A Robust Initialization Algorithm for k-Means Clustering in
Power Distribution Networks with PMU-based Adaptive
Protection System
Pooria Mohammadi and Hassan El-Kishky
Department of Electrical Engineering, University of Texas at Tyler, Tyler, TX, USA
ABSTRACT
The K-Means clustering is one of the most popular and
influential algorithms in data categorizing methods. K-
Means simple and straightforward formulation made it a
widely acceptable method in many fields and applications.
This simplicity comes with some prices such user defined
number of clusters, uniformly sized clusters and different
final clusters as of being sensitive to initial centroids. K-
means sensitiveness to initial centroids leads to different
clusters per execution with different and relatively long
iteration numbers. Different applications have their own
initialization and improvement techniques for k-means
relying on their particular data traits. Power systems
recently have been involving with data mining and
clustering due to fast increase in PMU uses for
supervisory, control and protection goals in smart grids.
Large amount of data streaming by PMU demands quite
simple method with minimum computational burden to
meet delay tolerance for various working phases and
expectations. This article presents an approach
significantly improving k-means clustering algorithm by
pre-analyzing the data and finding best initial centroids.
Extensive experiments have been made to verify the
approach robustness in reducing the number of iterations
and resulting in unique clusters in all executions.
Index Terms k-Means clustering, PMU, power
systems, Central Protection Unit
1 INTRODUCTION
K-means clustering is one of the most popular techniques in
data clustering. It has been utilized in applications such as data
mining, knowledge discovery, data compression and vector
quantization, pattern recognition and classification, medical
imaging, and many others involving experimental, statistical,
or just large amounts of information [1-3].
PMUs fast data stream demands a simple and adaptive
algorithm which is able to adapt to ongoing streams of
information. Xiong [3] shows that k-Means clustering tends to
produce clusters with relatively uniform sizes even by distinct
varied input data. Initial centroids in many applications can be
determined randomly from within the data [5,6]. In this study,
978-1-4799-4047-9/14/$31.00 2014 IEEE
however, random initial centroids may lead to noticeable
deviations in k-Means clustering results.
This article introduces a new initialization algorithm for k-
Means clustering. The paper is organized as follows: in
Section 2 k-Means is introduced; in Section 3, the proposed
initialization algorithm is presented; in Section 4, the results
and discussion are presented; and, finally, conclusions are
presented in Section 5.
2 K-MEANS CLUSTERING PRINCIPLES
Given as an input set of data which
consists of n objects with dimensions in space of and the
parameter which is the number of clusters that
input needs to be categorized in, k-Means clustering should
determine clusters including their vectors. k-Means does not
have a limit on an input objects number of dimensions. That
is, each can be represented as when
can be any positive integer. Clustering algorithm starts by
considering number of points within to start the algorithm
with cluster initial centroids which should be close to their
own clusters center of mass by the end of algorithm. One of
the disadvantages of k-Means is that it is sensitive to initial
centroids however on the other hand, this dependency relies
on the datas natural structure [6]. Choosing random centroids
may be quite sufficient for some types of data but results in
dissimilar outputs for PMU data, as shown in section 4 of the
paper.
The squared Euclidian distance is used for data clustering in
the basic k-means. Some k-means algorithms utilize different
distant computations which dont have a significant impact on
the results particularly in this application. An error function is
calculated in each iteration. Clustering is done through
minimization of the error function. Hence, either the a
function value or its change from previous iterations can be
used as the algorithm termination condition. The most
common termination condition, however, is the fact that no
more points will be transferred within the clusters and the k-
Means algorithm converges to a steady-state.
Both centroids and clusters groups of objects will be
updated in each of the iterations. Centroids play an important
role in k-means number of iterations to converge quality and
stability of the clustered results in different executions. The
252
centroids update with the mean of objects categorized in it from C-1 to C-2. After stabilizing in C-2 system encounters
cluster and is the number of objects in is given by a fault and enters to C-3, moving the system working
condition in Figure 2 from C-1 to C-2 is indicated by close
dots with small steps, however, dots are running from C-2 to
(1)
C-3, which indicates a fault. Also, power system and PMU
output data have concentration of data points in steady state
condition. That is why we have more than 5000 data points in
3 PROPOSED INITIALIZATION each of Cs sections while all the cyan parts are less than 1500
ALGORITHM number of data.
The k-Means clustering requires users to determine the
number of clusters and initial centroids. Both are challenging;
however, the number of clusters is beyond this study.
Each category of data has its particular traits due to the
nature of the system. Dashed lines in Figure 1 separate the
different phases of working in our system as the scenario
events are happening in 0 to 1 second simulation time span.
Plotted data fall into four categories in which three of them are
C-1, C-2, and C-3, which are related to the system when it is
stable. Narrow bands in Figure 1, indicated by dashed lines,
are representing the transient time where the system
encounters a change or fault and needs to pass a time to be
dampened. Figure 1 resulted from a stream set of data with
24031 3-D points yielded by a PMU. This is how , Figure 2. PMU5 VIIa 3-D data presentation.
relates to a specific moment and is a dot with its
position in system trace. Figure 2 presents the same data in a
3-D plotted environment using dots to curve the data. Cyan The two PMU data traits, data step length and
colored dots in Figure 2 are related to narrowly dashed line concentration, have been considered to find the best initial
times in Figure 1 and the same manner exists for Cs and their centroids. A discreet function is used to calculate every step
colors. One can deduce that Figure 1s narrow bands in trivial length of the system. The function is given by
times result in far spaces in actual 3-D space. That is, although (2)

transient parts take a small amount of time, the plotted system
using dots result in a bigger step length for dots in those time
periods. This phenomenon is true for faults, change in load, where , the length of is always one less than
DGs dis/connection, and any other disturbance or work the length of the data set. Figure 3 illustrates the resulted
condition alteration in power systems. function for the PMU output in Figure 1. Figure 3 and the
magnified section clearly present the fact that during steady
state operation, the data points step length barely exceeds 0.2
with an average in the range of 0.05. While the system is
changing its phase of operating point, this step change
significantly increase to more than one.

Figure 1. Typical PMU output data.

PMUs sample the phasors at a constant sampling rate

synchronized by satellite GPS. Hence, any change in Figure 2
dots step length is merely due to the system response which
can be categorized based on the changes rate, magnitude or
average. For instance, in Figure 2, the upstream utility grid is
feeding the system in C-1. Connecting a significant amount of Figure. 3. Discrete function D(i) representing the system step length.
DGs to the system changes the system working point, moving

253
 Using the function D and the PMUs data traits enabled us
to develop a more efficient algorithm to calculate the best
initial centroids. In this application, the total number of data is
24031, therefore, batches with sizes less than 100 can easily
be considered as noise or at most unnecessary. It should be
mentioned that we have reached the same results by ignoring
batches size 1000 but 100 will be the most conservative batch
reduction size.
The algorithm merges the batches with close average points.
This is not really critical in real systems as we are assuming to
have enough data during normal working condition, and
abnormal working condition has smaller periods of time
compared to normal steady state. At this stage, we have a
limited number of batches, and each of them includes sets of
data with a certain distance in their averages. These batches
will be organized in a rising manner based on their sizes.
Obviously, batches with larger sizes are more important, at
least indicating clusters with bigger size. Finally, the top k
ones of these batches will be chosen to represent the clusters
regions compared with Figure 1. Using basic algorithm has
taken 8 iterations while taking advantage of the proposed
initialization algorithm reduced it to three iterations.
Decreasing the iterations by one-third has a significant impact
on the clustering computational burden applications.
Figure 2 illustrates three plots of data (VII) and each of
them can be expanded to phases abc. The same thing is
applicable using the signals sequences. Now, any
combination of these data can be used for decision-making
and clustering processes regardless of the number of
dimensions. The proposed code has no limitation in data
dimensions which enables the CPU to utilize any combination
of preferred data.
C-3
180
C-1
100
0

Ia
and their center of mass will be calculated. These centroids C-2

will be used for k-means.

-100

4 RESULTS AND DISCUSSION

80 1.5
A simplified 8-bus system has been used in this study for 60
1
40
simulations and PMUs raw data extraction. SDFT PMU has 20 0.5
0 0
been used to estimate the voltage, current, and phases of the Ia Va
installed bus with specific sampling frequency. The 8-bus Figure 4. PMU5 data clustered using conventional algorithm.
system is radial, which is fed from the upstream utility grid
(UG) connection. In a downstream feeder, it has a large
C-3
amount of DG which connects at the scenarios second period. 180
The designed scenario aimed to show how the system working C-1

state changes and how the PMU data and subsequently 100

proposed algorithm responses change. Figure 1 can clearly

present the scenario starting from t=0 with normal condition 0
Ia

and just being fed from UG. At t=0.3 the downstream DGs
connect to the network improving the voltage, reducing the
sensed current, and significantly changing the phase in that
specific bus. At t=0.6 a bolted 3-phase happened in the middle
of the network which is obvious through the PMU output and
will be cleared at t=0.9. After fault clearance, the system still
is being fed from both sides. That is, the system has the same
state as the second period marked by C-2 in figure.
Figure 2 illustrates PMU5 VIIa data plotted based on the
known scenario periods. Hence, the transient parts are in cyan
and the actual clusters have been categorized by different
colors and arrows showing the cluster names which are the
same in other PMU5 cluster results. That is, the cyan colored
data is to clarify the borders and system states (on bus 5)
presenting in 3-D for the designed scenario. Figure 4
presents the clustered data using basic k-Means algorithm and
Figure 5 shows the resulted clusters using the proposed
algorithm. One can observe in Figure 5 that basic algorithm
could not differentiate the C-1 cluster region and has specified
parts of the transients to it. The actual C-1, however, has been
included in the green cluster section of C-2 which is not
correct. On the other hand, Figure 5 clusters using the
proposed algorithm have distinct borders and correct cluster
C-2
-100
80 1.5
60
1
40
20 0.5
0 0
Ia Va
Figure 5. PMU5 data clustered using proposed algorithm.
Figures 6 and 7 present clustering results for the same
scenario but using data from PMU3. Figure 7 depicts the
clustering result using the proposed algorithm for 3 scenario
periods. Connecting the DG at t=0.3 causing the system state
to go from C-1 to C-2 has a minor location change based on
bus 3 signals. Hence, installed PMU at bus 3 has been chosen
as the C-1 and C-2 are closely located to each other, which is
more challenging for clustering. Clustered data in Figure 7
show the correct distinction between C-1 and C-2 knots;
however, Figure 6 illustrates how k-Means may result in a
totally wrong result of random initial centroids. That is, both
C-1 and C-2 knots have been classified inside a single cluster
and parts of transient have been mistakenly considered as the
third cluster. Using basic k-Means method yields results by 9

254
iterations while the proposed algorithm again decreases it by
one-third. Data pre-analyzing, using proposed algorithm,
results in precise and stable clustering in any number of
executions. It has been mentioned before that k-Means
clustering using the basic algorithm does not have a stable
output. That is, Figures 4-7 are selected from many executions
which run by the authors. The basic code may even result in
really accurate clustering output or aim to a totally surprising
clustered data as of local minimums in objective function. The
number of iterations, however, never reaches less than two
times the iterations using the proposed algorithm.
180
100
C-3
C-1
0
Ia
Figure 8 presents a complicated scenario case which has
been clustered using the proposed algorithm. This cluster
result has been yielded by six iterations when the same
clusters region and iterations have been yielded in any
execution. On the other hand, the basic algorithm has
significantly different cluster regions per each execution with
some critical mistakes in categorizing two cluster knots in a
single cluster in many runs. Basic code took a number of
iterations with a minimum of twenty-one and a maximum of
forty-eight observed, and the iteration average of total runs is
thirty-four which is really high compared to six iterations
using the proposed algorithm.
5 CONCLUSIONS
This article presented the newly developed algorithm to pre-
analyze the data and calculate the initial centroids for

clustering using k-Means. k-Means clustering has a significant

popularity in many applications but each applications data
-100
have their own traits. Basic k-Means clustering simply uses
randomly chosen data to start as initial centroids.
0
This results in unstable and inaccurate clusters in large
0.5
C-2 250
iteration numbers. The proposed algorithm takes advantage of
1 200
100
150 power system PMU data characteristics to evaluate the scales
50
1.5 0 of the data and calculate the most efficient initial centroids.
Va Ia
Figure 6. PMU3 data clustered with conventioal k-Means This approach significantly reduces the number of iterations
and stabilizes the clustering output. That is, the clustering
regions are no longer varied executions. The effectiveness of
180 this algorithm can be used in CPU for power system
100 C-3 protection applications since the algorithm is simple and does
C-1
not add a significant computational burden to the system and
0 can be considered for fast and transient applications.
Ia

-100
REFERENCES
[1] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R.
Silverman, and A. Y. Wu, "An efficient k-means clustering algorithm:
0
Analysis and implementation," IEEE Trans. Pattern Anal. Mach. Intell.
0.5 vol.24, no.7, pp.881,892, Jul 2002.
C-2
200
250 [2] X. Hui, J. Wu, and J. Chen, "k-Means clustering versus validation
1
150 measures: A data-distribution rerspective," IEEE Trans. Syst., Man. B,
100
1.5 50
0 vol.39, no.2, pp.318-331, April 2009
Va Ia
[3] G. A. Jimnez-Estvez, L. S. Vargas, and V. Marianov, "Determination
Figure 7. PMU3 data clustered with proposed algorithm of feeder areas for the design of large distribution networks," IEEE
Trans. Power Del., vol.25, no.3, pp.1912-1922, July 2010
[4] M. J. Li, M. K. Ng, Yiu-Ming Cheung, and J. Z. Huang, "Agglomerative
fuzzy k-Means clustering algorithm with selection of number of
C-1 clusters," IEEE Trans. Knowl. Data Eng., vol.20, no.11, pp.1519-1534,
6 C-4
Nov. 2008
5
[5] G. F. Tzortzis and C. L. Likas, "The global kernel k -means algorithm
4 for clustering in feature space," IEEE Trans. Neural Netw., vol.20, no.7,
C-2 C-5 pp.1181-1194, July 2009
3
c

[6] A. Arvani and V. S. Rao, Detection and protection against intrusion on

I

2 smart grid systems Int. J. Cyber-Sec. Digit. Forensics, vol.3, no.1, pp.
1
34-38, April 2014.
C-3
0
8
6 10
8
4 6
2 4
2
Ib 0 0
Ia

Figure 8. PMU data clustered using proposed algorithm

255