IET Science, Measurement & Technology

Research Article

ISSN 1751-8822
Data-mining-based fault during power swing Received on 29th July 2015
Revised on 18th September 2015
identification in power transmission system Accepted on 29th September 2015
doi: 10.1049/iet-smt.2015.0169
www.ietdl.org

Aleena Swetapadma, Anamika Yadav ✉

Department of Electrical Engineering, National Institute of Technology, Raipur, C.G. 492010, India
✉ E-mail: [email protected]

Abstract: This study proposes a decision-tree-based scheme for detection and classification of fault during power swing in
double circuit transmission lines. The power swing may result due to switching in/out of heavy loads, switching of lines,
clearance of short-circuit faults, generator tripping or load shedding. The proposed decision tree approach makes the
discrimination among no fault situation/power swing and fault during power swing. The fundamental components of
currents and voltages and zero sequence currents measured at only one end of the double circuit line are used as
input to decision tree. To ascertain validity of the proposed scheme, it is tested for variation in fault type, fault
inception angle, fault location and fault resistance. The main advantage of this proposed scheme is that it detects fault
during power swing within half cycle time and classify the type of fault and identify the faulty phase also.

Nomenclature offering supervisory control to the operation of the conventional

distance relays is given in [6].
ANFIS Adaptive Neuro-Fuzzy Inference System An improved unbalanced fault phase selector during a power swing
ANN Artificial Neural Network based on series multi-resolution morphological gradient transform is
CB Circuit Breaker proposed in [7]. If a fault occurs during the power swing, the relay
DFT Discrete Fourier Transform should be able to detect the fault unblock the relay during power
DT Decision Tree swing. The procedure is easy to implement for unsymmetrical faults
IEEE The Institute of Electrical and Electronics Engineers but difficult to identify symmetrical fault during stable power swing.
SVM Support vector machines In [8], a method based on the measured three-phase active and
WT Wavelet Transform reactive powers for detecting symmetrical faults is proposed. A
frequency-domain approach for digital relaying of transmission line
faults mitigating the adverse effects of power swing on conventional
1 Introduction distance relaying is presented in [9]. A wavelet-neuro-fuzzy
combined approach for fault location is also discussed. PSB scheme
The transmission lines are exposed to vagaries of atmospheric has been suggested in [10] based on ANFIS utilising the variations
condition thus most prone to fault. Implemented distance relaying in the measured impedance, current, negative-sequence current and
scheme has problems in case of close in faults, power swings, voltage. A spectral energy function for fault detection during a
overloads, transient conditions, fault resistance, and reach up to 80– power swing using a novel time–frequency transform known as the
85% in first zone due to which other techniques are developed for S-transform which combines the elements of short-time Fourier
enhancing the performance of relaying. Faults in transmission lines transform and wavelet transform is used in [11]. Symmetrical fault
are detected and classified using different methods by different detection during the power swing is a challenge to protection
researchers. Power system network goes through power oscillations schemes. A novel method for detecting symmetrical faults occurring
called as power swings due to switching in/out of heavy loads, during a power swing presented in [12]. A method using differential
switching of lines and clearance of short-circuit faults, generator power is proposed which is calculated using differential components
tripping or load shedding. Power swing is a variation in power flow of voltage and current obtained from the autoregression technique
(variation of voltage and current) which occurs when generator rotor described in [13]. Intelligent detection of unstable power swing for
angles are advancing or retracting relative to each other. Owing to correct distance relay operation using S-transform and neural
variations in voltage and current, the apparent impedance may fall network is proposed in [14].
within the operating characteristic during the power swing which Decision tree is also used for fault classification in power transmission
may cause undesired tripping of transmission lines. Power swing lines [15, 16] but in these schemes the effect of power swing and fault
blocking (PSB) function is used to prevent mal-operation of during power swing has not been considered. Conventional fault
distance relays during swing. In case of distance relaying, a method detection techniques may fail to detect the fault during the power
by monitoring the voltage phase angle at the relay location to swing because of large variations in voltage and current even before
discriminate faults from power swings is proposed in [1, 2] where the fault. Owing to the symmetric nature of signals during the power
only single-phase faults have been simulated. Wavelet transform for swing, symmetrical faults are difficult to be detected. Thus after
detecting faults during power swings, taking advantage of the reviewing various techniques reported till date, there is a need of
presence of high-frequency components created on the voltage at efficient protection technique for transmission lines for detection of
inception of fault is described in [3]. A method using the dc faults during power swing. In this paper, a decision tree approach is
component of three-phase currents, extracted by the Prony method presented for detection of fault during power swing for both
for distance relaying is presented in [4]. Another wavelet-based symmetrical and unsymmetrical faults. Advantage of using decision
symmetrical fault detection method which needs the modal tree is that it can work on large dataset with minimum processing
transform in addition to WT is described in [5]. Classification and time. For splitting criteria Gini index is used which measures the
protection scheme based on support vector machines (SVMs) for impurity of dataset based on probability of different classes.

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
130 & The Institution of Engineering and Technology 2016
 class. Classification tree is built based on the splitting rule.
Different splitting criteria are used for classification from which
Gini splitting rule is widely used. Gini splitting rule uses a
function called impurity function denoted by I(n) for node n.
Impurity at node n is given in (1) according to [17]


k    
I (n) = − p w|i n log p wi |n (1)
i=1

where p(w|in) is the proportion of pattern allocated to class wi at node

n. Impurity function in Gini splitting rule is given by (2)


I (n ) = p(k|n)p(l|n) (2)
k=1

Fig. 1 Block diagram of binary decision tree used for classification where k, l = 1, …, k is the index of class and p(k|n) is the conditional
probability of class k at node n. Gini splitting rule obtains the largest
2 Decision tree and its application class from all the training samples and isolate it from rest of the
samples. After creating the decision tree, next task is to select the
Decision tree can be used as classification and regression algorithm correct tree for classification.
in different fields of pattern classification and pattern recognition. A
decision tree is a set of simple rules based on if–then conditions to 2.2 Choosing optimum tree
classify the data. On the basis of attributes of dataset decisions are
made to classify the data. Decision trees are built using a set of It is often possible to find a simpler tree that performs better than a
data with pre-assigned classes for all observation. It uses questions more complex tree. The decision trees consist of different levels and
to split the data into smaller sets. It uses all variables to find the sub-trees. Optimum tree is chosen by excluding unnecessary nodes
split which separates the data with maximum similarity. Block and sub-trees called as tree pruning. Tree pruning method used in
diagram of decision tree is shown in Fig. 1. Let A1, A2 and A3 be this paper is minimum error pruning. For the decision tree
the attributes of the input datasets and N1, N2 and N3 be the algorithm, the re-substitution error and cross-validation error
values of the attribute that is used for classification. Different estimates the accuracy of training set. It always decreases as the
classes are obtained depending on the decision of relations tree size grows but beyond a certain point, increasing the tree size
between A and N. Decision tree uses different steps to classify a increases the cross-validation error rate. A simple rule would be to
dataset which is described below. choose the tree with the smallest cross-validation error. After
training, test dataset are used for testing the trained network. Each
test case is designed on samples other than the training samples.
2.1 Creation of decision tree The advantages of decision tree is fast training compared with
other pattern recognition tools irrespective of the dataset. Training
Creation of decision tree involves constructing a tree from learning time is only a few seconds with better accuracy even for larger
samples until terminal nodes contain the samples of only one datasets.

Fig. 2 Flow diagram of the decision-tree-based scheme

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
& The Institution of Engineering and Technology 2016 131
Fig. 3 Single line diagram of the proposed power system network

3 Decision-tree-based fault during power swing fault is created on line 2, and after three cycles at 0.1 s, the fault is
detection and classification cleared by opening the respective circuit breakers at both ends of
line 2 to bring the system into power swing mode. At 0.2 s, a
The flowchart of the proposed protection scheme is shown in Fig. 2. fault is created on line 1 to get the situation of a fault during a
Proposed DT-based method first detects the fault during power power swing. A decision-tree-based relay is designed for both
swing. If there is fault, it will identify the fault phases and classify symmetrical and unsymmetrical faults during power swing using
the faults as described in Fig. 2. all the three cases discussed here.

3.1 Power system network used for study 3.3 Fault cases simulation

Using MATLAB 2010a, simulation model of 400 kV transmission The power system network is simulated by varying different fault
system is designed. The single line diagram of the power system parameters to generate different fault cases. For fault during power
network under consideration is shown in Fig. 3. The power system swing, various unsymmetrical faults LG, LLG, LL and
network consists of four synchronous generators of rating 4 × 13.8 symmetrical fault LLL at ten different fault locations (5, 15, 25,
kV, 250 MVA connected to four transformers of rating 4 × 13.8/ …, 95 km) with two fault resistances (0 and 100 Ω) and one fault
400 kV, 250 MVA and double circuit transmission line of length inception angle (0°) have been simulated. Three-phase currents
100 km connected to an infinite bus. At bus1 a 100 kW and 100 and voltages are then obtained at the relaying point for signal
kVar loads are connected. The fault breakers are used to simulate processing/feature extraction. The total number of fault cases used
various shunt faults with a specific fault inception angle and fault for testing is about 10,000 as given in Table 1.
resistance.
3.4 Feature extraction
3.2 Different cases of power swing under study
Feature extraction is necessary for implementing the
Different case studies which lead the power system into power swing decision-tree-based network. Input features given to the DT-based
mode are discussed below, which have been considered to evaluate network are the fundamental components of voltages, currents and
the performance of the proposed method for detection of fault during zero sequence components of currents of both circuits.
power swing. Three-phase currents and voltages are processed using discrete
Fourier transform block. It computes fundamental value of the
3.2.1 CASE I: Addition of load: When an extra load is switched input signal over a running window of one cycle of specified
in/added to the existing system, there is a change in power flow fundamental frequency. Zero sequence components of currents are
which leads to power swing. In the proposed method, a load 200 obtained by using the sequence analyser block.
MW is switched into the system, which causes increase in current,
voltage as well as power flow. In this situation, the conventional
distance relaying scheme may mal-operate. Hence some PSB 3.5 DT-based method for fault during power swing
devices are used to prevent mal-operation of relay and thereby detection
opening of the circuit breakers in power swing conditions.
However, there is a chance that a fault may occur in power swing After feature extraction next task is to create a training dataset for
mode and it is necessary to detect the fault during power swing. fault during power swing detection. Fault detection training
module is designed by taking samples of no faults, power swing
3.2.2 CASE II: Loss of synchronization in frequency: and different fault cases. Ten number of post fault samples of
Sometimes there is loss of synchronism between frequencies, that fundamental voltage and current signals for each fault case are
is, the frequency of a generator deviates from the nominal power extracted to form the input matrix for training of DT-based fault
system frequency. For example, say the original system is 400 kV
50 Hz but the frequency of the synchronous generator may Table 1 Cases of faults for both training and testing
decrease or increase due to switching in/out of heavy loads,
Parameters Training Testing
respectively. Consequently, the current signal also increases/
decreases which may be considered as a fault and may trip the
fault type LG, LLG, LL and LLL LG, LLG, LL and LLL
CB. In this situation, a fault may occur and it may consider as fault location ten locations: 5, 15, 25, 60 locations
power swing. …, 95 km
fault resistance 2R–0 and 100 Ω 0, 20, 40, 60, 80 and 100
3.2.3 CASE III: After fault clearance: The power system model Ω
fault inception 0° 0, 45, 90, 135, 180, 225,
in this case includes parallel transmission lines of 100 km connected angle 270, 315
to the source of 1000 MVA and infinite bus. At 0.04 s, a three-phase

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
132 & The Institution of Engineering and Technology 2016
Fig. 4 Decision tree for fault detection in fault during power swing

detection during power swing. Total 20 no fault samples and 240 (xii) class = F.
power swing samples are given to the training network along with (xiii) class = NF.
faulted samples. Thus the network inputs for fault during power (xiv) if x5 < 2013.49 then node 16 elseif x5> = 2013.49 then node 17
swing are fundamental voltage and current signals as given in (3). else F.
The output will be ‘1’ if no fault (NF) or power swing, ‘2’ if there (xv) class = NF.
is fault (F) which is shown in (4) (xvi) if x1 < 1911.31 then node 18 elseif x1> = 1911.31 then node 19
else F.
  (xvii) class = NF.
input = Vaf , Vbf , Vcf , Iaf 1 , Ibf 1 , Icf 1 , Iaf 2 , Ibf 2 , Icf 2 (3) (xviii) class = NF.
(xix) if x5 < 2009.56 then node 20 elseif x5> = 2009.56 then node 21
output = [‘1’ or ‘2’] (4) else F.
(xx) if x7 < 292,723 then node 22 elseif x7> = 292,723 then node 23
else F.
Here a decision-tree-based algorithm is proposed to detect fault
(xxi) if x4 < 1968.56 then node 24 elseif x4> = 1968.56 then node 25
during power swing so that the CB will operate correctly. Decision
else F.
tree used in this case are binary trees which decide using a
(xxii) if x3 < 1991.06 then node 26 elseif x3> = 1991.06 then node 27
particular condition whether the output will be ‘1’ or ‘2’. In this
else F.
case, the decision tree takes five levels to decide fault or no fault
(xxiii) class = F.
using the nine input attributes. The optimal decision tree obtained
(xxiv) class = F.
after training for fault during power swing detection is shown in
(xxv) class = NF.
Fig. 4. In Fig. 4, inputs x1, x2, …, x6 are the current signals of
(xxvi) class = F.
both circuits and x7, x8 and x9 are the voltage signals. Decision
(xxvii) if x1 < 2042.14 then node 28 elseif x1> = 2042.14 then node
tree is constructed using value of attributes according to splitting
29 else NF.
criteria. ‘NF’ represents the no fault situation and ‘F’ represents
(xxviii)class = NF.
the fault situation in the decision tree. Decision tree created for
(xxix) class = F.
fault during power swing detection can be shown as follows where
x1 = Ia1f, x2 = Ib1f, x3 = Ic1f, x4 = Ia2f, x5 = Ib2f, x6 = Ic2f, x7 = Vaf, x8 =
Vbf and x9 = Vcf:
Confusion matrix for training shown in Table 2 shows the
predicted value against the actual value. Diagonal elements contain
(i) if x2 < 1802.41 then node 2 elseif x2> = 1802.41 then node 3
the correctly classified data in all classes. Training accuracy of the
else F.
proposed decision tree method is 100%. Training time required for
(ii) if x1 < 1818.92 then node 4 elseif x1> = 1818.92 then node 5
detection of fault during power swing using DT is a few seconds,
else NF.
better than training time of other data-mining methods such as
(iii) if x8 < 264,744 then node 6 elseif x8> = 264,744 then node 7
ANN and SVM. After training the network, test fault cases
else F.
generated as per Table 1 are tested against the trained network to
(iv) class = NF.
check the performance of the network. If there is fault in system,
(v) class = F.
fault phase(s) are identified and classified.
(vi) class = F.
(vii) if x5 < 2140.72 then node 8 elseif x5> = 2140.72 then node 9
else F. 3.6 DT-based method for fault classification
(viii) if x7 < 276,519 then node 10 elseif x7> = 276,519 then node 11
else F. After fault during power swing detection, next task is to classify the
(ix) if x3 < 2062.92 then node 12 elseif x3> = 2062.92 then node 13 faults using DT-based network. Fault classification module is
else NF. designed individually for each phase of double circuit line DT-A1,
(x) class = F. DT-B1, DT-C1, DT-A2, DT-B2, DT-C2 and one for ground G
(xi) if x9 < 294,552 then node 14 elseif x9> = 294,552 then node 15 detection. Inputs to the fault classification module are fundamental
else F. components and zero sequence components of current signals. The

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
& The Institution of Engineering and Technology 2016 133
Table 2 Confusion matrix and accuracy of fault during power swing network inputs and outputs for classification are shown in (5) and
detection using DT (6). The output will be ‘1’ if no fault or power swings and ‘2’ if
Predicted value Actual value Training Training there is fault in any phase. Confusion matrix and training accuracy
for the entire fault classification module is shown in Table 3.
No fault/ Fault during Accuracy Time, s Actual class and predicted class of different phases during training
power swing power swing are given in Table 3. Overall accuracy of classification of faults
during training is 99.9%. Fig. 5 shows the optimal decision tree
no fault/ 260 0 100% 1.346 obtained for each phase for fault during power swing for fault
power swing
fault during 0 6000 – – phase identification. Figs. 5a–g show the optimal decision tree
power swing obtained for each phase A1, B1, C1, A2, B2, C2 and G. ‘NF’
represents the no fault situation and ‘F’ represents the fault
situation in the decision tree
 
Table 3 Confusion matrix and training accuracy of fault during power input = Ia f1 , Ib f1 , Ic f1 , Ia f2 , Ib f2 , Ic f2 , Iz1 , Iz2 (5)
swing classification for each phase

Classifiers Predicted Class Actual Class Training Training output = ‘1’ or ‘2’ (6)
No fault Fault Accuracy Time, s

DT-A1 no fault 260 0 100% 1.795 Decision tree for fault during power swing classification is
fault 0 6000 – – described below:
DT-B1 no fault 260 10 99.84% 1.317 Decision tree for Phase A1 where x1 = ia1:
fault 0 5990 – –
DT-C1 no fault 260 0 100% 1.335
fault 0 6000 – – (i) if x1 < 2102.03 then node 2 elseif x1> = 2102.03 then node 3
DT-A2 no fault 260 2 99.968% 1.671 else NF.
fault 0 5998 – – (ii) class = NF.
DT-B2 no fault 260 14 99.776% 1.289
fault 0 5986 – –
(iii) class = F.
DT-C2 no fault 260 7 99.888% 1.741
fault 0 5995 – – Decision tree for Phase B1 where x1 = ib1:
DT-G no fault 260 0 100% 1.892
fault 0 6000 – –
(i) if x1 < 1875.99 then node 2 elseif x1> = 1875.99 then node 3
else NF.

Fig. 5 Decision tree for fault phase identification for fault during power swing
a A1 phase
b B1 phase
c C1 phase
d A2 phase
e B2 phase
f C2 phase
g Ground G

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
134 & The Institution of Engineering and Technology 2016
Fig. 6
a Three-phase currents of circuit 1 of CASE I
b Output of DT-based fault detector of CASE I
c Three-phase currents of circuit 1 of CASE II
d Output of DT-based fault detector of CASE II
e Three-phase currents of circuit 1 of CASE III
f Output of DT-based fault detector of CASE III during power swing after 0.1 s

(ii) class = NF. (iii) class = F.

(iii) class = F.
Decision tree for Phase C2 where x1 = ic2:
Decision tree for Phase C1 where x1 = ic1:

(i) if x1 < 1649.77 then node 2 elseif x1> = 1649.77 then node 3 (i) if x1 < 1649.71 then node 2 elseif x1> = 1649.71 then node 3
else NF. else NF.
(ii) class = NF. (ii) class = NF.
(iii) class = F. (iii) class = F.

Decision tree for Phase A2 where x1 = ia2: Decision tree for Phase G where x1 = ix1 and x2 = iz2:

(i) if x1 < 2101.89 then node 2 elseif x1> = 2101.89 then node 3 (i) if x1 < 197.81 then node 2 elseif x1> = 197.81 then node 3 else
else NF. NF.
(ii) class = NF. (ii) if x1 < 34.9237 then node 4 elseif x1> = 34.9237 then node 5
(iii) class = F. else NF.
(iii) class = F.
Decision tree for Phase B2 where x1 = ib2: (iv) class = NF.
(v) if x2 < 80.8616 then node 6 elseif x2> = 80.8616 then node 7
(i) if x1 < 1874.74 then node 2 elseif x1> = 1874.74 then node 3 else NF.
else NF. (vi) class = F.
(ii) class = NF. (vii) class = NF.

Table 4 Fault detection time for symmetrical fault A1B1C1 in circuit 1

Fault location, km Fault resistance, Ω Fault inception angle, deg Fault cases Percentage accuracy Fault detection time, ms

Detection, % Classification, % Detection Classification

11–91 km in step of 10 0–10 in step of 2 0–360 in step of 45 CASE I 100 100 within<10 ms within<10 ms
– – – CASE II 100 100 within<10 ms within<10 ms
– – – CASE III 100 99.99 within<10 ms within<10 ms

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
& The Institution of Engineering and Technology 2016 135
Fig. 7
a Three-phase currents of circuit 1 of CASE I
b Output of DT-based fault detector of CASE I
c Three-phase currents of circuit 1 of CASE II
d Output of DT-based fault detector of CASE II
e Three-phase currents of circuit 1 of CASE III
f Output of DT-based fault detector of CASE III during A1B1C1 fault in circuit 1 at 94 km with Rf = 0Ω, Φi = 0° (ti = 0.2 s)

Fig. 8 Output of DT-based fault classification during

a A1B1C1 fault in circuit 1 of CASE I
b A1B1C1 fault in circuit 1 of CASE III at 95 km with Rf = 0Ω, Φi = 0° (ti = 0.2 s)

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
136 & The Institution of Engineering and Technology 2016
Table 5 Fault detection time for unsymmetrical fault
Fault location, km Fault resistance, Ω Fault inception angle, deg Fault cases Percentage accuracy Fault detection time, ms

Detection, % Classification, % Detection Classification

11–91 km in step of 10 0–10 Ω in step of 2 Ω 0–360 in step of 45 CASE I 100 100 within<10 ms within<10 ms
– – – CASE II 100 99.99 within<10 ms within<10 ms
– – – CASE III 100 99.99 within<10 ms within<10 ms

4 Results and discussion CASE III, respectively, when a fault occurs at 0.2 s or 200 ms
during power swing. Fig. 7b shows the output of fault during
The DT-based relay are required to be tested for all types of fault and power swing detector for CASE I which is low (0) and up to
variations in fault and network parameters. The test results show that 200 ms shows that there is no fault in the system although there is
fault during power swing is detected in less than half cycle time. The power swing in the system. After 200 ms output becomes high (1)
proposed DT-based fault detection time is estimated in all three at 205 ms and shows that there is fault in the system during power
power swing cases. The proposed scheme is also tested under swing. Therefore, fault detection time in this case is 205–200 =
different faulted conditions such as varying faults inception angle, 5 ms. Fig. 7d shows the output of fault during power swing
fault with high fault resistance, fault location, fault type etc. detector for CASE II. Fig. 7f shows the output of fault during
power swing detector for CASE III.
Faults are also classified in case of fault during power swing.
4.1 Performance of relay during power swing Output of DT-based fault classification during A1B1C1 fault in
circuit 1 at 94 km with Rf = 0 Ω, Φi = 0° is shown in Fig. 8.
To evaluate the operating time of the proposed DT-based relay, load Fig. 8a shows A1B1C1 fault in circuit 1 of CASE I. Phases A1,
increased to 200 MW in first case, frequency is changed to 51 Hz in B1 and C1 go high (1) after 207 ms, whereas other phases of
second case and at 0.04 s a three-phase fault is created on line 2, and circuit 2 are low (0) confirming that fault type is A1B1C1. Fig. 8b
after three cycles at 0.1 s, the fault is cleared by opening the shows during CASE III, A1B1C1 fault in circuit 1 at 94 km with
respective circuit breakers at both ends of line 2 to bring the Rf = 0 Ω, Φi = 0° (ti = 0.2 s). Phases A1, B1 and C1 go high (1)
system into power swing mode for the third case. The input after 200 ms, whereas other phases are low (0) and show that fault
current of circuit 1 and outputs during power swing are shown in type is A1B1C1. Results of some of the test cases are shown in
Fig. 6. Figs. 6a, c and e show currents of all phases of the circuit Table 4. Percentage accuracy of fault detection scheme is 100%
1 increase after power swing occur at 0.1 s in CASE I, CASE II and classification is 99.99% for all symmetrical faults tested.
and CASE III, respectively. Figs. 6b and d show that output of the
proposed DT-based relay is ‘1’ for case I and case II depicting it
as no fault condition, so it do not send trip to CB. Fig. 6f shows 4.3 Performance in case of fault during power swing for
‘2’ for duration 0.04–0.1 s during fault condition, thereafter unsymmetrical fault
clearance of fault brings the system to power swing mode. After
Proposed DT-based relay is also tested for unsymmetrical faults LL,
0.1 s, it shows output of ‘1’ showing that system is in no fault
LG, LLG and the results are given in Table 5. From Table 5, it can be
condition so that there will be no trip to CB.
observed that the fault detection and classification time for all
unsymmetrical fault cases are within half cycle time. Faults are
4.2 Performance in case of fault during power swing for also classified in case of fault during power swing in case of
symmetrical fault unsymmetrical faults.

Symmetrical faults are not easy to detect when the system is in power 4.4 Testing of the proposed method in IEEE 9-bus
swing mode. Proposed decision-tree-based relay tested for fault
system
during power swing in case of symmetrical faults for different
location, fault resistance and inception angle. Fault detection times Proposed data-mining-based method is tested with standard IEEE
are given in Table 4 for different fault cases for all three cases of 9-bus system. Line diagram of IEEE 9-bus system is shown in
power swing. The A1B1C1 fault case at 94 km from the relay Fig. 9. Relay is placed near to bus-8, so section between bus-8
location with Rf = 0 Ω at 0.2 s is shown in Fig. 7. Figs. 7a, c and e and bus-9 is considered as section to be protected. Different fault
show the three-phase currents of circuit 1 of CASE I, CASE II and situations are simulated by varying different fault parameters such
as fault types: LG, LL, LLG and LLL, fault locations: 5–95 km in
each section with a step of 20 km, fault inception angles: 0° and
fault resistances: 0, 50 and 100 Ω. Signal processing is done using
DFT same as above-described system. Design process for this
system is same as described above. The fundamental components
of each phase currents have been used as inputs to the fault
detection network and fault phase identification network and input
to ground identification network is zero sequence current. After
training the three DT networks for fault detection, fault phase
identification and ground identification, different test fault cases
are generated and tested against the trained network to check the
performance of the network.
Say a three-phase fault: ABC fault has been simulated in line
section between bus 8 and bus 9 at 91 km from relay location with
Rf = 0 Ω, Φi = 0° (ti = 200 ms) with +40% increase in load at 0.06 s
time. At 60 ms time, additional 40% load is connected at bus 8,
due to which system comes into power swing mode and the
current magnitude in all the three phases increases as shown in
Fig. 10a. When system is in power swing mode, a symmetrical
Fig. 9 Single line diagram of IEEE 9-bus system three-phase fault ABC is created at 200 ms time. Fig. 10b shows

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
& The Institution of Engineering and Technology 2016 137
Fig. 10 During +40% increase in load at 60 ms time at bus-8 of IEEE 9-bus system and ABC fault at 91 km with Rf = 0Ω, Φi = 0° (ti = 200 ms)
a Three-phase currents
b Output of DT-based fault detector
c Output of DT-based fault phase identifier, Phase A
d Phase B
e Phase C
f Ground G

Table 6 Test results of standard IEEE 9-bus system

Situations Increase in load at bus 8, % Fault type Fault location, km Fault resistance, Ω Detection output Time, ms

power swing +5 – – – 1 –
+15 – – – 1 –
+25 – – – 1 –
+35 – – – 1 –
fault during power swing +5 AB 3 1 2 3
+15 BC 13 5 2 3
+25 CA 23 10 2 4
+35 ABC 33 10 2 3
+5 AG 43 25 2 3
+15 BG 53 45 2 3
+25 CG 63 65 2 6
+35 ABG 73 75 2 3
+5 BCG 83 85 2 3
+15 CAG 93 95 2 3

the fault detection output of DT-based scheme which is ‘1’ up to fundamental components of current and zero sequence component
200 ms time, illustrating that system in no-fault condition. After of current are taken as input. Advantage of decision tree is that it
202 ms, the output becomes ‘2’ confirming that there is fault in is not affected by the size of dataset used for training. The
the system. Similarly Figs. 10c–f show the output of fault phase reliability of this scheme is not affected by different fault
identification modules A, B, C and ground G respectively. Some conditions such as variation in fault type, fault distance, fault
of the test results of the proposed scheme are shown in Table 6 for inception angle, fault resistance etc. Results demonstrate that the
power swing and fault during power swing cases. From Table 6, it proposed scheme effectively detects the fault and the fault type
can be observed that power swing and fault during power swing within half cycle time. Accuracy of the proposed method is
can be identified in standard IEEE 9-bus system using the 99.99% for all the tested fault cases.
proposed method accurately.

6 References
5 Conclusion
1 Mechraoui, A., Thomas, D.W.P.: ‘A new blocking principle with phase and earth
Decision-tree-based fault during power swing detection and fault detection during fast power swings for distance protection’, IEEE Trans.
Power Deliv., 1995, 10, pp. 1242–1248
classification scheme is proposed in this paper. Fundamental 2 Mechraoui, A., Thomas, D.W.P.: ‘A new principle for high resistance earth fault
components of current and voltages are taken as inputs to decision detection during fast power swings for distance protection’, IEEE Trans. Power
tree for fault during power swing detection. For fault classification, Deliv., 1997, 12, pp. 1452–1457

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
138 & The Institution of Engineering and Technology 2016
 3 Brahma, S.M.: ‘Distance relay with out-of-step blocking function using wavelet 11 Samantaray, S.R., Dubey, R.K., Chitti Babu, B.: ‘A novel time–frequency
transform’, IEEE Trans. Power Deliv., 2007, 22, pp. 1360–1366 transform based spectral energy function for fault detection during power
4 Lotfifard, S., Faiz, J., Kezunovic, M.: ‘Detection of symmetrical faults by distance swing’, Electr. Power Compon. Syst., 2012, 40, pp. 881–897
relays during power swings’, IEEE Trans. Power Deliv., 2010, 25, pp. 81–87 12 Mahamedi, B., Zhu, J.G.: ‘A novel approach to detect symmetrical faults occurring
5 Pang, C., Kezunovic, M.: ‘Fast distance relay scheme for detecting symmetrical during power swings by using frequency components of instantaneous three-phase
fault during power swing’, IEEE Trans. Power Deliv., 2010, 25, pp. 2205–2212 active power’, IEEE Trans. Power Deliv., 2012, 27, pp. 1368–1376
6 Seethalekshmi, K., Singh, S.N., Srivastava, S.C.: ‘A classification approach using 13 Rao, J.G., Pradhan, A.K.: ‘Differential power-based symmetrical fault detection
support vector machines to prevent distance relay maloperation under power swing during power swing’, IEEE Trans. Power Deliv., 2012, 27, pp. 1557–1564
and voltage instability’, IEEE Trans. Power Deliv., 2012, 27, pp. 1124–1133 14 Abidin, A.F., Mohamed, A., Shareef, H.: ‘Intelligent detection of unstable power
7 Zou, L., Zhao, Q., Lin, X., et al.: ‘Improved phase selector for unbalanced faults swing for correct distance relay operation using S-transform and neural
during power swings using morphological technique’, IEEE Trans. Power network’, Expert Syst. Appl., 2011, 38, pp. 14969–14975
Deliv., 2006, 21, pp. 1847–1855 15 Jamehbozorg, A., Shahrtash, S.M.: ‘A decision tree method for fault classification
8 Lin, X., Gao, Y., Liu, P.: ‘A novel scheme to identify symmetrical faults occurring in single circuit transmission lines’, IEEE Trans. Power Deliv., 2011, 25,
during power swings’, IEEE Trans. Power Deliv., 2008, 23, pp. 73–78 pp. 2190–2196
9 Reddy, M.J., Mohanta, D.K.: ‘Adaptive-neuro-fuzzy inference system approach for 16 Upender, J., Gupta, C.P., Singh, G.K.: ‘Statistical decision tree based fault
transmission line fault classification and location incorporating effects of power classification scheme for protection of power transmission lines’, Electr. Power
swings’, IET Gener. Transm. Distrib., 2008, 2, pp. 235–244 Energy Syst., 2012, 36, pp. 1–12
10 Zadeh, H.K., Li, Z.: ‘A novel power swing blocking scheme using adaptive 17 Breiman, L., Friedman, J.H., Olshen, R.A., et al.: ‘Classification and regression
neuro-fuzzy inference system’, Electr. Power Syst. Res., 2008, 78, pp. 1138–1146 trees’, (Chapman and Hall, London (UK), 1984)

IET Sci. Meas. Technol., 2016, Vol. 10, Iss. 2, pp. 130–139
& The Institution of Engineering and Technology 2016 139