Professional Documents
Culture Documents
1, JANUARY 2007
201
Manuscript received August 25, 2005; revised March 8, 2006. Paper no.
TPWRD-00503-2005.
The authors are with the Department of Electrical Engineering, Tampere
University of Technology, Tampere 33101, Finland (e-mail: jouni.pylvanainen@tut.fi; kirsi.nousiainen@tut.fi; pekka.verho@tut.fi).
Color version of Figs. 36 are available online at http://ieeexplore.org.
Digital Object Identifier 10.1109/TPWRD.2006.877075
202
B. Current Harmonics
The temperature evaluation methods presented in [1] and [2]
are suitable in loading situations, where no current harmonics
are involved. However, in some loading circumstances, the content of the current harmonics can be quite high. In the case, the
hot- spot temperature calculated with the conventional methods
adapted from loading guides [1] and [2] can differ quite greatly
from the real hot-spot temperature. Guidance to take into
account the harmonic losses with simple equations in loading
calculations can be found (e.g., in CENELEC harmonization
document 428.4 [8]). For similar purposes, equations from
valid IEEE standard C57.110 [9] and IEC standard 61378-1
[10] could be adapted.
1) Effects of the Harmonics: There are usually both current
and voltage harmonics occurring at the same time in the network, but in the studies only the effects of the current harmonics
are evaluated. This is because the effects of the increase of the
no-load losses due to harmonics in voltage are usually negligible
compared to the load losses, in general, or losses generated by
the current harmonics. The harmonics in the load current noticeably increase the losses of an ordinary delta-star-connected
transformer.
One way to take care of the generated extra losses due to current harmonics is to utilize -rated transformers. -rated transformers are designed with bearing the harmonic load in mind
and, in this way, better endure the effects of harmonics. However, these special transformers are quite expensive and a more
common solution is to derate a standard transformer designed
for sinusoidal loads. In this case, the amplitude and frequency
of the predominant current harmonics should be known to define the derating factor of the transformer. Guidance for the simplified calculation of the derating factor due to the current harmonics is in the studies adapted from CENELEC publication
HD 428.4 S1-1994 [8]. The utilized equations are depicted more
precisely in the following sections.
The extra losses caused by the harmonics are mainly due to
increased eddy current losses in the windings and connections
and due to stray losses in the conducting structural parts of the
transformer [3]. According to [8], the effects of the harmonics
(1)
are the eddy current losses at the fundamental frewhere
quency divided by the loss due to dc current equal to the root
mean square (rms) value of the sinusoidal current (both at reference temperature), is the harmonic order, and is an exponential constant which is dependent on the type of winding and
frequency. It should be noted that the presented parameter is
not the same as that used in standards [9] and [10], where a parameter is presented with two different values (one for eddy current losses inside the windings and one for stray losses outside
the windings). Another thing, which needs to be recognized, is
that the defined factor is not the same -factor, which is discussed, for example, in [3]. In (1), the used parameter is given
by (2)
(2)
is the magnitude of the th harmonic and
is the
where
magnitude of the fundamental current. The calculated factor
can easily be utilized in temperature and loss of life calculations
of a transformer.
C. Artificial Neural Networks (ANNs)
The ANNs can also be utilized in many ways for monitoring
purposes in distribution networks [11]. An ANN has been used,
for example, for thermal modeling of the energy cables [12].
The results presented in [12] have proven that ANNs can emulate thermal behavior of the network cables quite accurately
and the ANN models can be used for similar types of modeling
problems.
In the studies, ANN has been used to emulate hot-spot and
top-oil temperature behavior of the oil-immersed transformer
studied. The selected input parameters for the ANN are the same
as the input quantities used with calculation equations adapted
from the standards [1] and [2]. The selected ANN model is a dynamic two-layer perceptron neural network structure. To train
the network, a GaussNewton recursive prediction error algorithm is utilized. More information about the algorithms used is
available (e.g., in [13]).
The structure of the ANN has been optimized by testing different multilayer perceptron network structures and selecting
the optimal combination of hidden layer neurons. An example
of a schematic ANN structure, which has been divided into several layers, is shown as a blackbox model in Fig. 2. The first
level is the input layer, which consists of five different input vectors. The vectors are the measured ambient temperature
and three-phase currents , , , as well as one past output
of the ANN. The second layer is the hidden layer
value
with five hidden neurons and nonlinear activation functions. The
third layer specifies the output layer structure, which gives the
calculated temperature values as an output.
203
Fig. 2. Blackbox model of the ANN structure, which has been used in emulating the thermal behavior of the distribution transformers.
Fig. 3. Modeling results calculated with different calculation algorithms.
is the case, for instance, during the rapid load changes. It can
also be seen that sometimes the oil-temperature values calculated with [1] do not follow the measured values as accurately
as the values calculated with [2]. However, adjusting the calculation parameters again with the measured data could decrease
the error.
In general, it can be evaluated that both methods adapted from
[1] and [2] follow the measured top-oil temperature values accurately enough for the loss of life approximation. The error between the measured and calculated temperature values are in the
worst case below 4 and the difference has remained by only a
few moments. However, when the methods are utilized for accurate temperature evaluation, for example, when the methods are
used to define dynamically maximum loading limits, the calculation parameters should be adjusted properly according to the
transformer studied.
The achieved results are similar to the ones presented in other
studies related to this topic even if the other publications concentrate on power transformers [3], [10]. In conclusion, it can be
stated that the calculation methods described in standards IEC
60354 and IEEE C57.91 annex G can be utilized to determine
the condition of a transformer to some extent.
C. Results With Trained ANN
In order to study the accuracy of the trained ANN, the
measured and calculated top-oil temperatures have been compared. Fig. 4 presents an example of the ANN modeling with a
proper training set. Also, the results achieved with conventional
methods from [1] and [2] are shown in Fig. 4.
As can be seen from Fig. 4, the top-oil temperature modeled
with a properly trained ANN accurately follows the actual measured values. In the case of the proper training set, there is only a
negligible difference between the measured and the calculated
values. Also when utilizing simplistic equations adapted from
the standards, the results are, in the depicted case, quite accurate even if the modeling accuracy is not as precise all the time
as with a well-trained ANN.
However, the modeling results are not always as accurate
as shown in Fig. 4. The results achieved with ANN can differ
greatly from the measured values as depicted later in this paper.
The ANN could give misleading modeling results when the
204
Fig. 4. Results achieved with a properly trained ANN and equations adapted
from [1] and [2] compared to measured top-oil temperature values. The sampling period in the study is 10 min.
Fig. 6. Measured and calculated top-oil temperature with and without taking
into account current harmonics for a period of 6 weeks during which the changes
in compensation of harmonics took place.
training data used do not cover all of the input and output combinations which are possible for the process studied. Another
situation, when the ANN does not give accurate results, is in the
case when the process has changed after the training data for
ANN have been measured. If an ANN is not properly trained, the
use of ANN in continuous monitoring may lead, for example, to
false loading-level assumptions.
Measured information from the transformer is not always
available for ANN training in which case, the data measured
from other similar transformers could be utilized. In this case,
the ANN could be trained with one transformer data and the
trained ANN could be used, for example, to evaluate the temperature behavior of the other transformer on site, where data
cannot be measured. In Fig. 5, a case is depicted, where the ANN
is trained with one transformer data and the evaluation of the
ANN is done with other similarly transformer data. Generally,
the transformers studied can be considered to be similar even
if there is a slight difference in transformer losses. The nameplate loading value for the transformers studied is 1000 kVA.
The measurement period of 1 h is the same in both cases.
As can be seen from Fig. 5, also in this case, the ANN could
give quite accurate results. However, in this case, the difference
205
TABLE I
CONTENT OF HARMONICS IN THE EXAMPLE CASE
During the unchanged network situation, the ratio of the harmonics amplitude compared to the amplitude of the fundamental frequency and is predicted to be constant. The approximated values of harmonics during the presence of the serious
current harmonics are given in Table I.
By combining simple calculation methods described in [1]
and [3], the heating effects of harmonic currents in a transformer can be taken into account with the calculated factor .
The effects of the current harmonics can be clearly seen when
comparing the curves in Fig. 6. It can be seen that by utilizing
temperature calculations taking into account the effects of harmonics, the results are more accurate than without. However, it
can be seen that sometimes the results still differ greatly from
the measured values and more studies should be made for adequate thermal behavior modeling with nonsinusoidal loads.
The accuracy of the trained ANN has also been studied with
the same case and the results are also shown in Fig. 6. As can be
seen in this case, the output of the ANN does not follow accurately the measured temperature values after the load structure
has changed. This occurs because the ANN has been trained
with data recorded before the new installation. The input and
output correlation has changed after the measurement period
used in training.
E. Evaluation of the Results
The results calculated with equations adapted from loading
guides [1] and [2] are indicative, but it can be seen that with the
studied calculation algorithms, network utilities can achieve important temperature and aging information of their distribution
transformers. When a transformer is supplying nonsinusoidal
loads, rough estimates for the temperature behavior of the transformer can be achieved with equations adapted from standard
[3].
If the results achieved with a trained ANN are compared to the
results achieved with calculation models adapted from loading
guides [1] and [2], a well-trained ANN generally gives somewhat more accurate results. When proper training information
is used, there is only a negligible difference between the top-oil
temperature values achieved with trained ANN and the measured top-oil temperature. However, the error between the measured temperature value and the values calculated with methods
adapted from [1] and [2] can be minimized by measuring the
hot-spot or top-oil temperature from an adequate time period
from the studied transformer and by adjusting the calculation
parameters used with the information attained.
The studies with ANN have indicated that also neural networks can be an option when implementing an application for
temperature monitoring of oil-immersed transformers. It has
been shown that an ANN trained with a proper set of information
gives accurate estimates for the top-oil temperature of a transformer. However, the ANN should be used only when the accurate data from loading values, ambient, and top-oil/hot-spot
temperatures of the transformer is available for a proper time
period covering all possible loading situations. Also, if the load
characteristics have been changed and there is a possibility, for
instance, to current harmonics occurrence, ANN should be retrained with measured temperature values.
IV. ALGORITHM UTILIZATION WITH EXISTING NETWORK DATA
The studies have shown that different methods for distribution
transformer temperature and loss of life calculations are available. In most cases, the accuracy of the results is adequate, when
parameters for the methods adapted from [1] and [2] are selected
carefully or a proper training set is utilized with ANN. The results, of course, are also depending on the accuracy of the input
data used. In this chapter, some possibilities to utilize methods
also in a larger scale for condition monitoring purposes are depicted.
Fig. 7 shows some possibilities to implement the algorithms
in the existing information system. It can be seen that the calculational algorithms are possible to be implemented as an independent application or by integrating the algorithms into the existing supervisory control and data acquisition (SCADA)/distribution-management system (DMS) system. In these cases, data
have to be collected and stored widely from the network. The
other possibility is to implement algorithms as part of the relay
functionality, in which case data transfer is seldom needed and
monitoring can be done locally. More information of the pilot
implementations done during the research work is depicted in
[14]. All of the depicted options have both upsides and downsides and the implementation strategy should be selected according to the availability of the data and data-transfer possibilities.
The input information for temperature calculations (loading
values, ambient temperature, harmonics, etc.) can be adapted
from several different information sources as also depicted in
Fig. 7. The most accurate information for loss of life evaluation
206
V. CONCLUSION
Methods described in [1] and [2] can be used to evaluate the
temperature and aging of distribution transformers with some
accuracy. The input information could be adapted from different
data sources. The precision of the evaluation depends on the
input data and the calculation method utilized. To evaluate temperature values for a transformer supplying loads with severe
harmonics, the effects of current harmonics could be taken into
account. This could be done, for example, by using the equations presented in [8], [9], or [10]. In the selected approach, the
methods adapted from [8] have been used. In the case of harmonics, the information of the amplitudes of predominant current harmonics must be available in order to evaluate the excess
losses caused by the harmonics. For the transformer monitoring
purposes, also the ANN can be utilized when the ANN is properly trained. This requires measured data from the studied or
same type of transformer supplying similar loading. When utilized correctly, the ANN could give accurate temperature values
for loss of life calculations.
The analyzing methods studied can be utilized for large-scale
condition monitoring purposes. The utilized method should be
selected carefully, if accurate information is needed of the transformer. However, the methods could also be used with approximated loading information to give rough condition estimates of
the transformers for asset-management purposes. In the field, a
lot of research work has been conducted at TUT, and during the
recent work, the possibilities to utilize other relevant information for transformer condition monitoring purposes are studied.
REFERENCES
[1] IEC Guide for Loading Guide for Oil Immersed Transformers, CEI/IEC
Std. 60354, Sep. 1991.
[2] IEEE Guide for Loading Mineral-Oil-Immersed Transformers, IEEE
Std. C57.91-1995, Jun. 1995.
[3] L. W. Pierce, Transformer design and application considerations for
nonsinusoidal load currents, IEEE Trans. Ind. Appl., vol. 32, no. 3, pp.
170178, May/Jun. 1996.
[4] E. Simonson, Transformer ratings and transformer life, in , Inst.
Elect. Eng. Colloq. Transformer Life Management, Oct. 1998, pp.
7/17/7.
[5] L. W. Pierce, Predicting liquid filled transformer loading capability,
IEEE Trans. Ind. Appl., vol. 30, no. 1, pp. 170178, Jan./Feb. 1994.
[6] H. Nordman and M. Lahtinen, Thermal overload tests on a 400-MVA
power transformer with a special 2.5-p.u. short time loading capability, IEEE Trans. Power Del., vol. 18, no. 1, pp. 107112, Jan. 2003.
[7] D. Susa, M. Lehtonen, and H. Nordman, Dynamic thermal modeling
of distribution transformer, IEEE Trans. Power Del., vol. 20, no. 3, pp.
19191929, Jul. 2005.
[8] Three-Phase Oil-Immersed Distribution Transformers 50 Hz, From 50
to 2500 kVA, With Highest Voltage for Equipment Not Exceeding 36 kV.
Part 4: Determination of the Power Rating of a Transformer Loaded
With Non-Sinusoidal Currents, CENELEC HD 428.4 S1:1994 Std.,
Sep. 1994.
[9] IEEE Recommended Practice for Establishing Transformer Capability
When Supplying Nonsinusoidal Load Currents, IEEE Std. C57.110-98,
Jul. 1998.
[10] Convertor TransformersPart 1: Transformers for Industrial Applications, CEI/IEC Std. 61378-1, Sep. 1997.
[11] K. P. Wong, Artificial intelligence and neural network applications in
power systems, in Proc. Inst. Elect. Eng., 2nd Int. Conf. Advances in
Power System Control, Operation and Management Conference, Hong
Kong, China, Dec. 1993, vol. 1, pp. 3746.
[12] G. Wolff, H. Haubrich, and T. Seitz, Evaluation of non steady
state loading limits for network components using artificial neural
networks, in Proc. Inst. Elect. Eng., 2nd Int. Conf. Advances in Power
System Control, Operation and Management Conf., Hong Kong,
China, 1993, vol. 2, pp. 944949.
[13] M. Norgard, Neural Network Based System Identification Toolbox,
Ver. 2.0 Tech. Univ. Denmark, Denmark, Tech. rep. 00-E-891, Jan.
2000.
[14] K. Nousiainen and P. Verho, Monitoring of distribution transformer
integrated into network automation, in Proc. DA/DSM Distributech
Europe 2000, Vienna, Austria, Oct. 2000, p. 12.
Jouni K. Pylvninen was born in Pieksmki, Finland, on August 7, 1978. He received the M.Sc. degree in electrical engineering from Tampere University of Technology, Tampere, Finland, in 2002, where
he is currently pursuing the Ph.D. degree.
He has been conducting research work at the Institute of Power Engineering, Tampere University of
Technology, since 2003. The focus of his research
work is currently in distribution network component
condition modeling and asset management.
207
Pekka Verho was born in Hauho, Finland, in February 1966. He received the M.Sc. and Dr. of Engineering degrees in electrical engineering from Tampere University of Technology, Tampere, Finland, in
1991 and 1997, respectively.
From 1990 to 1997, he was a Research Engineer
with Tampere University of Technology. Currently,
he is a Research Manager with ABB, working in
ABB distribution automation. He is also a Part-Time
Professor with Tampere University of Technology.
His research interest is information system applications for electricity distribution.