You are on page 1of 9

EUROPEAN TRANSACTIONS ON ELECTRICAL POWER

Euro. Trans. Electr. Power (2011)


Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/etep.560

Application of interactive multiple models in thermal


modeling of power transformers

Asghar Akbari, Mehdi Allahbakhshi*,y and Reza Golpari


Department of Electrical Engineering, K.N. Toosi University of Technology, Tehran, Iran

SUMMARY
Power transformers are one of the most expensive equipments in power systems. The efficiency and optimal
utilization of these equipments are very important issues. Some factors, like temperature, moisture and
oxidation can affect the solid and liquid insulations and reduce the useful life of power transformer.
Temperature of a power transformer depends on the ambient temperature, present and past loading and
cooling classes. Proper thermal model is essential to predict temperature in different conditions of cooling
system. Such model leads to analyze useful life and improve loading and utilization of a power transformer.
In this contribution a thermal model for power transformers based on the new technique of Interactive
Multiple Modeling (IMM) is suggested to predict the top-oil temperature. Practical validation is carried out
with measured load and temperature data of a power transformer. Consequently the prediction error for the
conventional thermal model and the new suggested model based on IMM technique are compared.
Copyright # 2011 John Wiley & Sons, Ltd.

key words: interactive multiple modeling; power transformers; thermal modeling; top oil temperature

1. INTRODUCTION

Temperature plays a key role in insulation deterioration of power transformers. Thermal control
during service life is very important for optimum loading of these equipments. Critical points for
thermal stress in transformers are hottest spots which can predict loss of life of insulation systems.
Precise temperature estimation of the hottest spot is significant for evaluating and monitoring thermal
stresses of insulation system.
Aging wave of the electrical infrastructures and increasing failure rate of the electrical systems
have caused an increasing tendency for utilities to redesign their maintenance approaches. They are
changing maintenance strategy from fault and time based maintenances toward the condition based
maintenance applying on-line monitoring and diagnostic systems.
Temperature of the hot spot and top oil in transformer have direct relation with the ambient
temperature as well as the transformer loading condition. Nowadays direct measurement of the hot
spot temperature is possible using the new optical fiber sensors. But usage of these sensors has not still
become popular in manufacturing process and is not applicable for old transformers. Therefore,
measurement of winding temperature for thermal stress control and optimum loading of transformer
are performed indirectly using top oil temperature sensor[1]. This sensor is installed adjacent to a
thermal element (heater) and its temperature is proportional to the transformer current as depicted
in Figure 1.
Described method for temperature measurement is not precise enough in the new generation
of monitoring and diagnostic systems and novel methods based on precise thermal modeling are
necessary. Thermal aging is dependent on the operating temperature and a precise thermal model is

*Correspondence to: Mehdi Allahbakhshi, Department of Electrical Engineering, K.N. Toosi University of Technology,
Shariaati Street, Seyyed Khandan Bridge, P.O. Box: 163151355, Tehran 14317-14191, Iran.
y
E-mail: mehdiallahbakhshi@gmail.com

Copyright # 2011 John Wiley & Sons, Ltd.


A. AKBARI, M. ALLAHBAKHSHI AND R. GOLPARI

Figure 1. Indirect measurement of winding (hot spot) temperature in transformer.

required to predict the instantaneous top oil and consequently hot spot temperature to evaluate
thermal stresses in the insulation systems and loss of life (LOL)[2–3]. The other task for such model is
prediction of the temperature variation for optimum loading in order to extend lifetime of power
transformers.
Various researches have been done on developing and testing power transformer thermal models
[3–7]. These researches proposed individual linear or nonlinear models. Parameters of each
model should be estimated using parameter estimation techniques such as Least Square Error (LSE),
Prediction Error (PE), and Output Error[4]. But the developed models are less accurate than
desired as explained in Reference [8–10]. Applying each mentioned techniques, in some cases, gives
different and inconsistent results and parameters because these models do not accurately account
for dynamics of ambient temperature and cooling system modes and the thermal flow equations
in a complex system such as power transformers. Some efforts were performed to assess the
reliability of thermal models according to mentioned weak points[8,11]. In addition to these, Mao
et al. [9] found that the conventional methods are vulnerable against the embedded noise in
measured data.
In this contribution a novel approach based on Interactive Multiple Modeling (IMM) is proposed
for modeling thermal behavior of power transformers. It combines several thermal models to have an
accurate temperature prediction of top oil temperature and the hottest spot.

2. TRANSFORMER LINEAR THERMAL MODEL

A common thermal model which is used to predict top oil temperature is the model basically suggested
by IEEE and IEC[2–3]. Some improvements have been performed on the mentioned model for a
better and more accurate result[5–6]. The differential equation governing the top oil temperature rise,
which describes the dynamics of transformer load and the ambient temperature, can be written as
follows:

dutop
To ¼ utop þ uamb þ uu (1)
dt

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
IMM IN THERMAL MODELING OF POWER TRANSFORMERS

the solution of the differential equation gives:


  
utop ¼ uu þ uamb utop 1eðt=T0 Þ þ utop;i (2)

while uu and T0 are calculated as:


 2 n
K Rþ1
uu ¼ ufl (3)
Rþ1

Cufl
T0 ¼ (4)
Pfl

The parameters of the equations are defined as:


utop: top-oil temperature (8C),
ufl: top-oil rise over ambient temperature at rated load (8C),
uu: ultimate temperature rise (8C),
uamb: ambient temperature (8C),
utop,i: the initial value of TOT at t ¼ 0 (8C),
T0: thermal time constant at rated KVA (hours),
Pfl: full load power (Watt),
C: thermal capacity (Wh/8C),
n: a number between 0.8 up to 1 according to oil cooling system modes,
R: ratio of load loss to no-load loss at rated load,
K: ratio of specified load to rated load (I/Irated),
utop;i : initial top oil temperature.
Discretization of Equation (1) using the Euler backward approximation leads to:
dutop ½k utop ½kutop ½k1
 (5)
dt Dt
substituting Equations (3), (4) and (5) in Equation (1) gives the following recursive equation:
 2
Dtufl R I½k
utop ½k ¼ ToTþDt
o
utop ½k1 þ ToDt
þDt u amb ½ k  þ ðTo þDtÞðRþ1Þ Irated
(6)
Dtu
þ ðT0 þDtÞðflRþ1Þ ¼ K1 I 2 ½k þ K2 uamb ½k þ K3 utop ½k1 þ K4

The parameters K1 to K4 in Equation (6) are linear coefficients and can be estimated using
Least Square Error (LSE) method when measured data utop, load current I, and the ambient temperature
uamb for a period of time are available. Basically K2,K3 are dependent values but, to have a model with
more flexibility and degrees of freedom, they are assumed independent parameters.

3. INTERACTIVE MULTIPLE MODELING

The parameters of the thermal model which is described in the last section have significant variations
during service life because of several reasons such as noise, nonlinearity in the real model and
changes in cooling modes[8–11]. Variation of K1 to K4 was calculated using Recursive Least Square
with forgetting factor (RLSF) method as depicted in Figure 2. These fluctuations influence the
accuracy of model and lead erroneous results if the initial model is used for prediction of top oil
temperature.
In the proposed method, the history of model adaptation is calculated and if a condition comes
up which is closer to the previous conditions, the adjustment is performed again gradually and in
the meantime the prediction error will decrease. Several prepared models, optimized in various
time intervals, are applied simultaneously for accurate prediction of thermal behavior of power
transformers. This technique is known as Multiple Modeling as described in Reference[12]. For a brief

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
A. AKBARI, M. ALLAHBAKHSHI AND R. GOLPARI

Figure 2. Changing of the model parameter in 1000 hours of loading.

description to IMM, assume a time series xt, where t ¼ 0,1,. . .. For simplicity we take xt as a scalar and
assuming the existence of K predictors of the general form:
^xkt ¼ f ðxt1 ; . . .; xtM ; vk Þ;k ¼ 1; 2; . . .; K (7)
so M previous measured outputs are used to predict xt and its value depends on the applied model. For
the proposed model in Equation (6), M equals to 1.
Each predictor can be estimated from measured data of different intervals and related conditions
(i.e., different cooling modes OFAF, OFAN, and ONAN). The K predictors belong to a general family
f ð:; vÞwhere v is a parameter vector and the kth predictor is obtained by settingv
_k ¼ vk .
Assume the difference between actual observation xt and its kth prediction xt is a random variableekt .
The prediction error ekt forms a sequence of independent, identically distributed random variables with
zero mean. In the other words, if we denote mathematical expectation by E() then:
   
E ekt ¼ 0;E ekt eks ¼ s 2 :dðt; sÞ (8)
where d(t, s) is the Kronecker delta function. Let us also denote the probability density function of ekt by
gk ð:Þ (independent of t). Now, for the probability density of xt ^xkt and observation vectorx1 ; x2 ; :::; xt1 :
   
p xt ^xkt jx1 ; x2 ; . . .; xt1 ¼ g xt ^xkt
(9)
¼ gðxt f ðxt1 ; . . .; xtM ; vk ÞÞ

Assuming for all k, ekt has zero mean and Gaussian distribution with s k standard deviation. From
Bayes’ rule the following recursion equation can be written[12]:
 
;...;xtM ;vk Þj2
pkt1 :exp  jxt f ðxt1 2s 2
pkt ¼ K  k
 (10)
P k jxt f ðxt1 ;...;xtM ;vj Þj
2

pt1 :exp  2s 2
j¼1 j

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
IMM IN THERMAL MODELING OF POWER TRANSFORMERS

Figure 3. Block diagram of IMM method used for transformer thermal modeling.

Referring to Equation (10), three factors determine the value ofpkt . First, there is the absolute value
of the current prediction errorjxt f ðxt1 ; . . .; xtM ; vk Þj. Large prediction error results in smallpkt .
The next factor is the previous predictive performance of the kth predictor (pkt1 ). The last factor iss k ,
the larger s k causes the model to react slowly to the changes. It should be noticed thatpkt is not the
absolute predictive accuracy of the kth predictor, but the relative one. Even if the product of the
numerator terms is small, pkt can still remain close to one when the sum of products for other predictors
is small enough. If the kth predictor performs poorly but relatively better than the remaining predictors,
it will maintain a high posterior probability.
To have a more accurate prediction of desired values, the procedure of parameter calculation
is performed in different time intervals and the calculated models (and related parameters) are stored
in a reference table in order to be used appropriately. The question is ‘‘how many models have to be
considered?’’ and ‘‘how they can be selected in different time epochs as the best model?’’.
A key role in the IMM is the number of models to be combined. Thermal models of transformer in
different conditions are somewhat similar to each other. Therefore, limited number of models is
sufficient for the combination. Using numerous models lead to a time consuming procedure for
selecting procedure and huge data processing without a considerable improvement in prediction result.
Two methods have been suggested for implementing of multiple models during the prediction
procedure: Switching Multiple Modeling (SMM) and Interactive Multiple Modeling (IMM).
The SMM method uses the best predictor for the time series in each instance while the IMM uses
all predictors or models with appropriate probabilistic weighting functions to calculate the final
prediction. For the purpose of thermal modeling the IMM method showed more acceptable results and
so this method is utilized.
For K selected models, IMM method calculates conditional probabilitypkt for each model[12] and the
final prediction xkt is obtained by the summation of weighted predictions as follows:
X 
xkt ¼ xkt :pkt (11)

where xkt is the prediction of xkt made by the kth model or predictor. Figure 3 shows the prediction
procedure applying the IMM.

4. EXPERIMENTAL RESULTS

To investigate the validation of suggested thermal modeling method, 1000 samples of loading and
temperature data within 1000 hours for a 180 MVA, 63/230 kV transformer at a substation in Tehran,
Iran (Figure 4) during July to September were applied.
The measured top oil, ambient temperature and loading data for 1000 hours are depicted in Figure 5.

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
A. AKBARI, M. ALLAHBAKHSHI AND R. GOLPARI

Figure 4. 180 MVA, 63/230 kV transformer used for investigating thermal modeling.

At the first stage 800 initial data points were used to estimate the conventional model parameters
using least square method. This conventional thermal model is used to predict the top oil temperature
for the next points of 800 up to 850 hours as depicted in Figure 6. Root mean square error (RMSE)
criterion was used for comparing the results.
The conventional thermal model cause RMSE in the range of 1.5057 (8C) and a maximum error of
3.88 (8C) for depicted data in Figure 6.
At the second stage the 800 initial data points were divided into four separated time intervals and the
parameters of four individual models were calculated using the observed data for each time interval.
Calculated parameters are presented in Table I. To initiate probabilities (pkt1 ) for each model, the

Figure 5. The measured data for 1000 hours used for model validation.

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
IMM IN THERMAL MODELING OF POWER TRANSFORMERS

Figure 6. The measured and predicted temperature using conventional thermal modeling.

Table I. Parameter sets calculated for various time intervals.


Time interval K1 K2 K3 K4 Model no.
1–200 7.0986 0.1038 0.7682 8.0302 M1
201–400 8.3315 0.0638 0.7285 11.0607 M2
401–600 7.5726 0.0811 0.7642 8.9036 M3
601–800 6.7716 0.1565 0.7387 8.0467 M4

proposed algorithm were executed for the last 50 hours data (samples of 750–800 points). At the end of
this step, initial probabilities for each model were defined as shown in Figure 7. Proper models have the
higher probabilities and vice versa.
The IMM was applied with the four estimated parameter sets as four different models to predict the
top oil temperature for the next points of 800 up to 850 hours. Results are depicted in Figure 8.
By utilizing IMM method, the RMSE and the maximum error were decreased to 0.98 (8C) and
1.31(8C), respectively, which are meaningfully desirable comparing with the calculated performances
from the conventional method for the same data (RMSE ¼ 1.838C and maximum error ¼ 4.09 8C).

Figure 7. Running the algorithm to calculate the initial probabilities of multiple models.

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
A. AKBARI, M. ALLAHBAKHSHI AND R. GOLPARI

Figure 8. The measured and predicted temperature using IMM and LS methods.

5. CONCLUSIONS

Accuracy and adequacy of the predicted top-oil temperature is an essential task in loss of life
calculation and diagnostic systems of power transformers. A 2 or 3% improvement in top oil and hot
spot temperatures allows to optimally plan transformer purchases and planners can save million
dollars[13]. Due to some weakness in the conventional models, a combination methodology for models
based on Interacting Multiple Model was proposed. This method applies several locally (on
conditionally) fitted models using Bayes’ rule to have a more accurate and reliable top-oil temperature
prediction. The new method of IMM which is suggested in this paper showed considerable
improvement in prediction accuracy for thermal behavior of power transformer.

6. LISTS OF SYMBOLS AND ABBREVIATIONS

6.1. Symbols
C thermal capacity (Wh/8C)
K ratio of specified load to rated load (I/Irated)
I load current (Ampere)
n a number between 0.8 up to 1 according to oil cooling system modes
p probability density
Pfl full load power (Watt)
R ratio of load loss to no-load loss at rated load
T0 thermal time constant at rated KVA (hours)
uamb ambient temperature (8C)
ufl top-oil rise over ambient temperature at rated load (8C)
utop top-oil temperature (8C)
utop,i initial top-oil temperature
utop,i the initial value of TOT at t ¼ 0 (8C)
uu ultimate temperature rise (8C)
s standard deviation

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep
IMM IN THERMAL MODELING OF POWER TRANSFORMERS

6.2. Abbreviations
IMM Interactive multiple model
LSE Least square error
RLSF Recursive least square with forgetting factor
RMSE Root mean square error
SMM Switching multiple modeling

REFERENCES
1. Swift G, Molinski TS, Lehn W. A fundamental approach to transformer thermal modeling—Part II: field
verification. IEEE Transactions on Power Delivery 2001; 16(2):176–180.
2. IEC. Loading guide for oil-immersed power transformers. IEC-60354, 1991–2009.
3. IEEE. IEEE guide for loading mineral-oil-immersed transformers. IEEE Std C57.91.1995, Corrigendum, 2002.
4. Terry Morris A. Comparing parameter estimation techniques for an electrical power transformer oil temperature
prediction model. NASA Technical Memorandum 208974, 1999.
5. Lesieutre BC, Hagman WH, Kirtley JL. An improved transformer top oil temperature model for use in an on-line
monitoring and diagnostic system. IEEE Transactions on Power Delivery 1997; 12(1):249–256.
6. Swift G, Molinski TS, Lehn W. A fundamental approach to transformer thermal modeling—Part I: theory and
equivalent circuit. IEEE Transactions On Power Delivery 2001; 16(2):171–175.
7. Susa D, Lethonen M. Dynamic thermal modeling of power transformers: further development—Part I. IEEE
Transactions on Power Delivery 2006; 21(4):1961–1970.
8. Mao X. Transformer linear thermal models. Ph.D. Dissertation, Arizona State University 2005;Tempe.
9. Mao X, Tylavski DJ, McCulla GA. Assessing the reliability of linear dynamic transformer thermal modeling. IEE
Proceedings Generation, Transmission and Distribution 2006; 153(4):414–422.
10. Rivera LJ, Tylavski DJ. Acceptability of four transformer top-oil thermal model—Part II: comparing metrics. IEEE
Transactions on Power Delivery 2008; 23(2):866–872.
11. Rivera LJ, Tylavski DJ. Acceptability of four transformer top-oil thermal model—Part I: defining metrics. IEEE
Transactions on Power Delivery 2008; 23(2):860–865.
12. Petridis V, Kehagias A, Petrou L, Bakirtzis A, Kiartzis S, Panagiotou H, Maslaris N. A Bayesian multiple models
combination method for time series prediction. Journal of Intelligent and Robotic Systems 2001; 31(1–3):69–89.
13. Rivera LJ, Mao X, Tylavski D. Improving reliability assessment of transformer thermal top-oil model parameters
estimated from measured data. IEEE Transactions on Power Delivery 2009; 24(1):169–176.

Copyright # 2011 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2011)
DOI: 10.1002/etep

You might also like