See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/310620503

Fault location estimation for VSC-HVDC system using Artificial Neural Network

Conference Paper · November 2016

DOI: 10.1109/TENCON.2016.7848050

CITATIONS READS
21 414

3 authors, including:

Yew Ming Yeap Abhisek Ukil

Agency for Science, Technology and Research (A*STAR) University of Auckland
36 PUBLICATIONS 758 CITATIONS 247 PUBLICATIONS 4,838 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Abhisek Ukil on 20 October 2017.

The user has requested enhancement of the downloaded file.

 1

Fault Location Estimation for VSC-HVDC System

Using Artificial Neural Network
Somasundaram Vasanth, Yew Ming Yeap, Student Member, IEEE, and Abhisek Ukil, Senior Member, IEEE

Abstract—Recent developments in high voltage direct current the impedance merely based on the voltage and the current.
(HVDC) transmission have had a positive impact on the research In addition, the impedance-based method is considered to be
community. HVDC transmission offers superior power quality slow by the standard of VSC-HVDC system [4] as the lack of
and low transmission loss, however, the fault location and
clearance technology is not as mature as its AC counterpart. line inductance causes the DC fault to penetrate the network
This paper aims to propose a fault location technique on a 200 extremely fast. The fault analysis in [5]–[7] shows that pole-
km two-terminal VSC-HVDC system using wavelet transform to-pole fault can result in a fault current that rises to its peak
(WT) and artificial neural networks (ANN). The HVDC system is faster than AC fault and load change. For these reasons, HVDC
modeled in PSCAD and the results are treated with WT and ANN system needs to take on a different approach.
using MATLAB. The fault of concern in this paper is pole-to-
pole fault. The simulation is repeated by varying fault resistance A technique for fault location using transient fault signals
and location along 200 km to test the influence of these two in hybrid HVDC systems, based on terminal recorded data,
parameters on the proposed fault location method. The result is proposed by Hassan [8]. It involves the modification of
demonstrates reasonably high reliability in predicting the fault Rogowski coil, measuring surge times over a wide spectrum of
location with low error. currents, without saturating, and offers significant advantages
over transformers [9]. Takagi et. al [10] developed a new
Index Terms—HVDC system, VSC-HVDC, Wavelet analysis, fault locater, which determines reactance in a fault line, using
Neural networks, PSCAD, MATLAB.
single terminal voltage and current data, obtained using a
microprocessor. The algorithm is able to automatically correct
I. I NTRODUCTION load flow, fault resistance, and dissymmetry errors in the
transmission line. However, the fault locater is able to perform
AC transmission has limited transmission capacity and dis-
satisfactorily only for single faults in the transmission line, and
tance constraints. Besides that, two asynchronous AC networks
is unable to determine multi-fault locations simultaneously.
cannot be connected directly. To overcome these disadvan-
Most of the fault location methods are based on traveling wave
tages, there has been discussion to use HVDC as an alterna-
[11]. However, they require the detection of the wave arrival
tive. HVDC is used in bulk power transmission to minimize
time with high degree of precision, making them difficult to
losses and to improve power quality and load flow control.
be implemented. The problem is addressed by Song et. al’s
Compared to the conventional HVDC system, voltage source
method [12], in which the natural frequency is used as a
converter (VSC) based HVDC systems offers the advantages
criterion to locate fault.
such that the reactive power can be controlled independently
This paper is aimed to simulate, study and create a machine
without the need for reactive power compensation as in the
learning based algorithm to predict faults in the two terminal
classical HVDC link. Improvement in power electronics have
HVDC link. The two-terminal HVDC system is modeled
contributed immensely to the development of VSC based-
and pole-to-pole faults are simulated in PSCAD. The fault
HVDC system, eg. its high power application insulated-gate
result is reinterpreted in time-frequency domain using wavelet
bipolar transistor (IGBT) is used in the converters to convert
transform (WT). With the wavelet coefficient of the fault
AC into DC and vice-versa.
current signal as input data, an Artificial Neural Network
The fault location technology in HVDC system is in its
(ANN) is developed using MATLAB to predict the fault
adolescent stage. Numerous fault location technique in AC
location over the transmission line. The paper is organized as
system has been reviewed in [1]. The fault distance can be
follows. Section II presents the background of two-terminal
calculated based on voltage and current phasor estimation [2].
HVDC system modeling in PSCAD. The concepts of WT and
The method proposed by Novosel et. al [3] uses a lumped
ANN are briefly introduced in Section III and IV, respectively.
parameter model to represent all impedance behind the fault,
Section V presents the result of ANN processing the fault
applicable for short transmission lines. These information,
signal. Finally, conclusion is given in Section VI.
however, is not available to HVDC system. Due to the
absence of nominal frequency, it is impossible to compute
II. WAVELET ANALYSIS OF FAULT SIGNAL
The work was supported by the Start-Up Grant (M4081547.040), Nanyang
Technological University, Singapore. Similar to Fourier analysis, wavelet analysis is used to repre-
S. Vasanth and Y. M. Yeap are undergraduate and PhD student, respectively, sent the function in frequency domain. Unlike Fourier analysis
and A. Ukil is Assistant Professor in the School of Electrical & Electronic
Engineering, Nanyang Technological University (NTU), Singapore (Email: of signals which are localized only in frequency domain, the
[email protected], [email protected], [email protected]). wavelet analysis is able to extract the signal information in
 2

Fault Fault
signal feature
150 HVDC Signal Fault
ANN
Fault signal

100 system analysis location

50

1000 2000 3000 4000 5000 6000

Sample number
Wavelet: Detailed (freq) coefficients
Wavelet coefficient

0
Fig. 2. ANN process flow.
-1
50 100 150 200 250 300 350
Sample number VSC1 Idc12 Idc21 VSC2
Wavelet: Smoothed (time) coefficients AC1 T1 T2 AC2
Wavelet coefficient

600

400
200km
200
overhead line
50 100 150 200 250 300 350 Fig. 3. Simulated two-terminal HVDC system.
Sample number

prediction using ANN, which involves 3 layers consisting of

Fig. 1. Wavelet analysis of fault signal. 2 input nodes, 13 hidden nodes and 1 output node.
The proposed fault location method in this paper feeds the
XD coefficients generated from wavelet analysis to the ANN
both time and frequency domain, thereby allowing to track the
as the inputs. In the case of two-terminal system, the XD
change in the frequency components over time. Thus, wavelet
coefficients of line currents (Idc12 & Idc21 ) are of interest
analysis is useful in analyzing the functions with discontinuity
and necessary for the ANN to learn the fault features. The
or abrupt change [13].
algorithm developed to estimate the fault location in the
MATLAB is used to perform the wavelet analysis, with a
HVDC line uses train-br (Bayesian regularization) and 640
variety of functions that a WT can perform: signal denoising
training epochs, the ANN is commenced with 80% of the data
and data compression are of primary importance with respect
to be test data. It is in this stage where the ANN is given the
to the paper, the rise time of the fault current is a finer
two inputs and the corresponding fault location obtained from
parameter than the entire fault signal [13].
the simulation. In this stage, the ANN strives to establish a
Fig. 1 shows the result of a fault current signal being treated trend between the given inputs and the corresponding outputs.
with WT. The first plot is the fault signal generated from a After the training stage, the remaining 20% of the data is
simple simulation. When a signal is processed with WT, the used to test the ANN. In this stage, the generated output from
smoothed coefficient (XA) and the detailed coefficient (XD) the logic of the ANN is compared to the actual output obtained
are plotted and the first peak of XD coefficient is taken as from the simulation. The disparity in the actual values to the
an input to the neural network. In this case, each simulation predicted values is computed as an error vector.
contains 6002 data points in the time domain. Using the scale
of 4, the signal is downsampled by 24 , approximately yielding
IV. M ODELING AND SIMULATION
359 data points as a result. The spike observed in the plot of
XD suggests that there is high frequency component in the A two-terminal HVDC system, as depicted in Fig. 3 is
fault signal at that particular time. From the plot of XA, it can modeled with the technical specifications as indicated in Table
be seen that the signal resembles the original one implying I. A typical HVDC transmission system consists of an AC
successful reconstruction. source, a filter, transformer, converters and DC transmission
link.
Pole-to-pole faults are simulated on the aforementioned
III. A RTIFICIAL NEURAL NETWORK
200 km transmission line based on the parameters specified
Artificial Neural Networks (ANN) are identical to a cluster in Table I, the fault location is kept constant and the fault
of neurons in a human brain. The neural networks communi- resistance is varied from 0.01 Ω to 100 Ω, the corresponding
cates with each other in order to arrive at a decision. Networks currents Idc12 and Idc21 are plotted and the database is created
in ‘Neural Networks’ refer to linkage between neurons from using the values obtained in the simulation. The simulation is
different layers, the higher the intricacy of the problem, the performed on 13 different fault locations ranging from 25 km
higher is the requirement for number of hidden layers. Fig. to 175 km in steps of 12.5 km, this diverse range of data
2 shows a flow diagram of the steps associated with the data enables the ANN to learn and understand the dataset [14] and
 3

TABLE I
T ECHNICAL SPECIFICATIONS OF THE SIMULATED MODEL . : R=0.99532
System parameter Value
Data
Rated capacity 1200 MVA 0.4 Fit
Rated AC voltage 230 kV Y=T
Rated DC voltage 640 kV
Length of transmission 200 km

Output ~= 1*Target + -0.0015

Steady state frequency 60 Hz 0.2
Transformer ratio 230/370 kV
Switching frequency 1980 Hz
DC capacitance 5000 µF
0

analyze the trends developed with increasing fault resistance

or fault location. -0.2

V. R ESULT AND DISCUSSION

-0.4
The ANN is implemented to estimate the fault location
in the 200 km HVDC transmission line. Various statistical
parameters are used to measure the compatibility of ANN. -0.6
Table I summarizes the statistical error parameters obtained -0.6 -0.4 -0.2 0 0.2 0.4

from the ANN. Target

TABLE II
Fig. 4. Regression plot analysis.
S TATISTICAL E RROR PARAMETERS .

Error Value
M SE 20.5886 180
RM SE 4.5375
M AE 2.7544 actual values
160
M AP E 3.090 predicted values
MF E 0.0037 140
R 0.99532
R2 0.9906 120
Fault Location

100
As observed from Table II,
80
• The Mean Squared Error (M SE) is found to be 20.5886
km2 . 60

• The Root Mean Squared Error (RM SE) is found to be

40
4.5375 km which indicates that on an average there is
a 4.5375 km difference between the actual data and the 20

predicted data. 0
0 20 40 60 80 100 120 140 160 180 200
• The Mean Absolute Error (M AE) is found to be 2.7544
Number of Instances
km which indicates that on an absolute scale there is an
error of 2.7544 km predicted by the ANN. Fig. 5. A comparison between predicted and actual values.
• The Mean Forecast Error (M F E) is found to be 0.0037,
this value indicates that the ANN is over forecasting with
a factor of 0.0037. performed and the y-axis indicates the fault location in km. It
• The correlation coefficient (R) is found to be 0.99532 and can be observed that the ANN predicts with a higher accuracy
it indicates a good fit of prescribed values to the actual in the middle of the transmission line and it predicts at a lower
values. accuracy towards the sending end and the receiving end of the
2
• The coefficient of determination (R ) is found to be transmission line.
0.9906 and it also indicates a liner fit of the predicted
values to the actual values.
Fig. 4 shows the regression plot after running the ANN, R A. Comparative Analysis on identifying fault location at a
as mentioned above is the correlation coefficient, R value of constant fault resistance
1 indicates perfect correlation. In this case, R value is found The aim of this section is to compare and analyze the
to be 0.99532 indicating that the predicted values are in good accuracy of the fault location technique for different fault
fit with respect to the actual values. resistances and to identify the fault resistance at which the
Fig. 5 shows the deviation of the predicted values from the ANN is highly efficient. The fault resistance is held constant
actual values, the x-axis indicates the number of simulations and the fault location is identified using the ANN. Fig. 6 shows
 4

Compara ve analysis [5] M. Karthikeyan, Y. M. Yeap, and A. Ukil, “Simulation and analysis of
faults in high voltage dc (hvdc) power transmission,” In Proc. IECON
2014 - 40th Annual Conference of the IEEE Industrial Electronics
Society, Dallas, pp. 1786–1791, Oct 2014.
[6] Y. M. Yeap and A. Ukil, “Differentiation of fault and load change in
Correla on coe cient - R

hvdc system using amplitude tracking square wave,” In Proc. 2015 IEEE
Power Tech, Eindhoven, pp. 1–5, June 2015.
[7] Y. M. Yeap and A. Ukil, “Fault detection in hvdc system using short
time fourier transform,” In Proc. IEEE Power & Energy Society General
Meeting, Boston, July 2016.
[8] H. A. M. A. Hassan, “Fault Location Identificatoin For A VSC-HVDC
System With A Long Hybrid Transmission Medium,” Master’s thesis,
University of Pittsburgh, Pittsburgh, 2014.
[9] A. N. Pathirana, “Travelling Wave Based DC Line Fault Location
in VSC HVDC Systems,” Master’s thesis, University of Manitoba,
Winnipeg, 2012.
[10] T. Takagi, Y. Yamakoshi, M. Yamaura, R. Kondow, and T. Matsushima,
“Development of a new type fault locator using the one-terminal voltage
and current data,” IEEE Power Engineering Review, vol. PER-2, no. 8,
pp. 59–60, Aug 1982.
Fault resistance [11] P. Chen, B. Xu, and J. Li, “A traveling wave based fault locating system
for hvdc transmission lines,” In Proc. 2006 International Conference on
Power System Technology, Chongqing, pp. 1–4, Oct 2006.
Fig. 6. Comparing correlation coefficient at finding fault location.
[12] G. Song, X. Chu, X. Cai, S. Gao, and M. Ran, “A fault-location method
for vsc-hvdc transmission lines based on natural frequency of current,”
International Journal of Electrical Power & Energy Systems, vol. 63,
a plot of correlation coefficient of finding fault location at a pp. 347 – 352, 2014.
given fault resistance. [13] Y. M. Yeap and A. Ukil, “Wavelet based fault analysis in hvdc system,”
in In Proc. IECON 2014 - 40th Annual Conference of the IEEE Industrial
It can be observed that the ANN has a high correlation Electronics Society, Dallas, Oct 2014, pp. 2472–2478.
coefficient R at lower fault resistances, this can be attributed [14] M. M. Tawfik and M. M. Morcos, “Ann-based techniques for estimating
to the fact that the data set used to train the ANN has fault fault location on transmission lines using prony method,” IEEE Trans-
actions on Power Delivery, vol. 16, no. 2, pp. 219–224, Apr 2001.
location values corresponding to the lower fault location case.

VI. C ONCLUSION
In this paper the VSC-based HVDC system has been
examined. The two terminal VSC-based HVDC system is built
in PSCAD, in which the DC faults are simulated and the
corresponding fault currents are recorded. The experiment is
repeated for various values of fault resistances at various fault
locations along the 200 km transmission line.
The DC fault current data is processed with wavelet analysis
to generate the detailed coefficient, which is then used as input
to artificial neural network (ANN) to learn the fault pattern.
It is found that the ANN is able to locate the DC fault
with relatively high accuracy, as proven by high correlation
coefficient and low error. The study can be further extended
to multi-terminal network in which the influence of other
terminals can complicate the fault location. Reliability of the
suggested model can be verified by performing similar analysis
under different model parameters.

R EFERENCES
[1] M. M. Saha, R. Das, P. Verho, and D. Novosel, “Review of fault
location techniques for distribution systems,” In Proc. Power Systems
and Communications Infrastructures for the future, Beijing, September
2002.
[2] M. M. Saha, F. Provoost, and E. Rosolowski, “Fault location method for
mv cable network,” In Proc. Developments in Power System Protection,
2001, Seventh International Conference on (IEE), Amsterdam, pp. 323–
326, 2001.
[3] D. Novosel, D. Hart, Y. Hu, and J. Myllymaki, “System for locating
faults and estimating fault resistance in distribution networks with
tapped loads,” Nov. 17 1998, uS Patent 5,839,093. [Online]. Available:
http://www.google.com/patents/US5839093
[4] M. M. Xu, L. Y. Xiao, and H. F. Wang, “The state of art for fault loca-
tion in multi-terminal dc system,” Proc. Renewable Power Generation
Conference (RPG 2013), 2nd IET, Beijing, pp. 1–4, Sept 2013.

View publication stats