Sie sind auf Seite 1von 12

IET Generation, Transmission & Distribution

Research Article

High-impedance fault detection and ISSN 1751-8687


Received on 22nd October 2017
Revised 25th April 2018
classification in power system distribution Accepted on 23rd May 2018
E-First on 22nd June 2018
networks using morphological fault detector doi: 10.1049/iet-gtd.2017.1633
www.ietdl.org

algorithm
Moses Kavi1, Yateendra Mishra1 , Mahinda D. Vilathgamuwa1
1Faculty of Science and Engineering, School of Electrical Engineering and Computer Science, Queensland University of Technology, Brisbane,
Australia
E-mail: Yateendra.mishra@qut.edu.au

Abstract: This study presents a fast, secure and reliable algorithm for the detection and classification of power system
disturbances including high-impedance faults (HIFs). The proposed algorithm utilises mathematical morphology (MM)
techniques, where the non-linear MM characteristics are exploited by strategic cascading of appropriate filtering functions to
form a multistage morphological fault detector (MFD) for the extraction of features necessary for the characterisation of HIFs.
The target features of the HIF are the randomness and arc extinction and re-ignition/unsymmetrical characteristics. The
reliability and robustness in the extraction of the desired HIF features are enhanced by a weighted convex structuring element
designed based on the attributes of power system signals. The performance of the proposed algorithm is tested under different
types of disturbances including cases of HIFs on different contact surfaces. Moreover, the effectiveness of the algorithm is
tested under noise condition to demonstrate its performance of the proposed MFD. All tests are simulated using IEEE13 bus
test system.

1 Introduction characteristic. The most influential of all is (i) contact surface


material, which gives the non-linear voltage-current (V–I)
Any power system fault conditions or events resulting in low fault characteristic and, (ii) surface humidity, which dictates the level of
current magnitude not sufficient to overcome the threshold setting fault current magnitude [5]. Generally, HIFs most often result in
of the conventional overcurrent (OC) protection scheme are arcing creating an arc channel for the fault current to flow. The arc
classified as high-impedance faults (HIFs). Unlike bolted short- channel can be represented by an arc resistance, Rarc, which is
circuit faults, HIFs normally have high a HIF path which limits the
highly random and inversely proportional to the surface humidity
fault current magnitude. Occurrences of HIFs are frequently
[6]. The arc resistance restricts the magnitude of the fault current
experienced at medium-voltage to low-voltage networks. HIFs can
through the arc channel. The fault current is further restricted by
be grouped as (i) fallen energised conductor either broken or intact
the high impedance contact surface, whose resistance can be
making unwanted electrical contact with high impedance surfaces
represented by Rx. This gives an effective fault resistance, Rf = Rarc 
such as bitumen, concrete, sand, soil and so on, and (ii) energised
overhead conductor making unwanted contact with tree trunk and + Rx seen at the fault location. Different contact surfaces have
branches [1]. Such contact surfaces provide a HIF path to ground different Rx. which defines the non-linear (V–I) characteristic for
resulting in a fault current magnitude of typically between 10 and each surface material. Rarc and Rx are random, and usually very
50 Arms with erratic waveform. While the low HIF current high values, which defines the randomness and chaotic nature of
magnitude may not pose immediate significant threat to power the fault resistance.
system infrastructure; nonetheless, a fault condition exist. For arcing to occur, a separation (gap) between the energised
Moreover, if a HIF is allowed to persist, it increases risk of fire and conductor and high impedance contact surface must exist. The gap
safety hazard for people. Moreover, HIFs introduce harmonics, and can be represented by an air-gap breakdown voltage, Vbr, which
can degrade the quality of power supply in the long run [2], must be overcome by the conductor voltage for arcing to take
therefore must be detected and removed within a reasonable time. place. Considering that the conductor voltage is cyclic at 50 Hz,
The OC protection scheme in distribution network depends on there are two zero-crossings every half cycle, during which the arc
accurate measurement of fault current magnitude, and then makes a is extinguished at, then re-ignites at about the positive and negative
comparison with a predefined threshold to determine the trip time. peaks. Moreover, Vbr retains random value each half cycle resulting
If the fault current magnitude falls below the threshold, the event is in an asymmetrical fault current waveform with unequal positive
ignored. Considering that the threshold in OC protection is set at and negative half values [7, 8]. Furthermore, the fault current
around 2-to-3 times the prefault load current, it is impossible for gradually escalates to its maximum value due to slow initial
any conventional OC protection device, whose effectives are contact separation and maintains this value for several cycles
contingent on secure and reliable detection of increased fault giving the build-up and shoulder characteristics of the HIF current
current magnitude above a predefined threshold to detect HIFs to [9]. Moreover, the HIF current contains high frequency harmonics
suffice [3]. Thus, a scope for HIF detection and classification of between 2 and 10 kHz [10].
technique based, not only on the fault current magnitude but a Voltage and current signals are the primary quantities in fault
combination of other HIF characteristics exists. detection, and from the foregoing discussions, it can be concluded
HIF characteristics are highly random and non-linear which that HIFs do not manifest clearly in voltage and/or current
adds to the challenges in developing HIF detection and variation. Moreover, the discussion alludes to the fact that; HIFs
classification using pattern recognition and other feature extraction exhibit highly random and non-linear characteristics influenced by
techniques. There are several factors including, ground contact the environment and network condition. Furthermore, the physical
surface material, weather, humidity, network topology, load characteristics of HIF are specific to the condition at which the
condition and voltage level [4] that dictate the randomness in HIF
IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3699
© The Institution of Engineering and Technology 2018
fault occurred. This implies that it may be impossible to detect all and −ve peaks. As the rate of appearance of transients depends on
cases of HIFs based on a single technique that targets a single HIF the rate of change of the effective HIF impedance which is highly
signature. Thurs, for a secure and dependable HIF detection, an random, there is no guarantee that transients will occur at the
algorithm incorporating multiple techniques targeting more than points on the input signal thus robustness of the algorithm is
one HIF characteristic may be necessary. jeopardised. Moreover, the algorithm depends on slope detection,
Some of the success in identifying the characteristics of the HIF this means the presence of harmonics and noise will cause spikes in
is attributed to the algorithms and techniques proposed in the past. the output of the algorithm which can be easily confused with HIF,
The systems for HIF detection are classified as either mechanical and this has not been verified.
or electrical [11]. Mechanical protection using forced grounding to While all the proposed methods have shortcomings, they
activate the conventional protection system was introduced by contribute to the advancement of knowledge towards
Wester [12], however, this was abandoned due to reliability issues understanding the complex nature of the HIFs. As more research
and cost implication. Other methods such as lower threshold continues, a combination of different techniques and methods
settings were proposed [13], which compromised relay could evolve into the development of a universal system for
coordination resulting in nuisance/unwanted tripping. Electrical- reliable HIF detection and classification, hence the motivation for
based systems extract HIF characteristics in the time [14, 15], the work presented in this paper.
frequency [16] and time–frequency domain [17]. Development in The HIF detection and classification algorithm proposed in this
new and advance signal process techniques allowed for further paper is based on MM, a non-linear image/signal processing
improvement in the HIF detection and classification process. The technique to extract features from the voltage and current signal for
techniques involving expert systems and learning algorithms HIF detection and classification. The algorithm targets two HIF
including; decision-three based algorithm [18], Kalman filter [19], identifying characteristics, (i) randomness, due to randomly
artificial neural networks (ANNs) [20–22]; fuzzy logic and neuro- changing the effective fault resistance, Rf giving the erratic fault
fuzzy [23, 24] were proposed. Hybrid frequency–time-domain current and, (ii) arc extinguishing and re-ignition around the
techniques such as wavelet transform (WT) [25–27] and time- fundamental period giving a shoulder-shaped unsymmetrical fault
domain techniques such as mathematical morphology (MM) [28] current waveform. The proposed algorithm is a multistage
have also been proposed. The authors in [18] used a technique morphological filter, constructed from two non-linear MM filters
based on decision-tree to analyse the harmonic current magnitudes; called the morphological median filter (MMF) and the alternating
however, this does not perform well under noisy environments. In sequential filters (ASFs). MM is a non-linear image/signal
[19], Kalman filter is extended to incorporate support vector processing technique that analyses the topography of the input
machine, which is based on statistical learning theory to extract signals waveforms by means of a probing signal called the
features of magnitude and frequency of the fundamental and some structuring element (SE) in complete time domain. The SE is the
odd harmonics components. The technique requires a large dataset filtering signal that provides general functional attributes of the
for training which can become problematic if good dataset is not MM signal processing technique. MM can detect seemingly
available. The expert systems based on neural networks [20, 21], insignificant changes in the topography of the signal waveform
and probabilistic neural networks [22] are sensitive to frequency being investigated, thus making it convenient for the detection and
change and require a large training dataset as well as building classification of HIFs. The remaining of the paper is organised as
complex non-linear system by learning examples which be follows: Section 2 introduces the proposed algorithm and its
problematic with insufficient dataset or incorrect training. The mathematical derivations; Section 3 discusses the application of
authors in [23] proposed multiple techniques including WT, fuzzy morphological fault detector (MFD) algorithm in HIF detection
logic and ANN for HIF detection and feature extraction, whereas a and classification, in Section 4, simulation and results, in Section 5
genetic algorithm based on WT, principal component analysis and discussions are presented and conclusions in Section 6.
fuzzy logic is proposed in [24]. In [25] a technique based on WT is
proposed, and this is advanced by the authors in [26, 27] with
2 Proposed algorithm for HIF detection and
wavelet packet transform and boundary wavelet, respectively. The
effectiveness of WT is based on the appropriate selection of mother classification
wavelet. Moreover, the technique is not immune to noise and the In this section, the mathematical fundamentals of MM-based
effect of DC-offset. The most common physical characteristic of techniques and the design of the proposed algorithm including the
HIFs is AC electric arc. Arcing most often follows the onslaught of mathematical derivation of the MM filters used are presented.
HIFs, and it's the arcing phenomena combine with among others,
parameters such as contact surface material and weather condition 2.1 Preliminaries MM
that defines attributes of the HIF current. Most, if not all proposed
electrical methods attempt to extract features from the transients The genesis of all MM filters is two primitive MM transforms
present on the fault current induced by HIF for its classification. called ‘Dilation’ and ‘Erosion’. In MM processing of analogue
The HIF current has a time varying frequency spectrum, which signals, grey-scale dilation and erosion are applied [32].
makes the most frequency-based systems using Fourier transform Supposing, i is the fault current signal input with length x, and g is
(FT) incapable of analysing. Advance techniques using Kalman the SE with length s such that x > s, the grey-scale dilation and
filters to estimate the frequency components have been proposed erosion are obtained by algebraic addition and subtraction,
[28]. Amongst the signal processing techniques, WT/dynamic time respectively, defined by the following equations:
warping holds the most promise. WT is a time-scaling technique
where both frequency information and point of fault inception are iDIG = i ⊕ g(x) = max i(x − s) + g(s) (1)
s
captured and mapped in time. This makes WT effective in
analysing frequency spectrum with time-varying characteristics
[29, 30]. WT-based methods target transients induced by the fault iERG = i ⊖ g(x) = min i(x + s) − g(s) (2)
s
event. Since the transients due to HIF are subject to the damping on
the contact surface material, and if the fault input signals for where iDIG and iERG represent, respectively, the maximum and
feature extract are taken at the DN substation, it could result in minimum sets within the neighbourhood of the SE obtained
availability of insignificant transients which will compromise the through grey-scale dilation and erosion of the fault current signal
effectiveness of WT-based methods. Moreover, Ghaderi et al. [2] i(x). All other MM transformers that form the basis for any MM
report that despite having significant advantages, WT-based filters are derived from these fundamental transforms.
techniques are subject to (i) narrow frequency support, (ii) choice
of appropriate mother wavelet, and (iii) loss of feature resolution.
In [31] a technique based on MM is proposed where transients on
the fault voltage signal are analysed. However, the algorithm is
limited to detecting the transients occurring only at or near the +ve

3700 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018
2.2 Design of the MFD algorithm and mathematical ∀n = 0, 1, 2, …, N/2, and m = 1 is the centre point with all other
derivations values of m appearing symmetrically on both sides such that m ≤ n
must be observed. In the interest of maintaining computational
The HIF detection and classification technique presented in this efficiency, and moreover; considering the sampling rate used, an
paper proposes to use voltage and current signals measured at the SE of five points is sufficient for the real-time application. Using
distribution network substation. The measured voltage and current three, points of the new SE are determined by m = (2h + 1), ∀h = 0,
signals provide inputs to the HIF detection and classification 1, 2 such that m = 1 is the centre point, and points m = 3 and m = 5
algorithm centrally located at the substation. The occurrence/ located on both sides of the eccentrically decreasing SE with two
existence of HIFs is detected by continuously monitoring and linear but unequal slopes. To further enhance computational
analysing the extracted voltage and current signals for changes in efficiency, the five points SE is decomposed into two separate SEs
current and/or voltage magnitude and the presence of the targeted with three points denoted A1 and A2 where A1 = [cos1ϕ, 1, cos1ϕ]
HIF characteristics.
and A2 = [cos3ϕ, 1, cos3ϕ]. Decomposition of the SE suggests that
In the proposed algorithm, two MM filters, namely MMF and
ASF are strategically combined, and with the use of the designed the two components must be sequentially applied where A1 must be
SE to create the decomposed open-close alternating sequence dilated and/or eroded by A2 to cover the length of the SE [36]. Two
(DOCAS) to extract the HIF features from MFD signal output for other SEs designated B1 and B2, where B2 > B1 is composed with n 
HIF detection and classification. The mathematical derivations for = 2 for B2 = [cos3ϕ, cos1ϕ, 1 cos1ϕ, cos3ϕ] and n = 1 for B1 = 
the entire signal processing technique in creating the MFD output [cos1ϕ, 1, cos1ϕ] are also derived from (3). A1 and A2 are used in
are described herein. The MFD algorithm utilises 64 samples per the MMF section and while B1 and B2 are used in the ASF section.
cycle. For a 50 Hz system, a sampling frequency, fs = 3200 Hz is
The general SE defined in (4) and its decomposition as
required giving a sampling rate, ΔT = 312.5 µs. Thus, based on this described results in a weighted SE. Moreover, the sampling of the
this sampling rate, a sampled version of the input signal for a full fault signal results in pointwise weighted dilation and erosion at
fundamental cycle can be represented by the following equation: each sample by the elements of the SE. Weighted dilation and
erosion are implied in this application as shown in (5)–(8), and
π
f (n) = F0rn + F1sin n+α−θ ∀n = 0, …, N − 1 (3) thereafter.
N
2.2.2 Morphological median filter: The MMFs are averaging
where r = e−(ΔT/τ), and the first term, Iorn is the DC component filters that combine the dilation and erosion transforms
while the second term is the ac component. Instead of the full simultaneously with the same SE to produce an average version of
cycle, a quarter cycle of the input cycle has found to be sufficient the input signal. The decomposition of the SE into A1 and A2 gives
for fault detection; hence a window size of 16 samples such that N  rise to two stages of MMF in the MFD algorithm called MMF-
= 16 is used. Therefore, let F{n1, n2, n3, …, nN} represent a Stage 1 and MMF-Stage 2 where A1 and A2 used, respectively;
window in the domain of the fault signal f[n] to be processed by the such that the output of stage 1 is cascaded to stage 2. A window
MFD algorithm as a sliding window with each window separated segment in the fault signal f(n) is to be processed by the
by ΔT. The manner described herein alludes to the architecture of decomposed SEs, defined by Ak = [ak,m] ∀k = 1, 2 and m = 1, 2, 3.
the algorithm and the order in which the signal is processed to
The point wise simultaneous dilation and erosion of the fault signal
create the MFD output signal.
window by A1 are given by (5) and (6), respectively,
2.2.1 Design of the SE: The design and/or selection of an DI1 = f ⊕ A1 = max f n − a1, m /a1, m , ∀m = 1, 2, 3
optimum SE is fundamental to the MM algorithm in enhancing its m
(5)
effectiveness in extracting fault signal characteristics necessary for
the secure and reliable detection and classification of power system ER1 = f ⊖ A1 = min f n + a1, m /a1, m , ∀m = 1, 2, 3 (6)
faults. The SE can have different geometrical shapes such as line, m
square, disk or ball-shaped and may also include others bee line,
curve, triangle and so on. The use of flat or line SEs has been where DI1 and ER1 are defined as intermediate sets created by the
widely applied in a number of proposed MM algorithms in the dilation and erosion process in the MMF-Stage 1, and these are
detection of power system disturbances. However, Shih [33] cascaded to the MMF-Stage 2 where they are simultaneously
implies that the use of flat SE in some situation may produce dilated and eroded create the final sets denoted DI2 and ER2,
ambiguous results. While no specific guideline exists for the respectively, and defined by the following equations:
selection of appropriate geometry and size of the SE, the choice
depends on the particular application [34]. Thus, to detect the DI2 = DI1 ⊕ A2 = max DI1 n − a2, m /a2, m , ∀m = 1, 2, 3 (7)
occurrence of events in real-time applications, the design of the SE m
should capture attributes of the signal to be processed. Factors that
determine the selection of an optimal SE for specific application ER2 = ER2 ⊖ A1 = min ER2 n + a2, m /a2, m , ∀m = 1, 2, 3 (8)
m
include type and dimension of the signal, the frequency of
interference and the sampling rate [35]. Moreover, the length of the where DI1 and DI2 contain maximum selections of sets of sums
SE is defined by the sampling rate, thus, if the sampling rate is
from dilation operation while ER1 and ER2 contain the minimum
increased or decreased, then the length of SE must be increased or
decreased accordingly to match the change. While any sampling selections of set of differences from erosion operation within the
rate can be used, it is noted that the sampling rate dictates the neighbourhood A1 and A2 in stages 1 and 2 of the decomposed
window size as well, and thus an intuitive selection of the sampling MMF, respectively. The physical transformation on the signal is
rate to maintain an even number of samples in a window and to that, dilation increases the points on the signal while erosion
perform real-time detection of power system conditions is shrinks it. The processed signal can be reconstructed by taking the
desirable. Thus, the general structure of the SE used in the MFD average output of the decomposed MMF as defined by the
algorithm that captures the variation in faults signal inputs is following equation:
defined by (2) [32].
DI1+ER1+DI2+ER2
AvMMF = (9)
SE = cos mϕ (4) 4

where φ = 2πf0ΔT such that is the fundamental frequency, ΔT is the where AvMMF represents the magnitude transformation of the fault
sampling rate as previously defined are key features in the design signal input. It is assumed that the phase angle is maintained. Let
of the SE for real-time detection of any power system disturbance. fmmf(n) denote the transformed version of f(n) given by the
m represents the number of points in the SE where m = ((N/2) − n) following equation:
IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3701
© The Institution of Engineering and Technology 2018
π
f mmf (n) = F0mmf rn + F1mmf sin n+α−θ
N (10)
∀n = 0, …, N − 1

where F0mmf and F1mmf are the DC and AC components of the


MMF average output. The MFD algorithm is based on slope
detection to generate spikes in response to variation in the
topography of the fault signal waveform caused by the fault. The
distortion of the signal waveform from the transients following a
fault are emphasised by subtracting fmmf(n) from the original fault
input signal, f(n) to produce the initial fault detection signal given
in the following equation:
N−1
π
Δ f (n) = ΔF0rn + ∑ ΔF sin1
N
n+β (11)
n=1 Fig. 1  Response of DOCAS MFD for a simple AC signal
(a) Fault signal and average MMF output, (b) Difference fault signal, (c) MFD output
where Δf(n) = f(n)–fmmf(n) is the differential fault, and ΔF0 = F0– showing fault windows, (d) Regions for MFD spikes due to target HIF features
F0mmf, ΔF1 = F1–F1mmf and β = α–θ. ΔF0rn and ΔF1sin((π/N)n + β)
are the DC and AC components of ΔF(n) respectively. Δf(n) is A graphical illustration of the signal transformation to create the
processed further to create the MFD output for fault detection and MFD output can be demonstrated by considering a simple AC
classification. signal as shown in Fig. 1. In Fig. 1a, the input signal and its
transformed version at MMF stages are shown. The average MMF
2.2.3 Alternating sequential filter: The initial fault detection output shows the edge detection characteristic of the MMF filter,
signal in (11) is cascaded to the ASF layers to eventually create the where the edges of the data windows are visible. The edges are
MFD output. The ASF layers are constructed from two other MM fundamental to the fault detection functionality of the proposed
operations called the ‘Open’ and ‘Close’ transforms derived from algorithm. The difference fault signal obtained by subtracting the
dilation and erosion operations; such that average MMF output signal from the original fault signal where the
Δ f ○ B = ((Δ f ⊕ B) ⊖ B) and Δ f ● B = ((Δ f ⊖ B) ⊕ B) are edges are emphasised as shown in Fig. 1b. This signal provides
the open and close transforms, respectively. The open–close and input to the ASF layers where it is further transformed according to
close–open ASFs are developed based these transforms. The open (12) and (13), and signals extracted at each stage of the ASF layers
and close transforms are effective in the reduction of small positive to create the MFD output signal according to (14) and shown in
and negative regions on the fault signal, respectively. These Fig. 1c.
characteristics are effective in the reduction of any noise and
ripples that may exist in the initial fault signal. Their effectiveness 2.2.5 Fault detection windows: The MFD output signal in Fig.
is enhanced by simultaneously applying these filters in ASF mode. 1c is segmented into Fault Detection Windows. The process of
The ASFs are hierarchical filters where the transformation is summing the differences of closing and opening transformation at
performed iteratively in decreasing sequence starting with the each ASF stages resulted in the suppression of smaller spikes
largest SE. The SEs used at the ASF layers are B1(u) ∀ u = 1, 2, 3 within the fault window and giving visibility to the edge spikes.
and B2(υ) ∀ υ = 1, 2, 3, 4, 5 such that B2 > B1. The transformation The edge spikes defined the boundaries between the fault detection
windows. A fault detection window has a leading and tailing edge
of the difference fault signal, Δf at the ASF layers are, respectively,
spikes with one taller than other. The order in which they appear
defined by the following equations:
depends on the fault inception angle and where (in time) the first
sample was taken for processing. The edge spikes are relative to
OCASF = Δ f ○ B2 υ ● B2 υ ○ B1 u ● B1 u
(12) the instantaneous peak value of the fault signal. Both the tall and
∀υ = 1, 2, 3, 4, 5 and ∀u = 1, 2, 3 the short edge spikes can be considered, for fault detection,
however, in this application we will consider tall edge spikes to
COASF = Δ f ∙ B2 υ ∘ B2 υ ∙ B1 u ∘ B1 u detect variation in the fault signal magnitude. Thus, any variations
(13) in the height of the tall edge spikes directly correlate to the change
∀υ = 1, 2, 3, 4, 5 and ∀u = 1, 2, 3
in the fault signal magnitude when subjected to any disturbances.
where OCASF and COASF are sets created by the Open-close and The designed parameters of 64 samples per fundamental cycle
and a window size of 16 samples gives four fault windows per
Close-open alternating sequential filters, respectively.
fundamental period each having a time duration of 6.6875 ms
based on ΔT = 312.5 μs. This time duration is defined as twin = (N 
2.2.4 MFD output signal: The MFD signal is the sum of − 1)ΔT. The ending and beginning spikes of two adjacent windows
differences between the Close and Open operations, defined as appear consecutively and separated by ΔT. The tall edge spikes
stages of the ASF layers. The signal simultaneously transformed at then appear at an interval (2N − 1)ΔT = 9.6875 ms. This time
the ASF layers, and the transformed signal is extracted at each duration is denoted Tupdate in which the previous current magnitude
stage of the ASF layer, where a stage is defined by a closing (●)
and (○) operator giving four stages in each ASF layer. Stages 1–4 (maximum prefault MFD value of the tall edge spike) in the
go from left to right. Let's define the extracted signal at each stage memory is updated. The MFD value in the memory is used to
of the ASF layer as follows; OCASF layer: OCS(○) ∀S = 1, 3 and monitor the condition of the DN by comparing it with the current
value.
OCS(●) ∀S = 2, 4, where S denotes the stage number. Similarly, for The fault detection windows further serve the purpose of
the COASF layer: COS(●) ∀S = 1, 3 and COS(○) ∀S = 2, 4. Using capturing transient information generated by the disturbances.
this definition, and from (12) and (13), the MFD signal is During faults, transients occur, and the MFD algorithm generates
constructed as given below: spikes in response to the transient that grow inside the fault
windows within the time duration defined by Twin. These spikes are
CO1( ∙ ) − OC1( ∘ ) + OC2( ∙ ) − CO2( ∘ ) used for classifying the fault based on the nature and duration of
MFD = (14)
+⋯ + CO3( ∙ ) − OC3( ∘ ) + OC4( ∙ ) − CO4( ∘ ) the spikes as they span the fault windows. For HIF detection, the
MFD fault windows are partition into zones as shown in Fig. 1d to
where (CO1(●)–OC1(○)) = (Δf ●B2)–(Δf○B2), and (OC2(●)– extract information based on the appearance of the spikes that
CO2(○)) = ((Δf○B2) ●B2)–((Δf ●B2)○B2) and so on. correlate to the targeted HIF characteristics. This functionality of

3702 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018
Fig. 2  Flowchart of proposed HIF detection and declaration

the proposed HIF detection scheme is demonstrated by simulations detection window, and Imax(w + 1) be the MFDTall value
in Section 4. corresponding to the prefault current magnitude in the previous
window stored in the memory. The increase in fault current
3 Application of the DOCAS MFD output signal in magnitude is detected by comparing Imax(w) with Imax(w + 1). The
detection and classification HIFs per unit increase in current magnitude is defined by the following
equation:
The DOCAS algorithm continuously monitors the network
condition by testing the current signal against a threshold matric. Imax(w) − Imax(w + 1)
The threshold matric has minimum and maximum, iThresholdmin Iinc = (15)
Imax(w + 1)
and iThresholdmax, respectively. If the increase in current
magnitude, denoted Iinc is such that, Iinc > iThresholdmin, a where Iinc is the per unit increase in current magnitude. Under
disturbance detection (DD) flag is activated to alert the algorithm normal condition, Imax(w) = Imax(w + 1) such that Iinc = 0, the value
of an abnormal condition. Iinc is further tested against such that, if in the memory buffer is updated and the algorithm continuously
Iinc > iThresholdmax, then level 1, (L1) flag is activated to indicate monitor the network for changes in the current magnitude. Suppose
OC fault, and OC protection system is activated to act on this fault. a disturbance occurs at some time t, the parameter in (15) are
On the other hand, if Iinc < iThresholdmax, the level 2, (L2) flag is redefined as, Imax(w) = IFault, and Imax(w + 1) = IPrefault, such that
activated to initiate the HIF detection and classification procedure. Iinc = (IFault–IPrefault)/IPrefault. It is assumed that under fault
The sequence of operation and description of each classification condition, the current magnitude increases. In determining the
process in the algorithm in HIF detection is described by the value of the Threshold metric, it is noted that not all fault
flowchart in Fig. 2. conditions generate large fault currents, hence a minimum and
maximum threshold values, respectively denoted iThresholdmin and
3.1 Threshold classifier iThresholdmax must be determined. Based on (15), Iinc can have
The Threshold Classifier is used to (i) detect any disturbances and minimum and maximum values of 0 and 1.0 pu, respectively.
(ii) discriminate between fault normal operating conditions. A Considering that overcurrent threshold is set about two to three
threshold parameter based on the MFDTall edge spikes of the times the prefault load current, the OC threshold is iThreshold = 
IFault/2IPrefault, such that iThresholdmin and iThresholdmax are 0.5
DOCAS output from the current signal input is used to design the
threshold value. The corresponding MFD output is monitored for and 1.0, respectively. To compare Iinc with iThreshold, the
sudden change in the current magnitude. Since two tall edge MFD following scaling, iThresholdmin(1 + Iinc) is applied to Iinc, such
spikes appear in succession, Imax refers to the value of the taller that, Iinc_min = 0.5(1 + 0) = 0.5 and Iinc_max = 0.5(1 + 1) = 1.0. The
spike. Thus, let Imax(w) be the MFDTall output value corresponding iThresholdmin in consideration of noise is set at 15% giving, 0.5(1 
to the instantaneous prefault current magnitude in the present fault

IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3703
© The Institution of Engineering and Technology 2018
+ 0.15) = 0.575 pu. Thus, the threshold limits in the Threshold 3.3.1 Randomness: The randomness feature is extracted from the
Classifier are 0.575 > iThreshold > 1.0. edge spikes, MFDTall and MFDShort spikes as well as the random
Referring to the flowchart, if Iinc ≤ 0.575, no disturbance exists, spikes in the fault windows denoted MFDWindow and indicated as
and the MFDTall value in the memory is updated and the algorithm Win spikes in Fig. 1d. The edge spikes randomly vary in height due
continues monitoring the network. It must be noted that the to transients closer to the zero-crossing of the fundamental cycle
memory update occurs every Tupdate = 9.6875 ms. If the increase in and +ve and −ve peaks, respectively, while the window spikes
current magnitude corresponds to Iinc > 0.575 the activate DD flag randomly appear within the fault windows defined by Twin in Fig.
to put the algorithm on alert that a disturbance has occurred. Then 1c. The edge spikes can be separated from the other spikes to
Iinc is checked against iThresholdmax, and if Iinc > 1.0, the L1 flag is observe the variation. The edge spikes are defined by the following
activated to initiate OC protection. The OC protection part is not equations:
discussed in this paper. On the other hand, if Iinc < 1.0, then initiate
the Tdmax timer and keep checking until Tdmax is timed out. While 2kNΔT
MFDTall = ∀k = 0, …, ∞ − 1 (16)
on alert, the prefault MFDTall value in memory is held, and 2k N + 1 ΔT
subsequent MFDTall values are tested against this value. The
kNΔT
memory is only updated after the alert is cancelled by means of MFDShort = ∀k = 1, …, ∞ (17)
resetting the DD flag. k N + 1 ΔT
It must be mentioned that the periodical updating of the MFD
value in the memory makes the threshold parameter adaptive to The appearance of the elongated edges spikes is non-uniform and
changing network as well as load conditions. Hence, the threshold random, and can occur at any point from k = 0 to ∞ − 1 and k = 1,
parameter defined through this process can be used in any system. …, ∞, while HIF persists.
The third class of MFD spikes considered in HIF randomness
3.2 Timer classifier characteristic are the MFDWindow spikes. The MFDWindow spikes
are generated by those transients occurring away from the zero-
The timer classifier is defined by two time constrains denoted, Tini crossing and the +ve and −ve peaks. The MFDWindow spikes appear
and Tr, where Tini is the minimum time constrain to initiate HIF randomly inside the fault windows, and unlike MFD spikes due to
feature extraction and classification using the MFDv signal output transients from device switching and/or short-circuit faults, the
from the fault voltage signal input while Tr is the reset time, also MFDWindow spikes from HIF will prolong and span the entire
the maximum time constrain of the algorithm and occurs at 1.0 s length (in time) of the MFD output so long as the HIF persists. The
after Tini. The timer classifier is activated if 0.575 > Iinc < 1.0, and HIF feature extraction/classifier for the HIF randomness feature
the time delay from the point of fault inception to Tdmax defined as shown in the flowchart targets the regions defined in Fig. 1d to
the waiting time and denoted Tw must lapse before HIF extract the randomness feature by observing the MFDTall,
classification using the MFDv output can be initiated, thus Tini =  MFDShort and MFDWindow spikes.
Tdmax = 15.3125 ms. The waiting time delay, Tw is used to time out
any MFD spikes generated by switching transients from common 3.3.2 HIF arc extinction and re-ignition: The arc-extinction and
power system equipment and loads as these spikes could be re-ignition feature of the AC arc due to HIF is extracted by
confused for spikes relating to HIF, thus, a timer is set to allow observing the MFD output around the region defined by Arc
these spikes to naturally extinguish before initiating HIF feature Spikes in Fig. 1d. The HIF feature extraction/classifier for the HIF
extraction and classification. arc extinction and re-ignition feature shown in the flowchart targets
The timing for the timer classifier, defined as Tdmax is based on this region to extract the arc extinction and re-ignition feature by
observing the MFDArc spikes. The MFDArc spikes are naturally
¾ cycles +1 sample, which is also the time delay for the OC
protection system to respond to any OC. The Tdmax time delay is generated by the intersection of the current and voltage waveform
once every half cycle. If the voltage and current signal waveforms
based on the sampling rate used in the design of the algorithm.
are in phase then this happens at zero magnitude of each signal. If
Considering that the sampling rate, ΔT = 312.5 μs and a fault
they are not in phase, then at the point of intersection, the
window is size, N = 16, then ¾ cycles +1 sample gives Tdmax = (3N 
instantaneous voltage and current values are equal. In a 50 Hz
+ 1)ΔT = 15.3125 ms. This can also be determined in terms of system, this repeats every 0.01 s. A spike is generated in the MFD
MFD edge spikes. Let tf = nfΔT ∀nf = 0, …, N − 1 denote the fault output at the point of intersection that correlates to this
inception point, where nf is the faulted sample within a fault phenomenon. Let's denote the point of intersection as Tθ, then the
window. The first point of fault level detection, denoted t1 is the MFD spikes can be defined by the following equation:
beginning edge spike of the next consecutive window such that, t1 
= (N − nf + 1)ΔT ∀nf = 0, …, N − 1. Since there are five edge spikes, MFDArc = T θ + 2kNΔT ∀k = 0, …, ∞ − 1 (18)
the last point of detection is the beginning edge spike of the fourth
consecutive window after fault inception, denoted t5 = (3N − nf +  The AC arc extinguishes and re-ignites twice every have cycle. The
1)ΔT∀nf = 0, …, N − 1, such that Tdmax = t5. arc remains extinguished for a short period of time while waiting
Thus, the total time delay in HIF detection from the time of for the voltage to regain until it reaches the restriking level to
point inception is Tw + Tr = 15.3125 ms + 1.0 s = 1.0153125 s. overcome the gap separation break down voltage for arcing to
happen again. This phenomenon impacts the fault voltage at
specific and fixed period defined in (18). The arc extinction and re-
3.3 HIF classification using feature extraction from MFD ignition thus causes the MFD spikes defined as MFDArc to grow in
output signal
height when HIF occurs. Hence, the MFDArc spikes can be
The L2 flag is activated to initiate the HIF feature extraction from separated to observe the for the arc extinction and re-ignition HIF
the MFD output from the voltage signal input, denoted MFDv after characteristics.
Tdmax timer times out. The targeted HIF feature characteristics are
the randomness and HIF AC arc extinction and re-ignition resulting 3.3.3 Decision logic: Decision logic is used to test the existence
in the shoulder-shaped unsymmetrical fault current waveform. of the target HIF features. The flowchart shows the conceptualised
These HIF characteristics cause the variation MFD value of the HIF feature classifiers for the detection of HIF. The classifiers
edges as well as cause spikes to grow in specific regions of the operate simultaneously to extract the randomness and arc
MFD output signal. extinction and re-ignition features of the HIF from the target MFD
spikes. The HIF features must be detected within the upper time
limit constraint defined as Tr in the timer classifier. A decision

3704 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018
Fig. 3  Simulation systems
(a) IEEE 13 bus distribution feeder test system, (b) High impedance Arc model used in the simulations

Table 1 Contact surfaces in HIF simulations


Contact surface Vp, V Vn, A ΔV, V Rp = Rn, % IHIF, A
1 750 900 150 138 ± 10 15
2 1000 1175 175 98 ± 10 20
3 1200 1400 200 70 ± 10 25
4 1300 1550 250 43 ± 10 40
5 1400 1750 350 33 ± 10 50
6 1500 2000 500 23 ± 10 75
Contact surfaces: 1 – wet sand, 2 – dry sod, 3 – dry grass, 4 – wet sod, 5 – wet grass and 6 – reinforced concrete.

logic is used to test if the randomness and arc extinction and re- 4.1 HIF model
ignition HIF features exist. HIF can only be declared if both
features exist. The proposed decision logic is based on Boolean The HIF model used in the simulation is shown in Fig. 3b. A line
logic and a set of if then rules. The following rules constitute the voltage of 4.16 kV on the LV side represented by Vph is connected
decision logic: between the faulted phase and ground. The DC voltage sources, Vp
and Vn connected in series with respective diodes model the arc
Rule 1: if randomness;{MFDvTall AND MFDvShort AND voltage, and have unequal magnitudes. The resistance Rp and Rn
MFDvWindow} AND arc extinction/re-ignition{MFDvArc} are represent the arc resistance, and vary randomly to model the arcing
present, then initiate HIF alarm. phenomena of HIF giving an erratic unsymmetrical positive and
Rule 2: elseif randomness;{MFDvTall AND MFDvWindow} AND negative half cycle of the fault current.
arc extinction/re-ignition{MFDvArc} are present, then initiate HIF
alarm. 4.1.1 Contact surfaces: In the HIF model used, the parameters
Rule 3: elseif randomness;{MFDvShort AND MFDvWindow} AND Vp and Vn model the contact surfaces. It was experimentally shown
arc extinction/re-ignition{MFDvArc} are present, then initiate HIF by Emanuel et al. [6] that during HIF, the HIF current is
unsymmetrical with the positive half cycle having higher value
alarm.
than the negative half cycle, [37]. Hence, to model this
Rule 4: elseif no HIF alarm, then reset, Tr, DD flag and L2 flag. phenomenon, Vn must be greater than Vp(Vn > Vp), and Vn–Vp = 
ΔV, where ΔV is unsymmetrical voltage. Moreover, it was shown
HIF is only declared if both features are present. The condition
that, less densely packed contact surface (soil) yields a higher arc
for HIF must be detected and declared within the time limits, thus
voltage than contact surface with high density. Using this as the
it takes 1.0153125 s from the point of fault inception to reach a
guide, contact surfaces in [38] were modelled to obtain the
decision if HIF exists or not.
specified current magnitudes. Furthermore, the values for Rp and
Rn are specified in Table 1. These parameters randomly vary
4 Simulations and results between +10% of the specified steady-state values and represent
The IEEE Power Systems Relaying Committee recommends HIF the effective fault resistance for positive and negative half cycles,
studies to be conducted at voltage levels at 15 kV or below. Test respectively. Moreover, the rate of variation of the resistance used
and performance verification of the proposed algorithm were for surface is 40ΔT, 35ΔT, 30ΔT, 20ΔT, 15ΔT and 10ΔT going
conducted through simulation studies using the IEEE 13 bus test from contact surfaces 1 to 6.
system which was modelled in MATLAB/Simulink. The IEEE 13 The unequal positive and negative arc voltages and the
bus system shown in Fig. 3a is a heavily loaded unbalance system randomly changing arc resistances generate non-linear V–I
operating at 4.14 kV. Several cases of common power system loads characteristics in each of the contact surfaces. This phenomenon is
including capacitor and induction motor switching, a case of single graphically illustrated in Fig. 4 where the V–I characteristic curves
line-to-ground (SLG) fault and cases of HIFs were simulated at for each contact surface are shown.
different points on the test system. Voltage and current
measurements were taken at the substation via the secondary 4.2 Simulation results
windings of the VT and the CT, respectively. The efficiency of the
CT and VT have not been modelled, and included in the Simulation cases were performed to demonstrate the different
simulations. However, it is assumed that the VT and CT have classification process illustrated in the flowchart in Fig. 2.
similar saturation characteristics and operate within their linear
region as the fault current is very small, and both having turns 4.2.1 Threshold classification: Several power system conditions,
ratios of 1:500. This property is accounted for in the algorithm by including 100 kVar capacitor switching, 100 kW induction motor
diving each quantity by a factor of 500. switching, 50 kW step load increase and a case of SLG fault were
simulated at various points on the test system of Fig. 3a at different
inception time, Ts. The results of an increase in fault current

IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3705
© The Institution of Engineering and Technology 2018
Fig. 4  V–I characteristic curves of the simulated contact surfaces

Table 2 Time duration for the existence of the MFD spikes due to non-HIF transients
Event Node Phase Ts, s MFDiTall edge MFDvWindow
Value IInc, pu Tdw, ms
SLG 632 A 0.030 0.4066 1.8299 —
0.4279 1.9257
100 kVar cap switching on 611 A 0.030 0.1364 0.6058 4.6875
0.1111 0.5000
646 B 0.035 0.0821 0.5782 9.6875
0.0721 0.5077
652 C 0.037 0.1110 0.6160 12.031
0.0901 0.5000
675 ABC 0.033 0.1212 0.5455 7.0312
0.1111 0.5000
100 kW IM switching on 632 ABC 0.035 0.1936 0.8713 9.6875
0.1942 0.8740
675 ABC 0.037 0.1942 0.8740 7.0313
0.1950 0.8776
50 kW step Load Inc. 645 AB 0.037 0.1184 0.5329 2.4375
0.1179 0.5306
675 ABC 0.030 0.1328 0.5977 2.4375
0.1313 0.5910

magnitude and the existence of the window spikes for each A graphical illustration of the existence of the MFDvWindow
condition are tabulated in Table 2. The increase in fault current spikes for a case of capacitor switching, induction motor switching,
magnitude was computed using the process described in Section step load increase and SLG fault are shown in Fig. 5. In can be
3.1 using the MFDiTall edge spikes from the current signal input. observed that in each of these cases, the MFDvWindow spikes do not
The prefault MFDiTall values used computing the increases for each surpass the ¾ cycles +1 sample (15.3125 ms) limit. The longest
phase are 0.1111, 0.0710 and 0.0901, respectively, for phases A, B time duration of 12.031 ms is for the MFDvWindow spikes for
and C. Note that, in three-phase switching cases, the values in capacitor switching which exists for about two fault windows
Table 2 are phase A result. translating to half a cycle. Thus, the Time Classifier ensures that,
The results of the increase in fault current magnitude show that any non-HIF related MFDvWindow spikes are naturally extinguished
SLG sustained the highest increase of Iinc > 1.0. This is an within the time delay Tw = Tdmax.
overcurrent fault, and hence, HIF detection mechanism will not be
triggered. The alert will be removed when the fault is cleared by
4.2.2 HIF randomness and Arc extinction and re-ignition
the OC protection mechanism. All other conditions resulted in Iinc 
features classification: Several cases of HIFs on different contact
< 1.0 which shows that there are not OC conditions. However, the surfaces defined in Table 1 were simulated at various locations
following conditions including, capacitor switching at node 611 on (Lines) on the test system at a different fault inception time (Ts) to
phase A, node 646 on phase B, node 652 on phase C, all cases of
test the Randomness and the Arc extinction and re-ignition HIF
induction motor switching and three-phase step load increase at
features extraction/classifications. In all cases, the Threshold
node 675 resulted in Iinc > 0.575 which would activate the DD flag.
Classification to initiate HIF feature extraction/classification was
The waiting in the Threshold Classifier defined by Tdmax must time satisfied. The results from these simulations are tabulated in
out before HIF classification can be initiated. The time duration Table 3 which show the time of appearance of MFDvTall,
denoted Tdw is the time delay of the existence of the MDFvWindow MFDvShort and MFDvArc spikes of the faulted phases that went
spikes for each condition as shown in Table 2. It can be noted from above the prefault value of each category of the MFDv spikes
these results that the time duration for the MFDvWindow spikes for
according to the procedure defined in Section 3.1, however, applied
each case does not surpass Tdmax = 15.3125 ms which means all to MFDv output values. The prefault MFDvTall, MFDvShort, and
MFDvWindow spikes will be extinguished by the time HIF MFDvArc values for phases (Ph) A, B and C are, respectively, listed
classification using the MFDv is initiated. Moreover, the none below:
existence of the MFDvWindow spikes will not satisfy any rule for the
detection of HIF based on the Decision Logic Rules in Section MFDvTall: 0.4951, 0.5327, 0.4027
3.3.3. Therefore, the disturbances detected are not HIFs. MFDvShort: 0.1512, 0.0633, 0.3035
MFDvArc: 0.01090, 0.0130, 0.0109

3706 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018
Fig. 5  MFDvWindow spikes
(a) SLG fault condition, (b) 100 kVar cap switching, (c) 100 kW induction motor switching, (d) 50 kW step load increase

Table 3 Time of appearance of the MFDv spikes due to HIF transients


Contact surface Line Ph Ts, s MFDv Spikes Time of appearance of MFDv spikes, s
1st 2nd 3rd 4th 5th 6th 7th
wet sand (1) 602 A 0.37 MFDTall — — — — — — —
MFDShort 0.7147 0.7647 0.9647 — — — —
MFDArc 0.3984 0.4184 0.4384 0.4584 0.5384 0.6984 —
dry sod (2) 603 B 0.35 MFDTall 0.3947 0.4247 0.4947 — — — —
MFDShort 0.3797 0.4097 0.4397 0.4797 — — —
MFDArc 0.3663 0.3863 0.4063 0.4263 0.4463 0.4863 0.5063
dry grass (3) 605 C 0.35 MFDTall — — — — — — —
MFDShort — — — — — — —
MFDArc 0.3784 0.3984 0.4184 0.4384 0.4684 0.4984 0.5184
wet sod (4) 602 A 0.37 MFDTall — — — — — — —
MFDShort 0.4447 0.4647 0.7477 0.8647 0.9447 — —
MFDArc 0.3984 0.4184 0.4384 0.4484 0.4684 0.4784 0.4884
wet grass (5) 605 C 0.30 MFDTall — — — — — — —
MFDShort 0.3250 0.3450 0.3750 0.4050 0.4250 — —
MFDArc 0.3784 0.3984 0.4084 0.4184 0.4484 0.4584 0.4784
concrete (6) 602 A 0.37 MFDTall 0.4997 0.5800 0.5997 7997 — — —
MFDShort 0.3947 0.4050 0.4247 0.5950 0.7950 — —
MFDArc 0.4684 0.4884 0.5084 0.5284 0.5484 0.5684 0.5884

It must be noted that Table 3 only contains the time of duration of the HIF. As defined, the random appearance of the
appearance of the first seven MFDv spikes that went above the MFDvTall, MFDvShort and sustained appearance of MFDvWindow
lower threshold value within the upper time constrain (Tr). It must spikes indicate the HIF randomness characteristic. The randomness
be stated that, there is no Counter Classification to determine the is signified by the non-uniform time of appearance of the MFDvTall
minimum number of MFDv spikes to determine if HIF exists. The and MFDvShort spikes. The non-uniform time sequence can be
first seven MFDv spikes are recorded to show the random time easily observed in the time data recorded in Table 3 for all cases.
sequence of appearance of the spikes. Moreover, any assumptions The HIF arc extinction and re-ignition characteristic is signified by
of HIF detection for each test scenario are based on the recorded the MFDvArc, and it can be observed that in all cases, the arc
data in Table 3. It was noted during the simulation that in all cases, extinction and re-ignition characteristic is consistently present.
the random MFDvWindow spikes appeared and sustained for the This feature will be present in the presence of arcing.

IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3707
© The Institution of Engineering and Technology 2018
Fig. 6  DOCAS MFD response to HIF
(a) (i) HIF current signal, (ii) fault voltage, (iii) MFDv output; (b) (i) MFDvTall(1), (ii) MFDvTall(2); (c) (i) MFDvShort(1), (ii) MFDvShort(2), (iii) MFDvWindow and (d)
MFDvArc

The results in Table 3 must be tested cases by the case against ΔT and repeat at an interval of 2NΔT. In (17), the sequence of
the Decision Logic to determine if the HIF condition has been appearance of the two consecutive MFDShort edge spikes is shown
detected. Referring to the Decision Logic in Section 3.3.3, all cases where these spikes are separated by ΔT and repeat at an interval of
except HIF phase C of line segment 605 contacting with dry grass 2NΔT as well. The MFDTall edge spikes and MFDShort edge spikes
surface is not detected based on the first seven random MFDv are separated by the time delay of Twin (refer Fig. 1c).
spikes. While dispersed sustained MFDvWindow spikes were Extraction of features for HIF arc extinction and re-ignition
observed in this case, however, these by themselves with MFDvArc characteristic: The random variations in the heights of the
is insufficient to declare HIF. MFDvArc, spikes are observed in the regions defined as, arc spikes
The following can be observed from the result in Table 3. for the arc extinction and re-ignition characteristic. The sequence
Contact surfaces resulting in higher HIF current (refer Table 3) of appearance of the MFDvArc spikes is defined in (18).
resulted in more transients with the inception of HIF thus The parameter Tθ represents the point of intersection in the
generating more random MFDv spikes. This is visible in the results current and voltage waveforms, and under the network and load
for wet sand where no MFDvTall spikes are shown. In the case of conditions in which the simulations were conducted, the time delay
line 605, the point of fault is further away from the substation is Tθ = 0.0084375 s from T0 = 0 s was recorded for phase A in the
where the measurements were taken thus resulting in that HIF system.
condition not being detected. Furthermore, the dry grass contact The extracted MFDvTall, MFDvShort, MFDvWindow and MFDvArc
surface is overdamped, resulting in almost no transients to generate spikes from the MFDv output are graphically illustrated in Fig. 6.
the MFD spikes.
In Figs. 6a(i)–(iii), the HIF current signal, the faulted phase voltage
In order to make these observations, and record the data,
and MFDv output are shown, respectively. Figs. 6b(i) and (ii) show
MFDvTall, MFDvShort and MFDvArc spikes defined in (16), (17) and
the MFDvTall(1) and MFDvTall(2) spikes, respectively, and Figs.
(18), respectively, were extracted. Moreover, by separating these
spikes, the random MFDvWindow spikes extracted as well. The 6c(i)–(iii) show the MFDvShort(1), MFDvShort(2), and the random
target regions on the MFDv output for the extraction of HIF MFDvWindow spikes, respectively. The MFDv output shows the
randomness and arc extinction and re-ignition characteristics are random variation in the heights the edge spikes as the well
specified in Fig. 1d. sustained random appearance of the MFDvWindow spikes after the
Extraction of features for HIF randomness characteristic: The onslaught of HIF at Ts = 0.37 s. The target MFDv spikes must be
random variations in the heights of the MFDvTall, and MFDvShort detected over the entire time duration defined by Tr must, then
spikes and the random appearance of the MFDvWindow spikes are apply the threshold to determine if HIF randomness feature is
observed in the regions defined as, Tall Edge Spikes, Short Edge present or not. The arc extinction and re-ignition feature can be
Spikes and Win Spikes, respectively, for the randomness visual in Fig. 6d where the MFDvArc spikes are shown. These
characteristic. Equation (16) shows the sequence of appearance of spikes will also randomly increase in height due to the HIF AC arc
the two consecutive MFDTall spike. These spikes are separated by extinction and re-ignition about the zero-crossing of the
3708 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018
Fig. 7  DOCAS MFD response to HIF
(a) (i) HIF current signal, (ii) fault voltage, (iii) MFDv output; (b) (i) MFDvTall(1), (ii) MFDvTall(2); (c) (i) MFDvShort(1), (ii) MFDvShort(2), (iii) MFDvWindow, and (d)
MFDvArc

fundamental cycle. The random increase in the height of the despite the presence of noise, the target HIF features are still
MFDvArc signifies the occurrence of arcing, and must be observed present.
for the duration of Tr, then determine by means of the decision
logic if both the disturbance is HIF. 5 Discussions
Through the simulated cases, it was observed that those contact
4.3 Noise consideration surfaces resulting in low HIF current had low level of transients.
A noise signal was injected on phase A at node 672 giving a signal- Moreover, the distance of fault from the point of measurement, as
to-noise ratio of 46.42 dB on the voltage signal. A case of HIF on observed from the HIF case at line segment 605 on dry grass
the concrete surface was initiated at 0.37 s on phase A on line 602. resulted lower transients as well. Lower level of transients resulted
The effect of noise on the target MFDv spikes for HIF feature in less MFDv spikes. In all simulated cases, it was observed that
extraction are graphically illustrated in Fig. 7. In Fig. 7a, the HIF MFDv spikes were present. While it is not possible, based on the
current, the fault voltage and MFDv output are shown. The HIF decision logic rules to declare HIF based on this feature alone, the
current is impacted by the noise as shown in Fig. 7a(i), thus adding DOCAS MFD algorithm detected variation in the MFDvArc spikes
more randomness which generate spikes on the MFD output in Fig. in all cases.
7a(iii) even before the occurrence of the HIF. The two tall edge In considering the performance of the proposed algorithm under
spikes (MFDvTall(1) MFDvTall(2)) in Fig. 7b show no significant noise condition, it was observed that, while the method can be
difference in height after the HIF, thus it is not possible to effective in HIF detection under low noise condition, at higher
determine if HIF exists. However, considering the short edge noise level it could be difficult to clearly distinguish the impact of
spikes, (MFDvShort(1) and MFDvShort(2)), the increase in MFD HIF from noise. Thus, to improve detection under the noisy
values is profound and easily distinguishable from the noisy condition, the detection delay must be increased until more HIF
prefault condition. The short edge spikes increased in the height characteristics become apparent. Moreover, filtering to remove
above the prefault noise MFD values as shown in Figs. 7c(i) and higher frequency noise above the frequency spectrum of 2–10 kHz
(ii). The same effect is also observed at the window spikes in Fig. could improve the performance of the algorithm.
7c(iii), where the level of random MFD spikes increased from the In considering the practical implementation of the algorithm, it
prefault noisy condition. This is generally because the short MFD is necessary to maintain precise timing to extract the target MFDv
spikes target transients at the +ve and −ve peaks. Observing Fig. spikes for the randomness the arcing characteristics of HIF.
7a(1), the +ve and −ve peaks are impact by noise heavily than Moreover, the phase angle between the voltage and current must be
about the zero-crossing. This shows that the onslaught of HIF monitored to determine their point of intersection for the extraction
creates more chaos, thus the increased in the height of the short of the MFDvArc spikes. Thus, phasor measure units at the point of
MFD spikes and random window spikes. The arc extinction and re- measurement need to be installed to extract the information.
ignition spikes do exist in the presence of the noise as shown in
Fig. 7d. However, their rate of appearance is not regular. Thus,

IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710 3709
© The Institution of Engineering and Technology 2018
6 Conclusion [16] Aucoin, B.D., Russell, B.M.: ‘Distribution high impedance fault detection
utilizing high frequency current components’, IEEE Trans. Power Appar.
The DOCAS MFD algorithm presented in this paper utilises the Syst., 1982, PER-2, (6), pp. 46–47
[17] Huang, S.-J., Hsieh, C.-T.: ‘High-impedance fault detection utilizing a Morlet
MMF and the Open-close and Close-open alternating sequential wavelet transform approach’, IEEE Trans. Power Deliv., 1999, 14, (4), pp.
filters, and enhanced by a decomposed eccentrically decreasing 1401–1410
convex SE designed using power system signal properties to detect [18] Sheng, Y., Rovnyak, S.M.: ‘Decision tree-based methodology for high
and distinguish between very small, seemingly insignificant impedance fault detection’, IEEE Trans. Power Deliv., 2004, 19, (2), pp. 533–
536
changes in the input signal waveform to extract the target HIF [19] Samantaray, S.R., Dash, P.K.: ‘High impedance fault detection in distribution
randomness and arc extinction and re-ignition signature feeders using extended Kalman filter and support vector machine’, Eur. Trans.
characteristics. The algorithm has been tested under several cases Electr. Power, 2010, 20, (3), pp. 382–393
of power system operating and disturbances including several [20] Eissa, M.M., Sowilam, G.M.A., Sharaf, A.M.: ‘A new protection detection
technique for high impedance fault using neural network’. Proc. of Large
scenarios of HIFs on different contact surfaces and noise injection, Engineering Systems Conf. Power Syst., Halifax, NS, Canada, July 2006, pp.
and the results promise in HIF detection. However, it is concluded 146–151
that, further improvement can be made by incorporating effective [21] Baqui, I., Zamora, I., Mazon, J., et al.: ‘High impedance fault detection
noise filtering. Moreover, further tests considering the efficiency of methodology using wavelet transform and artificial neural networks’, Electr.
Power Res., 2011, 81, (7), pp. 1325–1333
the measurement VT and CT will improve the profile of the [22] Samantaray, S.R., Panigrahi, B.K., Dash, P.K.: ‘High impedance fault
algorithm. detection in power distribution networks using time-frequency transform and
probabilistic neural network’, IET Gener. Transm. Distrib., 2008, 2, (2), pp.
261–270
7 References [23] Etemadi, A.H., Sanaye-Pasand, M.: ‘High impedance fault detection using
[1] Torres, V., Guardado, J., Ruiz, H., et al.: ‘Modeling and detection of high multi-resolution signal decomposition and adaptive neural fuzzy inference
impedance faults’, Int. J. Electr. Power Energy Syst., 2014, 61, pp. 163–172 system’, IET Gener. Transm. Distrib., 2008, 2, (1), pp. 110–118
[2] Ghaderi, A., Ginn, H.L., Mohammadpour, H.A.: ‘High impedance fault [24] Haghifam, M.R., Sedighi, A.R., Malik, O.P.: ‘Development of a fuzzy
detection: a review’, Electr. Power Syst. Res., 2017, 143, pp. 376–388 inference system based on genetic algorithm for high-impedance fault
[3] Darwish, H.A., Izzularab, M.A., Taalab, A.M.I.: ‘Modeling and experimental detection’, IET Gener. Transm. Distrib., 2006, 153, (3), p. 359
verification of high impedance arcing fault in medium voltage networks’, [25] Bakar, A., Ali, M., Tan, C., et al.: ‘High impedance fault location in 11 kV
IEEE Trans. Dielectr. Electr. Insul., 2007, 14, (2), pp. 375–383 underground distribution systems using wavelet transforms’, Int. J. Electr.
[4] Kim, C., Russell, B.D., Watson, K.: ‘A parameter-based process for selecting Power Energy Syst., 2014, 55, pp. 723–730
high impedance fault detection techniques using decision making under [26] Mahari, A., Seyedi, H.: ‘High impedance fault protection in transmission lines
incomplete knowledge’, IEEE Trans. Power Deliv., 1990, 5, (3), pp. 1314– using a WPT-based algorithm’, Int. J. Electr. Power Energy Syst., 2015, 67,
1320 pp. 537–545
[5] Sedighi, A.-R., Haghifam, M.-R., Malik, O., et al.: ‘High impedance fault [27] Costa, F.B., Souza, B.A., Brito, N.S.D., et al.: ‘Real-time detection of
detection based on wavelet transform and statistical pattern recognition’, transients induced by high-impedance faults based on the boundary wavelet
IEEE Trans. Power Deliv., 2005, 20, (4), pp. 2414–2421 transform’, IEEE Trans. Ind. Appl., 2015, 51, (6), pp. 5312–5323
[6] Emanuel, A., Cyganski, D., Orr, J., et al.: ‘High impedance fault arcing on [28] Samantaray, S.: ‘Ensemble decision trees for high impedance fault detection
sandy soil in 15 kV distribution feeders: contributions to the evaluation of the in power distribution network’, Int. J. Electr. Power Energy Syst., 2012, 43,
low frequency spectrum’, IEEE Trans. Power Deliv., 1990, 5, (2), pp. 676– (1), pp. 1048–1055
686 [29] Banejad, M., Ijadi, H.: ‘High impedance fault detection: discrete wavelet
[7] Zamanan, N., Sykulski, J.K.: ‘Modelling arcing high impedances faults in transform and fuzzy function approximation’, J. AI Data Mining, 2014, 2, (2),
relation to the physical processes in the electric arc’, WSEAS Trans. Power pp. 149–158
Syst., 2006, 1, (8), pp. 1507–1512 [30] Varma, N., Ram, D., Anjaneyulu, D.: ‘Development of fault detection
[8] Sultan, A., Swift, G., Fedirchuk, D.: ‘Detecting arcing downed-wires using algorithm for high impedance faults in distribution network using
fault current flicker and half-cycle asymmetry’, IEEE Trans. Power Deliv., multiresolution analysis’, Int. J. Eng. Res. Technol., 2014, 3, (9), pp. 573–576
1994, 9, (1), pp. 461–470 [31] Gautam, S., Brahma, S.M.: ‘Detection of high impedance fault in power
[9] Santos, W., Lopes, F., Brito, N., et al.: ‘High impedance fault detection and distribution systems using mathematical morphology’, IEEE Trans. Power
location based on electromagnetic transient analysis’. Int. Conf. on Power Syst., 2013, 28, (2), pp. 1226–1234
Systems Transients (IPST2013), Vancouver, Canada, 18 July 2013, vol. 20 [32] Wu, Q.H., Lu, Z., Ji, T.: ‘Protective relaying of power systems using
[10] Thomas, M.S., Bhaskar, N., Prakash, A.: ‘Voltage based detection method for mathematical morphology’ (Springer-Verlag, London, UK, 2009)
high impedance fault in a distribution system’, J. Inst. Eng. B, 2016, 97, pp. [33] Shih, F.Y.: ‘Image processing and mathematical morphology, fundamentals
413–423 and applications’ (CRC Press, Boca Raton, FL, 2009)
[11] Chen, J., Phung, B., Zhang, D., et al.: ‘Study on high impedance fault arcing [34] Gautam, S., Brahma, S.M.: ‘Overview of mathematical morphology in power
current characteristics’. 2013 Australasian Universities Power Engineering systems – a tutorial approach’. Power & Energy Society General Meeting
Conf. (AUPEC), Hobart, TAS, Australia, 29 September–3 October 2013, pp. 2009 (PES'09), Calgary, AB, Canada, July 2009, pp. 1–7
1–6 [35] Tingfang, Y., Pei, L., Xiangjun, Z., et al.: ‘Application of adaptive generalized
[12] Wester, C.G.: ‘High impedance fault detection on distribution systems’. morphological filter in disturbance identification for power system
Papers Presented at the 42nd Annual Conf. Rural Electric Power Conf., 1998, signatures’. Int. Conf. on Power System Technology, 2006 (PowerCon 2006),
1998, pp. c5-1–5 Chongqing, China, October 2006, pp. 1–7
[13] Sharaf, A., Abu-Azab, S.: ‘A smart relaying scheme for high impedance faults [36] Shih, F.Y.-C., Mitchell, O.R.: ‘Decomposition of gray-scale morphological
in distribution and utilization networks’. 2000 Canadian Conf. on Electrical structuring elements’, Pattern Recognit., 1991, 24, (3), pp. 195–203
and Computer Engineering, Halifax, NS, Canada, May 2000, vol. 2, pp. 740– [37] Sedighi, A., Haghifam, M.: ‘Simulation of high impedance ground fault In
744 electrical power distribution systems’. Int. Conf. on Power System
[14] Calhoun, H., Bishop, M.T., Eichler, C.H., et al.: ‘Development and testing of Technology, 2010 (PowerCon 2010), Hangzhou, China, October 2010, pp. 1–
an electro-mechanical relay to detect fallen distribution conductors’, IEEE 7
Trans. Power Appar. Syst., 1982, PAS-101, (6), pp. 1643–1650 [38] Aucoin, B.M., Jones, R.H.: ‘High impedance fault detection implementation
[15] Huang, C.-L., Chu, H.-Y., Chen, M.-T.: ‘Algorithm comparison for high issues’, IEEE Trans. Power Deliv., 1996, 11, (1), pp. 139–148
impedance fault detection based on staged fault test’, IEEE Trans. Power
Deliv., 1988, 3, (4), pp. 1427–1435

3710 IET Gener. Transm. Distrib., 2018, Vol. 12 Iss. 15, pp. 3699-3710
© The Institution of Engineering and Technology 2018

Das könnte Ihnen auch gefallen