Professional Documents
Culture Documents
DOI 10.1002/etep.2257
RESEARCH ARTICLE
1
Universidade Tecnológica do Paraná, Cornélio
Procópio, Paraná, Brazil Summary
2
Universidade Estadual de Londrina, Much of the cost of manufacturing transformers is related to the complexity in the
Departamento de Engenharia Elétrica, Laboratório development of its project, which involves many variables, such as different types
de Automação e Instrumentação Inteligente, of materials, methods, and manufacturing processes. Also, it is notoriously difficult
Londrina, Paraná, Brazil
to establish standards that relate the transformer characteristics with these design var-
Correspondence
José Alexandre de França Departamento de
iables, since each, almost always, variable is calculated empirically. One of the most
Engenharia Elétrica, Universidade Estadual de important design variables is the internal temperature of the transformers, which
Londrina, Caixa Postal 10011, Londrina, Paraná, directly influences the lifetime of such equipments. So, a computational tool is under
Brazil.
Email: jaf@jaf.eng.br
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
Int. Trans. Electr. Energ. Syst. 2016; 1–10 wileyonlinelibrary.com/journal/etep Copyright © 2016 John Wiley & Sons, Ltd. 1
2 FINOCCHIO ET AL.
power and voltage limitations. In this sense, failures on these distributions were obtained in 3 different loads by the
equipments imply great financial and social losses. So, they coupled electromagnetic‐thermal model with the aid of a
must be designed and manufactured so as to meet the specific 2D transient finite element analysis, determined by the finite
usage needs, respecting the limits set in international stan- element method.
dards such as the IEC‐76 and the IEEE C57.12.50‐1981. This study proposes a tool to obtain the internal tempera-
Unfortunately, designing a transformer is not trivial. The ture of the windings, as well as winding losses and the total
importance of balancing lifetime, performance, and costs, losses, of dry‐type transformers. The applied methodology
besides the increasing development and evolution of various uses ANN in parts of the transformer design, where the rela-
materials, methods, and manufacturing processes, cause the tionship between the variables is not well defined and the
design of power transformers to be a highly complex process. design parameters are obtained empirically.23 The main
Part of this complexity is also due to the large number of vari- justification for using ANN is its ability to learn to relate a
ables and their interrelationships.3,4 complex set of data, which may be unnoticed by the
In the design of transformers, a major goal is to reduce designer.5 Moreover, these networks have a quick assessment
the internal power losses. These are related to the losses in and precision to resolve problems.24 The type of ANN
the winding and iron and are dependent on parameters such chosen was the feed‐forward multilayer perceptron. To train
as internal temperature, resistance of windings, working the ANN, we used a database with tests results of 225 similar
current, nominal voltage, load, and quality of the aluminum dry‐type transformers with power outputs between 15 and
used. Evidently, these losses affect the machine's temperature 150 kVA and voltage classes equal to either 15 or 25 kV.
regime, which directly influences the efficiency and lifetime Building data of the transformer under design and the
of the transformer.5 This is because the thermal stress is a environment temperature conditions, to which it will be
major cause of failures, since it causes the deterioration of subject to during operation, are inputs of the neural networks.
the insulating material.6 Therefore, estimating the value of In the data validation step, we used another 75 transformers
the internal temperature (or the value of the temperature with the same characteristics of those used in the training.
increase) is one of the most complex tasks in the transformer The results show that the technique is very promising. This
design.3,7,8 is because the results show that the proposed tool provides
Usually, transformers designers use either some design the designer the possibility of testing various design condi-
patterns curves or the thermal network method to estimate tions prior to manufacturing the transformer. Thus, the gener-
the temperature increase in the windings of this kind of ated results allow the adequacy of the constructive variables
equipment. Also, several studies have attempted to predict to the environmental condition of the equipment operation,
the internal temperature of the windings using empirical ensuring the construction of dry‐type transformers with better
and mathematical models and computational intelligence performance and lifetime. Evidently, this makes the design
techniques, which have obtained different degrees of suc- more efficient when it takes to preparation time and
cess.9 In these cases, the advantage is that the temperatures specificity.
can be foreseen when the operating conditions change, for The model presented is simple. However, it works
example, with the design of a new transformer. In this sense, because it applies only to the design of dry‐type transformers
several studies deal with oil transformers, so as to minimize based on the constructive technique of those transformers of
production costs and find core losses, winding losses, and the tests, ie, transformers of which test results were used to
no‐load losses.10–20 However, in the case of dry‐type trans- train the neural network. In this case, the neural network
formers, there are only a few thermal models available and was shown to be able to model the constructive technique
related researches.21 Besides, for example, Srinivasan and very well. Furthermore, by following the methodology
Krishnan22 also point out that there are few theoretical and discussed here, the readers may conduct their own tests and
practical studies on estimating hot spot temperature and train a neural network that models their own constructive
calculating the average temperature of the windings of dry‐ technique, which can lead to great results. Also, another
type transformers. So they developed a computational application of the proposed tool is to provide the initial
modeling for this purpose using statistical methods, such as parameters for a more complex design software. This can
multiple variable regression, multiple polynomial regression, be done rapidly and with low computational cost.
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 2 | M AT E R I ALS AN D M E T H O DS
variables cold resistance, ambient temperature, and tempera-
ture rise. The results were verified based on temperature rise Not all steps of a dry‐type transformer project have the same
tests in a dry‐type transformer of 55 kVA. In another degree of subjectivity or uncertainty. So, in this paper, to test
example, Lee et al21 developed a thermal model for foil‐type the use of ANN in modeling this type of problem, we opted
winding for ventilated dry‐type power transformers rated at to model the temperature in the windings (referred to as T)
2000 kVA, to calculate the electrical losses. Temperature of this type of transformer. Such quantity is a function of
FINOCCHIO ET AL. 3
the following variables: area of the conductors (in this paper, transformation ratio, and parameters related to the magnetiz-
referred to as Sections and given in mm2); thickness of the ing branch. Also, it is possible to analyze the nonsinusoidal
windings sheets (referred to as Thickness and given in mm); shape of the no‐load current and of the transient magnetizing
number of winding channels (Channels), and the electrical current. It is the safest test to determine those characteristics,
losses (referred to as LT and given in W). So, to calculate T, by applying the nominal operating voltage of the trans-
first it is necessary to know LT. In turn, the quantities that former. Figure 2 shows the block diagram of the no‐load
are part of the process of LT are defined as follows: high‐volt- test, characterized by the Aron connection (2‐wattmeters
age coil resistance (referred to as RHV and given in Ω); low‐ method). This test is characterized by open high‐voltage ter-
voltage coil resistance (RLV, Ω); the environment temperature minals (VH). VL are the low‐voltage terminals; W1 and W2
(TA, °C); the excitation current (IEX, A); the core losses or the are wattmeters (which indicate the losses in W); V represents
no‐load losses (referred to as LC, given in W); and the the voltmeter; A1, A2, and A3 are ammeters; and F is a fre-
winding losses (LAl, W). Thus, to calculate LT, it is necessary quency meter.
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 2.2 | Short‐circuit test
impedance (ZCC, given in %). This test is normally used to determine stray losses, but its
Based on the above, the developed tool as well as the results also allow to determine the short‐circuit impedance
stages where the neural networks are applied and their (ZCC), as well as the internal resistance and reactance of
purposes can be seen in Figure 1. The neural network on transformers. For this, the low‐voltage terminals (VL) are
the left (Winding Losses Network) is intended to estimate short circuited, and a voltage applied to the high‐voltage
the winding losses (LAl), a variable that is used in the next terminal is varied so that the primary current observed on
neural network (Total Losses Network), so as to estimate the the ammeter (Figure 3) is equal to the nominal current. This
total losses (LT). Since the system aims to determine the inter- voltage, referred to as VCC, represents from 10% to 15% of
nal temperature of the transformer, the Temperature Network the rated voltage (VR) of the high‐voltage side (VH). Under
takes as inputs, in addition to the conductor section, the thick- these conditions, the core losses, which vary with the square
ness of the core constituent plates and the number and dimen- of the voltage, can be neglected. Thus, through the readings
sions of ventilation channels, besides the output of the Total of power, voltage, and short‐circuit current, the variables of
Losses Network. All these networks are discussed in detail interest may be determined and applied in the determination
in Section 3. of the winding losses. More specifically, the impedance ZCC
For the training and validation of the 3 neural networks will be considered in percentage, where ZCC = 100VCC/VR.
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 2.3 | Winding resistance test
prepare the database, we performed the following tests with a The windings of the primary and the secondary of a trans-
3‐phase transformer in a laboratory25: no‐load test, short‐cir- former have an electrical resistance, as any conductor.
cuit test, and winding resistance test. These tests are Respectively, such resistances are usually indicated by R1,
discussed in detail below. 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
2.1 | No‐load test
as RHV. These resistances exert a double effect on the opera-
No‐load tests of transformers aim to determine the following tion of the transformer: determining an ohmic power loss in
quantities: core losses (LC), no‐load excitation currents (IEX), the primary and the secondary and producing a loss of energy
Sections
Thickness Temperature
Network T
Channels
RHV = RLV
FIGURE 1 Proposed system for the neural network FIGURE 2 Scheme of the no‐load test
4 FINOCCHIO ET AL.
4 | R E S U LTS
this slight variation in impedance is more difficult to be important because this quantity numerically represents the
observed experimentally, with this kind of detail, in tradi- relationship between the circuit voltage and the nominal volt-
tional designs of transformers. However, this information is 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-
FIGURE 7 Network behavior for the winding losses
cult to be observed in detail in experimental measurements
without the aid of the proposed tool.
FIGURE 9 Variation of the winding loss versus temperature FIGURE 10 Behavior of the total loss network
FINOCCHIO ET AL. 7
5 | C O NC LU S I ON
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 16. E.I Amoiralis, P.S. Georgilakis, M.A Tsili, and AG. Kladas. Global trans-
former optimization method using evolutionary design and numerical field
V voltmeter computation. IEEE Transactions on Magnetics, 45(3):1720–1723, March
2009. ISSN 0018–9464.
W watts
17. P.S. Georgilakis. Recursive genetic algorithm‐finite element method technique
for the solution of transformer manufacturing cost minimisation problem. Elec-
REFERENCES tric Power Applications, IET, 3(6):514–519, November 2009. ISSN 1751–8660.
1. Janusz Szczechowski and Krzysztof Siodla. Partial discharge localization on 18. Tsili MA, Kladas AG, Georgilakis PS. Computer aided analysis and design of
dry‐type transformers. In International Conference on High Voltage Engi- power transformers. Computers in Industry. 2008; ISSN 0166‐361559
neering and Application, pages 1–3, 2014. doi: 10.1109/ (4):338–350.
ICHVE.2014.7035432.
19. P. Georgilakis, N. Hatziargyriou, and D. Paparigas. AI helps reduce trans-
2. Mariano Berrogain and Martin Carlen. Dry‐type transformers for former iron losses. IEEE Computer Applications in Power, 12(4):41–46,
subtransmission. In 22nd International Conference and Exhibition on Elec- Oct 1999. ISSN 0895–0156.
tricity Distribution, pages 1–4, 2013. doi: 10.1049/cp.2013.1089. 20. Anastassia J. Tsivgouli, Marina A. Tsili, Antonios G. Kladas, Pavlos S.
Georgilakis, Athanassios T. Souflaris, and Annie D. Skarlatini. Geometry
3. Eleftherios Amoiralis, Pavlos Georgilakis, and Alkiviadis Gioulekas. An arti- optimization of electric shielding in power transformers based on finite
ficial neural network for the selection of winding material in power element method. Journal of Materials Processing Technology, 181(1–3):
transformers. In Grigoris Antoniou, George Potamias, Costas Spyropoulos, 159–164, 2007. ISSN 0924‐136. Selected Papers from the 4th Japanese‐Med-
and Dimitris Plexousakis, editors, Advances in Artificial Intelligence, volume iterranean Workshop on Applied Electromagnetic Engineering for Magnetic,
3955 of Lecture Notes in Computer Science, pages 465–468. Springer Berlin Superconducting and Nano Materials.
Heidelberg, 2006. ISBN 978–3–540‐34117‐8.
21. Lee M, Abdullah HA, Jofriet JC, Patel D. Temperature distribution in foil
4. Shiyou Wang, Youyuan Wang, and Xuetong ZHAO. Calculating model of winding for ventilated dry‐type power transformers. Electric Power Systems
insulation life loss of dry‐type transformer based on the hot‐spot temperature. Research. 2010; ISSN 0378‐779680(9):1065–1073.
In IEEE 11th International Conference on the Properties and Applications of 22. Srinivasan M, Krishnan A. Hot resistance estimation for dry type transformer
Dieletric Materials, pages 720–723, 2015. ISBN 978–1–4799‐8903‐4. using multiple variable regression, multiple polynomial regression and soft com-
5. A.K. Yadav, A. Azeem, A. Singh, H. Malik, and O. P. Rahi. Application puting techniques. American Journal of Applied Sciences. 2012;9(2):231–237.
research based on artificial neural network (ANN) to predict no‐load loss 23. Mohamed EA, Abdelaziz AY, Mostafa AS. A neural network‐based scheme
for transformer's design. In International Conference on Communication Sys- for fault diagnosis of power transformers. Electric Power Systems Research.
tems and Network Technologies, pages 180–183, June 2011. 2005; ISSN 0378‐779675(1):29–39.
6. Jian Sheng Chen, Hui Gang Sun, Chao Lu, Qing Zheng, and Yi Ren Liu. 24. N. Suttisinthong and C. Pothisarn. Analysis of electrical losses in transformers
Numerical calculation of dry‐type transformer and temperature rise analysis. using artificial neural networks. International MultiConference of Engineers
In Materials Processing and Manufacturing III, volume 753 of Advanced and Computer Scientists—IMECS, 2, 2014. ISSN 2078‐0958.
Materials Research, pages 1025–1030. Trans Tech Publications, 10 2013. 25. W. J. Bonwick and P. J. Hession. Fast measurement of real and reac-
doi: 10.4028/www.scientific.net/AMR.753-755.1025. 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.