Energy and Power Engineering, 2012, 4, 41-46

http://dx.doi.org/10.4236/epe.2012.41006 Published Online January 2012 (http://www.SciRP.org/journal/epe)

Power Transformer Top Oil Temperature Estimation with

GA and PSO Methods
Mohammad Ali Taghikhani
Engineering Department, Imam Khomeini International University, Qazvin, Iran
Email: [email protected]

Received December 8, 2011; revised December 30, 2011; accepted January 15, 2012

ABSTRACT
Power transformer outages have a considerable economic impact on the operation of an electrical network. Obtaining
appropriate model for power transformer top oil temperature (TOT) prediction is an important topic for dynamic and
steady state loading of power transformers. There are many mathematical models which predict TOT. These mathe-
matical models have many undefined coefficients which should be obtained from heat run test or fitting methods. In this
paper, genetic algorithm (GA) and particle swarm optimization (PSO) are used to obtain these coefficients. Therefore, a
code has been provided under MATLAB software. The effects of mentioned optimization methods will be studied on
improvement of adequacy, consistency and accuracy of the model. In addition these methods will be compared with the
Multiple-Linear Regression (M-L R) to illustrate the improvement of the model.

Keywords: Top-Oil Temperature (TOT); Genetic Algorithm (GA); Particle Swarm Optimization (PSO); Multiple
Linear Regression (M-L R)

1. Introduction is the instability of obtained coefficients from similar

experimental data. In this paper, the objective is propos-
Large power transformers are the most valuable assets in ing appropriate methods in order to attain consistence
electrical power networks. In order to improve trans- coefficients. To prove the efficiency of PSO and GA in
former utilization without thermal criteria violation such decreasing the range of coefficients changes metrics in-
as top oil temperature (TOT), and hottest spot tempera- troduced in [7] are used to assess adequacy, consistency,
ture (HST), TOT and HST need to be predicted accu- and accuracy of the model. Therefore, a code has been
rately in dynamic loading of transformer and maximum provided under MATLAB software. The organization of
steady state loading (SSLmax) [1,2]. Accurate TOT and the paper is as follows; mathematical models are studied
HST prediction allows system planners to plan optimally in Section 2, algorithms used for defining coefficients
for transformer purchases. Planners can save millions of through experimental data are discussed in Section 3,
dollars if even two or three percent improvement can be Section 4 illustrates coefficients obtained from algori-
achieved in TOT and HST prediction [3]. Some mathe- thms and finally, the model is evaluated in Section 5.
matical models are introduced for predicting TOT. Un-
defined coefficients of these mathematical models can be 2. Mathematical Models
obtained from heat run experiment or fitting methods
through experimental data such as multiple linear regres- 2.1. Top-Oil Temperature Rise over Ambient
sion method and optimization methods like PSO and GA Temperature
which will be studied in this paper. Other choices for This is a classical model for predicting TOT of power
TOT modeling are Neural Networks (NN) [4] and neuro- transformer. TOT rise over ambient temperature is de-
fuzzy systems [5,6]. Neural Networks methods are not fined in a differential equation as below [8]:
based on mathematical expression between TOT and
d 0
other variables but only are used for an appropriate map- T0   0  u (1)
ping among inputs and outputs. dt
In this paper, three models are introduced for predict- where 0 is top-oil temperature rise over ambient tem-
ing TOT. Then, GA and PSO are used so to define coef- perature, T0 is time constant at nominal load and u is
ficients of models through experimental data. One of the ultimate top-oil temperature rise due to load and is ex-
main challenges of power transformers thermal modeling pressed as the following equation:

Copyright © 2012 SciRes. EPE

42 M. A. TAGHIKHANI

 I 2  R 1
n
top  k   top  k  1
 u   fl    (2)
 R 1  t  I  k   R  1  1  (9)
2 1

      fl n  top  k    amb  k  n 
where fl is top-oil temperature rise over ambient tem- T0  R  1  
perature at nominal load, R is the ratio of load loss at
Because of the form of nonlinearity in Equation (9),
rated load to no-load loss, I is ratio of the specified load
top[k] appears implicitly on both sides of the equation
to rated load and n is oil cooling state exponent. Assum-
which makes training much difficult [7]. None of these
ing n ≈ 1, applying Euler discretion rule and after sim-
models perform adequately when using parameters ach-
plifying, TOT rise over ambient temperature is given in
ieved from test report. However, all of these models per-
below equation:
form adequately when their parameters are selected to
T0 t   fl  R optimally fit measured data [7]. When exponential coef-
0  k    0  k  1   I 2 k 
T0  t T0  t    R  1 ficient of oil is one, it means that pumps and fans are
(3) working at rated condition; in this situation swift model
t   fl
 is equal to the nonlinear top-oil model. Since fans and
T0  t    R  1 pumps are ON during the experiment, both of the models
And by substituting coefficients k1, k2 and k3: would have the same discrete model. The performance of
first model is not acceptable due to excluding variations
 0  k   k1  I 2  k   k2   0  k  1  k3 (4) of environmental temperature. It was mentioned that
Swift model and nonlinear model are equal so Equation
This simplified model does not take dynamic variation (6) is used as final model and coefficients will be ob-
of ambient temperature on TOT into account and in addi- tained through experimental data.
tion model accuracy is not acceptable.
3. Optimization Algorithms
2.2. Nonlinear Top-Oil Model
In this paper, three methods are used so to obtain the
Nonlinear Top-Oil Model proposed in [9,10] explains coefficients of Equation (6). Our target is comparing the
dynamic variation in ambient temperature and is defined results of proposed methods as well as the improvements
as below equation: in the limiting of variations in the coefficients.
dtop
T0  top  u   amb (5) 3.1. Multiple-Linear Regression (M-L R)
dt
The method used for multiple-regression is an extension
In fact, this model is the correlated form of the model
of which used for single regression. For a model with
proposed in IEEE. To use Euler discretion method and n
three independent variables and in scalar form, output is
≈ 1 we have:
rewritten as:
T0 t 
top  k   top  k  1   amb  k  y  y    k1  x1  k2  x2  k3  x3  k4   (10)
T0  t T0  t
where:
t   fl  R t   fl
  I k  
2 x1: Load value;
T0  t    R  1 T0  t    R  1 x2: Ambient temperature;
 k1  I 2  k   k2   amb  k   1  k2   top  k  1  k3 x3: TOT(k − 1);

y : Predicted TOT(k);
(6) y: Actual TOT(k);
: Error of prediction;
2.3. Swift Model k1, k2, k3: Coefficient to be determined.
Swift model proposed in [11] is the change in exponen- In vector form, Equation (10) can be written
tial coefficient of oil in nonlinear model: 
Y  Y    X  K  K0   (11)
dtop 1 
T0    top   amb   u
n (7) where Y is a 3  1 vector, X is a 3  k matrix of sampled
dt
variables, K is a k  1 vector of the coefficients, K0 is a 3
 k vector of constant scalar values and E is a 3  1 vec-
n
 1
 I 2  R 1
u   fnl    (8) tor of random errors. In order to determine coefficients,
 R 1 
those values are selected when the squared error between
After discretion of Equation (7): the actual TOT and the predicted TOT is minimized.

Copyright © 2012 SciRes. EPE

 M. A. TAGHIKHANI 43

This criterion can be expressed as: best previous position.

 2  Sociality: the tendency to move towards the neighbor-
min   yi  y   min   i2 (12)
hood’s best previous position.
The least-squares estimate coefficients as below: The velocity of each particle in the swarm is updated
by using the following equation:
K    X T  X 
1
 X T Y (13) 
vi t  1    c1  r1   xi  t   xi  t  
In order to find the coefficients that minimize the (15)
 c2  r2   g  t   xi  t  
squared error, Equation (12) can be solved with optimi-
zation algorithms such as GA and PSO. where,
vi(t): the velocity of particle i at time t;
3.2. Genetic Algorithm xi(t): the position of particle i at time t;
The genetic algorithm (GA) is an optimization and search w, c1, c2: user-supplied coefficients;
technique based on the principles of genetics and natural r1, r2: random values regenerated for each velocity
selection. GA allows a population composed of many update;
individuals to evolve under specified selection rules in a xˆi (t ) : the individual best candidate solution for parti-
state that maximizes or minimizes the fitness function cle i at time t;
[12,13]. N data-sets are selected in specific domain. g(t): the best swarm’s global candidate solution at time
Data-sets are substituted in fitness function and they are t.
scaled and then all of them are scored. Children of the The objective function is the same function in GA
next generation can be produced from parents in the cur- method.
rent generation according to the following methods:
 Selection of parents (Elite). 4. Coefficients Calculation
 Cross over. In order to increase precise of the model, all experimen-
 Mutation. tal data is converted to per unit values. Nominal values
The procedure will continue until one of the stopping are:
criteria is met. Some of the stopping criteria are dis-
cussed in below: P = 5 (MVA)
 A solution is found that satisfies minimum criteria. TOT = 55 (˚C)
 Allocated budget (computation time) reached.
 Fixed number of generations reached. ambient = 35 (˚C)
 Successive iterations no longer produce better results. Using per unit values is an essential fact which is
In this paper, fitness function is defined as the follow sometimes forgotten. For example, if values are not con-
equation: verted to per unit, load will have greater value comparing
Fittness Function to temperature. In this case, changes of load would be
greater comparing to thermal changes which results in-
i N i N (14)
   TOTpredicted  i   TOTactual  i       i 
2 2
appropriate effect of temperature on the model. In addi-
i 1 i 1 tion, using per unit values would decrease the range of
coefficients changes. Coefficients obtained from PSO,
where, N is the number of data used for finding of the
GA and linear regression in five time intervals are given
coefficients.
in Table 1. k1 is the most important coefficient since it
shows the effect of load on temperature. The range of
3.3. Particle Swarm Optimization (PSO)
coefficients changes is limited in an appropriate model.
Algorithm
In order to obtain global coefficients, all data is used.
In the PSO algorithm each individual is called a particle, Results are given in Table 2. In order to show the per-
and it is subjected to a movement in a multidimensional formance of the model as well as its precision and accu-
space that represents the belief space. Particles have racy in TOT modeling, limited data numbers are used to
memory and thus retain part of their previous state. There predict TOT in all ranges and then results are compared
is no restriction for particles to share the same point in with the original values. Estimated and actual values of
belief space, but their individuality is preserved in any TOT are compared in Figure 1. Table 3 shows error of
case. Each particle’s movement is the composition of an predicted TOT.
initial random velocity and two randomly weighted in- The load range used for testing the model is wide and
fluences [14]: results show that PSO has a good performance. Mean
 Individuality: the tendency to return to the particle’s relative error is not the only parameter which evaluates

Copyright © 2012 SciRes. EPE

44 M. A. TAGHIKHANI

Table 1. Coefficients of the model.

No. Data-set k4 k3 k2 k1 k1 Tolerance

1 0.05293951 0.04059169 0.39459100 0.51542845

2 0.05306756 0.04860554 0.39221286 0.50906984

PSO 3 0.05158763 0.10697031 0.36450539 0.48447931 2.99%

4 0.05382795 0.04260348 0.39145473 0.51444735

5 0.05185671 0.04227057 0.39546324 0.51368205

1 0.05112283 0.15620790 0.34168280 0.45287547

2 0.04491779 0.14171454 0.35643632 0.46037598

GA 3 0.01939867 0.20874699 0.34582685 0.43172918 3.75%

4 0.04911735 0.11572956 0.36144769 0.47483254

5 0.04426151 0.19501470 0.33340073 0.42992850

1 0.05293929 0.04059931 0.39458796 0.51542408

2 0.05306770 0.04860280 0.39221415 0.50907125

Multiple
3 0.05161140 0.06322293 0.39779668 0.47171893 5.7%
Linear-Regression
4 0.05382795 0.04060303 0.37145499 0.52544759

5 0.05185801 0.04225393 0.39546972 0.51369082

Figure 1. Comparison between predicted TOT and actual TOT.

Table 2. Obtained coefficients from all of data. Table 3. Error of predicted TOT.

k1 k2 k3 k4 Mean Relative Error (%) Max Relative Error (%)

PSO 0.53233 0.40959 0.00514 0.05289 PSO 0.0247 1.0803

GA 0.60565 0.47455 –0.14152 0.06259 GA 0.3962 2.5172

M-L R 0.53705 0.44082 0.00671 0.03300 M-L R 1.1355 1.9361

Copyright © 2012 SciRes. EPE

 M. A. TAGHIKHANI 45

performance of the model. High value of error in predic- 5.2. Consistency

tion causes a mistake in choosing the load of transformer
Consistency is a quantitative measuring of the model’s
and in many cases this mistake raises the oil temperature
ability and solution method to produce the same model
dramatically to a high degree and causes harmful effects.
parameter when training to use similar data. A standard
Predicting TOT higher than its real value makes the op-
deviation (STD) of parameters is used to examine con-
erator to reduce loading which is not harmful for trans-
sistency. For this purpose, “p” independent but similar
former. However, if the predicted value of TOT is fewer
data-sets are used for calculating coefficients. The STD
than its original value, operator would increase load and
of the model coefficients is calculated with Equation
it may cause actual TOT to become higher than its per-
(16):
mitted value while the predicted TOT is in permitted
 i 1  ki  k 
2
range. Repeating this action would accelerate the aging sp 
p
(16)
process.
Table 4 shows STD for each coefficient. It is clear that
5. Model Assessment standard deviations of coefficients obtained from opti-
mization algorithms have smaller values.
5.1. Adequacy
Adequacy measure whether the model has an appropriate 5.3. Accuracy
structure to capture the features of the process being The typical metric used for assessing model accuracy is
modeled. Residual versus fitted value diagrams are used R2. The R2 metrics measure how well the predicted val-
to examine the adequacy of the model [7]. A typical dia- ues (i.e., TOTpredicted) capture the variation of measured
gram of residual versus fitted value is shown in Figure 2. values (i.e., TOTactual):
Figure 2(a) shows example of a good adequacy but
  TOTpredicted  TOT 
n 2
Figure 2(b) means nonlinearity in the model and there
will be the need for having other variables. On the other i 1 SS R
R2   (17)
  TOTactual  TOT 
n
hand, some variables have not been considered in the 2 SST
model. R(i) verses TOT(i) is illustrated in Figure 3. It is i 1
obvious that the model is adequate and does not need where, SSR is the sum of residuals square and measures
additional variables. the variation of predicted values. The variable SST is the
total variation of measured values. The R2 value close to
1 indicates that values of the model closely match those
which are measured. Table 5 shows the values of R2 for
three models.

6. Conclusion
In this paper, three models are introduced for predicting
(a) (b) Table 4. STD of achieved coefficients.
Figure 2. Typical residual vs fitted value.
STD
Coefficient
Residual Vs. Fitted values PSO GA M-L R
2
k1 0.02611 0.03833 0.04129
1.5
k2 0.02608 0.02258 0.02145
1

0.5
k3 0.05708 0.07648 0.01923

k4 0.00184 0.02565 0.00183

R (i)

-0.5
Table 5. Obtained value of R2 from three models
-1
Model R2
-1.5
PSO 0.92609
-2
10 20 30 40 50 60 70 80 90 100
TOT (i) GA 0.92638
M-L R 0.92133
Figure 3. Achieved residual vs fitted value.

Copyright © 2012 SciRes. EPE

46 M. A. TAGHIKHANI

top oil temperature (TOT) in power transformers. GA 2000, pp. 1205-1211. doi:10.1109/61.891504
and PSO are used so to define coefficients of models [6] H. Nguyen, G. W. Baxter and L. Reznik, “Soft Comput-
through experimental data. PSO algorithm leads to the ing Techniques to Model the Top-Oil Temperature of
best performance with considering the achieved results. Power Transformers,” International Conference on Intel-
The main success is limited ranges of coefficients espe- ligent Systems Applications to Power Systems (ISAP),
Taiwan, 5-8 November 2007, pp. 1-6.
cially for k1 (effect of load on temperature). In addition, doi:10.1109/ISAP.2007.4441618
mean relative error becomes near zero. In the paper was
[7] L. Jauregui-Rivera and D. J. Tylavsky, “Acceptability of
shown that nonlinear model is a good model itself, but Four Transformer Top-Oil Thermal Models—Part I: De-
obtaining the coefficients with traditional method cause fining Metrics,” IEEE Transactions on Power Delivery,
inappropriate performances of the nonlinear model. New Vol. 23, No. 2, 2008, pp. 860-865.
optimization algorithms improve performances of this doi:10.1109/TPWRD.2007.905555
model related to multi-linear regression. Additionally, it [8] IEEE Standard, C57.91-1995, “IEEE Guide for Loading
was depicted that using optimization algorithms im- Mineral Oil Immersed Transformer,” 1996.
proves the model adequacy, consistency as well as accu- [9] B. C. Lesieutre, W. H. Hagman and J. L. Jr. Kirtley, “An
racy. Improved Transformer Top Oil Temperature Model for
Use in an On-Line Monitoring and Diagnostic System,”
IEEE Transactions on Power Delivery, Vol. 12, No. 1,
REFERENCES 1997, pp. 249-256. doi:10.1109/61.568247
[1] M. A. Taghikhani and A. Gholami, “Temperature Distri- [10] D. J. Tylavsky, X. L. Mao and G. A. McCulla, “Trans-
bution in ONAN Power Transformer Windings with Fi- former Thermal Modeling: Improving Reliability Using
nite Element Method,” European Transactions on Elec- Data Quality Control,” IEEE Transactions on Power De-
trical Power, Vol. 19, No. 5, 2009, pp. 718-730. livery, Vol. 21, No. 3, 2006, pp. 1357-1366.
doi:10.1002/etep.251 doi:10.1109/TPWRD.2005.864039
[2] Z. Radakovic, “Numerical Determination of Characteris- [11] G. Swift, T. Molinski, W. Lehn and, R. Bray, “A Funda-
tic Temperatures in Directly Loaded Power Oil Trans- mental Approach to Transformer Thermal Modeling—
former,” European Transactions on Electrical Power, Part I: Theory and Equivalent Circuit,” IEEE Transac-
Vol. 13, No. 1, 2003, pp. 47-54. tions on Power Delivery, Vol. 16, No. 2, 2001, pp. 171-
doi:10.1002/etep.4450130107 175. doi:10.1109/61.915478
[3] L. Jauregui-Rivera, X. L. Mao and D. J. Tylavsky, “Im- [12] R. L. Haupt and S. E. Haupt, “Practical Genetic Algo-
proving Reliability Assessment of Transformer Thermal rithms,” 2nd Edition, John Wiley & Sons Inc. Publication,
Top-Oil Model Parameters Estimated from Measured Hoboken, 2004.
Data,” IEEE Transactions on Power Delivery, Vol. 24, [13] V. Galdi, L. Ippolito, A. Piccolo and A. Vaccaro, “Pa-
No. 1, 2009, pp. 169-176. rameter Identification of Power Transformers Thermal
doi:10.1109/TPWRD.2008.2005686 Model via Genetic Algorithms,” Electric Power Systems
[4] K. P. Jouni, K. Nousiainen and P. Verho, “Studies to Research, Vol. 60, No. 2, 2001, pp. 107-113.
Utilize Loading Guides and Ann for Oil-Immersed Dis- doi:10.1016/S0378-7796(01)00173-0
tribution Transformer Condition Monitoring,” IEEE Trans- [14] W. H. Tang, S. He, E. Prempain, Q. H. Wu and J. Fitch,
actions on Power Delivery, Vol. 22, No. 1, 2007, pp. 201- “A Particle Swarm Optimizer with Passive Congregation
207. doi:10.1109/TPWRD.2006.877075 Approach to Thermal Modeling for Power Transformers,”
[5] Q. He, J. Si and D. J. Tylavsky, “Prediction of Top-Oil The 2005 IEEE Congress on Evolutionary Computation,
Temperature for Transformers Using Neural Networks,” Vol. 3, 2005, pp. 2745-2751.
IEEE Transactions on Power Delivery, Vol. 15, No. 4,

Copyright © 2012 SciRes. EPE