Electric Power Systems Research 228 (2024) 110055

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Characterizing transformer HV–LV winding FRA curves through derivation

of transfer functions from FRA data
Z. Zhang a, b, *
a
Department of Electronics & Computer Science, Southampton University, Southampton SO17 1BJ, UK
b
Department of Electrical and Electronic Engineering, Manchester University, Manchester M13 9PL, UK

A R T I C L E I N F O A B S T R A C T

Keywords: Mechanical malfunctions pose a significant threat to the reliability of power transformers. Frequency Response
Frequency response analysis Analysis (FRA) of transformer windings (TWs) serves as a predictive technology, allowing the assessment of
Power transformer mechanical conditions based on collected FRA data. Changes in the winding structure manifest in FRA charac­
Transfer function
teristics, making it possible to detect incipient faults. This paper presents the derivation of transfer functions (TF)
Transformer winding structure
equations through processing FRA data, aiming to investigate and interpret the characteristics of TWs. The
derived TFs are utilized to mathematically describe FRA curves, enabling the characterization of winding
structures by observing the FRA curves generated from these TF equations. Not only do the derived TFs yield very
similar FRA curves, but the TF estimation also provides valuable insights for assessing TW malfunctions. The
similarity analysis, using calculated RMSE and MAE values of 0.4807 and 0.2020, respectively, demonstrates a
strong correlation between the derived TF curves and the measured FRA curves. This affirms the validity of TF
estimation and underscores its significant impact on the advancement of FRA technology.

1. Introduction employing an RLC EC model to simulate winding deformations by

adjusting the values of the RLC parameters. This approach allows for the
The growing complexity of transmission and distribution system evaluation of winding failures that result in variations in the RLC values
topologies demands enhanced predictability in managing prosumer [7–15]. All these RLC EC models were obtained by conducting analyses
dispatch, necessitating the application of new techniques for state esti­ of both the transformer and its winding structures. The derivation of TFs
mation and the utilization of metaheuristics [1]. Specifically, the reli­ from the proposed EC models or through processing measured FRA data
ability of power transformers holds paramount importance within has been a longstanding technical challenge that has perplexed re­
transmission and distribution networks, as failures in these transformers searchers for decades. Establishing a connection between ECs and
can result in severe blackouts [2]. Consequently, preventive measures measured FRA data has presented a daunting conundrum that has
are imperative. Transformer predictive maintenance technologies have seemed insurmountable.
been developed, primarily utilizing methods such as dissolved gas The study in Ref. [16,17] explains that a power transformer’s mag­
analysis and FRA to assess and ensure the health and performance of netic core and winding structure can be represented as a complex
transformers, thus mitigating the risk of blackouts [3]. Since its proposal electrical network comprising resistances, self-inductances, ground ca­
by Dick and Erven in 1978 [4], FRA technology has been recognized as a pacitances, coupling inductances, and series capacitances. Understand­
powerful method for diagnosing TW deformations [5,6]. FRA data is ing the behaviour of this EC is crucial for interpreting the shape of the
obtained by measuring the response of TWs to excitation across a wide measured FRA curves. Currently, the widely used approach for
frequency range. This data contains diagnostic information about the analyzing FRA data is to compare FRA curves between the recently
complex equivalent circuit (EC) consisting of resistance, inductance, and measured data on a transformer and the fingerprint FRA data obtained
capacitance (RLC) components that represent the structure of the TWs from the same measurement setup, or by comparing FRA data collected
[5]. However, a direct connection between the RLC EC and FRA data has from identical units or different phases (or limbs) of the same trans­
yet to be established. Therefore, investigations typically involve former [5,6,18,19]. However, the effectiveness of interpreting FRA data

* Correspondence to: Department of Electronics & Computer Science, Southampton University, Southampton SO17 1BJ, UK.
E-mail address: [email protected].

https://doi.org/10.1016/j.epsr.2023.110055
Received 23 August 2023; Received in revised form 12 November 2023; Accepted 28 November 2023
0378-7796/© 2023 Published by Elsevier B.V.
Z. Zhang Electric Power Systems Research 228 (2024) 110055

heavily relies on empirical practices, which highlights the need for a established. This signifies that there are currently no techniques avail­
systematic approach to process and evaluate this data for assessing the able to directly correlate the electrical parameters of the RLC EC with
mechanical condition of a transformer. Despite significant advance­ FRA data. The primary goal of this study is to tackle and resolve this
ments in frequency-response measurement techniques, a practical and crucial issue. The proposed method to achieve this purpose consists of
objective methodology for interpreting FRA data remains unavailable. three steps. Firstly, the FRA experiment is conducted, and the FRA
measurement data is collected. Secondly, the measurement data is
processed to derive the TF equations, which serve as the bridge con­
necting the measurement data and the RLC EC network of the tested
Nomenclature
winding. Finally, the derived TF equations are used to determine the EC
Parameters
TF Transfer function network and calculate its RLC parameters. In this study, the main
s complex frequency research focus lies in deriving TFs to characterize the FRA curve, and the
k constant derivation of the RLC EC network will be presented in a subsequent
z zero paper.
p pole
ω angular frequency
The exploration of TF estimation aims to develop a systematic
ζ damping coefficient guideline for characterizing the features of FRA measurements. This
Amp amplitude of transfer function method will be utilized to gain a better understanding of the charac­
Phase phase angle of transfer function teristics of FRA technology. Previously, FRA curves were predominantly
TFD individual transfer function in denominator
assessed using visual aids [5,6]. However, this study is specifically
TFN individual transfer function in numerator
Abbreviations focused on interpreting FRA curves through mathematical analysis and
FRA Frequency response analysis techniques. One significant advantage of this approach is that the
TF Transfer function derived TF equations establish a direct correspondence with the seg­
EC Equivalent circuit ments of the FRA curves that represent the features of the EC. Further­
RLC Resistance, inductance, and capacitance
more, this method has the capability to unveil concealed circuit features
TW Transformer winding
HV High voltage within the FRA curves that were previously overlooked by earlier re­
LV low voltage searchers. This allows for a more comprehensive mathematical expla­
N Neutral nation of the intricacies presented in the FRA curve details. Ultimately,
RMSE Root mean squared error
this method holds the potential to derive the authentic EC of the RLC
MAE Mean absolute error
network and accurately calculate the electrical parameters of the RLC EC
for tested TWs using the individual TFs obtained from the FRA data.
Various types of mechanical defects and deformations in TWs are
expected to become evident in FRA curves. The accurate determination
To ensure objectivity in interpreting FRA data, various statistical of TFs has been a longstanding anticipation among scientists and engi­
indices have been developed and discussed in the literature review of neers. Notably, comprehending the specific physical implications of
FRA technology [20–23]. These indices employ statistical equations to distinct curve sections within FRA curves has posed a perplexing chal­
provide a numerical assessment of the correlation and similarity be­ lenge for researchers over the course of decades. This study has effec­
tween newly obtained FRA data and fingerprint data from transformers. tively tackled this predicament by offering mathematical elucidations
By capturing the variation in FRA data, these indices aim to predict for each unique curve section found within FRA curves. Furthermore,
mechanical faults in windings. However, measurement results can be the mathematical derivations of TFs have enabled the precise explana­
imprecise due to external factors like temperature and humidity. In such tion of peaks and troughs observed in the curves.
cases, these external factors influence the measurement data, and The pioneering accomplishments of this study encompass the initial
relying solely on index calculations can lead to inaccurate diagnoses. To attainment of distinct TFs corresponding to specific curve sections
address this issue, a proposition has been put forth to analyze a trend within FRA curves. This study has also succeeded in portraying the al­
curve of the results, thereby avoiding potential errors. terations observed in the curve sections of FRA curves by utilizing the
The analysis of TFs is pivotal in identifying faults within power TWs. parameters from the derived TFs. Undoubtedly, this study holds signif­
Each TW possesses a unique TF, often referred to as the "fingerprint" of icant importance as it addresses a technical gap by describing the
the transformer. Ref. [24] introduced a TW model based on the TF behaviour of FRA curves. Previous attempts to establish a successful link
approach, enabling the detection of faults such as buckling, displace­ between ECs and TFs with FRA data have been unsuccessful, primarily
ment, and short circuits within the winding. Typically, assessing wind­ due to the intricate nature of TWs. Thus, the primary objective of this
ing faults involves interpreting variations in TF. Both simulation and study is to characterize FRA curves and rectify existing mis­
experimental methods have been employed to interpret these variations, interpretations accurately. A vital advantage of this study is that the
allowing for the investigation of inter-turn faults in TWs [25]. Further­ derived TFs facilitate the simulation of FRA with exceptional accuracy,
more, Ref. [26] offered valuable insights, demonstrating that various closely matching the measured FRA data.
types of winding defects observed in real TWs can be used to analyze the
sensitivity of each component of TF. This includes the amplitude, phase, 2. FRA test specimen and measurement setup
real, and imaginary parts for different fault types, locations, or extents.
Moreover, Ref. [27] compared diverse mathematical approaches used to The FRA measurements collected from Autotransformer T7014 from
obtain the TF of a transformer, aiming to optimize the TF approach for National Grid were processed to derive the TF equations. This 400/275/
the online FRA technique. Through TF evaluation, the characteristics of 13 kV autotransformer features a three-limb core design without bolts,
TWs and their ECs can be analyzed, facilitating the predictive determi­ incorporating an interleaved disc series winding, an intershielded disc
nation of TW faults. common winding, and a helical tertiary winding. The FRA data for A
While significant research has been dedicated to the TF approach for phase high voltage (HV) to low voltage (LV), with other winding neu­
identifying TW defects, certain areas still warrant further exploration trals grounded and tertiary floated, was recorded on 01/08/98. The
and refinement. For instance, a comprehensive mathematical explana­ measurement setup is illustrated in Fig. 1. On the same day, measure­
tion of FRA curves remains an unaccomplished task, and a similar ments were also conducted for A phase Neutral (N) to LV and 3A2 to 3B2
detailed mathematical explanation of winding ECs has yet to be tertiary configurations, which will be discussed in future work. How­
ever, it is essential to note that the measurement configuration depicted

2
Z. Zhang Electric Power Systems Research 228 (2024) 110055

process the FRA data and obtain the vectors of coefficients, denoted as b
and a, which correspond to the numerator and denominator of Eq. (1).

bn sn + bn− 1 sn− 1 + … + b0
TF(s) = (1)
am sm + am− 1 sm− 1 + … + a0
FRA characteristics are mathematically represented by Eq. (1),
where the coefficients a and b correspond to the combination of RLC
elements in the EC network, and these coefficients are always real
numbers. The numerator of the TF is a polynomial of degree n, while the
denominator is a polynomial of degree m. The numerator and denomi­
nator polynomials can be factored into expressions of the product of
binomials. The zeros and poles of the TF are the roots of the numerator
and denominator polynomials, as represented in Eq. (2). Consequently,
Eq. (1) can be alternatively expressed in a factored form as shown in Eq.
Fig. 1. A phase HV to LV configuration for the FRA measurements, with the (2).
neutrals of other windings grounded and the tertiary windings left floating.
(s − z1 )(s − z2 )…(s − zn )
TF(s) = k (2)
in Fig. 1 eliminates the interaction between windings by grounding the (s − p1 )(s − p2 )…(s − pm )
neutrals of other windings and the magnetic core but are separated from The constant k is associated with the RLC network of the TW’s EC. As
the measured winding. Consequently, a shunt capacitance is formed the degrees of the polynomials of the numerator and denominator in­
between the tested winding and the magnetic core, along with the other crease, finding the roots of these polynomials becomes more chal­
windings. Exploring the interplay between windings is equally crucial lenging. However, the MATLAB built-in function roots [28] can be
for assessing mechanical conditions. This matter will be a focal point of utilized to calculate the roots of both the numerator and denominator of
investigation in upcoming research endeavors. Meanwhile, it is essential the TF. Due to the unique characteristics of the EC, the zeros and poles
to recognize that terminals and bushings wield a critical influence on can be complex conjugate numbers. Moreover, the complex conjugate
FRA measurements. In the current measurement setup, the impacts of roots can be used to construct two-order trinomials, which can then be
terminals and bushings are considered, but their effects should be employed to form another TF equation, as shown in Eq. (3).
examined individually in future studies. ∏ ∏( 2 )
The following conclusions are drawn based on the EC model
2
1 k (s + ωk ) l s + 2ζ l ωl s + ωl
TF(s) = k m ∏ (
∏ 2 ) (3)
described in Ref. [5]. At frequencies below 2 kHz, the FRA response is s i (s + ωi ) n s + 2ζ n ωn s + ωn
2

primarily influenced by the core effect. Between 2 kHz and 20 kHz, the
The parameters ωk , ωi , ωl , and ωn represent the corner fre­
FRA response is affected by the interaction between windings. From 20
quencies, while ζl and ζn denote the damping coefficients. These pa­
kHz up to 1 MHz, the winding structure becomes a significant factor in
rameters are directly associated with the RLC parameters of the TW’s
the frequency response. Beyond 1 MHz, the tap and grounding leads
EC. The multiplicative constant k is linked to the product of the pole and
impact the trace of the frequency response [5]. However, validating
zero corner frequencies. The term sm serves the purpose of representing
these explanations regarding the core and winding effects is challenging
pure inductances or capacitances, indicating when zeros or poles are
due to the uncertainty and lack of clarified TW structures. For this study,
equal to 0. Eq. (3) proves to be highly practical for characterizing the
the focus is solely on the FRA data collected from A phase HV to LV. In
FRA curves of TWs in relation to individual curve sections. This is due to
the measurement setup, five amplitude and phase angle pairs were
the direct linkage between the binomials and trinomials and the distinct
measured for different frequency ranges: 5 Hz–2 kHz, 50 Hz–20 kHz,
curve sections present within the FRA curves. In the TF equation, the
500 Hz–200 kHz, 5 kHz–2 MHz, and 25 kHz–10 MHz, respectively. In
binomials in the numerator correspond to single-zero responses, while
each region, a total of 400 frequency points is meticulously documented,
the trinomials in the numerator represent double-zero responses. On the
evenly distributed with specific sampling intervals of 0.005 kHz, 0.05
other hand, the binomials in the denominator correspond to single-pole
kHz, 0.5 kHz, 5 kHz, and 25 kHz, respectively. Subsequently, ten sets of
responses, and the trinomials in the denominator represent double-pole
measurement data were gathered and amalgamated to create a
responses. These binomials and trinomials are crucial in establishing the
comprehensive spectrum of FRA measurements, covering the entire
connection between the FRA data and the RLC elements of the TW’s EC.
dataset. This compilation involved eliminating overlapping data, spe­
cifically at low-frequency regions, to form a complete FRA dataset. The
4. FRA data processing procedures
amplitude signifies the ratio between the output voltage and the input
voltage, while the phase angle denotes the phase shift of the output
The process of deriving the TF equations from FRA data, as depicted
voltage concerning the input voltage induced by the inductances and
in Fig. 2, consists of three main steps. Firstly, the FRA data is imported
capacitance present in the winding EC. Subsequently, the curves of the
into the program, and the MATLAB built-in function invfreqs is
amplitude and phase angle responses of the studied winding can be
generated using MATLAB code for this study.

3. MATLAB functions and transfer function equations

The collected FRA measurements comprise raw data that contains

valuable mechanical information about TWs. However, without reliable
techniques, it is challenging to utilize this raw data effectively for
detecting TW mechanical conditions. The purpose of this study is to
extract the mechanical information embedded within the FRA data by
employing data analysis functions available in MATLAB. By processing
the data using these functions, valuable insights and findings can be
derived.
First, the MATLAB built-in function invfreqs [28] is employed to Fig. 2. Flowchart of FRA data processing.

3
Z. Zhang Electric Power Systems Research 228 (2024) 110055

the TF equation and those plotted using FRA data serving as a metric of
the accuracy and reliability of the derived TF equation. In future en­
deavors, the parameters obtained from the derived TF equation can be
utilized to create indices or establish ECs dedicated to diagnosing the
mechanical conditions of power transformers.

5. Simulation results

During the processing of FRA data, the FRA curves obtained from
measurements serve as a reference benchmark to assess the precision
and accuracy of the TF equations. A comparison is made between the
FRA curves generated from the FRA data and the TFs to identify the best-
matched TF equations. Once the appropriate TF is determined, the co­
efficients of each individual TF can be calculated. By combining all in­
Fig. 3. Comparison between the measured FRA amplitude curves for A phase dividual TFs, the complete TF equation is obtained. The characteristics
HV to LV and the modelled FRA amplitude curves. TFs are calculated for the of the individual TF equations can be observed by examining the FRA
selected frequencies in order to facilitate this comparison. curves produced through the integration of one additional individual TF
at a time. This allows for a deeper understanding of the behaviour and
impact of each individual TF equation.
The developed MATLAB code is utilized to process the FRA data for
the series winding from A phase HV to LV. The initial step of data pro­
cessing involves removing repetitive data at low frequencies and
creating a comprehensive range of FRA data spanning from 5 Hz to 10
MHz. To ensure effective use of the invfreqs function and obtain precise
coefficients, the frequency range is divided into five ranges. The co­
efficients obtained are then utilized to establish Eq. (1) for the selected
frequency ranges. The comparisons between the measured FRA curves
and the FRA curves generated from Eq. (1) for these frequency ranges
are illustrated in Figs. 3 and 4, respectively. Clearly, the FRA curves,
both for amplitude and phase angle, modelled using Eq. (1) and co­
efficients derived from the measured FRA data, exhibit remarkable
similarity when compared to the actual measured FRA curves. This
Fig. 4. Comparison between the measured FRA phase angle curves for A phase methodology has undergone rigorous testing across various types of TWs
HV to LV and the modelled FRA phase angle curves. TFs are calculated for the and numerous FRA datasets to ensure the reproducibility of results.
selected frequencies in order to facilitate this comparison. After obtaining the poles and zeros, the final individual TFs are
computed and presented in Tables 1 and 2 for the numerator and de­
employed to process the FRA data and obtain two vectors, b and a, which nominator, respectively. These individual TFs cover ranges from 5 Hz to
represent the coefficients of the polynomials in Eq. (1). Next, the 350 KHz, which are sufficient to characterize the FRA curves using the
MATLAB function freqs is utilized to derive Eq. (1), and the resulting derived TFs. However, the individual TFs from range four and range five
FRA curves are generated. By comparing these calculated FRA curves are excluded due to limitations in computer processing power and
with those obtained from the FRA data, the accuracy of the TF can be MATLAB capabilities. Including more individual TFs in these ranges
assessed. If the FRA curves generated from Eq. (1) do not align with would introduce significant calculation errors. This limitation will be
those from the FRA data, the order in invfreqs can be adjusted to achieve addressed in future work by utilizing a supercomputer. Thus, the
more precise results. partially completed TF for the transformer FRA, measured from 5 Hz to
Secondly, the MATLAB function roots can be employed to process the 350 KHz, can be obtained as follows:
vectors b and a, enabling the calculation of the zeros and poles in Eq. (2). ∏24
TFNn (s)
Subsequently, the TF equations in Eq. (2) can be derived by taking the TF(s) = k ∏25n=1 (4)
product of all the binomials formed by the zeros and poles. m=1 TFD m (s)

Lastly, the obtained zeros and poles undergo subsequent processing The constant k is determined by measuring the deviation between
to derive Eq. (3). Real roots are harnessed to create binomials, whereas the measured and modelled FRA amplitude curves and subsequently
complex conjugate roots are utilized to derive coefficients for double calculating k based on the amplitude definition. Additionally, the
zeros and double poles. These coefficients are then applied to formulate amplitude and phase angle can be computed using the following
trinomials, representing responses involving double zeros and double equations:
poles. The amalgamation of all binomials and trinomials results in a
comprehensive TF equation. This TF equation serves as the basis for Amp(TF) = 20log10 |TF| (5)
plotting FRA curves, with the deviation between the curves derived from

Table 1
Individual TFs of the numerator obtained from the FRA data for A phase HV to LV.
TFN1 - 5 TFN6 - 10 TFN11 - 15 TFN16 - 20 TFN21 - 24

s + 9.6285 s2 + (9.84e03)s + 1.66e08 s2 + (1.57e04)s + 9.12e09 s2 + (2.35e04)s + 5.06e10 s2 + (3.76e04)s + 6.58e11

s2 + (5.38e01)s + 3.26e06 s2 + (2.94e03)s + 4.32e08 s2 + (2.36e04)s + 1.28e10 s2 + (2.42e04)s + 6.01e10 s2 + (7.50e04)s + 8.81e11
s2 + (1.04e02)s + 5.38e06 s2 + (1.99e03)s + 5.33e08 s2 + (1.54e04)s + 1.84e10 s2 + (2.08e04)s + 9.42e10 s2 + (7.40e05)s + 6.99e11
s2 + (7.11e03)s + 4.05e07 s2 + (2.25e03)s + 8.08e08 s2 + (1.84e04)s + 2.22e10 s2 + (2.89e04)s + 3.04e11 s2 + (5.96e04)s + 1.14e12
s2 + (1.07e01)s + 1.56e08 s2 + (9.55e03)s + 6.22e09 s2 + (3.28e04)s + 4.22e10 s2 + (5.90e04)s + 4.11e11

4
Z. Zhang Electric Power Systems Research 228 (2024) 110055

Table 2
Individual TFs of the denominator obtained from the FRA data for A phase HV to LV.
TFD1 - 5 TFD6 - 10 TFD11 - 15 TFD16 - 20 TFD21 - 25

s + 0.0001 2
s + (1.02e04)s + 1.78e08 2
s + (8.74e03)s + 6.28e09 2
s + (2.34e04)s + 5.06e10 s2 + (5.78e04)s + 4.15e11
s + 27.5132 s2 + (1.67e03)s + 3.04e08 s2 + (1.54e04)s + 9.20e09 s2 + (2.41e04)s + 6.06e10 s2 + (3.74e04)s + 6.57e11
s2 + (5.93e01)s + 3.57e06 s2 + (3.01e03)s + 4.34e08 s2 + (2.13e04)s + 1.40e10 s2 + (7.94e07)s + 8.94e15 s2 + (7.52e04)s + 8.80e11
s2 + (6.66e03)s + 4.00e07 s2 + (2.06e03)s + 5.33e08 s2 + (1.48e04)s + 1.94e10 s2 + (2.14e04)s + 9.36e10 s2 + (7.87e05)s + 7.57e11
s2 + (1.08e01)s + 1.56e08 s2 + (4.08e03)s + 1.58e09 s2 + (3.20e04)s + 4.25e10 s2 + (2.82e04)s + 3.03e11 s2 + (5.98e04)s + 1.14e12

collective influence on the overall TF, the corresponding FRA curves for
each individual TF are scrutinized through their sequential integration.
Each individual TF is presumed to correlate with specific curve segments
within an FRA curve, signifying distinct stages or loop structures within
the RLC network of the TW’s EC. Thus, the exploration of these indi­
vidual TFs confers a noteworthy advantage in deriving the TW EC. The
interrelation between the individual TFs and EC topologies will be
elaborated upon in a forthcoming paper, illuminating the precise link­
ages between segment FRA curve sections, individual TFs, specific RLC
elements, and configurations within the EC of TWs.
In Figs. 7 and 8, it is observed that the combination of the constant k,
TFN1, TFD1, and TFD2, derived from the FRA data measured for the A
phase HV to LV, can produce FRA curves with remarkably high simi­
larity at frequencies lower than 100 Hz. These modelled curves closely
resemble the first sections of the measured FRA amplitude and phase
Fig. 5. Comparison between the FRA amplitude curves measured for A phase angle curves. Notably, the first curve sections at low frequencies in the
HV to LV and modelled using Eq. (4). FRA curve are determined by single-zero and single-pole responses,
which are distinct from the double responses exhibited by the other
Phase(TF) = angle(TF) (6) individual TFs. The observed discrepancy at the initial point of the phase
angle curve arises due to the parameters within the binomials.
The TF equation for the transformer FRA is formed by combining Enhancing accuracy can be attained by refining these binomial param­
individual TFs, encompassing single zero responses, single pole re­ eters; however, such refinement is not imperative for investigating the
sponses, double zero responses, and double pole responses. The com­ traits of TW FRA curves. Therefore, the parameters of the individual TFs
parisons between the measured FRA curves and the FRA curves listed in Tables 1 and 2 are the values computed using MATLAB code
generated from Eq. (4) are illustrated in Figs. 5 and 6, respectively. based on the measured FRA data.
It is evident that both the amplitude and phase angle FRA curves According to conventional wisdom, the initial curve segments
generated using the individual TFs exhibit remarkable similarities to spanning from the starting point to the first trough of the FRA curve are
their corresponding measured FRA curves. However, a certain level of traditionally thought to depict a single-pole response, signifying a low-
disagreement is observed at the third peak between the curves of TFs frequency inductance response. However, the author’s investigation
and the measured data. Two potential factors could contribute to this involved conducting simulations using a solitary single-pole response,
disparity: either the sample point deviates from the peak, or the calcu­ reveals significant disparities in both amplitude and phase angle re­
lated peak deviates from the actual peak. Nevertheless, it is important to sponses at the outset. Upon utilizing MATLAB functions for data pro­
note that this disparity does not undermine the investigation related to cessing, the outcomes indicated that the low-frequency response is
the derivation of the TF. The discrepancy demonstrated in Fig. 6 be­ portrayed by a combination of a single-zero response and two single-
tween the modelled and measured curves stems from limitations in pole responses. This inconsistency becomes particularly pronounced in
MATLAB and computer processing capabilities. Addressing this matter is the phase angle curve of the single-pole response, wherein a deviation is
essential when developing online instrumentation, necessitating the observable at the commencement of the linear segment in comparison to
utilization of dedicated microprocessors to mitigate such issues. the measured FRA curve. Given the intricate characteristics of power TW
To gain insight into the characteristics of individual TFs and their

Fig. 6. Comparison between the FRA phase angle curves measured for A phase Fig. 7. Comparison of the FRA amplitude curves between measured FRA data
HV to LV and modelled using Eq. (4). and modelled from the combination of k, TFN1, TFD1 and TFD2.

5
Z. Zhang Electric Power Systems Research 228 (2024) 110055

Fig. 8. Comparison of the FRA phase angle curves between measured FRA data Fig. 10. Comparison of the FRA phase angle curves between measured FRA
and modelled from the combination of k, TFN1, TFD1 and TFD2. data and modelled from the combination of k, TFN1, TFN2, TFD1 and TFD2.

ECs, it is rational to represent the initial curve segments of the FRA curve
with single responses that capture the complexities of the EC structure.
As a result, it is imperative to differentiate the winding EC topologies
corresponding to the initial FRA curve section from other EC structures
since the subsequent TFs manifest double responses.
When TFN2, representing a double-zero response, is multiplied by the
previous TF equation, the resulting FRA curves are depicted in Figs. 9
and 10. However, the amplitude of the trough in these curves does not
align with that of the measured FRA curve, as the subsequent individual
TFs have yet to be introduced. The amplitude is determined by the
interaction of all individual TFs. Additionally, by examining the phase
angle curve, it is evident that the phase angle reaches 90◦ directly,
indicating that TFN2 exhibits a 180◦ asymptote, considering the previous
combined TF has a − 90◦ asymptote. The dashed blue phase angle
response depicted in Fig. 10 lacks a sharp reversal due to the absence of
subsequent individual TFs. Fig. 11. Comparison of the FRA amplitude curves between measured FRA data
TFD3 is a double-pole response, observed by dividing it from the and modelled from the combination of k, TFN1, TFN2, TFD1, TFD2, and TFD3.
previous TF model. The FRA curves for both amplitude and phase angle,
comparing measured and calculated using the TF model, are depicted in
Figs. 11 and 12, respectively. The modelled amplitudes of the first
trough and peak differ from the measurements due to the absence of
subsequent individual TFs. The second section line exhibits a highly
steep slope because the frequencies of the first trough and the first peak
are closely situated, resulting in a condensed response line. As illustrated
in Fig. 12, the phase angle forms a sharp cone, subsequently reaching
− 90◦ and indicating that TFD3 possesses a − 180◦ asymptote, consid­
ering the previous combined TF with a 90◦ asymptote. Meanwhile, an
observation reveals that the amplitude response gradually shifts towards
the negative direction. This shift is primarily due to the prevalence of the
double-pole response, which governs the high-frequency response

Fig. 12. Comparison of the FRA phase angle curves between measured FRA
data and modelled from the combination of k, TFN1, TFN2, TFD1, TFD2,
and TFD3.

independently of any influence from subsequent individual TFs. Viewing

this from the perspective of ECs, it becomes apparent that the double-
zero and double-pole responses signify distinct EC topologies
compared to the single-zero and single-pole responses.
TFN3 represents a double-zero response and is utilized in the for­
mation of the TF model. The comparisons between the measured and
modelled FRA curves for both amplitude and phase angle are illustrated
in Figs. 13 and 14, respectively. The modelled amplitudes of the first
peak and the first and second trough align with the FRA measurements,
Fig. 9. Comparison of the FRA amplitude curves between measured FRA data indicating that TFN3 accurately accounts for the amplitude proportion
and modelled from the combination of k, TFN1, TFN2, TFD1 and TFD2. from the preceding individual TFs, thereby achieving appropriate

6
Z. Zhang Electric Power Systems Research 228 (2024) 110055

Fig. 13. Comparison of the FRA amplitude curves between measured FRA data Fig. 15. Comparison of the FRA amplitude curves between measured FRA data,
and modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and TFD3
and TFD3. and the above TF with TFD2’s corner frequency changing from 27.5132 to 50.

Fig. 14. Comparison of the FRA phase angle curves between measured FRA Fig. 16. Comparison of the FRA phase angle curves between measured FRA
data and modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, data, modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and
and TFD3. TFD3 and the above TF with TFD2’s corner frequency changing from 27.5132
to 50.
amplitudes. By analyzing the FRA phase angle curve shown in Fig. 14, it
is evident that the phase angle returns to 90◦ , suggesting that TFN3 including k, TFN1, TFN2, TFN3, TFD1, TFD2, and TFD3, with the corner
possesses a 180◦ asymptote in reference to the previous TF, which has a frequency of TFD2 varying from 27.5132 to 50. TFD2 corresponds to a
− 90◦ asymptote. Likewise, an observation indicates a direct shift of the single-pole response, indicating the presence of a singular L or C element
amplitude response towards the positive direction. This shift can be within the EC of a TW. Increasing the corner frequency of TFD2 in­
attributed primarily to the dominance of the double-zero response, troduces discrepancies at the beginning of the first sections of the FRA
which controls the high-frequency response in isolation from any in­ curves, affecting both the amplitude and phase angle responses,
fluence stemming from subsequent individual TFs. respectively. However, the subsequent sections of the FRA curves
The purpose of systematically presenting the individual TF equations remain unaffected by this parameter change. Hence, it becomes clear
is to facilitate a comparison with the corresponding sections of the FRA
curves. This allows for the establishment of relationships between the
RLC network stages, individual TFs, and the FRA curve sections. Each
RLC network stage should be associated with its respective individual
TF. Precisely, the second, third, and fourth FRA curve sections corre­
spond to TFN2, TFD3, and TFN3, respectively. By establishing the link
between the parameters of the RLC network and the individual TFs, it
becomes possible to investigate winding faults by studying the indi­
vidual TFs. Furthermore, by manipulating the parameters of the TFs, it is
possible to observe how these modifications influence the characteristics
of the FRA curves.
Certain mechanical malfunctions in TWs may lead to the deviations
in the RLC elements of TW ECs. The parameters of the individual TFs, as
presented in Tables 1 and 2, are intricately connected to the RLC ele­
ments of TW ECs. Consequently, the investigation into the effects of RLC
element variations on FRA curves can be accomplished by observing
how the parameters of individual TFs influence the modelled FRA Fig. 17. Comparison of the FRA amplitude curves between measured FRA data,
curves. The blue dashed lines in Figs. 15 and 16 represent the frequency modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and TFD3
response curves generated by the combination of individual TFs, and the above TF with TFN2’s ω2 changing from 3.26e06 to 2.26e06.

7
Z. Zhang Electric Power Systems Research 228 (2024) 110055

that alterations in a solitary L or C element within the EC of a TW can

result in deviations in the FRA curve. This prompts an exploration into
the RLC element, which signifies the mechanical condition of TWs,
through the investigation of TFs.
In Fig. 17, it is altering the value of TFN2’s ω2 from 3.26e06 to
2.26e06 that results in the leftward displacement of the first trough and
downward shift of the first section of the FRA amplitude curve. This
change occurs due to the decrease in the corner frequency, which sub­
sequently reduces the amplitude of the preceding section. The corre­
sponding FRA phase angle curve is illustrated in Fig. 18, where the
second phase angle curve section also shifts to the left. It is important to
note that changing the value of ω2 in TFN2 has no impact on the fourth
sections of both the amplitude and phase angle curves. This remarkable
revelation highlights that the double zero response has the capacity to
dictate both the amplitude and frequency of the trough within the FRA
Fig. 19. Comparison of the FRA amplitude curves between measured FRA data,
curve, as well as govern the amplitude of the curve segments preceding
modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and TFD3
that trough. However, its sensitivity diminishes when it comes to the and the above TF with TFD3’s ω2 changing from 3.57e06 to 4.57e06.
subsequent curve sections.
TFD3, being a double-pole response, has an opposing effect on the
FRA amplitude curve compared to TFN2, which is a double-zero
response. When the value of TFD3’s ω2 is increased, it causes a right­
ward shift of the first peak and a reduction in the amplitude of the
previous TF. This results in a downward trend in the amplitude curve, as
depicted in Fig. 19. Similarly, in Fig. 20, increasing the corner frequency
of TFD3 is observed to displace the first peak of the phase angle curve to
the right.
When the square of the corner frequency of TFN3, which represents a
double-zero response, is increased from 5.38e06 to 6.38e06, it results in
a rightward shift of the second troughs in both the amplitude and phase
angle curves. This shift can be observed in Figs. 21 and 22, respectively.
Additionally, the increment in the square of the corner frequency causes
a slight rise in the preceding FRA amplitude curve. At the same time, the
first trough and peak are also elevated. Consequently, it can be readily
deduced that the double pole response exerts control over both the
amplitude and frequency of the peak within the FRA curve while also Fig. 20. Comparison of the FRA phase angle curves between measured FRA
data, modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and
influencing the preceding amplitude response at the peak point. Both the
TFD3 and the above TF with TFD3’s ω2 changing from 3.57e06 to 4.57e06.
double-pole and double-zero responses align with distinct EC configu­
rations of the TWs. The pertinent L and C elements can be explored by
examining the frequency responses of the TFs.
The examination of FRA curves by altering the corner frequencies of
individual TFs yields exciting conclusions. The first section of the
amplitude curve is influenced by both single-zero and single-pole re­
sponses, as evidenced by the variations in the FRA curves corresponding
to changes in these single responses. While Ref. [5] previously described
the low-frequency regions as being affected by core effects and an
inductive element, the aforementioned observations and analysis offer a
different explanation for the properties of low-frequency FRA curves. It

Fig. 21. Comparison of the FRA amplitude curves between measured FRA data,
modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and TFD3
and the above TF with TFN3’s ω2 changing from 5.38e06 to 6.38e06.

is apparent that both the shunt and series RLC components of the
winding EC possess the potential to impact the first section of the FRA
curve. Another notable finding is that the troughs are determined by the
corner frequencies of the double-zero responses, while the peaks are
influenced by the corner frequencies of the double-pole responses.
Pinpointing the characteristic EC configurations of TWs linked to these
Fig. 18. Comparison of the FRA phase angle curves between measured FRA
double responses enables the attainment of precise interpretations
data, modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and concerning resonance and anti-resonance. This is accomplished through
TFD3 and the above TF with TFN2’s ω2 changing from 3.26e06 to 2.26e06. the application of circuit analysis theory and the utilization of TF

8
Z. Zhang Electric Power Systems Research 228 (2024) 110055

and TFD14 only affects the adjacent curves, leaving the rest of the FRA
curves unchanged. Evidently, the presence of ripples can be attributed to
a pair of peaks and troughs with corner frequencies that are relatively
close to each other. This proximity prevents the curve section between
them from developing into a smooth and distinct line. Hence, the sig­
nificance of ripples is comparable to that of peaks and troughs, as they
mirror analogous EC structures of the TWs representing the peaks and
troughs.
Thus far, the derived individual TFs have been examined by
comparing them to the measured FRA curves. The investigation of in­
dividual TFs that represent the section lines on the FRA curve is
intriguing because it helps establish a deeper understanding of the re­
lationships among FRA measurements, FRA curves, TFs, and the rele­
vant RLC EC network. Furthermore, a remarkable level of agreement
within the measurement range is observed. The TF model, which is
Fig. 22. Comparison of the FRA phase angle curves between measured FRA
based on the analysis and processing of FRA data, offers a precise and
data, modelled from the combination of k, TFN1, TFN2, TFN3, TFD1, TFD2, and
reliable technique for reproducing the modelled FRA curves and
TFD3 and the above TF with TFN3’s ω2 changing from 5.38e06 to 6.38e06.
exploring the characteristics of FRA technology in power transformers.

6. Discussion

The TF evaluation involves the analysis of FRA data through MAT­

LAB 2023b software. The simulation results are compared with FRA
measurements to validate the individual TF equations presented in Ta­
bles 1 and 2 for single poles, double poles, single zeros, and double zeros,
respectively. In Tables 1 and 2, the TFs for single poles and single zeros
are directly derived from binomials with real roots in the denominator
and numerator in Eq. (2), while the TFs for double poles and double
zeros are obtained by multiplying two binomials with complex conju­
gate roots. The comprehensive TF characterizing the FRA curves is
constructed by combining these four components. All detailed features
of amplitude and phase angle curves can be elucidated using these in­
dividual TF equations. The investigation focuses on autotransformer
windings, and similar outcomes can be anticipated for other TW types.
Fig. 23. Comparison of the FRA amplitude curves sampled between measured In this specific case study, TF derivation is accomplished using MATLAB
FRA data, modelled from the combination of all listed individual TFs, and TF built-in functions such as invfreqs, freqs, and roots [28]. The MATLAB
from the combination of all listed individual TFs by removing TFN13 & TFD14.
code is employed to combine, rearrange, segment, and process the
measured FRA data. The FRA curves, modelled using the derived TFs,
exhibit remarkable accuracy when compared to the actual FRA
measurements.
The accuracy of the modelled low-frequency curves, segmented from
the initial point to the first trough, is primarily determined by the real
roots of the single poles and zeros. This segment of the amplitude curve
represents a transition line with a slope of − 20 dB, indicative of a
structure resembling a single pole. However, modelling it with a single
pole results in an apparent deviation at the starting point of the ampli­
tude curve. In contrast, double poles and double zeros are represented
by trinomials with two parameters, ω and ζ. The FRA curves can be
conceptualized as numerous section lines oscillating between troughs
and peaks. The ω parameters in double poles and double zeros determine
the frequencies of these troughs and peaks. The lengths of the section
lines are dictated by pairs of adjacent troughs and peaks. When a pair of
adjacent troughs and peaks are close to each other, and the section curve
Fig. 24. Comparison of the FRA phase angle curves sampled between measured is not fully developed, a ripple forms. This phenomenon occurs when the
FRA data, modelled from the combination of all listed individual TFs, and TF ω parameters in a pair of double poles and double zeros have close
from the combination of all listed individual TFs by removing TFN13 & TFD14. values. The ζ parameters play a crucial role in determining the sharpness
of the troughs and peaks. Higher values of ζ result in sharper troughs and
investigation methods by processing FRA data. peaks in the FRA curves.
As depicted in Figs. 5 and 6, there are noticeable ripples on the FRA The FRA curves encompass both amplitude and phase angle curves,
curves between 10 and 100 kHz. These ripples, although often over­ and these two aspects are interrelated. Unlike many previous studies
looked in previous references, hold significant importance alongside the that solely focused on amplitude curves, neglecting the investigation of
peaks and troughs. Each ripple represents a pair of double-zero and phase angle curves [26], this study delves extensively into phase angle
double-pole responses, as can be observed in Figs. 23 and 24. It is worth curves when deriving TW equivalent models using the derived individ­
noting that removing TFN13 and TFD14 from the TFs eliminates the ual TFs. While existing TW equivalent models have been tested using
ripple within the magenta rectangle. Interestingly, the deletion of TFN13 Simulink software, a noteworthy observation is that most of them cannot
be verified when compared to the measured phase angle curves.

9
Z. Zhang Electric Power Systems Research 228 (2024) 110055

Consequently, the concealed information provided by phase angle through the integration of the TF evaluation technique to investigate
curves should not be overlooked. In this research, both phase angle and fault cases related to TWs. Moreover, given the rapid advancements in
amplitude curves are deemed equally significant. In amplitude curves, FRA research, opportunities for interdisciplinary collaboration and the
sections with a slope of 20 dB are attributed to one energy storage exploration of emerging methodologies are abundant for future in­
element, either inductance or capacitance, forming a zero in Eq. (2) vestigations. For instance, the incorporation of AI algorithms to refine
based on the position of L or C in TW’s EC. Consequently, the asymptotic the parameters of derived individual TFs holds promise. By recognizing
line in phase angle curves approaches 90◦ . Conversely, a pole in Eq. (2) and proactively addressing the limitations highlighted in this study,
generates a section curve with a slope of − 20 dB, and the corresponding researchers can significantly contribute to the ongoing development and
phase angle curve section extends its asymptote to − 90◦ . Therefore, refinement of our understanding of FRA curves. This, in turn, facilitates
single poles and single zeros contribute to a ± 20 dB amplitude curve the implementation of more reliable predictions of mechanical condi­
section and a ± 90◦ trendline in the phase angle curve. Similarly, double tions, aligning with the advancements in FRA technologies.
poles and double zeros contribute to a ± 40 dB amplitude curve section Nonetheless, for the studied cases, the characteristics of FRA can be
and a ± 180◦ trendline in the phase angle curve. However, when effectively observed by comparing the FRA curves obtained from both
calculating the slope of a specific amplitude curve section or deter­ FRA measurements and the individual TFs. Suppose the TFs can accu­
mining the trendline of a phase angle curve section, the results must be rately represent the FRA measurements. In that case, the coefficients of
determined by superimposing the TF of that curve section with all pre­ the TFs in Eq. (3), such as the constant term k, z and p in single responses,
vious individual TFs. This clarifies the absence of ±40 dB amplitude and ω and ζ in double responses, can be utilized to characterize the
curve sections and ±180◦ trendlines in phase angle curves observed in features of the FRA curves and aid in identifying potential faults in TWs.
this study. Additionally, these coefficients of the TFs are related to the RLC ele­
To assess the deviation of modelled FRA curves from the measured ments of the EC network, which enables the derivation of the RLC EC
FRA data, the metrics of root mean squared error (RMSE) and mean network. The derived RLC EC network, based on individual TFs, will be
absolute error (MAE) are employed. These numerical indices, as pro­ further discussed in the following paper, and demonstrate its utility in
posed in the literature, aim to advance technologies for detecting me­ assessing the mechanical conditions of TWs.
chanical deformations within a power transformer, minimizing
subjectivity. Technically, these indices offer an indication of similarity CRediT authorship contribution statement
between two traces. In this context, RMSE and MAE are utilized to
validate the derived TFs. Given that FRA curve sections can be repre­ Z. Zhang: Conceptualization, Methodology, Software, Writing –
sented by individual TFs, indicating that the sections are determined by original draft, Visualization, Investigation, Validation, Writing – review
the parameters of these TFs linked to troughs and peaks, the similarity & editing, Project administration.
analysis is simplified by comparing these features. The computation of
RMSE and MAE utilizes MATLAB functions, rmse and mae [28], based Declaration of Competing Interest
on the amplitude and frequency values at the troughs and peaks between
the modelled and measured FRA curves. The computed RMSE and MAE The author declare that they have no known competing financial
values are 0.4807 and 0.2020 based on the data extracted from the interests or personal relationships that could have appeared to influence
troughs and peaks in Fig. 5, respectively. These remarkable results from the work reported in this paper.
the similarity analysis validate the derived TFs through the processing of
measured FRA data. Data availability

7. Conclusions and future work The data that has been used is confidential.

The TF equations and their coefficients are computed using MATLAB

programming code based on the power transformer FRA data. The Acknowledgment
correlation between the FRA curves and TF equations allows for novel
interpretations of the FRA curves. Furthermore, the utility of TF equa­ The author is grateful to National Grid for providing the experi­
tions and their coefficients in detecting TW faults will be explored by mental data.
deriving RLC EC networks from the derived TFs.
The TF of transformer FRA encompasses a constant term, k, along References
with individual TFs representing single-zero response, double-zero
response, single-pole response, and double-pole response. Each of [1] M. Zafeiropoulou, I. Mentis, N. Sijakovic, A. Terzic, G. Fotis, T.I. Maris, V. Vita,
E. Zoulias, V. Ristic, L. Ekonomou, Forecasting transmission and distribution
these TFs corresponds directly to an individual section line on the FRA system flexibility needs for severe weather condition resilience and outage
amplitude and phase curves, except the first section line of the low- management, Appl. Sci. 12 (14) (2022) 7334.
frequency FRA curves, which is represented by the collective influence [2] V. Vita, G. Fotis, V. Chobanov, C. Pavlatos, V. Mladenov, Predictive maintenance
for distribution system operators in increasing transformers, Reliab. Electron 12 (6)
of all the single responses and the constant term k. In summary, from the (2023) 1356.
perspective of individual TFs, the new interpretation of FRA reveals that [3] G. Fotis, V. Vita, T.I. Maris, Risks in the European transmission system and a novel
the first sections of FRA curves are influenced by the combined effects of restoration strategy for a power system after a major blackout, Appl. Sci. 13 (1)
(2022) 83.
all single-zero and single-pole responses. The resonances and anti- [4] E.P. Dick, C.C. Erven, Transformer diagnostic testing by frequency response
resonances observed in the FRA curves are predominantly governed analysis, IEEE Trans. Power Appar. Syst. 97 (6) (1978) 2144–2153.
by the presence of double poles and double zeros, respectively. Addi­ [5] IEEE Guide for the Application and Interpretation of Frequency Response Analysis
for Oil-Immersed Transformers, IEEE Std. C57.149-2012, (2013) 1-72.
tionally, the ripples observed in the FRA curves correspond to local [6] P. Picher, S. Tenbohlen, M. Lachman, A. Scardazzi, P. Patel, Current state of
peaks and troughs with corner frequencies that are closer in proximity to transformer FRA interpretation: on behalf of CIGRE WG A2, 53, Procedia Eng. 202
each other. (2017) 3–12. Presented at 4th International Colloquium "Transformer Research
and Asset Management".
While this study has yielded valuable insights into TF evaluation for
[7] Z. Zhao, Y. Chen, Y. Yu, M. Han, C. Tang, C. Yao, Equivalent broadband electrical
FRA technology, it is imperative to acknowledge its limitations. The circuit of synchronous machine winding for frequency response analysis based on
constraints arising from a lack of information regarding the malfunction gray box model, IEEE Trans. Energy Convers. 36 (4) (2021) 3512–3521.
of TWs underscore the necessity for further research in this domain.
Future studies stand to gain by addressing these limitations, particularly

10
Z. Zhang Electric Power Systems Research 228 (2024) 110055

[8] J. Duan, Y. He, X. Wu, A space hybridization theory for dealing with data [20] W.C. Sant’Ana, C.P. Salomon, G. Lambert-Torres, L.E.B. da Silva, E.L. Bonaldi, L.E.
insufficiency in intelligent power equipment diagnosis[J], Electr. Power Syst. Res. D.L. de Oliveira, J.G.B. da Silva, A survey on statistical indexes applied on
199 (2021), 107363. frequency response analysis of electric machinery and a trend based approach for
[9] S. Arumugam, Investigations on using white noise as a test signal for performing more reliable results, Electr. Power Syst. Res. 137 (2016) 26–33.
frequency response measurements on transformers, Electr. Power Syst. Res. 202 [21] V. Nurmanova, Y. Akhmetov, M. Bagheri, A. Zollanvari, B.T. Phung, G.
(2022), 107586. B. Gharehpetian, Confidence level estimation for advanced decision-making in
[10] A. Abu-Siada, O. Aljohani, Detecting incipient radial deformations of power transformer short-circuit fault diagnosis, IEEE Trans. Ind. Appl. 58 (1) (2021)
transformer winding s using polar plot and digital image processing, IET Sci. Meas. 233–241.
Technol. 12 (4) (2018) 492–499. [22] Y. Akhmetov, V. Nurmanova, M. Bagheri, A. Zollanvari, G.B. Gharehpetian,
[11] N. Hashemnia, A. Abu-Siada, S. Islam, Improved power transformer winding fault A bootstrapping solution for effective interpretation of transformer winding
detection using FRA diagnostics–part 2: radial deformation simulation, IEEE Trans. frequency response, IEEE Trans. Inst. Meas. 71 (2022) 1–11.
Dielectr. Electr. Insul. 22 (1) (2015) 564–570. [23] W.C. Sant’Ana, G. Lambert-Torres, L.E. Borges da Silva, E.L. Bonaldi, L.E. de
[12] J. Ni, Z. Zhao, S. Tan, Y. Chen, C. Yao, C. Tang, the actual measurement and Lacerda de Oliveira, C.P. Salomon, J.G. Borges da Silva, Influence of rotor position
analysis of transformer winding deformation fault degrees by FRA using on the repeatability of frequency response analysis measurements on rotating
mathematical indicators, Electr. Power Syst. Res. 184 (2020), 106324. machines and a statistical approach for more meaningful diagnostics, Elect. Power
[13] S. Wang, S. Wang, H. Feng, Z. Guo, S. Wang, H. Li, A new interpretation of FRA Syst. Res. 133 (2016) 71–78.
results by sensitivity analysis method of two FRA measurement connection ways, [24] K. Patel, N. Das, A. Abu-siada, S. Islam, Power transformer winding fault analysis
IEEE Trans. Magn. 54 (3) (2017) 1–4. using transfer function, in: Proceedings of the 2013 Australasian Universities
[14] H. Zhang, S. Wang, D. Yuan, X. Tao, Double-ladder circuit model of transformer Power Engineering Conference (AUPEC), IEEE, 2013, pp. 1–3.
winding for frequency response analysis considering frequency-dependent losses, [25] X. Lei, J. Li, Y. Wang, S. Mi, C. Xiang, Simulative and experimental investigation of
IEEE Trans. Magn. 51 (11) (2015) 1–4. transfer function of inter-turn faults in transformer windings, Electr. Power Syst.
[15] A. Shintemirov, W.H. Tang, Q.H. Wu, Transformer core parameter identification Res. 107 (2014) 1–8.
using frequency response analysis, IEEE Trans. Magn. 46 (1) (2009) 141–149. [26] H. Tarimoradi, H. Karami, G.B. Gharehpetian, S. Tenbohlen, Sensitivity analysis of
[16] P. Jayarathna, S. Ediriweera, R. Samarasinghe, R. Lucas, Modelling the lumped different components of transfer function for detection and classification of type,
parameter network of a power transformer using sweep frequency response location and extent of transformer faults, Measurement 187 (2022), 110292.
analysis, Aust. J. Electr. Electron. Eng. 20 (4) (2023) 309–321. [27] F. Nasirpour, M.H. Samimi, H. Mohseni, Evaluation of online techniques utilized
[17] K.H. Ibrahim, N.R. Korany, S.M. Saleh, Effects of power transformer high- for extracting the transformer transfer function, Scientia Iranica 26 (2019)
frequency equivalent circuit parameters non-uniformity on fault diagnosis using 3582–3591 (Special Issue on machine learning, data analytics, and advanced
SFRA test, Ain Shams Eng. J. 13 (4) (2022), 101674. optimization techniques in modern power systems [Transactions on Comp. Scie.
[18] Measurement of frequency response, IEC 60076-18 Edition 1.0 2012-07 Power Engi.Elec. Engi. (D)]).
transformers –, Part 18 (2012). [28] Users Guide with MATLAB, The Mathworks, Inc, 2023.
[19] Mechanical condition assessment of transformer winding s: guidance, FRA
standardization, further improvements, CIGRE WG A2.26, (2008).

11