Electric Power Systems Research 130 (2016) 124–131

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Method for distributed generation anti-islanding protection based on

singular value decomposition and linear discrimination analysis
G. Marchesan ∗ , M.R. Muraro, G. Cardoso Jr, L. Mariotto, C.D.L. da Silva
Federal University of Santa Maria, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Anti-islanding protection is one of the most important requirements for the connection of distributed
Received 8 June 2015 generators in power systems. This paper proposes an algorithm to detect unintentional islanding in power
Received in revised form 23 July 2015 systems with distributed generation. It is based on the singular value decomposition and linear discrim-
Accepted 28 August 2015
ination analysis to differentiate frequency oscillations in synchronous generators caused by islanding
Available online 15 September 2015
from those caused by non-islanding events. The algorithm requires a very low number of mathematical
operations, which is suitable for relay purposes. This is possible because most of the operations are in the
Keywords:
training process and are made off-line. The performance of the proposed algorithm is evaluated for dif-
Anti-islanding protection
Frequency oscillation
ferent scenarios and load conditions in IEEE 34 Node Test Feeder. The algorithm is able to detect islanding
Distributed generation with active power mismatch of 1.6% of DG nominal power. The pattern recognition also prevents undue
Passive protection tripping, ensuring great robustness for the method.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction the disturbances inserted in the distribution system may cause

The distributed generation (DG) unintentional islanding can
cause life-threatening, power quality deterioration, due to poor
voltage and frequency regulation, and damage to system equip-
ment and its loads. For these reasons, the islanded operation of
distribution systems is normally not allowed and the anti-islanding
protection is necessary for the connection of DGs to distribution
networks. According to IEEE Std 1547 [1] islanding detection must
occur up to 2 s after the island formation.
To avoid said problems, many power utilities request reclosers
with transferred trip in the DG connection point. Other utili-
ties request dedicated feeders with transfer trip. Although these
communication-based methods are more effective than local tech-
niques, they can suffer with communication problems and its
implementation implies very high costs. The local methods were
proposed as alternatives to methods based on communication and
they can be divided into three categories: Active, hybrid and passive
methods. The active methods inject small signals in the distribution
system or force the DG to an abnormal situation, whilst the con-
nection to the system keeps it under normal conditions. In general
these methods have a small non detection zone (NDZ); however,
∗ Corresponding author. Tel.: +55 5599072819.
E-mail address: gutomarchesan@gmail.com (G. Marchesan).
power quality deterioration [2,3]. If the distribution system has
multiple generators connected very close to others with similar
techniques, they might cause interference into each other and
impair the performance of these techniques. In [4] the perform-
ances of active frequency drifty methods are evaluated for multi
inverter system. The non-detection zone increases when the inver-
ters try to drift the frequency in opposite directions. In [5] an
alternative solution for the interference problem has been pro-
posed for inverter based generations. The method injects a high
frequency signal in a master inverter and this signal is used in all
other slave inverters for island detection. The slaves operate in a
cancelation mode avoiding interference between inverters. Other
solution for multi-DG was proposed in [6]. The method aims to
estimate overall transient stiffness to distinguish prior- and post-
islanding. Each DG perturbs the system at different frequency, thus
avoiding the interference inter DG; however, these perturbations
can also cause power quality deterioration. An average absolute fre-
quency deviation value based active islanding detection technique
is proposed in [7]. The method has zero NDZ, detects stable island
formation without forcing the island to lose its stable operation,
and has a detection time up to 100 ms, however, the tolerance to
short circuits is not evaluated yet.
Due to lower cost of passive methods, it is common the use
of anti-islanding schemes using protection functions such as the
rate of change of frequency (ROCOF) [8], which is one of the

http://dx.doi.org/10.1016/j.epsr.2015.08.025
0378-7796/© 2015 Elsevier B.V. All rights reserved.
 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131 125

fastest passive protection algorithms used in relays for island- In the following sections, the mathematical fundamentals
ing detection. Other techniques such as under/over-frequency, regarding synchronous machine model, SVD, LDA, and the pro-
under/over-voltage [9], and vector-surge [8] are also used, although posed algorithm are presented, along with the simulations results.
these techniques are effective for islanding conditions with large
power imbalance, passive methods may fail or spend too much 2. Synchronous machine models
time detecting low power imbalance. In addition, events like short
circuits and switching of large blocks of load can cause islanding On a synchronous machine operating in steady state, the relative
erroneous islanding detection. In [10], an Island detection method position between rotor and resulting magnetic field remain almost
for inverter connected generators based on a dynamic estimator is constant. When a sudden disturbance occurs, the angle between
proposed, which measures the current amplitude and phase. them oscillates dynamically according to the swing equation given
Aiming to get the best sides of active and passive methods, by (1).
the hybrid methods often use a passive method that identifies
2H d2 ı dı
a transient condition and the DG starts to cause a disturbance +D = Pm − Pe (1)
to destabilize parameters in case of islanding. Hybrid techniques ω0 dt 2 dt
also have been proposed [11]. The technique changes GD active where ı is the relative rotor angle, t is the time, H is the generator
power only when it cannot differentiate clearly islanding from inertia constant, D is the damping coefficient, ω0 is the DG syn-
other events, thus making the islanding detection easier. In [12], chronous speed, and Pm and Pe are mechanical input and electric
the method uses a ROCOF relay to decide when the DG frequency power output of the DG, respectively.
set is changed. The main problem with hybrid methods is they still
depend on passive method threshold. 2.1. Frequency variation during non-islanding events
A passive technique based on decision tree was proposed in [13],
but it uses a very large set of parameters which hinders its imple- When a small disturbance occurs in the electrical system, the
mentation. Wavelet has also been used in an intelligent technique DG oscillates and returns to its original state after some time. The
proposed in [14,15]. The wavelets extract voltage and current fea- electrical power injected by DG in the distribution system can be
tures and use a decision tree to identify the islanding. The method written as
proposed by [14] uses a very large data set for training, which is
a hard work considering that the method should be retrained at Pe = Pmax sin ı (2)
each topology change. In [16], a wavelet design for island detec- A small perturbation ı in ı, from the initial operating position
tion is proposed. The algorithm is simpler than [14,15] and has ı0 can be represented by
lower computational effort using only the voltage and six wavelet
coefficients. ı = ı0 + ı (3)
Techniques based on synchronous machine oscillation fre-
Due to this perturbation, the swing equation (1) can be lin-
quency estimation are proposed by [17,18]. The methods
earized and rewritten as
respectively use windows of 350 ms and 500 ms for estimating
damping and oscillation frequency estimation, which demand too 2H d2 ı dı
+D + Ps ı = 0 (4)
much time for islanding detection purposes. The method proposed ω0 dt 2 dt
in [17] aims to estimate the signal parameters using a TLS-ESPRIT Ps is known as the synchronizing power coefficient and is defined
algorithm. The method needs to minimize a cost function which by the equation
demands a relatively high processing time. In [18], a different
solution for estimating damping and frequency of oscillation is Ps = Pmax cos ı0 (5)
proposed using Tufts–Kumaresan method. The method of [18]
Solving the differential equation shown in (4), [19] shows that
is not recursive, requiring much less computational effort; how-
the frequency deviation from nominal synchronous speed is given
ever, given its large window, the island detection time is more
by (6).
than 600 ms. Although it may be less than the 2 s required by
IEEE standard 1547 [1] this island detection time may be greater dı ωn ı(0) −ς ωt
ω = = − e n sin ωt
d
(6)
than recloser time. In general, utilities use auto reclosing times dt 1 − ς2
around 500 ms, which can result in out of synchronism reclos-
ing. where

In this paper, a faster algorithm using smaller windows than ωd = ωn 1 − ς2 (7)
previous methods is proposed. As well as [17,18], the proposed 
methodology uses the oscillation frequency to characterize island- D ω0
ing. However, the proposed methods do not use the value of the
ς= (8)
2 2HPs
oscillation frequency, but the shape of the electrical frequency. 
In other words, the adopted strategy is to use the well-known ω0
ωn = Ps (9)
information that during islanding the frequency behaves like an 2H
exponential or a low-frequency oscillation due to the governor From (6), one can see that the frequency is given by a damped
effect. During events where the DG is connected to the system, sinusoidal waveform.
like short circuits and load switching, the frequency oscillates
at the damped natural frequency. The frequency is decomposed 2.2. Frequency variation during islanding events
using singular value decomposition (SVD), and its components
are analyzed with linear discrimination analysis (LDA) using the During an islanding event, the DG loses connection with the
generalized Rayleigh quotient. The main contribution of pro- main system and, therefore, the synchronizing coefficient is 0. In
posed technique is an improvement in speed when compared this way, (4) can be rewritten as (10).
with [17,18]. This was possible since the approach uses a pat-
tern recognition technique with a small window, providing a faster 2H d2 ı dı
+D = P (10)
detection. ω0 dt 2 dt
126 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131

A - Non-Island B - Island In this way, V and U are determined by the eigen vectors of AT A
and AAT , respectively, and  is determined by the square root of
its eigen values. The values of  i are always non-negative; in this
way, the sign of U and V must be checked in order to ensure their
Frequency

Frequency
self-consistency.
As seen in the previous section, frequency during an islanding
is clearly different from other events that may happen during a
connected situation. For the constitution of the training set, the fre-
quency is analyzed half cycle (8.333 ms) after the instant when the
Time Time absolute value of frequency deviation of 60 Hz exceeds the thresh-
old of 0.05 Hz. In this way, as expressed in (20), f is a vector which
Fig. 1. Frequency behavior in: (A) non-island events; (B) island events. contains the absolute value of frequency deviation, where r is the
instant where the threshold is exceeded and m is the window size,
P is the power variation due to the islanding; in other words, equal to 10 cycles of 60 Hz (166.667 ms).
the transmitting power in the electrical system split point. In this     
case, P is assumed as constant during the islanding. P is assumed f = fr − fr  fr+1 − fr  . . . fr+m−1 − fr  T (20)
positive when the electrical power in the split point is flowing from
Matrix A can be rewritten as given in (21)
the main system to DG.
 
Since the rotor angle is synchronized with the stator mag- A = fni (1) ... fni (N) fi (1) ... fi (N) (21)
netic field before islanding, the two initial conditions for (10) are
ı(0) = 0 and dı(0)
dt
= 0. Solving (10), the equation for electrical fni is the frequency deviation in non-islanding events, fi is the
frequency deviation is obtained. frequency deviation during islanding events, and N is the total num-
ber of events of each case. In this paper, training set A is defined by
dı P
ω = = (1 − e−ω0 Dt/2H ) (11) the simulation of 24 cases. The non-islanding set is composed of
dt D
12 short circuits and load switching in several system positions.
Comparing (6) to (11), it can be observed that the frequency The islanding set is composed of 12 islanding with several lev-
of the DG behaves differently. During DG parallel operation with els of power mismatches. The algorithm should be trained for
the system, the frequency tends to oscillate at the damped natural each system; however, the number of simulations necessary are
frequency, ωd , as shown in Fig. 1A. Disregarding the voltage con- low. The tests show the algorithm performance does not improve
trollers, governors, and the load dynamic during islanding, which significantly for more than 24 cases involving different power mis-
may change due to voltage and frequency variation, the frequency matches and disturbances situations such as short circuits and load
does not oscillate during an islanding, but it is given by an expo- switching.
nential response, as shown in Fig. 1B. Applying the singular value decomposition in A, each column of
U can be seen as one base of A, and V* can be seen in a geometric
3. The singular value decomposition point of view as a scaling and rotation of base U to reconstruct
matrix A. In the next section, V* will be called E. According to
Singular value decomposition (SVD) is factorization of a matrix (13), the first singular values are the most significant; therefore, U
in to constitutive components. It is widely used in common can be limited to its firsts columns, and E to its firsts rows. In the
applications such as least square fitting data and pseudo-inverse calculations below, only the first 20 columns of U and the first 20
calculation. In electric power systems it was employed in some rows of E are being considered.
fields such as in [20,21]. In this paper, the SVD is used for pattern
recognition [22] and identifying islanding amongst other events 4. Linear discrimination analysis
involving distributed generation. The SVD decomposes an m by n
matrix A in a unitary matrix V n × n, a diagonal matrix  m × n, and Given the definition shown in (22)
a unitary matrix U m × m, as shown in (12).
E = ˙V ∗ (22)
A = U˙V ∗ (12)
Analyzing the construction of matrix A in (21), it can be deduced
The symbol * stands for the conjugate transpose. The diagonal
that the first N columns of E represent non-islanding events (Eni ),
matrix  is composed of non-negative singular values  i , where
and the last N columns represent islanding events (Ei ), as shown in
1 ≤ 2 ≤ 3 · · · ≤ r , r = min (m, n) (13) (23)
 
Considering T stands for the transpose matrix, we have E = Eni (1) ··· Eni (N) Ei (1) ··· Ei (N) (23)
T ∗ T ∗ ∗ ∗
A A = (U˙V ) (U˙V ) = V ˙U U˙V (14) This paper proposes the use of statistics about E in order to
T 2 ∗ characterize whether one event is an Islanding or not. The lin-
A A = V˙ V (15)
ear discrimination analysis was used with this goal. The LDA was
and proposed in [23] in the taxonomy context [22]. The main idea of
LDA is to take a high dimensional problem and transform it into a
AAT = (U˙V ∗ )(U˙V ∗ ) = U˙V ∗ V ˙U ∗
T
(16)
lower dimensional, with a transform W. Considering this, W should
T 2 ∗ project E in such a way that maximizes the distances inter-class
AA = U˙ U (17)
of the E data, and minimizes the distance intra-class of E data. To
Multiplying (15) and (17) on the right by V and U, respectively, achieve this, a criterion known as generalized Rayleigh quotient
two eigen value problems can be obtained. was used (24).
2
AT AV = V ˙ (18) W T SB W
2
W = argmax (24)
T
AA V = U˙ (19) W W T SW W
 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131 127

SW is the within class scatter matrix and SB is the between class

scatter matrix given by

SB = (i − ni )(i − ni )T (25)


N
T

N
T
SW = (Eni (j) − ni )(Eni (j) − ni ) + (Ei (j) − i )(Ei (j) − i )
j=1 j=1
(26)

ni and i are the mean values of E vectors according to

1
N
ni = Eni (j) (27)
N
j=1

1
N
i = Ei (j) (28)
N
j=1

The solution of the generalized Rayleigh quotient is given by the

eigen value problem

SB W = SW W (29)

W is given by the eigen vector with associated eigen value that has
the largest value. Being W is a unitary vector, the projection of E
over W is given by (30) and (31).
Fig. 2. Proposed island detection algorithm.
Pi = W T Ei (30)

Pni = W T Eni (31) waits a time equivalent to three cycles of 60 Hz, i.e., Th2 is 384 for
a sampling frequency of 7680 Hz.
If the mean value of Pi is greater than Pni , the threshold is given
by
6. Tests and results
min (Pi ) + max (Pni )
Th = (32)
2 To evaluate the performance of the proposed method, it has been
Otherwise, W should be multiplied by −1 and the threshold is used on the IEEE 34 node distribution test system, presented in
given by Fig. 3 [24]. A synchronous generator is connected through a trans-
former to bus 854. The transformer data are given in Table 1. The
min (−Pi ) + max (−Pi ) diesel generator controls the power factor to 0.98 inductive its data
Th = (33)
2 are presented in Table 2, and DG voltage and frequency regulators
are given in [25]. The excitation system model used in [25] is a static
5. The algorithm
excitation equivalent to IEEE type ST2 Model [26], and the gover-
nor was the same used in [27]. A 0.2 MVA load with 0.92 inductive
The definition of W transform and the threshold presented in
power factor is connected directly to the DG node.
previous section are made offline. In this section, the algorithm
The proposed method is compared with one of the best-known
that produces the DG trip signal is presented and discussed. See
islanding detection methods, the rate of change of frequency
flowchart presented in Fig. 2. The frequency is continually moni-
(ROCOF). The rate of change of frequency was calculated at each
tored, and is identified when the frequency deviation from 60 Hz
frequency sample by Eq. (36)
exceeds the threshold of 0.05 Hz.
If the frequency deviation exceeds the threshold, the vector con- df
= (fr − fr−1 )fsampling (36)
taining synchronous machine frequency measurements f, should dt
be projected in the bases of the training set. Taking f and multi- fsampling is the sampling frequency. When ROCOF exceeds the
plying by U−1 gives the coefficient C (34), which is equivalent to E threshold, a time counter starts. This temporization is important
obtained in the training set. Since U is a unitary, the transpose can
be used instead of the inverse. Table 1
Transformer data.
C = U T f (34)
Parameter Value
Applying the transform W over C, P can be obtained (35).
854-Point of common coupling
P = WT C (35) Three phase transformer
Rated power 3.0 MVA
If P is greater than Th, an islanding event is characterized and Nominal frequency 60 Hz
a timer counter Z starts; otherwise, it is a non-islanding event. Rated voltage 24.9/2.4 kV
Connection D/yn
When Z reaches the threshold Th2, the trip signal is sent to dis- Vector group Phase shift 1 × 30◦
connect the DG. The temporization given by Th2 is important to Positive sequence reactance (X1) 0.059371 p.u.
avoid an unwanted trip due to an misclassified event. A greater Positive sequence resistance (R1) 0.008667 p.u.
delay increases the robustness of the method, but also the detec- Zero sequence short circuit impedance 0.06 p.u.
Zero sequence short circuit resistance 0.0087 p.u.
tion time. In the studies carried out in this paper, the algorithm
128 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131

Table 3
Rate of change of frequency methods configuration.

ROCOF 1 ROCOF 2 ROCOF 3 ROCOF 4

df/dt (Hz/s) 0.500 1.500 2.500 0.500

Delay (s) 0.150 0.050 0.050 0.150
Voltage constraint (p.u) – – – 0.8

remains greater than 0.8 p.u. ROCOF 1, 2 and 3 do not use any
voltage restriction.
Two load conditions for the IEEE 34-bus distribution system
were considered, 100% and 50%. The loads shown in the test sys-
tem represent load condition 100%, whereas load condition 50% is
obtained by reducing them accordingly. Several islanding condi-
tions were tested and are shown in Table 4, which presents the line
switched, the load condition, the DG generated power, the active
and reactive switching interrupted power, and the protections trip-
ping time. It is possible to see that the proposed method did not fail
in any of the simulated cases. All algorithms based on rate of change
of frequency failed twice due to very low transmitted active power
on the switching point. The proposed technique island detection
time is generally around 220 ms, slightly greater than ROCOF 1.
Table 5 shows the methods’ performance during a single-phase
to ground short circuit sustained in the system for 350 ms. After
this time the fault line is disconnected, thus causing the DG island-
ing. Table 5 shows the short circuited bus and the fault resistance.
The islanding detection time is the difference between the protec-
tion trip times and 350 ms; in this way, negative times represent
protection trips before DG islanding, i.e., they represent failed trips.
The proposed method did not fail in any simulated case presented
in this table. ROCOF 1 failed once and had some detection times
Fig. 3. IEEE 34 Node Test Feeder [24]. greater than 500 ms. ROCOF 2 failed in almost all cases, presenting
negative islanding detection times. ROCOF 3 failed seven times. It
detected the islanding during the short circuit in four times and did
because of the high sensitivity of ROCOF protection, and helps to not trip during real islanding in three cases.
avoid unwanted Trips for short time transients in the distribu- Table 6 shows the algorithms’ performance for temporary phase
tion system, especially short circuits. The method was adjusted to to ground short circuit. The fault remains during 350 ms and
four typical settings, shown in Table 3. ROCOF4 operates if voltage disappears spontaneously without any switching. The proposed
algorithm as well as ROCOF 1 and ROCOF 4 worked well in all
simulated cases. ROCOF 2 and ROCOF 3 failed in 12 and 3 cases,
Table 2 respectively.
Generator parameters.
Table 7 presents the tests of load switching, caused by the open-
Parameter Value Parameter Value ing lines. The ROCOF 2 failed in case 5 identifying an islanding
Reference machine Not flag Direct axis reactance 1.56 p.u.
erroneously. The proposed method and ROCOF methods worked
(Xd ) well in all tests.
Mode of local Voltage Quad. axis reactance 1.06 p.u. Fig. 4 presents the time detection for islanding caused by switch-
voltage (Xq ) ing in bus 802.
controller
In Fig. 4A the power is changed in such a way that the inter-
Dispatch – voltage 1.0 p.u. Direct axis transient 0.26 p.u.
reactance (Xd  ) rupted reactive power remains 0 and active power changes from
Nominal apparent 3.125 MVA Direct axis 0.15 p.u. 0.016 to 0.36 p.u. regarding DG nominal power. The interrupted
power subtransient reactance power is changed by a load introduced in bus 802. One can see
(Xd  ) the proposed method is able to detect the small active power mis-
Nominal voltage 2.4 kV Quadrature axis 0.15 p.u.
subtransient reactance
match, 0.016 p.u. In Fig. 4B the active power mismatch is set to 0
(Xq  ) and the reactive power is changed. The methods based on ROCOF
Power factor 0.8 Direct axis 3.7 s did not identify the island for any of the tested cases. However, the
short-circuit transient proposed methodology was able to detect islanding with reactive
time-constant (Td  )
power mismatch up to 0.12 p.u. Although the proposed method
Connection yn Direct axis short-circuit 0.05 s
subtransient may be slower than ROCOF for large active power mismatch, it
time-constant (Td  ) is much more efficient for small values, detecting islanding were
Inertia time 1.071 s Quadrature axis 0.05 s other methods were not able to do. Also, as shown in Fig. 4A, the
constant (rated short-circuit proposed method can be faster than ROCOF 1 and ROCOF 4 for small
to Sgn) H subtransient
time-constant (Tq  )
power mismatches.
Leakage reactance 8.8% Main flux saturation – 0.17 p.u. Due to frequency pattern recognition, the proposed method
Sg10 avoids the nuisance tripping that would happen in other frequency
Rotor type Salient pole Main flux saturation – 0.60 p.u. based relays such as ROCOF and Under/Over frequency. This is
Sg12
an advantage since, for instance, in case of a big generation trip
 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131 129

Table 4
Performance of islanding detection methods during islanding events.

Operating characteristic of the system Islanding detection time (s)

Opened line Load (%) PG (MW) POP (MW) QOP (MVAr) Proposed ROCOF 1 ROCOF 2 ROCOF 3 ROCOF 4

800–802 100 2.5 −0.38 −0.11 0.221 0.150 0.050 0.050 0.150
830–854 100 2.5 −0.75 −0.18 0.221 0.150 0.050 0.050 0.150
800–802 50 2.5 −1.32 −0.67 0.221 0.150 0.050 0.050 0.150
830–854 50 2.5 −1.61 −0.71 0.221 0.150 0.050 0.050 0.150
800–802 100 1.0 1.12 0.13 0.220 0.150 0.050 0.050 0.150
830–854 100 1.0 0.72 0.04 0.218 0.150 0.050 0.050 0.150
800–802 50 1.0 0.05 −0.31 0.248 Not det. Not det. Not det. Not det.
830–854 50 1.0 −0.13 −0.49 0.396 Not det. Not det. Not det. Not det.

Table 5
Performance of islanding detection methods during phase to ground short-circuits, sustained for 350 ms, and followed by islanding.

Operating characteristic of the system Islanding detection time (s)

Short circuit bus Opened line Z fault (

) Load (%) PG (MW) Proposed ROCOF 1 ROCOF 2 ROCOF 3 ROCOF 4

802 802–806 0 100 2.5 0.350 0.150 −0.223 0.307 Not det.
802 802–806 60 100 2.5 0.205 0.150 −0.230 0.050 Not det.
816 816–824 0 100 2.5 0.327 0.189 0.323 0.578 Not det.
816 816–824 60 100 2.5 0.213 0.150 −0.230 Not det. Not det.
830 830–854 0 100 2.5 0.389 0.208 0.440 Not det. Not det.
830 830–854 60 100 2.5 0.232 0.150 −0.230 Not det. Not det.
802 802–806 0 50 2.5 0.194 0.401 −0.223 0.050 Not det.
802 802–806 60 50 2.5 0.202 Not det. −0.230 0.050 Not det.
816 816–824 0 50 2.5 0.220 0.490 0.050 0.050 Not det.
816 816–824 60 50 2.5 0.219 0.540 −0.230 0.050 Not det.
830 830–854 0 50 2.5 0.220 0.490 0.050 0.050 Not det.
830 830–854 60 50 2.5 0.219 0.527 −0.230 0.050 Not det.
802 802–806 0 100 1.0 0.219 0.128 −0.227 −0.219 Not det.
802 802–806 60 100 1.0 0.220 0.150 −0.229 0.050 Not det.
816 816–824 0 100 1.0 0.219 0.150 0.050 0.050 Not det.
816 816–824 60 100 1.0 0.220 0.150 −0.230 0.050 Not det.
830 830–854 0 100 1.0 0.219 0.150 0.050 0.050 Not det.
830 830–854 60 100 1.0 0.220 0.150 −0.230 0.050 Not det.
802 802–806 0 50 1.0 0.192 0.123 −0.300 −0.221 Not det.
802 802–806 60 50 1.0 0.197 0.150 −0.230 0.050 Not det.
816 816–824 0 50 1.0 0.211 0.150 −0.223 0.050 Not det.
816 816–824 60 50 1.0 0.194 0.150 −0.231 −0.223 Not det.
830 830–854 0 50 1.0 0.217 0.150 0.050 0.050 Not det.
830 830–854 60 50 1.0 0.194 0.150 −0.231 −0.222 Not det.

Table 6
Performance of islanding detection methods during temporary phase to ground short circuit, 350 ms.

Operating characteristic of the system Islanding detection time (s)

Short circuit bus Z fault Load (%) PG (MW) Proposed ROCOF 1 ROCOF 2 ROCOF 3 ROCOF 4

830 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
830 60 100 2.5 Not det. Not det. 0.120 Not det. Not det.
852 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
852 60 100 2.5 Not det. Not det. 0.124 Not det. Not det.
842 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
842 60 100 2.5 Not det. Not det. 0.125 Not det. Not det.
830 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
830 60 50 1.0 Not det. Not det. 0.120 Not det. Not det.
852 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
852 60 50 1.0 Not det. Not det. 0.123 Not det. Not det.
842 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
842 60 50 1.0 Not det. Not det. 0.124 Not det. Not det.
830 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
830 60 100 2.5 Not det. Not det. 0.120 Not det. Not det.
852 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
852 60 100 2.5 Not det. Not det. 0.124 Not det. Not det.
842 0 100 2.5 Not det. Not det. Not det. Not det. Not det.
842 60 100 2.5 Not det. Not det. 0.125 Not det. Not det.
830 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
830 60 50 1.0 Not det. Not det. 0.119 0.128 Not det.
852 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
852 60 50 1.0 Not det. Not det. 0.122 Not det. Not det.
842 0 50 1.0 Not det. Not det. Not det. Not det. Not det.
842 60 50 1.0 Not det. Not det. 0.123 Not det. Not det.
130 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131

Fig. 4. (A) Anti island methods time detection and NDZ for active power variation. (B) Anti island methods time detection and NDZ for reactive power variation.

Table 7 References
Load switching tests.

Operating characteristic of the system [1] IEEE standard for interconnecting distributed resources with electric
power systems. IEEE Std 1547-2003 vol. no. (2003) 1, 28, doi:10.1109/
Case Opened line Load (%) PG (MW) POP (MW) QOP (MVAr) IEEESTD.2003.94285.
[2] D. Velasco, C. Trujillo, G. Garcera, E. Figueres, An active anti-islanding method
1 854–852 100 2.5 1.511 0.107 based on phase-PLL perturbation, IEEE Trans. Power Electron. 26 (4) (2011)
2 834–842 100 2.5 0.565 −0.376 1056–1066.
3 854–852 50 2.5 0.754 −0.381 [3] Jae-Hyung Kim, Jun-Gu Kim, Young-Hyok Ji, Yong-Chae Jung, Chung-Yuen
4 834–842 50 2.5 0.285 −0.593 Won, An islanding detection method for a grid-connected system based
5 854–852 100 1.0 1.507 0.112 on the Goertzel algorithm, IEEE Trans. Power Electron. 26 (4) (2011)
6 834–842 100 1.0 0.563 −0.374 1049–1055.
7 854–852 50 1.0 0.75 −0.375 [4] L.A.C. Lopes, Yongzheng Zhang, Islanding detection assessment of multi-
8 834–842 50 1.0 0.284 −0.558 inverter systems with active frequency drifting methods, IEEE Trans. Power
Deliver. 23 (1) (2008) 480–486.
[5] D. Reigosa, F. Briz, C. Blanco, P. Garcia, J. Manuel Guerrero, Active islanding
detection for multiple parallel-connected inverter-based distributed genera-
tors using high-frequency signal injection, IEEE Trans. Power Electron. 29 (3)
(2014) 1192–1199.
in a large DG penetration scenario, the DG may help the system
[6] M. Al Hosani, Z. Qu, H.H. Zeineldin, A Transient stiffness measure for island-
in the recovering process. However, a large perturbation on the ing detection of multi-DG systems, IEEE Trans. Power Deliver. 30 (2) (2015)
generation or transmission system may cause frequency variations 986–995.
[7] P. Gupta, R.S. Bhatia, D.K. Jain, Average absolute frequency deviation value
similar to those present in case of islanding, producing a unde-
based active islanding detection technique, IEEE Trans. Smart Grid 6 (1) (2015)
sirable tripping. Therefore, the Standard IEEE 1547 [1] allows the 26–35.
system operator to specify the frequency setting and time delay [8] W. Freitas, Wilsun Xu, C.M. Affonso, Zhenyu Huang, Comparative analysis
for underfrequency trips down to 57 Hz. In these cases, the settings between ROCOF and vector surge relays for distributed generation applications,
IEEE Trans. Power Deliver. 20 (2) (2005) 1315–1324.
of the proposed method should take this recommendation in to [9] IEEE Recommended Practice for Utility Interconnected Photovoltaic (PV) Sys-
account. tems: IEEE Standard 929-2000.
[10] M. Al Hosani, Zhihua Qu, H.H. Zeineldin, Development of dynamic estimators
for islanding detection of inverter-based DG, IEEE Trans. Power Deliver. 30 (1)
(2015) 428–436.
7. Conclusion [11] P. Mahat, Zhe Chen, B. Bak-Jensen, A hybrid islanding detection technique using
average rate of voltage change and real power shift, IEEE Trans. Power Deliver.
24 (2) (April 2009) 764–771.
This paper proposes a technique for islanding detection based [12] M. Khodaparastan, H. Vahedi, F. Khazaeli, H. Oraee, A novel hybrid islanding
on singular value decomposition and linear discrimination analy- detection method for inverter-based DGs using SFS and ROCOF, IEEE Trans.
sis. During islanding, the synchronous generator oscillates at very Power Deliver. 99 (2015) 1,1.
[13] K. El-Arroudi, G. Joos, I. Kamwa, D.T. McGillis, Intelligent-based approach to
slow frequency due to governors’ actions or the frequency growth islanding detection in distributed generation, IEEE Trans. Power Deliver. 22 (2)
exponentially when the governors are unable to correct it. How- (2007) 828–835.
ever, while connected to the main grid, the DG oscillates at a higher [14] N.W.A. Lidula, A.D. Rajapakse, A pattern recognition approach for detecting
power islands using transient signals – part I: design and implementation, IEEE
frequency. The method uses a pattern recognition methodology Trans. Power Deliver. 25 (4) (2010) 3070–3077.
that detects the frequency signature during islanding and sends [15] N.W.A. Lidula, A.D. Rajapakse, A pattern-recognition approach for detecting
a trip signal to the synchronous generator switcher. It uses smaller power islands using transient signals – part II: performance evaluation, IEEE
Trans. Power Deliver. 27 (3) (2012) 1071–1080.
windows than those used by methods that seek to estimate the
[16] S. Alshareef, S. Talwar, W.G. Morsi, A new approach based on wavelet design
frequency of oscillation and damping coefficient, providing faster and machine learning for islanding detection of distributed generation, IEEE
tripping. Several simulations were performed on different scenar- Trans. Smart Grid 5 (4) (2014) 1575–1583.
[17] H.H. Zeineldin, T.A. Galil, E.F.E. Saadany, M.M.A. Salam, Islanding detection of
ios with islanding and non-islanding events, and the proposed
grid connected distributed generators using TLS ESPRIT, Int. J. Elect. Power Syst.
methodology proved to be more sensitive than ROCOF islanding Res. 77 (2006) 155–162.
protection. The proposed algorithm is able to detect islanding with [18] M. Bakhshi, R. Noroozian, G.B. Gharehpetian, Anti-islanding scheme for syn-
power mismatches smaller than 1.6% and does not trip for the chronous DG units based on Tufts–Kumaresan signal estimation method, IEEE
Trans. Power Deliver. 28 (4) (2013) 2185–2193.
vast majority of non-islanding events. In short, the main advan- [19] Hadi Saadat, Power System Analysis, McGraw-Hill International, 1999.
tages brought by the proposed algorithm over other methods are [20] M. Dehghani, B. Shayanfard, A.R. Khayatian, PMU ranking based on singular
robustness for load switching and short circuits, small NDZ, and fast value decomposition of dynamic stability matrix, IEEE Trans. Power Syst. 28
(3) (2013) 2263–2270.
tripping.
 G. Marchesan et al. / Electric Power Systems Research 130 (2016) 124–131
[21] K. Ben-Kilani, M. Elleuch, Structural analysis of voltage stability in power sys-
tems integrating wind power, IEEE Trans. Power Syst. 28 (4) (2013) 3785–3794.
[22] J. Nathan Kutz, Data-Driven Modeling & Scientific Computation: Methods for
Complex Systems & Big Data, Oxford University Press, Inc., New York, NY, USA,
2013.
[23] R.A. Fisher, The use of multiple measurements in taxonomic problems, Ann.
Eugen. 7 (1936) 179–188.
[24] IEEE, Distribution System Analysis Subcommittee. IEEE 34 Node Test Feeder
[S.l.] IEEE (2010).
131
[25] K.E. Yeager, J.R. Willis, Modeling of emergency diesel generators in an 800
megawatt nuclear power plant, IEEE Trans. Energy Convers. 8 (3) (1993)
433–441.
[26] IEEE committee report, excitation system models for power system stability
studies, IEEE Trans. Power Appar. Syst. PAS-100, 2 (1981) 494, 509.
[27] L.N. Hannett, F.P. de Mlello, G.H. Tylinski, W.H. Becker, Validation of nuclear
plant auxiliary power supply by test, IEEE Trans. Power Appar. Syst. PAS-101
(9) (1982) 3068–3074.