J. Cent. South Univ. Technol.

(2011) 18: 190195

DOI: 10.1007/s1177101106795

Power interconnected system clustering with

advanced fuzzy C-mean algorithm
WANG Hong-mei()1, KIM Jae-Hyung1, JUNG Dong-Yean2, LEE Sang-Min3, LEE Sang-Hyuk3
1. School of Mechatronics, Changwon National University, Changwon 641-773, Korea;
2. Daeho Tech Co. Ltd., Changwon 641-465, Korea;
3. Department of Electronics Engineering, Inha University, Incheon 402-751, Korea
Central South University Press and Springer-Verlag Berlin Heidelberg 2011
Abstract: An advanced fuzzy C-mean (FCM) algorithm was proposed for the efficient regional clustering of multi-nodes
interconnected systems. Due to various locational prices and regional coherencies for each node and point, modified similarity
measure was considered to gather nodes having similar characteristics. The similarity measure was needed to contain locational
prices as well as regional coherency. In order to consider the two properties simultaneously, distance measure of fuzzy C-mean
algorithm had to be modified. Regional clustering algorithm for interconnected power systems was designed based on the modified
fuzzy C-mean algorithm. The proposed algorithm produces proper classification for the interconnected power system and the results
are demonstrated in the example of IEEE 39-bus interconnected electricity system.
Key words: fuzzy C-mean; similarity measure; distance measure; interconnected system; clustering

1 Introduction
In the regional management of interconnected
network systems, the efficient and economical operation
of the networked systems in terms of system coherency
was essential [1]. Hence, the research of system
coherency has been made by numerous researchers [23].
However, most of the studies were emphasized on the
dynamic grouping [4]. Hence, a novel approach to
partition the total system into several regions considering
locational information, such as locational cost, loss and
regional distances, was needed.
Grouping the locations in a networked system with
similar locational prices has been proposed considering
the regional coherency. Locational costs in a networked
system imply the price at which the good was consumed
at each location. Due to the physical characteristics of
the transmission network of the systems, the good was
lost as it was transmitted from supplying locations to
consuming locations, and an additional supply must be
provided to compensate for the loss. Also, the
transmission network of the systems has a capacity
limitation preventing full uses of cheap production.
Therefore, locational prices at each point or node, was
differently decided depending on the network topology

and supplydemand configuration.

Similarity measure between data was introduced by
LIU [5]. Later, explicit expression of fuzzy entropy and
similarity measure were also followed [67]. Furthermore,
relation between fuzzy entropy and similarity measure
was shown continuously [89]. To consider locational
prices and regional coherency simultaneously, it was
needed to extend the conventional similarity measure.
Hence, to design the measure of two properties not only
locational prices but also regional coherency should be
considered. In the previous researches, similarity
measure was designed through distance measure or fuzzy
entropy [10]. Well-known Hamming distance was also
used to construct similarity measure. With only
conventional similarity measure, unpractical results
could be followed. Emphasizing the only way of
locational prices or regional coherency, results were
originated by the structure of single variable similarity
measure. Hence, the addition of regional coherency made
it possible to complete the modified similarity measure.
Then, fuzzy C-mean (FCM) was introduced, and
similarity measure was proposed. Proposed similarity
measure was designed by the distance measure.
Similarity measure was modified with additional
regional coherency to apply for power interconnected
system. And illustrative example was shown.

Foundation item: Work supported by the Second Stage of Brain Korea 21 Projects; Work(2010-0020163) supported by Priority Research Centers Program
through the National Research Foundation (NRF) funded by the Ministry of Education, Science and Technology of Korea
Received date: 20100328; Accepted date: 20100624
Corresponding author: LEE Sang-Hyuk, PhD, Professor; Tel: +82552133884; E-mail: leehyuk@inha.ac.kr

J. Cent. South Univ. Technol. (2011) 18: 190195

191

2 Fuzzy C-mean clustering

Fuzzy C-means clustering was proposed by
BEZDEK in 1973 as an improvement over hard C-mean
(HCM) [10]. FCM played a roll of partitioning arbitrary
n vectors into c fuzzy groups, also finding a cluster
center for each group such that a cost function of
similarity measure was maximized, or dissimilarity
measure was minimized. Well-known facts about FCM
and HCM indicated that FCM employed fuzzy
partitioning such that a data point could belong to several
groups with the degree of membership grade between 0
and 1.
2.1 Preliminaries
Fuzzy C-mean was introduced briefly as follows.
These results could be found in Ref.[11]. Membership
matrix U was defined as
c

uij = 1, j = 1,

L, n

i =1

(1)

The cost function for FCM was constructed by

c

J (U , c1 ,L , cc ) = J i = uijm dij2
i =1

(2)

i =1 j =1

where uij is denoted between 0 and 1, ci is also the center

of fuzzy group i, dij=|cixj| satisfies the Euclidean
distance between i-th cluster center and the j-th data
point xj, and m is defined by the weighting value. With
Lagrange multiplier, the necessary conditions for Eq.(2)
to reach a minimum were found in Ref.[11] as
n

ci =

j =1
n

uijm x j

j =1

and uij =
uijm

1
c

dij

(d
k =1

2.2 Similarity measure with distance function

To define the similarity between data sets, it was
essential to consider similarity measure. Similarity
measure was proposed by LIU [5]. Axiomatic properties
of the similarity measure were proposed in Definition 1.
Of course, there were lots of similarity measures that
satisfied Definition 1 [1213].
Definition 1: A real function s: P2R+ or F2R+ is
a similarity measure if s has the following properties:

(S1) s(A, B)=S(B, A), A, B F ( X )

(S2) s(D, DC)=0, D P( X )
(S3) s(C, C)= max A, BF s ( A, B), C F ( X )
(S4) A, B, C F ( X ), if A B C, then s(A, B)
s(A, C) and s(B, C)s(A, C)
where F(X) denotes a fuzzy set and P(X) is a numeric set.
Similarity measure was proposed with the distance
measure as follows.
Theorem 1: For any set A, B F(X) or P(X) if d
satisfies Hamming distance measure, then
s(A, B)=42d((A I B), [1]X)2d((A U B), [0]X)

(3)

is a similarity measure between set A and B.

Proof 1: Proof was followed by proving that Eq.(3)
satisfied the similarity definition. (S1) means the
commutativity of set A and B, hence it was clear from
Eq.(3) itself.
From (S2),
s(A, AC)=42d((A I AC), [1]X)2d((A U AC), [0]X)
then 2d((A I AC), [1]X) and 2d((A U AC), [0]X) were the
maximum values between A and arbitrary set. For
arbitrary sets A and B, inequality of (S3) was obtained by
s(A, B)=42d((A I B), [1]X)2d((A U B), [0]X)
42d((D I D), [1]X)2d((D U D), [0]X)=s(D, D)
Inequality was satisfied from

2 /( m 1)

kj

With these results, well-known FCM algorithms

could be listed as
Step 1: Initialize membership matrix U.
Step 2: Calculate ci, i=1, , c.
Step 3: Computate Eq.(2). Stop if either it is below a
certain tolerance.
Step 4: Computate a new U.
Now for minimizing Eq.(2), the less the distance
dij=|cixj| was, the smaller the cost function became.
Hence, distance values have a relation to the similarity
between two data points. Finding similarity could be
determined from the types of data, time series signal,
image, sound, etc. Then, similarity measure for the fuzzy
sets was introduced, and the proposed similarity measure
could be applied to the power interconnected problem.

d((A I B), [1]X)d((D I D), [1]X) and

d((A U B), [0]X)d((D U D), [0]X)
Finally, A, B, C F ( X ), if A B C,
s(A, B)=42d((A I B), [1]X)2d((A U B), [0]X)=
42d(A, [1]X)2d(B, [0]X)
42d(A, [1]X)2d(C, [0]X)=s(A, C)
Similarly, s(B, C)s(A, C) was obtained easily through
d(B, [0]X)d(C, [0]X) and d(B, [1]X)d(A, [1]X).
Therefore, proposed similarity measure Eq.(3)
satisfied similarity measure. Similarly, numerous
similarity measures could be obtained, and it was also
designed with the distance measure.
Theorem 2: For any set A, B F(X) or P(X), if d
satisfies Hamming distance measure, then
s(A, B)=22d((A I BC), [0]X)2d((A U BC), [1]X)
is also similarity measure between set A and B.

(4)

J. Cent. South Univ. Technol. (2011) 18: 190195

192

Proof 2: Proofs were also followed similarly as

Theorem 1.

3 Similarity consideration with additional

information
To apply FCM with dij=|cixj|, it was required that
dij had to satisfy the similarity property. Hence, dij was
considered the modified similarity measure in Eq.(2).
The proposed similarity measure has to be gathered for
the point that has similar characteristic values. However,
the large power connected system has multiple
characteristics for each node. For example, locational
prices and geometric information (location) were the
representative characteristics. It could happen that two
nodes had very similar locational prices even they were
located far away. Then, it was not realistic to gather even
though they had similar valued measure. To overcome
this drawback, another similarity measure had to be
required to gather data properly. With Eqs.(3) and (4),
similarity measure was considered:
s2(A, B)=2/(1+d)

(5)

where d denotes the geometrical distance value.

Combining similarity measure as
s(A, B)=1s1(A, B)+2s2(A, B)

(6)

where s1(A, B)=42d((A I B), [1]X)2d((A U B), [0]X), 1

and 2 are the weighting values.
Usefulness of Eq.(6) was verified directly, and
properties of s1(A, B) were proved in Theorem 1 and
Theorem 2. Usefulness for the similarity of s2(A, B)
could be verified as follows:
Commutative property of distance was the same,
hence (S1) was easily shown. From (S2), distance of A
and AC was the longest, hence s2(A, AC) satisfied the
minimum value. For all A, B P(X), inequality of (S3)
was proved by
s2(A, B)=2/(1+d(A, B))2/(1+d(D, D))=s2(D, D)
In the above, d(D, D) is the smallest value, i.e., zero.
So (S3) could be verified.
Finally, A, B, C F ( X ) and A B C,
s2(A, B)=2/(1+d(A, B))2/(1+d(A, C))=s2(A, C)
where d(A, C) was longer than d(A, B). Similarly,
s2(B, C)=2/(1+d(B, C))2/(1+d(A, C))=s2(A, C)
was also satisfied. Hence, it could be verified that s2(A, B)
satisfied Definition 1.
Then, similarity measure extension for the multiple
facts could be possible. Furthermore, more flexible
design for multiple facts could also be possible by
adjusting the weighting values:

s2(A, B)=1s1(A, B)+2s2(A, B)++nsn(A, B)=

n

i si ( A, B)
i =1

As mentioned before, for the power interconnected

system clustering problem, multiple considerations have
to be followed. Because each node has multiple
representative values such as geometric information
(location), locational prices or other values could be
followed. Hence, the extension of similarity measure
would be useful to evaluate the similarity of complex
node information.
Then, Eq.(6) could be used as the similarity
measure for the particular points which have locational
prices and geometric information at the same time.

4 Illustrative example
As an illustrative application, the interconnected
electricity system was considered. The IEEE reliability
test system which was prepared by the reliability test
system task force of the application of probability
methods subcommittee in 1996 [14] is considered a test
system. In the test system, 39 nodes (buses) and 10
generators are contained and each bus has its own
locational price and information.
In networked electricity systems, due to the physical
characteristics of the electricity transmission network,
electricity is lost when it is transmitted from supplying
nodes (i.e., supplying buses) to consuming nodes
(consuming buses), and additional generation must be
supplied to provide energy in excess of that consumed by
customers. Moreover, the capacity limitation of the
transmission network of electricity systems prevents full
uses of system wide cheap electricity. Therefore, the
electricity price at each node, i.e., the price at which the
electricity is consumed at each node is differently
decided depending on the network topology and energy
configuration [15].
The electricity prices at each node are defined as
locational prices at each node and the locational prices
represent the locational value of energy, which includes
the cost of electricity and the cost of delivering it, i.e, the
delivery losses and network congestion. In Table 1, each
locational price is illustrated for the 39 buses, and per
unit geometrical information for nodes is also shown.
Locational prices of each nodes were from
24.98 $/(kWh) to 55.00 $/(kWh), and 39 locational
information is represented through 2-dimensional plane
at which the plane is assumed to be flat. In order to apply
FCM for the clustering problem of Fig.1, replacing s(A,
B) in Eq.(6) into dij=|cixj| in Eq.(2), then the similarity
of locational prices and locations can be assigned by s1(A,
B) and s2(A, B), respectively. Combined similarity
measures are constituted as follows with the proposed

J. Cent. South Univ. Technol. (2011) 18: 190195

193

Table 1 Locational prices and location (normalized geometric information) for each bus
Bus
No.

Locational
price/($kW1h1)

Location

Bus
No.

Locational
price/($kW1h1)

Location

Bus
No.

Locational
price/($kW1h1)

Location

29.21

( 0.9, 9 )

14

41.74

(6.6, 6)

27

51.45

(4.6, 3.5)

28.53

(0.6, 6.2)

15

43.79

(6.6, 4.9)

28

55.00

(2.7, 1.5)

31.40

(3, 7.5)

16

45.84

(6.5, 4)

29

55.00

(2.7, 0.8)

32.78

(4.7, 7.5)

17

47.90

(5, 4.5)

30

28.53

(0, 6.2)

37.57

(7, 7.6)

18

46.40

(4.2, 6)

31

38.26

(8.3, 6.6)

38.26

(8.5, 7.6)

19

45.84

(6.9, 2.8)

32

40.00

(11.3, 5.8)

37.81

(9.6, 8.4)

20

45.84

(6.9, 1.7)

33

45.84

(8, 1.7)

37.35

(8.5, 9.1)

21

45.84

(8.7, 2.8)

34

45.84

(5.5, 1)

30.56

(6.1, 9.5)

22

45.84

(10, 2.8)

35

45.84

(10, 1.6)

10

40.00

(10.8, 5.8)

23

45.84

(11.1, 2.8)

36

45.84

(11.1, 1.6)

11

39.42

(9.7, 6.3)

24

45.84

(8.2, 4.3)

37

24.98

(0.7, 3.7)

12

40.00

(11.1. 7.1)

25

24.98

(1.4, 4.7)

38

55.00

(2.7, 0)

13

40.58

(8.5, 5.5)

26

55.00

(2.7, 3)

39

29.88

(3.4, 9.5)

Fig.1 Networked electricity system with 39 nodes (buses) and 10 generators

measure.
Consider power interconnected systems in Fig.1. At
first, 39 buses are partitioned to the 3 groups, and the
result is illustrated in Fig.2. 39 buses are shown in 3
dimensional space. X-Y plane represents the locational
information and height means the locational price. Three
dimensional 39 vectors are projected to the X-Y plane. In
Fig.2 the result with only locational information is

obtained. Three groups are marked by , and .

This indicates that three groups are considered with only
locational information because weighing values are
considered as 1=0 and 2=1. This strict condition is
satisfied for only location consideration. Hence, to
satisfy more flexible request, the locational price and
locational information have to be considered
simultaneously. Next, with the variations of weighting

J. Cent. South Univ. Technol. (2011) 18: 190195

194

values 1 and 2, clustering results are obtained in Fig.3.

By varying both values of 1 and 2, 9 nodes change
their clusters. Two were changed into group .
One is changed into group . Finally, 6 are
changed into group . Clustering results are illustrated
in Fig.3. It is very interesting to find the changes in
groups. Their changes all happen at the boundary areas.
However, other changes do not happen for further
variation of weighting values 1 and 2 till 0.27 and 0.73,
respectively. The result is shown in Fig.4, which shows
exactly the same result of Fig.3.
Then, the difference of clustering result between
Fig.3 and Fig.4 when the weighting values are 1=0.27
and 2=0.73 is shown in Fig.4.

Fig.2 Clustering by only location consideration (1=0, 2=1)

Fig.3 Clustering by two consideration (1=0.2, 2=0.8)

As the value 1 approaches to 0.27 and 2 to 0.73,

even any special changing cannot be noticed in Fig.4,
there are so many changes near the cluster boundaries at
each iteration. Explicitly, when 1 approaches from 0.2
to 0.27 and 2 from 0.8 to 0.73, data clustering is
considered by unstable status, i.e., not determined.
However, when 1 and 2 are over 0.27 and 0.73,
respectively, node clustering is convergent 3 groups
without changing any node.

5 Conclusions
1) Fuzzy C-mean clustering method is modified to
apply for the clustering of interconnected power system.
Conventional Euclidean distance used in fuzzy C-mean
is changed into the modified similarity measure. The
proposed similarity measure is applicable to the multisimilarity problem. Furthermore, the usefulness of the
proposed multi similarity measure is verified.
2) For grouping of the power interconnected
networked system in terms of an appropriate similarity
measure, not only locational prices but also locational
information should be properly considered in the
formulation of the similarity measure. Hence, a modified
similarity measure accompanied with regional
information is proposed, followed by example on the
IEEE reliability test system to verify the usefulness of
the proposed idea of the modified similarity measure.
From the results, the coherency between the degree of
similarity level and the number of clusters can be
checked.
3) The convergence of clustering is verified by the
variation of weighting values of each similarity measure.
Locational price and geometric information weighting
values converge to 0.27 and 0.73, respectively. It is also
found by simulation result that some nodes change near
the boundary of each group.

References
[1]

LI W, BOSE A. A coherency based rescheduling method for dynamic

security [J]. IEEE Transactions on Power Systems, 1998, 13(3):
810815.

[2]

JOO S K, LIU C C, JONES L E, CHOE J W. Coherency and

aggregation techniques incorporating rotor and voltage dynamics [J].
IEEE Transactions on Power Systems, 2004, 19(2): 10681075.

[3]

GALLAI A M, THOMAS R J. Coherency identification for large

electric power systems [J]. IEEE Transactions on Circuits and
Systems, 1982, CAS-29(11): 777782.

[4]

WU F F, NARASMITHAMURTHI N. Coherency identification for

power system dynamic equivalents [J]. IEEE Transactions on
Circuits and Systems, 1983, CAS-30(3): 140147.

Fig.4 Changes happened continuously with group boundary

(1=0.27, 2=0.73)

[5]

LIU Xue-cheng. Entropy, distance measure and similarity measure of

fuzzy sets and their relations [J]. Fuzzy Sets and Systems, 1992, 52:

J. Cent. South Univ. Technol. (2011) 18: 190195

305318.
[6]

195
[12]

FAN J L, XIE W X. Distance measure and induced fuzzy entropy [J].

dissimilarity measures for fuzzy sets on the basis of distance measure

Fuzzy Set and Systems, 1999, 104: 305314.

[7]

FAN J L, MA Y L, XIE W X. On some properties of distance

[J]. International Journal of Fuzzy Systems, 2009, 11(2): 6772.

[13]

measures [J]. Fuzzy Set and Systems, 2001, 117: 355361.

[8]

[9]

17831786.
[14]

reliability test system task force of the application of probability

LEE S H, KIM J M, CHOI Y K. Similarity measure construction

methods subcommittee [J]. IEEE Transactions on Power Systems,

Artificial Intelligence, 2006, 4114: 952958.

[11]

The IEEE Reliability Test System-1996. A report prepared by the

4114: 134139.
using fuzzy entropy and distance measure [J]. Lecture Notes in
[10]

LEE S H, RYU K H, SOHN G Y. Study on entropy and similarity

measure for fuzzy set [J]. IEICE Trans Inf & Syst, 2009, E92/D(9):

LEE S H, CHEON S P, KIM J. Measure of certainty with fuzzy

entropy function [J]. Lecture Notes in Artificial Intelligence, 2006,

LEE S H, PEDRYCZ W, SOHN G Y. Design of similarity and

1999, 14(3): 10101020.

[15]

ALLAN R N, BILLINTON R, ABDEL-GAVAD N M K. The IEEE

BEZDEK J C. Fuzzy mathematics in pattern classification [D].

reliability test system - extensions to and evaluation of the generating

Applied Math Center, Cornell University, Ithaca, 1973: 5055.

system [J]. IEEE Trans on Power Systems, 1989, PWRS-1: 17.

JANG J S R, SUN C T, MIZUTANI E. Neuro-fuzzy and soft

computing [M]. Upper Saddle River, Prentice Hall, 1997: 425427.

(Edited by YANG Bing)