Professional Documents
Culture Documents
Technologies
article has been accepted for publication in a future issue of this journal, but has not been fully edited.IET Smart Grid
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.
Abstract: The integration of the distributed power generation into a distribution system comes with several system
problems. One of the teething problems related to system protection is islanding detection. Various anti-islanding
techniques based on feature evaluation were proposed in the recent past. However, they overlook the need of justifying
the selection of a particular detection feature among all the possible measures. In this article, a wrapper feature selection
approach is proposed where a modified multi-objective differential evolution algorithm (MMODEA) is coupled with a
kernel-based extreme learning machine classifier (K-ELM). To select the optimum features, five standard objective
functions have been considered, such as dependability, security, accuracy, F-measure, and the number of features. About
1864 cases have been generated from the designed IEEE 13 bus system to extract the sensitive features. IEEE 1547
standards have been considered while designing and testing IEEE 13 bus system against islanding. The selected optimal
features detect the islanded condition decisively for both synchronous and inverter-based DGs. The features also validate
their performance under noisy environment accurately with lesser computational time.
Passive schemes are mostly preferred because of
1. Introduction their reduced design complexity, cost, and PQ issues. In this
Incorporation of distributed generators (DGs) using method of detection, the system parameter plays a
the microgrid concept at the distribution level has gained significant role. In literature, several parameters have been
momentum due to the exponential increase in demand and adopted in the passive anti-islanding schemes. They can be
environmental factors. Energy generation from distributed stated as voltage, frequency, current, power, harmonics, and
energy resources (DERs) has registered a growth of 23.7% their derivatives [6]-[10]. Apart from that, vector surge [11],
of a global generation of eco-friendly energy in the year phase shift [12], and change of harmonic impedance [13] are
2015 [1]. However, the deployment of DGs with the utility also suggested as detecting parameters. As the techniques
raises many technical issues concerned with the protection correspond to a single feature, they exhibit a large non-
and control strategies. One of the most critical protection detection zone (NDZ) with lower accuracy. Further, a
issues is Islanding. It is a scenario in which DGs are combination of two or more features is analyzed together to
disconnected from the main grid but continue to power the reduce the NDZ and attain higher accuracy. For example,
local loads connected to it. This may lead to various the rate of change of voltage and real power shift [14],
undesirable issues like reduction of power quality (PQ) Sandia frequency shift and rate of change of frequency [15].
restoration problems, and voltage and frequency instability. In literature, several detection techniques proposed
As per the IEEE 1547 standard, islanding condition should till date focus on minimizing the NDZ, which are deployed
be detected within 2 seconds from its inception [2]. So, all and tested for a particular type of DG [16,17]. However, a
the DGs should be accompanied by an anti-islanding relay, challenging task arises while designing a passive anti-
operating as an islanding detection device. islanding scheme protecting both the type of DGs
Islanding can be broadly classified into two types as simultaneously [4].
intentional and unintentional islanding. Unintentional Generally, passive techniques deal with three basic
islanding is a major concern and should be detected properly steps, such as feature extraction, feature selection (FS), and
in order to safeguard the utility personnel, DGs and end-user classification. In a practical scenario, the extraction of
equipment from out-of-phase reclosing [3]. Thus, the anti- numerous features intensifies the computational burden.
islanding schemes implemented for the detection are Thus, the selection of appropriate features with a minimum
generally categorized as remote, active, and passive schemes. number turns out to be a significant point of concern.
Remote scheme implements a communication path to detect Furthermore, the selection of optimum features reduces the
the islanding event. This process of detection is highly time taken for training and testing of a dataset for
expensive because of the equipment cost. Passive technique classification. In the process of feature selection task, the
practices the concept of measuring the system parameters redundant features are eliminated which helps to improve
and equating these values with the set of limiting values to the output performance of the classifier. Hence, it becomes
detect the islanding instant. Whereas, the active technique essential to design an algorithm to select an appropriate
implements the idea of continuous disturbance injection into feature set, to attain an efficient islanding detection and an
the system under study [4]. These added up small ensured protection to the DGs incorporated with the utility.
disturbances to the system, enhance the instability in the Some of the FS approaches implemented previously for the
system parameters for faster detection of islanding instance islanding detection problem are cited in [4],[18]-[20].
[5]. This article concentrates on the selection of sensitive
features to support faster as well as precise detection of the
islanding instances. Most of the papers present different 1. Offline selection of optimal features for islanding
detection techniques directly with selected features without detection, using modified multi-objective differential
mentioning the basis of selection. Therefore, the FS strategy evolution and kernel-based extreme learning machine.
has been focused on and detailed in the manuscript. FS 2. Computation of sensitive feature vectors through five
technique is broadly categorized into the wrapper and filter standard objectives such as dependability, security,
approaches. Wrapper FS methods use a classifier learning accuracy, F-measure, and the number of features.
process to optimize the objective function, conversely filter 3. Kernel-based extreme learning machine has been used to
FS methods do not. Thus, wrapper FS methods are more cross-validate the performance of selected feature vector
accurate and robust, compared to the filter FS methods in the noisy environment.
which are considered to be faster. A comparative study on 4. 45 features are extracted through simulation of modified
the wrapper and the filter FS method is presented in [21]. In IEEE 13 bus system integrated with synchronous and
this literature, the wrapper FS method has been considered inverter-based distributed generators for feature selection
to select the most sensitive features for islanding detection. algorithm.
Selection of optimum feature is a tricky job because 5. The method is numerically simple and operationally faster
the features hold a complicated relationship among with minimum NDZ.
themselves. Moreover, the wide-ranging search space The manuscript is organized as follows: Section 2
intensifies the difficulty in the selection and the increasing focuses on the design of the system under study.
number of features increases the search space exponentially Subsequently Sections 3 and 4 discuss the cases and features
[22]. Taking into account the above-specified problem, a considered in this study, respectively. Further, the idea of
proficient search technique is essential. Thus in 1997, Storn feature selection using a multi-objective differential
and Price proposed differential evolution (DE) which is a evolutionary algorithm and proposed methodology
population-based global optimization approach. Being an respectively are highlighted in Section 5 and 6. Section 7
efficient global search technique, it also possesses the analyses and discusses the effectiveness of the results
advantages like execution simplicity, reliability, and high obtained for clear discrimination of the islanding and non-
robustness. Therefore, DE has been extensively used by the islanding situations. At last the final summary concludes in
researchers in many diverse fields for feature selection. A Section 8.
detailed review of DE for FS task was presented by Ali et al.
[23]. 2. System under Study
In order to handle an optimization problem, several The distribution system designed to analyze the
aspects of a dataset are required to be taken care of. A single proposed scheme follows the standards of IEEE-13 bus test
objective function, solely could not tackle several attributes system [26]. The considered radial distribution feeder is
of a problem adequately. Therefore, a multi-objective integrated with three different types of DG. Fig.1.a.
optimization is preferred over a single objective represents the schematic layout of the system untaken for
optimization approach and is immensely used in many fields the study.
for FS. Taking into account the wrapper FS approach, The 115(Δ)/4.16(Y-grounded) kV substation
various classification methods have been implemented in the transformer serves the 13 bus test feeder. The feeder covers
last decade by several researchers. Extreme learning a distance of about 8200ft and is uniformly loaded with light
machine (ELM) is a latest classifier which follows a single loads. Following the layout of the IEEE 13 bus system, the
feedforward network [24]. As the ELM tends to be a recent shunt capacitors are placed at feeder ends. A 200kVAR
approach, it is yet to be explored in the field of FS. shunt capacitor is connected to the three phases at node 6
Moreover, the ELM possesses an attractive advantage of and another of 100kVAR is placed at the C phases of node
non-iterative linear solutions which tends to speed up the 13. The system specifications and the operating parameters
magnitude to 5 and 6 times in comparison with the multi- are presented in Table.1.
layer perceptron (MLP) and support vector machine (SVM)
respectively [25]. A kernel function based ELM (K-ELM) is Table 1 System specification & ratings
an enhancement to the ELM by improving the algorithm's Parameters Values
stability and by providing a robust unified solution. Q.F 1
Furthermore, the K-ELM does not require parameter tuning
Frequency 60 Hz
and possesses a minimum of computational complication.
Transformer T1 4.16kV/575V
Thus, the article proposes an MMODEA coupled K-ELM
T2 4.16kV/1300V
classifier as an anti-islanding scheme, addressing the
T3 4.16kV/575V
protection against both the synchronous based DGs and
T4 115kV/4.16kV
inverter based DGs. Further for feature selection, three
Loads L1 1.5MW, 0.9657MVar
objective functions such as dependability, security, and F-
L2 2.25MW, 1.0875MVar
measure have been considered along with the two basic
L3 2MW, 0.9657MVar
objective functions such as a maximum of accuracy and
Shunt C1 200kVar
minimum of feature number. A dataset of 1864 instances
Capacitor C2 100kVar
has been generated and applied to the proposed algorithm.
The outcomes of the selected feature set reveal the
effectiveness of the proposed algorithm by resulting an
accurate detection of the islanding events.
The major contributions of the present study are as
follows:
2
represent the number of objective functions opted for the The 𝛽 matrix is the only parameter to be computed in ELM.
optimization. A continuous minimization of n number of The least squares estimation is implemented to calculate the
objective functions leads to Pareto optimal solutions. The set matrix.
of non-dominated Pareto optimal solutions form a Pareto 𝛽 = 𝐻∗𝑇 (7)
front. This Pareto-front helps in selecting the best solutions here, 𝐻 ∗ = (𝐻𝑇 𝐻)−1 𝐻 𝑇 called as Moore-Penrose
according to the user's requirement. generalized pseudo-inverse of the matrix H. The ELM
follows simple procedures of three steps as:
5.3. Modified Multi-Objective Differential Evolution 1. Arbitrary selection of hidden parameter.
2. Evaluate the hidden layer matrix ′𝐻′.
A brief discussion of the proposed MMODEA is 3. Evaluate the output weight by means of (7).
provided in this section. MMODEA follows the same steps ELM performs significantly faster, as the entire
of classical DE except for certain changes, as we are dealing process of evaluation is completed within a single iteration.
with multiple objectives. On closely analyzing the selection
step of the DE algorithm, it can be realized that this step
Start
may pass over certain good solutions. Thus, the selection
step is modified according to an elitism based non-
dominating sorting from NSGA-II [32]. Specify control parameters
The modified selection step merges the target and
trial vectors, resulting in a population size of 2𝑃. The non-
domination principle is implied to assign a rank and sort the
solutions of the 2𝑃 population. Hence the non-domination Iteration
Stop
< NIter
rank and crowding distance helps in obtaining the top-best No
solutions of 𝑃 size [32].
The proposed MMODE algorithm executes for 'NIter' Yes
number of iterations to acquire the most excellent feature
Generate ‘P’ population
subset according to the objective functions. The flowchart of
the proposed algorithm for FS is presented in Fig.2.
No Store the best feature
5.4. ELM Generation
subset as per the
<G
objective function
ELM characterizes to be a feedforward artificial
neural network with a hidden layer. An introductory layout Yes
of the ELM is presented in Fig.3. ELM has acquired Generate mutant vector Mutation
enormous consideration amongst the researchers due to its
simplification, reliability, and faster learning speed in
Generate trial vector by performing
comparison with the other techniques, such as support crossover mechanism on target and
vector machine, back propagation, etc. [33]. ELM possesses mutant vector Crossover
a significant characteristics of selecting the input weight and
bias arbitrarily. Moreover, the algorithm also assists the
analytical determination of the output weights via Moore Merge trial and target vector
resulting ‘2P’ population
Penrose generalized pseudo-inverse [34]. ELM aims to
achieve the least training error along with a minimum norm
of output weights. Therefore, it presents an optimal solution Evaluate each instances w.r.t each
after excelling over the conventional issues such as local objective function for fitness
minima, learning rate, training period and stopping criterion. calculation
The output ′𝑦′ obtained from the ELM, having ′𝐿′ number
of hidden layers can be stated as in (5): Assign rank based on non-dominated
sorting as per the fitness value of Selection
𝑦 = ∑𝐿𝑖=1 𝛽𝑖 𝑔𝑖 (𝑥) = ∑𝐿𝑖=1 𝛽𝑖 𝐺(𝜔𝑖 , 𝑏𝑖 , 𝑥) = 𝐻𝛽 (5)
‘2P’ population
here, 𝑔𝑖 (𝑥) = 𝐺(𝜔𝑖 , 𝑏𝑖 , 𝑥) relates to the output for the ith
hidden layer. Where ′𝑥′ signifies the input samples, ′𝜔𝑖 ′ and
′𝑏𝑖 ′ presents the arbitrarily initiated weight and bias Sort the 2P population as per rank
respectively, ′𝐻′ and ′𝛽′ signifies the hidden layer matrix
and the output weights respectively [35]. For N individual New population created by selecting
sample, the matrices can be represented as: top-most ‘P’ population based on
𝐻𝛽 = 𝑇 (6) non-dominated sorting and crowding
where, distance
ℎ(𝑥1 ) 𝐺(𝜔1 , 𝑏1 , 𝑥1 ) ⋯ 𝐺(𝜔𝐿 , 𝑏𝐿 , 𝑥𝑁 ) Fig.2. Flowchart for proposed MMODE algorithm
𝐻=[ ⋮ ]=[ ⋮ ⋮ ⋮ ] ,
ℎ(𝑥𝑁 ) 𝐺(𝜔1 , 𝑏1 , 𝑥𝑁 ) ⋯ 𝐺(𝜔𝐿 , 𝑏𝐿 , 𝑥𝑁 ) 𝑁×𝐿
𝛽1 𝑇1
𝛽 = [ ⋮ ] , 𝑇 = [ ⋮ ] and T is the target matrix.
𝛽𝑁 𝐿×1 𝑇𝑁 𝑁×1
5
𝐼 𝐼
𝑦(𝑥) = ℎ(𝑥). 𝐻𝑇 ( + 𝐻𝐻 𝑇 )−1 𝑇 = [𝐾 𝑇 (𝑥, 𝑥1 ), 𝐾 𝑇 (𝑥, 𝑥2 ), . . . 𝐾 𝑇 (𝑥, 𝑥𝑁 )]. ( + Ω𝐸𝐿𝑀 )−1 𝑇 (9)
𝜆 𝜆
𝑆𝑢𝑚𝑚𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑟𝑖𝑔ℎ𝑡𝑙𝑦 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒𝑠 𝑇𝑃+𝑇𝑁
𝐴= = (13)
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒𝑠 𝑇𝑃+𝐹𝑃+𝑇𝑁+𝐹𝑁
7. Result Analysis
The proposed islanding detection approach has been
tested considering almost every scenario of very often
occurring disturbances in the distribution network. A dataset
comprising of 784 islanding instances and 1080 non-
islanding instances is obtained. The dataset of 1864
instances is examined by the proposed feature selection
algorithm (MMODEA-KELM) to obtain the best possible
feature subset on the basis of specified four objective
functions. The obtained feature subsets classifying the
islanding and non-islanding instances efficiently are
presented in Table.5 (Appendix). Amongst all the best-
performing feature subsets, optimal feature vector selection
becomes a major concern.
a
It can be evidently pointed out that the obtained
feature vector comprises two or more features to validate the
specified objective functions. Thus, in order to reduce the
computational burden during feature extraction, the feature
vector with the reduced number of features and best
performance should be selected. Furthermore, before
selecting the feature vector, it must ensure an efficient
performance in a noisy environment.
Thus, the feature vectors are cross-examined under
noisy environment through the K-ELM classifier. The
performance of the feature vectors corresponding to 20db
and 30db noise can be analyzed from Table.6 (Appendix).
Results indicate that the performance of the obtained
vectors in 30db noisy environment has a performance above b
90%. But a significant reduction in the performance can be
observed under 20db noisy environment.
Feature subsets at the 20db noise, are analyzed and
evaluated on the basis of performance of objective functions
and number of features in the subset. The most efficient
feature subset is concluded to be [F10 F11 F39], retaining
the dependability, security, accuracy, and F-measure as
94.77%, 97.12%, 96.13% and 95.37% respectively.
In order to validate the obtained feature subset, each
feature from the feature vector is further analyzed. The
visual representation of the selected individual feature is
shown in a comparative form on the basis of islanding and
non-islanding event. The feature F-10 i.e. the ratio between
negative sequence voltages to that of positive sequence c
voltage measured at DG1 is shown in Fig.4 for both
islanding and non-islanding events. The islanding event with
the active power mismatch of ±10% is shown in Fig.4.a.
Varying only the reactive power by ±5% is represented in
Fig.4.b. Further, the variation in Fig.4.c represents the
mismatch made on both active and reactive power by ±20%
and ±1% respectively. In islanding cases, small variations of
power mismatch are represented so as to clearly signify the
evidence for the negligible NDZ. A comparison of non-
islanding cases like capacitor switching, load switching and
faults are presented in Fig.4.d while all the cases experience
the disturbance at 0.5 seconds.
Likewise, various islanding cases and non-islanding
cases for feature F-11 and F-39 is presented in Fig.5 and d
Fig.6 respectively. The similar observation has been Fig.4. Islanding and Non-Islanding case study for feature F-
obtained at another target DG location (DG2 and DG3), 10;
however, due to space limitation, the results are not (a) Islanding Case: only P variations (b) Islanding Case:
presented in this article. only Q variations (c) Islanding Case: P and Q variations (d)
Non-Islanding Comparisons.
7
a a
b b
c c
d d
Fig.5. Islanding and Non-Islanding case study for feature F- Fig.6. Islanding and Non-Islanding case study for feature F-
11 39
(a) Islanding Case: only P variations (b) Islanding Case: (a) Islanding Case: only P variations (b) Islanding Case:
only Q variations (c) Islanding Case: P and Q variations (d) only Q variations (c) Islanding Case: P and Q variations (d)
Non-Islanding Comparisons Non-Islanding Comparisons
c
Fig.8. Features plot
(a) F-10 verses F-11 (b) F-11 verses F-39 (c) F-10 verses F-
39.
Table 7 Comparative analysis of the proposed technique with the existing techniques.
Ref Type of DG If feature Extracted Features Detection Performance-based on the
implemented selection used no. of considered speed objective function.
features for (milliseconds) A D S FM
islanding
detection
[16] MS No 11 11 - 91.6% 100% 83.3% 92.2%
[17] MI No 21 21 - 95.0% 100% 90.0% 95.7%
[18] MS Yes 11 3 - 100% 100% 100% 100%
[19] MS & MI Yes 21 4 180 100% - - -
[20] MS & MI Yes 27 11 30 97.5% 98.0% 97.1% -
[4] MS & MI Yes 16 4 125 100% - - -
Proposed MS & MI Yes 45 3 75 100% 100% 100% 100%
Approach
*MS: Multiple Synchronous DG; MI: Multiple Inverter-based DG; A: Accuracy; D: Dependability; S:Security; FM: F-Measure.
Table 8 Analysis of the proposed technique with the existing techniques based on similar test model and test cases.
Ref Features Detection Performance-based on the objective Performance-based on the objective
for speed function. function at the 20db noise
islanding (millisecond) A D S FM A D S FM
detection
[16] 11 80 89.69% 99.9% 82.22% 90.25% 82.83% 88.26% 78.88% 75.21%
[17] 21 55 96.19% 100% 93.42% 93.66% 93.61% 92.47% 94.44% 92.40%
[18] 3 125 100% 100% 100% 100% 94.08% 93.75% 95.37% 92.50%
[19] 4 180 98.64 100% 97.22% 98.43% 95.22% 94.38% 95.83% 92.82%
[20] 11 30 96.62% 98.08% 95.55% 96.12% 94.25% 91.83% 96.01% 92.48%
Proposed 3 75 100% 100% 100% 100% 96.13% 94.77% 97.12% 95.37%
Approach
[3] Llaria, A., Curea, O., Jiménez, J., et al.,: 'Survey on
8. Conclusion microgrids: unplanned islanding and related inverter control
With an intention to obtain precise performance of techniques', Renewable energy, 2011, 36, (8), pp 2052-2061.
the detection technique for both synchronous and inverter-
based DG, the offline selection of sensitive features for the [4] Chandak, S., Mishra, M., & Rout, P. K.,: 'Hybrid
system is focused on and discussed in details. In this study, islanding detection with optimum feature selection and
forty-five features are extracted from the system designed. minimum NDZ', International Transactions on Electrical
The dataset extracted is optimized according to the five Energy Systems, 2018, e2602.
objective functions using MMODEA, to find the set of the
best features so as to minimize the NDZ and operate [5] Li, C., Cao, C., Cao, Y., et al.,: 'A review of islanding
accurately in diversified operating conditions. To study the detection methods for microgrid', Renewable and
robustness of the selected feature for islanding detection, Sustainable Energy Reviews, 2014, 35, pp 211-220.
different islanding and non-islanding scenarios are
examined through K-ELM classifier in both ideal and noisy [6] Zeineldin, H. H., and James L. Kirtley.: 'A simple
environments. Evaluating the feature vectors on the basis of technique for islanding detection with negligible
minimizing the number of features and maximizing the nondetection zone', IEEE Transactions on Power Delivery,
performance parameters in noisy environments, the most 2009, 24, (2), pp 779-786.
efficient feature subset is concluded as [F-10 F-11 F-39]. It
is also verified that the time taken to detect islanding based [7] Zeineldin, H. H., and James L. Kirtley Jr.: 'Performance
on K-ELM using selected feature vector is approximately 75 of the OVP/UVP and OFP/UFP method with voltage and
milliseconds after the occurrence of an islanding event. This frequency dependent loads', IEEE Transactions on Power
represents the swiftness in detecting of an islanding scenario. Delivery, 2009, 24,(2), pp 772-778.
change of voltage phase angle', IET Generation, [22] Guyon, I. and Elisseeff, A.,:'An introduction to variable
Transmission & Distribution, 2015, 9, (15), pp 2337-2350. and feature selection', Journal of machine learning
research, 2003, 3, pp 1157-1182.
[10] Massoud, A.M., Ahmed, K.H., Finney, S.J., et al.,:
'Harmonic distortion-based island detection technique for [23] Ali, M., Siarry, P. and Pant, M.,: 'An efficient
inverter-based distributed generation', IET Renewable differential evolution based algorithm for solving multi-
Power Generation, 2009, 3, (4), pp 493-507. objective optimization problems', European journal of
operational research, 2012, 217, (2), pp 404-416.
[11] Freitas, W., Huang, Z. and Xu, W.,: 'A practical method
for assessing the effectiveness of vector surge relays for [24] Tang, J., Deng, C. and Huang, G.B.,: 'Extreme learning
distributed generation applications', IEEE Transactions on machine for multilayer perceptron', IEEE transactions on
Power Delivery, 2005, 20, (1), pp 57-63. neural networks and learning systems, 2016, 27, (4), pp
809-821.
[12] Hung, G.K., Chang, C.C. and Chen, C.L.,: 'Automatic
phase-shift method for islanding detection of grid-connected [25] Akusok, A., Björk, K.M., Miche, Y., et al.,: 'High-
photovoltaic inverters', IEEE Transactions on energy performance extreme learning machines: a complete toolbox
conversion, 2003, 18, (1), pp 169-173. for big data applications', IEEE Access, 2015, 3, pp 1011-
1025.
[13] Merino, J., Mendoza-Araya, P., Venkataramanan, G., et
al.,: 'Islanding detection in microgrids using harmonic [26] Kersting, W.H.,: 'Radial distribution test feeders', IEEE
signatures', IEEE Transactions on Power Delivery, 2015, 30, Transactions on Power Systems, 1991, 6, (3), pp 975-985.
(5), pp 2102-2109.
[27] Gagnon R.,: 'Detailed model of a doubly-fed induction
[14] Mahat, P., Chen, Z. and Bak-Jensen, B.,: 'A hybrid generator (DFIG) driven by a wind turbine', The MathWork,
islanding detection technique using average rate of voltage January, 2006.
change and real power shift', IEEE Transactions on Power
delivery, 2009, 24, (2), pp 764-771. [28] Turbine, H.,: 'Turbine Control Models for System
Dynamic Studies, Working Group on Prime Mover and
[15] Khodaparastan, M., Vahedi, H., Khazaeli, F., et al.,: 'A Energy Supply Models for System Dynamic Performance
novel hybrid islanding detection method for inverter-based Studies', Transactions on Power Systems, 1992. 7, (1), p 9.
DGs using SFS and ROCOF', IEEE Transactions on Power
Delivery, 2015. [29] Gagnon R.,: 'Wind Farm - Synchronous Generator and
Full Scale Converter (Type 4) Detailed Model', The
[16] El-Arroudi, K., Joos, G., Kamwa, I., et al.,: 'Intelligent- MathWork, January, 2006
based approach to islanding detection in distributed
generation', IEEE transactions on power delivery, [30] Storn, R. and Price, K.,: 'Differential evolution–a
2007, 22,(2), pp 828-835. simple and efficient heuristic for global optimization over
continuous spaces', Journal of global optimization, 1997, 11,
[17] Faqhruldin, O.N., El-Saadany, E.F. and Zeineldin, (4), pp 341-359.
H.H.,: 'Islanding detection for multi DG system using
inverter based DGs', IEEE Power & Energy Society General [31] Xue, B., Zhang, M. and Browne, W.N., :'Particle
Meeting, July 2013, pp 1-5. swarm optimization for feature selection in classification: a
multi-objective approach', IEEE transactions on cybernetics,
[18] Samantaray, S.R., El-Arroudi, K., Joos, G., et al.,: 'A 2013, 43, (6), pp 1656-1671.
fuzzy rule-based approach for islanding detection in
distributed generation', IEEE transactions on power delivery, [32] Deb, K., Pratap, A., Agarwal, S.,et al.,: 'A fast and
2010, 25, (3), pp 1427-1433. elitist multiobjective genetic algorithm: NSGA-II', IEEE
transactions on evolutionary computation, 2002, 6, (2), pp
[19] Faqhruldin, O.N., El-Saadany, E.F. and Zeineldin, 182-197.
H.H.,: 'A universal islanding detection technique for
distributed generation using pattern recognition', IEEE [33] Mishra, M., M. Sahani, and P. K. Rout. "An islanding
Transactions on Smart Grid, 2014, 5, (4), pp 1985-1992. detection algorithm for distributed generation based on
Hilbert–Huang transform and extreme learning
[20] Kar, S. and Samantaray, S.R.,: 'Data-mining-based machine." Sustainable Energy, Grids and Networks, 2017, 9,
intelligent anti-islanding protection relay for distributed pp 13-26.
generations', IET Generation, Transmission &
Distribution, 2014, 8, (4), pp 629-639. [34] Zhai, M.Y., Yu, R.H., Zhang, S.F., et al.,: 'Feature
selection based on extreme learning machine', IEEE Conf.
[21] Xue, B., Zhang, M., Browne, W.N., et al.,: 'A survey on on Machine Learning and Cybernetics, July 2012 , Vol. 1,
evolutionary computation approaches to feature pp 157-162.
selection', IEEE Transactions on Evolutionary Computation,
2016, 20, (4), pp 606-626. [35] Huang, G.B., Zhou, H., Ding, X. et al.,: ' Extreme
learning machine for regression and multiclass
11
classification', IEEE Transactions on Systems, Man, and [36] Wong, C.M., Vong, C.M., Wong, P.K. et al.,: 'Kernel-
Cybernetics, Part B (Cybernetics), 2012, 42,(2), pp 513-529. Based Multilayer Extreme Learning Machines for
Representation Learning', IEEE Transactions on Neural
Networks and Learning Systems, 2016.
Appendix
Table 4 Features Extracted
Notation Features (in p.u.) Symbolization of Feature
F-1 Root Mean Square of Voltage Vrms
F-2 D-Axis Voltage Vd
F-3 Q-Axis Voltage Vq
F-4 Root Mean Square of Current Irms
F-5 D-Axis Current Id
F-6 Q-Axis Current Iq
F-7 Negative Sequence Voltage Vneg
F-8 Positive Sequence Voltage Vpos
F-9 Zero Sequence Voltage Vzero
F-10 Negative Sequence Voltage / Positive Sequence Voltage Vneg/ Vpos
F-11 Zero Sequence Voltage / Positive Sequence Voltage Vzero/ Vpos
F-12 Negative Sequence Current Ineq
F-13 Positive Sequence Current Ipos
F-14 Zero Sequence Current Izero
F-15 The rate of change of Negative Sequence Voltage dVneg/ dt
F-16 The rate of change of Negative Sequence Current dIneg/ dt
F-17 Frequency F
F-18 The rate of change of frequency df/dt
F-19 Active Power P
F-20 The rate of change of Active Power dP/dt
F-21 Reactive Power Q
F-22 Negative Sequence Active Power Pneg
F-23 Negative Sequence Reactive Power Qneg
F-24 The rate of change of Negative Sequence Active Power dPneg/dt
F-25 The rate of change of Negative Sequence Reactive Power dQneg/dt
F-26 Zero Sequence Active Power Pzero
F-27 Zero Sequence Reactive Power Qzero
F-28 The rate of change of Zero Sequence Active Power dPzero/dt
F-29 The rate of change of Zero Sequence Reactive Power dQzero/dt
F-30 The rate of change of Frequency over Active Power df/dP
F-31 The rate of change of Voltage over Reactive Power dV/dQ
F-32 Inverse-Hyperbolic Secant of Negative Sequence Voltage 2 )⁄𝑉 2 )
ln ((1 + √1 − 𝑉𝑛𝑒𝑔 𝑛𝑒𝑔
12
Table 5 Obtained feature subsets through the proposed MMODE coupled K-ELM algorithm.
Feature Without Noise
Dependability Security Accuracy F-Measure
[F10 F11] 100% 100% 100% 100%
[F11 F30] 100% 98.88% 99.35% 99.24%
[F10 F11 F39] 100% 100% 100% 100%
[F11 F30 F39] 100% 99.53% 99.73% 99.68%
[F11 F30 F39 F43] 100% 100% 100% 100%
[F11 F30 F31 F39 F43] 100% 99.62% 99.78% 99.74%
[F10 F11 F18 F31 F39 F43] 100% 100% 100% 100%
Table 6 Performance of the feature subsets under 30db and 20db noise, to select the optimum feature vector.
Feature At 30 dB Noise
Dependability Security Accuracy F-Measure
[F10 F11] 98.72% 98.61% 98.65% 98.41%
[F11 F30] 96.04% 98.59% 97.51% 97.03%
[F10 F11 F39] 99.87% 99.62% 99.73% 99.68%
[F11 F30 F39] 95.28% 98.33% 97.04% 96.44%
[F11 F30 F39 F43] 95.91% 98.14% 97.21% 96.65%
[F11 F30 F31 F39 F43] 97.44% 98.33% 97.96% 97.57%
[F10 F11 F18 F31 F39 F43] 98.72% 98.88% 98.81% 98.59%
Feature At 20 dB Noise
Dependability Security Accuracy F-Measure
[F10 F11] 92.72% 96.11% 94.68% 93.62%
[F11 F30] 90.68% 95.46% 93.45% 92.09%
[F10 F11 F39] 94.77% 97.12% 96.13% 95.37%
[F11 F30 F39] 91.45% 95.46% 93.77% 92.51%
[F11 F30 F39 F43] 92.21% 95.37% 94.04% 92.87%
[F11 F30 F31 F39 F43] 90.68% 93.70% 92.43% 90.97%
[F10 F11 F18 F31 F39 F43] 94.13% 94.13% 95.60% 94.73%
13