Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Department of Electrical Engineering and Electronics The University of Liverpool, Brownlow Hill, Liverpool L69 3GJ, UK
a r t i c l e i n f o
Article history:
Received 10 April 2008
Received in revised form 19 June 2009
Accepted 2 March 2010
Available online 13 April 2010
Keywords:
Transformer winding
Mathematical model
Bacterial swarming algorithm
Frequency response analysis
Parameter identication
a b s t r a c t
The paper discusses an improved modelling of transformer windings based on bacterial swarming algo-
rithm (BSA) and frequency response analysis (FRA). With the purpose to accurately identify transformer
windings parameters a model-based identication approach is introduced using a well-known lumped
parameter model. It includes search space estimation using analytical calculations, which is used for the
subsequent model parameters identication with a novel BSA. The newly introduced BSA, being devel-
oped upon a bacterial foraging behavior, is described in detail. Simulations and discussions are presented
to explore the potential of the proposed approach using simulated and experimentally measured FRA
responses taken from two transformers. The BSA identication results are compared with those using
genetic algorithm. It is shown that the proposed BSA delivers satisfactory parameter identication and
improved modelling can be used for FRA results interpretation.
2010 Published by Elsevier B.V.
1. Introduction
Power transformer is a major apparatus in a power system, and
its correct functioning is vital to system operation. It is therefore
very necessary toclosely monitor their in-service behavior, inorder
to avoid catastrophic failures and costly outages and improve the
management of maintenance and servicing.
Among various techniques applied to power transformer condi-
tion monitoring, frequency response analysis (FRA) is suitable for
reliable winding displacement and deformation assessment and
monitoring. It has been established upon the fact that frequency
response shape of a transformer winding in high frequencies
depends on changes of its internal distances and proles, which
are concerned with its deviation or geometrical deformation [1].
However, the interpretation of FRA data is mainly conducted
manually by trained experts. Measured FRA traces are compared
with the references taken from the same winding during previous
tests or from the corresponding winding of a sister transformer,
or from other phases of the same transformer. The shifts in reso-
nant frequencies and magnitude of FRA traces are believed to be
indicators of a potential winding deformation. However, the ques-
tion of potential deformation location in a winding is still required
to be investigated [2].
A range of research activities have been undertaken to uti-
lize FRA in the development of suitable mathematical models of
Corresponding author. Tel.: +44 151 7944535; fax: +44 151 7944540.
E-mail address: qhwu@liv.ac.uk (Q.H. Wu).
transformer windings. Consideringthesimpliedequivalent model
of transformer winding, various experimental research was per-
formed with the purpose to observe the model behaviors in the
frequency domain [3,4]. A winding equivalent model and an iden-
tication method of transformer equivalent circuit were proposed
in [5,6], where equivalent circuits of transformer winding for the
low, medium and high frequency ranges were discussed and its
frequency responses were compared with experimental data in
order to identify the models parameters. These models represent
the overall windings by combinations of single lumped elements:
inductances, resistances and capacitances. This allows estimat-
ing only the overall winding parameters in a particular frequency
range, which makes these models unsuitable for deformation anal-
ysis of each winding section.
Thecalculationof internal parameters plays animportant part in
accurate simulations of transformer winding frequency behavior.
Modelling of a real winding in order to obtain frequency responses,
being close to experimental ones, is an extremely complex task
since a detailed transformer model must consider each turn or sec-
tion of a winding separately. The reason is the uctuation of real
winding parameters such as inductances and resistances per turn
length as well as interturn capacitances. The insulation property
deviation should also be taken into account, which is frequency
dependent.
In [7] efcient procedures to calculate turn self inductances,
mutual inductances and capacitances were proposed which
demanded additional experimental tests and knowledge of geo-
metric and physical characteristics of a transformer. A transfer
functionapproachis usedin[2] tostudythe discriminating changes
0378-7796/$ see front matter 2010 Published by Elsevier B.V.
doi:10.1016/j.epsr.2010.03.001
1112 A. Shintemirov et al. / Electric Power Systems Research 80 (2010) 11111120
Fig. 1. Equivalent circuit of single-phase power transformer windings.
introduced into a winding physical model. In [811] analytical
expressions areusedtoestimateparameters of anequivalent model
basedonthe geometryof windings. The well-knownnite-element
methodwas appliedin[12,13] for more precise calculationof wind-
ing parameters for an equivalent circuit model. These techniques
show higher degree of accuracy compared with experiment mea-
surements.
However, in industry measurements it is not always possible to
conduct additional tests for precise measurements of transformer
geometry or insulation parameter estimation. Recently, evolution-
aryalgorithms wereutilizedtoovercomesuchdifculties, that offer
a way to identify model parameters using limited measurement
data. Regarding transformer winding modelling, at rst, two simi-
lar simplied winding model parameter identication approaches
using particle swarm optimizer (PSO) [14] and genetic algorithms
(GAs) [15] were proposed in [16,17] respectively. However, only
simplied one-winding lumped parameter models were consid-
ered in these work for parameter identication. The questions of
initial estimationof the model parameters to establishsearchspace
for the evolutionary algorithms were not discussed either.
A variety of well-established biologically inspired computa-
tional methodologies has emerged in the past a few years, such
as GAs, PSO, evolutionary programming (EP) [18], bacterial forag-
ing algorithm (BFA) [19], etc. However, data processing in these
algorithms may be time consuming, especially when a large num-
ber of multi-dimensional variables need to be optimized. Thus, it
leads to a slowconvergence rate and reluctant application in many
problems involving a large number of parameters to be optimized,
primarily because of the huge computational burden imposed. A
keyadvanceinthis eldwill thereforebemet byasignicant reduc-
tion in the computational time-costs whilst further improving the
efciencyof global researchcapabilities of thesealgorithms [20,21].
In this paper an improved modelling of transformer windings is
presentedbyusing a model-basedidenticationapproachtoderive
the parameters of transformer windings models. The approach,
rstly introduced in [16,17], is further modied with a novel bac-
terial swarming algorithm (BSA) [21,22], presented and utilized to
undertake parameter identication. The newly introduced BSA is
based on BFA, which incorporates ideas fromthe modelling of bac-
terial foraging patterns [22]. Simulation studies and discussions
are presented to explore the potentials of the proposed modelling
approach.
The remainder of this paper is organized as follows: Section 2
describes a lumped parameter mathematical model of transformer
windings utilized in this study, then in Section 3 the analyti-
cal expressions for model parameter estimation are presented.
BSA and a model-based identication approach are introduced in
Section 4. Subsequently, results of parameter identication using
simulated and experimentally measures frequency responses from
two transformers are presented and discussed in Section 5. Finally,
conclusions are given in Section 6.
2. Lumped parameter model of transformer windings
Since resonances of an FRA trace are related to the values
of capacitances and inductances within a transformer winding,
lumped-parameter equivalent circuit models have beenwidely uti-
lized to analyze frequency domain behaviors. Each section of an
equivalent circuit usually represents one or a few discs in the case
of a disc type windings as well as one or a few turns for helical
type windings [911]. Therefore, despite of the model simplicity in
comparison with those based on traveling wave theory [20,23,24]
and multiconductor transmission line theory [2528], it retains a
physical veracity and can be useful for frequency response simula-
tion in a restricted frequency diapason up to 1 MHz. Fig. 1 shows a
typical equivalent circuit of single-phase power transformer wind-
ings [812], where the following notations for section parameters
of a high voltage (HV) and a low voltage (LV) windings are in use:
K
HV
and K
LV
series capacitances of HV and LV winding sections
A. Shintemirov et al. / Electric Power Systems Research 80 (2010) 11111120 1113
C
HV
and C
LV
ground capacitances of HV and LV winding sections
C
HVLV
capacitances between HV and LV winding sections
C
HVLV
conductance between HV and LV winding section
g
HV
and g
LV
series conductances of HV and LV winding sections
C
HV
and C
LV
ground conductances of HV and LV winding sections
L
HV
and L
LV
inductances of HV and LV winding sections
R
HV
and R
LV
resistances of HV and LV winding sections
M
km
mutual inductance between the k
th
and m
th
winding sec-
tions
n number of sections in one winding
The mathematical descriptionof the model infrequencydomain
is usually given in a matrix form of nodal equations applying the
rst and second Kirchhoffs laws [811]:
Y U = A I +QU
0
,
Z I = A
1
U+PU
0
,
(1)
where vectors Uand I represent node voltages and branch currents
respectively and voltage U
0
denotes an input sinusoidal signal.
Each element of the admittance matrix Y is a combination of
admittances, corresponding to g K and C C parallel branches
respectively as shown in Fig. 1. Using matrix notations for the
capacitance and conductance matrices, Y = jC +G, where is an
angular frequencyandj denotes theimaginaryoperator. Thebranch
impedance matrix Z = jL +Rconsists of the self- and mutual sec-
tional inductances L and M, being combined into the matrix L, and
equivalent section resistances R, building the matrix R [8].
The incidence matrix A consists of 1, 0 and 1 and serves to
link nodal voltages with branch currents. Matrices Q and P are
formed as a component vector of Y and Amatrices correspondingly
inaccordancewithterminal connections of thewindingmodel such
as external voltage source imposition, node grounding, internodal
connections, etc. [911,29].
FromEq. (1), the branch currents and nodal voltages vectors can
be expressed as [8,9,11,29]:
U = (Y +AZ
1
A
1
)
1
(AZ
1
P +Q)U
0
,
I = Z
1
(A
1
U+PU
0
).
(2)
An output signal can be chosen arbitrarily fromthe variable vec-
tor Uor I in order to obtain the transfer function H(j) of a winding
in frequency domain, i.e. the ratio of nodal voltages and inductive
branch currents with respect to the applied input voltage.
3. Estimation of model parameters
Model parameters are usually estimated using physical
dimensions of a winding. In practice some simplication and
approximations of winding geometrical structures are accepted
which allow to apply analytical formulae [8,9]. On the other
hand, geometrysimplications canbe avoidedusingnite-element
method [12,13] for parameter calculation. In both cases the
frequency dependent behavior of resistive elements should be
accounted as well as frequency dependent insulation properties.
In this study, the analytical expressions for initial estimation of the
model parameters are presented.
3.1. Capacitance and conductance
One of the common methods to calculate the ground and inter-
winding capacitances C between windings, tank (or core) is to use
the expression for cylindrical capacitance having an axial height of
the model section [9,10,13].
On the other hand, the evaluation of series capacitances K
depends on winding types. For instance, for a disc type winding the
series capacitance is determined by the amount of stored energy
in the disc, which can be estimated by assuming the equal voltage
drop across each disc and the existence of equipotential surfaces
in the interdisc space as stated in [9,10,13,30]. Transformer insula-
tionmaterials at capacitance calculationare usually representedby
effective dielectric permittivities calculated as proposed in [9,13]
usingdielectric material referencesources or additional test results.
It is known that the insulation conductivity is frequency depen-
dent due to dielectric losses characterized by tan. Therefore,
expressions for series and ground conductances g and G can be
obtained using the well-known formulae [27,28]:
C = C tan
g
,
g = K tan
s
,
(3)
where tan
g
and tan
s
are the effective loss tangents of the insula-
tion between winding and ground, and the intersection insulation
respectively.
3.2. Inductance and resistance
Research reported in [8,29] assumes that in the high frequency
region above 10 kHz the core effect is not signicant and can be
neglected. Hence, the self and mutual inductances of model were
calculated using air-core case expressions. This method showed a
high degree of accuracy in comparison with the results performed
on several experimental transformers with the core removed and
substituted by a hollow metal cylinder.
However, according to [13] the core effect has to be taken into
account during inductance calculation in order to accurately model
practical transformers. This can be achieved using the analytical
expressions derived by Wilcox [31,32] as described below, which
are adopted in this study.
In Fig. 2, two coils representing the k
th
and m
th
sections of a
transformer winding are illustrated. These coils have N
k
and N
m
turns respectively wounded concentrically on a magnetic core of
radius b at intersection distance z. Each coil is characterized by the
average a, internal u
1
and external u
2
radii respectively, and cross-
section weight wand height h. The mutual impedance between the
coils is given as [32]:
Z
km
= sL
km
+Z
1(km)
+Z
2(km)
, (4)
where s denotes the Laplace transform operator, L
km
corresponds
to the mutual inductance upon the air-core assumption, Z
1(km)
rep-
resents the impedance due to the ux conned to core and Z
2(km)
is the impedance owing to leakage ux upon introducing the core.
The secondandthirdterms of Eq. (4) are thoroughly denedin[32].
Regarding the rst term, L
km
, it is proposed to use the following
approximate formula [32]:
L
km
=
0
N
k
N
m
_
u
k
u
m
2
k
__
1
k
2
2
_
K(k) (k)
_
, (5)
where
0
is a free space permeability, K(k) and (k) are complete
elliptic integrals of the rst and the second kinds respectively, and
k =
_
4u
k
u
m
z
2
+(u
k
+u
m
)
2
. (6)
However, a more precise calculation of the mutual inductances
can be achieved by the following expression [31]:
L
km
= 2
0
N
k
N
m
u
k
u
m
_
0
l
1
(u
m
)K
1
(u
k
) cos(z)d, (7)
where K
1
and l
1
are the modied Bessel functions.
In both Eqs. (5) and (7) in the case of self impedance calculation,
e.g. Z
kk
, z = 0.2235(h +w) [32].
1114 A. Shintemirov et al. / Electric Power Systems Research 80 (2010) 11111120
Fig. 2. Two coil sections on a core.
The calculation of winding resistance is the one of the major
challenges due to eddy current effect in winding conductor and
core. There are a lot of methods being proposed and utilized
[8,11,12] for the resistance calculation, among them, the Dowells
approach [33] is one of the most referenced:
R= R
dc
z
_
sinh(2z) +sin(2z)
cosh(2z) cos(2z)
+
2(p
2
1)
3
sinh(z) sin(z)
cosh(z)+cos(z)
_
,
(8)
where R
dc
is the DC resistance of one-winding section, p is the
number of layers in the section and
z=
_
4
_
3]4
d
3]2
t
1]2
, (9)
in which d is the conductor equivalent diameter and t is the dis-
tance between the centers of two adjacent conductors. The skin
penetration depth can be found as follows:
=
_
2
r
o
, (10)
where o and
r
are the conductor conductivity and relative per-
meability, respectively.
4. Model-based identication approach with bacterial
swarming algorithm
4.1. Bacterial foraging algorithm
In the past a few years, the development of evolutionary algo-
rithms received great attention in the computational intelligence
community worldwide. The BFA [19] is one of the emerging opti-
mization methods, utilizing an optimization model for E. coli
bacterial foraging to mimicry the self-adaptability of bacteria in
the group searching activities.
In general, an E. coli bacteriumhas a control systemthat enables
it to search for food molecules (nutrient) and try to avoid noxious
substances. This activity of individual meretriciously agellated
bacteria, which is called chemotaxis, can be described in terms
of run intervals during which the microbe swims approximately
in a straight line interspersed with tumbles so that the organism
undergoes a randomreorientation, and it alternates between these
two modes of operation in its entire lifetime.
E. coli bacteria also demonstrate a particularly interesting
groupbehaviorcellcell communication, whichis a process that
allows bacteria tosearchfor similar cells intheir close surroundings
using secreted chemical signaling molecules called autoinducers.
Other bacteria release the same autoinducers in response. One-
cell organisms in effect become multi-cellular organisms and can
respond together [34].
A BFA optimization model operates by a population (set) of
bacteria executing processes of chemotaxis, reproduction and dis-
persion. A chemotactic process in BFA consists of Tumble and
Run steps for each bacterium in the population, and a step
tness is evaluated at each step of the process. The total t-
ness of each bacterium is calculated as the sum of the step
tness during its life, which is obtained after all chemotactic
steps.
In the reproduction process, all bacteria are sorted in reverse
order according totheir tness. Only the rst half of populationsur-
vive and a surviving bacteriumsplits into two identical ones, which
occupy the same positions in the environment at 1
st
step. Thus,
A. Shintemirov et al. / Electric Power Systems Research 80 (2010) 11111120 1115
the population of bacteria remains constant in each chemotactic
process.
The dispersion process prevents the optimization process to be
trapped around local optima or initial positions. In BFA, the disper-
sion event happens after a xed number of reproduction processes.
A bacterium is chosen, according to a preset probability p
ed
, to
be dispersed and moved to another position within the environ-
ment. BFA adopts a strategy of individual-based attraction and
repellent for cellcell communication, which largely increases
its computational complexity [19].
BFA is claimed to have a satisfactory performance in optimiza-
tion problems and has been applied to power ow optimization
[35] and stabilizers design [36] for power systems. However, the
algorithm has limitations, e.g. it is unable to obtain satisfactory
results in a certain range of optimization problems, in particular,
high dimensional and multi-modal problems, as some parts of the
foraging process are articially set in such a way that the charac-
teristics of the problem (landscape) were ignored [21].
Comparative studies have demonstrated that a recent develop-
ment uponthebacterial foragingstudy, BSA, outperforms BFAinthe
aforementioned problems. The numerical comparison study [21]
has shown that BSA has demonstrated a superior performance in
comparison with some popularly used algorithms, such as PSOand
Fast Evolutionary Programming. Therefore, this algorithm is used
in this study to perform intelligent optimization with the purpose
to identify transformer winding model parameters andis described
in detail in the next subsection.
4.2. Bacterial swarming algorithm
The main difference between BSA and BFA is the absence of the
reproduction process in BSA. All bacteria are kept in the population
with only their positions updated in the search domain accord-
ing to their tness values. The BSA model executes a combination
of both chemotactic and group-based attraction and dispersion
processes, which is described below in details.
In the chemotactic process, a unit walk with random direction
represents a Tumble and a unit walk with the same direction of
the last step indicates a Run. The process consists of one step of
Tumble and followed by N
s
steps of Run, depending on the variation
of environment.
In the process of Tumble, the position of the i
th
bacteriumcan be
represented as
0
i
(j +1, r) = 0
i
(j, r) +C
i
(j)
i
(j), (11)
where 0
i
(j, r) indicates the position of the i
th
bacterium at the j
th
chemotactic step in the r
th
iteration loop; (j) is the direction angle
of the j
th
step for bacterium i, and it is a random angle generated
within a range of [0, 2] [21]. C
i
(j) is the length vector of a unit walk
for the i
th
bacterium at the j
th
chemotactic step in the r
th
iteration
loop, which is dened as follows [21]:
C
i
(j) =
_
C
init
8 : if Tumble
D
1
r
1
8, : if Run
, (12)
where C
init
is the step size for a unit walk, D
1
is a constant, r
1
is a
randomnumber, r
1
[0, 1], and 8 is the length vector of the bound-
aries of the search domain, depending on a particular optimization
problem.
The tness of the i
th
bacterium at the j
th
chemotactic step is
represented by j
i
(j, r). If j
i
(j +1, r) is better than j
i
(j, r), then the
process of Run follows, which can be represented by:
0
l+1
i
(j +1, r) =
0
l
i
(j +1, r) +C
i
(j)
i
(j) (13)
where
0
l
i
denotes the position of the i
th
bacterium in the l
th
step of
Run, 1 l N
s
, 1 j N
c
, 1 r N
r
. The notation N
s
corresponds
to the swimlength limits when it is on a gradient, N
c
is the number
of chemotactic steps per bacteria lifetime and N
r
is the number
of iteration steps. This process continues until j
l+1
i
(j +1, r) is not
better than j
l
i
(j +1, r) [21].
The group-based attraction and dispersion process is adopted
to mimicry cellcell communication. The individual with most
energy (best tness value) gainedinN
c
chemotactic steps is dened
as the best cell and its position 0
p
(r) is kept for updating the
positions of other bacteria in the next selection process. A certain
percentage of bacteria are involved in an attraction action, which
are selected according to a probability p
a
. Based on their current
positions and the global best position, the positions of bacteria
being attracted are recalculated as follows [21]:
0
i
(1, r +1) = 0
i
(N
c
, r) +r
2
D
2
(0
p
(r) 0
i
(N
c
, r)), (14)
where D
2
is a constant and r
2
a randomnumber, r
2
[0, 1]. The rest
bacteria are dispersed to positions around the best individual with
a randomly chosen mutation step and a mutation angle, using the
following equation:
0
i
(1, r +1) = 0
p
(r) +r
3
8j (15)
where r
3
is a random number, r
3
[0, 1] and j is a random angle
chosen from [0, 2].
In summary, a BSA process can be briey expressed in the form
of a sequence of the following operations:
(i) Random generation of the initial population.
(ii) Performing the chemotactic process for tness evaluation of
each bacterium in the population.
(iii) Performing the group-based attraction and dispersion pro-
cess.
(iv) Repeating steps iiiv until a termination criterion is met.
(v) Presentation of the best bacteriumin the population as the BSA
output.
4.3. Model-based identication approach
The model-based learning approach is based on searching of the
optimal model parameters by minimizing the difference, i.e. tness,
between reference frequency responses and simulated model out-
puts. It is achieved by measuring the errors between the original
responses and the model outputs. Therefore, for each individual
(bacterium) of a population in BSA, its total tness value is given as
follows:
min
S
j=1
||H
0
(
j
) H(
j
)|| w
j
, (16)
where H
0
(
j
) and H(
j
) R
1
are the reference and simulated with
the identied parameters frequency responses at frequency
j
, j =
1, . . . , S, where S is the number of frequency points involved in BSA
learning process and w
j
is the relative weight of the j
th
point.
Due to iterative nature of evolutionary algorithms, processing
a large number of data points can greatly slow down a learning
process. In the case of FRA, frequency responses are characterized
mainlybyresonant andantiresonancefrequencies andcorrespond-
ingmagnitudevalues. Therefore, as proposedin[17], thedimension
of processed FRAdata can be reduced by selection of points of reso-
nance andantiresonance andits vicinities for more speedy analysis,
which are weighted accordingly.
The following steps are performed for parameter identication
of the transformer windings model: