Professional Documents
Culture Documents
3, JUNE 2015
Abstract—A new high-impedance fault (HIF) detection method than the load current level. Since it is difficult to detect HIFs,
using time-frequency analysis for feature extraction is proposed. the exposed energized conductor may result in fire, a tremen-
A pattern classifier is trained whose feature set consists of current dous release of energy, and a threat to human life [1]. Although
waveform energy and normalized joint time–frequency moments.
The proposed method shows high efficacy in all of the detec- reports state 5% to 10% of all distribution network faults are
tion criteria defined in this paper. The method is verified using HIF [2], the actual percentage has to be higher, since only the
real-world data, acquired from HIF tests on three different mate- HIFs ending in bolted faults are included in the event recorder
rials (concrete, grass, and tree branch) and under two different statistics.
conditions (wet and dry). Several nonfault events, which often The nature and physical properties of the HIF are well docu-
confuse HIF detection systems, were simulated, such as capacitor
switching, transformer inrush current, nonlinear loads, and mented [3]. A comprehensive survey of HIF detection methods
power-electronics sources. A new set of criteria for fault detection is presented in [4]. The low-current magnitude, associated arc,
is proposed. Using these criteria, the proposed method is evaluated and current asymmetry are the dominant characteristics of these
and its performance is compared with the existing methods. These faults. Although the current magnitude of HIFs falls into non-
criteria are accuracy, dependability, security, safety, sensibility, detection zones of network protection systems, their nonsta-
cost, objectivity, completeness, and speed. The proposed method
is compared with the existing methods, and it is shown to be tionarity and random behavior have shown promising results in
more reliable and efficient than its existing counterparts. The defining signatures for HIF [5]. However, a method that can per-
effect of choice of the pattern classifier on method efficacy is also fectly detect HIF in different conditions and networks has not
investigated. been introduced yet.
Index Terms—High-impedance fault (HIF), power distribution HIF detection methods can be classified into four classes
faults, principal component analysis, protection, statistical joint based on the analysis domain they exploit to define HIF sig-
moment, time-frequency analysis. nature. These analysis domains are time-domain analysis,
frequency-domain analysis, time-scale analysis, and time-fre-
quency analysis [4]. The time-domain analysis emphasizes
I. INTRODUCTION irregularity of the HIF waveform. Methods based on the power
dissipation factor [6] and fractal techniques [7] are considered
simple time-domain-based methods. In [8] and [9], Sarlak et al.
0885-8977 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
GHADERI et al.: HIF DETECTION IN THE DISTRIBUTION NETWORK USING THE TIME-FREQUENCY-BASED ALGORITHM 1261
high-frequency support, subjectivity to the choice of the mother The desirable properties of TFD and their constraints are de-
wavelet, and its loss of feature resolution (less interpretable). fined as follows:
Time–frequency analysis (TFA) is the fourth set of analysis 1) Time Marginal: Integration of the TFD over frequency
domain. TFA has shown high efficacy for identifying discon- gives the “Instantaneous Power” ( )
tinuities, repeating patterns, and nonstationarity aspects of sig-
nals [17][18]. TFA has been successfully applied to different (3)
power system applications, such as power-quality (PQ) assess-
ment [19][20], and direction finding for capacitor switching dis-
turbances [21]. Because of the advantages that TFA presents 2) Frequency Marginal: Integration of the TFD over time
over other methods, namely, time and frequency localization, gives the “Energy Spectrum” ( )
objectivity to the signal choice, coherent time-frequency sup-
port, and interpretable features, an HIF detection method uti- (4)
lizing TFA for feature extraction can potentially reveal hidden
information of the HIF waveforms.
In this paper, a novel method based on TFA and joint moment 3) Global Energy: Integration of the TFD over the entire
calculation is proposed for HIF detection. A new set of criteria time–frequency plane yields the “Signal Energy” ( )
for the evaluation of the fault detection algorithms is defined.
The detection method gives comparable results with the existing (5)
methods in accuracy and security, while showing improvement
in the rest of the criteria. The extracted features based on TFA
are interpretable and objective, and their mathematical descrip- 4) Reduced Interference: Due to the bilinearity of the IAC,
tion based on the joint moment calculation is complete. introducing “artifacts” (interference) is inevitable in generating
TFDs. If the kernel has a low-pass filter characteristic in the
II. THEORETICAL FOUNDATION OF THE METHOD Doppler-lag domain , the interference could be reduced.
Since the disturbances in power systems are characterized by
A. Time–Frequency Analysis the presence of multiple frequency components over a short du-
ration, keeping high time–frequency resolution, while avoiding
Any protection technique can be projected to a pattern classi-
artifacts is of great significance in their analysis [21].
fication task [16], which consists of three steps, that is: 1) mea-
Although WVD satisfies the first three constraints (
surement; 2) feature extraction; and 3) classifier. In order to de-
), a large proportion of interference could result in a poor in-
tect HIF, it is necessary to develop a more precise feature ex-
terpretation of a signal.
traction, and classification technique.
In 1989, Choi and Williams [25] proposed a Gaussian kernel,
In the method presented in this paper, TFA is deployed as the
called Choi-Williams distribution (CWD) that satisfies all four
method for feature extraction. TFA was motivated by the need
constraints. Therefore, the CWD is chosen as the TFD in this
to describe nonstationary signals, where the Fourier transform
paper. This 2-D distribution of the nonstationary signal reveals
proves ineffective. Several time–frequency representation
hidden time–frequency behavior, which later will be mathemati-
methods are designed to mathematically describe these signals,
cally described using statistical joint moments. The CWD kernel
namely, short-time Fourier transform (STFT), Wigner–Ville
is defined as follows:
distribution (WVD), and Choi–Williams distribution (CWD)
[22][23]. To systematically design proper time–frequency dis- (6)
tribution (TFD), Cohen's generalization of the quadratic TFDs
is used [22]. Cohen proved that one can relate the desirable where is a parameter that is handled to achieve proper char-
properties of TFDs to constraints on its kernel acteristics of the distribution.
A. High-Impedance Fault
(1) In order to collect data for HIF, a test was conducted in the
high-current research laboratory. The test circuit is shown in
where is a 2-D function (in Doppler-lag domain), called Fig. 1. Three different materials (tree branch, grass surface, and
kernel. is the time–frequency distribution of the signal. concrete surface) in two different conditions (wet and dry) were
The constant time cross section of time–frequency distribution tested, and their current and voltage magnitude were recorded
( ) represents the frequencies available at time with a digital data recorder (sampled at 20 kHz). The voltage
, and its frequency cross section ( ) represents level of 7.2-kV line to ground is chosen for the test, since this
the times when the frequency occurred. And is the voltage level is one of the dominant voltage levels in distribution
instantaneous auto-correlation of the signal , respectively. networks. The circuit specifications are explained in Table I.
is defined as A high-speed camera (with time scale) was used to
the detailed behavior of arc initialization, tracking, extinction,
(2) and possible explosion. Since the HIF test is not an experiment
1262 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 30, NO. 3, JUNE 2015
TABLE I
TEST CIRCUIT SPECIFICATION
Fig. 3. Current waveform, normalized SIDs, and their TFDs for different events in the distribution network: (a) high-impedance fault, (b) capacitor switching, (c)
transformer inrush current, and (d) harmonic load.
(7)
(8)
Step 2: Time–Frequency Distribution: Next, time–frequency where and are non-negative integers. Also, is the signal
distributions are generated from the quarter-cycle SID signals. energy defined in (5), which is used to normalize the moments.
1264 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 30, NO. 3, JUNE 2015
, and are called mean time, and mean frequency, and systems event provides distinguishable 2-D signatures, which
are formulated as follows: are mathematically described in the feature set of (14).
Step 4: Principal Component Analysis (PCA): The objective
of PCA is to find a set of orthogonal components that minimizes
the mean square error of the reconstructed data and represents
(9) the original data with fewer components, which reduces the di-
mension of the data. The ratio of an eigenvalue ( ) to the sum
of eigenvalues expresses the loss of information (LOI) due to
(10) the elimination of the th eigenvector, as given in (15). In other
words, the algorithm keeps the features that more strongly rep-
resent the signatures of different events
It should be noted that the signal time duration ( ) and signal
frequency bandwidth ( ) can be found with the calculation of (15)
( ), and ( ), respectively.
By keeping all of the joint moments of a time–frequency dis-
tribution, one can preserve all of its information. This is due where is the length of the feature set.
to the fact that any time–frequency distribution can be recon- To determine the optimum number of principal components
structed using a unique set of joint time–frequency moments (PC), LOI is calculated since the number of PCs is sequentially
and vice-versa [22]. This is proved using the mathematical def- increased, starting from the one with the largest eigenvalue.
inition of characteristics function, that is, a moment generating Step 5: Pattern Classifier: In this paper, the support vector
function. The characteristic function is a double Fourier trans- machine (SVM) shows the highest reliability compared to other
form of the TFD from the time–frequency domain to the classifiers. SVM is a linear discriminant classifier, which relies
Doppler-lag domain and is defined as follows: on the preprocessing of the data in higher dimension [16]. A
Gaussian nonlinear ( ) function is deployed to map
the feature set to a higher dimension. It can be proved that in
(11) a sufficiently higher dimension, data from a two-class classifier
can always be separated [33]. The detection algorithm separates
two conditions, namely, (for HIF), and (for NHIF). As-
Taylor expansion coefficients of CF are the JTFM sume there are samples from the class, and samples from
the class. After SVM mapping, a linear discriminant func-
(12) tion is trained in a way that maps the feature vectors to a
linear class boundary by maximizing the criterion function. The
boundary solution of this maximization routine is
This means that the TFD is uniquely reconstructed from
JTFM and vice-versa, due to the correspondence between (16)
Fourier couples and between Taylor series couples [16], [33].
In other words, the 2-D time–frequency signature of any where and are the sample mean vectors and and
waveform could be uniquely reconstructed from its set of joint are the scatter matrices of the H-class and NH-class, re-
moments. Therefore, we utilized JTFM as a 1-D and faithful spectively. and are calculated as follows:
representation of the 2-D TFD with the following definition:
(17)
(13)
(18)
where is the joint time–frequency feature set. And and
are non-negative integers.
In the next step, the feature set will be normalized using its where is the th mapped sample in the H-class, and is the
statistical descriptions as given in (14). This normalization pre- number of samples in the H-class.
vents features with large magnitudes from dominating the total and are calculated the same way as and ,
variance respectively, with two differences: is replaced with ,
which is the th sample in the NH class, and is the number of
(14) samples in the NH class.
where is the normalized feature set, and and are the av- V. ALGORITHM PERFORMANCE EVALUATION
erage and standard deviation of the features over the sampling In this section, the efficiency of the proposed technique on
set, respectively. real-world data is discussed. Two-hundred forty HIF cases and
The unique characteristics of this feature set are that it is ca- 120 NHIF cases were used to train and test the algorithm. The
pable of keeping all of the 2-D time–frequency patterns of the criteria for performance evaluation and result will be discussed
signal in a 1-D stream of data. As shown in Fig. 3, each power in the following subsections.
GHADERI et al.: HIF DETECTION IN THE DISTRIBUTION NETWORK USING THE TIME-FREQUENCY-BASED ALGORITHM 1265
A. Fault Detection Evaluation Criteria (FDEC) by dividing the number of right prediction of nonfaults over the
total number of predicted nonfaults
The complexity of the HIF phenomena and their detection
techniques necessitate a complete set of criteria for their evalu-
ation. Although tripping the HIF ensures public safety, it would (23)
disconnect important loads, such as traffic lights, elevators, and
hospitals. In order for utilities to conduct a risk analysis on this 5) Sensibility: is required to emphasize the risk of tripping
tradeoff, two criteria are proposed—safety and sensibility. Due sensitive loads, for example, hospitals. Sensibility (SN) can be
to the fact that HIF detection techniques have higher complexity, formulated as the fraction of the number of right prediction of
some criteria are also needed to compare their efficiency. faults, over the total number of predicted faults
Defining power system protection algorithms as a pattern
classification task provides the opportunity to evaluate the pro- (24)
tection algorithm using the well-defined assessment methods
for classifiers, for example, “Confusion Matrix” (CM). For These five measures are named “reliability criteria”, and are
two-class classifiers, as of the fault detection algorithms, CM is desired to be as high as possible. The next four measures weigh
a 2 2 matrix defined as the general performance of the method and emphasize their ro-
bustness and investability.
(19) 6) Cost: takes into account the computational complexity
(operational burden, training complexity) and hardware costs
(additional measurement devices and communication infra-
where “ ” is the number of detected faults, “ ” is the structure).
number of not detected faults, “ ” is the number of nonfaults 7) Objectivity: highlights whether the technique is ob-
that mistakenly activate the relay, and “ ” is the number of jective to the type of fault, and the network topology. This crite-
healthy conditions that do not mistakenly activate the relay. rion guarantees that the same method is able to be applied to dif-
Based on the confusion matrix, five criteria are defined for ferent applications and network topologies. In other words, this
the evaluation of the reliability of the HIF detection algorithm. criterion is rather a qualitative philosophy behind the protection
These criteria are: 1) accuracy; 2) dependability; 3) security; 4) scheme than a criterion that ought to be verified quantitatively.
safety; and 5) sensibility. In addition, three criteria for method 8) Completeness: demonstrates whether the technique
efficiency and practicality evaluation are proposed. These preserves the information during its feature extraction/feature
criteria are: 6) cost; 7) objectivity; and 8) completeness of the selection routine. This criterion guarantees that no important
detection method. These parameters form comprehensive fault information is discarded during the process, but rather com-
detection evaluation criteria (FDEC), which are defined as presses it. If the completeness criterion is not met, a new ob-
follows. served sample might necessitate the modification of the method.
1) Accuracy: emphasizes the all-inclusive performance of 9) Speed: is the time window of the data that is needed
the method, by means of dividing the number of right decisions for the method to make the crucial decision for HIF, because of
over the total number of decisions its possible safety hazards and intermittent nature. The higher
the number of power frequency cycles it takes, the higher the
(20) number of possible arc ignitions. As a result, the number of data
cycles needed for the algorithm to perform can be a measure of
2) Dependability: is defined as the proneness of the pro- speed of the protection technique. This can be formulated as the
tection scheme to detect and isolate the fault, which can be for- fraction of the duration of one cycle of power frequency current
mulated as the fraction of the number of detected faults to the ( ) over the time it takes for the method to detect the fault ( )
actual number of faults
(25)
(21)
B. Feature Extraction
3) Security: illustrates the method ability to selectively de- JTFM features were extracted by utilizing (13). For each one-
tect the fault, and not to trip for nonfault conditions. It is the fourth of the cycle, the signal length is 83 samples, and
fraction of the number of predicted nonfaults over the actual frequency bins are also set to 83 levels. JTFM size (
number of nonfaults ) is the set of one order of magnitude smaller than
in order to obtain an acceptable Taylor expansion estimation of
the characteristic function. This results in 36 features in each
fourth cycle and 144 features in the whole cycle. Due to the pos-
4) Safety: is needed to reflect the method ability to isolate sible loss of information through the windowing and normaliza-
the faults that are hazardous to the public, for example, HIF in a tion process, several common signal-processing features are ex-
crowded area and islanding [24]. It is the ability of the method tracted before processing. These five features are energy ( ),
to be vigilant for fault occurrence and to make sure that faults mean time ( ), mean frequency ( ), time duration ( ),
are not mistaken for nonfaults. This criterion can be calculated and bandwidth ( ) of the total cycle of the current. These are
1266 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 30, NO. 3, JUNE 2015
TABLE III
PCA EFFICIENCY FOR DIFFERENT NUMBERS OF PCS
C. Feature Selection Fig. 5. The 2-D representation of HIF and NHIF in the feature domain.
D. Pattern Classifier
In this step of the algorithm, a hyperplane boundary has to be
found in the 9-D feature plane, using a pattern classifier tech-
nique. Since it is not possible to represent the feature set in nine
dimensions, they are represented in a 2-D feature plane as in
Fig. 5. The magnitude of each feature for fault class (HIF) and
nonfault class (NHIF) is demonstrated. It can be observed that a
linear discriminant function cannot successfully separate these
classes.
In this paper, the SVM classifier is utilized to separate two Fig. 6. Comparison of different pattern classifiers for HIF detection.
classes by means of a linear hyperplane in the higher dimen-
sion feature domain. Training an SVM has low computational
burden, with high reliability and stability. Fig. 6 shows the effi- SVM classifier with nine principal components has the highest
ciency of several classifiers with lower complexity (linear dis- reliability.
criminant function) and higher complexity (quadratic discrim-
inant function, naive Bayes, and artificial neural network) than E. Comparison
SVM. This comparison is performed using the detection criteria Table IV demonstrates the comparison between the proposed
(FDEC) introduced in Section V-A. For the purpose of training time–frequency-based technique, and recent HIF detection
the classifier, 80% of the feature sets are randomly selected and methods, based on FDEC. If the information for extracting the
the rest are dedicated to the test set. As shown in Fig. 6, the criterion was not provided by this paper, not available (NA)
GHADERI et al.: HIF DETECTION IN THE DISTRIBUTION NETWORK USING THE TIME-FREQUENCY-BASED ALGORITHM 1267
TABLE IV
COMPARISON OF THE PROPOSED METHOD WITH THE RECENTLY PROPOSED TECHNIQUES
[18] B. Ghoraani and S. Krishman, “Discriminant non-stationary signal fea- Amin Ghaderi (S’12) received the B.Sc. degree in
tures clustering using hard and fuzzy cluster labeling,” EURASIP J. electrical engineering from the University of Tehran,
Advances Signal Process., vol. 2012, no. 1, pp. 1–20, Nov. 2012. Tehran, Iran, in 2008, the M.Sc. degree in power
[19] W. Min and A. V. Mamishev, “Classification of power quality events systems engineering from Iran University of Science
using optimal time-frequency representations-part 1: Theory,” IEEE and Technology, Tehran, in 2011, and is currently
Trans. Power Del., vol. 19, no. 3, pp. 1488–1495, Jul. 2004. pursuing the Ph.D. degree in electrical engineering
[20] Y.-J. Shin, E. J. Powers, M. Grady, and A. Arapostathis, “Power quality from the University of South Carolina, Columbia,
indices for transient disturbances,” IEEE Trans. Power Del., vol. 21, SC, USA.
no. 1, pp. 253–261, Jan. 2006. He is now working on nonperiodic load current
[21] Y.-J. Shin, E. J. Powers, W. M. Grady, and A. Arapostathis, “Signal compensation using power-electronics-based de-
processing-based direction finder for transient capacitor switching dis- vices. His main research interests include power
turbances,” IEEE Trans. Power Del., vol. 23, no. 4, pp. 2555–2562, quality, power system protection, and power electronics.
Oct. 2008.
[22] L. Cohen, Time-Frequency Analysis. Upper Saddle River, NJ, USA:
Prentice-Hall, 1995.
[23] B. Boashash, Time-Frequency Signal Analysis and Processing; A Com- Hossein Ali Mohammadpour (S’10) received the
prehensive Reference. New York, USA: Elsevier, 2003. B.Sc. and M.Sc. degrees in electrical engineering and
[24] A. Ghaderi, A. Esmaeilian, and M. Kalantar, “A novel islanding de- power systems from Iran University of Science and
tection method for constant current inverter based distributed genera- Technology, Tehran, Iran, in 2006 and 2009, respec-
tions,” in Proc. Int. Conf. Environment Elect. Eng., May 2011, pp. 1–4. tively, and is currently pursuing the Ph.D. degree in
[25] J. Jeong and W. J. Williams, “Kernel design for reduced interfer- electrical engineering at the University of South Car-
ence distributions,” IEEE Trans. Signal Process., vol. 40, no. 2, pp. olina, Columbia, SC, USA.
402–412, Feb. 1992. His primary research interests include power sys-
[26] C. Hory, N. Martin, and A. Chehikian, “Spectrogram segmentation by tems stability, power electronics, renewable energy,
means of statistical features for non-stationary signal interpretation,” flexible ac transmission system technologies, and
IEEE Trans. Signal Process., vol. 50, no. 12, pp. 2915–2925, Dec. electric ship system modeling and analysis.
2002.
[27] C. Kocaman, O. Ozgonenel, M. Ozdemir, and U. K. Terzi, “Calculation
of fundamental power frequency for digital relaying algorithms,” in
Proc. Develop. Power Syst. Protect., Mar. 2010, pp. 1–5.
[28] H. Edels, “ Properties, theory of the electric arc,” Proc. Inst. Elect. Eng. Herbert L. Ginn, III (M’14) received the M.S.
A: Power Eng., vol. 108, no. 37, pp. 55–69, Feb. 1961. and Ph.D. degrees in electrical engineering from
[29] B. Tacer and P. J. Loughlin, “Non-stationary signal classification Louisiana State University, Baton Rouge, LA, USA,
using the joint moments of the time-frequency distributions,” Pattern in 1998 and 2002, respectively.
Recogn., vol. 31, no. 11, pp. 1635–1641, Nov. 1998. Currently, he is an Associate Professor with the
[30] I. Navarro, F. Sepulveda, and B. Hubias, “A comparison of time, fre- Electrical Engineering Department at the University
quency, ICA based features and five classifiers for wrist movement of South Carolina, Columbia, SC, USA. His areas
classification in EEG signals,” in Proc. IEEE Int. Conf. Eng. Med. Biol. of specialization are power-electronics applications
Soc., Mar. 2005, vol. 4, pp. 2118–2121. in energy systems, power phenomena and compen-
[31] P. Paatero and U. Tapper, “Positive matrix factorization: A non-neg- sation in nonsinusoidal systems, and power quality.
ative factor model with optimal utilization of error estimates of data His current research activities focus on the control
values,” Environmetrics, vol. 5, pp. 111–126, 1994. and coordination of power-electronic converters in special case distribution sys-
[32] R. J. Prokop and A. P. Reeves, “A survey of moment-based techniques tems, such as self-contained vehicular systems and microgrids.
for unoccluded object representation and recognition,” CVGIP: Graph-
ical Models Image Process., vol. 54, no. 5, pp. 438–446, Sep. 1992.
[33] B. Ghoraani and S. Krishnan, “Time-frequency matrix feature extrac-
tion and classification of environmental audio signals,” IEEE Trans. Yong-June Shin (S’98–SM’04) received the B.S.
Audio, Speech, Language Process., vol. 19, no. 7, pp. 2197–2209, Sep. (Hons.) degree in electrical engineering at Yonsei
2011. University, Seoul, Korea, in 1996, the M.S. degree
[34] S. Gautam and S. M. Brahma, “Detection of high impedance fault in electrical engineering from The University of
in power distribution systems using mathematical morphology,” IEEE Michigan, Ann Arbor, MI, USA, in 1997, and the
Trans. Power Syst., vol. 28, no. 2, pp. 1226–1234, May 2013. Ph.D. degree in electrical and computer engineering
[35] M. Carpenter, R.R. Hoad, T. D. Bruton, R. Das, S. A. Kunsman, and J. from The University of Texas at Austin, Austin, TX,
M. Peterson, “Staged-fault testing for high impedance fault data collec- USA, in 2004.
tion,” in Proc. Annu. Conf. Protect. Relay Eng., Apr. 2005, pp. 9–17. Upon his graduation, he joined the Department
[36] C. G. Wester, “High impedance fault detection on distribution sys- of Electrical Engineering, The University of South
tems,” in Proc. Rural Elect. Power Conf., 1998, pp. 1–5. Carolina, Columbia, SC, USA, as an Assistant
Professor. He was promoted to Associate Professor with tenure in 2011. He
joined the School of Electrical and Electronic Engineering, Yonsei University,
Seoul, Korea as Associate Professor in 2012. His current research interests are
characterized by the application of novel digital-signal-processing techniques
to a wide variety of important transient and nonlinear problems in smart electric
power grids.
Prof. Shin is a recipient of the United States National Science Foundation
CAREER award in 2008, and the General Electric Korean-American Education
Commission Scholarship.