Computation of Sensitivity-Based Islanding

Detection Parameters for Synchronous Generators

Tin Rabuzin
Fabian Hohn
Lars Nordström
Division of Electric Power and Energy Systems
KTH Royal Institute of Technology
Stockholm, Sweden
{rabuzin, fhohn, larsno}@kth.se

Abstract—Significant penetration levels of distributed energy can endanger line workers, cause unsynchronised reclosing or
resources increase the likelihood of continued operation of a impair the voltage and frequency control in the island [5]. To
power system island after an islanding event. It is important prevent the consequences of islanding and possibly ensure a
to employ adequate islanding detection methods to mitigate the
adverse effects of unintentional islanding and possibly transition safe transition to an islanded mode of operation, a quick and
to a safe islanded mode of operation. This paper focuses on the reliable islanding detection method (IDM) is needed.
computational aspects of a class of methods that utilizes a change IDMs can be broadly classified into local and remote tech-
in the sensitivity parameters as an indicator of islanding events. niques [6]. Local IDMs can be further classified into passive
It is shown that the inherent properties of the measurement and active methods. Passive IDMs base their algorithms on
signals cause numerical issues for the computation of the sensi-
tivity parameters. The paper also analyses three algorithms for a detectable change in the measurements of the system’s
estimation of the coefficients that overcome the numerical issues. electrical quantities after islanding, such as rate of change of
The performance of the algorithms has been demonstrated using frequency (ROCOF), rate of change of active power [7], rate
synthetically generated measurements. In addition, the data from of change of output power over frequency [8], vector shift,
field experiments has been used to further illustrate the practical and under/over voltage [9]. The inherent drawback of passive
viability of the algorithms.
methods is the existence of a non-detection zone (NDZ), i.e.
Index Terms—distributed generation, islanding detection, the passive methods fail to detect islanding if the imbalance
Kalman filter, least squares, sensitivity computation of the production and consumption in an island is sufficiently
small [10].
Active methods have been utilized to reduce the NDZ
I. I NTRODUCTION
by introducing non-destructive perturbations to the system.
The reduction of greenhouse gas emissions through sus- It is assumed that the perturbations disturb the balance in
tainable and efficient energy production has been one of the the island and cause a detectable change in the system’s
targets of the European 2020 strategy [1]. Many of the EU response, when a DER is islanded. While active methods
countries have adopted and implemented measures to achieve have a smaller NDZ compared to passive methods, they might
the 2020 targets. For example, Sweden reports that 54% of its cause undesirable system disturbances. Furthermore, interfer-
generated electric power in 2017 originated from renewable ence issues may arise if more than one DER employs an
energy sources [2]. The aforementioned environmental goals, active detection method in an island [11]. The active methods
along with the technological developments and economic which are applicable to the case of synchronous generators
opportunities, have facilitated the integration of distributed are reactive export error detection [12], impedance monitoring
energy resources (DERs) in distribution networks [3]. Despite [13], and the methods introducing positive feedback loop to
many benefits that DERs bring, several adverse effects on the the generator’s control systems [14].
utility grids caused by DERs have been observed, one of which Remote techniques rely on the ability of acquiring distant
is the possibility of islanding [4]. measurements or signals. Examples of such IDMs are com-
Power system islanding is an event that causes a part of a parison of ROCOF [15], transfer-trip schemes [16], power
network to be electrically separated from the utility grid and, line signaling [17], and slip frequency and angle method [18].
subsequently, the newly created electrical island continues to These methods typically have smaller NDZs when compared
be energized by local DERs. Unintentional islanding events to local methods, but the cost of communication infrastructure
can have a negative influence on the safe and stable operation often makes remote methods economically infeasible.
of a network. If unintentional islanding is not detected, it This paper focuses on a class of local IDMs, such as those
discussed in [8], [12], and [19] that use computed sensitivity
This work was supported by the SamspEL program of the Swedish Energy coefficients of the generator’s electrical quantities. Although
Agency. the use of these IDMs had been shown to be promising, the

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 Uref + consider a case in which one of the references is continuously
Perturbation − Ug perturbed using a small nondestructive periodic signal in order
+
to ensure persistent excitation of the system. The response
of the system is measured at the terminals of the generator
AVR Exciter
External Grid based on which the sensitivity parameters, introduced in the
+
PCC following subsection, can be computed. The change in the re-
+
Pref
Governor Turbine sponse of the system, reflected in the change of the sensitivity
Xe
− parameters, can serve as an indication of an islanding event.
Pg Load

Fig. 1. Unintentional islanding test configuration A. Sensitivity Parameters

The sensitivity parameters, which can be computed from
the system’s response, are:
practical implementation aspects of the computational algo-
• rate of change of frequency over active power
rithms have been overlooked. However, sensitivity coefficients
• rate of change of active power over voltage
have previously been used, mainly for voltage stability and
• rate of change of reactive power over voltage
control problems, and different methods have been developed
• rate of change of reactive power over frequency
to compute them from measurements. For example, in [20]
the sensitivities have been computed by low pass filtering and While all of the parameters above produce a change in their
differentiation of the measurements smoothed by a moving value upon islanding, this paper further analyses the rate of
average filter. Furthermore, [21] proposes a least squares change of reactive power over voltage as a representative
approach to computing sensitivity parameters. A generalized case due to its physical interpretability. Moreover, the reactive
least squares method has been used in [22] and [23] proposes power and voltage control loops produce faster responses to
to use compressed sensing to estimate sensitivities. the changes of the reference values due to the smaller time
One of the contributions of this paper is to identify the chal- constants involved in the control loops resulting in faster
lenges in computing the sensitivity coefficients and demon- islanding detection.
strate how these challenges can degrade the performance The reactive power produced by a generator connected to a
of the discussed class of IDMs. When compared to previ- network approximated by its Thévenin equivalent is:
ously proposed sensitivity estimation approaches, the intended Ug2 − Ug E cos(ϕ)
application of the computed coefficients and the associated Qg = , (1)
Xe
requirements are different in this paper. Therefore, the paper
analyses three estimation algorithms which can successfully where E and Xe are the Thévenin-equivalent voltage magni-
estimate the coefficients considering the following conditions. tude and reactance, respectively. Ug is the generator’s terminal
Firstly, we assume that only local measurements are avail- voltage magnitude and ϕ is the angle difference between the
able and that they are reported at higher rates than the generator terminal and the Thévenin-equivalent voltage. The
measurements from phasor measurement units (PMUs) or sensitivity of the produced reactive power to changes in the
smart meters as in [21]–[23]. Furthermore, this paper assumes generator’s terminal voltage is:
the existence of a weak although persistent system excitation 
∂Qg  2Ugo E cos(ϕ)
originating from perturbations of generators’ reference values. θq = = − (2a)
∂Ug  Ug =Ugo Xe Xe
The estimation algorithms also need to be capable of real-time Qg =Qgo
execution, to provide fast tracking of sensitivity coefficients to 2
Ugo Ugo 2
− Ugo E cos(ϕ)
avoid out-of-step reclosing and to operate successfully under = + (2b)
Ugo Xe Ugo Xe
the assumptions stated above.
Ugo Qgo
In addition, the paper evaluates the performance of the = + , (2c)
algorithms both by synthetically generated measurements and Xe Ugo
real measurements taken at a synchronous generator in a with Ugo and Qgo being the steady-state generator voltage and
distribution network during islanding experiments. reactive power, respectively. Synchronous generators are typi-
cally operated close to unity power factor and thus the second
II. P ROBLEM D EFINITION element in (2c) is considered to be negligible. Furthermore,
Consider an equivalent model of a synchronous generator Xe in the grid-connected mode of operation is much smaller
connected to a network at the point of common coupling when compared to the islanded mode of operation. Therefore,
(PCC) as depicted in Fig. 1. There are two main control the increase of the equivalent network reactance as seen from
systems at every synchronous generator; the one involving a generator after islanding causes a reduction in the sensi-
an automatic voltage regulator (AVR) and an exciter, and the tivity parameter, which indicates that an islanding event has
other comprised of a turbine and a governor. Using these occurred. Furthermore, by knowing the equivalent reactance
two systems, small perturbations to the references of the of the network, a threshold for the sensitivity parameter can
active power, reactive power or voltage can be introduced. We be determined

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 30 Before presenting and discussing the estimation algorithms,
20 the estimation problem is reformulated by manipulating (4)
resulting in:
10
∆y(n) = θ∆x(n) + v(n), (5)
0
where it is assumed that the measurement errors of both
-10 ∆x(n) and ∆y(n) can be grouped and represented by white
Finite differences
-20 Savitzky-Golay filter noise v(n) ∼ N (0, σv2 ). The implications of this assumptions
True value of q
are discussed in the following section. Using (5), sensitivity
-30
0 0.05 0.1 0.15 0.2 computation is restated as a minimization problem and as a
state estimation problem in the following subsections.
Fig. 2. Parameter θq computed by finite differences and SG filter
A. Recursive Least Squares (RLS)
The least squares estimate of the parameter θ can be
B. Computational Issues obtained by minimizing the following cost function:
The detection parameters, listed in the previous subsection, N
can be generically represented as the following parameter:
X 2
f (θ) = [∆y(i) − θ∆x(i)] , (6)
∂y(t) i=1
θ= , (3)
∂x(t) where N is the number of available measurements. Differ-
where the signals x(t) and y(t) can be any of the aforemen- entiation of the cost function with respect to the parameter θ
tioned measurements of the generator’s electrical quantities. and setting the derivative equal to zero yields the least squares
The straightforward way of estimating the parameter θ in estimator:
#−1 N
practical implementations is to perform approximations of
"N
X X
2
the sensitivity parameter by finite differences of the sampled θ̂ = ∆x(i) ∆x(i)∆y(i). (7)
signals, x(n) and y(n), as follows: i=1 i=1

y(n) − y(n − 1) ∆y(n) The measurements in this application are acquired sequen-
θ≈ = , (4) tially, which implies that the number of available measure-
x(n) − y(n − 1) ∆x(n)
ments, N , grows with time thus requiring, in theory, an
such that t = nTs where Ts is sampling interval and n ∈ Z. infeasible amount of memory. Furthermore, the parameter θ
Performing numerical differentiation using (4) accentuates can change over time, conversely to the inherent assumption
the omnipresent measurement noise, which can make the of the least squares estimator in (7).
estimation of the parameter θ unreliable. Furthermore, the To adapt the estimation for real-time execution and take
denominator of (4) may result in small quantities and if the into account the time-varying behavior of θ, the recursive least
algorithm is implemented using finite-precision arithmetic, this squares algorithm can be used. The RLS estimator is given by
may result in memory overflows. the following set of recursive equations [25]:
A snapshot of measurements was taken at a generator whose
voltage reference was periodically perturbed, with the true θ̂(n) = θ̂(n − 1) + R(n)−1 ∆x(n)
value of the sensitivity parameter being θq ≈ 3.8. Fig. 2
h i
× ∆y(n) − ∆x(n)θ̂(n − 1) (8a)
shows the parameter θq estimated by (4) and by dividing
measurements differentiated by the Savitzkey-Golay (SG) filter R(n) = λR(n − 1) + ∆x(n)2 (8b)
[24]. SG filtering is one of the possible ways of obtaining
smoothed derivatives of a signal. It can be observed that the The forgetting factor λ is used to improve the tracking
finite difference method results in an unstable estimation of capability of the least-squares estimator. If λ < 1 more weight
θq due to the presence of noise. SG filter somewhat improves is given to newer measurements, i.e. older measurements are
the estimation of θq . However, similarly as in the case of “forgotten”.
finite differences, the division with the small values of the
B. Blockwise Least Squares (BLS)
denominator, ∆x, at some time instances causes spikes in the
estimated parameter. This indicates that neither differentiation As pointed out by the authors in [26], RLS suffers from
using finite difference nor differentiation using signal smooth- slow tracking capabilities for time-varying system parameters.
ing provide satisfactory results for an islanding detection To overcome this drawback, one can compute the least squares
application. estimates over a sliding window of measurements similarly as
in (7) using:
III. E STIMATION A LGORITHMS " #−1
n n
The preceding section shows that there is a need for more
X X
θ̂(n) = ∆x(i)2 ∆x(i)∆y(i) (9)
suitable algorithms for the estimation of sensitivity parameters. i=n−W S+1 i=n−W S+1

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
where W S is the window size. Using BLS, improved tracking BLS
speed is expected since all of the old measurements outside Kalman filter
4 RLS
of the sliding window are not taken into account.
Typically, BLS would be computationally expensive in real-
time applications due to the operation of matrix inversions 2
involved in the estimation of vectors of parameters. In this
application, all of the involved variables are scalars, and 0
0 0.5 1 1.5 2 2.5
therefore the computational demand is much lower. However,
compared to RLS, this approach does require more memory
since W S measurements need to be stored at each time instant. Fig. 3. Arithmetic mean of the estimated θ̂q over 500 experiments

C. Kalman Filter (KF) IV. S IMULATION AND E XPERIMENTAL R ESULTS

The sensitivity parameter can also be computed by esti- This section presents test results used to evaluate the
mating the state of a discrete-time state-space model where performance of the algorithms described in Section III. We
parameter changes are modelled as a random walk: analyse convergence rates of the algorithms and the influence
of measurement noise on the estimation results. Furthermore,
θ(n + 1) = θ(n) + w(n), (10) the algorithms were tested using real measurements acquired
2 during islanding events and the blocking scheme was tested
with white noise w(n) ∼ N (0, σw ), i.e. we consider (10) to
using the measurements acquired during a synchronisation
be the state equation and (5) to be the output equation of the
procedure of an asynchronous machine.
state-space model. The parameter θ, now treated as a state
variable, can be estimated using the recursive KF equations: A. Convergence Analysis
h i To compare the convergence properties of the estimation
θ̂(n) = θ̂(n − 1) + K(n) ∆y(n) − ∆x(n)θ̂(n − 1) (11a)
algorithms, the following synthetic measurements were gener-
P (n − 1)∆x(n) ated:
K(n) = (11b)
σv2 + ∆x(n)2 P (n − 1)
2
Q(n) = Qgo + Qm sin (2πf T n) + eq (n) (13a)
P (n) = P (n − 1) − K(n)P (n − 1)∆x(n) + σw (11c)
U (n) = Ugo + Um sin (2πf T n) + eu (n) (13b)
RLS can be viewed as an instance of a KF with specific where T = 1.5ms is the sampling time interval and f
assumptions about the noise variances, i.e. KF is a more gen- is the frequency of the simulated sinusoidal perturbations
eral filter. Although the possibility of tuning noise covariances with the period of 130ms. The vector of parameters X =
makes KF more flexible, this comes at the cost of higher  T
Q0 Qm U0 Um changes as follows:
computational complexity compared to RLS. ( T
0.05 0.02 1 0.005 for 0s ≤ t < 1s
D. Blocking Scheme X=  T
0.01 0.005 1 0.01 for 1s ≤ t < 2.5s
The linearisation of the system in (2) is valid only in the
The change in the parameters effectively causes the true value
proximity of the equilibrium. Consequently, when the system
of the sensitivity parameter to decrease from θq = 4 to
experiences transitions, neither the model of the system in
θq = 0.5 at t = 1s. Zero mean white Gaussian noise processes
(5) nor the estimated values of θ will be valid. Therefore,
(eq and eu ) were added to both of the signals such that the
a blocking scheme is needed to prevent maloperation of the
signal-to-noise ratio (SNR) of the sinusoidal component of the
IDM. To block the IDM during transients, we propose to use
signal for t < 1s is 37dB and 49dB for reactive power and
the RMS value of the residuals:
voltage signals, respectively. Generated measurements were
v
u
u 1 Xn then differentiated and the algorithms were used to estimate
B(n) = t ε(i)2 (12) θq . The parameters of the algorithms were tuned so that the
W Sb mean and variance of the steady-state estimates for all of
i=n−W S b +1
the algorithms were equal. This experiment was repeated 500
where ε(n) = ∆y(n) − ∆x(n)θ̂(n − 1) are the residuals times and the arithmetic mean of the estimates of θ̂q was
obtained from the respective algorithms and W Sb is the calculated at each time instant.
window size over which B(n) is calculated. If W Sb is set The results of this procedure are shown in Fig. 3. It can
to be equal to the period of the injected perturbations and be observed that BLS has the highest convergence rate due
the threshold for B(n) is set above its steady-state value, it to the fact that the transient caused by a sudden change of
is essentially required that the filters need to be in steady θq is forgotten as soon as the sliding window moves past it.
state over a period of the perturbation signal in order for the This is not true for the recursive algorithms, RLS and KF,
estimates to be considered valid. since the transient is not “forgotten” as quickly as in the case

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 4 1.02

3 1.01 Generator Disconnection

2 1 Islanding
1
0.99
0 0 0.5 1 1.5 2 2.5
5 10 15 20 30 40 50 60 70 80
0.03
0.02
Fig. 4. Effect of noise in the independent variable on estimates of θq
0.01
0
4.5 -0.01
0 0.5 1 1.5 2 2.5

4
Fig. 6. Measurements during an islanding event with ∆P ≈ 3kW
3.5

C. Field Measurements
5 10 15 20 30 40 50 60 70 80
The performance of the estimation algorithms was also
Fig. 5. Effect of noise in the dependent variable on estimates of θq tested by using measurements taken at a distribution level
generator. The nominal power and voltage of the generator
are Sn = 20.5MVA and Un = 11kV, respectively. The true
of BLS. However, the error during the transition period of value of the sensitivity parameter as seen from the synchronous
BLS is higher than in the case of RLS and KF. Furthermore, generator is θq ≈ 3.8. The thresholds for the estimated
flexibility of KF when compared to RLS can be observed. parameter θ̂q (n) and the blocking signal B(n) were set to
Although both of the algorithms have the same average steady- 1 and 0.001, respectively. To ensure persistent excitation, the
state performance, the state covariance was tuned so that KF generator voltage reference was being perturbed by a periodic
converges faster to the steady-state value than RLS. signal of 4Hz.
The measurements of the reactive power and voltage used in
the estimation of θq were obtained from the excitation system
B. Influence of Measurement Noise of the generator, which performs the following preprocessing
steps. The AC stator voltage and current measurements are
It can be observed in Fig. 3 that the estimates of all of low-passed using a second order infinite impulse response
the algorithms are biased, i.e. on average they do not result (IIR) filter with a cutoff frequency at 600Hz. In the succeeding
in the parameter’s true value. The cause of the bias is the step, the RMS signals of the low-passed measurements are
presence of noise in the measurements of the independent calculated and the produced reactive power is estimated.
variable ∆x(n), which is not taken into account. This issue is Lastly, second order band-stop IIR filters are applied to the
known as errors-in-variables problem. However, the reliability RMS signals of the voltage and reactive power measurements
of IDMs considered in this paper is not reduced if the bias with attenuated frequencies at 50Hz and 100Hz. In addition,
present in the measurements is not significant, as shown below. the reactive power measurement is low-passed with a cutoff
To quantify the impact of noise on the estimates, the frequency at 100Hz. These band-rejected RMS voltage and re-
system in (13) was simulated in steady-state, i.e. with no active power measurements are the field measurements Ug (n)
parameter changes with various SNRs. Fig. 4 shows a box and Qg (n), respectively, and considered in the following
plot of estimated values of θ with respect to different SNRs analysis. Note, that the perturbation frequency of 4Hz has been
of the independent variable while keeping the SNR of the chosen in order to avoid impairments by the IIR filters.
dependant variable, ∆y(n), fixed. It can be observed that A part of the distribution network to which the generator
the increase of SNR of the independent variable reduces the is connected was islanded with different values of power
bias in the estimates. In particular, SNRs higher than 40dB imbalance in the newly created island, listed in the first column
would not significantly reduce the reliability of the steady- of Table I. Measurements of Ug and Qg shown in Fig. 6 were
state estimates. Similarly, Fig. 5 demonstrates the effect of taken during the islanding event with an active power imbal-
SNR of the dependent variable on the estimates of θ while ance in the island of ∆P ≈ 3kW. These measurements were
keeping SNR of the independent variable fixed. The change provided to the estimation algorithms presented in Section III
of SNR of the dependant variable does not affect the bias and the estimation results are shown in Fig. 7. It can be seen
of the estimates. However, the increase of SNR reduces the that the estimated values start decreasing after the islanding
variance of the estimates, which increases their reliability. event. At the same time, the blocking signal B(n) starts rising

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 BLS 1.06
3
KF
2 RLS 1.05
1
0 1.04
Detection Threshold
-1
0 1 2 3 4 5 6 7 8
0 0.5 1 1.5 2 2.5
-3
10
4
BLS 0.2
KF
RLS 0.15
2 Blocking Threshold
0.1

0 0.05
0 0.5 1 1.5 2 2.5 0 1 2 3 4 5 6 7 8

Fig. 7. Estimation of θq for an islanding event with ∆P ≈ 3kW Fig. 8. Measurements during synchronisation of an asynchronous generator

and blocks the operation of the islanding detection methods.

5
After the initial transient, both the estimates of θq and B(n)
fall below the threshold levels, which results in a successful 0 BLS
islanding detection. Detection Threshold KF
The same tests were done for different imbalances in the -5 RLS
created island and the detection times of the algorithms are 0 1 2 3 4 5 6 7 8
shown in Table I. The detection times of BLS and KF are 0.015
BLS
TABLE I 0.01 KF
I SLANDING DETECTION TIMES FOR FIELD DATA RLS

∆P [kW] tRLS [ms] tBLS [ms] tKF [ms] 0.005 Blocking Threshold

3 240 228 228 0

8 336 228 252 0 1 2 3 4 5 6 7 8
15 264 228 228
60 240 228 228 Fig. 9. Estimation of θq during synchronisation of an asynchronous generator
270 312 192 228

mostly influenced by the window size of the blocking scheme, and BLS perform similarly. However, the tracking capability
W Sb , and thus the times do not differ much over different is not the only measure by which the algorithms should
cases. This is, however, not the case for RLS, whose detection be judged since the algorithms have to be augmented by a
time is mostly influenced by its slower convergence rate. blocking scheme in order to ensure stability during transients.
To demonstrate the stability of the IDM, θq is estimated It was shown that the islanding detection times, do not differ
using field measurements taken during a synchronisation pro- significantly between KF and BLS since they are mostly
cedure of an asynchronous generator located at the same influenced by the window size used in the blocking scheme.
busbar as the synchronous machine. The measurements are However, the detection times of RLS were longer than both KF
shown in Fig. 8 and the estimation results in Fig. 9. The and BLS and influenced mostly by its poor tracking capability.
results show stable estimates of θq until t ≈ 3.2s when
Furthermore, although it is possible to address the errors-
the synchronisation is started. During the synchronisation,
in-variables problem with more complex algorithms such as
estimates are considered to be invalid since B(n) surpasses
total least squares, this is not necessary at sufficiently high
the threshold levels. When the steady-state is reached again,
SNR of the independent variable. It was shown that the bias
the estimates are above the threshold level resulting in the
is negligible above a certain level of SNR, which was also
stable operation of the IDM.
confirmed by accurate estimates of θq obtained from real
V. D ISCUSSION measurements in steady-state.
The preceding section demonstrates the performance of the The computational complexity and memory requirements
estimation algorithms. First, it was shown that the tracking of the algorithms have to be taken into account in real-
capabilities of different algorithms are not the same. Tuned time systems. RLS with the poorest tracking capability has
such that the steady-state estimates are similar for all of the the lowest computational and memory requirements. KF has
algorithms, the RLS was shown to be the slowest while KF higher computational and memory requirements than RLS but

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
lower than BLS. Similarly to other protection functions, the [6] P. Mahat, Z. Chen, and B. Bak-Jensen, “Review of islanding detection
design of an IDM has to balance between the dependency and methods for distributed generation,” in 2008 Third International Con-
ference on Electric Utility Deregulation and Restructuring and Power
stability of the method. The balance is reflected in the choice Technologies, April 2008, pp. 2743–2748.
of the method and its hardware requirements as well as in the [7] M. A. Redfern, O. Usta, and G. Fielding, “Protection against loss of
tuning of the algorithms’ parameters. utility grid supply for a dispersed storage and generation unit,” IEEE
Trans. Power Del., vol. 8, no. 3, pp. 948–954, July 1993.
One of the parameters that can significantly influence the [8] F. S. Pai and S. J. Huang, “A detection algorithm for islanding-prevention
performance of the method is the threshold for the sensitivity of dispersed consumer-owned storage and generating units,” IEEE Trans.
Energy Convers., vol. 16, no. 4, pp. 346–351, Dec 2001.
parameter. As shown in Section II, it should be set according [9] J. Vieira, D. Correa, W. Freitas, and W. Xu, “Performance curves of
to the expected equivalent impedance of an islanded system voltage relays for islanding detection of distributed generators,” IEEE
which can vary with the current operation conditions of the Trans. Power Syst., vol. 20, no. 3, pp. 1660–1662, Aug 2005.
[10] Z. Ye, A. Kolwalkar, Y. Zhang, P. Du, and R. Walling, “Evaluation
network. Therefore, the threshold essentially determines the of anti-islanding schemes based on nondetection zone concept,” IEEE
“size” of the island for which the IDM can operate. Trans. Power Electron., vol. 19, no. 5, pp. 1171–1176, Sep. 2004.
[11] A. Timbus, A. Oudalov, and C. N. M. Ho, “Islanding detection in smart
VI. C ONCLUSIONS grids,” in 2010 IEEE Energy Conversion Congress and Exposition, Sep.
2010, pp. 3631–3637.
[12] J. E. Kim and J. S. Hwang, “Islanding detection method of distributed
This paper demonstrated the numerical challenges in com- generation units connected to power distribution system,” in 2000
puting the sensitivity parameters for islanding detection using International Conference on Power System Technology (PowerCon),
local measurements. It also analysed the performance of vol. 2, Dec 2000, pp. 643–647 vol.2.
[13] P. O’Kane and B. Fox, “Loss of mains detection for embedded genera-
three algorithms for sensitivity estimation that circumvent the tion by system impedance monitoring,” in Sixth International Conference
numerical issues. The problem was formulated as an overde- on Developments in Power System Protection, March 1997, pp. 95–98.
termined system of linear equations for which the solution [14] P. Du, J. K. Nelson, and Z. Ye, “Active anti-islanding schemes for
synchronous-machine-based distributed generators,” IEE Proceedings -
was obtained using the two least squares algorithms, and as Generation, Transmission and Distribution, vol. 152, no. 5, pp. 597–606,
a state estimation problem for which Kalman filter was used. Sep. 2005.
Synthetic and field measurements were used to validate the [15] C. G. Bright, “COROCOF: comparison of rate of change of frequency
protection. a solution to the detection of loss of mains,” in 2001 Seventh
algorithms. The results show that the longest detection times International Conference on Developments in Power System Protection
are obtained with RLS algorithm while BLS and KF perform (IEE), April 2001, pp. 70–73.
similarly since they are constrained by the blocking function. [16] R. A. Walling, “Application of direct transfer trip for prevention of DG
islanding,” in 2011 IEEE Power and Energy Society General Meeting,
The problem formulation in this paper can serve as a basis July 2011, pp. 1–3.
for more complex IDMs where measurements from remote [17] W. Xu, G. Zhang, C. Li, W. Wang, G. Wang, and J. Kliber, “A power
locations can be used. Furthermore, the interference issues in line signaling based technique for anti-islanding protection of distributed
generators—part I: Scheme and analysis,” IEEE Trans. Power Del.,
the system with multiple DGs that inject perturbations into the vol. 22, no. 3, pp. 1758–1766, July 2007.
network can be analysed. Although the paper demonstrated the [18] J. Mulhausen, J. Schaefer, M. Mynam, A. Guzman, and M. Donolo,
case of a synchronous generator, the same algorithms could “Anti-islanding today, successful islanding in the future,” in 2010 63rd
Annual Conference for Protective Relay Engineers, March 2010, pp.
be applied in the cases of inverter-interfaced production units. 1–8.
However, further work is need to evaluate the performance of [19] S. Raza, H. Arof, H. Mokhlis, H. Mohamad, and H. A. Illias, “Passive
the algorithms in these settings. Additionally, future work also islanding detection technique for synchronous generators based on
performance ranking of different passive parameters,” IET Generation,
includes refinements of the estimation algorithms by taking Transmission Distribution, vol. 11, no. 17, pp. 4175–4183, 2017.
into account the discussed errors-in-variables problem and [20] R. Leelaruji, L. Vanfretti, and M. S. Almas, “Voltage stability monitoring
improvements of the blocking scheme by, e.g. performing using sensitivities computed from synchronized phasor measurement
data,” in 2012 IEEE Power and Energy Society General Meeting, July
whiteness tests. 2012, pp. 1–8.
[21] Y. C. Chen, A. D. Dominguez-Garcia, and P. W. Sauer, “Measurement-
ACKNOWLEDGEMENTS based estimation of linear sensitivity distribution factors and applica-
tions,” IEEE Trans. Power Syst., vol. 29, no. 3, pp. 1372–1382, May
The authors would like to thank Advensys Engineering Ltd. 2014.
[22] C. Mugnier, K. Christakou, J. Jaton, M. D. Vivo, M. Carpita, and
who provided the field measurements used in this work. M. Paolone, “Model-less/measurement-based computation of voltage
sensitivities in unbalanced electrical distribution networks,” in 2016
R EFERENCES Power Systems Computation Conference (PSCC), June 2016, pp. 1–7.
[23] P. Li, H. Su, C. Wang, Z. Liu, and J. Wu, “PMU-based estimation of
[1] European Commission (EC), “Europe 2020: a strategy for smart, sus- voltage-to-power sensitivity for distribution networks considering the
tainable and inclusive growth,” 2010, COM (2010) 2020. sparsity of Jacobian matrix,” IEEE Access, vol. 6, pp. 31 307–31 316,
[2] Swedish Energy Agency, “Energy in Sweden - an overview,” 2019. 2018.
[3] G. Pepermans, J. Driesen, D. Haeseldonckx, R. Belmans, and [24] R. Schafer, “What is a Savitzky-Golay filter? [lecture notes],” IEEE
W. D’haeseleer, “Distributed generation: definition, benefits and issues,” Signal Process. Mag., vol. 28, no. 4, pp. 111–117, July 2011.
Energy Policy, vol. 33, no. 6, pp. 787–798, 2005. [25] L. Ljung, System Identification: Theory for the User, 2nd ed., ser.
[4] N. Jenkins, J. Ekanayake, and G. Strbac, Distributed Generation, ser. Prentice Hall information and system sciences series. Prentice Hall
Energy Engineering Series. Institution of Engineering and Technology, PTR, 1999, pp. 363–364.
2010, pp. 12–18. [26] J. Jiang and Y. Zhang, “A revisit to block and recursive least squares for
[5] M. Bollen and F. Hassan, Integration of Distributed Generation in the parameter estimation,” Computers and Electrical Engineering, vol. 30,
Power System, ser. IEEE Press Series on Power Engineering. Wiley, no. 5, pp. 403–416, 2004.
2011, pp. 340–342.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020