Sie sind auf Seite 1von 9

Swarm and Evolutionary Computation 36 (2017) 1826

Contents lists available at ScienceDirect

Swarm and Evolutionary Computation

journal homepage:

Estimation of transformer parameters from nameplate data by imperialist MARK

competitive and gravitational search algorithms

H.A. Illiasa, , K.J. Moua, A.H.A. Bakarb
Department of Electrical Engineering, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
UM Power Energy Dedicated Advanced Centre (UMPEDAC), Level 4, Wisma R & D UM, University of Malaya, Jalan Pantai Baharu, 59990 Kuala Lumpur,


Keywords: Accurate determination of parameters in power transformer equivalent circuit is important because it can
Power transformers inuence the simulation results of condition monitoring on power transformers, such as analysis of frequency-
Imperialist competitive algorithm response. This is due to inaccurate simulation results will yield incorrect interpretation of the power
Gravitational search algorithm transformer condition through its equivalent circuit. Works on development of transformer models have been
Swarm optimisation
widely developed since the past for transient and steady-state analyses. Estimating parameters of a transformer
Equivalent circuit
using nameplate data without performing a single experiment has been developed in the past. However, the
average error between the actual and estimated parameter values in the past work using Particle Swarm
Optimisation (PSO) and Genetic Algorithm (GA) is considerably large. This signies that there is a room for
improvement by using other optimisation techniques, such as state of the art methods which include
Heterogeneous Comprehensive Learning PSO (HCLPSO), LSHADE-EpSin, Imperialist Competitive Algorithm
(ICA), Gravitational Search Algorithm (GSA) and others. Since ICA and GSA have advantages over GA and PSO,
in this work, estimation of transformer parameters from its nameplate data was proposed using ICA and GSA.
The results obtained using ICA and GSA was compared to those using GA and PSO to determine the parameters
of transformer equivalent circuit. The results show that GSA performs the best as it gives the lowest average
error compared to PSO, GA and ICA. Therefore, the proposed technique using GSA and ICA can give a better
accuracy than PSO and GA in estimating the parameters of power transformers. The proposed method can also
be applied to estimate parameters of three-phase transformers from their nameplate data without disconnecting
them from the grid for testing.

1. Introduction mental testing using optimisation methods has been developed in the
past [1,1518]. Optimisation methods have become popular among
Power transformer is one of the most important components in power system applications recently [1921].
power system networks. When there is a transformer failure, the eects Mohamed et al. developed a method to determine the parameters of
are very signicant, which include power outage and costly repairs. one-phase transformer equivalent circuit from its nameplate data with
Each transformer has its own transfer function, which is widely used to the absence of experiment test using evolutionary computations
analyse its internal condition without dismantle its part. This transfer techniques [1]. In this work, PSO and GA were employed to minimise
function can be measured using analysis of frequency-response (FRA) the objective function, which is the sum of square error of the
or determined using equivalent circuit of a transformer. Therefore, transformer nameplate data. The square error is calculated as the
accurate estimation of parameters in power transformer equivalent dierence between the actual nameplate data and estimated data. The
circuit is of great importance because it can inuence the condition results show that GA and PSO managed to determine the equivalent
monitoring method of power transformers [15]. Since the past, many circuit parameters of transformers with good accuracy. It was also
works have been performed to construct transformer modelling for found that the convergence rate of PSO is faster compared to GA. PSO
transient and steady-state analyses [610] and to analyse physical and GA have been widely used since the past as optimisation algorithm
phenomena occurring in power transformers [1114]. Estimating in various applications, including power engineering [2229].
transformer parameters from nameplate data without physical experi- Meister et al. presented the use of least squares method to estimate

Corresponding author.
E-mail address: (H.A. Illias).
Received 17 October 2016; Received in revised form 14 March 2017; Accepted 31 March 2017
Available online 07 April 2017
2210-6502/ 2017 Elsevier B.V. All rights reserved.
H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

the model parameters of a transformer [30]. This method employs R1 jX1 I2' R2' jX2'
polynomial function approximations to track the transformer model
parameters by minimising the distance between the approximation and I1 I0
real functions. Third degree polynomial was found to be corresponded Ic Im
to optimum estimation of the parameters of magnetizing ux of Joule V1 Rc jXm V 2'
losses and leakage ux while polynomial of second degree can
represent the core losses-related parameter. However, the model that
has been developed cannot be generalized for other similar transformer
Fig. 1. Equivalent circuit of a transformer by referring to the primary winding.
ratings. For example, when this model was applied for other similar
transformer ratings, there is a large error between the estimated
parameter values and the experimental data. This is due to the alternative method to determine the power transformer parameters
equation of Joule losses, dispersion ux, magnetizing ux and core from its nameplate data without measurement in practice.
losses are not the same for dierent units of transformer with similar
ratings. 2. Transformer equivalent circuit and optimisation
Based on Soliman et al. work in [7], an alternative method over techniques
conventional short and open circuit tests to determine the N-windings
transformer parameters at power frequency in online mode was 2.1. Transformer equivalent circuit
developed. The method was based on linear least errors square
algorithm, which used digitized samples of the input voltage and Fig. 1 shows a two-winding transformer equivalent circuit under
current and the output voltage and current of the transformer wind- stable condition by referring to the primary side, where
ings. From this method, the transformer parameters that have been
successfully estimated are the winding and core parameters. R1 = Primary winding resistance
Thilagar et al. have also performed three-winding transformer R2 = Secondary winding resistance referred to primary side
parameter estimation using genetic algorithm (GA) [31]. They for- X1 = Primary leakage reactance
mulated the objective function to be minimised as the sum of square X2 = Secondary leakage reactance referred to primary side
error percentage between the measured and estimated primary voltage Rc = Resistance corresponding to core losses
and input power. From the results, the performance of transformer Xm = Magnetizing reactance
parameters estimation obtained using GA at dierent load points is I1 = Primary current
quite satisfactory. From this research, it was found that GA could I2 = Secondary current referred to primary side
provide a very accurate solution by nding the global optimum V1 = Terminal voltage on the primary side
solutions whereas conventional optimisation techniques converged to V2 = Secondary terminal voltage referred to primary side
local optimum solutions. I0 = No-load current
Subramanian et al. estimated an equivalent circuit parameter of a
three-winding Transformer using GA and Bacterial Foraging Algorithm In this work, the optimum parameters of the transformer are
(BFA) [15]. The estimated equivalent circuit parameters using BFA determined using ICA and GSA. The parameters to be determined
have been used to generate load test data. The estimated parameters of are R1, X1, R2, X2, Rc and Xm. The objective function to be minimised
the transformer using BFA and GA were compared with the measure- by the optimisation methods is the summation of square error between
ment data. It is concluded that BFA gives better performance in the actual value and estimated value of all transformer nameplate
optimisation over GA because BFA has estimated the transformer nominal parameters, which are I1, I2 and V2. It can be written as
parameters with lower error than GA. The proposed method using BFA E = (I1actual I1est )2 + (I2 actual I2 est )2 + (V2 actual V2 est )2 (1)
gives advantages such as lower mathematical burden, high quality
solution, accurate estimate, less computational time and fast conver- where I1actual, I2actual and V2actual are the actual nominal name-
gence. plate values while I1est, I2est and V2est are the estimated value using
In the past, PSO and GA have been applied to estimate transformer optimisation methods. The actual values are determined from the
parameters from nameplate data without experiment tests [1]. actual open-circuit and short-circuit tests on the transformer.
However, the average error between the actual and estimated para-
meter values is between 10% and 24%, which is considerably large. 2.2. Imperialist Competitive Algorithm (ICA)
This signies that there is a room for improvement by using other
optimisation techniques, such as Imperialist Competitive Algorithm ICA is a new evolutionary algorithm for optimisation which is
(ICA) and Gravitational Search Algorithm (GSA). In [32], ICA shows inspired by imperialistic competition [3440]. Imperialism is the
considerable improvement over GA in nding the optimum solution in policy of extending the power and rule of a government beyond its
less computational time with the same population size and iterations. own boundaries. The imperialist states compete to increase the number
On the other hand, GSA has the ability to nd near global optimum of their colonies by taking possessions of each other colonies and
solution and provides better results than other nature-inspired algo- expanding their empires across the globe. This competition will result
rithms [33]. in a development of the powerful empires while the collapse of weaker
Since ICA and GSA have advantages over GA and PSO, in this work, ones. In this algorithm, the assimilation policy is modelled by moving
a method using ICA and GSA to determine single-phase transformer the colonies toward the imperialists. Imperialist and colonies are used
equivalent circuit parameters from its nameplate data without experi- to represent the possible solutions for an optimisation problem. In ICA,
ment test is proposed. ICA and GSA were applied to full the name- the most important step to reach the optimum global solution is
plate data by minimising an objective function. The proposed techni- imperialistic competition, which the weakest colony in the weakest
que was tested on three single-phase transformers with dierent empire is picked and having a competition between all empires to
ratings and also comparison with the previously reported works. possess this colony. The more powerful the empire, the more likely it
From comparison between dierent methods and the previous work will possess this aforementioned colony.
in [1], the method that can estimate the parameters of transformer For ICA, the algorithm is laid down based on the concept of
equivalent circuit with the lowest error compared to the actual values imperialism. Imperialist states will compete strongly in order to
can be identied. Hence, the proposed method could be used as an increase the number of colonies possessed by each imperialist and

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

Start recording Transformer Nameplate 2.2.2. Assimilation

Data (I1, I2, V2) In this process, all colonies will move towards the imperialist. x is a
random variable with uniform distribution which equals to
x ~ U (0, d ) (4)
Generate initial empires for R1, X1, R2, X2, RC, Xm
where is a random number greater than 1 and between 1 and 2 (1 <
2). > 1 will cause the colonies to move closer to the imperialist
Assimilation: Move colonies to imperialist from both sides and d is the distance between imperialist and colony.

2.2.3. Exchange position of imperialist and colony

When the colonies are moving towards the imperialist, the colony
Any colony in an empire has
lower cost than imperialist? may reach a spot in the search space which has a lower cost than the
imperialist. If this happens, the position of imperialist will be ex-
changed with the best colony. The algorithm will continue to move the
No colonies towards the new position of the imperialist.
Exchange position of that colony with imperialist
2.2.4. Revolution
Some colonies are randomly selected and change their positions to
Revolution allow for more freedom in the search space. The ICA parameter
p_revolve controls the degree of revolution.
The total power of an empire is calculated by
Calculate total cost of all empires
Total cost,
TCn = Cost (imperialistn ) + mean [Cost (colonies of empiren )] (5)
Imperialistic competition: Give the weakest
colony from weakest empire to the empire having where is a positive number which is less than 1. If is larger, the
highest probability to possess it colonies will play a larger role in calculating the total cost of an empire
and vice versa. The total power of an empire is inversely proportional to
No its total cost since the problem is a minimisation problem (a lower cost
Any empire with no colony? represents a better solution).

Yes 2.2.5. Imperialistic competition

Eliminate that empire Imperialistic competition between all the empires is the basis of
ICA. In this process, all empires will compete to possess the weakest
colony in the weakest empire. The total cost of n-th empire in
No normalized is given by
Termination criteria satisfied?
NTCn = TCn maxi (TCi ) (6)
where TCn is the total cost of n-th empire and NTCn is the normalized
Record best fitness value total cost of n-th empire.
and final R1, X1, R2, X2, Rc
and Xm values NTCn
ppn = N
i =1 NTCi (7)
End The possession probability of each empire is
Fig. 2. Flowchart of ICA for parameter estimation. P = [ pp1, pp2 , pp3 , ppNimp ] (8)
Also, a vector R which is the same size as P will be created, in which
also extend their empire around the world. Fig. 2 shows a owchart of
the elements are distributed random numbers uniformly.
ICA algorithm. The explanation of each step is as follows [36]:
R = [r1, r2, r3, , rNimp] (9)
2.2.1. Initialize the empires Next, vector D is formed by subtracting R from P
The algorithm starts with an initial population generated randomly.
D = P R = [D1, D2, D3, , DNimp] (10)
An array of variables which are called country will be formed. In this
work, the dimension of the optimisation problem, Nvar is 6. A country During the competition, powerless empires will collapse and the
is 1 Nvar array, dened as powerful empire possesses the colonies. Imperialistic competition aims
to converge to a state where there is only one empire left and its
Country = [R1, X1, R2, X2, Rc , Xm] (2)
colonies have the same cost as the imperialist and in the same position.
A country is divided into colonies and imperialists; and both of
them will be combined together to form a number of empires. The 2.2.6. Eliminate the weakest empire
number of imperialist, Nimp is specied in the algorithm. The number An empire is assumed to be collapsed when there is no colony left
of colonies is equal to due to imperialistic competition. As such, the powerless empire from
Ncolony = Ncountry Nimp the population is removed.

The colonies among the imperialist are divided based on their 2.2.7. Convergence
power, where the initial number of colonies in an empire is directly After a certain amount of iterations, there will be only one empire
proportional to its corresponding imperialist's power. left, as that empire is the most powerful empire among the other

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

Start 2.3.1. Initialisation

GSA begins with an initial populations consisting of masses
generated randomly. In this work, the masses are the transformer
parameters, which are
Define input data and GSA parameters
Massi = [R11, X12 , R2 3, X2 4 , Rc 5, Xm 6] (11)
for i =1, 2, Nmass
Generate initial masses Nmass is the number of masses in the population as dened by the

2.3.2. Evaluate the tness of each mass

Evaluate the fitness of each mass The tness for each mass is calculated using Eq. (1).

2.3.3. Update gravitational constant, best mass, worst mass and

Update the gravitational constant, best, worst and inertia mass at each iteration
inertia mass In order to control the searching accuracy, the gravitational
constant, G is initialized at the beginning and is reducing in value with
iteration. The equation of G is given by
Calculate the total force in iter
different direction G (iter ) = G0 exp
max iter (12)
Calculate the acceleration and velocity
iter = current iteration
= user specied constant
Go = initial value of gravitational constant
Update masses position
max_iter= total number of iterations

No The inertia mass, Mi is updated as follows by assuming all masses

Meeting end of criterion? are equal:
Mai = Mpi = Mii = Mi
fiti (iter ) worst (iter )
mi (iter ) =
Record best fitness value best (iter ) worst (iter ) (13)
and final R1, X1, R2, X2, Rc
and Xm values mi (iter )
Mi (iter ) = Nmass
j =1 mj (iter ) (14)
ti(iter) = tness value of mass i at iteration iter
Fig. 3. Flow chart of GSA.
mi(iter) = mass of i at current iteration

empires. The other empires have collapsed during imperialistic com-

Since this is a minimisation problem, the best(iter) and worst(iter)
petition. Also, the colonies in the most powerful empire will have the
are dened as
same position and same tness as the imperialist. The program will
terminate when the number of imperialist, Nimp =1 or the maximum best (iter ) = minj {1, ..., Nmass} fit j (iter ) (15)
number of iterations has reached.
worst (iter ) = maxj {1, ..., Nmass} fit j (iter ) (16)

2.3. Gravitational Search Algorithm (GSA) 2.3.4. Calculation of the total force in dierent directions
The gravitational force, Fij of mass i due to mass j at current
Gravitational Search Algorithm (GSA) was developed by Rashedi iteration iter is calculated by
et al. [4144]. GSA is a new stochastic search algorithm which is based
Mi (iter )
on mass interactions and Newton laws. In GSA, the population Fij (iter ) = G (iter ) (Massj (iter ) Massi (iter ))
Rij (iter ) Rpower + (17)
individuals are referred as masses and their performances are mea-
sured by their position masses. Each mass has position, active where
gravitational mass, inertial mass and passive gravitational mass. The
position of the mass is related to a solution of the problem. The Mi = inertial mass of mass i
gravitational and inertia mass of each mass will be determined by using Mj = the inertial mass of mass j
a tness function. All masses attract each other by a gravity force, = small constant
causing a global movement of each mass towards the heavier masses, Rij(iter) = Euclidian distance between i and j masses specied as
where they correspond to a better solution, hence it will move slower
Rij (iter ) = Massi (iter ), Massj (iter ) (18)
than the lighter mass. In this way, the exploitation of GSA is 2

Fig. 3 shows a owchart of GSA to estimate the transformer Rpower = power of Euclidean distance
parameters. The steps of GSA are explained as follows [43,45,46]:

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

2.3.5. Calculation of acceleration and velocity 3.1. 4 kVA, 50 Hz, 250/125 V, one-phase transformer
The acceleration of the mass i at current iteration, iter in dth
dimension, aid (iter) is dened by Table 1 shows the parameters of 4 kVA, 50 Hz, 250/125 V, one-
phase transformer at full load (RL =15.625) for the actual data,
Fi d (iter )
ai d (iter ) = obtained using ICA and GSA methods from this work and PSO and GA
Mii (iter ) (19)
from [1]. The actual values of transformer parameters were obtained
d from short-circuit and open-circuit tests, as reported in [1]. From this
Fi (iter) is the total force that acts on mass i of dth dimension and
is dened by table, it can be seen that GSA yields the lowest percentage of average
error for estimated transformer parameters, followed by ICA, GA and
Fi d (iter ) = j mass
kbest , j i
randj Fij d (iter )
(20) nally PSO. Average error is the best measure to compare these
algorithms because it indicates how close the results obtained using
where randj is a random number between the interval [0,1]. Kbest, a these algorithms compared to the actual measurement parameters.
function of iteration which is the set of rst K masses with the lowest Table 2 shows the full load data for 4 kVA, 50 Hz, 250/125 V, one-
power loss and biggest mass, is introduced to this algorithm to phase transformer for the actual data, obtained using ICA and GSA
compromise between exploration and exploitation. This is to ensure methods from this work and PSO and GA from [1]. These values were
that exploration must fade out and exploitation must fade in with lapse obtained using the transformer parameter values from Table 1. The
of iterations. It means all masses apply force to each other in the parameters from the actual measurement at 50 Hz from [1] are also
beginning and only one mass applies force to others in the end of the shown in Table 1 in the Actual row. From this table, it can be
iterations. Ko, the initial value of Kbest is set at the beginning and observed that the average error for transformer data at full load is the
decreased with iterations and hence Kbest decreases linearly with smallest for PSO, followed by GA, ICA and nally GSA. Although GSA
iterations. The variable nal_per in Kbest equation denes how many yields the largest average error for transformer data at full load, the
percent of masses apply force to the others in the last iteration. The average error of estimated transformer parameters using GSA is the
next velocity of a mass is calculated by lowest. However, the average error from Table 2 for GSA is still within
vi d (iter + 1) = randi vi d (iter ) + ai d (iter ) (21) reasonable range, which is less than 1%. Thus, based on the results in
Tables 1, 2, GSA can be considered to perform best in estimating the
parameter values for 4 kVA, 50 Hz, 250/125 V, one-phase transformer
2.3.6. Update masses position compared to GA, PSO and ICA.
The next position of a mass is updated by The voltage regulation under dierent loading was determined
using the estimated parameters generated by ICA and GSA and was
Massi d (iter + 1) = Massi d (iter ) + vi d (iter + 1) (22) compared with the voltage regulation calculated using the actual
transformer parameters as shown in Fig. 4. From this gure, it can
be observed that the estimated voltage regulation using GSA and ICA
2.3.7. Convergence
only diers slightly from the actual voltage regulation.
The algorithm ends when the maximum number of iterations is
After the parameters for 4 kVA, 50 Hz, 250/125 V, one-phase
transformer have been identied, they can also be used to determine
the parameters at 60 Hz. This can be done by rst calculating the value
3. Results and discussion of L for each X1, X2 and Xm at 50 Hz. Then, X1, X2 and Xm at 60 Hz
can be obtained from these values of L at 60 Hz. Hence, the actual
In this section, results of tests using ICA and GSA to estimate three parameter values of X1, X2 and Xm for 4 kVA, 60 Hz, 250/125 V, one-
dierent single-phase power transformers are presented. The results phase transformer are 0.24 , 2.4127 and 900 respectively. The
using ICA and GSA were compared with the real parameter values of values for R1, R2and Rc are the same for 50 Hz and 60 Hz transformer.
each transformer to determine the error of estimation.
For each optimisation method, dierent combinations of parameter
values were tested until the summation of square error E between the 3.2. 10 kVA, 50 Hz, 500/125 V, one-phase transformer
actual value and estimated values of all transformer nameplate
nominal parameters, calculated using Eq. (1) is the lowest. Each Table 3 shows the parameters of 10 kVA, 50 Hz, 500/125 V, one-
combination set was tested for 10 times and the average value of E phase transformer at full load for the actual data, obtained using ICA
was calculated. For ICA, dierent combinations of the movement and GSA methods from this work and PSO and GA from [1]. The
constant , revolution degree p_revolve and total cost constant were parameters from the actual measurement at 50 Hz from [1] are also
tested. For GSA, dierent combinations of gravitational constant G0, shown in Table 3 in the Actual row. Similar to 4 kVA transformer, it
iteration constant and power of Euclidean distance Rpower were can be seen that GSA yields the lowest percentage of average error,
tested. These tests were performed to demonstrate that the best followed by ICA, PSO and GA. Comparing the average error with 4 kVA
settings of ICA and GSA were used. transformer, the average error for each method is less for 10 kVA

Table 1
Parameter values for 4 kVA, 50 Hz, 250/125 V, one-phase transformer.

Parameter R1 () X1 () R2 () X2 () Rc () Xm () Average error (%)

Actual [1] 0.4 0.2 0.4 2 1500 750

PSO [1] 0.587 0.2554 0.209 1.602 1476 738
PSO error (%) [1] 46.75 27.7 47.75 19.9 1.6 1.5 24.2
GA [1] 0.598 0.226 0.336 1.957 1410 707
GA error (%) [1] 49.5 13.0 16.0 2.15 6.0 5.73 15.4
ICA 0.430 0.202 0.394 2.500 1200 700
ICA error (%) 7.50 0.86 1.41 24.99 20 6.67 10.2
GSA 0.425 0.203 0.415 2.399 1426 745.30
GSA error (%) 6.16 1.54 3.86 19.94 4.93 0.63 6.2

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

Table 2
Full load data for 4 kVA, 50 Hz, 250/125 V, one-phase transformer.

Transformer data I1 (A) I2 (A) V2 (V) Eciency (%) Average error (%)

Actual [1] 14.0813 13.6893 235.8759 83.9990

Computed with PSO [1] 14.0818 13.6972 237.2283 84.0927
PSO error (%) [1] 0.0036 0.0580 0.5733 0.1116 0.1866
Computed with GA [1] 14.1035 13.6654 234.2131 83.6697
GA error (%) [1] 0.1577 0.1746 0.7049 0.3920 0.3573
Computed with ICA 13.8729 13.6196 235.8917 83.9312
ICA error (%) 1.4799 0.5091 0.0067 0.0807 0.5191
Computed with GSA 13.8077 13.5916 235.407 83.8756
GSA error (%) 1.9431 0.7135 0.1988 0.1469 0.7506

6.5 meters is achieved using GSA as shown in Table 5. Hence, GSA also
performs the best in estimating parameter values for 15 kVA, 50 Hz,
Voltage regulation (%)

GSA 2400/240 V, one-phase transformer compared to GA, PSO and ICA.

5.5 Referring to Fig. 6, the estimated voltage regulation for 15 kVA
5 transformer using GSA and ICA also diers only slightly from the
actual voltage regulation.
4.5 From the optimisation convergence curve in Fig. 7, it can be
4 observed that GSA converges slightly faster than ICA. The tness value
using ICA reached a constant value after 9 iterations while the tness
3.5 value using GSA researched a constant value after 3 iterations only.
3 Moreover, it is noted that GSA yields smaller average error for the
50 60 70 80 90 100 estimated transformer's parameters than ICA, PSO and GA for three
Loading (%) dierent ratings of single- transformer. Hence, from the results
Fig. 4. Voltage Regulation at dierent loading levels for the 4 kVA transformer. obtained in this work, in overall, it can be concluded that GSA can
give better performance than ICA, PSO and GA in estimating three
transformer. dierent rating of single-phase transformers equivalent circuit para-
Using transformer parameter values from Table 3, the full load data meters.
for 10 kVA, 50 Hz, 500/125 V, one-phase transformer for the actual
data, ICA and GSA methods from this work and PSO and GA from [1] 3.4. Comparison on results between three dierent single-phase
were obtained as shown in Table 4. Similar to 4 kVA transformer, the transformer ratings
average error for transformer data at full load is the smallest for PSO,
followed by GA, ICA and nally GSA but the lowest average error of From the results of three dierent single-phase transformer ratings,
estimated transformer parameters is obtained using GSA. Hence, GSA it can be seen that GSA gives the lowest average error for the estimated
is the most suitable method to estimate parameter values for 10 kVA, transformer parameters for all transformer ratings compared to ICA,
50 Hz, 500/125 V, one-phase transformer over GA, PSO and ICA. PSO and GA. The ability of GSA to nd the global optimum solutions
Referring to Fig. 5, the estimated voltage regulation using GSA and ICA more accurately than other algorithm can explained by the attractive
diers only slightly from the actual voltage regulation. power of GSA, which is the way the masses communicate and move-
ment of all masses towards the heavier ones, which is a better
3.3. 15 kVA, 50 Hz, 2400/240 V, one-phase transformer candidate solution.
The error improvement of GSA and ICA over PSO and GA in terms
Table 5 shows the parameters of 15 kVA, 50 Hz, 2400/240 V, one- of estimated transformer parameters is summarized in Tables 7 and 8.
phase transformer at full load for the actual data, from ICA and GSA From the results obtained, it can be concluded that both ICA and GSA
methods in this work and PSO and GA in [1]. The parameters from the can give better performance than PSO and GA in estimating the
actual measurement at 50 Hz from [1] are also shown in Table 5 in the parameters of equivalent circuit for single-phase transformers from
Actual row. GSA again yields the lowest percentage of average error, their nameplate data without any experiment test. However, GSA yields
followed by ICA, GA and PSO. Referring to Table 6, similar to 4 kVA larger error improvement over PSO and GA compared to ICA.
and 10 kVA transformers, the average error for transformer data at full Several tests have been performed on ICA and GSA using dierent
load is the smallest for PSO, followed by GA, GSA and nally ICA. initial conditions. The initial conditions are the initial estimated values
However, the lowest average error of estimated transformer para- for R1, X1, R2, X2, RC, Xm. It was found that the nal results of the

Table 3
Parameter values for 10 kVA, 50 Hz, 500/125 V, one-phase transformer.

Parameter R1 () X1 () R2 () X2 () Rc () Xm () Average error (%)

Actual [1] 0.9 0.94 1.6 0.44 700 250

PSO [1] 0.811 0.8608 1.678 0.7540 713 314.2
PSO error (%) [1] 9.89 8.4 4.875 1.9 13.34 25.66 10.6775
GA [1] 1.025 0.8 1.507 0.493 651.5 204.4
GA error (%) [1] 13.9 14.9 5.8 12.05 6.92 18.26 11.9717
ICA 0.8 0.8 1.5 0.4259 692.48 255
ICA error (%) 11.11 14.89 6.25 3.21 1.07 2.00 6.4233
GSA 0.8001 0.8119 1.5004 0.4236 695.54 251.35
GSA error (%) 11.10 13.63 6.22 3.74 0.64 0.54 5.9775

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

Table 4
Full load data for 10 kVA, 50 Hz, 500/125 V, one-phase transformer.

Transformer data I1 (A) I2 (A) V2 (V) Eciency (%) Average error (%)

Actual [1] 19.688 19.299 491.729 95.241

Computed with PSO [1] 19.7012 19.3189 487.9916 95.2314
PSO error (%) [1] 0.0670 0.0993 0.7601 0.0108 0.2343
Computed with GA [1] 19.6072 19.1979 487.847 94.6334
GA error (%) [1] 0.4104 0.5274 0.7895 0.6386 0.5915
Computed with ICA 19.0427 18.2217 455.5431 94.5961
ICA error (%) 3.2774 5.582 7.3589 0.6771 4.2239
Computed with GSA 19.0417 18.2207 455.5186 94.5956
GSA error (%) 3.2827 5.5871 7.3639 0.6777 4.2278

2.5 Table 6
Full load data for 15 kVA, 50 Hz, 2400/240 V, one-phase transformer.
Voltage regulation (%)

GSA Transformer I1 (A) I2 (A) V2 (V) Eciency (%) Average

2 data error (%)

Actual [1] 6.2 6.2 2383.8 98.5

Computed with 6.2004 6.2008 2384.7 98.52
1.5 PSO [1]
PSO error (%) 0.0056 0.0128 0.0385 0.0202 0.01928
Computed with 6.2017 6.2001 2381.8 98.527
GA [1]
1 GA error (%) 0.0278 0.0011 0.0845 0.0280 0.0354
50 60 70 80 90 100
Loading (%) Computed with 6.1653 6.1387 2375.662 98.4801
Fig. 5. Voltage regulation at dierent loading levels for the 10 kVA transformer.
ICA error (%) 0.559 0.9893 0.3414 0.0202 0.4775
Computed with 6.1693 6.1393 2375.917 98.4804
estimated parameters do not depend on the initial conditions. The GSA
initial conditions only aect the number of iteration taken for the GSA error (%) 0.4949 0.9787 0.3307 0.0198 0.4560
algorithms to converge. It was found that GSA converges faster than
ICA under most of the dierent initial conditions that have been tested
in this work. Actual
Voltage regulation (%)

0.6 GSA
4. Conclusions
In this work, estimation of parameters from the nameplate data for
single-phase transformers of three dierent ratings has been success-
fully proposed using Imperialist Competitive Algorithm (ICA) and 0.4
Gravitational Search Algorithm (GSA). From results comparison
between ICA and GSA with PSO and GA, it was found that GSA and
ICA yield lower average error of estimated transformer parameters 50 60 70 80 90 100
compared to the actual data than to PSO and GA. GSA also converges Loading (%)
faster than ICA and shows higher error improvement over PSO and GA
Fig. 6. Voltage regulation at dierent loading levels for the 15 kVA transformer.
compared to ICA. Therefore, the proposed technique using GSA gives
the best accuracy results in estimating the parameters of transformer
application of ICA and GSA in this work is only for 6 parameters to be
equivalent circuit. The proposed technique could also be applied to
optimised. Works on using more than 6 parameters to be optimised can
determine the parameters of three-phase power transformers from its
be a future work. Also, since the results from the past work that was
nameplate data in the absence of actual experiment tests by repeating
used for comparison with this work reported parameters from the
the same procedures on each phase of the transformer.
actual measurement at 50 Hz only, using broader frequency range
In the work, up to 6 parameters were optimised. Hence, the limit of

Table 5
Parameter values for 15 kVA, 50 Hz, 2400/240 V, one-phase transformer.

Parameter R1 () X1 () R2 () X2 () Rc () Xm () Average error (%)

Actual [1] 2.45 3.14 2 2.2294 105,000 9106

PSO [1] 2.25 4.082 2.2 1.8526 99,517 9009
PSO error (%) [1] 8.16 30 10 16.9 5.22 1.07 11.8917
GA [1] 2.76 3.414 1.68 1.846 97,001 8951
GA error (%) [1] 12.6 8.72 16 17.2 7.6 1.7 10.6367
ICA 2 3 1.8 2 120,000 9200
ICA error (%) 18.37 4.46 10.00 10.29 14.29 1.03 9.7390
GSA 2.00 3.11 1.81 2.26 104,281 9094.87
GSA error (%) 18.35 0.80 9.68 1.36 0.68 0.12 5.1653

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

[6] A. Shintemirov, W. Tang, Q. Wu, Transformer core parameter identication using

0.072 frequency response analysis, IEEE Trans. Magn. 46 (2010) 141149.

Fitness value
0.5 0.07
[7] S. Soliman, R. Alammari, M. Mostafa, On-line estimation of transformer model
ICA parameters, Conf. Large Eng. Syst. (2004) 170178.
GSA [8] S.-D. Cho, Three-phase transformer model and parameter estimation for ATP, J.
0.4 0.066 Electr. Eng. Technol. 1 (2006) 302307.
[9] S.D. Mitchell, J.S. Welsh, Modeling power transformers to support the interpre-
2 3 4 5 6 7
tation of frequency-response analysis, IEEE Trans. Power Deliv. 26 (2011)
Fitness value

0.055 27052717.
0.3 0.054 [10] J. Martinez, B. Mork, Transformer modeling for low-and mid-frequency transients-

Fitness value
0.053 a review, IEEE Trans. Power Deliv. 20 (2005) 16251632.
0.052 [11] E.E. Mombello, K. Moller, New power transformer model for the calculation of
0.2 0.051 electromagnetic resonant transient phenomena including frequency-dependent
0.05 losses, IEEE Trans. Power Deliv. 15 (2000) 167174.
[12] S. Jazebi, F. d. Le, Duality-based transformer model including eddy current eects
0.1 2 3 4 5 6 7 in the windings (and x00F)IEEE Trans. Power Deliv. 30 (2015) 23122320.
[13] N. Chiesa, B.A. Mork, Transformer model for inrush current calculations: simula-
tions, measurements and sensitivity analysis (H. H. K, x00F, and idalen)IEEE
0 Trans. Power Deliv. 25 (2010) 25992608.
0 10 20 30 40 [14] G.L. Alegi, W.Z. Black, Real-time thermal model for an oil-immersed, forced-air
Iteration cooled transformer, IEEE Trans. Power Deliv. 5 (1990) 991999.
[15] S. Subramanian, S. Padma, Bacterial foraging algorithm based parameter estima-
Fig. 7. Optimisation convergence curve for the 15 kVA transformer. tion of three winding transformer, Energy Power Eng. 3 (2011) 135.
[16] M. Valii, M. Bigdeli, F. Hojjatiparast, Transient model parameters identication of
transformer based on PSO algorithm, in: Proceedings of the 3rd International
Table 7
Conference on Electric Power and Energy Conversion Systems (EPECS), pp. 15,
Percentage error improvement of GSA and ICA over PSO for transformer parameter
[17] M. Bigdeli, E. Rahimpour, Estimation of simplied transient model parameters
using genetic algorithm, 18th Iran. Conf. Electr. Eng. (ICEE) (2010) 722726.
Error 4 kVA 10 kVA 15 kVA [18] S. Bagheri, R. Eatnejad, A. Salami, Transformer winding Parameter identication
improvement transformer transformer transformer based on frequency response analysis using hybrid wavelet transform (WT) and
over PSO simulated annealing algorithm (SA) and compare with genetic algorithm (GA),
Indian J. Sci. Technol. 7 (2014) 614621.
GSA (%) 18.0 4.7 6.7 [19] S. Nagalakshmi, N. Kamaraj, Comparison of computational intelligence algorithms
ICA (%) 14.0 4.3 2.2 for loadability enhancement of restructured power system with FACTS devices,
Swarm Evolut. Comput. 5 (2012) 1727.
[20] D. Guha, P.K. Roy, S. Banerjee, Load frequency control of interconnected power
system using grey wolf optimization, Swarm Evolut. Comput. 27 (2016) 97115.
Table 8 [21] Y. Arya, N. Kumar, BFOA-scaled fractional order fuzzy PID controller applied to
Percentage error improvement of GSA and ICA over GA for transformer parameter AGC of multi-area multi-source electric power generating systems, Swarm Evol.
estimation. Comput.
[22] M. Abido, Optimal power ow using particle swarm optimization, Int. J. Electr.
Error 4 kVA 10 kVA 15 kVA Power Energy Syst. 24 (2002) 563571.
improvement transformer transformer transformer [23] H.A. Illias, X.R. Chai, H. Mokhlis, Transformer incipient fault prediction using
over GA combined articial neural network and various particle swarm optimisation
techniques, PloS One 10 (2015) e0129363.
GSA (%) 9.2 6.0 5.5 [24] M.R. AlRashidi, M.E. El-Hawary, A survey of particle swarm optimization
ICA (%) 5.2 5.5 0.9 applications in electric power systems, IEEE Trans. Evolut. Comput. 13 (2009)
[25] R. Hassan, B. Cohanim, O.D. Weck, G. Venter, A comparison of particle swarm
optimization and the genetic algorithm, Proc. 1st AIAA Multidiscip. Des. Optim.
cannot be veried in this work. Hence, at this stage, the limit of Spec. Conf. (2005) 1821.
application of ICA and GSA in this work is only at 50 Hz. Works on [26] M. Morad, M. Abedini, A combination of genetic algorithm and particle swarm
optimization for optimal DG location and sizing in distribution systems, Int. J.
using broader frequency range can be a future work.
Electr. Power Energy Syst. 34 (2012) 6674.
This work has focused on optimisation methods. Hence, more work [27] E. Pourjafari, H. Mojallali, Predictive control for voltage collapse avoidance using a
in the future can be done to determine if the applicability of the modied discrete multi-valued PSO algorithm, ISA Trans. 50 (2011) 195200.
proposed methods are valid for more complicated problems, such as [28] I. Poultangari, R. Shahnazi, M. Sheikhan, RBF neural network based PI pitch
controller for a class of 5-MW wind turbines using particle swarm optimization
the applications of ICA and GSA in cable fault location and distributed algorithm, ISA Trans. 51 (2012) 641648.
generation sizing and placement in large bus systems. [29] I. Pan, S. Das, Fractional order fuzzy control of hybrid power system with
renewable generation using chaotic PSO, ISA Trans. 62 (2016) 1929.
[30] D. Meister, M.A.G.D. Oliveira, The use of the least squares method to estimate the
Acknowledgement model parameters of a transformer, Int. Conf. Electr. Power Qual. Util. (2009) 16.
[31] S.H. Thilagar, G.S. Rao, Parameter estimation of three-winding transformers using
The authors thank the University of Malaya and Malaysia Ministry genetic algorithm, Eng. Appl. Artif. Intell. 15 (2002) 429437.
[32] H. Towsyfyan, S.A. Adnani-Salehi, M. Ghayyem, F. Mosaedi, The comparison of
of Higher Education (MOHE) for supporting this work through MOHE imperialist competitive algorithm applied and genetic algorithm for machining
HIR grant (H-16001-D00048) and PPP grant (PG028-2016A). allocation of Clutch assembly, International J. Eng. 26 (2013) 1485.
[33] Y. kumar, G. Sahoo, A review on gravitational search algorithm and its applications
to data clustering & classication, Int. J. Intell. Syst. Appl. 6 (2014) 7993.
References [34] D.E. Goldberg, Genetic Algorithm in Search, Optimization and Machine Learning,
Addison 1, Wesley Publishing Company, Reading, MA, 1989.
[1] M. Mossad, M. Azab, A. Abu-Siada, Transformer parameters estimation from [35] P. Grabusts, Evolutionary alorithms at choice: from GA to GP, in: Proceedings of
nameplate data using evolutionary programming techniques, IEEE Trans. Power the International Scientic and Practical Conference Environment Technology
Deliv. 29 (2014) 21182123. Resources (2015) 185192.
[2] J. Ramrez-Nio, A. Pascacio, Acoustic measuring of partial discharge in power [36] E. Atashpaz-Gargari, C. Lucas, Imperialist competitive algorithm: an algorithm for
transformers, Meas. Sci. Technol. 20 (2009) 115108. optimization inspired by imperialistic competition, IEEE Congr. Evolut. Comput.
[3] F. Marek, F. Jakub, Application of the transfer function to recognition of resonance (2007) 46614667.
overvoltages in transformer windings, Meas. Sci. Technol. 21 (2010) 125702. [37] A. Kaveh, Imperialist Competitive AlgorithmAdvances in Metaheuristic Algorithms
[4] H.R.-R. Joaqun, M.-P. Fernando, A..-C. Jos, J.-V. David, G.R.-P. Edgar, Direct for Optimal Design of Structures, Springer, 2014, pp. 349368.
optical techniques for the measurement of water content in oilpaper insulation in [38] L.D.S. Coelho, L.D. Afonso, P. Alotto, A modied imperialist competitive algorithm
power transformers, Meas. Sci. Technol. 22 (2011) 065706. for optimization in electromagnetics, IEEE Trans. Magn. 48 (2012) 579582.
[5] J.M.E. Vicente, A.J.J. Rezek, A.T.L. de Almeida, C.A.M. Guimares, Automation of [39] R.M. Amini, A. Safari, S.N. Ravadanegh, Optimal model of PDIG based microgrid
the temperature elevation test in transformers with insulating oil, ISA Trans. 47 and design of complementary stabilizer using ICA, ISA Trans. (2016).
(2008) 2531. [40] H. Shabani, B. Vahidi, M. Ebrahimpour, A robust PID controller based on

H.A. Illias et al. Swarm and Evolutionary Computation 36 (2017) 1826

imperialist competitive algorithm for load-frequency control of power systems, ISA algorithm, Inf. Sci. 179 (2009) 22322248.
Trans. 52 (2013) 8895. [44] C. Li, J. Zhou, Semi-supervised weighted kernel clustering based on gravitational
[41] S. Talatahari, B.F. Azar, R. Sheikholeslami, A. Gandomi, Imperialist competitive search for fault diagnosis, ISA Trans. 53 (2014) 15341543.
algorithm combined with chaos for global optimization, Commun. Nonlinear Sci. [45] E. Rashedi, H. Nezamabadi-Pour, B.S. Saryazdi, BGSA: binary gravitational search
Numer. Simul. 17 (2012) 13121319. algorithm, Nat. Comput. 9 (2010) 727745.
[42] B. Xing, W.-J. Gao, Imperialist Competitive AlgorithmInnovative Computational [46] S. Mirjalili, S.Z.M. Hashim, H.M. Sardroudi, Training feedforward neural networks
Intelligence: A Rough Guide to 134 Clever Algorithms, Springer, 2014, pp. using hybrid particle swarm optimization and gravitational search algorithm, Appl.
203209. Math. Comput. 218 (2012) 1112511137.
[43] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, GSA: a gravitational search