energies

Article
PMU Measurement-Based Intelligent Strategy for
Power System Controlled Islanding
Yi Tang 1, *, Feng Li 1 , Chenyi Zheng 1 , Qi Wang 1 and Yingjun Wu 2
1 School of Electrical Engineering, Southeast University, Nanjing 210096, China; lifeng_ee@seu.edu.cn (F.L.);
220162328@seu.edu.cn (C.Z.); wangqi@seu.edu.cn (Q.W.)
2 College of Automation, Nanjing University of Posts and Telecommunications, Nanjing 210023, China;
ywu@njupt.edu.cn
* Correspondence: tangyi@seu.edu.cn; Tel.: +86-025-8379-0617

Received: 28 November 2017; Accepted: 3 January 2018; Published: 7 January 2018

Abstract: Controlled islanding is an effective remedy to prevent large-area blackouts in a power

system under a critically unstable condition. When and where to separate the power system are the
essential issues facing controlled islanding. In this paper, both tasks are studied to ensure higher time
efficiency and a better post-splitting restoration effect. A transient stability assessment model based
on extreme learning machine (ELM) and trajectory fitting (TF) is constructed to determine the start-up
criterion for controlled islanding. This model works through prompt stability status judgment with
ELM and selective result amendment with TF to ensure that the assessment is both efficient and
accurate. Moreover, a splitting surface searching algorithm, subject to minimal power disruption,
is proposed for determination of the controlled islanding implementing locations. A highlight of this
algorithm is a proposed modified electrical distance concept defined by active power magnitude
and reactance on transmission lines that realize a computational burden reduction without feasible
solution loss. Finally, the simulation results and comparison analysis based on the New England
39-bus test system validates the implementation effects of the proposed controlled islanding strategy.

Keywords: controlled islanding; transient stability; machine learning; splitting surface searching
algorithm

1. Introduction
A modern power system operates close to its technical limits because of expansions in scale
and load demand. This operation makes a power system more vulnerable to serious disturbances.
To prevent blackouts evolved from rotor angle instability conditions, system separation measures
are implemented and show an effect in certain scenarios [1]. With advances in the development of
measuring and controlling techniques, controlled islanding strategies adapted to different scenarios
are of increasing interest [2–4].
For controlled islanding [5], the splitting initiation criterion, an algorithm for splitting
surface determination and stability control for an isolated sub-group system, are necessary in the
implementation and are usually discussed separately in the literature. Moreover, determining when
and where to separate the power system is primarily studied because of the crucial impact of these
factors on the success rate of controlled islanding. Splitting initiation criterion usually relies on prompt
transient stability status judgment of the power system, which can be categorized into two types
according to which physical model is considered. Rapid time-domain simulation [6] and transient
energy function methods [7] are representatives of model-based methods.
With widely-configured advanced metering infrastructure [8], dynamic data during a power
system transient period can be obtained within milliseconds. On this basis, the use of model-free
methods has expanded rapidly because of their acceptable accuracy and accredited high processing

Energies 2018, 11, 143; doi:10.3390/en11010143 www.mdpi.com/journal/energies

Energies 2018, 11, 143 2 of 15

speed. On one hand, implicit information in trajectories, e.g., the rotor angle trajectory and the active
power trajectory of a generator, is excavated from the aspect of physical [9–11] and mathematical [12,
13] characteristics. In [9,10], post-disturbance dynamics are analyzed on phase planes ∆ω–∆δ and
P–∆δ to realize stability assessment. These methods rely on correct mapping of an actual power
system, the complexity of which increases vastly with the expansion of the power system scale.
In [12,13], a curve extrapolation technique, including Taylor series expansion and a pattern recognition
method, is applied for fast stability assessment. These methods demonstrate promising performance
in short-term assessment; however, their results in long-term assessment are questionable because
they do not account for power system non-linear characteristics in a dynamic process. On the other
hand, machine learning techniques, including artificial neural networks (ANN) [14,15], support vector
machine (SVM) [16], decision tree (DT) [17,18], and extreme learning machine (ELM) [19], have been
widely utilized for transient stability assessment, showing promising performance. In [18], an adaptive
controlled islanding measure assisted by the DT method is proposed to solve the problem of “when
to island”. However, the accuracy of these machine learning based methods is closely related to the
training sample scale and quantity, as well as the training method.
Three types of methods are used to determine the splitting surface. Slow coherency theory is the
theoretical support for the first type [2,20–22]. This type of method reduces the feasible solution space
by identifying, in advance, slowly coherent groups of generators with which the dynamic behavior
of each generator is considered. However, the time-consuming problem caused by evaluation of
higher order states of the system and iterative calculations is noteworthy. The second type of method
relies on graph theory [23–25]. These methods consider the power system as an undirected graph,
and cut sets of splitting lines are determined by satisfying certain constraints. Although these methods
achieve an increase in computing efficiency, problems with respect to feasible solution loss in the graph
reduction process and rationality of the computing results deserve further research. In [23], breadth
first search (BFS) and depth first search (DFS) were applied to searching for the optimal solution with
minimal power imbalance. In [24], a three-phase method based on an ordered binary decision diagram
(OBDD) highlighted the superiority of graph theory. In [25], the calculation complexity is considerably
reduced with spectral clustering; however, the procedure on generator coherency constraint processing
is controversial [26]. The third type of method is based on an intelligent optimal solving algorithm,
which can help to speed up the calculations and determine globally-optimal solutions that are as close
to the truly optimal solutions as practically possible [27,28].
In most of the relevant literature, research on controlled islanding strategies are conducted
from the aspect of either the start-up criterion or optimal splitting surface determination. However,
a complete controlled islanding strategy should include both aspects in practical implementation.
Moreover, increased configuration of phasor measurement unit (PMU) devices enables a situational
awareness capability of a power system, and an efficient data-processing method is necessary to
achieve such a capability.
Therefore, the main objective of this paper is to propose a controlled islanding strategy with high
efficiency and reliability. For the start-up criterion, a transient stability assessment model aimed at
maximizing the benefits of model-free methods is constructed by integrating machine learning and the
trajectory fitting (TF) method. Furthermore, an optimal splitting surface determining algorithm based
on graph theory is proposed; this algorithm avoids feasible solution loss using the modified network
reduction method.
The remainder of this paper is organized as follows: In Section 2, key issues about controlled
islanding strategy design is discussed. The transient stability assessment model for the start-up
criterion and the searching algorithm for the optimal splitting surface are presented in Sections 3 and 4,
respectively. The implementing structure and process of the proposed controlled islanding strategy is
explained in Section 5. Case studies and the conclusions are presented in Sections 6 and 7, respectively.
Energies 2018, 11, 143 3 of 15

2. Controlled Islanding Strategy Design

In contrast with conventional passive islanding, controlled islanding is an online, centralized,
and globalized implementation method for system separation. This method determines the splitting
locations rapidly with globally obtained operation information and constraints, e.g., power imbalance
and power-flow disruption. Hence, controlled islanding always shows outstanding performance.
Determining both when and where to separate the power system is essential for controlled
islanding. The splitting action occasion is influential on the dynamics of post-splitting systems.
The splitting surface selection determines the operational state of the isolated system and the
subsequent control measures.
Regarding the task of “when”, prompt transient stability assessment is an effective tool.
Model-based transient stability assessment methods, e.g., time simulation and transient energy
function, are restricted by excessive computing time or high complexity in the analysis of a large-scale
power system. Machine learning-based methods have been widely applied in stability assessment,
exhibiting good performance. Among these machine learning methods, the ELM algorithm has been
demonstrated to be useful for transient stability assessment because of its high training efficiency
and preferable accuracy. In [19], the ELM was verified to provide incorrect judgments at relatively
high possibility, if its output is within a certain interval. This result indicated that the accuracy
of the ELM-based transient stability assessment model could be enhanced further with specifical
amendments. In this paper, the TF method in [13] is adopted as the enforcement tool for amendments.
Hence, the start-up criterion for controlled islanding strategy is determined using a transient stability
model constructed based on the ELM and the TF method. The model realizes a balance between
computing speed and accuracy through a proposed coordinating mechanism.
Regarding the task of “where”, it aims to determine a suitable islanding surface to ensure the
stability of the post-splitting subgroup system. In this paper, “suitable solution” corresponds to the
splitting surface with minimal power imbalance and power-flow disruption, where minimal power
imbalance is regarded as the objective function in the optimal splitting surface calculation. Theoretically,
the searching algorithm requires a large calculation quantity because of the extensive possibility of
splitting line sets; this situation is known as a typical NP-hard (non-deterministic polynomial) problem.
For either graph theory-based methods or slow coherency theory-based methods, determining the
splitting surface is intended to search for generator groups with weak connections. Hence, the possible
solution space can be pre-filtered with an evaluation of the electrical connection strength. The electrical
distance with line reactance only is such an evaluation index; however, this index’s static feature may
lead to loss of feasible solutions in the network reduction. In this paper, a modified electrical distance
defined by the ratio of the transmission line reactance and the active power measurements (xij /Pij ) is
proposed for solution space reduction.

3. Hybrid Transient Stability Assessment Model for Activation Criterion

Conventional passive islanding relies on local electrical information to function. Controlled
islanding makes a decision based on global system information, and its action state depends on
transient stability assessment results.
In this part, a hybrid transient stability assessment model based on the ELM and the TF method is
constructed, as shown in Figure 1. The proposed model can provide support for making a controlled
islanding decision with post-disturbance system stability assessment results. If the assessment result is
stable, then subsequent measures for controlled islanding (e.g., splitting surface searching and splitting
actions) will not be triggered. Otherwise, the splitting surface searching algorithm will be initiated at
once with generator grouping information.
Energies 2018, 11, 143 4 of 15
Energies 2018, 11, 143 4 of 15
Energies 2018, 11, 143 4 of 15

Offline Constructing Online Transient Stability

for Hybrid
Offline Model
Constructing Assessment
Online Procedure
Transient Stability
Data for Hybrid Model
Processing Procedure Assessment Procedure
Post-disturbance
Data Processing Procedure
Feature Selection PMU Measurements
Post-disturbance
with Feature
Fisher Discrimination
Selection PMU5Measurements
Cycles Sampling
with Fisher Discrimination 5 Cycles Sampling
ELM Transient Stability Reliable
Training Assessment
ELM TransientModel
Stability Reliable
with Selected Features

Unreliable
Training ELM based Classifier
Assessment Model
with Selected Features

Unreliable
ELM based Classifier
15~20 Cycles Sampling
Knowledge Base
15~20
TFCycles Sampling
Transient Stability
Knowledge
On-line Base
Recorded Data
Assessment
TF Model
Transient Stability
On-line Recorded Data
Historical Data Trajectory Fitting
Assessment Model
Historical Data TrajectoryAmendment
Fitting
Off-line Simulation Data
Off-line Simulation Data Transient Stability
Amendment
Assessment Result
Transient Stability
Data Processing Procedure Assessment Result
Data Processing
Hierarchical Procedure
Clustering
forHierarchical
Rotor AngleClustering
Trajectories Start-up Criterion
for Rotor Angle Trajectories for Start-up
Controlled Islanding
Criterion
for Controlled Islanding
Figure 1. Construction and work-flow of the hybrid transient stability assessment model.
Figure 1.
Figure Construction and
1. Construction and work-flow
work-flow of
of the
the hybrid
hybrid transient
transient stability
stability assessment
assessment model.
model.
3.1. ELM Theory
3.1. ELM
3.1. ELM Theory
Theory
The ELM is a single hidden layer feed-forward neural network algorithm proposed by Huang
[29]. The ELMisisawith
asingle
single hiddenlayer
layer feed-forward
networkneural network algorithm proposed by Huang
The ELM hidden feed-forward neural network algorithm proposed by Huang [29].
Compared a conventional neural algorithm, the ELM algorithm determines the
[29]. Compared
Compared
parameters
with with anodes
a
of hidden conventional
conventional neural network
neural
using a random
network
assignment algorithm,
algorithm, the
and inverse the ELM
ELM algorithm
algorithm
calculation determines
determines
process, the
the
rather than
parameters
parameters of
of hidden
hiddennodes
nodesusing
usinga random
a random assignment
assignment and inverse
and
the time-consuming tuning process using the gradient descent method, thereby guaranteeing its high calculation
inverse process,
calculation rather
process, than
rather
the
than time-consuming
the time-consuming
training speed. tuning process
tuning usingusing
process the gradient
the gradientdescent method,
descent thereby
method, guaranteeing
thereby its high
guaranteeing its
training
high speed.samples with N sets can be represented as follows:
training
Training
speed.
Training samples
Training samples with
with NN sets
sets can
can bebe represented
represented as as follows:
follows:
ℵN {( xi , ti ) | xi ∈ R n , ti ∈ R m }N
N
= (1)
ℵ N ℵ=N {({x(ix, it,it)|
= i ∈R
i )x| ix∈ ∈RRmm}i}==11iN=1
Rnn,,tii ∈ i
(1)
(1)
where xi is a n × 1 input vector, and ti is a m × 1 target vector. A single-hidden layer network with
where xxii is
activation
where aa nn ×
isfunction
× 11 input vector,
θ(x) can
input and ttii is
be modeled
vector, and isas: m×
aa m × 11 target vector. A
target vector. A single-hidden
single-hidden layer
layer network
network with
with
activation function
activation function θ(x)θ(x) can
can be
be modeled
modeled as: as:
N

∑ β ϑ ( w , b ,=
N x ) o= , j 1, , N (2)
∑
i i i i j j
N
β ϑ ( w , b ,=
i =1 x ) o= , j 1, , N (2)
e
i =1∑ βi ϑi (wi , bi , x j ) = o j , j = 1, . . . , N
i i i i j j
(2)

where N represents the number of hidden nodes, wi = [ wi1 , wi2 , , win ] and bi are parameters of
i = 1 T

where N  represents the number of hidden nodes, wi = [ wi1 , wi2 , , win ]T and bi are parameters of
the hidden nodes, and β i = [ β i1 , β i2 , , β in ]T is the weighting vector connecting T the i-th hidden nodes
where N represents
the hidden the number
nodes, and β i = [ β i1 ,of
β i2hidden
, , β inT]nodes,
T i = [ wi1 , wi2
is thewweighting vector in ] and bi the
, . . . , wconnecting are i-th
parameters of the
hidden nodes
e
and the nodes,
hidden output and
nodes. The
βi = [ βstructure
, β , . . . ,of
β the
] ELM
is the algorithm
weighting isvector
shown in Figure 2.
connecting the i-th hidden nodes
and the output nodes. The structure
i1 i2
ofinthe ELM algorithm is shown in Figure 2.
and the output nodes. The structure of the ELM algorithm is shown in Figure 2.
Hidden
layer
Hidden
Input
)
, x jlayer
layer ( wi , bi
Input
θi ) β
layer ( wi , bi , x j
θi β Output
layer
Output
layer
. Oj
. Oj
. .
. .
.
.

Figure 2. Structure of the single hidden layer feed-forward neural network.

Figure 2. Structure of the single hidden layer feed-forward neural network.
Figure 2. Structure of the single hidden layer feed-forward neural network.
Energies 2018, 11, 143 5 of 15

For each training sample, the calculated outputs are expected to remain the same as the actual
N
results, which can be represented as ∑ ko j − t j k = 0. Therefore, β i , bi , wi is to be determined to satisfy
j =1
N
e
∑ β i ϑi (wi , bi , x j ) = t j , j = 1, . . . , N, which can be written in matrix form as follows:
i =1

Hβ = T (3)

where:
 
ϑ ( w i , bi , x 1 ) ϑ (w Ne , b Ne , x1 )
H (w1 , . . . , w Ne , b1 , . . . , b Ne , x1 , . . . , x Ne ) = 
 .. .. 
,
. ··· . 
ϑ ( w i , bi , x N ) ϑ (w Ne , b Ne , x N ) N × Ne
 
T  T 
β t1
 .1 
β= .  T =  ... 
 
 . 
βTe e tTN N ×m
N
N ×m

Hence, the training process is to determine βi , wi , and bi . For the ELM training algorithm, the
parameters wi and bi are fixed before training with random values. βi is the only undetermined
parameter. If the number of hidden neurons is equal to the number of training samples, then βi can
be calculated easily [29]. Although the number of hidden neurons is less than the number of training
samples, generally, precise values of βi , wi . and bi , may not exist. The mathematical model can be
transformed to minimize the cost function given below, where β i , bi , wi (the optimal approximate
solution) are to be determined:
!2
N L
E= ∑ ∑ βei ϑ(wei · Xj + ebi ) − t j (4)
j =1 i =1

For fixed w bi , βei the approximate solution can be easily determined [29].
ei , e
The ELM-based transient stability assessment model is constructed through three main steps.
First, historical samples and offline simulation samples, which consist of system features, are collected
for feature selection. The recorded actual data can be used to supplement the prior knowledge base to
ensure the adaptation to a practical power system. Second, Fisher discrimination is implemented to
determine the features that are closely related to the system transient stability status from the initial
features with samples in the prior knowledge base. In this manner, the required data scale is reduced
and the computing efficiency is improved.
In this paper, the initial features (listed in Table 1) can be obtained by PMU measurements.

Table 1. Initial features for Fisher discrimination evaluation.

Notation Description
δi rotor angle variation of generator i
Vi , θ i voltage amplitude and phase angle variation of bus i
Pin,i , Qin,i active and reactive power injection variation of bus i
PL,i , QL,i active and reactive power flow of line i
Finfo (ftime and fduration ) fault information (fault time and fault duration)
L, Lnode load level

Finally, the selected features are regarded as significant features and applied for training of the
ELM-based classifier. Once the classifier is established, the ELM transient stability assessment model
Energies 2018, 11, 143 6 of 15

is constructed. In practical implementation, the transient stability status is judged by the assessment
model according to the PMU measurements of the significant features within a short time delay.

3.2. ELM-Based Transient Stability Assessment Model

As mentioned above, the accuracy of a machine learning technique-based method is greatly
influenced by the scale and quality of the training samples, as well as the training approach. To improve
the accuracy, the TF method [13] is applied to amend the outputs of the ELM-based stability assessment
model in certain cases; this approach indicates a coordination mechanism exists in the hybrid transient
stability assessment model. The TF method can be constructed using the following two steps:

1. Generator rotor angle trajectories in various scenarios are collected from simulation and historical
samples and comprise the prior sample database for one certain generator. The difference between
samples is measured by the vector distance and is defined as follows:
!1
T −1 2
f ( xi,T , x j,T ) = ∑ (xi,t − x j,t )2 (5)
t =0

where xi,T , xj,T are samples i and j with data of T moments. The samples are considered
approximate if the vector distance between them is less than a certain threshold.
2. Approximate trajectory samples are grouped using the hierarchical clustering algorithm.
The trajectory having the minimum vector distance with other trajectories is defined as the
standard trajectory of the group. Standard trajectories comprise the standard trajectory pattern
library, and the measured trajectory is subsequently matched with pattern trajectories in this
pattern library to predict the rotor angle trajectory tendency. If the rotor angle of any generator is
beyond 180◦ with the predicted trajectory, then it is judged as a transient instability. Otherwise,
it is stable in the transient process.

As the essential component of the hybrid transient stability assessment model, the coordination
mechanism is designed for sampling time and result amendment. The ELM- and the TF-based
assessment model require five cycles and 15–20 cycles of sampling time, respectively. If the
measurements are sufficient in quantity and quality, then the assessment model is initiated immediately.
With the conclusion that output of the ELM-based model is problematic in a certain interval in [19],
the result amendment procedure is triggered if the output of the ELM-based model is in an unreliable
interval. In this case, the TF-based method is used for this amendment procedure and determines the
final transient stability assessment result. The determined result is taken as the start-up criterion to
decide whether to start up the controlled islanding surface searching algorithm and the splitting action.

4. Optimal Splitting Surface Search Algorithm

The splitting surface searching approach is always used to determine weakly-connected generator
groups, among which an out-of-synchronization state typically occurs. In this section, a modified
electrical distance index is used to determine those nodes having vague connection with separated
generator groups, which could significantly simplify the solution space without feasible solution loss.
The ‘optimal’ algorithm involves satisfying some other constraints and considering an active power
imbalance of the isolated system as the objective function in the proposed searching algorithm.

4.1. Nodes Classification

Conventionally, a power system is considered to be an undirected graph with edge weight and the
accumulated value of the edge weights between nodes is regarded as the electrical distance. In most
cases, the reactance of lines is defined as the edge weight for network simplification, and nodes in the
electrical network can be divided into two types. Nodes near one certain group of coherent generators
Energies 2018, 11, 143 7 of 15

in electrical
Energies 2018, distance
11, 143 are labeled as normal nodes (NN). Otherwise, if nodes share a similar connection 7 of 15
Energies
with 2018, 11, groups
multiple 143 of coherent generators, then they are classified as public nodes (PN). 7 of 15

However, considering
However, considering reactance
reactance only
onlyininthetheelectrical
electrical distance
distance is not complete,
is not theoretically.
complete, The
theoretically.
However, considering reactance only in the electrical distance is not complete, theoretically. The
detailed
The detailedanalysis of aof
analysis specific casecase
a specific is shown below:
is shown below:
detailed analysis of a specific case is shown below:
Figure33shows
Figure showsa astructure
structureofof a simple
a simple power
power system,
system, where
where thethe reactance
reactance of two
of two lineslines is equal,
is equal, i.e.,
Figure 3 shows a structure of a simple power system, where the reactance of two lines is equal,
1 =x
xi.e., x2, and
x12 ,=and the the
loadload on buses
on buses m and
m and n are
n are S12S
S1 = = 2S 2, where
2 , where S1S=1 =2S2S
2 ,2,i.e.,
i.e.,the
theactive
activepower
powertransmitted
transmitted
i.e., x1 = x2, and the load on buses m and n are S1 = 2S2, where S1 = 2S2, i.e., the active power transmitted
betweennode
between nodei iand
andmmisistwice
twiceas
asmuch
muchas asthat
thatbetween
betweennode nodeiiandandn. n. Therefore,
Therefore,thetheelectrical
electricaldistance
distance
between node i and m is twice as much as that between node i and n. Therefore, the electrical distance
betweennodes
between nodesi iandandmmisisequal
equaltoto that
that between
between nodes
nodes i and
i and n when
n when only only reactance
reactance is considered
is considered in the in
between nodes i and m is equal to that between nodes i and n when only reactance is considered in
the electrical
electrical distance.
distance. However,
However, this approach
this approach neglectsneglects
the effectthe of
effect of transmitted
transmitted power power on theInlines.
on the lines. this
the electrical distance. However, this approach neglects the effect of transmitted power on the lines.
In thisthe
paper, paper,
ratiothe ratio
of the of the reactance
reactance and the
and the active active
power of power
the lines ofisthe lines is considered
considered as the edgeasweight
the edge to
In this paper, the ratio of the reactance and the active power of the lines is considered as the edge
weight to
evaluate evaluate
the electricalthedistance.
electricalWe
distance.
take theWe take test
IEEE-9 the IEEE-9
system testas an system
exampleas an
to example
show theto show the
advantage
weight to evaluate the electrical distance. We take the IEEE-9 test system as an example to show the
advantage
of this approach of thisasapproach
follows: as follows:
advantage of this approach as follows:
x1 m
x1 m
S1=P1+jQ1
G S1=P1+jQ1
G S2=P2+jQ2
i x2 S2=P2+jQ2
i x 2
n
n
Figure3.3.A
Figure Asimple
simplepower
powersystem
systemfor
forillustration.
illustration.
Figure 3. A simple power system for illustration.
The effect of the edge weight on the results of determining the splitting surface is analyzed on
effect of
The effect ofthe
theedge
edgeweight
weighton onthe
theresults
results
ofof determining
determining thethe splitting
splitting surface
surface is analyzed
is analyzed on
on the
the IEEE-9 test system [25]. Figure 4 shows the edge weight defined with xij and xij/Pij of the IEEE-9
the IEEE-9
IEEE-9 test test system
system [25].[25]. Figure
Figure 4 shows
4 shows the edge
the edge weight
weight defined
defined with with xijxand
xij and ij /P x
ij of
ij /Pthe
ij of the IEEE-9
IEEE-9 test
test system in (a) and (b), respectively. The minimum accumulating value of the edge weight from
test system
system in (a)inand
(a) (b),
andrespectively.
(b), respectively. The minimum
The minimum accumulating
accumulating value ofvalue of theweight
the edge edge weight from nodefrom 9
node 9 to nodes 1 and 3 are calculated, as summarized in Table 2.
node
to 9 to1nodes
nodes and 3 1areand 3 are calculated,
calculated, as summarized
as summarized in Tablein
2. Table 2.
6 7 8 0.0586 6 7 8 0.0689
6 7 8 0.0586 6 7 8 0.0689
2 3 2 3
2 0.1700 3 2 0.2931 3
0.1700 0.2931
5 9 5 9
5 9 5 9
0.0920 0.3286
0.0920 0.3286
4 4
4 0.0576 4 0.0800
0.0576 0.0800
1 1
1 (a) 1 (b)
(a) (b)
Figure 4. Graph-model of the IEEE-9 test system. (a) edge weight with xij index; (b) edge weight with
Figure 4. 4. Graph-model
Graph-model of
of the
the IEEE-9
IEEE-9 test
test system.
system. (a)
(a) edge
edge weight with xxij index;
weight with (b) edge weight with
xij/Pij index.
Figure ij index; (b) edge weight with
x ij/Pij index.
xij /Pij index.
Table 2. Minimum edge weight cumulative summation.
Table 2. Minimum edge weight cumulative summation.
Table 2. Minimum edge weight cumulative summation.
Minimum Accumulating Edge Weight Value
Edge Weight Definition Minimum Accumulating Edge Weight Value
Edge Weight Definition Node 9 to 1 Node 9 to 3
Edge Weight xDefinition Node 9 to
Minimum 1
Accumulating NodeWeight
Edge 9 to 3 Value
ij 0.1496 0.2286
xij 0.1496
Node 9 to 1 0.2286
Node
xij/Pij 0.4086 0.36209 to 3
xij/Pij 0.4086 0.3620
xij 0.1496 0.2286
Table 2 indicates that edge weight defined with xij has smaller weight
x ij /P ij 0.4086 value from node 9 to 1,
0.3620
Table 2 indicates that edge weight defined with xij has smaller weight value from node 9 to 1,
i.e., node 9 is closer to node 1 than to node 3. While nodes 9 to 3 have relatively stronger connections
i.e., node 9 is closer to node 1 than to node 3. While nodes 9 to 3 have relatively stronger connections
for the edge
Table weight defined
2 indicates that edgeby xweight
ij/Pij. This difference
defined with xmay
ij has
lead to a loss
smaller of feasible
weight solutions
value from nodein network
9 to 1, i.e.,
for the edge weight defined by xij/Pij. This difference may lead to a loss of feasible solutions in network
reduction
node for the
9 is closer to reactance-only
node 1 than to node scenario. Hence,
3. While further
nodes 9 toresearch is necessary
3 have relatively to consider
stronger reactance
connections for
reduction for the reactance-only scenario. Hence, further research is necessary to consider reactance
not only in the electrical distance.
the edge weight defined by xij /Pij . This difference may lead to a loss of feasible solutions in network
not only in the electrical distance.
In this
reduction forpaper, xij/Pij is defined
the reactance-only as the edge
scenario. Hence,weight
further of research
the undirected graph-model.
is necessary Reduction
to consider reactance of
In this paper, xij/Pij is defined as the edge weight of the undirected graph-model. Reduction of
the only
not islanding
in thesolution
electricalspace is executed using the following steps:
distance.
the islanding solution space is executed using the following steps:
1. Calculating the minimum accumulative edge weight among nodes using the Floyd algorithm. The
1. Calculating the minimum accumulative edge weight among nodes using the Floyd algorithm. The
electrical distance between the load nodes and the coherent generator groups can be written as:
electrical distance between the load nodes and the coherent generator groups can be written as:
Energies 2018, 11, 143 8 of 15

In this paper, xij /Pij is defined as the edge weight of the undirected graph-model. Reduction of
the islanding solution space is executed using the following steps:

1. Calculating the minimum accumulative edge weight among nodes using the Floyd algorithm. The
electrical distance between the load nodes and the coherent generator groups can be written as:

1 Nm
Nm j∑
Di → m = Wij (6)
=1

where Nm is the number of generators in coherent group m, and Wij is the minimum accumulative
edge weight between load node i and generator node j in group m.
2. The difference in the electrical distance between load node i and two generator groups m1 and
m2 can be estimated as:
γ = Di → m 1 − Di → m 2 (7)

A threshold ε is pre-determined according to statistical data and operational experience, with the
public nodes occupying approximately 20% of the whole nodes. Node i is defined as the public node
of the two generator groups when |γ| < ε. If |γ| > ε, then node i is defined as a normal node and node
i is classified as a closely connected generator group according to the positive-negative value of γ.

4.2. Optimal Splitting Surface Searching Algorithm

On basis of node classification, a normal node is assigned to a corresponding generator group.
The essence of searching the optimal splitting surface becomes distribution of the public nodes into
suitable generator groups. To determine the optimal splitting surface, the minimum active power
imbalance and the minimal power-flow disruption are applied as the objective functions in Equations
(8) and (9), respectively:
min ∆P = | PG − PL | (8)
 
min ∑  Pij  (9)
i ∈V1 ,j∈V2

where PG and PL denote the active power of the generators and the loads in the isolated system
respectively, and Pij is the active power transmitted between nodes i and j. The minimum active
power imbalance of the isolated system determines whether the isolated system can recover from
load-generation imbalance. Minimum power-flow disruption can reduce the disruption caused by
variation of the power-flow dispatch. In this paper, the minimum active power imbalance of the isolated
system is taken as the objective function. The set of public nodes is distributed into suitable generator
groups using the Breadth First Search (BFS) algorithm for optimal splitting location determination.

5. Controlled Islanding Implementing Strategy

The problems of when and where to separate the power system are solved by implementing the
proposed strategy. In this strategy, the splitting action is supposed to be executed once the splitting
location is determined by the optimal splitting surface searching algorithm. The implementation of the
proposed controlled islanding strategy is shown in Figure 5; the details of the steps are as follows:

Step 1: Start the prediction program after a disturbance is detected.

Step 2: Capture the PMU measurements required by the hybrid transient stability assessment model.
Note that the sampling time is different for the ELM- and for the TF-based models in this
hybrid assessment model.
Step 3: Initiate the splitting criterion-deciding procedure using the hybrid transient stability
assessment model. In this model, the amending procedure using the TF model is activated if
the output of the ELM model shows uncertainty.
Energies 2018, 11, 143 9 of 15

Step 4: Judge whether to start the optimal islanding surface searching algorithm according to the
splitting criterion, i.e., the system stability status. If the power system tends to lose stability,
then
Energies 2018, go to step 5; otherwise, go to step 1.
11, 143 9 of 15
Step 5: The optimal splitting surface searching program starts. Series procedures, including
Step 5: system
The optimal splitting
undirected surface
graph searchingnodes
construction, program starts. Series
classification, andprocedures, including
optimal solution system
calculating,
undirected
are conducted.graph construction, nodes classification, and optimal solution calculating, are
conducted.
Step 6: The optimal splitting surface is determined, and the splitting action is immediately executed.
Step 6: The The program
optimal ends,
splitting surface
and the is determined,
next round begins. and the splitting action is immediately
executed. The program ends, and the next round begins.

Start

Monitoring Splitting surface

system stability status determining algorithm
Surface searching
No
Any disturbance? Constructing system
undirected graph
Yes
Classifying nodes to
PMU Measurements reduce solution space

Hybrid stability Search of optimal

assessment model splitting surface

Splitting Criterion End

Yes
Result is unreliable?

No Curve fitting
No
Status is unstable?

Yes

Figure 5. Flowchart of the proposed controlled islanding implementation strategy.

Figure 5. Flowchart of the proposed controlled islanding implementation strategy.

It can be inferred from Figure 5 that the hybrid transient stability assessment model and the
optimal
It cansplitting surface
be inferred fromsearching
Figure 5algorithm are essential
that the hybrid transient parts of theassessment
stability proposed strategy.
model and In the
the
hybrid model,
optimal splitting thesurface
ELM method searchingoutput assessment
algorithm result is taken
are essential in most
parts of cases andstrategy.
the proposed the TF method
In the
amendment
hybrid model, is the
taken
ELM in certain
methodcasesoutputto realize optimization
assessment result is of the in
taken combination
most casesof andcomputing speed
the TF method
and accuracy.
amendment Moreover,
is taken in certainthecases
splitting surface
to realize with a minimized
optimization active power
of the combination imbalance
of computing speedcanandbe
determined rapidly via reduction of the possible solution space. Sequentially, fast and reliable
accuracy. Moreover, the splitting surface with a minimized active power imbalance can be determined
detection
rapidly viaofreduction
transientof instability
the possibleandsolution
determination of the optimal
space. Sequentially, splitting
fast surface
and reliable is realized,
detection thereby
of transient
achieving and
instability the purpose of the of
determination proposed
the optimalstrategy.
splitting surface is realized, thereby achieving the purpose
of the proposed strategy.
6. Case Study
6. Case Study
The New England 39-bus test system is used as the testbed for validation of the proposed
The New
controlled England
islanding 39-busTotest
strategy. systemthe
construct is used as the
transient testbed
stability for validation
assessment model,of10,000
the proposed
samples
controlled islanding strategy. To construct the transient stability assessment
are generated using the Monte Carlo method; 9000 of these samples are applied for training model, 10,000 samples
and
are generated using the Monte Carlo method; 9000 of these samples are
trajectory clustering, and the rest are applied for testing. The transient stability status will beapplied for training and
trajectory
determined clustering,
to be unstableand the rest as
as soon aretheapplied for testing.
rotor angle The transient
of any generator exceedsstability status willwith
180° compared be
determined to be unstable as soon as the rotor angle of any generator exceeds 180 ◦ compared with the
the reference generator.
reference generator.
Initial features for feature selection are summarized in Table 1. These initial features are
processed features
Initial via Fisher for feature selection to
discrimination aredetermine
summarized thein significant
Table 1. These initial features
features. Figure are processed
6 shows the
via
importance value for each feature, and the top 100 features of the total are chosen as the inputs for
Fisher discrimination to determine the significant features. Figure 6 shows the importance value for
each feature,
the ELM and the
classifier top 100 features of the total are chosen as the inputs for the ELM classifier training.
training.
Energies 2018, 11, 143 10 of 15
Energies 2018, 11, 143 10 of 15
Energies 2018, 11, 143 10 of 15
δ V θ Pin Qin PL QL Finfo L Lnode
0.12
δ V θ Pin Qin PL QL Finfo L Lnode
0.12
0.1
0.1

Value
0.08

Value
0.08

Importance
0.06

Importance
0.06
0.04
0.04 Importance Value Threshold of
0.02 SignificantValue
Importance Features: 0.00191
Threshold of
0.02 Significant Features: 0.00191
0.00191
0
0.00191
00 50 100 150 200 250 300
0 50 100 Feature
150 No. 200 250 300
Feature No.
Figure 6. Features
Figure 6. Features selection with Fisher
selection with Fisher discrimination.
discrimination.
Figure 6. Features selection with Fisher discrimination.
Figure 6 shows that most significant features are dynamic information, e.g., voltage magnitude
Figure 66 shows
Figure shows that
that most
most significant
significant features
features are
are dynamic
dynamic information, e.g.,
e.g., voltage magnitude
and phase angle variation, generator rotor angle variation, andinformation,
fault information. voltage
These magnitude
significant
and
and phase
phase angle variation, generator rotor angle variation, and fault information. These
These significant
features areangle
usedvariation,
as the ELM generator
trainingrotor
and angle
modelvariation, and fault information.
inputs. Examinations on the ELM-based significant
transient
features are
features are used
used as the
the ELM
asmodel ELM training and
and model inputs. Examinations on
on the
the ELM-based transient
stability assessment showtraining
that accuracymodel inputs.
can reach Examinations
97.46% with 1350 hidden ELM-based
nodes transient
at maximum.
stability assessment
stability assessmentmodel
modelshowshowthat
thataccuracy
accuracycan
can reach97.46%
97.46%with
with 1350 hidden nodes
at at maximum.
Misjudged cases of the ELM-based transientreach 1350
stability assessment hidden
model nodes
are small maximum.
in quantity,
Misjudged
Misjudged cases
cases of the ELM-based transient stability assessment model are small in quantity,
but are inevitable. Theofdistribution
the ELM-based transient
of the output stability
value forassessment
correct andmodel are small
incorrect in quantity,
assessment of the
but are
but are inevitable.
inevitable. The distribution
The distribution of the output
output value
value for
for correct
correct and
and incorrect
incorrect assessment
assessment of of the
the
ELM-based model is further studied,ofasthe
shown in Figure 7.
ELM-based model is further studied, as shown in
ELM-based model is further studied, as shown in Figure 7. Figure 7.
Correctly Judged Cases Wrongly Judged Cases
2
Correctly Judged Cases Wrongly Judged Cases
2
1.8
Model

1.8
Model

1.6
Assessment

1.6
1.4
Assessment

1.4
1.2
1.2
1
of ELM

1
of ELM

Unreliable

0.8
Unreliable

0.8
Value

0.6
Value

0.6
0.4
Output

Interval

0.4
Output

Interval

0.2
0.2
0
0 100 200 300 400 500 600 700 800 900 1000
0
0 100 200 300 400 Testing
500Samples
600 700 800 900 1000
Testing Samples
Figure 7. ELM-based stability assessment model output distribution with testing samples.
Figure
Figure 7.
7. ELM-based
ELM-based stability
stability assessment
assessment model
model output
output distribution
distribution with
with testing
testing samples.
samples.
This finding indicates that the output value for the most misjudged case is within a specific
Thisranging
interval finding indicates that the output value forunreliable
the most misjudged casepaper,
is within a specific
This findingfrom 0 to 1.0,
indicates which
that is defined
the output as the
value for the most interval.
misjudged In this the unreliable
case is within a specific
interval
interval ranging
is set as from
(0, 0 to 1.0, which is defined as the unreliable interval. In this paper, the unreliable
0.8).
interval ranging from 0 to 1.0, which is defined as the unreliable interval. In this paper, the unreliable
interval is set
In the as (0,transient
hybrid 0.8). stability assessment model, the TF method is applied to amend the ELM
interval is set as (0, 0.8).
In
assessmentthe hybrid
results transient
in the stabilityinterval.
unreliable assessment Whenmodel, the TF
number methodnodes
of hidden is applied
of theto amend the ELM
In the hybrid transient stability assessment model, the TF method is applied toELM
amendalgorithm
the ELM is
assessment
set as 500, the results in
accuracythe unreliable interval. When number of hidden nodes of the ELM algorithm is
assessment results in the is improved
unreliable from 94.1%
interval. When(ELM number only model) nodes
of hidden to 100% (hybrid
of the ELM model).
algorithm The
is
set as 500,
results show the accuracy
that 355 setsisofimproved
1000 test from 94.1%
samples with (ELM only model)
problematic outputs to are
100% (hybridusing
processed model).the The
TF-
set as 500, the accuracy is improved from 94.1% (ELM only model) to 100% (hybrid model). The results
results
based thatshow that
model. 355 method
sets of 1000 test samples with problematic outputs results
are processed using ofthe TF-
show 355ThesetsTFof 1000 testrectifies
samples the ELM-based
with problematicmodel assessment
outputs are processed63using times, 45TF-based
the which
based model.
are corrected The
from TF method
the stable rectifies
condition the ELM-based
to the unstable model assessment results 63 times, 45 of which
model. The TF method rectifies the ELM-based model condition.
assessmentThe expected
results computing
63 times, time are
45 of which for
are
each corrected
test sample from the stable
increases from condition
0.041 s to theonly
(ELM unstable
model)condition.
to 0.136 The
s expected
(hybrid computing
model). Compared timewith
for
corrected from the stable condition to the unstable condition. The expected computing time for each
each test sample
the ELM-only increases
method, thefrom 0.041 stime
computing (ELM is only model)
amodel)
slightly to 0.136 sthe(hybrid model). Compared with
test sample increases from 0.041 s (ELM only tolonger,
0.136 s but
(hybridaccuracy
model). is highly
Compared improved.
with the
the ELM-only method, the computing time is a slightly longer, but the accuracy is highly improved.
ELM-only method, the computing time is a slightly longer, but the accuracy is highly improved.
6.1. Case 1: Assessment by ELM Model is Reliable
6.1. Case 1: Assessment by ELM Model is Reliable
In this case, a three-phase short circuit fault is set to occur at 0 s in the middle of lines 23–24 and
In thisatcase,
is cleared 0.4 s.aThe
three-phase
generators short
formcircuit fault is set
two coherent to occur
groups: at34,
{33, 0 s 35,
in the
36} middle
and {30,of31,lines 23–24
32, 37, 38, and
39},
is cleared at 0.4 s. The generators form two coherent groups: {33, 34, 35, 36} and {30, 31, 32, 37, 38, 39},
Energies 2018, 11, 143
Energies 2018, 11, 143 11 of 15 11 of 15

in which generators 33–36 are the critical crew. Identification of the instability using the ELM-based
6.1. Case 1: Assessment by ELM Model is Reliable
transient stability assessment model requires 0.036 s. On the basis of the node classification
In this case, a three-phase short circuit fault is set to occur at 0 s in the middle of lines 23–24 and is
mentioned above, the evaluation value for the electrical distance between load nodes and two
cleared at 0.4 s. The generators form two coherent groups: {33, 34, 35, 36} and {30, 31, 32, 37, 38, 39},
generator
in groups are listed
which generators in Table
33–36 are the3.
critical crew. Identification of the instability using the ELM-based
transient stability assessment model requires 0.036 s. On the basis of the node classification mentioned
above, theTable 3. Difference
evaluation value for of
thethe minimum
electrical edgebetween
distance weights cumulative
load nodes andsummation.
two generator groups
are listed in Table 3.
Nodes γ Nodes γ Nodes γ
Table 3.1Difference
0.0392 11
of the minimum 0.0332
edge 21
weights cumulative−0.0749
summation.
2 0.0276 12 0.0318 22 −0.0772
Nodes Nodes Nodes
3 0.0221
γ
13 0.0318
γ
23 −0.0859
γ
14 0.0369
0.0392 14
11 0.0295
0.0332 24
21 −0.0790
− 0.0749
2 0.0276 12 0.0318 22 −0.0772
35 0.0379
0.0221 15
13 -0.0699
0.0318 25
23 0.0300
−0.0859
46 0.0372
0.0369 16
14 -0.0721
0.0295 26
24 0.0442
−0.0790
57 0.0372
0.0379 17
15 −0.0663
-0.0699 27
25 0.0442
0.0300
6 0.0372 16 -0.0721 26 0.0442
78 0.0379
0.0372 18
17 −0.0587
−0.0663 28
27 0.0545
0.0442
89 0.0547
0.0379 19
18 −0.0764
−0.0587 29
28 0.0551
0.0545
10
9 0.0327
0.0547 20
19 −0.0803
−0.0764 29 0.0551
10 0.0327 20 −0.0803

The threshold ε is set as 0.035. Therefore, nodes 2, 3, 10, 11, 12, 13, 14, and 25 are defined as public
The threshold ε is set as 0.035. Therefore, nodes 2, 3, 10, 11, 12, 13, 14, and 25 are defined as public
nodes. Normal nodes 15, 16, 17, 18, 19, 20, 21, 22, 23, and 24 belong to the generator group {33, 34, 35,
nodes. Normal nodes 15, 16, 17, 18, 19, 20, 21, 22, 23, and 24 belong to the generator group {33, 34, 35,
36}. The 36}.
calculated optimal
The calculated splitting
optimal surface
splitting surfacesolution isshown
solution is shown in in Figure
Figure 8 with
8 with the redthe red dashed
dashed line line
(Scheme(Scheme
1). 1).

G G
37
30
25 26 28 29
2 Scheme2 27 38
1 3 18 G
17 Scheme 1
G 21
39 15 16

4
24 G
14 36

5 13 23
6 12
9 19

11 20 22
7 10

8 31 32 34 33 35
G G G G G

Figure 8. Optimal
Figure 8. Optimalislanding surface
islanding surface (case
(case 1). 1).

It can beIt inferred fromfrom

can be inferred Table 3 that
Table 3 thatnode
node27 hasaashorter
27 has shorter electrical
electrical distance
distance from
from the the generator
generator
group {30, 31, {30,
group 32, 31,
37, 32,
38,37,
39}38,than fromfrom
39} than {33,{33,
34,34,
35,35,36}.
36}.InInthe
thealgorithm defining
algorithm defining thethe reactance
reactance as as the
edge weight, node 27 is classified into the generator group {33, 34, 35, 36}, and the final splitting
the edge weight, node 27 is classified into the generator group {33, 34, 35, 36}, and the final splitting
solution is shown in Figure 8 with the green solid line (Scheme 2). To show the strength of the proposed
solution is shown in Figure 8 with the green solid line (Scheme 2). To show the strength of the
algorithm, the algorithm defining reactance as the edge weight is contrasted from the aspect of active
proposed algorithm,
power imbalancetheandalgorithm
power-flow defining
disruption reactance
in Table 4. as the edge weight is contrasted from the
aspect of active power imbalance and power-flow disruption in Table 4.

Table 4. Comparison of the islanding surface searching algorithms.

Islanding Strategy Active Power Imbalance (p.u.) Power-Flow Disruption (p.u.)

14–15, 3–18, 17–27 0.329/0.130 0.805
14–15, 17–18, 26–27 0.901/1.360 4.913
Energies 2018, 11, 143 12 of 15

Table 4. Comparison of the islanding surface searching algorithms.

Islanding Strategy Active Power Imbalance (p.u.) Power-Flow Disruption (p.u.)

14–15, 3–18, 17–27 0.329/0.130 0.805
14–15, 17–18, 26–27 0.901/1.360 4.913

The table indicates that algorithm considering only reactance in electrical distance may lead to
loss of feasible
Energies 2018,
2018, solutions and, thus, affect the final optimal solution. In this condition, power12flow
11, 143
143 of 15
15
Energies 11, 12 of
disruption is higher than the result found using the proposed algorithm.
Figures 999and
Figures
Figures and10
and 10show
10 showthe
show thetrajectories
the trajectories
trajectories of
ofof the
thethe critical
critical generator
generator
critical group
group
generator {33,
{33,{33,
group 34, 36}
34, 35,
34, 35, using
35, 36} using
36} using the
the two
the
two islanding
islanding strategies.
strategies. Variation
Variation of the of the generator
generator rotor rotor
angle angle
shows shows
that the that the
islanding islanding
solution
two islanding strategies. Variation of the generator rotor angle shows that the islanding solution solution
presented
presented
in this paper
presented in this
in this paper
paper
retains retains
system system
stability
retains system stability better.
better.
stability better.

300
300 Generator Group 33-36
Generator Group 33-36
Angle/°

250
RotorAngle/°

250
200
200
150
150
GeneratorRotor

100
100
Generator

50
50
00
-50
-50
00 0.2
0.2 0.4
0.4 0.6 0.8
0.6 0.8 1.0
1.0 1.2
1.2 1.4
1.4
Time/s
Time/s
Figure 9.
Figure
Figure 9. Variation
9. Variation of
Variation of the
of the generator
the generator rotor
generator rotor angles
rotor angles (strategy
angles (strategy 1).
(strategy 1).
1).

300
300 Generator Group 33-36
Generator Group 33-36
Angle/°

250
RotorAngle/°

250
200
200
150
GeneratorRotor

150
100
100
Generator

50
50
00
-50
-50
00 0.2
0.2 0.4
0.4 0.6 0.8
0.6 0.8 1.0
1.0 1.2
1.2 1.4
1.4
Time/s
Time/s
Figure 10. Variation
Figure Variation of
of the
the generator
generator rotor
rotor angles
angles (strategy
(strategy 2).
2).
Figure 10.
10. Variation of the generator rotor angles (strategy 2).
6.2. Case
6.2. Case 2:2: ELM
ELM Output
Output Located
Located in in the
the Unreliable
Unreliable Interval
Interval
6.2. Case 2: ELM Output Located in the Unreliable Interval
In this
In this case,
case, aa three-phase
three-phase short
short circuit
circuit fault
fault occurs
occurs at
at 0.1
0.1 ss at
at the
the middle
middle ofof line
line 6–7
6–7 and
and continues
continues
for In sthis
0.4 case,
until it isacleared.
three-phase
This short
fault circuit
leads tofault occurs at 0.1 s atbetween
desynchronization the middlethe of line 6–7 and
generator groupcontinues
{31, 32}
32}
for 0.4 s until it is cleared. This fault leads to desynchronization between the generator group {31,
for 0.4
and the s until it
the remaining is cleared.
remaining generators. This
generators. Itfault leads
It takes to desynchronization
takes approximately
approximately 0.3 0.3 ss of between
of computing the
computing time generator
time toto ensuregroup
ensure an {31, 32}
an accurate
accurate
and
and the remaining
assessment usingusing the generators.
the TF
TF method It takes
method because approximately
because thethe ELM
ELM output 0.3
output iss of computing
is located
located in
in the time to
the unreliableensure an
unreliable interval accurate
interval [0,
[0, 0.8].
0.8].
assessment
assessment
The using the
optimal TF method
splitting solution because
is the ELM
shown in output
Figure 11, ismarked
locatedwith
in the
theunreliable
red interval
dashed line [0, 0.8].
(Scheme
The optimal splitting solution is shown in Figure 11, marked with the red dashed line (Scheme
1). Thesolution
The optimalissplitting solution
contrasted withistheshown
the resultsin Figure
(Scheme11,2marked
in Figure
Figure with
11)the red dashedbyline
determined the(Scheme
algorithm 1).
1). The solution is contrasted with results (Scheme 2 in 11) determined by the algorithm
The solution
presented in is
in the contrasted
the literature with
literature [30], the
[30], whose results
whose objective (Scheme
objective function 2 in
function is Figure
is the 11)
the minimumdetermined
minimum power-flow by the algorithm
power-flow disruption;
disruption; aa
presented
presented
comparison of in the
of the literature
the results
results is [30],
is shown whose
shown in objective
in Table
Table 5. 5. function is the minimum power-flow disruption;
comparison
a comparison of the results is shown in Table 5.
Table 5.
Table 5. Comparison
Comparison of
of islanding
islanding surface
surface searching
searching algorithms.
algorithms.

Islanding Strategy
Islanding Strategy Power-Flow Disruption
Power-Flow Disruption (p.u.)
(p.u.) Time (ms)
Time (ms)
3–4, 9–39, 14–15
3–4, 9–39, 14–15 1.2733
1.2733 ≈11
≈11
3–4, 8–9, 14–15
3–4, 8–9, 14–15 1.2678
1.2678 ≈40
≈40

It is
It is apparent
apparent that
that the
the algorithm
algorithm that
that determines
determines the
the minimum
minimum cut
cut ensures
ensures the
the minimum
minimum power-
power-
flow disruption.
flow disruption. Compared
Compared withwith this
this algorithm,
algorithm, the
the proposed
proposed algorithm
algorithm in
in this
this paper
paper can
can also
also avoid
avoid
Energies 2018, 11, 143 13 of 15
Energies 2018, 11, 143 13 of 15

G G
37
30
25 26 28 29
2 27 38
1 3 18 G
17
G 21
39 15 16

Scheme 2
4
24 G
Scheme 1 36
14

5 13 23
6 12
9 19

11 20 22
7 10

8 31 32 34 33 35
G G G G G

Figure 11. Optimal islanding surface (case 2).

Figure 11. Optimal islanding surface (case 2).

The validation results illustrate that the ELM-based transient stability assessment model is
Table 5. Comparison of islanding surface searching algorithms.
computationally efficient. The TF method is able to amend the problematic results of the ELM-based
assessment model; Islanding
however,Strategy
this amendment requires
Power-Flow a slightly
Disruption (p.u.) longer
Time computing
(ms) time. The
algorithm proposed for optimal splitting location determination reduces the solution space using the
3–4, 9–39, 14–15 1.2733 ≈11
modified electrical distance
3–4, 8–9,concept
14–15 while not losing 1.2678
any feasible solutions. ≈40

7. Conclusions
It is apparent that the algorithm that determines the minimum cut ensures the minimum
In this paper,
power-flow a controlled
disruption. Compared islanding strategy
with this consisting
algorithm, of prompt
the proposed transientinstability
algorithm assessment
this paper can also
and optimal splitting surface determination was described.
avoid high power-flow disruption by taking active power on transmission lines into consideration
(1).the
in A concept
hybrid of transient stability
the electrical assessment
distance. Using the model basedalgorithm,
proposed on the ELM and TFspace
the solution methods was
is reduced
based constructed to determine the start-up criterion. In this model, the ELM method is used for prompt
on the electrical distance, resulting in a reduced computing time while providing a consistent
stability
(nearly assessment,
identical) result. and the TF method is employed to enhance the reliability of the assessment
results.
The validation results illustrate that the ELM-based transient stability assessment model is
(2). A rapid algorithm
computationally comprised
efficient. The TF of network
method reduction
is able and optimal
to amend solution searching
the problematic results ofwas
the proposed
ELM-based to
determine
assessment the however,
model; splitting surface. In this algorithm,
this amendment requires afeasible
slightly solution loss in thetime.
longer computing network reduction
The algorithm
process
proposed forisoptimal
avoidedsplitting
by modifying
locationthedetermination
evaluation index of the the
reduces electrical
solutiondistance.
space using the modified
electrical distance concept while not losing any feasible solutions.
Using the proposed method, both the start-up criterion and the optimal splitting surface can be
determined
7. Conclusions online with high reliability. The results showed that the proposed controlled islanding
strategy is effective in preventing transient instability. As the scenarios considered in this work were
In this paper, a controlled islanding strategy consisting of prompt transient stability assessment
designed optimally, the sample generation problem and data acquisition risks in a realistic scenario
and optimal splitting surface determination was described.
were ignored. Further research can be dedicated to these problems to achieve reliable performance
of theAproposed
(1). controlled
hybrid transient islanding
stability strategy
assessment in a practical
model based on implementation.
the ELM and TF methods was constructed
to determine the start-up criterion. In this model, the ELM method is used for prompt stability
Acknowledgments: This work was supported in part by the National Science Foundation of China under grants
assessment, and the TF method is employed to enhance the reliability of the assessment results.
51577030 and 61503197, Science and Technology Project of SGCC (XT71-16-032）, and the National Key Research
(2). A rapid algorithm
and Development Program comprised
of China inofBasic
network reduction
Research and optimal
Class under solution searching was proposed
grant 2017YFB0903000.
to determine the splitting surface. In this algorithm, feasible solution loss in the network reduction
Author Contributions: Yi Tang and Qi Wang contributed in developing the concepts of this research study.
process is avoided by modifying the evaluation index of the electrical distance.
Feng Li and Chenyi Zheng performed this research study. Yingjun Wu provided technical support in the
simulations of this research study. All of the authors read and approved the final manuscript.
Using the proposed method, both the start-up criterion and the optimal splitting surface can be
Conflicts of Interest:
determined The authors
online with declare no The
high reliability. conflict of interest.
results showed that the proposed controlled islanding
strategy is effective in preventing transient instability. As the scenarios considered in this work were
References
designed optimally, the sample generation problem and data acquisition risks in a realistic scenario
were
1. ignored.
Taylor, Further
C.W.; research
Erickson, D.C. can be dedicated
Recording to thesethe
and analyzing problems to achieve
July 2 cascading reliable
outage. performance
IEEE of
Comput. Appl.
the proposed controlled islanding strategy in
Power 1997, 10, 26–30, doi:10.1109/67.560830. a practical implementation.
2. You, H.; Vittal, V.; Wang, X. Slow Coherency-Based Islanding. IEEE Trans. Power Syst. 2004, 19, 483–491.
3. Senroy, N.; Heydt, G.T. A conceptual framework for the controlled islanding of interconnected power
systems. IEEE Trans. Power Syst. 2006, 21, 1005–1006, doi:10.1109/TPWRS.2006.873009.
Energies 2018, 11, 143 14 of 15

Acknowledgments: This work was supported in part by the National Science Foundation of China under grants
51577030 and 61503197, Science and Technology Project of SGCC (XT71-16-032), and the National Key Research
and Development Program of China in Basic Research Class under grant 2017YFB0903000.
Author Contributions: Yi Tang and Qi Wang contributed in developing the concepts of this research study. Feng Li
and Chenyi Zheng performed this research study. Yingjun Wu provided technical support in the simulations of
this research study. All of the authors read and approved the final manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Taylor, C.W.; Erickson, D.C. Recording and analyzing the July 2 cascading outage. IEEE Comput. Appl. Power
1997, 10, 26–30. [CrossRef]
2. You, H.; Vittal, V.; Wang, X. Slow Coherency-Based Islanding. IEEE Trans. Power Syst. 2004, 19, 483–491.
[CrossRef]
3. Senroy, N.; Heydt, G.T. A conceptual framework for the controlled islanding of interconnected power
systems. IEEE Trans. Power Syst. 2006, 21, 1005–1006. [CrossRef]
4. Franco, R.; Sena, C.; Taranto, G.N.; Giusto, A. Using synchrophasors for controlled islanding—A prospective
application for the Uruguayan power system. IEEE Trans. Power Syst. 2013, 28, 2016–2024. [CrossRef]
5. Li, L.; Liu, Y.; Mu, H.; Yu, Z. Out-of-Step splitting scheme based on PMUs. In Proceedings of the 3rd
International Conference on Deregulation and Restructuring and Power Technologies (DRPT 2008), Nanjing,
China, 6–9 April 2008.
6. Wang, S.; Lu, S.; Zhou, N.; Lin, G.; Elizondo, M.; Pai, M.A. Dynamic-feature extraction, attribution, and
reconstruction (DEAR) method for power system model reduction. IEEE Trans. Power Syst. 2014, 29,
2049–2059. [CrossRef]
7. Bhui, P.; Senroy, N. Real-Time Prediction and Control of Transient Stability Using Transient Energy Function.
IEEE Trans. Power Syst. 2017, 32, 923–934. [CrossRef]
8. Bian, D.; Kuzlu, M.; Pipattanasomporn, M.; Rahman, S. Analysis of communication schemes for Advanced
Metering Infrastructure (AMI). In Proceedings of the 2014 IEEE Power and Energy Society General Meeting,
National Harbor, MD, USA, 27–31 July 2014.
9. Shrestha, B.; Gokaraju, R.; Sachdev, M. Out-of-Step protection using state-plane trajectories analysis.
IEEE Trans. Power Deliv. 2013, 28, 1083–1093. [CrossRef]
10. Hazra, J.; Reddi, R.K.; Das, K.; Seetharam, D.P.; Sinha, A.K. Power grid transient stability prediction using
wide area synchrophasor measurements. In Proceedings of the 2012 3rd IEEE PES Innovative Smart Grid
Technologies Europe (ISGT Europe 2012), Berlin, Germany, 14–17 October 2012.
11. Gurusinghe, D.R.; Rajapakse, A.D. Post-disturbance transient stability status prediction using synchrophasor
measurements. IEEE Trans. Power Syst. 2016, 31, 3656–3664. [CrossRef]
12. Ohura, Y.; Suzuki, M.; Yanagihashi, K.; Yamaura, M.; Omata, K.; Nakamura, T.; Mitamura, S.; Watanabe, H.
A predictive out-of-step protection system based on observation of the phase difference between substations.
IEEE Trans. Power Deliv. 1990, 5, 1695–1704. [CrossRef]
13. Liu, X.D.; Li, Y.; Liu, Z.J.; Huang, Z.G.; Miao, Y.Q.; Jun, Q.; Jiang, Q.Y.; Chen, W.H. A novel fast transient
stability prediction method based on PMU. In Proceedings of the 2009 IEEE Power and Energy Society
General Meeting, Calgary, AB, Canada, 26–30 July 2009.
14. Sobajic, D.J.; Pao, Y.-H. Artificial neural-net based dynamic security assessment for electric power systems.
IEEE Trans. Power Syst. 1989, 4, 220–228. [CrossRef]
15. Amjady, N.; Majedi, S.F. Transient stability prediction by a hybrid intelligent system. IEEE Trans. Power Syst.
2007, 22, 1275–1283. [CrossRef]
16. Zhou, Y.; Wu, J.; Yu, Z.; Ji, L.; Hao, L. A hierarchical method for transient stability prediction of power
systems using the confidence of a SVM-based ensemble classifier. Energies 2016, 9, 778. [CrossRef]
17. Amraee, T.; Ranjbar, S. Transient instability prediction using decision tree technique. IEEE Trans. Power Syst.
2013, 28, 3028–3037. [CrossRef]
18. Vittal, V.; Senroy, N.; Heydt, G.T. Decision Tree Assisted Controlled Islanding. IEEE Trans. Power Syst. 2006,
21, 1790–1797. [CrossRef]
Energies 2018, 11, 143 15 of 15

19. Xu, Y.; Dong, Z.Y.; Meng, K.; Zhang, R.; Wong, K.P. Real-time transient stability assessment model using
extreme learning machine. IET Gener. Transm. Distrib. 2011, 5, 314–322. [CrossRef]
20. Yusof, S.B.; Alden, R.T.H.; Rogers, G.J. Slow coherency based network partitioning including load buses.
IEEE Trans. Power Syst. 1993, 8, 1375–1382. [CrossRef]
21. Yang, B.; Vittal, V.; Heydt, G.T. Slow-coherency-based controlled islanding—A demonstration of the approach
on the August 14, 2003 blackout scenario. IEEE Trans. Power Syst. 2006, 21, 1840–1847. [CrossRef]
22. Song, H.; Wu, J.; Wu, K. A wide-area measurement systems-based adaptive strategy for controlled islanding
in bulk power systems. Energies 2014, 7, 2631–2657. [CrossRef]
23. Wang, X.M. Slow Coherency Grouping Based Islanding Using Minimal Cut Sets and Generator Coherency
Index Tracing Using the Continuation Method. Ph.D. Thesis, Iowa State University, Ames, IA, USA, 2005.
24. Sun, K.; Zheng, D.Z.; Lu, Q. Splitting strategies for islanding operation of large-scale power systems using
OBDD-based methods. IEEE Trans. Power Syst. 2003, 18, 912–923. [CrossRef]
25. Ding, L.; Gonzalez-Longatt, F.M.; Wall, P.; Terzija, V. Two-step spectral clustering controlled islanding
algorithm. IEEE Trans. Power Syst. 2013, 28, 75–84. [CrossRef]
26. Lin, Z.Z.; Wen, F.S. Discussion on “two-step spectral clustering controlled islanding algorithm”. IEEE Trans.
Power Syst. 2014, 29, 413. [CrossRef]
27. El-Zonkoly, A.; Saad, M.; Khalil, R. New algorithm based on CLPSO for controlled islanding of distribution
systems. Int. J. Electr. Power Energy Syst. 2013, 45, 391–403. [CrossRef]
28. Liu, L.; Liu, W.; Cartes, D.A.; Chung, I.Y. Slow coherency and Angle Modulated Particle Swarm Optimization
based islanding of large-scale power systems. Adv. Eng. Inform. 2009, 23, 45–56. [CrossRef]
29. Huang, G.-B.; Zhu, Q.-Y.; Siew, C.-K. Extreme learning machine: Theory and applications. Neurocomputing
2006, 70, 489–501. [CrossRef]
30. Tang, F.; Jia, J.; Wang, B.; Liao, Q.; Wang, J.; Zhu, Z.; Liu, Y. Study on a fast controlled partition scheme based
on improved Dinic max-flow strategy. In Proceedings of the 2015 IEEE Power and Energy Society General
Meeting, Denver, CO, USA, 26–30 July 2015.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).