1

Parameter Identification of the Transformer

Winding based on Least-Squares Method
E. S. Jin, L. L. Liu, Z. Q. Bo and A Klimek

correctness of the identification results are also verified by

Abstract-- The principles of the equivalent circuit principle simulation test. It has a great potential for the application to
and the magnetic property theory are not affected by the inrush online identification.
current and over-excitation, which enable protection to act
rapidly and reliably and implement ideas different from Index Terms-- Parameter identification; Least-squares
differential protection. However, the leakage inductance and method; Leakage and resistance; Transformer protection; Inrush
resistance of each winding are required as priori parameters. At current .
present, there is no accurate measurement method for the
transformer winding parameters. Therefore, it is very important I. INTRODUCTION

A
for a method to accurately estimate the leakage inductance and
resistance of the transformer. A new identification algorithm for
t present, the power system has gone through rapidly
transformer winding parameters is presented in this paper. Based development in China. The requirements for the power
on the electromagnetic relation between the primary and the system protection are becoming higher and higher with the
secondary sides, the equation for transformer flux balance can be build-up of more EHV transmission lines and installation of
derived based on the transformer model. Through the elimination large generating capacity units. More information is required
of the ratio of flux change with respect to time change from the to meet the need for accurate parameter identification. The
equation, the discrete flux balance equations can be obtained
through discrete process. Using the derivative of voltage and
study on a more reliable and sensitive transformer protection
current as the input and output variables, the leakage must integrate two state variables - voltage and current. The
inductances and resistances of every winding as the identification principles of magnetic property, the equivalent circuit theory
parameters, an identification model for transformer is built. Then and other new transformer protection techniques all take two
parameters can be identified using this model. The least squares variables – the current and voltage into account at the same
method based on the orthogonal decomposition is used to identify time. Better results can be expected with these techniques
the winding parameters in this paper. A modified orthogonal
method is adopted which is able to improve the numerical than that only using single electricity information. The
stability of Classical Gram-Schmidt orthogonal method in the principle of magnetic property technique is to identify the
calculation. The requirement for the input signal is also inrush current by making use of the feature that the
discussed. The input signal should include sufficient number of transformer has different flux characteristics during inrush
frequencies with sufficient exciting time. If a system has n and internal fault. The equivalent circuit theory gives the
parameters to be identified, continuous excitation signal should
interrelated equations of voltage and current of the primary
contain at least n/2 different frequency components. Considering
the actual operating conditions, this paper uses data from both and secondary windings. Using whether the expressions can
sides of transformer when it is switched on with no-load and be established to determine inrush current, the method is not
over-excitation to identify the parameters. The simulation results affected by inrush and over-excitation. As a result, new
of electromagnetic transient program EMTP show that the protection ideas different from differential protection can be
parameters identified by the method proposed in this paper are established [1, 2]. However, all the above methods need to
very close to the true values with vary small errors. When data of
know the leakage inductances and resistances of primary and
transformer switching on with no-load is employed, the error of
％
resistance is round about 1 and the error of inductance is less
secondary windings as priori parameters. At present, there is
％
than 0.1 . When data of over-excitation is employed, the error no accurate measurement method for these transformer
％
of resistance is under 3 and the error of inductance is about 0.3 winding parameters. Therefore, the precise calculation of the
％ . The characteristic of the multi-value and convergence of least transformer leakage inductances and resistances has become a
squares method are greatly improved by the method proposed in matter of urgency.
this paper. In a certain range, more accurate identification As to the transformer parameter identification, many
parameters can be achieved by using the longer data length. The research works have been conducted, among which the use of
recursive least squares (RLS) method to carry out online
This work was sponsored by AREVA T&D Automation & Information parameter identification of the transformer winding
Systems UK LTD.
E. S. Jin and L. L. Liu are with Electrical Engineering Institute, Northeast
parameters is able to accurately estimate leakage inductances
Dianli University, Jilin, China (e-mail: [email protected], [3, 4]. The RLS method is adopted in [3] to identify both
[email protected]). leakage inductances and resistances, and uses the degree of
Z. Q. Bo and A Klimek is with AREVA T&D Automation & Information the parameter variations to determine whether a system is in
Systems, Stafford, UK (e-mail: [email protected]).

©2008 IEEE.
 2

fault condition. This method does not require to detect the

inrush current and is simple and fast, but with larger error in
identified resistance. The literature [4] uses RLS method to
estimate values of leakage inductance but requires the actual
measurement of resistances, requiring high precision
measurement meter, which introduces many factors for errors.
In addition, the messages provided by new data are swapped
over by the old data when the RLS is used on the on-line Fig. 1. Single-phase double-windings transformer
identification. The information obtained from the new data
relatively declines, which result in the gradually loss of its self [u1(n∆T) + u1((n -1)∆T)] [u2 (n∆T) + u2 ((n -1)∆T)]
−K
amendment ability of the algorithm. Thus, estimation values 2 2
of parameters may deviate far from the true values and can [i1 (n∆T ) + i1 ((n -1)∆T )] [i (n∆T ) + i2 ((n -1)∆T )]
= R1 + KR2 2
2 2
not renovate, resulting the "saturation" phenomenon [5]. To
[i1(n∆T) −i1((n-1)∆T)] [i (n∆T) −i2 ((n -1)∆T)]
solve this problem, a forgotten factor is proposed in [6] to +L1 + KL2 2 (3)
∆T ∆T
overcome this deficiency. The basic thinking is to add a
forgotten factor to old data in order to reduce the amount of B. Parameters identification of transformer windings
old data and add new data information. But the introduction The identification model is established by using the current
of forgotten factor is more susceptible to noise interference, derivative as the input, the variable of left side of equation as
causing the increase of error. the output, and the leakage inductances and resistances of
A new transformer parameter identification method to every winding as the identification parameters. The least
solve the above mentioned problem is described in this paper. squares method based on the decomposition of the Cartesian
The requirement for the input signals is discussed. This is used to identify transformer parameters, which is able to
method is able to accurately identify both leakage inductances greatly improve the multi-valued and convergence of least-
and resistances of the transformer windings. Therefore, the squares method.
stability of the algorithm is greatly enhanced and the multi- Supposing that
value and convergence of least squares method is greatly [u1(n∆T) +u1((n-1)∆T)] [u2 (n∆T) +u2 ((n -1)∆T)]
improved. EMTP simulation and dynamic experimental y= −K
2 2
simulation prove that the identification error produced by this [i1 (n∆T ) + i1 ((n -1)∆T )]
algorithm is very small. α1 =
2
[i2 (n∆T ) + i2 ((n -1) ∆T )]
II. PARAMETERS IDENTIFICATION OF TRANSFORMER WINDINGS α2 = K (4)
2
[i1 (n∆T ) − i1 ((n -1)∆T )]
A. Transformer model α3 =
∆T
Fig. 1 shows a double-winding single-phase transformer [i2 (n∆T ) − i2 ((n - 1)∆T )]
α4 = K
model. According to the electromagnetic relation of primary ∆T
and secondary, transformer flux balance equation can be The Beyond linear equations are derived by
described by Cx=y (5)
d i1 dΦ Where x=[R1,R2,L1,L2]T is a four-dimensional vector,
U 1 = L1 + R1i1 + N 1 (1)
dt dt C=[ α 1(1) , α (1) (1) (1) ] and y is a known vector.
2 ,α 3 ,α 4
di2 dΦ The four listed Vectors in matrix C are all m dimension.
U 2 = L2 + R 2i 2 + N 2 (2)
dt dt The rank of C matrix is r ≤ 4. It can always be divided into
Where L1, L2, R1, R2, N1, N2 represent the leakage the following form:
inductance, resistance and turns of primary and secondary C QU ＝ (6)
winding, respectively; u1, u2, i1, i2 are the voltage and current Where Q is m × 4 order; the ranks of C and U are both r.
of primary and secondary winding. Φ is for the public flux of Thus,
both sides. x=UT(UUT)-1(QTQ)-1QTy (7)
Eliminate dΦ/dt in equations (1) and (2), supposing the A method using QU decomposition approach to C is given
turn ratio k=N1/N2, here:
Discrete flux balance equations are derived as: The classical Gram-Schmidt orthogonal method can not
offer good numerical stability. A modified orthogonal method
is used here. The following flowchart shows the processing
steps[7].
 3

Fig. 3. Simulation system single line diagram

If only the data from normal operation is used, the input

signal is too simple. At this time the transformer model can
not be sufficiently excited and the parameters can not be
identified. The data from inrush current and over-excitation
conditions contain obvious harmonic components and can be
used as the identification data to obtain more satisfactory
results, including switch on with load. However, considering
the actual operating conditions, the transformer is not allowed
to be switched on with load, therefore the fault data is difficult
to be obtained. This paper uses the date from both sides of
transformer when switching on with no-load to identify
parameters.
Table I shows the identification results using data of 100
cycles when switching on at 0°, 30°, 60° with no-load.
From the results of table I, we can see that the error of
％
resistance is round about 1 and the error of inductance is
％
less than 0.1 . So the error is very small when the data of
switching on with no-load is used.
Table II shows the identification results using over-
excitation data which is produced by raising the voltages to
1.4 and 1.5 times of the rated voltage.
From the data of table II, we can see that the error of
％
resistance is under 3 and the error of inductance is about
％
0.3 .
Fig. 2. QU orthogonal decomposition method flowchart Error is calculated by the following equation:
Error = [(Identification value-true value) / true value] * 100%
III. CASE STUDIES
The current and voltage of transformer is generated by TABLE I
THE IDENTIFICATION RESULTS WHEN USING THE DATA OF CLOSING WITHOUT
using the electromagnetic transient simulation program LOAD
(EMTP). Fig. 3 shows the single-phase representation of the (a) The identification results when using the data of switching
simulation system. The rated capacity of transformer is 55 without load at 0 degrees
MVA and rated voltage is U1/U2 = 88.912/12.702 kV. The real value identification value error
sampling frequency is 192 points per cycle, a second-order R1 (Ω) 0.940770 0.949537 0.9319 ％
Butterworth low-pass filter is used with a cut-off frequency R2 (Ω) 0.019200 0.019343 0.5113 ％
1920 Hz. L1 (H) 0.076405 0.076362 0.0534 ％
The input signal must meet certain conditions in order to L2 (H) 0.001559 0.001559 0.0383 ％
enable parameters to be identified. The minimum
requirements are: the whole process must be constantly (b)The identification results when using the data of switching without
load at 30 degrees
stimulated by the input signal during the time period of
real value identification value error
identification. That is to say, input signal must be sufficiently R1 (Ω) 0.940770 0.944287 0.3738 ％
incentive for all modals and the frequency spectrum of input R2 (Ω) 0.019200 0.019381 1.0372 ％
signal must cover the process spectrum during the process. L1 (H) 0.076405 0.076324 0.1016 ％
Therefore, the input signal should include a sufficient band of
frequencies with a certain excitation time. If a system has
L2 (H) 0.001559 0.001558 0.0246 ％
number n parameters to be identified, for the harmonic of (c) The identification results when using the data of closing
each frequency components of the input signal ωi, the without load at 60 degrees
corresponding frequency response results in a real R(ωi) and real value identification value error
an imaginary part Im(ωi). Two corresponding equations with R1 (Ω) 0.940770 0.934422 0.6745％
two unknown parameters can be derived. Therefore, in order R2 (Ω) 0.019200 0.019343 0.7468％
to identify n parameters, continuous excitation signal should L1 (H) 0.076404 0.076173 0.3063％
contain at least j≥ n/2 different frequency components. L2 (H) 0.001559 0.001559 0.0095％
 4

1
using the data of switching without load at 0 deg
TABLE II

/real value
0.9 using the data of switching without load at 30 deg
THE IDENTIFICATION RESULTS WHEN USING THE DATA OF OVER EXCITING using the data of switching without load at 60 deg
(a) The identification results at 140%rated voltage 0.8

real value identification value error

％
0.7

identification value
R1 (Ω) 0.940770 0.915586 2.6768
R2 (Ω) 0.019200 0.019768 2.9570 ％ 0.6

L1 (H) 0.076405 0.076267 0.1771 ％ 0.5

L2 (H) 0.001559 0.001562 0.1718 ％

0.4
Ll

real value
0.3
(b) The identification results at 150%rated voltage
0.2
real value identification value error
％

R1 (Ω)

100×
0.940770 0.934151 0.7034 0.1

R2 (Ω) 0.019200 0.019457 1.3407 ％

％
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
L1 (H) 0.076405 0.076162 0.3150
％
data windows 4
x 10
L2 (H) 0.001559 0.001563 0.2723
(b) The identification error of L1

Fig. 4 shows the errors of parameters identification when 6

the data of switching with no-load is used. Fig. 5 shows errors using the data of switching without load at 0 deg

/real value
using the data of switching without load at 30 deg
of parameters identification when the data of over-excitation 5 using the data of switching without load at 60 deg
data is used. In the figures, the abscissa axis signifies

sampling points, the longitudinal coordinate shows the error

identification value
4
of data identification results from the beginning to the point.
The results show that the parameters identified by the method 3
proposed by this paper are very close to the true values.
Errors can be very small. In a certain range, using the longer

real value

2 R2
data length can gain more accurate identification results.

1
3
100×

using the data of switching without load at 0 deg

/real value

using the data of switching without load at 30 deg

0
2.5 using the data of switching without load at 60 deg 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
data windows 4
x 10


identification value

2 (c) The identification error of R2

0.4
1.5 using the data of switching without load at 0 deg
/real value

0.35 using the data of switching without load at 30 deg

using the data of switching without load at 60 deg
real value

1 R1 0.3


identification value

0.5 0.25
100×

L2
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
data windows 4
0.15
x 10
real value

(a) The identification error of R1 0.1

100×

0.05

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
data windows 4
x 10

(d) The identification error of L2

Fig. 4. The identification errors when the data of switching with no-load
 5

the error of Rl
10
5
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4
x 10
the error of Ll

0.5
0
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Fig. 6. Dynamic power system model
4
x 10
the error of R2

20
TABLE III
10
DYNAMIC MODULUS TEST RESULTS
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

R1 (Ω) R2 (Ω) R1+R2’ (Ω)

4
x 10 Error (%)
the error of L2

0.5
0 0.862421 0.650243 3.463397 3.25
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0.726261 0.652206 3.335088 0.56
data windows
x 10
4 0.750473 0.679046 3.466658 3.35
0.811048 0.630982 3.334979 0.57
(a) Identification results at 140%rated voltage
TABLE IV
the error of Rl

40 DYNAMIC MODULUS TEST RESULTS

20
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 L1 (H) L2 (H) L1+L2’ (H) Error (%)
4
x 10 0.024228 0.004817 0.043498 3.31
the error of R2 the error of Ll

1
0.024888 0.004819 0.045520 2.47
0.5
0 0.023867 0.004867 0.043336 2.93
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4 0.025183 0.004512 0.043234 2.68
x 10
50

V. CONCLUSION
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4
x 10
This paper presents a new parameter identification method
the error of L2

0.5 using the least squares method based on orthogonal

0
-0.5
decomposition to identify the leakage inductances and
0 0.2 0.4 0.6 0.8 1 1.2
data windows
1.4 1.6 1.8
4
2
resistances of transformer winding. The requirements for
x 10
input signals are analyzed and the method is able to accurately
(b) Identification results at 150% rated voltage identify the resistance and leakage values of the transformer.
Fig. 5. The identification errors when use the data of over exciting EMTP simulation studies have verified the accuracy of this
algorithm with very small error. When these identified
IV. EXPERIMENT RESULTS USING REAL TRANSFORMER parameters are applied to the algorithms of equivalent circuit,
This algorithm is tested by transformer data from the magnetic properties principle and the equivalent instantaneous
dynamic simulation laboratory of the Northeast Dianli inductance, which use both voltage and current to
University. The basic diagram of the experimental system is discriminate inrush current, not only the sensitivity for
shown in Fig. 6. The experimental transformer is a Y-Y internal fault detection and correct operation rate can be
connected three single-phase transformer group. Single-phase improved greatly, but also the transformer protection is able
parameters: the rated capacity of 2KVA, rated voltage of to avoid the influence of inrush current and can operate
U1/U2 = 440V/220V, short-circuit voltage of 13.59%; short- rapidly and reliably. This algorithm lays a good foundation for
circuit loss of 0.07%. The sampling frequency is 64 points per online identification of transformer parameters and offers
cycle. Obtained from short-circuit tests, transformer winding great potential for practical application.
parameters are
r1+r2’=3.3541Ω L1+L2’=0.0419H. ， [1]
VI. REFERENCES
Z. G. Hao, B. H. Zhang, and Y. L. Chu, "The present state and
The identification results obtained by the algorithm are
development of the inrush current identification technology of
very close to the real values. The results are shown in table III transformer," Transformer, vol. 42(7), 2005.
and table IV respectively. [2] Y. Z. Tang, and J. L. Xu, "Analysis and evaluation of several current
The increase in test error for the dynamic tests is mainly discrimination of the inrush current in the transformer protection," Power
Automation Equipment, vol. 1, pp. 1-7, 1995.
from the following reasons: [3] X. F. Xiong, X. L. Deng, and D. You, "The computer-based transformer
(1) Error produced by VT and CT; protection based on parameter identification," Power System Automation,
(2) Error caused by channel of fault recorder; vol. 23(11), pp. 18-21, 1999.
[4] Z. G. Hao, B. H. Zhang, and Y. L. Chu, "The transformer protection
(3) Rounding error generated when converting graphic files
principle based on the balance equation of the equivalent circuit," China
into data files; Electrical Engineering Journal, vol. 26(10), pp. 67-72, 2006.
(4) Error produced by measuring meter during short-circuit [5] L. D. Pan, and Y. D. Pan, System identification and model construction,
tests. Beijing: Chemical Industry Press, 2004.
 6

[6] The transformer protection based on adaptive fault component and Li-Li Liu is studying for her M.S. degree at Northeast Dianli University, China.
parameter identification, Chongqing: Chongqing University, 1999. Her research interest is power system protection.
[7] Y. C. He, Y. C. Su, and X. S. Bao, Concise Guide to computational
mathematics, Shanghai: People's Education Press, 1980. Zhiqian Bo received his BSc degree from the Northeastern University, China in
[8] B. M. Ge, X. H. Yu, and X. Z. Wang, "A new algorithm on the 1982 and PhD degree from The Queen's University of Belfast, UK in 1988
discrimination of the inrush current based on the equivalent transient respectively. Presently, he is with AREVA T&D Automation & Information
inductance," Power System Automation, vol. 28(7), 2004. Business.
Andrew Klimek is a licensed Professional Engineer with B.Sc., M.Sc. degrees
VII. BIOGRAPHIES in Electrical Engineering. He has 35 years international experience in the
electric power industry. His experience includes a number of positions ranging
En-Shu Jin received the B.S. and M.S. degrees from Northeast DianLi from university teaching and consulting, to engineering and business
University, China in 1994 and 1999, respectively, and the Ph.D degree from management. Andrew has served as a project manager, marketing manager and
Chonbuk National University, Korea in 2005. Now, she is an associate professor general manager and has executed various power systems projects. Currently
at Northeast Dianli University, China. Her research interest is power system Andrew is the technical Director of the AREVA T&D Automation &
protection. Information Business.