A Universal Islanding Detection Technique for

Distributed Generation Using Pattern Recognition

ABSTRACT
This paper proposes an Anti-islanding protection method, proposed in the
literature, is distributed generation (DG) type dependent or in other words
works efficiently for a specific DG type (synchronous or inverter based). Here
we investigate the possibility of developing an efficient universal islanding
detection method that can be applied to both inverter and synchronous-based
DG. The proposed method relies on extracting a group of features, from
measured data simulated for both types of DGs, from which the best features are
selected for islanding detection. A random forest (RF) classification technique
is used to detect islanding and non-islanding situations with an objective of
minimizing the non-detection zone as well as avoiding nuisance DG tripping
during non-islanding conditions. Islanding and non-islanding cases were
generated for the IEEE 16-bus system and used to train and test the proposed
technique. The results show that the proposed methodology has zero non-
detection zone, high accuracy, and fast response when applied to both types of
DGs independently of the size of the island. The simulation work will be carried
out using MATLAB/Simulink.

INTRODUCTION
Islanding is a condition in which part of the distribution network is
disconnected from the reminder of the grid, and yet is powered by one or more
distributed generation (DG) connected to it. Islanding detection methods can be
divided into three main categories: active, passive, and communication- based
technique. Active methods rely on introducing a small noise signal into the
system. Under normal conditions, this signal will not cause large deviation in
system parameters. However, during islanding, this signal will be amplified and
thus facilitating islanding detection.

Over the past years, various islanding detection methods have been proposed
with an overall aim of minimizing the non-detection zone (NDZ). Islanding
detection methods, proposed in previous literature, are DG dependent or in
other words have been tested and developed for a specific type of DG interface.
For example, recent islanding detection techniques such as Sandia frequency
shift(SFS), positive-feedback-based active islanding detection, frequency drift
anti-islanding, high-frequency signal injection, and Bayesian passive islanding
detection have proven to have small or even negligible NDZ for inverter-based
DG. These methods would either fail to operate properly when applied to
synchronous-based DG or cannot be integrated in the same way. On the
contrary, islanding detection methods such as decision tree (DT)-based
islanding detection technique, Fuzzy rule-based approach for islanding
detection, pattern recognition approach for detecting power islands using
transient signals and synchronous distributed generation islanding protection
using intelligent relays have been designed specifically for synchronous-based
DG. Similarly, since inverter-based and synchronous-based DG react differently
to an islanding condition, such methods might not operate efficiently for
inverter-based DG. Over/under frequency and over/under voltage protection)
(OFP/UFP and OVP/UVP) and rate of change of frequency (ROCOF) are
passive islanding detection methods that can be considered universal in the
sense that they can be applied to both types of DGs. Such methods suffer from
large NDZ especially for inverter-based DG. Communication-based islanding
detection is also considered a universal approach but as mentioned earlier is a
costly solution for islanding detection. Recently, artificial intelligence (AI)
techniques have been proposed for islanding detection. The main advantage of
AI techniques is the capability of determining the most suitable combination of
features/parameters as well as thresholds that can reduce significantly the non-
detection zone of islanding detection methods. An intelligent islanding detection
technique that uses decision trees in order to extract threshold values for eleven
system parameters was proposed. However, the method was not capable of
capturing all possible islanding events; therefore, a fuzzy rule-based approach
was developed which yielded very accurate results in terms of islanding
detection. Another passive islanding detection technique was proposed where a
Bayesian classifier was used to detect islanding based on 64 parameters
calculated by ESPRIT. The proposed technique used voltage and frequency
waveforms in order to generate the features and was fed to a naïve Bayesian
classifier to distinguish islanding scenarios. A pattern recognition approach was
implemented for islanding detection. Voltage and current transient signals were
discretized by discrete wavelet transformation in order to extract features.

A decision tree was modeled in order to discriminate between islanding

and non-islanding events based on the energy content in the wavelet
coefficients. For all the above techniques AI techniques showed high accurate
results in detecting islanding if proper features were extracted and a proper
classifier was used. Nevertheless, all the techniques were implemented and
tested based on one DG type (synchronous or inverter). This paper proposes an
efficient universal islanding detection method that can be applied to both
inverter and synchronous based DG. A feature selection algorithm is
implemented to determine the most salient features/parameters for detecting
islanding that: 1) would result in minimal NDZ, 2) is robust to other non-
islanding events. A random forest classification algorithm is trained using data
generated from the IEEE 16-bus system considering both types of DG interface.
The accuracy of the proposed technique is determined using cross validation
and is compared against other classification approaches.
PROBLEM STATEMENT
This paper proposes an efficient universal islanding detection method that
can be applied to both inverter and synchronous based DG. A feature selection
algorithm is implemented to determine the most salient features/parameters for
detecting islanding that: 1) would result in minimal NDZ, 2) is robust to other
non-islanding events. A random forest classification algorithm is trained using
data generated from the IEEE 16-bus system considering both types of DG
interface.

LITERATURE SURVEY
G.-K. Hung, C.-C. Chang, and C.-L. Chen, proposed “Automatic phase-
shift method for islanding detection of grid-connected photovoltaic inverters,”.
The traditional frequency-shift methods for islanding detection of grid-
connected PV inverters—the active frequency drift method and the slip-mode
frequency-shift method—become ineffective under certain paralleled RLC
loads. The automatic phase-shift method is proposed in this paper to alleviate
this problem. The method is based on the phase shift of the sinu-soidal inverter
output current. When the utility malfunctions, the phase-shift algorithm keeps
the frequency of the inverter terminal voltage deviating until the protection
circuit is triggered. Simulation and experiments are performed for verification

H. H. Zeineldin and S. Conti, proposed “Sandia frequency shift parameter

selection for multi-inverter systems to eliminate non-detection zone”. Among
frequency drift islanding detection methods, Sandia frequency shift (SFS) is
considered as one of the most effective methods in detecting islanding
conditions for grid connected photovoltaic (PV) inverters. The performance of
the SFS method during an islanding condition and its non-detection zone (NDZ)
depends to a great extent on its parameters. Furthermore, the capability of the
SFS method to detect an islanding condition deteriorates with multiple PV
inverters. A mathematical formula is derived to aid protection engineers in
determining the optimal setting of the SFS islanding detection parameters with
multiple inverter-based distributed generation (DG), such as PV systems, to
eliminate the NDZ. The derived formula is applied to multiple DG systems
equipped with the over frequency/under frequency protection, active frequency
drift and SFS islanding detection methods and is verified through NDZ analysis
and simulation results on PSCAD/EMTDC. The derived formula provides an
effective guideline for designing frequency drift methods in multi-inverter-
based DG systems.

H. H. Zeineldin and M. M. A. Salama, proposed “Impact of load

frequency dependence on the NDZ and performance of the SFS islanding
detection method”. Sandia frequency shift (SFS) falls under the active islanding
detection methods that rely on frequency drift to detect an islanding condition
for inverter-based distributed generation. Active islanding detection methods
are commonly tested on constant RLC loads where the load’s active power is
directly proportional to the square of voltage and is independent on the system
frequency. Since the SFS method relies primarily on frequency to detect
islanding, the load’s active power frequency dependence could have an impact
on its performance and the nondetection zone (NDZ). In this paper, the impact
of the load’s active power frequency dependence on the performance of the SFS
method, during an islanding condition, is analyzed. A NDZ model that takes
into account the load’s frequency dependence parameter is derived
mathematically and validated through digital simulation. The results show that
the load’s frequency dependence has a significant impact on the NDZ of the
SFS method and thus is an important factor to consider when designing and
testing this method.

L. Lopes and H. Sun, proposed “Performance assessment of active

frequency drifting islanding detectionmethods,”. Islanding detection is a
mandatory feature for grid-connected inverters. The effectiveness of passive
islanding detection methods (IDMs) is usually demonstrated by means of
nondetection zones (NDZs) represented in a power mismatch space (ΔP versus
ΔQ). Active frequency drifting IDMs have been shown to provide improved
performance but their theoretical NDZ cannot be described in the ΔP versus ΔQ
space for a general RLC load. This paper shows that a load parameter space
based on the values of the quality factor and resonant frequency of the local
load (Qf versus f0) can be used in these cases. It employs a single curve to
represent the NDZ of frequency drifting IDMs for any RLC loads. Equations
that represent NDZs of three common active IDMs in the Qf versus f0 load
parameter space are derived and it is shown that the slip mode frequency shift
and the Sandia frequency shift IDMs can be designed to guarantee islanding
detection for equivalent RLC loads with a quality factor smaller than a design
value. The accuracy of the NDZs is verified with simulation and experimental
results.

P.Du, Z.Ye,E.Aponte, J. Nelson, andL. Fan,, proposed “Positive-

feedback-based active anti-islanding schemes for inverter-based distributed
generators: Basic principle, design guideline, and performance analysis,”. The
positive-feedback-based anti-islanding schemes invented by the authors are
highly effective in preventing islanding without causing any degradation in
power quality. This paper presents the basic principles of these schemes and
their design guidelines. Moreover, their performance is investigated for
inverter-based distributed generators (DGs). The parametric study reveals the
factors significantly influencing the performance of these schemes. The
simulation results demonstrate the effectiveness of these schemes.

A. Yafaoui, B. Wu, and S. Kouro, proposed “Improved active frequency

drift anti-islanding detection method for grid connected photovoltaic systems,”.
As more distributed generators join the utility grid, the concern of possible
undetected islanding operation increases. This concern is due to the safety
hazards this phenomenon imposes on the personnel and equipment. Passive
anti-islanding detection methods monitor grid parameters to detect islanding,
whereas active methods inject a perturbation into the current waveform to drive
these parameters out of limit when islanding occurs. The performance of active
methods, such as conventional active frequency drift (AFD), is limited by the
amount of total harmonic distortion (THD) they inject into the grid, which
defines its nondetection zone. In this paper, an improved AFD anti-islanding
method is presented based on a different current distortion injection waveform.
The proposed method generates 30% less THD compared to classic AFD,
resulting in faster island detection and improved nondetection zone. The
performance of the proposed method is derived analytically, simulated using
Matlab and verified experimentally using a prototype setup. A single-phase
grid-tied photovoltaic distributed generation system is used for the simulation
and experimental setup, and considered as potential application.

H. Karimi, A. Yazdani, and R. Iravani,, proposed “Negative-sequence

current injection for fast islanding detection of a distributed resource unit”. This
paper presents an active islanding detection method for a distributed resource
(DR) unit which is coupled to a utility grid through a three-phase voltage-
sourced converter (VSC). The method is based on injecting a negative-sequence
current through the VSC controller and detecting and quantifying the
corresponding negative-sequence voltage at the point of common coupling of
the VSC by means of a unified three-phase signal processor (UTSP). UTSP is
an enhanced phase-locked loop system which provides high degree of immunity
to noise, and thus enable islanding detection based on injecting a small (3%)
negative-sequence current. The negative-sequence current is injected by a
negative-sequence controller which is adopted as the complementary of the
conventional VSC current controller. Based on simulation studies in the
PSCAD/EMTDC environment, performance of the islanding detection method
under UL1741 anti-islanding test is evaluated, and its sensitivity to noise, grid
short-circuit ratio, grid voltage imbalance, and deviations in the UL1741 test
parameters are presented. The studies show that based on negative-sequence
current injection of about 2% to 3%, islanding can be detected within 60 ms
even for the worst case scenario.

M. Ropp, M. Begovic, A. Rohatgi, G. A. Kern, R. H. Bonn, Sr, proposed

“Determining the relative effectiveness of islanding detection methods using
phase criteria and nondetection zones,”. Islanding of a utility-interactive
photovoltaic (UIPV) system occurs if the UIPV system continues to power a
section of the utility system after that section has been disconnected from the
utility source. Since islanding creates hazards for personnel and equipment,
UIPV systems are required to detect and prevent it. It is desirable to have a
simplified method of determining which islanding detection methods (IDMs)
are most effective. In this paper, a previously described method for finding the
nondetection zones (NDZs) of IDMs is experimentally verified. This method is
used to determine the NDZs of several common IDMs. These results indicate
that, of the IDMs discussed in this paper: (1) Sandia Frequency Shift (SFS) is
most effective; and (2) the worst-case loads are low-power loads that are near
resonance at the line frequency and have a large capacitance and small
inductance (a high value of the quality factor Q).

METHODOLOGY
Here, in this paper the base system(IEEE 16-bus distribution system)
which is given below shows the system with a DG connected at the point of
common coupling (PCC). There is a local RLC load of 100-kW connected at the
PCC. The DG and the load are connected to the remaining distribution system
by a 100-kVA 24.9-kV/480-V transformer.
The control unit for Inverter based DG system works as follows: the voltage is
measured at the PCC point and fed to the phase-locked loop (PLL) to determine

the frequency and the phase angle . In addition, the current at the PCC is
measured and fed into the Park transformation to get the dq component of the
current. After extracting the dq components of the current, the i q and id are
compared to iqref and idref and the difference between them were used as an input
to the proportional-integral controller. Finally, in order to control the inverter-
based DG switches, the frequency with the output of the proportional-integral
controller were used as input to the pulse width modulation (PWM).

The idref is generated by a fuzzy controller which controls the dc voltage and
thereby real power flow.

A wide range of simulated cases were performed in order to provide the

classifier with solid information in order to dif ferentiate between islanding and
non-islanding conditions. In order to simulate islanding conditions, different
mismatches in generated power by the local DG and consumed power by the
local load inside the island were simulated. The range of active power mismatch
was from 0% up to ±30% and reactive power mismatch were between 0% up to
±5%. This range for mismatch has been chosen since it represents the hardest
cases to detect. On contrary, capacitor bank switching, load switching, and
motor switching were performed in order to produce the non-islanding
scenarios. The proposed islanding detection technique can works with single
(inverter/synchronous) DG or multiple DG units connected together at the same
feeder. To clarify this, the multiple cases for generating the island and they are
as follows:

1. Disconnecting from the PCC to generate a small island that includes load and
a DG only.

2. Disconnecting from middle of the system to generate a larger island that

includes local loads and multiple DG.

3. Disconnecting from the main substation to create a large island that includes
all the loads in the system with multiple DG.

In each of the proposed island the following scenarios were followed:

1) Scenario 1: All inverter DG system.

2) Scenario 2: All synchronous DG system

3) Scenario 3: Multiple DG types connected in the system.

In this paper, a combination of various features/system parameters have been

chosen on inverter-based and synchronous based-DG. The proposed universal
islanding detection method has three main phases which include feature
extraction, feature selection, and classification. A feature selection algorithm is
used to determine the most effective combination of features that can be used
for detecting islanding accurately and that can be applied to both inverter-based
and synchronous- based DG. Sequential feature selection methods such as
forward feature selection (FFS) and backward feature selection (BFS) are used
to select the best features to be used for islanding detection. Last, a RF
algorithm is used to distinguish islanding from non-islanding events.

It includes two main stages: 1) the extraction and selection of the feature. 2)
Training the classifiers and detecting islanding.

The methodology followed in this paper and described as follows:

1) 200 cases were simulated using MATLAB/SIMULINK in order to cover all
possible islanding and non-islanding conditions.

2) Using voltage and current waveforms at the PCC of the targeted island, 21
features were extracted to distinguish between islanding and non-islanding.

3) Using FFS and BFS feature selection, the top four features were selected.

4) The process is repeated for all the events until data matrix is built.

5) Four-fold cross validation technique was used to train and test the classifier.

6) Using training subset, RF classifier was trained to distinguish between

islanding a non-islanding.

7) Using the testing subset, the accuracy and required detection time of the RF
classifier was measured.

8) The process was repeated four times to calculate the average accuracy for
more robust islanding detection technique.

REFERENCE
IEEE Application Guide for IEEE Std 1547, IEEE Standard for Inter-
connecting Distributed Resources With Power Systems, IEEE Std. 1547.2–
2008, 2009.

G.-K. Hung, C.-C. Chang, and C.-L. Chen, “Automatic phase-shift method for
islanding detection of grid-connected photovoltaic inverters,” IEEE Trans.
Energy Conversion, vol. 18, no. 1, pp. 169–173, Mar. 2003.

H. H. Zeineldin and S. Kennedy, “Sandia frequency-shift parameter selection to

eliminate nondetection zones,” IEEE Trans. Power Del., vol. 24, no. 1, pp. 486–
487, Jan. 2009.

H. H. Zeineldin and S. Conti, “Sandia frequency shift parameter selection for

multi-inverter systems to eliminate non-detection zone,” Renewable Power
Generat., IET, vol. 5, no. 2, pp. 175–183, Mar. 2011.
H. H. Zeineldin and M. M. A. Salama, “Impact of load frequency dependence
on the NDZ and performance of the SFS islanding detection method,” IEEE
Trans. Ind. Electron., vol. 58, no. 1, pp. 139–146, Jan. 2011.

L. Lopes and H. Sun, “Performance assessment of active frequency drifting

islanding detectionmethods,” IEEE Trans. Energy Conversion, vol. 21, no. 1,
pp. 171–180, Mar. 2006.

P.Du, Z.Ye,E.Aponte, J. Nelson, andL. Fan, “Positive-feedback-based active

anti-islanding schemes for inverter-based distributed generators: Basic
principle, design guideline, and performance analysis,” IEEE Trans. Power
Electron., vol. 25, no. 12, pp. 2941–2948, Dec. 2010.

A. Yafaoui, B. Wu, and S. Kouro, “Improved active frequency drift anti-

islanding detection method for grid connected photovoltaic systems,” IEEE
Trans. Power Electron., vol. 27, no. 5, pp. 2367–2375, May. 2012.

H. Karimi, A. Yazdani, and R. Iravani, “Negative-sequence current injection for

fast islanding detection of a distributed resource unit,” IEEE Trans. Power
Electron., vol. 23, no. 1, pp. 298–307, Jan. 2008.

L. Lopes and H. Sun, “Performance assessment of active frequency drifting

islanding detectionmethods,” IEEE Trans. Energy Conversion, vol. 21, no. 1,
pp. 171–180, Mar. 2006