Busbar Protection Using a Wavelet Based ANN

Ahmad Abdullah
Electric Power Engineers, Inc
Austin, TX 78738
Email: [email protected]
This paper seeks to provide; fault protection for the bus, back
Abstract—This paper presents a new application of wavelet up protection for adjacent lines and classification between
based artificial neural networks to the field of high voltage transients occurring on adjacent lines identifying the line
busbar protection. Any transient event type -whether fault or causing the event.
not- causes high frequency components to be generated and
The application of ANN for classification of faults has been
imposed on the fundamental frequency current. Those
components propagate from the line causing them passing given in [4]. Wavelets applications for power systems have
through the protected bus bar to the other lines connected to the been introduced in [5]. Wavelets have proven to be a very
same bus. In this paper, it is shown that those components useful tool for transmission line fault detection and
captured at any line connected to the bus can be used not only to classification. Application of wavelets has received much
detect internal and external bus faults but also to identify the attention for classification of power quality disturbance in [6],
faulted line in case of external faults. A scheme will be presented
[7], [8] and [9]. In these applications and others, energy [10]
that uses the current from any of the lines connected to the bus to
detect internal and external bus faults, classify transients on based methods seem to dominate the application of wavelets
adjacent lines and identify the line that is causing the transient in power systems. In energy based methods, the energy of
disturbance. Modal transformation is used to transform phase travelling waves can be quantized in time because of the
quantities to modal quantities. Discrete Wavelet Transform localization property of wavelets which makes it very useful in
(DWT) is used to extract high frequency components of the two fault location application [11]. The application of ANN to bus
aerial modes of the current measured. A feature vector consisting
protection is given in [12] where currents are fed to ANN for
of level 3 details coefficients of the two aerial mode currents is
used to train a feedforward neural network with one hidden training and validation. In [13] a wavelet based scheme is
layer. Results show that very accurate classification can be made presented for busbar protection where currents going in and
using one eighth of a cycle of post event data. out of bus zone are analyzed using DWT. DWT [14] produces
a series of details coefficients for each level which are used to
Index Terms-- ART neural networks, Classification infer the polarity and the direction of currents thus issuing a
algorithms, Discrete wavelet transforms, Wavelet coefficients, trip signal when a fault occurs.
Modal analysis, Power system faults, Current transformers,
lightning. This paper uses only one set of the three phase currents
from any of the lines terminating into the bus for bus
I. INTRODUCTION protection and backup protection for lines connected to the
bus. Modal transformation [15] is used to transform phase
P ower system bus protection has been traditionally
accomplished with voltage or current differential schemes
[1]. In current differential scheme, current transformers are
signals to modal signals. DWT is used to extract one eighth of
a cycle of post-event data. It will be shown that a subset of the
modal currents carries enough information for bus protection
installed on incoming and outgoing transmission lines.
and backup protection for lines terminating into that bus. It
Currents from CTs are then summed for a fault criterion. If a
will be also shown that the same subset can be used to classify
fault is outside the bus (external fault) then current summation
transient events on the lines terminating into the bus and to
equals to zero, if the currents do not add to zero then the fault
identify the line causing the transient disturbance. The feature
is on the bus (internal fault). The major disadvantage of
vector used to train a feedforward ANN [16] is a vector
current differential schemes is CT saturation which greatly
consisting of level 3 details coefficients of the two aerial
impairs the method [2]. High impedance current differential
modes [11] of currents stacked on top of each other. The
schemes have been proposed to resolve CT saturation [2],
transient event types studied in this paper are lightning, faults
however they introduce other problems. The problem of CT
and line switching. The paper is organized as follows.
saturation is eliminated completely at the source by the use of
Methodology is given in section II. EMTP model and ANN
linear couplers with voltage differential principles. Voltages
training are provided in section III. Results are shown in
are used instead of currents for fault discrimination. Currents
section IV while conclusions are presented in the last section.
in the feeders are converted to voltages using linear couplers
[3] which produce linear input output voltage relation over a
II. METHODOLOGY
very large range of input currents.
Traditional bus protection schemes do not provide backup In this paper it is shown that the information present in the
for incoming and outgoing feeders nor can they discriminate transient current available at a local substation caused by
between various transient events occurring on adjacent lines. sudden network topology changes contains sufficient
information for classification not only between internal and
978-1-5090-6617-9/17/$31.00 ©2017 IEEE
external bus faults but also between transients on adjacent using the feature vector above, one can tell which line is being
lines which effectively provides back up for lines terminating switched. If one performs only lightning cases, then using the
into the bus. Any change of system topology – caused by same feature vector, one can identify the line being hit by a
faults for example- causes a traveling wave to be generated strike. Finally, the aggregate of all transient event types can be
travelling from the point of change towards the ends of the distinguished from each other; this means if one trains ANN
line. In the simplest case this wave will be just a step but in using all transient events, then we can have three ANN output
reality will be accompanied by oscillating components. Those classes each of which corresponds to fault, line switching and
high frequency oscillating transient components are a function lightning. In summary, the procedure is as follows:
of the initial and boundary conditions at the location of the 1. Currents are decoupled using modal analysis.
event causing the change of topology. Fourier analysis is not 2. Apply DWT to the two aerial modes of one eighth of a
suitable for analyzing such waveforms because the oscillations cycle of post event information.
will be distorted and attenuated as they arrive at the line 3. Build feature vector consisting of details coefficients
terminals. However, applying discrete wavelet transform will of level 3 of the two aerial modes on top of each other
enable us to see both the spectral and temporal variations of to train the neural network.
those high frequency components. 4. A neural network of suitable size is selected for
The information from the high frequency components is training.
extracted using DWT at level 3. Level 3 is used for Training is accomplished using transient scenarios. Those
computational efficiency as the feature vector -under the scenarios include: faults, line switching and lightning.
sampling frequency used in this paper- with level 1 or 2 will
have a lot of coefficients that will make ANN training III. SIMULATION PLATFORM
impossible with modern day PCs because of the high RAM This section describes the EMTP model used and the cases
memory needed. For example, training an ANN with used to train the neural network and how the data was
MATLAB neural network toolbox needs 30 GB of RAM with prepared as input for ANN.
Level 2 coefficients. Before DWT application, phase currents
are decoupled from each other using modal analysis [15]. The A. ATP/EMTP Model Description
modal matrix is calculated at 10 kHz even though it is The topology of the system under study is presented in Fig.
frequency dependent. The author has tried different 1. This topology is part of IEEE 118 bus case [21]. Three
frequencies for the modal matrix and has not seen any effect phase currents are being measured from bus 23 to bus 25. Any
on classification. Since current transformers have a bandwidth other current set can be used but focus is casted towards this
of 100 kHz [17] which could go up to 400 kHz [18] or even set due to paper length requirement. Selection for bus 23 for
500 kHz [19], the author chooses to be on the conservative study is because it has almost all power system components
side and select the sampling rate such that the maximum around it. Bus 23 has three incoming feeders into it, namely
frequency available in the signal is 100 kHz. The time step line connecting bus 23 to bus 22 (line 23-22), line 23-32 and
used in EMTP simulations is 1μs which should theoretically line 23-24. The outgoing feeder is line 23-25. Bus 25 has a
give a maximum frequency of 500 kHz in sampled output generator connected to it through a step up transformer and
according to Nyquist theory. However, it is given in [20] that power transformer connecting it to bus 26. Bus 26 has another
the maximum frequency in the EMTP simulations is only one generator connected to it via a step up transformer.
fifth of that, so the maximum frequency is 100 KHz.
After decoupling phase quantities, DWT is applied to the
transmission line three phase currents chosen to convert them
from modal currents to a series of coefficients that will be
used for training of the neural networks. The event is detected
once a change of level 3 coefficients is detected. In theory any
level can be used as given in [33] but the author focuses on
level 3 for paper length requirements. Once the event is
detected, a window of one eight of a cycle of post event
information is used for neural network training. The outcome
of the DWT is a series of coefficients for each mode. The
coefficients of the two aerial current modes are stacked on top
of each other to build one vector that will be used to train the Fig. 1. Portion of the system under study
ANN. ANN of an appropriate size is selected for
classification. ANN is then trained using various scenarios for In the simulations, the author used very general tower
transients on lines. If one performs internal and external fault configurations for the lines around the bus being protected.
simulations on the bus and its adjacent lines then using the The line connecting buses 23 and 25 is taken from [22], while
feature vector built above, one can tell whether the bus is the lines surrounding it are taken from [23]. All tower
faulted or can identify the faulted line. If one only performs configurations have been scanned and inserted in the paper for
line switching cases on the lines terminating on the bus then the convenience of the readers. All of those towers are
modeled as lines with frequency dependent parameters and a in [26] due to their high importance for transient studies.
ground return while all others lines in the system are modeled Surge arresters have been used to protect all transmission lines
as frequency dependent models with no ground wires. We in the study area. The characteristics of those arresters are
apply transient cases only to lines 23-25, 23-22, 23-24, 23-32 selected according to the guide in [27].
and 25-27. We don’t aim to study transients beyond those
B. Creation of Transient Cases and Preparation of Cases for
lines.
Training of ANN
The argument in the paper is that high frequency
components of the currents measured at a local relay not only
can detect fault and non-fault condition on a protected bus but
also can detect fault and non-fault conditions on adjacent
lines. A large number of simulations have to be carried out to
show the validity of this argument. The author presents a very
general network topology with different tower configurations.
Creation of transient cases has been automated using the
toolbox released in [28]. For this paper, one needs to create
cases of faults, line switching and lightning. I will first
describe the creation of fault cases then line switching cases
and after that proceed to describe the lightning strike cases.
At this point it should be emphasized that the bus under study
is bus 23 which is bordered by 4 adjacent lines namely line
23-22, line 23-32, line 23-24, line 23-25. The author also
studies transients for line 25-27. The author trains ANN for
the bus under study and the 5 lines just mentioned above only.
The next subsections describe creation of transient cases for
faults, line switching and lightning and creation of feature
vector for training.
1) Creation of Transmission line Fault Cases
Fault cases are created in batches; a batch for each line.
Each batch has fault parameters, these parameters are the
following: incipient angle, fault resistance, fault location and
fault type. All types of faults have been created i.e., AG, BG,
CG, AB, BC, CA, ABG, CBG, ACG, ABC, ABCG. Incipient
angles are from 0 to 350 degrees in 10 degrees increments.
Fault resistance assumes the values: 0, 20Ω, 100Ω and 1000Ω.
The distance take the following values: 5%, 15%, 35%, 50%,
65%, 80% and 95% which are all percentages of total line
length. A total of 8066 cases per batch have been created
which gives a total of 40330 cases. Fault arc although
simulated but was not taken into account during training. This
is because fault arc causes little or almost no distortion to
phase currents which is consistent with the results in [29] and
[30]. The author verified the findings of [29] and [30] using
the fault arc model given in [31].
Fig. 2. Tower configurations of lines in the area 2) Creation of Bus fault Cases
under study. All graphics has been scanned from the A total number of 1120 bus fault cases have been generated.
references given in the paper. These faults include all types of faults mentioned in subsection
III.B.1, with the values of fault resistances and incipient
All generators in the system are assumed unregulated
angles provided in the same subsection. The bus protection
machines with no exciters or governors. This done mainly zone includes one meter from each line that terminates on the
because attention is given to one eighth of a cycle of post bus. This is due to EMTP limitation since the time step will
transient data during which exciters’ and governors’ time need to be very small to account for shorter line lengths when
constants are large compared to the period under study. not included within the protection zone, i.e., a fault within the
All power transformers have been modeled by ATP Hybrid first one meter of any line terminating on the protected bus is
Transformer Model according to [24] with typical parameters considered an internal bus fault.
provided by ATPDraw [25]. 3) Creation of Line Switching Cases
Bus capacitance and transformer stray capacitance are For line switching cases, a batch is created for each line
accounted for in the simulation platform with the values given giving a total of 5 batches. Each batch consists of two smaller
batches; each of the smaller batches corresponds to switching the author finds that a 100% classification was possible. This
by one of the circuits breakers in each terminal. For example, means that the feature vector can tell which line is being
line 23-22 would have a smaller batch for switching using the switched if all transient events were just switching.
breaker installed on terminal 23 and another smaller batch for B. Line Identification in Case of Lightning
switching using the breaker on terminal 22. The variable in
If all transient events were just lighting strike cases, then the
switching cases is the moment of switching. Switching times
feature vector built the way described in section III.C was able
ranges from 0 to 360 degrees in 2 degrees increment. This
to determine the line being hit by a strike by 100% accuracy.
way, 360 cases per batch had been created, giving a total of
An ANN with only 40 neurons in the hidden layer was enough
1800 cases. for 100% accurate classification.
4) Creation of Lightning Strike Cases
For lightning strike cases, a batch for each line has been C. External and Internal Bus Faults
created. The author uses ATP Heidler type lightning with In this case, all transient events are faults; the feature vector
rising time equal to 4 μs and a τ equal to 10 μs. The author built the way described in section III.C; use trial and error to
kept the rising and tailing times constant during simulations get an appropriate size of ANN that gives good classification
but changing their amplitude which was set to 5 kA, 10 kA, results. If one is only interested in knowing whether a fault is
15KA, 20 kA and 30 kA. Striking distances were the same as an internal bus or external bus fault, then an ANN with 20
the ones used for fault batches: 5%, 15%, 35%, 50%, 65%, neurons in the hidden layer gives 100% classification
80% and 95% which are all percentages of total line length. accuracy. If one needs to see if ANN can recognize the
The incipient angle was 0, 90,180, 270 and 330 degrees. This adjacent line being faulted, then an ANN with 80 neurons is
amounts to 630 cases per batch giving a total of 3150 cases. necessary for 100% accurate classification.
No lightning strike cases have been generated for the bus D. Transient Event Type Classification
under study because busbars are either indoor or when they In this section, the author investigates whether the feature
are outdoor they are protected either by shield wire or vector proposed can identify the transient event type. For this
lightning mast (high lattice structure with a spike on top) and purpose, the author trains the ANN with all transient cases
sometimes combinations of both depending upon type of given in subsections III.B.1, III.B.2, III.B.3 and III.B.4 such
layout of substation. that its output layer has three output classes: one for switching,
C. ANN Training and Feature Vector Building another one for lightning and a third one for faults. Building
the ANN this way and using 50 neurons in the hidden layer
Once transient cases are available from simulations, the gives the classification output in table (1).
author has a subroutine that transforms pl4 files to mat files A relay that is programmed such that it processes the local
using MATLAB. After that MATLAB is invoked to import currents using modal analysis building the feature vector as
simulations into workspace, another subroutine is used to the authors proposed can first identify the event type as given
transform phase signals to modal signals. At this stage, the in table (1). Once the event is identified, the feature vector is
author applies DWT to the two aerial mode currents then
extracts the details coefficients corresponding to level 3 aerial Table 1
modes. After that, the feature vector is built using the two
aerial modes’ coefficients of the currents stacked on top of Transient Event Type Correct Classification Accuracy -
each other. It should be noted that the feature vector can be Numbers are Percentages.
built by stacking the modes in any order but one has to be
consistent in building the vector, i.e., one method has to be Faults 100
chosen throughout training or one will get contradicting Lightning 99
results. It should be clear at this point that using only two Switching 98.9
modes as pointed in [33] instead of three phase quantities Total Correct Accuracy 99.9
makes the proposed approach faster and requires less
computational power for training as opposed to other methods then applied to corresponding ANN of the corresponding type
that use all voltages and currents for training [32]. given in IV.A, IV.B or IV.C to determine the line causing such
Finally, the author randomly divides the data set choosing transient event.
to have 70% of all cases for training, 15% for validation and
the last 15% for testing. V. CONCLUSIONS
A new application of wavelet based ANN for bus protection
IV. RESULTS has been presented in this paper. The contributions of the
This section provides classification results for faults, line paper are as follows:
switching and lighting.  The scheme provides ultra-high speed protection with
only one eighth of a cycle of post event data.
A. Line Identification in Case of Switching
 The approach doesn’t suffer from CT saturation which
In case of line switching, the feature vector is built as is considered an improvement over modern differential
described in section III.C with level 3 details of the two aerial schemes. This is because it has been shown in [32] that
mode currents stacked on top each other. After training is high frequency transients are measured well before CT
accomplished using ANN with a hidden layer of 40 neurons,
 saturation. Since only one eighth of a cycle of post [11] Magnago, F.H.; Abur, A., "Fault location using wavelets," Power
Delivery, IEEE Transactions on , vol.13, no.4, pp.1475,1480, Oct 1998
event data is used, the results of [32] apply very well
[12] Xiucheng Dong; Han Han; Mei Shu, "Research of bus protection based
for case in hand. on artificial neural network," Intelligent Control and Automation, 2008.
 It has been shown that the backup protection can be WCICA 2008. 7th World Congress on , vol., no., pp.4261,4264, 25-27
provided to adjacent lines. Backup protection is June 2008
[13] N. Perera, A.D.; Rajapakse, D. Muthumuni, “Wavelet Based Transient
provided by differentiating internal and external bus Directional Method for Busbar Protection”, International Conference on
faults via ANN classification. Power Systems Transients (IPST2011) in Delft, the Netherlands June
 The algorithm can not only detect faults, but also can 14-17, 2011
classify transient disturbances on adjacent lines [14] Albert Boggess, Francis J. Narcowich, A First Course in Wavelets with
Fourier Analysis, College Station, TX: Wiley 2009
identifying the line causing the transient event. [15] Hedman, D.E., "Propagation on Overhead Transmission Lines I-Theory
It should be noted that even though training has been done of Modal Analysis," Power Apparatus and Systems, IEEE Transactions
with regards to a certain tower configuration for line 23-25, on , vol.84, no.3, pp.200,205, March 1965
the author has tried other tower configurations and confirmed [16] Jacek M. Zurada, Introduction to Artificial Neural Systems, Louisville,
KY: West Group, 1992
that the approach in this paper is applicable to all of them [17] H. Lee, “Development of an Accurate Travelling Wave Fault Locator
which is consistent with what is given in [33], i.e., the feature Using the Global Positioning System Satellites,” proceedings of the 20th
space defined in this paper is applicable to all transmission Annual Western Protective Relay Conference, Spokane, WA, October
line towers used in practice. 1993
[18] Douglass, D.A., "Current Transformer Accuracy with Asymmetric and
The ANN architecture given in this paper cannot detect High Frequency Fault Currents," Power Apparatus and Systems, IEEE
faults that are caused by lightning strikes on the lines that Transactions on , vol.PAS-100, no.3, pp.1006,1012, March 1981
terminates into the protected bus. This issue has been rectified [19] M. A. Redfern, S. C. Terry, F. V. P. Robinson, and Z. Q. Bo, “A
in [34]. The effect mutual coupling of parallel lines has been Laboratory Investigation Into the use of MV Current Transformers for
Transient Based Protection,” proceedings of the 2003 International
ignored in this research and is a subject for further Conference on Power Systems Transients (IPST), New Orleans, LA,
investigation. Further research is still needed to identify the September–October 2003
cases that have been misclassified for another transient event [20] José Carlos G. de Siqueira, Benedito D. Bonatto, José R. Martí, Jorge A.
type in table (1). The methodology given in this paper can Hollman, Hermann W. Dommel, “A discussion about optimum time step
size and maximum simulation time in EMTP-based programs”,
detect line faults as fast it can detect bus faults. To allow line International Journal of Electrical Power & Energy Systems, Volume 72,
relays to clear the line fault first, a time delay has to be used to November 2015, Pages 24-32
delay the backup protection provided by the proposed [21] Power Systems Test Case Archive, [online]
methodology. However, a mechanism need to be implemented https://www.ee.washington.edu/research/pstca/pf118/pg_tca118bus.htm
[22] Gashimov, A.M.; Babayeva, A.R.; Nayir, A., "Transmission line
to detect clearing of the line fault via the line relay using local transposition," Electrical and Electronics Engineering, 2009. ELECO
data only to avoid tripping the bus unnecessarily which is a 2009. International Conference on , vol., no., pp.I-364,I-367, 5-8 Nov.
topic under study. 2009
Lastly, events outside the lines directly terminating on the [23] Alstom Network Protection & Automation Guide, May 2011, Page 70
[24] Sung Don Cho, “Parameter estimation for transformer modeling”, PhD
bus have not been considered as they may mislead the current dissertation, Michigan Technological University, 2002.
scheme and further research is still needed to include them in [25] ATPDraw, [Online]. Avalable: http://www.atpdraw.net
ANN training. [26] A. Greenwood, Electrical Transients in Power Systems, Wiley-
Interscience,1991,chapter 13
[27] Arrester Selection Guide, [online]
VI. REFERENCES http://www.arresterworks.com/resources/selection_guide.php
[1] A.R. Warrington, Protective Relays Their Theory and Practice Power [28] Abdullah, A, “ATPMAT: An Open Source Toolbox for Systematic
Systems, vol. I. Springer, 3rd edition. Creation of EMTP Cases in ATP Using Matlab,” proceedings of the
[2] P.M. Anderson, Power System Protection, Wiley-IEEE Press, 1st edition 2015 North American Power Symposium (NAPS), Charlotte, NC,
[3] C. Mason, The Art and Science of Protective Relaying, Wiley, 1 edition. October 4-6 2015
[4] Kezunovic, M.; Rikalo, I., "Detect and classify faults using neural nets," [29] Martin, F.; Aguado, J.A., "Wavelet-based ANN approach for
Computer Applications in Power, IEEE , vol.9, no.4, pp.42,47, Oct 1996 transmission line protection," Power Delivery, IEEE Transactions on ,
[5] Robertson, D.C.; Camps, O.I.; Mayer, J.S.; Gish, William B., "Wavelets vol.18, no.4, pp.1572,1574, Oct. 2003
and electromagnetic power system transients," Power Delivery, IEEE [30] Perera, N.; Rajapakse, A.D., "Recognition of Fault Transients Using a
Transactions on, vol.11, no.2, pp.1050, 1058, Apr 1996 Probabilistic Neural-Network Classifier," Power Delivery, IEEE
[6] Gaouda, A.M.; Kanoun, S.H.; Salama, M.M.A.; Chikhani, A.Y., Transactions on , vol.26, no.1, pp.410,419, Jan. 2011
"Wavelet-based signal processing for disturbance classification and [31] Terzija, V.V.; Koglin, H.-J., "On the modeling of long arc in still air and
measurement," Generation, Transmission and Distribution, IEE arc resistance calculation," Power Delivery, IEEE Transactions on ,
Proceedings- , vol.149, no.3, pp.310, 318, May 2002 vol.19, no.3, pp.1012,1017, July 2004
[7] Gaouda, A.M.; Salama, M.M.A.; Sultan, M.R.; Chikhani, A.Y., [32] N. Perera, A. D. Rajapakse, and T. Buchholzer, “Isolation of faults in
"Application of multiresolution signal decomposition for monitoring distribution networks with distributed generators,” IEEE Trans. Power
short-duration variations in distribution systems," Power Delivery, IEEE Del., vol. 23, no. 4, pp. 2347–2355, Oct. 2008.
Transactions on, vol.15, no.2, pp.478, 485, Apr 2000 [33] Abdullah, Ahmad. "Towards a new paradigm for ultrafast transmission
[8] Gaouda, A.M.; Salama, M.M.A.; Sultan, M.R.; Chikhani, A.Y., "Power line relaying." 2016 IEEE Power and Energy Conference at Illinois
quality detection and classification using wavelet-multiresolution signal (PECI). IEEE, 2016.
decomposition," Power Delivery, IEEE Transactions on, vol.14, no.4, [34] Abdullah, Ahmad. "A wavelet entropy approach for detecting lightning
pp.1469, 1476, Oct 1999 faults on transmission lines." 2016 IEEE/PES Transmission and
[9] Santoso, S.; Powers, E.J.; Grady, W.M.; Hofmann, P., "Power quality Distribution Conference and Exposition (T&D). IEEE, 2016.
assessment via wavelet transform analysis," Power Delivery, IEEE
Transactions on, vol.11, no.2, pp.924, 930, Apr 1996
[10] Ibrahim, W.R.A.; Morcos, M.M., "Artificial intelligence and advanced
mathematical tools for power quality applications: a survey," Power
Delivery, IEEE Transactions on , vol.17, no.2, pp.668,673, Apr 2002