1

A Recursive Method for Traveling-Wave

Arrival-Time Detection in Power Systems
R. Jalilzadeh Hamidi, Member, IEEE, H. Livani, Member, IEEE

 detect ATs. If the selected wavelet is more similar to the

Abstract— This paper proposes a novel recursive method for arriving-wave shape, DWT will more accurately detect the AT
detecting the first arrival time (AT) of traveling waves (TWs) in [8]-[12]. Thus, researchers developed customized wavelets. In
power grids to enhance the fault-location methods relying on [13], the characteristic frequencies of a test system for
TWs. This method depends on the adaptive discrete Kalman
filter (ADKF). It estimates the parameters of a high time-
different fault locations are determined. Then, based on the
resolution voltage or current measurement and generates pre-fault analysis, the Morlet wavelet is amended for more
residuals (innovation sequence). Both measurement noises and accurate detection of ATs. In [14], wavelets are customized
TWs can result in an abrupt change in the residuals. The based on the time differences between sequential arriving
proposed method pinpoints the probable abrupt change and waves for enhancement of AT-detection accuracy. This
distinguishes whether it is caused by noises or arriving waves. As method customizes wavelets by solving an unconstrained
the proposed method is recursive, it is proper for implementation
in on-site digital fault locators for real-time applications.
optimization problem, relying on db wavelet to find the initial
For evaluation of the proposed method, EMTP-RV and the guess for the optimization problem. In [15], the wavelet
real-time digital simulator (RTDS) are utilized to perform the transform together with modulus maxima are used to identify
transient simulations. The results are then analyzed in the polarities of TWs. Then, the identified TW polarities are
MATLAB. The proposed method and the state-of-the-art AT- utilized for enhancing the detection and identification of
detection methods in the prior literature are compared, and the single-line-to-ground (SLG) faults in distribution systems.
sensitivity analysis demonstrates that the measurement noises As the selection of the mother wavelet largely influences
and fault parameters have less influence on the proposed method
efficiency in comparison to the existing AT-detection methods. the performance of the TW-based fault-location methods [16],
the other methods for AT detection are addressed in the
Index Terms— Abrupt change detection, fault location, literature.
transient detection, traveling-wave arrival time. In [17], Park’s transform is utilized for real-time AT
detection. In addition, it is improved to be less prone to power
system imbalances and measurement noises. However, Park’s
I. INTRODUCTION transform curtails modal decomposition of the measurements
using Clarke’s transform since it requires three-phase
A FTER the occurrence of a fault, it is necessary for electric
utilities to locate the fault and repair the damage for
restoration of service. It is well-established that the fault-
measurements.
In [18] and [19], the application of Hilbert transform (HT)
location methods relying on traveling waves (TWs) are in the detection of ATs is addressed. The performance of HT
superior to the impedance-based methods, especially for cross- in AT detection is investigated in [20] and [21], and it is
linked PolyEthylene (XLPE) insulated or cross-bonded cables concluded that HT is to some extent more efficient compared
[1]-[3]. to DWT. S-transform is utilized for AT detection in [22] and
A considerable number of TW-based fault-location methods [23], and it is assessed that S-transform is capable of detecting
rely on the arrival times (ATs) of the fault-induced TWs to ATs more efficiently in comparison with DWT.
estimate the fault location. In this regard, several methods are In [24], the application of short-time matrix pencil method
proposed for AT detection in the literature. In [4] and [5], a (STMPM) in AT detection is proposed. The arriving waves in
method for AT detection is proposed based on the low signal-to-noise ratio (SNR) measurements can be detected
differentiator-smoother method. The authors of [6] addressed utilizing STMPM. However, STMPM pinpoints the ATs
the application of sliding discrete Fourier transform (SDFT) in through matrix algebra, requiring high processing ability.
AT detection. However, SDFT is not able to detect ATs with Therefore, its implementation in on-site digital fault locators
adequate time resolution that increases the error in fault seems to be challenging.
location. After the introduction of discrete wavelet transform The abovementioned AT-detection methods either are
(DWT) for AT detection in [7], various mother wavelets such sensitive to the influencing parameters (e.g., noises) or require
as Haar and Daubechies (db) at various scales are used to high computational capacity. For improving these
shortcomings, this paper proposes an AT-detection method
with the following advantages: i) It is appropriate for on-site
Reza Jalilzadeh Hamidi is with the EE Dept. of Arkansas Tech University microprocessor-based fault locators, considering that the
and Hanif Livani is with EBME department of the University of Nevada, proposed method is computationally efficient as it is fully
Reno (emails: [email protected], [email protected]).
 2

recursive. ii) In comparison with the existing methods, the

proposed method is relatively robust against measurement
noises. iii) Fault parameters and power system specifications
change the shape of the fault-induced arriving waves.
Therefore, pre-assessment of fault-induced TWs is necessary
for DWT-based methods in contrast to the proposed method.
EMTP-RV and the real-time digital simulator (RTDS) are
utilized to perform transient simulations. Then, the results are
processed using MATLAB. The performance of the proposed
and the existing methods (e.g., DWT, HT, and STMPM) are
compared in terms of the following influencing parameters:
measurement noises, fault inception angle (FIA), sampling
frequency (𝐹𝑠 ), and non-ideal faults (NIF) [25]. The simulation
results demonstrate that the proposed method is less
susceptible to the influencing parameters. Therefore, the
performance of the TW-based fault-location methods
depending on the first ATs will improve if the proposed
method is utilized as an alternative to the existing AT-
detection methods. Fig. 1. Typical aerial mode of three-phase high-frequency measurements. The
The organization of this paper is as follows: in Section II, voltage level is 230 kV and the sampling frequency is 200 kHz. (a) Voltage
the method is described. In Section III, two test cases, one measurement. (b) Current measurement.
illustrative example, and sensitivity analysis are provided. In
Sections IV and V, comparison and conclusion are presented, a variance of 𝜎 2 .
respectively.
2. Parameter Estimation
The ADKF is utilized for estimating the parameters of AC
II. METHODOLOGY
voltage or current waveforms in pre-fault time. As the power
Referring to Figs. 1(a) and 1(b), after the occurrence of a system voltages and currents may consist of harmonics, (1)
fault, abrupt changes (also referred to as transients) in the represents the true values of the measurements.
related voltage and current measurements arise due to the 𝑁ℎ
fault-induced arriving waves. Then, the corresponding 𝑧𝑘𝑡𝑟𝑢𝑒 = ∑ 𝐴ℎ sin(𝑘ℎ𝜔0 𝑇𝑠 + 𝜙ℎ ) (1)
ℎ=1
voltages and currents start drifting from their normal
where 𝐴ℎ and 𝜙ℎ are respectively the amplitude and phase
trajectories. The proposed method objective is to detect the
angle of the ℎ-th harmonic, 𝑘 is the sample number, 𝜔0 [rad/s]
first AT of the TWs. In this regard, the aerial mode of high-
frequency measurements are continuously fed to the proposed is the power system angular frequency, 𝑇𝑠 [s] is the sampling
method. Before the occurrence of a fault, the proposed method time-step, and 𝑁ℎ represents the number of the existing
estimates the measurement parameters and noise variance harmonics in the power system. (1) is extended as
through the adaptive discrete Kalman filter (ADKF). In pre- 𝑁ℎ
fault duration, the ADKF generates residuals (the differences 𝑧𝑘𝑡𝑟𝑢𝑒 = ∑ 𝐴ℎ cos(𝜙ℎ ) sin(𝑘ℎ𝜔0 𝑇𝑠 )
ℎ=1 (2)
between the measured and estimated samples), which + 𝐴ℎ sin(𝜙ℎ ) cos(𝑘ℎ𝜔0 𝑇𝑠 )
comprise measurement noises. Therefore, the residuals are
Applying the definitions in (3) to (2), yields the state-
Gaussian. However, after the occurrence of a fault, when
arriving waves reach the line ends, the residual sizes become variable representation of 𝑧𝑘𝑡𝑟𝑢𝑒 as provided in (4), [26] and
large, resulting in an abrupt change in the residual sequence. [27],
The proposed method first detects the abrupt change. As an 𝑥 2ℎ−1 = 𝐴ℎ cos(𝜙ℎ )
abrupt change can be caused by noises or arriving waves, the 𝑥 2ℎ = 𝐴ℎ sin(𝜙ℎ )
proposed method then distinguishes if the abrupt change is 2ℎ−1 , ∀ℎ = 1,2, … , 𝑁ℎ (3)
produced by the arriving waves or noises. Fig. 2 shows the 𝑎𝑘 = sin(𝑘ℎ𝜔0 𝑇𝑠 )
2ℎ
proposed method flowchart, and its steps are described as { 𝑎𝑘 = cos(𝑘ℎ𝜔0 𝑇𝑠 )
follows, where 𝑥 denotes the states (2𝑁ℎ states are required for a signal
1. Measurement of One New Sample with 𝑁ℎ harmonics) and 𝑎 indicates the coefficients.
𝑁ℎ
During pre-fault time, measured samples consist of two
𝑧𝑘𝑡𝑟𝑢𝑒 = ∑ 𝑎𝑘2ℎ−1 𝑥 2ℎ−1 + 𝑎𝑘2ℎ 𝑥 2ℎ (4)
parts that can be shown in the form of 𝑧̃𝑘 = 𝑧𝑘𝑡𝑟𝑢𝑒 + 𝛾𝑘 , where ℎ=1
𝑧̃𝑘 is the 𝑘-th measured sample, 𝑧𝑘𝑡𝑟𝑢𝑒 is the true value of the and (4) can be shown in matrix form as
sample, and 𝛾𝑘 ~𝒩(0, 𝜎 2 ) is a zero-mean Gaussian noise with 𝑧𝑘𝑡𝑟𝑢𝑒 = 𝑯𝐾 𝒙𝑘 (5)
2𝑁
where 𝑯𝐾 = [𝑎𝑘1 , 𝑎𝑘2 , … , 𝑎𝑘 ℎ ] is the measurement coefficient
vector, 𝒙𝑘 = [𝑥 1 , 𝑥 2 , … , 𝑥 2𝑁ℎ ]𝑇 is the measurement state
 3

vector, and (. )𝑇 indicates matrix transposition. The state-space

representation of the measurement is
𝒙 = 𝝓𝑘 𝒙𝑘 + 𝒘𝑘
{ 𝑘+1 (6)
𝑧̃𝑘 = 𝑯𝑘 𝒙𝑘 + 𝛾𝑘
where 𝝓 is the state transition matrix, 𝒘 is the process noise
vector, and 𝛾 is the measurement noise. The algorithm of the
ADKF is as follows, [28],
−1
𝑮𝑘 = 𝑷− 𝑇 − 𝑇 ̂
𝑘 𝑯𝑘 (𝑯𝑘 𝑷𝑘 𝑯𝑘 + 𝑅𝑘−1 )
̂𝑘 = 𝒙
𝒙 ̂− ̂−
𝑘 + 𝑮𝑘 (𝑧̃𝑘 − 𝑯𝑘 𝒙𝑘)
−
𝑷𝑘 = (𝑰 − 𝑮𝑘 𝑯𝑘 )𝑷𝑘 (7)
−
̂
𝒙𝑘+1 = 𝝓𝑘 𝒙𝑘̂
{ 𝑷− 𝑇
𝑘+1 = 𝝓𝑘 𝑷𝑘 𝝓𝑘 + 𝑸
where 𝑮, 𝑰, 𝑷, and 𝑸 are respectively the gain, identity, error
covariance, and process noise covariance matrices. (. )−
denotes prediction. It is realistic to assume 𝝓𝑘 = 𝑰, ∀𝑘 =
1, 2, 3 …, considering the high sampling rate and slow-
changing power system parameters [26]. 𝑅̂ is the estimated
measurement noise variance. It is scalar since the ADKF
successively receives the measured samples (i.e., one sample
of the measurement aerial mode at any given time).
3. Buffering Time
Since the ADKF gradually updates the states (𝑥) by
receiving successive samples, a short period (buffering time) Fig. 2. The flowchart of the proposed method.
is considered for the ADKF results to become valid.
For detection of the abrupt change and identification of its
4. Measurement Noise Estimation source (i.e., noises or arriving waves), normalization of the
It is necessary to predict the measurement noise variance for residuals is necessary. Therefore, the average of residuals is
the discrete Kalman filter (DKF). However, environmental required. In this regard, (11) is used for recursively calculating
noises and measuring device accuracy are inconsistent [29], the average of the residuals after receiving each sample.
reducing the DKF performance. Conversely, the ADKF (Appendix A provides formulations resulting in (11)).
estimates the measurement noise variance that leads to more
accurate results. The measurement noise variance is updated 𝑟̅𝑘 = 𝜆𝑟̅𝑘−1 + (1 − 𝜆)𝑟𝑘 , 𝑟̅0 = 0 (11)
as [28] where 𝜆 ∈ (0,1) is the forgetting factor.
𝑅̂𝑘 = 𝑅0 , ∀𝑘𝑇𝑠 ≤ 𝑏 (8.a) 6. Normalization of Residuals
{ 1 𝑚 The normalized residuals are required for the proposed
𝑅̂𝑘 = ∑ 𝑟𝑘−𝑖 2
− 𝑯 𝑘 𝑷− 𝑇
𝑘 𝑯𝑘 , ∀𝑘𝑇𝑠 > 𝑏 (8.b)
𝑚 𝑖=1
method to detect the abrupt change. The residuals are
where 𝑅0 is the initial value of the measurement noise Gaussian random variables (RVs), therefore, they are
variance, 𝑏 [s] is the buffering time, 𝑚 is the length of the normalized as follows,
residual history required for updating the measurement noise, 𝑟𝑘𝑁 = (𝑟𝑘 − 𝑟̅𝑘 )/𝑅̂𝑘 (12)
and 𝑟 is the residual, which is defined as where 𝑟𝑘𝑁 is the normalized value of the 𝑘-th residual.
𝑟𝑘 = 𝑧̃𝑘 − 𝑯𝑘 𝒙̂𝑘 (9)
7. Detection of an Abrupt Change in the Residual Sequence
A digital processor requires saving a long history of 𝑚 A transient in the high-frequency measurements leads to an
residuals for calculation of (8.b). Moreover, the digital abrupt change in the residuals (𝑟𝑘 ,), and consequently, in the
processor requires running the “add” operation 𝑚 − 1 times to normalized residuals (𝑟𝑘𝑁 ). For detection of the probable
calculate the summation of 𝑚 residuals in (8.b), which is time abrupt change in the normalized residuals, the two-sided
consuming. Therefore, we reformulated (8.b) to calculate the exponentially weighted moving average (EWMA) is selected
estimated measurement noise variance (𝑅̂𝑘 ) in a recursive as a high-gain control chart. The EWMA is capable of
manner as given in (10). The formulations resulting in (10) are pinpointing the abrupt change, while it is relatively less prone
provided in Appendix A. to release false alarms due to the measurement noises [30] and
𝑅̂𝑘 = 𝛬(𝑅̂𝑘−1 + 𝑯𝑘−1 𝑷− 𝑇 2
𝑘−1 𝑯𝑘−1 ) + (1 − 𝛬)𝑟𝑘−1
[31]. The two-sided EWMA is formulated as
(10)
− 𝑯 𝑘 𝑷−
𝑘 𝑯𝑘
𝑇
𝑔𝑘 = max{𝜓𝑔𝑘−1 + (1 − 𝜓)𝑟𝑘𝑁 , 0} , 𝑔𝑘 > 𝑇 → 𝑎𝑙𝑎𝑟𝑚
{ (13)
where Λ ∈ (0,1) is the forgetting factor. 𝑔𝑘′ = max{𝜓𝑔𝑘−1
′
− (1 − 𝜓)𝑟𝑘𝑁 , 0} , 𝑔𝑘′ > 𝑇 → 𝑎𝑙𝑎𝑟𝑚
5. Recursive Average of Residuals where 𝑔𝑘 and 𝑔𝑘′ (with an initial values of zero, 𝑔0 , 𝑔0′ = 0)
are the outputs of the EWMA to rising and falling abrupt
 4

changes, respectively. 𝜓 ∈ (0,1) is the control-chart factor,

and 𝑇 is the threshold, which is selected based on the
cumulative distribution function (CDF) of the normal
Gaussian probability density function (PDF). The selection of
the threshold based on the desired confidence level is
described in Appendix A.
Fig. 3. One-line diagram of Test Case 1.
8. Tentative Arrival Time
TABLE I
Smaller values for 𝜓 and 𝑇 increase the gain of the EWMA THE HARMONIC CONTENTS OF THE VOLTAGE MEASUREMENT
control chart, enabling it to pinpoint smaller abrupt changes in Vh2 Vh3 Vh5 Vh7 Vh9 Vh11
normalized residuals (𝑟𝑘𝑁 ). However, high-gain (short- Harmonic Level [%]
0.2 1 0.5 0.5 0.5 0.5
window) control charts are prone to release false alarms due to 𝜙ℎ2 𝜙ℎ3 𝜙ℎ5 𝜙ℎ7 𝜙ℎ9 𝜙ℎ11
Harmonic Phase [⁰]
measurement noises. Accordingly, it is essential to confirm 2 5 3 3 3 3
that the source of the detected abrupt change is arriving waves. 𝑁
and the elements of 𝑟𝑝𝑟𝑒 are no longer normal Gaussian.
Therefore, the AT corresponding to the detected abrupt change
Hence, the Chi-square testing technique is applicable to check
is considered a tentative AT (𝑇𝐴𝑇 = 𝑘𝑇𝑠 ), and its validity
the goodness of the fit between the during-fault measured
should be verified. For this purpose, low-gain (long-window) −
filters can be utilized, which are robust against measurement (𝑧̃𝑘+𝑖 ) and the predicted (𝑧̂𝑘+𝑖 ) samples as
𝑛
𝑁 2
noises. However, they fail to accurately pinpoint abrupt 𝑌=∑ (𝑟𝑘+𝑖 ) (16)
changes (i.e., the time resolution of a low-gain filter is low). If 𝑖=1
the abrupt change is a result of noises, then the long-window where 𝑌 is an RV with Chi-square distribution with (𝑛 − 1)
filter does not flag. However, in case of arriving waves, the degrees of freedom.
𝑦𝑡
long-window filter flags. To be able to use long-window 2
∫ 𝜒𝑛−1 (𝜁) 𝑑𝜁 = 𝐶𝐿 (17)
filters, a sufficient set of normalized residuals (𝑟𝑘𝑁 ) is required. 0
2
Therefore, Steps 9 and 10 are considered. where 𝜒𝑛−1 is the Chi-square PDF with (𝑛 − 1) degrees of
9. Signal Prediction freedom, 𝜁 is the integration variable, 𝐶𝐿 is the confidence
As low-gain filters require a set of samples, the next 𝑛 level, and 𝑦𝑡 is the largest acceptable value for 𝑌.
samples after the detection of the abrupt change are predicted If 𝑌 > 𝑦𝑡 , the elements of the normalized prediction-based
𝑁
based on the last estimated states (𝒙 ̂𝑘 ) as residual set (𝑟𝑝𝑟𝑒 ) are not normal Gaussian, implying that 𝑒 ≫
−
𝑍𝑝𝑟𝑒 = {𝑧̂𝑘+1 −
, 𝑧̂𝑘+2 −
, … , 𝑧̂𝑘+𝑛 } (14) 0. Therefore, a fault occurred, and 𝐴𝑇 = 𝑇𝐴𝑇.
where 𝑍𝑝𝑟𝑒 is the set of predicted samples. 𝑧̂𝑘+𝑖 −
= If 𝑌 ≤ 𝑦𝑡 , the elements of the normalized prediction-based
𝑁
residual set (𝑟𝑝𝑟𝑒 ) are normal Gaussian, implying that 𝑒 ≈ 0.
̂𝑘 , ∀𝑖 = 1,2, … , 𝑛. The size of the set, 𝑛, should be
𝑯𝑘+𝑖 𝒙
properly selected based on the low-gain filter, which is Therefore, the EWMA released a false alarm. Thus, the
described in Step 11. EWMA resets and the procedure continues from Step 1.

10. Normalized Prediction-Based Residuals III. CASE STUDY AND DISCUSSION

It takes time for protection systems to disconnect a faulty
transmission line (TL). Therefore, the transient recorders A. Test Case 1
(TRs) at the ends of the TL can measure the during-fault The double-ended TW-based fault-location method in [7] is
voltages or currents before the protection system clears the adopted to validate the proposed method. However, the
fault. Hence, a set of normalized residuals based on the proposed method is applicable to a wide range of TW-based
−
predicted (𝑧̂𝑘+𝑖 ) and the during-fault measured samples (𝑧̃𝑘+𝑖 ) fault-location methods utilizing the first ATs (e.g., [2], [8],
are generated as and [10]).
𝑁 𝑁 𝑁 𝑁 Test Case 1 is shown in Fig. 3, and its details are given in
𝑟𝑝𝑟𝑒 = {𝑟𝑘+1 , 𝑟𝑘+2 , … , 𝑟𝑘+𝑛 } (15)
𝑁 Appendix B based on [32]. The TL is a transposed line with
where 𝑟𝑝𝑟𝑒 is the normalized prediction-based residual set,
the length and voltage level of 100 mi and 350 kV,
𝑁
𝑟𝑘+𝑖 = (𝑟𝑘+𝑖 − 𝑟̅𝑘 )/𝑅̂𝑘 , and the prediction-based residuals are respectively. The frequency-dependent TL model is utilized in
−
defined as 𝑟𝑘+𝑖 = (𝑧̃𝑘+𝑖 − 𝑧̂𝑘+𝑖 ), ∀𝑖 = 1, 2, … , 𝑛. simulations. It is assumed that the voltage contains the
11. Verification of the Tentative AT (TAT) harmonics given in Table I. Both ends of the TL are equipped
The normalized prediction-based residuals are in the form with TRs. Electronic voltage transformers, such as optical
𝑁
of 𝑟𝑘+𝑖 = 𝑣𝑘+𝑖 + 𝑒𝑘+𝑖 , where 𝑣𝑘+𝑖 ~𝒩(0,1) is the VT/RCVT and C-divider, are able to measure the voltage at
measurement noise and 𝑒𝑘+𝑖 is the error in the (𝑘 + 𝑖)-th high sampling-rates up to 2 MHz [33]. The applications of
normalized residual. Referring to Figs. 1(a) and 1(b), before optical voltage sensors in electric power systems are addressed
the occurrence of a fault (before arriving waves), 𝑒𝑘 ≈ 0, and in [34] and [35], and it is concluded that high sampling-rate
𝑁
the elements of 𝑟𝑝𝑟𝑒 are normal Gaussian RVs. However, voltage measurements are practically feasible. In this paper,
when a fault occurs, the measurements start drifting from their similar to [36], we assumed that the measuring devices are
normal trajectory. Therefore, 𝑒 starts becoming large, 𝑒𝑘 ≫ 0, ideal to focus on the proposed method performance. GPS
 5

signal receivers are utilized for time stamping of the high

time-resolution measurements.
The aerial-mode TW velocity is 183,486 mi/s, which is
equivalent to a traveling-time (𝜏) of 545 μs. The two-ended
method for locating the fault is [7]
𝐸𝐹𝐿 = (𝐿 − 𝑣 Δ𝑡)/2 (18)
where 𝐸𝐹𝐿 is the estimated fault location in miles, 𝐿 [mi] is
the TL length, Δ𝑡 = 𝑡2 − 𝑡1 [s] and 𝑡1 and 𝑡2 are respectively
the ATs at Bus 1 and Bus 2. 𝑣 [mi/s] is the TW velocity. The
error in fault-location is calculated as
𝑒𝑟𝑟% = |𝑅𝐹𝐿 − 𝐸𝐹𝐿|/𝐿 × 100 (19)
where 𝑒𝑟𝑟 [%] is the fault-location error, and RFL is the real
fault location in miles.

1) Illustrative Example
The following parameters are selected for the proposed
method: 𝑅0 = 0.005 × 𝑉𝐴 where 𝑉𝐴 is the voltage amplitude
(285.774 kV), 𝑞 = 1, 𝑏 = 20 ms (i.e., the first 4000 samples
as 𝐹𝑠 = 200 kHz), Λ = 0.99, 𝜆 = 0.99, 𝜓 = 0.1. 𝑇 = 1.488 is Fig. 4. SLG fault with 𝑍𝑓 = 10 Ω at 50 mi from Bus 1 at 𝑡 = 250 ms. (a) The
selected for a confidence level of 95%. 𝑛 = 10 is selected for measured voltage and its estimated value in the aerial mode. (b) The
the Chi-Square test. An SLG fault with a fault impedance of normalized residuals during the buffering time. (c) The estimated and
measured voltages before the occurrence of the fault. (d) The normalized
10 Ω (𝑍𝑓 = 10 Ω) occurs 50 mi away from Bus 1 at 𝑡 = 250 residuals after the buffering time. (e) and (f) The fault-induced arriving wave
ms. The voltage measurements contain zero-mean Gaussian and its corresponding normalized residuals. (g) and (h) The entire
noises with 𝜎 2 = 0.01 × 𝑉𝐴 . The aerial mode of the measured measurement and normalized residuals during the procedure.
voltages (using Clarke’s transform) is fed to the proposed
method.
Fig. 4(a) shows the measured (𝑧̃𝑘 ) and estimated
measurements (𝑧̂𝑘 ) in the beginning of the procedure during
the buffering time (𝑡 ≤ 20 ms). The estimated voltage quickly
converge to the measured voltage since the sampling rate is
high and the states update fast. Fig. 4(b) shows the normalized
residuals (𝑟𝑘𝑁 ) during the buffering time. It is noticed that the
normalized residuals become smaller as the states are updated.
After the buffering time, the estimated voltage closely follows
the measured voltage and the normalized residuals deviate
around zero, corresponding to a zero-mean Gaussian noise, as
shown in Figs. 4(c) and 4(d). The fault-induced arriving waves
reach Bus 1 at 250.275 ms as Fig. 4(e) depicts. As Fig. 4(f)
shows, at 𝑡 = 250.275 ms, the normalized residual caused by
the arriving wave becomes large. Therefore, the EWMA
determines 𝑇𝐴𝑇 = 250.275 ms. Then, the proposed method
predicts the next 10 samples for times
250.280, 250.285, … , 250.325 ms. After measurement of the
next 10 samples, the normalized prediction-based residual set
𝑁 Fig. 5. Fault-location errors related to an SLG fault 50 mi away from Bus 1,
(𝑟𝑝𝑟𝑒 ) is generated. Chi-square testing technique determines
𝑍𝑓 = 10 Ω, 𝐹𝐼𝐴 = 150 , and 𝐹𝑆 = 200 kHz. (a) 𝜎 2 = 0.0075𝑉𝑝 . (b) 𝜎 2 =
that the elements of the normalized prediction-based residual 0.01𝑉𝑝 . (c) 𝜎 2 = 0.0125𝑉𝑝 .
𝑁
set (𝑟𝑝𝑟𝑒 ) do not belong to the normal Gaussian noise.
Therefore, the low-gain filter verifies that the abrupt change is proposed method with the existing AT-detection methods
produced by an arriving wave. Therefore, 𝐴𝑇 = 𝑇𝐴𝑇 that based on Test Case 1.
equals 250.275 ms. Figs. 4(g) and 4(h) illustrate the entire
procedure. In Fig. 4(h), the double-head arrow indicates that 1) Noise Sensitivity
as the ADKF updates the measurement noise variance, the Fault-induced TWs mitigate along TLs. Also, the
normalized residuals become smaller, improving the proposed environmental and measuring device noises skew the
method efficiency. measurements. Therefore, it is preferable that AT-detection
methods be able to accurately detect arriving waves in low-
B. Sensitivity Analysis Based on Test Case 1 SNR measurements. For evaluation of the proposed method, it
The following subsections evaluate and compare the is assumed that an SLG fault takes place 50 mi away from Bus
 6

TABLE II
FIA INFLUENCE OVER FAULT LOCATION ERROR
Transient FIA [⁰]
RFL
Detection 5 6 7 8
[mi]
Method Error in Fault-Location [%]
1 F(†) F F F
DWT, db-4,
5 F F F F
scale 1
50 F F F 1.120
1 F F 1.021 1.060
DWT, db-4,
5 F 1.109 1.015 1.001
scale 2
50 F 0.388 0.382 0.381
1 F F F F Fig. 7. The effect of fault-impedance on the arriving waves at Bus 1, 𝑅𝐹𝐿 =
DWT, Haar 5 F F F 0.601 10 mi from Bus 1, 𝐹𝑠 = 200 kHz, 𝑍𝑓 = 𝑅𝑓 + 𝑗𝑋𝑓 = 2 Ω, and 𝐹𝐼𝐴 = 90⁰.
50 F F 0.820 0.433
1 F F 0.227 0.214
HT 5 F 0.188 0.168 0.160
50 F 0.074 0.071 0.066
1 0.960 0.420 0.404 0.395
STMPM 5 0.528 0.418 0.392 0.367
50 0.271 0.216 0.201 0.194
1 F 0.455 0.404 0.395
The Proposed
5 0.414 0.391 0.391 0.370
Method
50 0.223 0.188 0.176 0.191
(†)
F means that the method fails to detect the AT.
Fig. 8. The effect of fault-impedance on fault-location errors.

Table II. The fault location is in the middle of the line, 𝑍𝑓 = 5

Ω, 𝐹𝑆 = 200 kHz, and the additive noise is 𝒩(0,0.01𝑉𝐴 ).
According to the results given in Table II, STMPM is able to
locate the faults with 𝐹𝐼𝐴 ≥ 5⁰ and the proposed method can
find the location of the faults with 𝐹𝐼𝐴 ≥ 6⁰. However, it is
worth noting that in the case of a larger attenuation factor
(attenuation factor is equal to 𝑒 −𝛼𝐿 in which 𝛼 is the
attenuation constant [mi-1] and 𝐿 is the TL length [mi], [37]) or
more severe noise-contaminated measurements, a fault with a
larger FIA can be located.

3) Sampling Frequency (𝐹𝑆 )

Fig. 6. Fault location errors with respect to the sampling frequency. (a) The
fault is 5 mi away from Bus 1 (𝑅𝐹𝐿 = 5 mi). (b) The fault is 10 mi away from
A higher sampling frequency results in a higher time-
Bus 1 (𝑅𝐹𝐿 = 10 mi). resolution. Therefore, ATs can be detected more precisely.
However, the attenuation factors corresponding to higher
1, 𝐹𝑆 = 200 kHz and the following parameters are selected to
frequency contents of TWs increase [37], especially for cables
provide low-SNR fault-induced arriving waves: 𝑍𝑓 = 10 Ω
[2] and [24]. Accordingly, the efficiency of AT-detection
and 𝐹𝐼𝐴 = 15⁰. The statistical analyses of the fault-location methods with respect to the sampling frequency is studied and
errors based on one thousand Monte Carlo simulations the results are provided in Fig. 6. Two fault locations (i.e., 5
(MCSs) are shown in Fig. 5. The fault-location errors for the and 10 mi from Bus 1) are considered with 𝑍𝑓 = 5 Ω, 𝐹𝐼𝐴 =
additive Gaussian noises with 𝜎 2 = 0.0075𝑉𝐴 , 𝜎 2 = 0.01𝑉𝐴 ,
90⁰, and the additional noise is 𝒩(0,0.01𝑉𝐴 ). It is noticed that
and 𝜎 2 = 0.0125𝑉𝐴 are shown in Figs. 5(a), 5(b) and 5(c), the sampling frequency drastically detracts from the efficiency
respectively. It is noticed that the proposed and STMPM of DWT with db-4 at scale 2 and HT. The performance of the
methods are able to detect the arriving waves regardless of proposed method and db-4 at scale 1 to some extent decreases
their relatively low SNRs. However, DWT using db-4 at scale as the sampling frequency decreases. However, sampling
1 cannot accurately detect the arriving waves as Fig. 5 depicts. frequency has a minor effect on STMPM and Haar.
DWT with db-4 at scale 2 is able to detect the arriving waves
with a noise variance less than 0.01𝑉𝐴 as Fig. 5 shows. DWT 4) Non-Ideal Faults (NIF)
with Harr as well as HT can detect the ATs for a noise The fault impedance (shown with 𝑍𝑓 in Fig. 3) with
variance less than 0.0125𝑉𝐴 . different 𝑋𝑓 /𝑅𝑓 ratios generates fault-induced TWs with
different shapes, affecting AT-detection methods [2] and [24].
2) Fault Inception Angle (FIA)
As the faults with lower FIAs generate smaller TWs, it is An SLG fault occurs at 20 mi from Bus 1 with 𝐹𝐼𝐴 = 90⁰,
more probable for fault-induced TWs to fade away in noisy 𝐹𝑆 = 200 kHz, and the additional noise is 𝒩(0,0.01𝑉𝐴 ). The
measurements. Thus, the proposed method is compared to the absolute value of the fault impedance is constant (𝑍𝑓 = 𝑅𝑓 +
existing methods concerning FIA, and the results are given in 𝑗𝑋𝑓 = 2Ω), however, the following values are considered for
 7

𝑋𝑓 /𝑅𝑓 ratios: 0.05, 0.1, and 0.2. Fig. 7 shows the arriving-
wave shapes at the Bus 1. It is noticed that the arriving-wave
shapes are slightly dissimilar; therefore, AT-detection methods
estimate different ATs. One thousand MCSs are run for each
𝑋𝑓 /𝑅𝑓 ratio. The fault-location errors of different AT-
detection methods are shown in Fig. 8. It is noticed that db-4
at scale 1 and to a lesser extent at scale 2 are influenced by
NIFs. However, NIFs have a relatively low impact on Haar
and HT. The proposed method and STMPM are robust against
NIFs.
C. Test Case 2: Real-Time Digital Simulator (RTDS)
Implementation
The schematic of Test Case 2 is shown in Fig. 9(a). The line
topology is based on [36], and its details are provided in Fig. 9. (a) The schematic of the test system in RTDS. (b) The schematic of
the hardware for converting the analog measurements to digital and sending
Appendix B. The TL is transposed and its length is 200 mi.
them to Matlab.
The Bergeron line model, which is frequency dependent, is
utilized in RTDS simulations. The system voltage and
frequency are respectively 230 kV and 60 Hz. The voltage
contains the harmonics given in Table I. The TW velocity is
178,430 mi/s [36].
The simulations are performed using RTDS with a digital-
to-analog (D/A)-in-the-loop. Fig. 9(b) shows the schematic of
the used hardware. The three-phase voltage measurements at
both ends of the TL (shown in Fig. 9(a) as “Measurement 1”
and “Measurement 2”) are sent to the D/A of the RTDS. The
D/A produces analog voltages based on the simulation results
at 1 MHz. The output of the D/A is then connected to a data
acquisition (DAQ), “IOTech DAQ/3000”, to provide time-
stamped digital samples at a rate up to 1 MHz [38]. The
digitalized measurements are then sent to a computer and
processed in MATLAB. Fig. 10. (a) and (b) show the arriving waves at Buses 1 and 2, respectively
(𝐹𝑠 = 200 kHz). (c) and (d) show the corresponding normalized residuals and
Figs. 10(a) and (b) respectively show the measurements at ATs.
Buses 1 and 2 at a sampling rate of 200 kHz for a solid SLG
fault at 75 miles from Bus 1. Figs. 10(c) and (d) show the compared and the results are given in Table IV. The
normalized residuals corresponding to the measurements at performance of DWT using db-4 at scale 1 is not preferable
Buses 1 and 2, respectively. The proposed method detects the against all the influencing parameters except for 𝐹𝑠 . However,
ATs at Bus 1 as 𝑡1 = 74.945 ms, and at Bus 2 as 𝑡2 = 75.220 DWT using db-4 at scale 2 is comparatively robust against
ms. Therefore, with reference to (18) and (19), Δ𝑡 = 275 µs, FIA. The efficiency of Haar deteriorates against low FIAs.
and the estimated fault location is 𝐸𝐹𝐿 = 75.465 mi, and However, Haar efficiency does not deteriorate with lower
𝑒𝑟𝑟 = 0.23%. The measurement noise variances estimated by sampling frequencies. The errors in the estimated fault-
the ADKF are almost 2080 V (i.e., 0.011VA ) and 575 V (i.e., locations using HT are smaller compared to DWT. However,
0.003𝑉𝐴 ) at Buses 1 and 2, respectively, which are due to the HT results in inefficient outcomes in the case of low-sampling
RTDS D/A, DAQ system, and the connections. measurements. None of the influencing parameters affects
Several experiments for the DAQ sampling frequencies of STMPM. The proposed method results in reliable outcomes
200 and 500 kHz and different fault locations are carried out. against the influencing parameters, except for 𝐹𝑠 .
The average fault-location errors based on one hundred The proposed method is also compared to STMPM in terms
experiments for each fault location and 𝐹𝑠 are provided in of computational burden since STMPM is considerably
Table III. The average fault-location errors corresponding to capable of detecting ATs, especially for low-SNR and low-𝐹𝑆
measurements. The CPU and RAM of the computer are Intel
the higher 𝐹𝑠 are smaller. In addition, it is noticed that the
Xeon E5420 and 32 GB, respectively. In both the proposed
average fault-location errors related to the faults closer to Bus
and STMPM methods, calculation of the sine and cosine
1 are larger.
functions are fulfilled through lookup tables, and divisions by
constants are replaced with multiplication (e.g., ÷ 4 is
IV. COMPARISON replaced with × 0.25) since it decreases the computation time.
Based on the simulation outcomes, the efficiency of the The measurement shown in Fig. 4(g) for 0 ≤ 𝑡 ≤ 100 ms is
existing and proposed AT-detection methods are qualitatively considered. As 𝐹𝑠 = 200 kHz, 100 ms is equal to 20,000
 8

𝐴𝑘−1 can be written as

𝐴𝑘−1 = 𝑚(𝑅̂𝑘−1 + 𝐵𝑘−1 )
TABLE III
AVERAGE FAULT-LOCATION ERRORS IN TEST CASE 2 (21)
2
𝑍𝑓 [Ω]
AFL [mi] from Bus 1 after the measurement of the next sample (𝑘-th), 𝑟𝑘−1
𝐹𝑠 [kHz]
5 10 25 50 100 (calculated and saved in the previous step) is added to both
200 Error [%] 0.87 0.81 0.32 0.26 ≈0 sides of (21) as
2
500 Error [%] 0.50 0.28 0.17 0.08 ≈0 2
𝐴𝑘−1 + 𝑟𝑘−1 = 𝑚(𝑅̂𝑘−1 + 𝐵𝑘−1 ) + 𝑟𝑘−1
2
⏟ (22)
200 Error [%] 0.88 0.81 0.32 0.26 ≈0 𝑙ℎ𝑠
5
≈0 where 𝑙ℎ𝑠 = ∑𝑚+1 2
𝑖=1 𝑟(𝑘)−𝑖 , and the window length (𝑚)
500 Error [%] 0.50 0.28 0.18 0.07
200 Error [%] 0.91 0.83 0.32 0.27 ≈0
10 increases by 1. (20) for the 𝑘-th sample is
500 Error [%] 0.52 0.29 0.18 0.07 ≈0
1 𝑚+1
𝑅̂𝑘 = ∑ 𝑟2 𝑯 𝑘 𝑷−
−⏟ 𝑇
𝑘 𝑯𝑘
TABLE IV 𝑚 + 1 ⏟ 𝑖=1 (𝑘)−𝑖 (23)
AT-DETECTION METHODS IN QUALITATIVE COMPARISON 𝐵 𝑘
𝐴𝑘
Detection Method FIA NIF 𝐹𝑠 Noise replacing 𝐴𝑘 in (23) with 𝑙ℎ𝑠 in (22), yields
DWT, db-4, Scale 1 W(†) W RR(†) W 1
DWT, db-4, Scale 2 RR W W W 𝑅̂𝑘 = 𝑙ℎ𝑠 − 𝐵𝑘 (24)
𝑚+1
DWT, Haar W RR R(†) RR
HT R RR W RR Referring to (22), (24) is
2
𝑚 𝑟𝑘−1
𝑅̂𝑘 = (𝑅̂𝑘−1 + 𝐵𝑘−1 ) +
STMPM R R R R
−⏟ 𝑯 𝑘 𝑷− 𝑇
𝑘 𝑯𝑘 (25)
The Proposed Method R R RR R 𝑚+1 𝑚+1
(†) 𝐵𝑘
“W”, “R”, and “RR” are abbreviations for weak, robust, and relatively
robust, respectively. by defining 𝛬 = 𝑚/(𝑚 + 1), and referring to (20),
samples. The total running times of STMPM and the proposed 𝑅̂𝑘 = 𝛬(𝑅̂𝑘−1 + 𝑯𝑘−1 𝑷− 𝑇 2
𝑘−1 𝑯𝑘−1 ) + (1 − 𝛬)𝑟𝑘−1
(26)
method for processing 20,000 samples are 244.539002 s and − 𝑯 𝑘 𝑷−
𝑘 𝑯𝑘
𝑇

0.088102 s (88.102 ms), respectively. Thus, the proposed that is the recursive form of (8.b) as presented in (10).
method is considerably faster compared to STMPM. • Recursive Average of Residuals: the fixed forgetting factor
average is defined as [30], [31]
V. CONCLUSION 𝑘
Accurate detection of traveling-wave arrival times (ATs) 𝑟̅𝑘 = (1 − 𝜆) ∑ 𝜆(𝑘−𝑖) 𝑟𝑖 (27)
𝑖=1
plays a vital role in fault location based on ATs. Thus, this where 𝜆 ∈ (0,1). Therefore,
paper proposes a novel method for detecting the first AT of 𝑟̅1 = (1 − 𝜆)𝑟1 ,
TWs in power grids. The proposed method pinpoints the
𝜆(1 − 𝜆)𝑟1 + (1 − 𝜆)𝑟2 ,
𝑟̅2 = ⏟
abrupt changes in the residuals (innovation sequence)
𝜆𝑟̅1
generated by an adaptive discrete Kalman filter (ADKF)
⏟2 (1 − 𝜆)𝑟1 + 𝜆(1 − 𝜆)𝑟2 + (1 − 𝜆)𝑟3 , …, and
𝑟̅3 = 𝜆
through the exponentially weighted moving average (EWMA) 𝜆𝑟̅2
as a high-gain control chart. Then, the Chi-square filter ⏟𝑘−1 (1 − 𝜆)𝑟1 + 𝜆𝑘−2 (1 − 𝜆)𝑟2 + ⋯ + (1 − 𝜆)𝑟𝑘 .
𝑟̅𝑘 = 𝜆
verifies the released alarm validity. The proposed method is 𝜆𝑟̅𝑘−1
recursive. Therefore, it is proper for implementation in on-site Therefore, 𝑟̅𝑘 = 𝜆𝑟̅𝑘−1 + (1 − 𝜆)𝑟𝑘 , and it is (27) in a
microprocessor-based fault locators managing real-time recursive manner as presented in (11).
streams of high time-resolution measurements. The transient • The EWMA threshold: Referring to (13), and considering
simulations are performed utilizing EMTP-RV and the real- that 𝑟𝑘𝑁 ~𝒩(0, 1), then
time digital simulator (RTDS). The proposed method is 𝑔1 , 𝑔1′ ~𝒩(0, (1 − 𝜓)2 ),
compared with the state-of-the-art AT-detection methods in
𝑔2 , 𝑔2′ ~𝒩(0, 𝜓 2 (1 − 𝜓)2 + (1 − 𝜓)2 ),
the literature using MATLAB. According to the sensitivity
𝑔3 , 𝑔3′ ~𝒩(0, 𝜓 4 (1 − 𝜓)2 + 𝜓 2 (1 − 𝜓)2 + (1 − 𝜓)2 ),…, and
analysis, the performance of the proposed method is
satisfactory for AT detection. The measurement noises, fault 𝑔𝑘 , 𝑔𝑘′ ~𝒩(0, ∑𝑘𝑗=1 𝜓 2(𝑗−1) (1 − 𝜓)2 ).
inception angle (FIA), and non-ideal faults (NIFs) Based on the properties of geometric series and as 𝜓 < 1,
comparatively less affect the proposed method. However, (1−𝜓)2
lower sampling frequencies (𝐹𝑆 ) to some extent reduce the 𝑔𝑘 , 𝑔𝑘′ ~𝒩 (0, ) , ∀𝑘 ≫ 1. Therefore, using the normal
1−𝜓2
accuracy of the proposed method. Gaussian PDF, the value of the threshold 𝑇 can be selected.
For example, for 𝜓 = 0.1, 𝑔𝑘 , 𝑔𝑘′ ~𝒩(0,9/11), ∀𝑘 ≫ 1. Thus,
APPENDIX A 𝑇 = 1.488 indicates that 𝑔𝑘 and 𝑔𝑘′ are Gaussian noises for
• Recursive Measurement Noise Estimation: For any given 95% and the released alarm is false for 5%.
(𝑘 − 1)-th sample satisfying (8.a), (8.b) is [27]
1 𝑚
APPENDIX B
𝑅̂𝑘−1 = ∑ 𝑟(𝑘−1)−𝑖 2
−⏟𝑯𝑘−1 𝑷− 𝑇
𝑘−1 𝑯𝑘−1
𝑚 ⏟ 𝑖=1 (20) • The topology of the TL used in Test Case 1 is shown in Fig.
𝐵 𝑘−1
𝐴𝑘−1
B. 1 [32], and its details are as follows: The DC resistances of
 9

[13] A. Borghetti, M. Bosetti, M. Di Silvestro, C. A. Nucci, and M.

Paolone, “Continuous-wavelet transform for fault location in
distribution power networks: definition of mother wavelets inferred
from fault originated transients,” IEEE Trans. Power syst., vol. 23,
no. 2, pp. 380 – 388, Apr. 2008.
[14] P. E. Argyropoulos and H. Lev-Ari, “Wavelet customization for
improved fault-location quality in power networks,” IEEE Trans.
Power Del., vol. 30, no. 5, pp. 2215-2223, Oct. 2015.
[15] X. Dong, J. Wang, S. Shi, B. Wang, B. Dominik, and M. Redefern,
“Traveling wave based single-phase-to-ground protection method for
Fig. B. 1. The topology of the TL used in Test Case 1, the unit is meter. power distribution system,” CSEE Journal of Power and Energy
Systems, vol. 1, no. 2, pp. 75-82, Jun. 2015.
[16] F. B. Costa, A. H. P. Sobrinho, M. Ansaldi, and M. A. D. Almeida,
“The effects of the mother wavelet for transmission line fault detection
and classification,” in Proc. 3rd IYCE, Leiria, Portugal, Jul. 2011.
[17] F. V. Lopes, D. Fernandes, and W. L. A. Neves, “A traveling-wave
detection method based on Park’s transformation for fault locators,”
IEEE Trans. Power Del., vol. 28, no. 3, pp. 1626-1634, Jul. 2013.
[18] A. Bernadic and Z. Leonowicz, “Fault location in power networks with
mixed feeders using the complex space-phasor and Hilbert–Huang
transform,” Elec. Pow. and Ener. Syst., vol. 42, pp. 208-219, 2012.
[19] Y. Hao, Q. Wang, Y. Li, and W. Song, “An intelligent algorithm for
Fig. B. 1. The topology of the TL used in Test Case 2, the unit is meter. fault location on VSC-HVDC system,” Inter. Journal of Elec. Power
& Energy Syst., vol. 94, pp.116-123, Jan. 2018.
the phase conductors and shield wires are 0.03214 and 1.6216 [20] Z. Moravej, M. Movahhedneya, G. Radman, and M. Pazoki, “Effective
Ω/km, respectively. The outside diameters of the phase fault location technique in three-terminal transmission line using
Hilbert and discrete wavelet transform,” in Proc. EIT, Dekalb, IL, pp.
conductors and shield wires are 4.06908 and 0.98044 cm, 170-176 May. 2015.
respectively. [21] S. S. Gururajapathy, H. Mokhlis, and H. A. Illias, “Fault location and
• The topology of the TL utilized in Test Case 2 is depicted in detection techniques in power distribution systems with distributed
generation: A review,” Renewable and Sustainable Energy Reviews,
Fig. B. 2 [36]. The phase conductors are ACSR 636.0 MCM vol. 74, pp. 949-958, Jul. 2017.
26/7, and the shield conductors are ACC 3/8’ HS 7 wires. [22] H. Hasanvand, A. Parastar, B. Arshadi, M. R. Zamani, and A. S.
Bordbar, “A comparison between S-transform and CWT for fault
location in combined overhead line and cable distribution networks,”
REFERENCES
in Proc. EPDC, Karaj, Iran, Apr. 2016.
[1] IEEE Guide for Fault Location Techniques on Shielded Power Cable [23] C. Xi, Q. Chen, and L. Wang, “A single-terminal traveling wave fault
Systems, IEEE Standard 1234, 2007. location method for VSC-HVDC transmission lines based on S-
[2] R. J. Hamidi and H. Livani, “Traveling wave-based fault location transform,” in Proc. APPEEC, Xi'an, China, pp. 1008-1012, Oct.
algorithm for hybrid multi-terminal circuits,” IEEE Trans. Power Del., 2016.
vol. 32, no. 1, pp.135-144, Feb. 2017. [24] R. J. Hamidi, H. Livani, and R. Rezaeisarlak, “Traveling-wave
[3] S. L. Zimath, M. A. Ramos, J. S. Filho, J. M. Beck, and N. Mueller, detection technique using short-time matrix pencil method,” IEEE
“Traveling wave-based fault location experiences,” in Proc. Trans. Power Del., vol. 32, no. 6, pp. 2565-2574, Mar. 2017.
Protective Relay Eng., College Station, TX, pp. 1-7, Mar.-Apr. 2010. [25] L. Xun, L. Shungui, H. Ronghui, A. Jingwen, A. Yunzhu, C. Ping, and
[4] M. Ando, E. O. Schweitzer, and R. A. Baker, “Development and field- X. Zhengxiang, “Study on accuracy traveling wave fault location
data evaluation of single-end fault locator for two-terminal HVDC method of overhead line-cable hybrid line and its influencing factors,”
transmission lines, Part I: data collection system and field data,” IEEE in Proc. Chinese Automation Congress (CAC), Jihan, China, Oct.
Trans. Power Apparatus and Systems, vol. PAS–04, no. 12, pp. 3524– 2017.
3530, Dec. 1985. [26] A. A. Girgis, W. B. Chang, and E. B. Makram, “A digital recursive
[5] M. Ando, E. O. Schweitzer, and R. A. Baker, “Development and field- measurement scheme for on-line tracking of power system
data evaluation of single-end fault locator for two-terminal HVDC harmonics,” IEEE Trans. Power Del., vol. 6, no. 3, pp. 1153-1160, Jul.
transmission lines, Part II: Algorithm and Evaluation,” IEEE Trans. 1991.
Power Apparatus and Systems, vol. PAS–104, no. 12, pp. 3531–3537, [27] M. M. Giray and M. S. Sachdev, “Off-nominal frequency
Dec. 1985. measurement in electric power systems,” IEEE Trans. Power Del.,
[6] R. H. Salim, K. R. C. de Oliveira, and A. S. Bretas, “Fault detection vol. 4, no. 3, pp. 1573-1578, Jul. 1989.
in primary distribution systems using wavelets,” in Proc. IPST, Lyon, [28] R. K. Mehra, “Approaches to adaptive filtering,” IEEE Trans.
France, Jun. 2007. Automatic Control, vol. 17, no. 5, pp. 693-698, 1972.
[7] F. H. Magnago and A. Abur, “Fault location using wavelets,” IEEE [29] IEEE Guide for Power-Line Carrier Applications, IEEE Standard
Trans. Power Del., vol. 13, no. 4, pp. 1475-1480, Oct. 1998. 643-2004, 2004.
[8] C. Y. Evrenosoglu and A. Abur, “Travelling wave based fault [30] Fredrik Gustafsson, “Change detection,” in Adaptive filtering and
location for teed circuits,” IEEE Trans. Power Del., vol. 20, no. 2, pp. change detection, John Wiley & Sons, Ltd, West Sussex, England,
1115-1121, Apr. 2005. 2000.
[9] L. Peretto, R. Sasdelli, and R. Tinarelli, “On uncertainty in wavelet [31] D. A. Bodeman, “Adaptive estimation with change detection for
based signal analysis,” IEEE Trans. Instrum. Meas., vol. 54, no. 4, pp. streaming data,” Ph. D. Thesis, University of London and the Imperial
1593–1599, Aug. 2005. College, Nov. 2014.
[10] M. Korkali and A. Abur, “Robust fault location using least-absolute- [32] EMTP-RV, Example for Line Data, [Online]. Available:
value estimator,” IEEE Trans. Power Syst., vol. 28, no. 4, pp. 4384- http://www.emtp-software.com/
4392, Nov. 2013. [33] Instrument Transformers - The Use of Instrument Transformers for
[11] L. Yuansheng, W. Gang, and L. Haifeng, “Time-domain fault-location Power Quality Measurement. IEC/TR Standard 61869-103, 2012.
method on HVDC transmission lines under unsynchronized two-end [34] F. Rahmatian, “Brief overview of optical current and voltage sensors
measurement and uncertain line parameters,” IEEE Trans. Power Del., in the electric power industry,” NASPI Distribution Task Team, Jun.
vol. 30, no. 3, pp. 1031-1038, Jun. 2015. 2017.
[12] M. Shafiullah, M. A. Abido, and Z. Al-Hamouz, “Wavelet-based [35] L. Coffeen, J. McBride, and D. Cantrelle, “Initial development of EHV
extreme learning machine for distribution grid fault location,” IET bus transient voltage measurement: an addition to on-line transformer
Gen., Trans., and Dist., vol. 11, no. 17, pp. 4256 - 4263, Dec. 2017.
 10

FRA,” EPRI Substation Equipment Diagnostics Conference, Orlando,

FL, Mar. 2008.
[36] F. V. Lopes, K. M. Silva, F. B. Costa, W. L. A. Neves, and D.
Fernandes, Jr., “Real-time traveling-wave-based fault location using
two-terminal unsynchronized data,” IEEE Trans. Power Del., vol. 30,
no. 3, pp. 1067-1076, Jun. 2015.
[37] H. W. Dommel, “Overhead transmission lines,” in EMTP Theory
Book, USA: BPA, 1995, ch 4, pp. 64-149.
[38] Available online: https://www.mccdaq.com/products/pdaq3s