> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Power System Islanding Detection & Identification
using Topology Approach and Decision Tree
Rui Sun Zhongyu Wu
IEEE Student member IEEE Student member
Abstract- Power system islanding contingency has always been
one of the most severe wide area system failures. The recent
Hurricane Gustav landed in west coast in 2008 was one exact case to
illuminate the importance and severity of this kind of contingencies.
Longtime the researches of islanding situations focus on studying
the coherence groups of the generations under an islanding
contingency. In this paper, a new method in detecting power system
islanding contingencies using both the systems topological
structure and real-time system dynamic state variables is presented
and tested. An islanding severity index concept is given for ranking
the severity of the islanding cases.
Index Terms Phasor measurements, PMU placement
strategy, Islanding contingency, Decision tree algorithm, islanding
severity index
I. INTRODUCTION
T he research of power system islanding contingencies
identification and observation has started for years. An
islanded power system indicates that the connections
between the main grid and the region where the islanding
occurs are lost and the area is fully isolated from the system.
The isolation can be introduced by multiple reasons: a false trip
of a malfunction break; accidents caused by human or animals
and natural disasters like fire or hurricane. The islanding
duration could be short as seconds when the system recovers
from breaker reclosing, or lasting permanently (for the landfall
of Hurricane Gustav in New Orleans and LA in 2008, the
islanded system survived for 33 hours after isolation [1]). The
radial network system is more vulnerable because the system is
operating under N-1 criteria that losing a single branch may
lead to islanding contingency.
The islanding contingency could bring enormous loss to both
the power utilities and consumers. However the islanding
situation itself is difficult to be distinguished from other events
before. Unlike other contingencies, the islanding influences
both the stability and structure of the system. Sometimes new
islands could survive for a short period of time since its
formation and precise tuning between generation and loads is
needed immediately for every isolated area and a final reunion
is required. Failure to track the islands will leave the system
operating isolated and unsynchronized and eventually breaking
down.
In recent years, the phase measurement units are widely used
in the major power system grids in the U.S. The primary
function of the PMUs is to monitor the grid with real time data
and protect the system from losing stability. They also play an
important role when islanding cases happen. According to the
reports from Energy Transmission, Inc and TEPCO [1, 2, 3],
the phase measurement system successfully detected the danger
978-1-4577-1002-5/11/$26.00 2011 IEEE
1
Virgilio A. Centeno
IEEE Senior Member
of island forming and sent alerts/commands to protection
devices in several severe contingencies during the 2008 Gustav
Hurricane.
The PMUs offer several types of measurements including
voltage phasors from PTs, current phasors from CTs and the
breakers status. These data could all be used for islanding
identification. TEPCO have developed a scheme detecting
power system islanding using the voltage angle differences
between PMUs [2, 4]. According to their research, the
frequency scheme is not as sensitive as the angle scheme in
their tests [4]. While in the 2008 Hurricane Gustav case, the
protection judgments were made upon frequency
measurements [1, 3].
The study of power system islanding in the power system has
started for years. N. Senroy and V. Vittal has identified and
analyzed the behaviors of controlled islands based on forming
coherence groups of the generations under an islanding
contingency [10]. R. Diao and V. Vittal defined islanding cases
using slow coherent generator grouping and minimum power
imbalance on a study of the 16100 bus Eastern power system of
America [11]. In both studies the decision trees are trained to
recognize the general characteristics of the islanding
contingencies and for further detection. The decision tree is
also used for the determination of the best PMU locations.
While generating the islanding contingencies, Senroy and Diao
selected to create 3-phase bolted faults with following line
tripping using enumeration method and Diao tripped multiple
500kV lines to create severe contingencies. In another similar
study of cascading hidden failures, H. Wang and J. S. Thorp has
examined over 170,000 blackouts by characterizing the cases as
a tree-search problem and defined a random search algorithm
with certain thresholds [9].
In this paper, an alternative method in detecting power
system islanding contingencies using both the systems
topological structure and real-time system dynamic state
variables is presented and tested. The proposed method gathers
system information from the PMU measurements and generates
decision tree in detecting abnormal system operation behaviors
related to islanding contingencies. An islanding severity index
concept is given for ranking the severity of the islanding cases.
II. METHODOLOGY
To analyze the islanding contingencies in a study system, we
may want to study the existed or potential islanding cases. By
using the decision tree algorithm, the general characteristics of
this kind of contingencies could be summarized through
analysis. However the actual islanding cases may be limited
and potential island locations are obscure to tell. In the
proposed method a potential islanding location study on the
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

study system models based on their topological structure is

carried out and generating sufficient islanding cases is
expected.
The method has split into three procedures. The first step is to
determine a strategy at concept level, to distinguish islanding
cases from normal system operations and other kind of
contingencies using system information and measurements that
can be obtained. This strategy will help us judge the
contingencies we create in the simulation procedure; in the
second step sufficient islanding cases are expected to
generated, dynamic simulations of the islanded power system
Fig. 1: 8-bus 3-machine test system
are created, a modified islanding formation method is
introduced to distinguish reasonable islanding cases from the
The system structure data can be expressed in several
pool; the simulation data are analyzed in step 3, the decision
matrices:
tree algorithm is used to look for the inner relationships among
Generator matrix: G = [1 5 7] ;
the data and conclude the common characteristics of the
islanding cases and their distinctions from the non-islanding Incidence matrix:
cases. And the islanding cases generated in the second step will 1 1 1 0 0 0 0 0
also be analyzed consequently. 1 1 1 1 0 1 0 0

1 1 1 0 0 0 0 0
A. System Islanding Detecting Strategy:
0 1 0 1 1 0 0 1
For a power system, based on the system topological structure A= ;
0 0 0 1 1 0 0 0
and breaker status, the incidence matrix of the system is
obtained. When the incidence matrix is squared and all 0 1 0 0 0 1 1 0
non-zero elements are changed to 1, the 2nd step depth system 0 0 0 0 0 1 1 0
observability matrix [5] is obtained. And if the incidence matrix
is cubed and all non-zero elements are changed to 1, the 3rd step 0 0 0 1 0 0 0 1
depth system observability matrix is obtained and so on. The nth
step depth system observability matrix has the same rank as the The longest path: 5-4-2-6-7 or 8-4-2-6-7, so the Q = 4.The
system incidence matrix. The non-zero off-diagonal element incidence matrix A needs four iterations (Q=4) to reach full 1
Ai,j in the nth observation depth matrix represents that the bus i matrix by multiplying itself.
can be reached from bus j within n branches. The upper diagonal shows the connection conditions, there
After multiplying the incidence matrix by certain times, all are 8 lines in the system and the total tripping combination is
elements in the Depth matrix will turn to be 1 in a power system 255, if we only concern tripping 4 or fewer lines, the total
with N buses. The certain time is said Q. Q can be obtained combination number is 162.
through the system incidence matrix analysis. Taking any two When applying islanding analysis, we randomly trip 1 and 2
random buses from the system, recording the bus numbers in line(s) in the system:
the shortest path of all existed line paths between the two buses,
repeat this to every possible bus-pair combinations in the
system and find out the largest bus quantity . This number is the
maximum iteration number Q mentioned before. For a system
with lines tripped, in incidence matrix view it equals to
replacing some 1s in the matrix with 0s. If the tripped line isnt
the shortest path between any two buses, the Q will not be
Tripping the line circled in the upper diagonal part of the
influenced, otherwise the Q need to be recalculated.
incidence matrix will create an island while tripping others will
And for system containing islands, theres no possibility to
not. Several cases are raised below:
make the all elements in the depth matrix to be 1. The stopping
Case 1: Tripping line 2-4 in the system, this will cause 2
rule is set as the k-1th depth matrix is equal to the kth depth
islands. When multiplying the incidence matrix, the 3rd depth
matrix; k is always no larger than Q. This is easy to prove: in
matrix is equal to the 4th, which means the final depth matrix is
the islanding case all paths are lost between some bus-pairs.
reached. The 3rd depth matrix shows 4 blocks in the diagonal:
The largest iteration number k equals to the longest bus path
Bus 1-3, 4-5, 6-7 and 8. The off-diagonal blocks show the
among those shortest ones of all remained bus-pairs, which is a
relationships between each diagonal bus group. When elements
subset of the original system. Thus k is no larger than Q. The
are 1, this means the two blocks are in the same island,
final depth matrix will contain lots of blocks with elements
otherwise are not. This offers a way to distinguish islands. Its
inside each block are either all 0 or 1. From this we can easily
easy to find the system is split into two islands: 1-3, 6, 7 and 4,
find out how many islands are formed and how they forms.
5, 8.
The principle of this method is illustrated using an 8-bus
3-machine test system shown in Fig. 1.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1 1 1 0 0 1 1 0
1 1 1 0 0 1 1 0
1 1 1 0 0 1 1 0
0 0 0 1 1 0 0 1
(C1 )3 =
none zero1 0 0 0 1 1 0 0 1
1 1 1 0 0 1 1 0
1 1 1 0 0 1 1 0
0 0 0 1 1 0 0 1
Case 2: Tripping line 1-2 in the system, this wont cause an
island. The matrix ends up in full 1.
Case 3: Tripping line 1-2, 4-8 in the system. The generator
matrix is used to ensure every island contains at least one
generator thus make it an island. The computation stops at the
5th multiplying. Its easy to see the system is split into 2 parts:
Bus 1-7 and 8. However checked with the generator matrix,
Bus 8 is not a generator bus so no island is created.
Using this method, we can easily generate and identify any
islanding cases by manually creating contingencies in a study
system, regardless of the system size.
The islanding detecting strategy is purely working with the
system structure information. Failure to achieve a full-1 matrix
demonstrates a physical structure separation in the system. If
the exact structure of the power system is known and can be
updated within seconds (i.e. all breakers and branches status in
the system are monitored by PDCs), this method itself could be
an efficient tool in system islanding determination. However, in
most power systems such information could not be obtained
thoroughly or simultaneously, which means at the most time,
we need to judge an islanding contingency with limited
synchronized system measurements that could be monitored
through the PMUs, such as voltage and current variables.
B. The modified enumeration method in islanding
contingencies formation
In practical, the islanding contingencies are formed by
environmental disasters like hurricanes or fires; or cascading
breaking down that created by a list of interacting hidden
failures. Most likely the directly involved buses and branches
are which cause the islanding are within a certain region. For
the environmental disaster cases, the region is mostly
determined by the geographical distances of the buses in the
wide area system; for the cascading hidden failure cases, the
region is generally decided by the electrical distances.
However both two islanding factors can be described by
converting them to the bus observation depth matrix. For a
power system, based on the system structure and breaker status,
the incidence matrix of the system is obtained. As mentioned in
Section A, the system observability matrix is easily obtained by
incidence matrix multiplication. The certain geographical and
electrical distance could be equivalently converted by certain
observation depth n. n might be changing due to the differences
in practical branch lengths and different test requirements for
the length of the hidden failure chains.
To locate a potential islanding case, we trip one line to test the
system from every single branch, and then move to multiple
lines. The multiple lines are randomly selected from the system
3
which will create a huge number of combinations. The
complete testing combinations for tripping up to 3 lines in a
1000 bus, 1500 branch system could be more than 500 million;
which is not desirable for simulation. There are too many
unreasonable combinations are counted including some
unreasonable combinations that may introduce an islanding
condition. By applying the observation depth selection, the
enumeration random branches tripping is limited only within
certain observation depth, in other words, inside certain
geographical and electrical distances. Those useless
combinations are eliminated, tremendous time is saved. By
setting observation depth n = 5, the combinations number
decreases to around 50 thousand for the same system
mentioned above.
Fig. 2: Various branch tripping locations
As shown in Fig. 2, for example, tripping line A will create an
island in the system, tripping lines A and B will cause the same
islanding case and this combination is reasonable because the
geographic and electrical distances between the two lines are
close; however tripping line A and C is an less-likely
combination which leads to the same result but not realistic. By
the increase of the system scale, the unreasonable islanding
combinations will increase significantly; the modified depth
first method will eliminate those cases.
C. Decision Tree Algorithm:
A decision tree is a decision support tool that uses a tree-like
graph or model of decisions and their possible consequences,
including chance event outcomes, resource costs, and utility.
Decision trees are commonly used in operations research,
specifically in decision analysis, to help identify a strategy most
likely to reach a goal [6]. In the propose method, system state
variables at different operating conditions will be used as the
input for the decision tree, the DT algorithm would look for the
inner relationships among the data and conclude the common
characteristics of the islanding cases and their distinctions from
the non-islanding cases.
The formation of the decision tree is through the CART, a
professional data-mining software by Salford Systems, Ltd.
The tree type is selected as classification tree. When in the
testing, two methods are used: a 10% test sample method and
10-fold cross validation method. The former method uses 10%
of the input data for testing; the decision tree is formed using
the other 90% data. While in the latter method the whole
learning data is divided into 10 pieces equally. The program
creates 10 decision trees, for each run, 9 pieces of the data are
used for forming the tree and the other piece is used for testing.
Each piece of the data is used as the test sample once. Then the
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

program optimizes the results and generates the best decision
tree. This method is strict and has a very good accuracy.
Through the decision tree generation, three questions are
expected to be answered:
1) The critical locations where the primary splitting parameters
are located;
2) How much difference of the same system variables between
islanding cases and non-islanding cases;
3) How many system variables are needed to create an efficient
decision tree in detecting islanding contingencies.
model is used. The islanding area is limited within the Virginia
states region and a total number of 190 islanding contingency
cases are generated and recorded, with 400 non-islanding
contingency and normal operating cases served as the control
group (300 normal operation cases; 100 short circuit, rapid load
changing and line tripping cases). The channels selected are the
voltage and current measurements at the 50kV level buses
inside Virginia area. Table 2 shows the decision tree prediction
success percentages at different tree sizes using both 10% test
sample method and 10-fold cross validation method.

TABLE 2: DECISION TREE PREDICTION SUCCESS

III. RESULTS & DISCUSSION
Tree size Variable Prediction success Relative
The coverage test results of the modified depth first method (Nodes number) importance cost
in islanding contingencies formation are shown in Table 1. Learn Test
4 16 99.78% 97.73% 0.067
TABLE 1: COVERAGE TEST FOR DEPTH FIRST METHOD 10% test
3 10 98.655 95.45% 0.133
Enumeration Method Modified Depth First Method sample
Test Model 2 6 94.62% 90.91% 0.267
Islanding Depth = 2
(Max. Combo 10-fold 4 18 99.59% 98.16% 0.038
case Combos Islands Coverage
tripping = 3) s tested cross 3 12 98.57% 96.73% 0.078
formed tested formed %
IEEE-14 1350 253 999 143 56.52 validation 2 6 94.29% 93.06% 0.179
IEEE-30 11521 1024 3246 245 23.93 Node 1
Class = 1
IEEE-57 79157 80 6434 20 25 A314912 <= 0.02
Class Cases %
IEEE-118 0 300 61.2
16110 890 2185 49 5.51 1 190 38.8
(up 2 lines) W = 490.00
N = 490
IEEE-300
83845 13786 4073 513 3.72
(up 2 lines) A314912 <= 0.02 A 314912 > 0.02
Node 2 Terminal
Class = 0 Node 4
Q314907_314912 <= -0.01 Class = 1
Modified Depth First Method Class Cases %
0 300 91.5
Class Cases %
0 0 0.0
Depth = 3 Depth = 4 Depth = 5 1 28 8.5
W = 328.00
1 162 100.0
W = 162.00
Combos Islands Cover Combos Islands Cover Combos Islands Cover N = 328 N = 162

tested formed age % tested formed age % tested formed age %

1343 246 97.2 1350 253 100 - - - Q314907_314912 <= -0.01 Q314907_314912 > -0.01
Node 3 Terminal
8414 701 68.5 11136 963 94.0 11521 1024 100 Class = 0
P314927_314928 <= 0.01
Node 3
Class = 1
19246 34 42.5 37219 49 61.3 54882 64 80 Class Cases %
0 300 97.7
Class Cases %
0 0 0.0
3976 97 10.9 5964 182 20.5 8092 286 32.1 1 7 2.3
W = 307.00
1 21 100.0
W = 21.00
N = 307 N = 21
7450 1039 7.5 11865 1660 14.0 17048 2760 20.0
P314927_314928 <= 0.01 P314927_314928 > 0.01

For small system, the modified depth first method is no other Terminal
Node 1
Terminal
Node 2

than enumeration method. Both the methods use the same Class = 0
Class Cases %
Class = 1
Class Cases %
0 298 100.0 0 2 22.2
computing concept and speed. When the system is getting 1 0 0.0
W = 298.00
1 7 77.8
W = 9.00
complex, more unreasonable islanding combinations are N = 298 N=9

introduced. While taking more time in computation, the

enumeration method could not distinguish them efficiently. On
the other hand, the modified depth first method will only work
on forming islands in the given region, which both prevents
recording the unreasonable combinations and saves enormous
time.
The IEEE-118 system is used for decision tree algorithm. Set
the observability depth = 4, the total islanding contingencies
recorded are 182 for tripping up to 2 lines and 7412 for tripping
up to 3 lines.
All contingency dynamic simulations are done using Python
compiled PSS\E (32.03 version). The PSS\E simulation data
are then converted into excel files for the CARTs use.
The decision tree algorithm has been carried out on several
system models. The results showed that this method works very
well with large scale non-equivalent systems. In the simulation,
a modified MMWG 57,000 bus Eastern coast power grids
Fig. 3: Decision tree splitting diagram using 10-fold cross validation
The primary split variables are referred to as having the most
effluence to the prediction accuracy so the places where the
variables measured at are also considered the most proper
locations for PMU installations. In this simulation, all state
variables are selected among 30 500kV buses, which are the
locations that the PMUs will be installed in and primarily
monitor. The measurements at those locations show exactly
what the PMUs may see when an island form in a lower voltage
area and how this island would influence the major 500kV
power delivery system. For 18 variable importances at 500kV,
only 4 locations are required to achieve a full observation.
Through the analysis to this east coast full-loop model,
three islanding contingency categories has been found and
studied.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Fig. 4: Machine angles oscillation of category 1 cases
Category 1:
1) Machines in the islands are angle coherent;
2) The islanded area remains stable although it is physically
separated from the main grid;
3) The islanded machine angles have a relatively larger
oscillation in a different direction compared with the main
grids;
Fig. 5: Machine angles oscillation of category 2 cases
Category 2:
1) Machines in the islands are angle coherent;
2) The islanded area lost stability after the contingency;
3) The islanded machine angles go up because a mismatch
between loads and generation; the angles in the main grid go to
the other direction. Processive protection actions needed.
Fig. 6: Machine angles oscillation of category 3 cases
5
Category 3:
1) Machines in the islands are angle coherent;
2) The islanded area remains stable after the contingency and
the islanded machine angles go along with the main grid
although it has already been physically separated;
3) Some machines near the islanding boundary are influenced
more by the contingency. Sometimes the generators related to
the islanded area lost stability.
The Islanding Severity Index:
The islanding severity index, an alternative of the PMU
location determination is described in this section.
Voltage stability index (VSI) is widely used in power system
stability analysis. However, theres no specific severity index
for judging islanding cases yet. The islanding cases, which vary
from the normal contingencies, show the diversity of the
structure of the system. In some separated region, the system
may stay balanced power generation and load requests. This
does not mean the region is stable, on the contrast; the system
loses its interchange ability between each area and becomes
extremely vulnerable for further power oscillation. This may
not be reflected from common system state variables for normal
VSIs. While deciding the islanding severity, the power
mismatches and the machines inertia are selected as the state
variables in the proposed method.
And the Islanding Severity Index is defined as:
1) Long term: the summation of the power mismatch over the
machines total inertia time constant in each isolated area.
N
M Ni
mislanding ( ISI ) = i H i = H ij
i =1 H i j =1
Hi is the total inertia time constant of the isolated area I; N is
the number of isolated areas in the islanding contingency.
2) Instant: the summation of the power mismatch over the
machines total real power generation in each isolated area.
N Ni
M
mislanding ( ISI ) = i Pi = Pij
i =1 Pi j =1
Pi is the real power generation in the isolated area i; N is the
number of isolated areas in the islanding contingency.
As its literal expression, the instant index shows the instant
impact of the isolated area will have towards the main grid and
how severe it could be exactly at the time the contingency
occurs. The protection scheme may not have enough time to
react to the contingency so this index would show the
maximum hazard the contingency could cause. The long term
severity index could be used in determining how severe the
systems operating status is from the normal operating criteria
long after the systems getting isolated and each islanded areas
working individually for a while, if the contingency is too
massive to be handled in time. At that time, all possible
protection schemes should be carried out. Excessive
generations and less-important loads are tripped, branches are
tempted to reclose and switching shunts are put into use to
balance the voltage magnitudes and angles.
Table 3 shows 10 typical islanding contingency cases taken
from the database. According to the computation, category 2
islanding cases have relatively larger ISI instant values than the
category 1 and 3 cases. Islanding case with a higher instant ISI
value show larger angle oscillations; while it doesnt
necessarily mean it has a higher longterm ISI value which is
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

related to the generators inertia and have effect to their power [2] Hashiesh, F.; Mostafa, H.E.; Mansour, M.M.; Khatib, A.-R.; Helal, I.;
Wide area transient stability prediction using on-line Artificial Neural
re-dispatch characteristic. Those Island cases with huge Networks, Electric Power Conference, 2008. EPEC 2008. IEEE
generation have correspondingly much larger longterm ISI Canada, 6-7 Oct. 2008 Page(s):1 7
values however a large ISI value does not indicate the machines [3] Kolluri, S.; Mandal, S.; Galvan, F.; Island formation in entergy power
in the island will lose stability. There are also some meter grid during Hurricane Gustav, Thomas, M.; Power & Energy Society
General Meeting, 2009. PES '09. IEEE, 26-30 July 2009 Page(s):1 5
islanding cases in the system; case 7-10 all have interarea [4] T.; Yasuda, T.; Takahashi, O.; Kaminaga, M.; Imai, S.; Islanding
tielines connected with the islanded area. Although in these protection system based on synchronized phasor measurements and its
cases they have relatively high instant and longterm ISI values, operational experiences, Ohno, Power and Energy Society General
the islanded areas are still stable, however the influences to the Meeting - Conversion and Delivery of Electrical Energy in the 21st
Century, 2008 IEEE, 20-24 July 2008 Page(s):1 5
system is more severe than category 1 cases, which are also [5] Chakrabarti, S.; Kyriakides, E.; Optimal Placement of Phasor
shown as the more severe angle oscillations of the boundary Measurement Units for Power System Observability, Power Systems,
generators. IEEE Transactions on, Volume 23, Issue 3, Aug. 2008 Page(s):1433
1440
TABLE 3: ISLANDING SEVERITY INDEX ANALYSIS [6] Ming Zhou; Centeno, V.A.; Phadke, A.G.; Yi Hu; Novosel, D.; Volskis,
Mismatches H.A.R.; A preprocessing method for effective PMU placement studies,
Island
Case (Initial) Inertia Cate ISI_ ISI_lo Electric Utility Deregulation and Restructuring and Power Technologies,
Generation
# Combo gory Instant ngterm 2008. DRPT 2008. Third International Conference on, 6-9 April 2008
(Initial) P Q
1 165+j18.3 -63.42 8.469 15.3 1 0.384 4.145 Page(s):2862 2867
2 210+j14.8 169.98 6.196 12.28 2 0.809 13.842 [7] Ishibashi, A.; Imai, M.; Omata, K.; Sato, S.; Takagi, T.; Nakachi, Y.;
3 1744-j254.2 -429.2 -385.8 20.22 1 0.246 21.226 Ogawa, S.; New type of islanding detection system for distributed
4 606+j74.4 556.23 52.487 27.54 2 0.918 20.197 generation based on voltage angle difference between utility network and
5 348+j84.8 316.02 73.682 26 2 0.908 12.155 distributed generation site, Developments in Power System Protection,
6 275+j9.9 46.582 0.069 15.3 1 0.169 3.0446 2004. Eighth IEE International Conference on, Volume 2, 5-8 April 2004
7 227-j12.6 92.111 -32.85 6.82 3 0.406 13.506 Page(s):542 - 545 Vol.2
8 227-j12.6 84.76 -36.06 6.82 3 0.373 12.428 [8] Wang, H. ; Thorp, J.S.; Optimal Locations for Protection System
9 147.5-j11.9 95.738 -53.60 6.232 3 0.649 15.362 Enhancement: A Simulation of Cascading Outages, Power Delivery,
10 147.5-j11.9 118.67 -47.22 6.232 3 0.805 19.042
IEEE Transactions on, Volume 16 , Issue 4 Oct 2001 Pages(s): 528 533
[9] Senroy, N. ; Heydt, G.T. ; Vittal, V. ; Decision Tree Assisted Controlled
Islanding, Power systems, IEEE Transaction on, Volume 21, Issue 4,
IV. CONCLUSIONS Nov. 2006 Page(s):1709 - 1797
[10] Ruisheng Diao ; Vittal, V. ; Kai Sun ; Kolluri, S. ; Mandal, S. ; Galvan, F.
In this paper, the power system islanding contingency ; Decision Tree Assisted Controlled Islanding for Preventing Cascading
analysis method is presented and tested. The islanding Events Power Systems Conference and Exposition, 2009. PSCE '09.
IEEE/PES, 15-18 March 2009, Page(s): 1 - 8
detecting strategy is using the systems topological structure [11] Hashiesh, F.; Mostafa, H.E.; Mansour, M.M.; Khatib, A.-R.; Helal, I.;
and real-time breaker status for the judge of whether islands Wide area transient stability prediction using on-line Artificial Neural
occur or not. In the islanding contingency simulation Networks, Electric Power Conference, 2008. EPEC 2008. IEEE
procedure, a modified enumeration method is introduced in Canada, 6-7 Oct. 2008 Page(s):1 - 7
[12] Jerel Alan Culliss, A Method for PMU-Based Reconfigurable
selecting reasonable potential islanding locations. The Monitoring, VT graduate thesis, 2009
geographical and electrical distances are converted into the [13] Power Systems Engineering Research Center, Optimal Placement of
observation depth and unrealistic islanding cases are Phasor Measurement Units for State Estimation, Texas T&M
successfully eliminated. The decision tree algorithm is used to University, 2005
[14] Jian Chen; Abur, A.; Improved bad data processing via strategic
analyze the distinction between islanding contingencies and placement of PMUs, Power Engineering Society General Meeting,
other operating conditions. Simulations are carried out on east 2005. IEEE, 12-16 June 2005 Page(s):509 - 513 Vol. 1
coast full-loop system and an average predication success of
98% is achieved for a full scale tree; the simulation results VII. BIOGRAPHIES
demonstrate the decision tree algorithm is effective in the Rui Sun (S10) received his B.S. degree from Southeast University, Nanjing,
islanding judgment for large scare power system models. China, in 2007 and M.S. degree from Clemson University, in 2009,
Furthermore, the important variables and primary splitters respectively. He is currently pursuing the Ph.D. degree from the Department of
could help in deciding the PMU location. The concept of the Electrical and Computer Engineering, Virginia Polytechnic Institute and State
University, Blacksburg. His main research interests are power system state
islanding severity index, an alternative of the PMU location estimation and power system wide area measurement.
determination is also introduced and tested using simulated
islanding contingencies. Zhongyu Wu (S10) received her B.S. degree from Southeast University,
Nanjing, China, in 2008 and M.S. degree from Virginia Tech, in 2009. She is
currently pursuing the Ph.D. degree from the Department of Electrical and
V. ACKNOWLEDGEMENTS Computer Engineering, Virginia Tech, Blacksburg. Her main research interests
The authors gratefully acknowledge the contributions of are power system wide area measurement, power system state estimation and
power system protection.
Arun G. Phadke, James S. Thorp, Craig Crider and Matt
Gardner from Dominion Virginia Power for their work on this Virgilio A. Centeno (M92--SM06) received the M.S. and Ph.D. degrees in
document. electrical engineering from Virginia Polytechnic Institute and State University
(Virginia Tech), Blacksburg, in 1988 and 1995, respectively. He worked as a
Project Engineer at Macrodyne, Inc., Clifton Park, NY, in the development of
VI. REFERENCES phasor measurement units from 1991 to 1997. He joined the faculty of Virginia
[1] Galvan, F.; Mandal, S.; Thomas, M.; Phasor Measurement Units (PMU) Tech as a Visiting Professor in the fall of 1997 and became an Associate
instrumental in detecting and managing the electrical island created in the Professor in 2007. His area of interest is wide area measurement and its
aftermath of Hurricane Gustav, Power Systems Conference and applications.
Exposition, 2009. PES '09. IEEE/PES, 15-18 March 2009 Page(s):1 4