Received: 10 November 2015 Revised: 22 May 2016 Accepted: 10 July 2016

DOI 10.1002/etep.2257

RESEARCH ARTICLE

Neural networks applied to the design of dry‐type transformers:

an example to analyze the winding temperature and elevate the
thermal quality
Marco A. Ferreira Finocchio1 | José Joelmir Lopes2 | José Alexandre de França2* |
Juliani Chico Piai2 | José Fernando Mangili Jr2

1
Universidade Tecnológica do Paraná, Cornélio
Procópio, Paraná, Brazil
2
Universidade Estadual de Londrina,
Departamento de Engenharia Elétrica, Laboratório
de Automação e Instrumentação Inteligente,
Londrina, Paraná, Brazil
Correspondence
José Alexandre de França Departamento de
Engenharia Elétrica, Universidade Estadual de
Londrina, Caixa Postal 10011, Londrina, Paraná,
Brazil.
Email: jaf@jaf.eng.br
Summary
Much of the cost of manufacturing transformers is related to the complexity in the
development of its project, which involves many variables, such as different types
of materials, methods, and manufacturing processes. Also, it is notoriously difficult
to establish standards that relate the transformer characteristics with these design var-
iables, since each, almost always, variable is calculated empirically. One of the most
important design variables is the internal temperature of the transformers, which
directly influences the lifetime of such equipments. So, a computational tool is under
development whose purpose is to automate the design of cast resin dry‐type trans-
formers and minimize the time taken for its completion, by means of artificial neural
networks. In this article, we present some results of this tool, which relate some geo-
metrical parameters of the specific transformer with the temperature of its windings.
For this, the winding losses and total losses are also estimated. The training of the
neural network was done with the test data of about 300 dry‐type transformers from
the same manufacturer, so, with the same constructive features. Nevertheless, our
results show that the technique is promising because the neural networks were well
fitted and its results present errors lower than 1% when compared with data from tests
with real transformers. Evidently, the proposed methodology is dependent on the
constructive technique; that is, once trained, the neural network can be used only
in the design of dry‐type transformers of the same constructive technique as those
whose results of the tests were used to train the network. Nevertheless, even if only
used to provide the initial parameters for a more complex design tool, the proposed
method is useful since it is rapid and has low computational cost.

KE YWO RD S

artificial intelligence, epoxy resin, magnetic losses, transformers design automation,

winding losses, winding temperature

1 | IN T RO D U C T IO N robustness; the possibility of installation close to the load

Power transformers are essential components in electric
power distribution systems and industries. Suitable for indoor
installations that require safety and reliability, the dry‐type
transformers are the focus of this paper. These transformers
have smaller dimensions than the ones with oil insulation.
Also, they have advantages such as a higher mechanical
point; does not require transformation wells for oil collection,
fire fighting systems, fireguard walls, and oil supervision
accessories; and the lowest level of internal partial discharges
(owing to its vacuum encapsulation).1 Therefore, they are
typically used where there is the presence of people, for
example, in industries, residential buildings, and hospitals.2
However, dry‐type transformers are more expensive and have

Int. Trans. Electr. Energ. Syst. 2016; 1–10 wileyonlinelibrary.com/journal/etep Copyright © 2016 John Wiley & Sons, Ltd. 1
2 FINOCCHIO
power and voltage limitations. In this sense, failures on these
equipments imply great financial and social losses. So, they
must be designed and manufactured so as to meet the specific
usage needs, respecting the limits set in international stan-
dards such as the IEC‐76 and the IEEE C57.12.50‐1981.
Unfortunately, designing a transformer is not trivial. The
importance of balancing lifetime, performance, and costs,
besides the increasing development and evolution of various
materials, methods, and manufacturing processes, cause the
design of power transformers to be a highly complex process.
Part of this complexity is also due to the large number of vari-
ables and their interrelationships.3,4
In the design of transformers, a major goal is to reduce
the internal power losses. These are related to the losses in
the winding and iron and are dependent on parameters such
as internal temperature, resistance of windings, working
current, nominal voltage, load, and quality of the aluminum
used. Evidently, these losses affect the machine's temperature
regime, which directly influences the efficiency and lifetime
of the transformer.5 This is because the thermal stress is a
major cause of failures, since it causes the deterioration of
the insulating material.6 Therefore, estimating the value of
the internal temperature (or the value of the temperature
increase) is one of the most complex tasks in the transformer
design.3,7,8
Usually, transformers designers use either some design
patterns curves or the thermal network method to estimate
the temperature increase in the windings of this kind of
equipment. Also, several studies have attempted to predict
the internal temperature of the windings using empirical
and mathematical models and computational intelligence
techniques, which have obtained different degrees of suc-
cess.9 In these cases, the advantage is that the temperatures
can be foreseen when the operating conditions change, for
example, with the design of a new transformer. In this sense,
several studies deal with oil transformers, so as to minimize
production costs and find core losses, winding losses, and
no‐load losses.10–20 However, in the case of dry‐type trans-
formers, there are only a few thermal models available and
related researches.21 Besides, for example, Srinivasan and
Krishnan22 also point out that there are few theoretical and
practical studies on estimating hot spot temperature and
calculating the average temperature of the windings of dry‐
type transformers. So they developed a computational
modeling for this purpose using statistical methods, such as
multiple variable regression, multiple polynomial regression,
and computational techniques, such as artificial neural net-
works (ANN) and adaptive neuro‐fuzzy inference system.
In that paper, Srinivasan and Krishnan22 used as input
variables cold resistance, ambient temperature, and tempera-
ture rise. The results were verified based on temperature rise
tests in a dry‐type transformer of 55 kVA. In another
example, Lee et al21 developed a thermal model for foil‐type
winding for ventilated dry‐type power transformers rated at
2000 kVA, to calculate the electrical losses. Temperature
ET AL.
distributions were obtained in 3 different loads by the
coupled electromagnetic‐thermal model with the aid of a
2D transient finite element analysis, determined by the finite
element method.
This study proposes a tool to obtain the internal tempera-
ture of the windings, as well as winding losses and the total
losses, of dry‐type transformers. The applied methodology
uses ANN in parts of the transformer design, where the rela-
tionship between the variables is not well defined and the
design parameters are obtained empirically.23 The main
justification for using ANN is its ability to learn to relate a
complex set of data, which may be unnoticed by the
designer.5 Moreover, these networks have a quick assessment
and precision to resolve problems.24 The type of ANN
chosen was the feed‐forward multilayer perceptron. To train
the ANN, we used a database with tests results of 225 similar
dry‐type transformers with power outputs between 15 and
150 kVA and voltage classes equal to either 15 or 25 kV.
Building data of the transformer under design and the
environment temperature conditions, to which it will be
subject to during operation, are inputs of the neural networks.
In the data validation step, we used another 75 transformers
with the same characteristics of those used in the training.
The results show that the technique is very promising. This
is because the results show that the proposed tool provides
the designer the possibility of testing various design condi-
tions prior to manufacturing the transformer. Thus, the gener-
ated results allow the adequacy of the constructive variables
to the environmental condition of the equipment operation,
ensuring the construction of dry‐type transformers with better
performance and lifetime. Evidently, this makes the design
more efficient when it takes to preparation time and
specificity.
The model presented is simple. However, it works
because it applies only to the design of dry‐type transformers
based on the constructive technique of those transformers of
the tests, ie, transformers of which test results were used to
train the neural network. In this case, the neural network
was shown to be able to model the constructive technique
very well. Furthermore, by following the methodology
discussed here, the readers may conduct their own tests and
train a neural network that models their own constructive
technique, which can lead to great results. Also, another
application of the proposed tool is to provide the initial
parameters for a more complex design software. This can
be done rapidly and with low computational cost.
2 | M AT E R I ALS AN D M E T H O DS
Not all steps of a dry‐type transformer project have the same
degree of subjectivity or uncertainty. So, in this paper, to test
the use of ANN in modeling this type of problem, we opted
to model the temperature in the windings (referred to as T)
of this type of transformer. Such quantity is a function of
FINOCCHIO ET AL.
the following variables: area of the conductors (in this paper,
referred to as Sections and given in mm2); thickness of the
windings sheets (referred to as Thickness and given in mm);
number of winding channels (Channels), and the electrical
losses (referred to as LT and given in W). So, to calculate T,
first it is necessary to know LT. In turn, the quantities that
are part of the process of LT are defined as follows: high‐volt-
age coil resistance (referred to as RHV and given in Ω); low‐
voltage coil resistance (RLV, Ω); the environment temperature
(TA, °C); the excitation current (IEX, A); the core losses or the
no‐load losses (referred to as LC, given in W); and the
winding losses (LAl, W). Thus, to calculate LT, it is necessary
first to calculate LAl. In turn, the main variables that influence
LAl are as follows: the resistances of the high‐voltage and
low‐voltage coils, the temperature TA, and the short‐circuit
impedance (ZCC, given in %).
Based on the above, the developed tool as well as the
stages where the neural networks are applied and their
purposes can be seen in Figure 1. The neural network on
the left (Winding Losses Network) is intended to estimate
the winding losses (LAl), a variable that is used in the next
neural network (Total Losses Network), so as to estimate the
total losses (LT). Since the system aims to determine the inter-
nal temperature of the transformer, the Temperature Network
takes as inputs, in addition to the conductor section, the thick-
ness of the core constituent plates and the number and dimen-
sions of ventilation channels, besides the output of the Total
Losses Network. All these networks are discussed in detail
in Section 3.
For the training and validation of the 3 neural networks
developed in this paper, we used a database with tests results
of 300 dry‐type transformers, with power outputs from 15 to
150 kVA and voltages classes equal to either 15 or 25 kV.7 To
prepare the database, we performed the following tests with a
3‐phase transformer in a laboratory25: no‐load test, short‐cir-
cuit test, and winding resistance test. These tests are
discussed in detail below.
2.1 | No‐load test
No‐load tests of transformers aim to determine the following
quantities: core losses (LC), no‐load excitation currents (IEX),
Sections
Thickness Temperature
Network
Channels
RHV = RLV
IEX Total Losses
Network
Lc
RHV
RLV Winding
Losses
TA Network
ZCC
FIGURE 1 Proposed system for the neural network
3
transformation ratio, and parameters related to the magnetiz-
ing branch. Also, it is possible to analyze the nonsinusoidal
shape of the no‐load current and of the transient magnetizing
current. It is the safest test to determine those characteristics,
by applying the nominal operating voltage of the trans-
former. Figure 2 shows the block diagram of the no‐load
test, characterized by the Aron connection (2‐wattmeters
method). This test is characterized by open high‐voltage ter-
minals (VH). VL are the low‐voltage terminals; W1 and W2
are wattmeters (which indicate the losses in W); V represents
the voltmeter; A1, A2, and A3 are ammeters; and F is a fre-
quency meter.
2.2 | Short‐circuit test
This test is normally used to determine stray losses, but its
results also allow to determine the short‐circuit impedance
(ZCC), as well as the internal resistance and reactance of
transformers. For this, the low‐voltage terminals (VL) are
short circuited, and a voltage applied to the high‐voltage
terminal is varied so that the primary current observed on
the ammeter (Figure 3) is equal to the nominal current. This
voltage, referred to as VCC, represents from 10% to 15% of
the rated voltage (VR) of the high‐voltage side (VH). Under
these conditions, the core losses, which vary with the square
of the voltage, can be neglected. Thus, through the readings
of power, voltage, and short‐circuit current, the variables of
interest may be determined and applied in the determination
of the winding losses. More specifically, the impedance ZCC
will be considered in percentage, where ZCC = 100VCC/VR.
2.3 | Winding resistance test
The windings of the primary and the secondary of a trans-
former have an electrical resistance, as any conductor.
Respectively, such resistances are usually indicated by R1,
R2, R3 and R4, R5, R6 (1 for each phase). In this paper, we
consider R1, R2, and R3 as being equal to each other and
referred to as RLV. Similarly, R4, R5, and R6 are referred to
as RHV. These resistances exert a double effect on the opera-
tion of the transformer: determining an ohmic power loss in
the primary and the secondary and producing a loss of energy
T
LT
LAl
FIGURE 2 Scheme of the no‐load test
4 FINOCCHIO
FIGURE 3 Scheme of the short‐circuit test
by Joule effect. To reduce this winding loss to limits
established by technical standards, it is necessary to make
such resistances small enough. This is done by appropriately
choosing the section of the winding conductors. In
conducting this test, the electrical resistances of each winding
must be registered, phase to phase, as well as the terminals
where they have been measured and their respective temper-
atures. This procedure is performed by means of the Wheat-
stone bridge. It is also important to note that the resistance
values should always be obtained at the reference tempera-
ture, which is generally set at 115°C for dry‐type
transformers.
After obtaining the experimental data and preparing the
database, these data have been used for training the neural
networks, as shown below.
3 | IMP LEM ENTAT ION OF THE PRO POSED
TOOL
All neural networks implemented in this work have an input
layer, a single hidden layer, and 1 output layer with a single
neuron. In addition, for each of the neural networks, the heu-
ristic defined by Kolmogorov26 was applied to determine the
maximum number of neurons in the hidden layer, considering
a feed‐forward multilayer perceptron network. Since the
structure of Figure 1 has been defined, it was necessary to
train each neural network individually. For this, from the total
of the transformers design data, only 75% were used. The
remaining 25% were used for validation of this training. Also,
this training was performed by the modified Levenberg‐
Marquardt algorithm as learning rule.27,28,24 In the Winding
Losses Network and Total Losses Network (Figure 1), we
used the tangent sigmoid activation function.5 In the Temper-
ature Network (Figure 1), in turn, we used the hyperbolic tan-
gent activation function.29
3.1 | Winding losses network
The architecture of the neural network responsible for identi-
fying the processes related to the transformer winding losses
is shown in Figure 4. The variables that make up the input
vectors of this neural network were defined through the
ET AL.
FIGURE 4 Neural network architecture for the transformer winding losses
short‐circuit and windings resistance tests. Among these vari-
ables, TA represents the environment temperature in Celsius.
On the other hand, the output vector of the neural network
consisted of a single variable, LAL, which represents the
winding loss in watts. Based on the method of Kolmogorov
and considering this number of inputs, the maximum number
of neurons in the hidden layer was set equal to 9. However, it
was empirically determined that the best results were
obtained when we used only 5 neurons in the hidden layer.
3.2 | Total losses network
This neural network was used to identify processes related to
the transformer's total losses, which include both the winding
losses (considered variables) and the core losses (considered
fixed). Its architecture is shown in Figure 5. The network
input vectors were defined from the short‐circuit and no‐load
tests, besides the power lost due to the winding losses, esti-
mated by neural network discussed in the previous section.
More specifically, the input variables of this network were
defined as follows: RHV, RLV, TA, IEX (which represent the
excitation current in amperes), and LAl and LC (which repre-
sent the core loss in W). However, owing to their similarities,
RHV and RLV were considered equal.1 In turn, the output vec-
tor of the neural network has a single output, LT, which rep-
resents the total loss. Finally, based on the Kolmogorov
method, considering the number of inputs, the maximum
number of neurons in the hidden layer was set to 11. How-
ever, empirically, it was determined that the best results were
obtained when we used 10 neurons in this layer.
3.3 | Temperature network
Figure 6 shows the neural network used to estimate the
temperature. The input variables of this neural network are
as follows: the cross‐sectional area of the conductors used in
the windings (in mm2), referred to as Sections; thickness of
the winding sheet (in mm), referred to as Thickness;
1
It is possible because the resistances of the windings are balanced per phase
in both high and low voltages, since it is required by standard. The trans-
former leaves the factory with balanced windings. Their imbalance is gener-
ated by the load, since the system phases may be unbalanced.
FINOCCHIO ET AL. 5

do not influence the internal temperature of the transformers.

Instead, simply, it is not necessary to include them in the neu-
ral network input parameters because, as stated, these param-
eters correspond to the construction methodology of each
manufacturer. Thus, with respect to the ventilation ducts,
the only unknown is, in fact, the number of ducts.
In the following sections, the results of training of all
developed neural networks are presented.

4 | R E S U LTS

In this section, we present the results of the training and exe-

cution of the 3 neural networks implemented in the present
work.

4.1 | Winding losses network

Figure 7 shows the results of the winding loss generalization
done by the Winding Losses Network. In this graph, the
values obtained experimentally and the ones obtained by
FIGURE 5 Architecture of the neural network for the transformer total
the neural network are compared. One can check that the
losses
relative error is very small. In fact, it was calculated that the
and the number and dimensions of ventilation channels average relative error μ = 0.267%, the standard deviation
(Channels) and the output of the Total Losses Network σ = 0.204%, and the variance just σ2 = 0.042%. These results
(Section 2), referred to as LT. In the output of this neural show that the neural approach could adequately model the
network, the estimated variable is the temperature inside the winding losses processes.
transformer, which indicates the temperature rise of the wind- After validating the neural network, we examined the
ings. This temperature must be added to the environment one, relationship between the winding losses process and the
thus determining the temperature of the winding in nominal short‐circuit impedance and temperature. The results are
operating conditions. Again, using the Kolmogorov method, shown in Figures 8 and 9. Figure 8 shows the behavior of
the indicated maximum number of neurons in the hidden layer the winding loss relating to the short‐circuit impedance vari-
is 9. However, the final implemented hidden layer has only ation. In this simulation, the environment temperature was
8 neurons. considered constant, and mean values were used for the resis-
Unfortunately, the presented model allows only to foresee tance of high‐voltage and low‐voltage bobbins (RHV and RLV,
the temperature of a single point of the transformer. However, see neural network architecture in Figure 4). One can observe
when performing a test on the transformer, for example, a in this graph that in the intervals between 1.6% and 1.8% of
short‐circuit test, one can verify that the temperature of the short‐circuit impedance, there is an increasing winding loss.
transformer is not homogeneous, existing in higher‐tempera- Beyond this range, the curve remains constant. That behavior
ture zones.4,1 Therefore, in the training of the neural network, is due to the thermal effect inside the transformer. Generally,
it is important not to consider an average temperature. This is
because during normal operation of the transformer, the
temperature in a hot spot may increase excessively, to the point
of damaging the equipment.8,2 So, naturally, in this work, for
the training of the neural network, the considered temperature
is the one with the hottest point in the transformer. In turn, the
hottest point was determined experimentally with the aid of a
set of thermocouples during short‐circuit tests.
It is also important to note that the proposed methodology
is dependent on the constructive method of the transformer.
Especially, with respect to the ventilation ducts, dry‐type
transformers based on a given construction method have no
ventilation ducts with different geometries and positions.
Normally, such parameters are well defined in the construc-
tive methodology.6 This does not mean that these parameters FIGURE 6 Architecture of the feedforward multilayer perceptron network
6 FINOCCHIO
this slight variation in impedance is more difficult to be
observed experimentally, with this kind of detail, in tradi-
tional designs of transformers. However, this information is
FIGURE 7 Network behavior for the winding losses
FIGURE 8 Variation of the winding loss versus impedance
FIGURE 9 Variation of the winding loss versus temperature
ET AL.
important because this quantity numerically represents the
relationship between the circuit voltage and the nominal volt-
age of the transformer. In turn, Figure 9 shows the maximum
winding loss relating to the variation in the transformer tem-
perature. To generate this graph, we considered average
values for ZCC and for the high‐voltage and low‐voltage
impedances, while varying the room temperature. This graph
shows that there is a direct dependency between the tempera-
ture and the winding losses. In the intervals between 12°C
and 15°C, the values of the winding loss have rapid and uni-
form increases, reaching 1250 W, which is an accepted value
according to standards.
From this temperature until 34°C, the losses remain con-
stant. This is due to the inertia of the thermal process, in
which the saturation threshold of the material is reached
and then the loss becomes constant with the temperature
increase. This identification is very important for the designer
to select the most suitable type of aluminum, based on the
knowledge of the temperature maximum limit related to the
degree of purity of the material. However, it would be diffi-
cult to be observed in detail in experimental measurements
without the aid of the proposed tool.
4.2 | Total losses network
Figure 10 shows the results of the total loss generalization
done by the Total Losses Network, comparing the values
obtained experimentally with the ones obtained by the neural
network. One can see that the output values of the neural net-
work approached the values obtained in laboratory tests,
which show the success of this application for this type of
problem. Comparing experimental data with the output of
the neural network, one can see that the relative error is very
small. In fact, it was calculated that the average relative error
μ = 0.774%, with standard deviation σ = 0.567%, and vari-
ance σ2 = 0.322%.
FIGURE 10 Behavior of the total loss network
FINOCCHIO ET AL. 7

After validating the Total Losses Network, we examined

the dependence of the total losses regarding the no‐load
losses, winding losses, and excitation current, as presented
in Figures 11–13, respectively. For this, with the aid of the
validated neural network, we simulated the total losses
considering the minimum, average, and maximum values of
the following quantities: RHV, RLV, IEX, LAL, and LC (Section
2). In each graph, 3 curves are shown, which represent the
maximum, medium, and minimum of the input variables. It
is noteworthy that the simulations with the maximum and
average values are hypothetical, so as to identify possible
problems in the project, both in the constructive aspects and
in the quality of the material used.
Figure 11 shows the behavior of the total loss relating to
the no‐load loss, considering 3 different values for RHV, RLV,
TA, IEX, and LAL. In this graph, one can verify a linear trend FIGURE 12 Variation of the total loss versus winding loss
in all 3 curves, which indicates a thermal effect. It was also
observed that the curves of “maximum values” and “medium
values” show total loss values close to each other, even
considering the average and maximum values of the quanti-
ties already mentioned. This behavior shows that the estima-
tion of phenomena was performed satisfactorily, because if
there were any nonlinearity in the simulations, it would
represent that the network had not learned the process for
various situations. In this regard, it is essential to mention that
the simulated values in the curves are in disagreement with
the values recommended by standard. However, this condi-
tion is an extreme project situation that provides valuable
information to the designer, since it identifies the worst
condition to the transformer's design. This kind of detail
would be almost impossible to verify through either conven-
tional tools or experimental testing. This type of identifica-
tion is also valuable so as to provide information to the
transformers designer so that the project will be set neither FIGURE 13 Variation of the total loss versus excitation current
below the blue curve (minimum values) nor above the black
curve (maximum values). This is because it is preferable to at an operation midpoint. Thus, it will continue to be efficient
dimension the transformer so that it has a better performance at a point above or below the normal.
Figure 12 shows the behavior of the total loss relating to
the winding loss, considering different values for the quanti-
ties RHV, RLV, TA, IEX, and LC. One can observe that the
behavior of the total losses is similar to that in Figure 11.
The difference is that the average and maximum values are
farther. This analysis shows the sensitivity of the developed
neural tool. This observation is also very particular, because
it presents the contribution of the winding losses over the
total losses of a transformer.
In Figure 13, the total loss behavior with respect to the
excitation current is presented, considering 3 different values
of RHV, RLV, TA, LAl, and LC. In Figure 13, one can see the
influence of the excitation current in the overall result of
the total loss. This detail would be almost impossible to be
observed with conventional tools, owing to the complexity
found in the mathematical modeling of the phenomenon.
Generally, the weighing of the excitation current is restricted
FIGURE 11 Variation of the total loss versus no‐load loss only to no‐load core losses.
8 FINOCCHIO ET AL.

5 | C O NC LU S I ON

Currently, the design of dry‐type transformers is almost

entirely based on the designers' experience. Typically, several
attempts are necessary so as to achieve satisfactory results.
This is mainly due to the large number of variables involved
in calculations and to the fact that the relationships between
these variables are not well defined. In this article, we evalu-
ated the use of ANN in modeling the influence of various
design parameters on the internal temperature of dry‐type
transformers. For this, we used results of practical tests with
300 transformers. Qualitative and quantitative results showed
that ANN were able to model, with excellent accuracy, the
relationship between various design parameters and the
losses and internal temperature rise in dry‐type transformers.
Thus, by means of simulations using the validated neural
FIGURE 14 Evolution of the error in the Temperature Network training networks, one could see details that are almost impossible
process
to be observed with conventional tools, owing to the
complexity in the mathematical modeling of phenomena
involved. Thus, the developed system can be used to study
the phenomena that define and relate the variables involved
in the design. Evidently, this will have a direct influence on
the quality of future transformers designs. This is because,
based on the notion of behavior extracted via neural
networks, it will be possible to obtain closer‐to‐real functions
to model the interrelationships between the design parame-
ters. Evidently, this work can be extended so that not only
the internal temperature but also other dry‐type transformers
features could be modeled with the aid of ANN. Certainly,
this would minimize the time taken in the design and increase
the precision when estimating the final characteristics of this
type of transformer. Unfortunately, the proposed methodol-
ogy is dependent on the constructive method of the trans-
former. However, by following the methodology discussed
FIGURE 15 Comparison of internal temperature values so as to test the here, the readers may conduct their own tests and train a
generalization capacity neural network that models their own constructive technique.

4.3 | Temperature network

6 | L IST O F S Y M B O L S A N D
The evolution of the error in the Temperature Network train- A B B R E V I AT IO N S
ing process as a function of the number of iterations is shown
in Figure 14. The training stopping criterion was the stabiliza-
tion of the mean square error. With the completion of the net- 6.1 | Abbreviations
work training, the project initial parameters estimated by the
network are compared with the values actually used in the pro- ANN artificial neural networks
jects, through quantile‐quantile graphs.30,19 A comparison Channels number of winding channels
between the results estimated by the Temperature Network
FEA finite element analysis
and the real values obtained in the tests with 75 transformers
(used to validate the proposed system) is shown in the Sections area of the conductors

quantile‐quantile graphic of Figure 15, for the purpose of val- Thickness thickness of the windings sheets
idating and testing the generalization quality of the network.
In this graph, the output of the neural network fits the experi-
6.2 | Symbols
mental data with a coefficient of determination R2 = 0.988.
These results confirm that the neural network is well fitted,
having thus a good generalization. These facts demonstrate μ average relative error
the ability of the Temperature Network to solve the problem. A1, A2, A3 ammeters
FINOCCHIO ET AL. 9

7. L.H. Geromel and C.R. Souza. The application of intelligent systems in power
IEX no‐load excitation current transformer design. In IEEE Canadian Conference on Electrical and Computer
LAL winding losses Engineering, volume 1, pages 285–290, 2002.
8. M. A. Arjona, C. Hernandez, R. Escarela‐Perez, and E. Melgoza. Thermal
LC core losses analysis of a dry‐type distribution power transformer using fea. In Interna-
LT electrical losses tional Conference on Electrical Machines, pages 2270–2274, 2014. doi:
10.1109/ICELMACH.2014.6960501.
2
R coefficient of determination
9. Z. R. Radakovic and M. S. Sorgic. Basics of detailed thermal–hydraulic model
RHV high‐voltage coil resistance for thermal design of oil power transformers. IEEE Transactions on Power
Delivery, 25(2): 790–802, April 2010. ISSN 0885–8977. doi: 10.1109/
RLV low‐voltage coil resistance TPWRD.2009.2033076.
R1, R2, R3 electrical resistances of primary phase of the windings 10. A Khatri, H. Malik, and O.P. Rahi. Optimal design of power transformer using
genetic algorithm. In International Conference on Communication Systems
R4, R5, R6 electrical resistances of secondary phase of the windings
and Network Technologies, pages 830–833, May 2012.
σ standard deviation 11. Adly AA, Abd‐El‐Hafiz SK. A performance‐oriented power transformer
TA environment temperature design methodology using multi‐objective evolutionary optimization. Jour-
nal of Advanced Research. 2015;60(3):417–423.
σ2 variance 12. Rashtchi V, Bagheri A, Shabani A, Fazli S. A novel pso‐based technique for
VCC voltage at the common collector optimal design of protective current transformers. COMPEL—The interna-
tional journal for computation and mathematics in electrical and electronic
VH high‐voltage terminals engineering. 2011;30(2):505–518.
VL low‐voltage terminals 13. S. Tamilselvi and S. Baskar. Modified parameter optimization of distribution
transformer design using covariance matrix adaptation evolution strategy.
VR rated voltage International Journal of Electrical Power & Energy Systems, 61(0):208–218,
W 1 , W2 wattmeters 2014. ISSN 0142–0615.
14. M.A Arjona, C. Hernandez, and M. Cisneros‐Gonzalez. Hybrid optimum
ZCC short‐circuit impedance
design of a distribution transformer based on 2‐D FE and a manufacturer
F frequency meter design methodology. IEEE Transactions on Magnetics, 46(8):2864–2867,
Aug 2010. ISSN 0018–9464.
kV kilovolt amperes
15. R.A Jabr. Application of geometric programming to transformer design. IEEE
kVA kilovolt‐amps Transactions on Magnetics, 41(11):4261–4269, Nov 2005. ISSN 0018–9464.

T Temperature
V voltmeter
W watts
REFERENCES
1. Janusz Szczechowski and Krzysztof Siodla. Partial discharge localization on
dry‐type transformers. In International Conference on High Voltage Engi-
neering and Application, pages 1–3, 2014. doi: 10.1109/
ICHVE.2014.7035432.
2. Mariano Berrogain and Martin Carlen. Dry‐type transformers for
subtransmission. In 22nd International Conference and Exhibition on Elec-
tricity Distribution, pages 1–4, 2013. doi: 10.1049/cp.2013.1089.
3. Eleftherios Amoiralis, Pavlos Georgilakis, and Alkiviadis Gioulekas. An arti-
ficial neural network for the selection of winding material in power
transformers. In Grigoris Antoniou, George Potamias, Costas Spyropoulos,
and Dimitris Plexousakis, editors, Advances in Artificial Intelligence, volume
3955 of Lecture Notes in Computer Science, pages 465–468. Springer Berlin
Heidelberg, 2006. ISBN 978–3–540‐34117‐8.
4. Shiyou Wang, Youyuan Wang, and Xuetong ZHAO. Calculating model of
insulation life loss of dry‐type transformer based on the hot‐spot temperature.
In IEEE 11th International Conference on the Properties and Applications of
Dieletric Materials, pages 720–723, 2015. ISBN 978–1–4799‐8903‐4.
5. A.K. Yadav, A. Azeem, A. Singh, H. Malik, and O. P. Rahi. Application
research based on artificial neural network (ANN) to predict no‐load loss
for transformer's design. In International Conference on Communication Sys-
tems and Network Technologies, pages 180–183, June 2011.
16. E.I Amoiralis, P.S. Georgilakis, M.A Tsili, and AG. Kladas. Global trans-
former optimization method using evolutionary design and numerical field
computation. IEEE Transactions on Magnetics, 45(3):1720–1723, March
2009. ISSN 0018–9464.
17. P.S. Georgilakis. Recursive genetic algorithm‐finite element method technique
for the solution of transformer manufacturing cost minimisation problem. Elec-
tric Power Applications, IET, 3(6):514–519, November 2009. ISSN 1751–8660.
18. Tsili MA, Kladas AG, Georgilakis PS. Computer aided analysis and design of
power transformers. Computers in Industry. 2008; ISSN 0166‐361559
(4):338–350.
19. P. Georgilakis, N. Hatziargyriou, and D. Paparigas. AI helps reduce trans-
former iron losses. IEEE Computer Applications in Power, 12(4):41–46,
Oct 1999. ISSN 0895–0156.
20. Anastassia J. Tsivgouli, Marina A. Tsili, Antonios G. Kladas, Pavlos S.
Georgilakis, Athanassios T. Souflaris, and Annie D. Skarlatini. Geometry
optimization of electric shielding in power transformers based on finite
element method. Journal of Materials Processing Technology, 181(1–3):
159–164, 2007. ISSN 0924‐136. Selected Papers from the 4th Japanese‐Med-
iterranean Workshop on Applied Electromagnetic Engineering for Magnetic,
Superconducting and Nano Materials.
21. Lee M, Abdullah HA, Jofriet JC, Patel D. Temperature distribution in foil
winding for ventilated dry‐type power transformers. Electric Power Systems
Research. 2010; ISSN 0378‐779680(9):1065–1073.
22. Srinivasan M, Krishnan A. Hot resistance estimation for dry type transformer
using multiple variable regression, multiple polynomial regression and soft com-
puting techniques. American Journal of Applied Sciences. 2012;9(2):231–237.
23. Mohamed EA, Abdelaziz AY, Mostafa AS. A neural network‐based scheme
for fault diagnosis of power transformers. Electric Power Systems Research.
2005; ISSN 0378‐779675(1):29–39.

6. Jian Sheng Chen, Hui Gang Sun, Chao Lu, Qing Zheng, and Yi Ren Liu.
Numerical calculation of dry‐type transformer and temperature rise analysis.
In Materials Processing and Manufacturing III, volume 753 of Advanced
Materials Research, pages 1025–1030. Trans Tech Publications, 10 2013.
doi: 10.4028/www.scientific.net/AMR.753-755.1025.
24. N. Suttisinthong and C. Pothisarn. Analysis of electrical losses in transformers
using artificial neural networks. International MultiConference of Engineers
and Computer Scientists—IMECS, 2, 2014. ISSN 2078‐0958.
25. W. J. Bonwick and P. J. Hession. Fast measurement of real and reac-
tive power in three‐phase circuits. Science, Measurement and
10 FINOCCHIO ET AL.

Technology, IEE Proceedings A, 139(2):51–56, Mar 1992. ISSN 0960‐ 30. Blom G. Statistical Estimates and Transformed Beta‐Variables. John Wiley
7641. & Sons;1958.
26. Kůrková V. Kolmogorov's theorem and multilayer neural networks. Neural
Networks. 1992; ISSN 0893‐60805(3):501–506.
27. M.T. Hagan and M.‐B. Menhaj. Training feedforward networks with the
How to cite this article: Finocchio, M. A. F., Lopes,
marquardt algorithm. IEEE Transactions on Neural Networks, 5(6):989–993, J. J., de França, J. A., Piai, J. C., and Mangili, J. F., Jr
Nov 1994. ISSN 1045‐9227. (2016), Neural networks applied to the design of
28. Haykin SO. Neural Networks and Learning Machines. 3rd ed. dry‐type transformers: an example for analysis of
Prentice Hall;2008 .ISBN ISBN‐10: 0131471392; ISBN‐13: 978‐
winding temperature and increase in thermal qua-
0131471399
29. Madahi SSK, Salah P. A neural network based method for cost estimation 63/
lity, Int. Trans. Electr. Energ. Syst., doi: 10.1002/
20 kV and 132/20 kV transformers. Journal of Basic and Applied Scientific etep.2257
Research. 2012;939–945.