques occurred five times (phases), one training per class in order to identify whether the record belongs to a certain class (or not). Therefore, the training for class "A" was performed leading the technique to "learn" what is (and what is not a record from class "A", i.e., B, C, D, E and F). Then, another training was performed for class "B", removing the data from set "A", and leading the technique to "learn" what is a record from class "B" (and what is not "B", i.e., C, D, E, and F) and so on for classes "C", "D" and "E". When the last training was done (class "E"), if a record is not classified as "E", it automatically becomes "F" (last class). When a new record needs to be classified, this record must be "introduced" to all the stages of the technique and its classification will be done according to the corresponding one. Following the techniques and its specific features are approached.

A. ARTIFICIAL NEURAL NETWORKS

A back-propagation learning algorithm was used in the application of the ANN. It was implemented in Visual Basic 6.0. Each trained ANN had three inputs (C5, C7 and C8) for each input layer; hidden layers (with number of neurons varying between "1" and "20") and one neuron in the output layer (that will indicate the class). For all of them (hidden and output layers) the sigmoid-logistic function was used as the activation function.

The network was trained five times; the initial weight set varied randomly in the interval (-1,1). In total, there were 1,500 tests (three stages of the three-fold method × five classification phases - "A" to "F" - × five tests with initial weights × variation of neurons in the hidden layer 1-20). The training was done once one of the three following condition was met: 1,000 iterations; mean square error less than or equal to 10-4; or number of records incorrectly classifies as equal to zero.

Concerning the problem discussed in this paper, the percentage of accuracy in the training of this technique was of 99.88% considering the three stages, and 99.67% in the test. Table 9 shows the result for the feeders classification obtained with the application of this technique. In Table 9, as well as in the other tables to be introduced, the "Voting Classification" column (last column) indicates the classification with the highest occurrence in the former columns, in other words, the statistical mode.

Table 9 Feeder Classification Results: ANN technique. Source: Authors. 

Feeder Stratified Three-fold Procedure
1st stage 2nd stage 3rd stage Voting Classification
AA A A A A
AB C C C C
AC D D D D
AD C C C C
AE B B B B
AF F F F F
AG A A A A
AH A A A A
AI E E E E
AJ A A A A
AK A A A A
AL A A A A

Even though the AA, AG, AJ, AK and AL feeders already have their classification defined once they were directly classified in the quality label, they were also introduced to the networks, that way confirming their classification. Therefore, six feeders have "A" classification, one feeder, "B" classification, two feeders, "C" classification, one feeder "D" classification, one feeder, "E" classification, and one feeder, "F" classification.

B. SUPPORT VECTOR MACHINES

The SVM technique searches for a plane that has the same distance for the elements of both classes, thus controlling the capacity of the decision function in the search. The SVM uses the Kernel function (function in the algorithm that calculates the classification function) for sets in which the data are not linearly separable. It is done in order to project this data in a space, called "features space", where it is possible to separate them linearly through one extra dimension. Therefore, even though it is not linearly separable in the pattern input space, it is in the feature space, as illustrated in Figure 10 (Vanipk, 1995, 1998; Burges, 1998).

Figure 10 Input space and feature space. Source: Góes, 2012

For the application of the SVM technique, it was used the svmtrain function of the MATLAB 7.9.0 software with the parameters: "kernel_function: linear"; "optimization method: Sequential Minimal Optimization"; "tolerance to the training method: 10-3"; "parameters of the multilayer perceptron kernel: [-1,1]".

Also, it was used with two matrices in the arguments: "Examples" and "Answers", according to equation (1). The "Examples" matrix has the three inputs Ci values (C5, C7 and C8) in its columns, and the "Answer" matrix has only one column with the range value that each pattern (line of the "Examples" matrix) has as its answer (classes "A" to "F").

Next, a set of test, described here in the form of a matrix "NewExamples", and the result of "Training" with the svmclassify function, as shown in equation (2), were used in order to verify the percentage of correct classifications of these new data.

It should be noted that the arguments used in the training for the svmtrain function are thedefaults on MATLAB 7.9.0 once the sets of ranges on the quality label are linearly separable by a plane. A total of 15 tests (3 training stages × 5 classification phases) were performed in this technique. The accuracy percentage in the technique training was of 100%, considering the three stages of the three-fold procedure and 99.55% in the test. Table 10 shows the results of the classifications achieved with the SVM application.

Table 10 Feeder classification results: SVM technique. Source: Authors. 

Feeder Stratified Three-fold Procedure
1st stage 2nd stage 3rd stage Voting Classification
AA A A A A
AB C C C C
AC D D D D
AD C C C C
AE A A A A
AF F F F F
AG A A A A
AH A A A A
AI D E E E
AJ A A A A
AK A A A A
AL A A A A

Table 10 indicates that the feeders voting classification is seven with "A" classification, none with "B" classification, one with "E" classification, and one with "F" classification. Feeders AA, AG, AJ, AK and AL, which had already been classified in the quality label, were also classified correctly with this technique.

C. GENETIC ALGORITHM

The GA metaheuristic was used in order to determine a plane so that in each of the semi-spaces determined by it may have only one of the sets of each stage of the application, according to the features described in the beginning of Section 4.3.1. We have to emphasize that the training sets, in the way we defined them, are linearly separable.

The value of the fitness function was established by an algorithm that determines three points that will define such plane, in which the coordinates of each point are the alleles of the individuals. Additionally, the fitness function looks for a plane equidistant to both training sets, considering the differences of the distances between two points (in different sets) that are closer to the established plane. The greater the difference between the distances, the greater the penalty applied to fitness will be.

Figure 11 shows this algorithm for the fitness calculation, where X is a vector in which each coordinate represents an allele of the individual of the population composed of nine alleles with values belonging to the set of real numbers. Thus, the three first alleles represent the coordinates of a point P 1, the next three alleles are the coordinates of point P 2, and the last three are the coordinates of point P 3; CL 1 is the data set "class 01" (for example, "A") and CL 2 is the data set "class 02" (not "A"); and k is an element that belongs to CL 1CL 2; EP(α) is the equation of the plane defined by P 1, P 2 and P 3. In this way:

  1. i) if kCL 1 then k must belong to the subspace below the plane α, and thus EP(k) should have a negative value;

  2. ii) if kCL 2 then k must belong to the upper subspace the plane α, and thus EP(k) should have a positive value;

  3. iii) Dist 1 and Dist 2 are initialized with high values so that algorithm determines if the plane is equidistant, or nearly equidistant, the two training sets (CL 1 and CL 2).

Figure 11 Pseudocode for fitness calculation. Source: Authors. 

In Figure 11, for each individual of the GA population, it is defined a plane (P 1, P 2, P 3) and it is verified to which subspace (determined by plane α) each element k of the training set belong (using EP(α)); after that, it is calculated the distance between k and the plane α. The value of the fitness function is determined by the difference of z 1 (number of elements correctly classified) and z 2 (penalty which that seeks to determine "how much" this plane is equidistant from the training sets).

For the application of the GA technique, it was used the gatool of the MATLAB 7.9.0 software with "penality" of 0.1. The arguments for training which achieved the best results were the following: "population type: double vector"; "population size: 20"; "fitness scaling: rank"; "selection function: stochastic uniform"; "crossover fraction: 0.8"; "crossover function: scattered"; "stopping criteria (generations): 100"; "stopping criteria (stall generations): 50"; and "stopping criteria (function tolerance): 10-6". It was used three stopping criteria in such a way that when a first one was reached, the procedure was finished.

It is worthwhile to remember that the "crossover scattered" works in the following way: the default crossover function creates a random binary vector and selects the genes where the vector is a "1" from the first parent, and the genes where the vector is a "0" from the second parent, and combines the genes to form the child. For example, if p 1 and p 2 are the parents: p 1 = [abcdefgh]; p 2 = [12345678] and the binary vector is [11001000], the function returns the following child [ab34e678].

It was performed a total of 45 tests (three stages of the three-fold method × five classification phases - "A" to "F" - × three tests with different populations) with the parameters described above. The percentage of correctness in the training was of 100% considering the three stages of the three-fold method, and 99.11% in the test. Table 11 shows the results of the classification of the feeders achieved through the application of this technique. The feeders AA, AG, AJ, AK and AL also confirm the classification previously achieved. Thus, there are six feeders with "A" classification, one feeder with "B" classification, four feeders with "C" classification, none with "D" classification, none with "E" classification, and one with "F" classification.

Table 11 Feeder Classification Results: GA technique. Source: Authors. 

Feeder Three-fold Procedure
1st stage 2nd stage 3rd stage Voting Classification
AA A A A A
AB E C C C
AC E C C C
AD B B C B
AE A A A A
AF F F F F
AG A A A A
AH E C C C
AI E C C C
AJ A A A A
AK A A A A
AL A A A A

4.4 Result Analysis

The analysis of the results is the last stage of the KDD process and it is performed in this study by comparing the classifications achieved in the three techniques that were applied. Table 12 shows the result of the achieved classification (column "voting classification" of Tables 9 to 11). In addition, this Table 12 also shows a column named "voting classification" that indicates the result of higher occurrence among the techniques, which is assumed in this analysis to be the most adequate one to solving the problem.

Table 12 Comparison between classifications achieved through ANN, SVM and AG techniques. Source: Authors. 

Feeder ANN SVM AG Voting Classification
AA A A A A
AB C C C C
AC D C C C
AD C C B C
AE B A A A
AF F F F F
AG A A A A
AH A A C A
AI E E C E
AJ A A A A
AK A A A A
AL A A A A

An analysis of Table 12 indicates that seven of the 12 feeders (AA, AB, AF, AG, AJ, AK and AL) are equally classified in all the techniques. The others achieved the same classification only in two techniques: two of them (AC and AE), with the same classification in SVM and GA techniques and three of them (AD, AH and AI), with the same classification in the ANN and SVM techniques.

It is possible to note when comparing each technique with the classification assumed to be adequate ("voting classification" column), the GA technique shows three feeders (AD, AH, AI) with distinct classification. Two of them are in non-neighbor ranges. According to this technique the feeder AD has "C" classification when the proper classifications is "A", and the feeder AI was classified with "C" through the GA when "E" is the proper one. For the latter, the result achieved in GA is very distant from that of the other two techniques that have the same results as the proper classification.

In the classification introduced by the ANN technique there are two feeders (AC and AE) that have different classification from the one in "voting classification", however, in neighbor ranges. The proper classification for the AC feeder is "C", but the ANN classified it as "D". The AE feeder was classified by the ANN as "B" when the proper classification is "A". Finally, SVM technique shows results that are identical to the ones from the "voting classification" column, which makes it the most adequate technique for this study.

The proper feeder classification resulted in seven feeders with "A" classification, none with "B" classification, three with "C" classification, none with "D" classification, one with "E" classification and one with "F" classification. (Figure 12).

Figure 12 Quality label with feeder classification. Source: Authors. 

The values shown for each feeder in Figure 12 express events occurrences in each of the C5, C7 and C8, classes. The label shows non-explicit knowledge when it analyses theses values, as for example, the classification of feeders AI and AF, that is because values of C5, C7 in AI are greater than the ones in AF. This could indicate lower quality to AI when compared to AF, but since C8 has lower values for AI, the technique indicates that AF quality is lower than AI. Thus, the methodology developed and applied in this study reveals non-explicit knowledge in the concessionaire database to an unprecedented real problem: the PQ considering voltage sags.

5 CONCLUSIONS

In this work PQ classification for feeders is proposed as recommended by ANEEL. The classification is assigned based on voltage sag frequencies of 10 different combinations of remnant voltage and duration. Six quality levels have been set (from "A" to "F"), and for each category there is an upper and a lower frequency for each of the 10 combinations presented on Table 3.

As there are frequencies for 10 different combinations, in several cases frequencies for a given feeder would not fit into the range for a single classification requiring additional analysis for the correct classification. For this task, based on KDD process the search for alternatives, RNA, SVM and GA metaheuristics have been used and results are compared, the best results obtained through SVM technique. The final result is an individual classification for each feeder and inscribed into a label. This label would inform PQ quality for each feeder, helping for a faster and more precise management.

Methodology has been applied for a substation unit equipped with 12 feeders. Although this number could be adequate for the study conducted here, it is strongly suggested the gathering of new information, from other sites before implementing the final classification parameters and printed labels issued.

From the results achieved in the classification of PQ considering voltage sags in distribution systems, the concessionaire will be able to analyze their performance, identifying low quality feeders and applying mitigation measures so that these feeders have their classification "leveraged". Therefore, this study may be continued through the identification, for example, for the reason why AF feeder has an "F" quality.

The quality label may be alternative to the ANEEL demand should this methodology be applied to annual data of Power distribution substations. As already mentioned in the introduction of this study, ANEEL's demand does not define performance patterns concerning voltage sags, but instead, it affirms that "concessionaires should follow up the performance of monitored bus bars and make them available on an annual basis". And this information may be used as a reference to consumer units being serviced by the Distribution System.

Concerning the techniques, the SVM showed the most adequate results when compared to the "voting classification", followed by ANN in the search for the classification of the feeders that were not directly classified in the quality label and which also classified correctly previously classified feeders. The versatility of the approached methodology allows its application in problems of various fields of knowledge through many different techniques. Finally, there could be a comparison of the techniques used with others like, for example, the ones presented in Kuncheva (2004), in order to maximize the accuracy to the methodology here proposed.

6 ACKNOWLEDGEMENT

This study is an integral part of the P&D project number 2866-019/2007 - Event Classification for Power Quality, approved by ANEEL and developed in partnership with COPEL (Companhia Paranense de Energia Elétrica) and UFPR (Universidade Federal do Paraná).

REFERENCES

ADEPOJU GA, OGUNJUYIGBE SOA & ALAWODE KO. 2007. Application of Neural Network to Load Forecasting in Nigerian Electrical Power System. The Pacific Jounal of Science and Technology, 8(1): 68-72. [ Links ]

ANEEL - AGÊNCIA NACIONAL DE ENERGIA ELÉTRICA. 2008. Procedimentos de Distribuição de Energia Elétrica no Sistema Elétrico Nacional - PRODIST: Módulo 8 - Qualidade da Energia Elétrica. [ Links ]

ANEEL - AGÊNCIA NACIONAL DE ENERGIA ELÉTRICA. 2000. Resolução no. 24, de 27 de janeiro de 2000. [ Links ]

ANGÉLICO BA, MENDONÇA M, ABRÃO T & ARRUDA LV. 2013. A Comparative Analysis of three Metaheuristic Methods applied to Fuzzy Cognitive Maps Learning. Pesquisa Operacional, 33(3): 443-465. [ Links ]

ARORA R & BHATIA R. 2012. Optimization of Automation in Fuzzy Decision Rules. Second International Conference on Advanced Computing & Communication Technologies. Haryana, Índia, p. 41-45. [ Links ]

ATMANI B & BELDJILALI B. 2007. Knowledge Discovery in Data base: Induction Graph and Cellular Automaton. Computing and Informatics, 26(1): 171-197. [ Links ]

BANG J, DHOLAKIA N, HAMEL L & SHIN S. 2009. Customer Relationship Management and Knowledge Discovery in Databases. Encyclopedia of Information Science and Technology. 2 ed., p. 902-907. [ Links ]

BISWAL B, BEHERA HS, BISOI R & DASH PK. 2012. Classification of power quality data using decision tree and chemotactic differential evolution based fuzzy clustering. Swarm and Evolutionary Computation, 4: 12-24. [ Links ]

BRACHMAN RJ & ANAND T. 1994. The Process of Knowledge Discovery in Data bases: A First Sketch. KDD Workshop, Seattle, Washington, USA, 1-12. [ Links ]

BRADWAJ BK & PAL S. 2011. Mining Educational Data to Analyze Students'Performance. International Jounal of Advanced Computer Science and Applications, 2(6): 63-69. [ Links ]

BURGES CJC. 1998. A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery, 2(1): 121-168. [ Links ]

CACIOTTA M, GIARNETTI S & LECCESE F. 2009. Hybrid Neural Network System for Electric Load Forecasting of Telecomunication Station. XIX IMEKO World Congress - Fundamental and Applied Metrology. Lisboa, Portugal, p. 657-661. [ Links ]

CASTEREN JFLV, ENSLIN LHR, HULSHORST WTJ, KILNG WL, HAMOEN MD & COBBEN JFG. 2005. A customer oriented approach to the classification of voltage dips. The 18th International Conference and exhibition on Electricity Distribution - CIRED, Turin, Italy, 1-6. [ Links ]

COBBEN JFG & CASTEREN JFL. 2006. Classification Methodologies for Power Quality. Electrical Power Quality & Utilization Magazine, 2(1): 11-17. [ Links ]

DASH PK, PADHEE M & BARIK SK. 2012. Estimation of power quality indices in distributed generation systems during power islanding conditions. Electrical Power and Energy Systems, 36(1): 18-30. [ Links ]

DING CH & DUBCHAK I. 2001. Multi-class protein fold recognition using support vector machines and neural networks. Bioinformatics, 17(4): 349-358. [ Links ]

FAYYAD U, PIATETSKY-SHAPIRO G, SMYTH P & UTHURUSAMY R. 1995. Advances in Knowledge Discovery & Data Mining. 1st. ed. American Association for Artificial Intelligence, Menlo Park, California. [ Links ]

FRAWLEY WJ, PIATETSKY-SHAPIRO G& MATHEUS CJ. 1991. Knowledge Discovery in Data bases - An Overview. Knowledge Discovery in Data bases, 13(3): 1-30. [ Links ]

GENCER O, OZTURK S & ERFIDAN T. 2010. A new approach to voltage sag detection based on wavelet transform. Electrical Power and Energy Systems, 32(2): 133-140. [ Links ]

GÓES ART. 2012. Uma metodologia para a criação de etiqueta de qualidade no contexto de Des coberta de Conhecimento em Bases de Dados: aplicação nas áreas elétrica e educacional.145 pages. Thesis in Numerical Methods in Engineerin, Federal University of Parana, Curitiba, Paraná, Brazil. [ Links ]

GOLDBERG DE. 1989. Genetic algorithms in search, optmization, and machines learning. Addison-Wesley Publishing Company, Inc. Massachusetts. [ Links ]

HAN J & KAMBER M. 2006. Data Mining: Concepts and Techniques. 2nd ed. Morgan Kauffmann Publishers. [ Links ]

HAYKIN S. 1999. Neural Networks - A Comprehensive Foundation. 2nd ed. Prentice Hall, New Jersey. [ Links ]

HOLLAND JH. 1992. Adaptation in natural and artificial systems. 2nd ed. Cambridge, USA: MIT Press. [ Links ]

KAEWARSA S, ATTAKITMONGCOL K & KULWORAWANICHPONG T. 2008. Recognition of power quality events by using multiwavelet-based neural networks. Electrical Power and Energy Systems, 30(4): 245-260. [ Links ]

KAPPOR R & SAINI MK. 2011. Hybrid demodulation concept and harmonic analysis for single/multiple power quality events detection and classification. Electrical Power and Energy Systems, 33(10): 1608-1622. [ Links ]

KHEMCHANDANI R, JAYADEVA & CHANDRA S. 2009. Knowledge based proximal support vector machines. European Journal of Operational Research, 195(3): 914-923. [ Links ]

KUNCHEVA LI. 2004. Combining Pattern Classifiers: Methods and Algorithms. John Wiley and Sons, Inc., New York. [ Links ]

LI S & KUO S. 2008. Knowledge discovery in financial investment for forecasting and trading strategy through wavelet-based SOM networks. Expert Systems with Applications, 34(2): 935-951. [ Links ]

LIU L, QI H & LI D. 2012. A research for Data Mining Technology based in Fuzzy Neural Network. Advanced Materials Research, 433-440: 2509-2512. [ Links ]

LIU S, TIAN X & ZHANG Z. 2010. Process planning knowledge discovery in the process data base. In: International Conference on Computer Application and System Modeling, Taiyuan, 11(1): 370-373. [ Links ]

MAO S, WAN W, WANG Y, WANG Z & YU H. 2011. The application of an improved BP artificial neural network in distributed data mining. IET International Conference on Smart and Sustainable City, Shanghai, China. [ Links ]

MCCULLOCH WS & PITTS W. 1943. A logical calculus of the ideas immanent in nervous activity. Bulletin Mathematical Biology, 5(4): 115-133. [ Links ]

MITCHELL T. 1997. Machines Learning. McGraw Hill. [ Links ]

MORAVEJ Z, BANIHASHEMI SA & VELAYATI MH. 2009. Power quality events classification and recognition using a novel support vector algorithm. Energy Conversion and Management, 50(12): 3071-3077. [ Links ]

OLESKOVICZ M, COURY DV, CARNEIRO AAFM, ARRUDA EF, DELMONT O & SOUZA SA. 2006. Estudo comparativo de ferramentas modeANNs de análise aplicadas à qualidade da energia elétrica. Revista Controle & Automação, 17(3): 331-341. [ Links ]

OLIVEIRA ACM, CHAVES AA & LORENA LAN. 2013. Clustering Search. Pesquisa Operacional, 33(1): 105-121. [ Links ]

OZGONENEL O, YALCIN T, GUNEY I & KURT U. 2013. A new classification for power quality events in distribution systems. Electric Power Systems Research, 9: 192-199. [ Links ]

RODRIGUES TB, MACRINI JLR & MONTEIRO EC. 2008. Seleção de variáveis e classificação de padrões por redes neurais como auxìlio ao diagnóstico de cardiopatia isquêmica. Pesquisa Operacional, 28(2): 285-302. [ Links ]

SAINI MK & KAPOOR R. 2012. Classification of power quality events - A review. Electrical Power and Energy Systems, 43(1): 11-19. [ Links ]

SASSI RJ. 2012. An hybrid architecture for clusters analysis: rough sets theory and self-organizing map artificial neural network. Pesquisa Operacional, 32(1): 139-164. [ Links ]

SETIONO R, BAESENS B & MUES C. 2009. A note on knowledge Discovery using neural networks and its application to credit card screening. European Journal of Operational Research, 192(1): 326-332. [ Links ]

STEINER MTA, SOMA NY, SHIMIZU T, NIEVOLA JC & STEINER NETO PJ. 2006. Abordagem de um problema médico por meio do processo de KDD com ênfase à análise exploratória dos dados. Revista Gestão & Produção. 13(2): 325-337. [ Links ]

SUNG AH & MUKKAMALA S. 2003. Identifying important features for intrusion detection using support vector machines and neural networks. Symposium on Applications and the Internet, p. 209-216. [ Links ]

VAPNIK V. 1998. Statistical Learning Theory. John Wiley & Sons, Inc., New York. [ Links ]

VAPNIK V. 1995. The Nature of Statistical Learning Theory. Springer-Verlag, New York. [ Links ]

Received: April 28, 2013; Accepted: August 28, 2014

*

Corresponding author.

Creative Commons License This is an open-access article distributed under the terms of the Creative Commons Attribution License