Electrical Power and Energy Systems 49 (2013) 243–252

Contents lists available at SciVerse ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier .com/locate / i jepes

An ANFIS-based fault classification approach in power distribution system

J. Zhang a, Z.Y. He b,⇑, S. Lin b, Y.B. Zhang a, Q.Q. Qian b

a State Grid Energy Research Institute, Beijing, Chinab Department of Electrical Engineering, Southwest Jiaotong University, Chengdu, SC, China

a r t i c l e i n f o

Article history:Received 5 August 2010Received in revised form 11 December 2012Accepted 20 December 2012Available online 26 February 2013

Keywords:Fault classificationNeutral non-effectively groundeddistributionAdaptive Neural Fuzzy Inference System(ANFIS)Wavelet transform

0142-0615/$ - see front matter � 2013 Elsevier Ltd. Ahttp://dx.doi.org/10.1016/j.ijepes.2012.12.005

⇑ Corresponding author. Tel.: +86 2887602445; faxE-mail address: hezy@swjtu.cn (Z.Y. He).

a b s t r a c t

Fault classification is very important for power system operation because it is the premise of faultanalysis process. In this paper, an ANFIS (Adaptive Neural Fuzzy Inference System) based fault classificationscheme in neutral non-effectively grounded distribution system is proposed. The transient currents areobtained by wavelet transform after faults occur. According to the statistic characteristic of transientcurrents in different fault types, the fault identifiers are defined. The fault identifiers can characterizethe traits of fault type and show different disciplinarian in different fault types. They are inputted intothree ANFISs to obtain the fault type. The proposed approach only needs the voltages and currentsmeasured at substation, and can identify ten types of short-circuit fault accurately. The simulation modelis established in PSCAD/EMTDC environment, and the performance of the proposed approach is studied.The results show that it has high accuracy. Besides, the adaptability of proposed approach to the neutralcompensated grounding system, different network configurations and so on are verified throughsimulation. Through simulation, the proposed approach exhibits good performance.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Once fault occurs in power system, fast and accurate faultclassification is very important for post-fault analysis and powersupply restoration. On the one hand, fault type information is thepremise of fault location. Correct fault classification helps faultdiagnosis system select appropriate fault location scheme. On theother hand, as the automation of distribution network advances,a mass of data is uploaded after fault. So, it is impossible for oper-ators to classify fault correctly by their experience.

A lot of work has been focused on the fault classification prob-lem in transmission network [1–6]. It is probably due to the impor-tance of fault classification for relay operations. As the amount ofpower carried by distribution networks increases during the pastdecades, the need of effective fault classification scheme becomesmore and more necessary for operators. The research of fault clas-sification in distribution network can be summarized to two cate-gories approximately: (1) the schemes using stationary electriccomponents [7–9], and (2) the schemes using transient electriccomponents [10,11]. The method proposed in [7] is widely usedin industry. It utilizes the angular relationships among stationarycomponents to identify fault type in isolated distribution. Theangular relationships are defined by fuzzy membership functionsin [8], afterward, the fault type is obtained by logic reasoningmechanism. As fast sampling devices are installed in modern

ll rights reserved.

: +86 2887600861.

power systems, the acquisitions of transient electric signals arepossible. It highly supports the transient-based post-fault analysistechniques, such as fault location [10], fault line identification [12]and so on. In the paper [10], the transient currents are obtained bystationary wavelet transform. Afterward, by comparison withenergies of transient currents, the fault type is identified.

In this paper, an ANFIS-based fault classification approach indistribution network is proposed. It uses transient fault signals toactualize ten types of short-circuit fault classification (AG, BG,CG, ABG, ACG, BCG, ABC/ABCG, AB, AC and BC). The whole classifi-cation process contains 3 steps: the transient fault signals are ex-tracted by WT (Wavelet Transform) at first. At second, theextracted signals are calculated for the statistic quantities, whichare called FIs (Fault Identifiers) here. At third, the FIs are inputtedinto three ANFISs (Adaptive Network-based Fuzzy Inference Systems)to obtain the final result. The simulation model is established inPSCAD/EMTDC environment. The classification accuracy of pro-posed approach is verified under different fault circumstances.Moreover, the adaptability of proposed approach to different dis-tribution operations is thoroughly studied. The results show thatits adaptability is good.

The remainders of this paper are constructed as follows. Thewhole structure of proposed approach is introduced in section 2for clearness. The regulars of FIs in different fault types are studiedin section 3. The ANFIS-based fusion process is introduced in sec-tion 4. The classification accuracy of proposed approach is exhib-ited in section 5. The adaptability of proposed approach isstudied in section 6. At last, the conclusion is given in section 7.

244 J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252

2. The whole structure

Although it is important how to trig fault classificationapproach, it is not the certain in this paper. The method proposedin the paper [10] can be employed to trig the fault classificationapproach and find the fault inception time. The structure ofclassification approach is depicted in Fig. 1.

In Fig. 1, the inputs of fault classification are the zero-sequencevoltage u0(t) and the fault-components of three-phase currents i00a,i00b and i00c . They are calculated by formulas (1) and (2),

u0ðtÞ ¼ uaðtÞ þ ubðtÞ þ ucðtÞ; ð1Þi00pðtÞ ¼ ipðtÞ � ipðt � TÞ: ð2Þ

where p stands for phase-a, b, and c, ip(t) is the phase-p current inthe first basic cycle after fault inception at the secondary windingof transformer, ua(t), ub(t) and uc(t) are the three bus voltages inthe first cycle after fault inception, and T is the basic cycle (20 ms).

As can be seen from Fig. 1, u0(t) and i00p are preprocessed by FFT(Fast Fourier Transform) and WT separately. Moreover, the statisticquantities are calculated for WT-extracted signals. The 2 stepsabove aim at constructing the FIs, which are inputted into ANFISsto obtain the classification result. The FIs and their fusion processare very important for fault classification, because FIs must havedistinguished features in different fault types, and the fusion pro-cess should fuse FIs effectively to obtain the correct results.

3. Fault Identifiers (FIs)

When fault occurs in power system, transient oscillation isdominant. To utilize the transient electric signals, WT techniqueis employed for its flexible time–frequency focus. By comparisonwith Daubechies series wavelets, the quadratic spline wavelet isselected as the mother wavelet in this work because of its goodperformance in fusion process. The coefficients of wavelet filter gi-ven by formulas (3)–(5) are typical, and are adopted in this work.

lo ¼ ½ 0:125 0:375 0:375 0:125 � �ffiffiffi2p

; ð3Þ

hi ¼ ½1:0 �1:0 � �ffiffiffi2p

; ð4Þ

hii ¼ ½0:015625 0:109375 0:34375 � 0:34375

� 0:109375 � 0:015625� �ffiffiffi2p

: ð5Þ

In (3)–(5), lo is the coefficient of low-pass filter in waveletdecomposition and reconstruction calculation, and hi and hii are

Wavelet

FFT

Construct fault

Ia''(t)ib''(t)ic''(t)

u0(t)

Bus

Bus

Feeders

Source

Trans-former

c

Fig. 1. Structure of fa

the coefficients of high-pass filter in wavelet decomposition andreconstruction calculation respectively.

The current i00p is implemented 3 levels of wavelet decomposi-tion. The detail coefficients of 2 and 3 levels are reconstructed.The reconstructed signal is denoted by i0p. As long as the samplingrate is 10 kHz, it can be known that the frequency band of i0pranges from 625 Hz to 2.5 kHz. The frequency band employs tran-sient components as much as possible, while avoids the interfer-ence of the 3rd–11th harmonics which are normal in powersystem.

Fig. 2 shows the original currents and WT-extracted currents inACG fault occurring at 0.1 s.

It can be seen from subfigure (b) in Fig. 2 that the WT-extractedsignals of faulty phases (phase-a and -c) vary more intensively thanthe WT-extracted signal of healthy phase (phase-b). The descrip-tion is qualitative, so the quantitative description should be foundout. As long as i0p is the signal in one basic cycle, it has 200 samplingpoints. There would be 600 inputs if i0a, i0b and i0c are inputted intoANFISs directly. This is unpractical. Since the statistic quantitiescan characterize the shape and energy of signals, six statistic quan-tities are selected from 12 typical statistic quantities [13], and theyare calculated by,

s�p ¼sp

smaxp ¼ a; b; c; ð6Þ

qa;b ¼Eði0ai0bÞ � Eði0aÞEði

0bÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

Eði0aÞ2 � E2ði0aÞ

q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiEði0bÞ

2 � E2ði0bÞq

��������������; ð7Þ

qa;c ¼Eði0ai0cÞ � Eði0aÞEði

0cÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

Eði0aÞ2 � E2ði0aÞ

q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiEði0cÞ

2 � E2ði0cÞq

��������������; ð8Þ

qb;c ¼Eði0bi0cÞ � Eði0bÞEði

0cÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

Eði0bÞ2 � E2ði0bÞ

q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiEði0cÞ

2 � E2ði0cÞq

��������������: ð9Þ

In formula (6),

sp ¼1

n� 1

Xn

i¼1

ði0pðnÞ � Eði0pÞÞ !1

2

; n ¼ 200; ð10Þ

smax ¼maxðspÞ: ð11Þ

ANFIS

Fusion identifiers

AG,BG,CG,

ABG,ACG,BCG,ABC/

ABCG,AB,AC,BC.

Fault classfication

Statisticalculation

ult classification.

Time t (s)

Cur

rent

i (A

)

0.08-100

-Phase C

-Phase B

-Phase A

-50

0

50

100

0.1 0.2

(a)

Sampling Points

Cur

rent

i (A

)

170-8

0

6

200 400

-Phase C

-Phase B

-Phase A

(b) Fig. 2. Three phase currents, (a) the original, and (b) the WT-extracted.

J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252 245

In formulas (7)–(9), E(x) is the mathematical expectation ofvariable x.

As can be seen, s�a, s�b and s�c are the normalized standard devia-tions of three phases, qa;b, qa;c and qb;c are the correlation coeffi-cients between two phases.

Besides 6 statistic quantities above, the U0 (zero-sequenceamplitude of fundamental component) is calculated by FFT fromu0(t). It is used to judge whether the fault is grounded or not.

The 6 statistic quantities and U0 are defined as FIs. Take thewaves in Fig. 2 for example, the FIs are: s�a ¼ 0:788, s�b ¼ 0:305,s�c ¼ 1, qa;b ¼ 0:595, qa;c ¼ 0:967, qb;c ¼ 0:773. As can be seen, s�aand s�c are bigger than s�b, and the qa;c is bigger than qa;b and qb;c .To widely study the performance of FIs in different fault types,AG, ABG, AB and ABC/ABCG faults are assumed in the isolated dis-tribution. The fault conditions are the combinations of six faultresistances (0 X, 50 X, 100 X, 200 X, 500 X and 1 kX) and 6 FIAs(Fault Inception Angles) (0�, 18�, 36�, 90�, 126� and 162�). The FIAis defined as the angle of phase-a. As can be seen, the total numberof fault condition in each fault type is 36. The variances of FIs areshown in Fig. 3.

From Fig. 3, these four fault types can be distinguished by com-parison of FIs in 3 steps:

� AG and ABG faults can be discriminated from AB and ABC/ABCGfaults, because U0 is non-zero from subfigures (a) and (b) inFig. 3 and zero from subfigures (c) and (d). This is because U0

is zero when zero-sequence loops do not exist (AB fault) or sym-metric fault (ABC/ABCG) occurs. In addition, it has been alsonoted that U0 decreases as the fault resistance increases.

� AG and ABG can be distinguished from each other, because s⁄ ofthe faulty phase is bigger than s⁄ of the healthy phases (subfig-ures (a) and (b)), and q between two faulted phases is biggerthan the others (subfigure (b)).� AB and ABC/ABCG can be distinguished from each other,

because the s⁄ of the faulty phase is always bigger than thatof the healthy phase (subfigures (c) and (d)), and q betweentwo faulty phases is visibly bigger than others in AB fault (sub-figure (d)) while symmetric fault does not have such regular(subfigure (c)).

4. Fusion process

In the classification process in section 3, the word ‘‘bigger’’, ‘‘vis-ibly bigger’’ and ‘‘zero’’ are heuristic and linguistic. So, the faultclassification can be considered as a fuzzy inference problem. Infuzzy inference, the definitions of membership functions are notan easy work. The changeful fault conditions make their definitionsdifficult. The combination of fuzzy logic with an architecture de-sign of ANN has lead to the creation of ANFIS which has been usedfor every plant [14–19]. The ANFISs can adjust the parameters ofmembership functions adaptively from history data and take theuncertainties (fault resistances, FIAs, etc.) into account well. Thisis of utmost importance since it is well known that the constantthresholds for all fault conditions are very difficult to choose.

In this work, three ANFISs are employed for different purposes.They work as Fig. 4 shown.

From Fig. 4, ANFIS1 is used to identify whether fault is AG, BGCG, LLG (Line-to-Line-to-Ground) or others. If the result from AN-FIS1 is ‘‘LLG’’, ANFIS2 would be used further. If the result is ‘‘oth-ers’’, ANFIS3 would be used. Because the outputs of ANFISs arenumeral, their corresponding fault type is shown in Fig. 4. The in-put vectors of ANFIS1, ANFIS2 and ANFIS3 are C1, C2 and C3 respec-tively. They are the combinations of FIs and shown as follows,

C1 ¼ ½ s�a s�b s�c U0 � ð12Þ

C2 ¼ C3 ¼ ½qa;b qa;c qb;c s�a s�b s�c �: ð13Þ

The inputs and outputs of ANFISs are shown in Fig. 5.The general structure of ANFIS with two inputs is shown in

Fig. 6 [14].As can be seen from Fig. 6, ANFIS can be demonstrated in five

layers: the first layer actualizes the fuzziness of inputs; the sec-ond layer calculates the firing strength of each rule; the thirdlayer normalizes the firing strengths; the fourth layer determinesthe consequent parameters of the rule; the fifth layer computesthe output of the fuzzy system by summing up the outputs ofthe fourth layer.

The feedforward equations of ANFIS in Fig. 6 are shown asfollows,

wi ¼ lAiðxÞ � lBiðyÞ i ¼ 1;2: ð14Þ

wi ¼ wi=ðw1 þw2Þ i ¼ 1;2: ð15Þ

f1 ¼ p1xþ q1yf2 ¼ p2xþ q2y

�; ð16Þ

f ¼ w1f1 þw2f2; ð17Þ

where x and y are the inputs of ANFIS, l is the membership functionof Ai and Bi which are fuzzy sets associated with the inputs, wi de-notes the firing strength of a rule, wi is the normalized wi, pi and qi

are the parameters sets of the consequent parameters, fi is a func-tion, and f computes the overall output.

1# 15#0.4

0.5

0.6

0.7

0.8

0.9

1.0

Fault Cases

18°; 0~1k

Fault Cases

0.8

0.9

1.0

Cor

rela

tion

Coe

ffic

ient

s

Fault Cases

0

25

50

Zer

o-se

quen

ce V

olta

ge U

0

*as*bs*cs

,a bρ

,a cρ,b cρ

1# 15# 30# 36# 1# 15# 30# 36#

0.0

0.2

0.4

0.6

0.8

1.0

Fault Cases

Stan

dard

dev

iatio

ns (

p.u)

Fault Cases

1.0

Cor

rela

tion

Coe

ffic

ient

s

Fault Cases

0

10

30

Zer

o-se

quen

ce V

olta

ge U

0

0.0

0.2

0.4

0.6

0.8

20

*as*bs*cs

,a bρ

,a cρ,b cρ

1# 15# 1# 15# 1# 15#

0.0

0.2

0.4

0.6

0.8

1.0

Fault Cases Fault Cases

1.0

Cor

rela

tion

Coe

ffic

ient

s

Fault Cases

0.00Zer

o-se

quen

ce V

olta

ge U

0

0.0

0.2

0.4

0.6

0.8

0.01

*as*bs*cs

,a bρ

,a cρ,b cρ

1# 15# 1# 15# 1# 15#

0.0

0.2

0.4

0.6

0.8

1.0

Fault Cases Fault Cases

1.0

Cor

rela

tion

Coe

ffic

ient

s

Fault Cases

0.00Zer

o-se

quen

ce V

olta

ge U

0

0.0

0.2

0.4

0.6

0.8

0.01*as*bs*cs

,a bρ

,a cρ,b cρ

1# 15# 1# 15# 1# 15#

30# 36#

30# 36# 30# 36# 30# 36#

30# 36# 30# 36# 30# 36#

30# 36# 30# 36# 30# 36#

Stan

dard

dev

iatio

ns (

p.u)

Stan

dard

dev

iatio

ns (

p.u)

Stan

dard

dev

iatio

ns (

p.u)

(a)

(b)

(c)

(d)Fig. 3. Variance of FIs in different fault types. (a) AG fault, (b) ABG fault (c) ABC/ABCG fault, and (d) AB fault.

246 J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252

In this work, ANFISs are established in MATLAB environment[20]. The ANFIS1 has 4 inputs, and ANFIS2 and ANFIS3 both havesix inputs. Each input of ANFIS has two ‘‘Gauss’’ membership func-tions. The ANFISs are initialized by grid partition method. From theknowledge of ANFIS, there are 16 (24) rules in ANFIS1 and 64 (26)rules in ANFIS2 and ANFIS3. Moreover, the ‘‘and’’, ‘‘or’’ and ‘‘defuzz-ification’’ methods in ANFISs are selected as ‘‘product’’, ‘‘max’’ and‘‘wtaver’’ respectively.

5. Train and test of ANFISs

5.1. Simulating model

For the absence of field data, the simulating model is estab-lished in PSCAD/EMTDC environment to produce fault data [21].Its configuration is referenced by IEEE-34 workbench. BecauseIEEE-34 workbench is neutral directly grounded system [22], some

ANFIS 1

1(AG)

2(BG)

3(CG)

4(ABG)

5(ACG)

Vector C2

4(LLG)ANFIS 2 0

(Others)

7(ABCG/

ABC)

8(AB)

Vector C3

ANFIS 3

6(BCG)

9(AC)

10(BC)

over

Vetor C1

Fig. 4. Process of fault classification.

ANFIS 2

ρa,b

4, 5 or 6

ρa,c ρb,c sa* sb

* sc*

ANFIS 3

ρa,b

7, 8, 9 or 10

ρa,c ρb,c sa* sb

* sc*

ANFIS 1

1, 2, 3, 4 or 0

sa* sb

*sc* U0

Fig. 5. Inputs and outputs of ANFISs.

A1

xA2

B1

yB2

layer 1 layer 2 layer 3 layer 4 layer 5

N

N

f1

f2

Σ f

w1

x

w2

w1

w2

y

x y

w1 f1

f2w2

II

II

Fig. 6. General structure of ANFIS.

J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252 247

modifications are implemented for simulating neutral non-effec-tively grounded system. The configuration of model is shown inFig. 7.

In Fig. 7, E is the external source with impedance Zs =2.3 + j18.4 X. T-1 is transformer and T-2 is Zig-Zag transformer insubstation. The transformers adopt the high frequency model in[23]. The compensate reactance L = 7.8 H, over compensatedsystem by 10%. K is the switch which controls the neutralgrounding style of the system. F denotes the fault location and Rf

is the fault resistance. The line parameters are: zero-sequenceimpedance z0 = 0.52 + j1.5 X/km, positive- and negative-impedance z1 = z2 = 0.26 + j0.37 X/km, zero-sequence susceptanceb0 = j0.15 � 10�5 s/km, positive- and negative-susceptance b1 = b2

= j0.31 � 10�5 s/km. The length of lines and the loads are as sameas those in [22].

5.2. Train of ANFISs

To obtain enough data to train ANFISs, ten types of short-circuitfault with different conditions are assumed in the establishedmodel. The fault conditions are listed in Table 1.

All fault cases are simulated in the isolated system (with switchK open). The obtained data is used to train ANFISs in MATLAB envi-ronment. In train step, the stop criteria are that the train epochsreaches 10,000 or the RMSE (Root Mean Square Error) is smallerthan 0.0003. The convergence curves of ANFISs after train areshown in Fig. 8.

From Fig. 8, the train of ANFIS1 and ANFIS3 is completed at8328 and 703 epochs respectively, since the minimum RMSE isreached. The train of ANFIS2 is stopped when the maximumepochs are reached, and the final RMSE is about 0.0067.

5.3. Test of ANFISs

To validate the classification accuracy of trained ANFISs, 3600fault cases which are different from those in train step are simu-lated. Their fault conditions are listed in Table 2.

As can be seen in Table 2, the maximum fault resistance(1.5 kX) is more than half of that in train. It is considered to bedifficult for ANFISs, because the extrapolation ability of neuralnetworks is not as good as its interpolation ability.

Fig. 7. Simulating model in PSCAD/EMTDC.

Table 1Different fault conditions in train step.

Faulted bus 802,810,814,822,826,856,864,844,838,840Fault resistance (X) 0,50,100,200,500,1000FIA (�) 0,18,36,90,126,162Total number 3600

0.14

RM

SE

0.00100000

Epoches

ANFIS1

ANFIS2

ANFIS3

Fig. 8. Convergence curves of ANFISs after train.

Table 2Different fault conditions in test step.

Faulted buses Fault resistance (X) FIA (�) Number

806,816,818,830 0, 5,10,25,40,50 0,18,54,72,108,144 1440812 0, 50,100,200,500,1000 0,18,36,90,126,162 360832,834 20,70,150,300,600,1200 18,54,72,90,108,144 720848,860,862 20,70,150,300,1200,1500 18,54,72,90,108,144 1080Total 3600

99.0

99.2

99.4

99.6

99.8

100.0

Acc

urac

y R

ate

(%

)

99.1

99.3

99.5

99.7

99.9

BGCG

LLGOthersAG

ABGABC/

ABCGACG

BCGAB

ACBC

ANFIS1

ANFIS2

ANFIS3

100 100

99.4

99.8

100

99.4

100

99.7

100 100 100 100

Fig. 9. Statistic results of test cases.

The approach proposed in this paper

The approach proposed in [10]

65

75

80

95

100

90

85

70

Acc

urac

y R

ate

(%

)

BGCGAG

ABGABC/ABCG

ACGBCG

ABAC

BC

Fig. 10. Comparison of two approaches.

248 J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252

All fault cases in Table 2 are simulated. The obtained data isused to test ANFISs. Because the outputs of ANFIS are not alwaysintegers, the round off method is adopted here. From the testresults, the accuracy of ANFISs is very good. Due to the spaceconstraint, it is impossible to list all results in the paper. Hence,the statistic results are shown in Fig. 9.

The accuracy rate d in Fig. 9 is defined by,

d ¼ Right numbers from ANFIS classification360ðTotal numbers in each fault typeÞ � 100% ð18Þ

From Fig. 9, the minimum d is 99.4% in ABG faults which means 2out of 360 ABG faults are misclassified. In addition, through ourobservation, the classification results are right when fault resistanceis 1.2 kX and 1.5 kX. It reveals that the trained ANFISs have goodextrapolation ability.

In comparison with the other proposed approaches [8–11], thiswork needs high sampling devices and complicated calculation. Ifit does not have advantages, it would not be practical. So, the pro-posed approach is compared with the approach in the paper [10]for its transient-based either and recent publication. This approach

-Phase C

-Phase B

-Phase A

Time t (s)

-150

-50

0

50

150

0.1 0.30.2

(a)

-Phase C

-Phase B

-Phase A

Sampling Points

Cur

rent

i (A

)

-0.4

0

0.4

31 210

0.2

-0.2

Cur

rent

i (A

)

(b)Fig. 11. Currents of three phases, FIA = 70�, fault resistance is 70 X, (a) the original,and (b) the WT-extracted.

-160 0 160-2

0

2

Current (A)

Vol

tage

(kV

)

Fig. 12. V–I character of arc model.

98.0

98.4

98.8

99.2

99.6

100.0

Acc

urac

y R

ate

(%

)

98.2

98.6

99.0

99.4

99.8

BG CGAG

98.6

98.3

100ANFIS1

Fig. 13. Statistic results under arc faults.

J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252 249

extracts the energies of transient currents and compares them withthe thresholds to obtain the fault classification result. There arethree thresholds which are k3 (the minimum classification index),k4 (the minimum relative classification index) and k5 (the minimumzero-sequence current relation). They equal 2000, 2 and 30 respec-tively in our work. The fault cases listed in Table 2 are used to com-pare two approaches, and the results are shown in Fig. 10.

As can be seen from Fig. 10, the approach proposed in this paperexhibits dominant performance. For the approach proposed in thepaper [10], it is difficult to find appropriate thresholds which areadaptable to different fault conditions. Take ACG fault at 848 nodeof the simulating model for example, the original currents andWT-extracted currents are shown in Fig. 11.

Through calculation from (b) in Fig. 11, the energies of threephases are Ea = 1.89, Eb = 1.9 and Ec = 1.57. Since the energy ofhealthy phase Eb is even bigger than those of faulty phases Ea

and Ec, the approach is confused. However, the approach proposedin this paper is right in such fault case.

6. Adaptability analysis

6.1. Neutral compensated grounding distribution

Let the switch K close in Fig. 7, then the distribution modelbecomes neutral compensated grounding. The fault cases inTable 2 are simulated again and the obtained data is used to verifythe proposed approach.

Although the proposed approach is trained in the isolated net-work, the results show that it is hardly affected by the change ofneutral grounding style. The statistic results are similar to thosein Fig. 9, so they are not repeated here. As we know, the responseof suppress coil in high frequency band is quite slow in comparisonwith that of the distributed capacitances along feeder. So, the WT-extracted current signals are quite same as those in the isolateddistribution during the first post-fault cycle. It is intuitively whythe test results are quite same as those in the isolated distribution.

6.2. Non-linear fault resistance

In practice, fault always accompanies arc phenomenon. Thefault resistance of arc is non-linear and varies with wind speed,environmental temperature and so on. For study on the impactof the non-linear fault resistance, The arc model and its extinctioncondition introduced in the paper [24] are employed in this work.There are three parameters to initialize arc model, which are s0

(the initial time constant), l0 (the initial length) and a (the negativeexponent). They equal 0.25 ms, 0.2 m and �0.4 respectively. Thosevalues are considered to be typical in power distribution network.The simulating V–I character of arc model is presented in Fig. 12.

In comparison with the result reported in the paper [25], Fig. 12shows that the established arc model is very effective. Because SLGfault parameters are only provided in the paper [24], 1080(360 � 3) SLG test cases contained in Table 2 are simulated with

95

96

97

98

99

100100100 100

Acc

urac

y R

ate

(%

)

ANFIS1

ANFIS2

ANFIS3

99.799.5 99.699.699.6

96.1

94.3

100100100100 99.8

250 J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252

the arc fault. The accuracy of proposed approach to identify SLG(Single-Line-to-Ground) faults is verified, and the statistic resultsare shown in Fig. 13.

Because ANFIS1 actualizes the SLG faults classification, theaccuracy rate of ANFIS1 is only shown in Fig. 13. In comparisonwith Fig. 9, Fig. 13 shows that the non-linear fault resistance doesnot decrease the accuracy rate monotonically since it is enhancedin CG faults. In AG and BG faults, the accuracy rate decreasesslightly. In BG fault, it reaches 98.3% which means 6 of 360 casesare misclassified. The result is acceptable.

In conclusion, the non-linear fault resistance has little impacton the proposed approach in SLG faults.

Fig. 14. Different network configurations.

O94

a b c d O a b c d O a b c d

Fig. 15. Statistic results under different network configurations.

6.3. Different network configurations

Switch operations are common in distribution network to actu-alize load balance, power loss minimum and other purposes. So,the configuration of distribution network is not constant. The faultclassification approach must show good performance regardless ofdifferent network configurations. For the purpose, four differentnetwork configurations shown in Fig. 14 are considered to verifythe proposed approach.

As can be seen from Fig. 14, the configurations are obtainedfrom the original in Fig. 7, i.e. some laterals of the original config-uration are deleted to form the new configurations. The toughestcase is (d) in Fig. 14, because the 86% line length of total is deletedand only 5 nodes (800, 802, 806, 808 and 810) are left.

For comparison with the results in different configurations, tentypes of short-circuit fault are assumed at 806 node, and the faultconditions are the combinations of 10 fault resistances (0 X, 50 X,100 X, 200 X, 400 X and 700 X) and 10 FIAs (0�, 18�, 36�, 90�, 126�and 162�). The obtained data is used to verify the proposed ap-proach. The statistic results are shown in Fig. 15.

In Fig. 15, the symbol ‘O’ on x-axis denotes the original config-uration, and the symbols ‘a’, ‘b’, ‘c’ and ‘d’ correspond to the config-urations in Fig. 14.

As can be seen from Fig. 15, the more laterals are deleted, themore the accuracy rate decreases. In (a) and (b) configurations,there is no obvious decrease of the statistic results. In (c) and (d)configurations which are 41% and 14% of the total lines left inthe original network, the accuracy rate decreases. The minimumaccuracy rate is 94.3% when ANFIS2 classifies LLG faults in (d) con-figuration. The results are acceptable, because they are above 90%.

6.4. Different load levels

To study the impact of different load levels on the proposed ap-proach, two load levels are considered which are ‘‘half loading’’ and‘‘double loading’’. The term ‘‘half loading’’ indicates that the loadsat all buses are reduced by 50% uniformly, and the term ‘‘doubleloading’’ indicates that the loads at all buses are doubled uni-formly. The test cases in Table 2 are simulated again under eachload level. The obtained data is used to verify the proposed ap-proach, and the statistic results are shown in Fig. 16.

In comparison with Fig. 9, the subfigure (a) in Fig. 16 shows thatthe decrease of load has little influence on the proposed approach.Moreover, the accuracy is benefit from the light load level, becausethe accuracy rate increases in CG faults and ABG faults. In compar-ison with Fig. 9, the subfigure (b) in Fig. 16 shows that the heavyload decreases the accuracy rate of proposed approach remarkably,especially in SLG faults. The minimum accuracy rate is 83.3% when

99.0

99.2

99.4

99.6

99.8

100.0

Acc

urac

y R

ate

(%

)

99.1

99.3

99.5

99.7

99.9

BGCG

LLGOthersAG

ABGABC/

ABCGACG

BCGAB

ACBC

ANFIS1

ANFIS2

ANFIS3

100 100 100

99.7

100

99.6

100

99.7

100 100 100 100

99.5

(a)

80

84

90

94

98

100

Acc

urac

y R

ate

(%

)

82

86

92

96

BGCG

LLGOthersAG

ABGABC/

ABCGACG

BCGAB

ACBC

ANFIS1

ANFIS2

ANFIS3

89.2 88.9

83.3

99.3100 99.7 100 99.7 99.6 100 99.5 100

88

(b) Fig. 16. Statistic results under different load levels. (a) half loading, and (b) doubleloading.

Table 3Source impedance conditions undertaken for study.

No. Impedance conditions

1 SE + ve seq. impedance magnitude is reduced by 70%, rest unchanged2 SE + ve seq. impedance magnitude is reduced by 70%, rest unchanged3 SE + ve seq. impedance angle is made 20�, rest unchanged4 SE + ve seq. impedance angle is made 50�, rest unchanged5 SE + ve seq. impedance magnitude is doubled, rest unchanged

99.3

99.7

99.5

BGCG

LLGOthersAG

ABGABC/

ABCGACG

BCGAB

ACBC

ANFIS1

ANFIS2

ANFIS3

100

99.4

99.9

100 100

99.7

100.0100

99.5

100 100 100 100

99.4

99.6

99.8

99.9

(a)

99.3

99.7

99.5

99.9

BGCG

LLGOthersAG

ABGABC/

ABCGACG

BCGAB

ACBC

ANFIS1

ANFIS2

ANFIS3

100

99.4

100.0 100

99.7

99.5

100 100 100 100 100 100

99.4

99.6

99.8

99.7

(b)

Acc

urac

y R

ate

(%

)A

ccur

acy

Rat

e (

%)

Fig. 17. Statistic results under different source impedances. (a) condition No. 1, and(b) condition No. 2.

J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252 251

ANFIS1 identifies CG faults. It means 60 out of 360 faults are mis-classified. The result is unacceptable. Fortunately, faults are in-clined to happen when the load are heavy in the field, and thismeans much fault data under heavy load can be obtained in historyand used to train ANFISs. Since the train samples contain a lot ofthose in heavy load level, the performance of ANFIS can be highlyimproved.

6.5. Variance of source impedance

The source impedance depends on the system condition at theSE (Substation End). To study the impact of source impedance onthe proposed approach, the impedance at SE has been varied asoutlined in Table 3 [8].

Under each of impedances given in Table 3, the test cases listedin Table 2 are simulated. Comparison of the results thus obtainedwith those obtained with constant source impedance, the variation

of source impedance has very negligible impact on the proposedapproach. The accuracy rate obtained under condition Nos. 1 and2 is shown in Fig. 17.

Comparison of Fig. 17 with Fig. 9, their statistic results are quitesame. The results obtained under condition Nos. 3, 4 and 5 are assame as those shown in Fig. 17, so they are not repeated here.

7. Conclusion

In this paper, wavelet technique is employed to extract thetransient signals in the interested frequency band. The statisticquantities are used to construct FIs. An ANFIS-based fault classifi-cation approach in distribution network is proposed. Through alarge number of simulation studies, the main conclusions of thiswork are given as follows:

� The constructed FIs can effectively characterize the fault type.� The proposed approach only needs three phase currents and

zero-sequence voltage. It can accurately classify ten types ofshort-circuit faults in distribution. Moreover, it can actualizehigh resistance fault classification. The highest resistance testedin this paper is 1.5 kX.� Although the ANFISs are trained in the isolated distribution, it

can be adaptable to the neutral compensated grounding distri-bution without any change. Besides, the variances of sourceimpedance and fault resistance (arc fault) have little impacton the proposed approach.

252 J. Zhang et al. / Electrical Power and Energy Systems 49 (2013) 243–252

� When the change of network configurations is not intensive, theaccuracy rate of proposed approach is as same as that obtainedin the original configuration. As the network configurationchanges more intensively, the accuracy rate decreases moreobviously. However, the accuracy rate keeps above 90% in ourwork.� With regard to the impact of load level, the light load has little

impact on the proposed approach, but the heavy load decreasesthe accuracy rate remarkably. Fortunately, faults tend to hap-pen when load is heavy. A great deal of history fault data inheavy load can be used to train ANFISs and improves theiraccuracy.

In our future work, the performance of proposed approachshould be improved through increasing the train samples of heavyload from the field. The history fault data in power system will becollected, and used to verify the proposed approach.

References

[1] Upendar J, Gupta CP, Singh GK. Statistical decision-tree based faultclassification scheme for protection of power transmission lines. Int. J. Electr.Power Energy Syst. 2012;36(1):1–12.

[2] Valsan SP, Swarup KS. High-speed fault classification in power lines: theoryand FPGA-based implementation. IEEE Trans. Ind. Electron.2009;56(5):1793–800.

[3] Safty SE, El-Zonkoly A. Applying wavelet entropy principle in faultclassification. Int. J. Electr. Power Energy Syst. 2009;31(10):604–7.

[4] El-Zonkoly AM, Desouki H. Wavelet entropy based algorithm for faultdetection and classification in FACTS compensated transmission line. Int. J.Electr. Power Energy Syst. 2011;33(8):1368–74.

[5] Parikha UB, Dasb B, Maheshwarib R. Fault classification technique for seriescompensated transmission line using support vector machine. Int. J. Electr.Power Energy Syst. 2010;32(6):629–36.

[6] Bhowmika PS, Purkaitb P, Bhattacharyac K. A novel wavelet transform aidedneural network based transmission line fault analysis method. Int. J. Electr.Power Energy Syst. 2009;31(5):213–9.

[7] D.Q. Hou, N. Fischer, Deterministic High-Impedance Fault Detection and PhaseSelection on Ungrounded Distribution Systems, in: IEEE/IAS Ind. & Com. Pow.,Edmonton, vol. 2006, pp. 112–122.

[8] Das B. Fuzzy logic-based fault-type identification in unbalanced radial powerdistribution system. IEEE Trans. Power Del. 2006;21(1):278–85.

[9] Thukaram D, Khincha HP, Vijaynarasimha HP. Artificial neural network andsupport vector machine approach for locating faults in radial distributionsystems. IEEE Trans. Power Del. 2005;20(2):710–21.

[10] Salim RH, Oliveira Kd, Filomena AD, Resener M, Bretas AS. Hybrid faultdiagnosis scheme implementation for power distribution systems automation.IEEE Trans. Power Del 2008;23(4):1846–56.

[11] Dwivedi UD, Singh SN, Srivastava SC. A wavelet based approach forclassification and location of faults in distribution systems. Annual IEEEIndia, vol. 2. KANKUR; 2008. p. 488–93.

[12] Dong XZ, Shi SX. Identifying single-phase-to-ground fault feeder in neutralnoneffectively grounded distribution system using wavelet transform. IEEETrans. Power Del. 2008;23(4):1829–37.

[13] D. Stefanoiu, F. Lonescu, A fuzzy-statistical reasoning model for bearings faultdiagnosis, in: IEEE Systems, Systems Man and Cybernetics, vol. 7, 2002, pp. 6–11.

[14] Erenturk K. ANFIS-based compensation algorithm for current-transformersaturation effects. IEEE Trans. Power Del. 2009;24(1):195–201.

[15] Reddy MJ, Mohanta DK. Adaptive-neuro-fuzzy inference system approach fortransmission line fault classification and location incorporating effects ofpower swings. IET Gener. Transm. Dis. 2008;2(2):235–44.

[16] Etemadi AH, Sanaye-Pasand M. High-impedance fault detection using multi-resolution signal decomposition and adaptive neural fuzzy inference system.IET Gener. Transm. Dis. 2008;2(1):110–8.

[17] Reddy M, Mohanta DK. Performance evaluation of an adaptive-network-basedfuzzy inference system approach for location of faults on transmission linesusing Monte Carlo simulation. IEEE Trans. Fuzzy Syst. 2008;16(4):909–19.

[18] Zhang Y, Zhou Q, Sun CX, Lei SL, Liu YM, Song Y. RBF neural network andANFIS-based short-term load forecasting approach in real-time priceenvironment. IEEE Trans. Power Syst. 2008;23(3):853–8.

[19] Makarand SB, Zafar JK, Hiralal MS, Ram LS. Adaptive neural fuzzy inferencesystem for the detection of inter-turn insulation and bearing wear faults ininduction motor. IEEE Trans. Ind. Electron. 2007;54(4):250–8.

[20] User’s Guide (MATLAB� Compiler™), second ed., Math Works Inc., Natick,2000.

[21] PSCAD/EMTDC User’s manual, Manitoba HVDC Research Centre Inc.,Winnipeg, 2003.

[22] W.H. Kersting, Radial distribution test feeders, in: IEEE Power EngineeringSociety Winter Meeting, Columbus, vol. 2, 2001, pp. 908–912.

[23] Borghetti A, Bosetti M, Silvestro MD, Nucc CA, Paolone M. Continuous-wavelettransform for fault location in distribution power networks: definition ofmother wavelets inferred from fault originated transients. IEEE Trans. Powersyst. 2008;23(2):380–8.

[24] M. Kizilcay, P. Seta, Digital simulation of fault arcs in medium-voltagedistribution networks, in: PSCC 15th Power Systems Computation, Liege, vol.2005, pp. 22–26.

[25] Michalik M, Rebizant W, Lukowicz M, Lee SJ, Kang SH. High-impedance faultdetection in distribution networks with use of wavelet-based algorithm. IEEETrans. Power Del. 2006;21(4):1793–802.