Electrical Power and Energy Systems 61 (2014) 208–218

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Modified parameter optimization of distribution transformer design

using covariance matrix adaptation evolution strategy
S. Tamilselvi ⇑, S. Baskar
Electrical and Electronics Engineering Department, Thiagarajar College of Engineering, Madurai 625 015, Tamil Nadu, India

a r t i c l e i n f o a b s t r a c t

Article history: Optimal transformer design (TD) is a complex multi-modal, multi-objective, mixed-variable and non-lin-
Received 16 April 2013 ear problem. This paper discusses the application of Covariance Matrix Adaptation Evolution Strategy
Received in revised form 18 March 2014 (CMA-ES) for distribution TD, minimizing four objectives; purchase cost, total life-time cost, total mass
Accepted 21 March 2014
and total loss individually. Two independent variables; voltage per turn and type of magnetic material
are proposed to append with the usual TD variables, aiming at cost effective, reduced weight, and energy
efficient TD. Three case studies with three sets of TD vectors are implemented on 400 KVA, 20/0.4 KV
Keywords:
transformer to demonstrate the superiority of Modified Design Variables (MDV), in terms of cost savings,
Distribution transformer design
CMA-ES
material savings, and loss reduction. Simulation results of CMA-ES provide better TD on comparison with
Magnetic material conventional transformer design procedure, branch and bound algorithm tailored to a mixed-integer
Purchase cost non-linear programming, Self Adaptive Differential Evolution (SaDE), and real coded GA (RGA). Statistical
Total life time cost analysis has proven the faster convergence and consistency of CMA-ES. Moreover, NSGA-II is applied for
Voltage per turn solving multi-objective TD optimization problem with the aim of providing tradeoff between conflicting
TD objectives.
Ó 2014 Published by Elsevier Ltd.

Introduction generation needed to accommodate the losses, operating cost,

All power utilities are much worried about the high failure rate
of distribution transformers. Transformers can be expected to
operate 20–30 years. But losses in the form of heat reduce the trans-
former life by causing damage to the insulation. Finding ways to
decrease the losses of the transformer is an important factor
towards reducing transformer failure rate, costs and CO2 emissions.
Total life time cost (TLTC) is the total life cycle cost which considers
the future operating costs of a unit over its lifetime, brought back
into present day cost and then added to its total purchase price. A
transformer with low TLTC is expected to have low losses and
longer life. Selecting energy efficient distribution transformers
‘SEEDT’ project also concluded that electricity distribution compa-
nies and commercial and industrial users should use the TLTC
method for making transformer purchasing decisions. Energy
efficient transformers are more expensive, but use less energy,
resulting in lesser TLTC [1]. So designing a transformer based on
purchase cost, without considering the future losses is therefore
not a right decision. TD based on minimizing TLTC alone can drop
off the transformer failure rate by reducing losses, amount of power
⇑ Corresponding author. Tel.: +91 9003225642.
E-mail addresses: tseee@tce.edu (S. Tamilselvi), sbeee@tce.edu (S. Baskar).
emission of green house gases and environmental cost, among the
other TD objectives. But this TLTC minimization objective is rarely
attended in most of the literatures.
From the overview of research papers in TD, efforts are focusing
on prediction of specific transformer characteristics, techniques
adopted for transformer design optimization, transformer post
design performance and modeling and recent trends on trans-
former technology. In a nutshell, TD optimization problem remains
an active area [1]. TD optimization can be minimization of no-load
loss [2,3], minimization of load loss [4], maximization of efficiency
[5–8], maximization of rated power [9], minimization of mass [9] or
minimization of cost [5,6,10–19], based on the objective functions.
Comparison of non-linear programming techniques for opti-
mum design of transformers was presented in [20]. Optimization
done by geometric programming for the minimization of total
mass of the transformer, [9] provided solution for low and high fre-
quency transformers. But this method finds difficulty in combina-
tion with cost estimation algorithm and requires mathematical
model.
Mixed integer programming (MIP) in combination with finite
element method (FEM) [10], MIP in combination with branch
bound algorithm (BBA) [11], bacterial foraging algorithm (BFA)
[5] and simulated annealing technique [12] have been adopted

http://dx.doi.org/10.1016/j.ijepes.2014.03.039
0142-0615/Ó 2014 Published by Elsevier Ltd.
 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218 209

Nomenclature

TD transformer design GLV weight of LV conductor in kg

CMA-ES Covariance Matrix Adaptation Evolution Strategy GHV weight of HV conductor in kg
TLTC total life time cost GSS weight of sheet steel in kg
MIP mixed integer programming GOIL weight of mineral oil in kg
FEM finite element method GINS weight of insulating paper in kg
BBA branch bound algorithm GDS weight of duct strips in kg
BEM boundary element method GCORR weight of corrugated panel in kg
BBA-MINLP BBA-mixed integer non-linear programming CTM total main materials cost of transformer in Euro
BFA bacterial foraging algorithm CTLT total life time cost of transformer in Euro
MC manufacturing cost Twt total mass of transformer in kg
MM magnetic material Tlos total loss of transformer in Watts
TMM type of magnetic material Crem cost of transformer remaining materials in Euro
GA genetic algorithm Clab labor cost in Euro
CTDP conventional transformer design procedure AA no load loss cost rate in Euro per Watts
LV low voltage PNLL designed no-load loss in Watts
HV high voltage BB load loss cost rate in Euro per Watts
MDV Modified Design Variables PLL designed load loss in Watts
SD Standard Deviation Sm sales margin in percentage
CT computation time dvi=1,..3 design vector
k
NA not applicable ov output variables
LMT length of mean turn Uk designed short circuit impedance in percentage
EAs evolutionary algorithms Tcl thickness of core leg in mm
SaDE Self Adaptive Differential Evolution Hdiss total heat dissipated by convection and radiation
RGA Real coded Genetic Algorithm through cooling arrangement.
MOP multi-objective optimization problem C(g) covariance matrix ‘C’ at generation g
MOEAs Multi-Objective EAs N normal search distribution
POF Pareto Optimal Front r(g) overall standard deviation; step size at generation g
NSGA-II non-dominated sorting genetic algorithm-II m(g) mean value ‘m’ of the search distribution at generation g
CMA Covariance Matrix Adaptation Rn linear transformation
ES Evolution Strategy n search space dimension, i.e., number of decision vari-
SymbolsPGNLL ables
guaranteed no-load loss in Watts g iteration number
ðgþ1Þ
PGLL guaranteed load loss in Watts Xk kth offspring from generation g + 1; search point
Uks guaranteed short circuit impedance in percentage k number of sampling points
e1 IEC standard tolerance for no-load and load loss con- l number of selected best individuals in the population
ðgÞ
straints in % pc evolution path at generation g
ðgÞ
e2 IEC standard tolerance for total loss and impedance pr conjugate evolution path at generation g
voltage constraints in % FF fitness function
CFE,i=1,..10 unit cost of ith magnetic material in Euro per kg gmax maximum number of generations
CCOND unit cost of conductor in Euro per kg zmax objective function value of the worst feasible solution
CSS unit cost of sheet steel in Euro per kg z(x) objective function value of the feasible solution
COIL unit cost of mineral oil in Euro per kg errj(x) normalized constraint violation
CINS unit cost of insulating paper in Euro per kg Feval Maximum number of function evaluations
CDS unit cost of duct strips in Euro per kg H Hessian matrix
CCORR unit cost of corrugated panel in Euro per kg Pt parent population of MOP
xt=1,2.n tth design variable Qt offspring population of MOP
zp=1,..,4 pth objective function Rt combined population of MOP
hj=1,2..q jth inequality constraint fi=1,..,M ‘M’ objective function values of MOP
GFE,i=1,..10 weight of ith magnetic material in kg Np population size in MOP problem

for the minimization of main material cost (CTM) of transformer,
without considering the transformer losses. However, all these
methods have got their own drawbacks. MIP-FEM is sensitive to
the selection of value range of design variables and fails to find
the global optimum. MIP-BBA is time consuming, since the number
of nodes in a branching tree is too large. The drawbacks of simu-
lated annealing technique are finding difficulty in extending itself
to the multi-objective case and long searching time to find the
optimum. When the search space and complexity grow exponen-
tially in scalable problems, basic BFA would not be suitable.
Hybrid FEM with boundary element method (BEM) [21], trial
and error based heuristic approach [13] have been implemented
for the minimization of transformer manufacturing cost (MC),
which is mere sum of CTM and labor cost, excluding the operational
cost.
BBA tailored to a mixed integer non-linear programming (MIN-
LP) [15], numerical field analysis technique in combination with
BEM [19] and multiple algorithm based hybrid approach [23] have
addressed the minimization of TLTC of transformer. These papers
have overcome the above said weaknesses, by including losses in
the objective function calculation. But the numerical field analysis
technique has the disadvantage of complex mesh size in 3D config-
urations [22]. Besides, minimization of only one TD objective is
considered [19,23] resulting in a single solution, which may not
210 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218
suit the requirements of all the categories of decision makers. Type
of magnetic material (TMM) is also not utilized as a design variable
in the optimization process [15,19,23].
A novel gamma approach [24] and BFA based optimal design [6]
have been proposed for solving the multi-objective TD optimiza-
tion problem. But the authors failed to consider a three phase
transformer and have not taken TLTC as one of the TD objectives.
Generally, the objective functions of the TD problem are non-
differentiable, non-convex, mixed-variable, non-linear, and
multi-modal and it is very difficult to obtain an optimal solution.
Furthermore, TD calculations require access to several look-up
tables’ data for the evaluation of specific core loss at various flux
densities, winding gradient, oil gradient, and heat transmission.
Such complex analytical calculations interacting with graphical
data are not handled accurately by the derivative based methods
discussed above and thus the optimal solution is not guaranteed
for the TD optimization problem solved by the analytical methods.
Apart from the deterministic methods, soft computing tech-
niques such as genetic algorithm (GA) and neural network are also
employed for the TD optimization. GA has been applied for the
minimization of MC, incorporating TMM as design variable [14],
transformer cost plus running cost minimization [16,17], and opti-
mal placement of distribution transformers [18]. Neural network
technique has been applied for the prediction and minimization
of power losses [22], and minimization of no load loss [2,3]. But
these papers used single optimization method for TD which lim-
ited the optimization performance of their approach. A more
recent approach to adapting the search direction is the Covariance
Matrix Adaptation Evolution Strategy (CMA-ES) proposed by Han-
sen and Ostermeier [25]. Its important property is invariance
against the linear transformations in the continuous search space,
when compared to other evolutionary algorithms (EAs).
CMA-ES is a continuous (real-parameter) EA that generates new
population members by sampling from a probability distribution
that is constructed during the optimization process. Owing to the
learning process, CMA-ES is invariant to rotation and scaling of
the coordinate system, reliably adapts well to ellipsoidal functions,
and significantly improves convergence rate especially on non-sep-
arable and/or badly scaled objective functions. CMA-ES thus finds a
global or near optimal minimum without using derivatives of the
objective functions. CMA-ES is a strong optimizer that outper-
formed its other similar learning algorithms in CEC2005 bench-
mark functions [26,27] and BBOB-2009 [28]. Hence in this paper
CMA-ES is applied for solving this complex TD optimization
problem.
The main contributions of the paper are: (a) application of CMA-
ES for TD optimization for the first time, assuring accuracy, consis-
tency and convergence; (b) incorporation of TMM as one of the
design variables for representing 10 different materials; (c) inclu-
sion of variable, voltage per turn in place of low voltage (LV) turns;
(d) implementation of three case studies to show the superiority of
MDV; (e) optimization of four different objectives such as minimi-
zation of purchase cost, minimization of TLTC, minimization of
total mass, and minimization of total loss, individually suggesting
the designer a set of optimal transformers instead of single solu-
tion, so that he can choose which of them best fits the requirement
of the customer and application under consideration; (f) compari-
son of simulation results with recent report [15], conventional
transformer design procedure (CTDP) [29], Self Adaptive Differen-
tial Evolution (SaDE), and Real coded Genetic Algorithm (RGA);
(g) use of multi-objective TD optimization using NSGA-II.
This paper is organized as follows: Sections Design of distribu-
tion transformer, CMA-ES algorithm an overview, CMA-ES based
TD optimization, Multi-objective TD optimization, Computational
results and Conclusion the paper.
Design of distribution transformer
Preliminary input for TD
The design procedure is presented for three phase oil immersed
shell type wound core distribution transformer. The specified
information consists of desired input variables [29] required for
the transformer design employing analytical formulae to calculate
the transformer parameters. These transformer variables include
transformer rating, design requirements on guaranteed no-load
loss (PGNLL), guaranteed load loss (PGLL), guaranteed short circuit
impedance (Uks), minimum full load efficiency, maximum temper-
ature rise, voltage regulation, tolerances e1, (e2), core stacking fac-
tor, mass density of core, magnetization curve of the magnetic
material (MM), specific core loss data [30] for various maximum
magnetic flux densities at 50 Hz frequency for 10 magnetic mate-
rials (M3-0.27, M4-0.27, MOH-0.23, MOH-0.27, 23ZDKH90,
23ZH90, 23ZH95, 27ZDKH95, 23ZDMH85, and 27ZDMH), resistiv-
ity of the conductor material (copper) at the maximum specified
temperature, type of internal and external winding, typical practi-
cal values for insulation of conductor, distance and insulation
between windings and core, mass density of conductor, distance
between two adjacent cores, maximum ambient and winding tem-
perature, direction space factor for turns and layer, HV taps, etc.
Copper sheet is used for LV conductor and copper wire is used
for HV conductor.
Mathematical model
Mathematically, multi-variable, non-linear, multi-objective,
mixed-variable and mixed programmed transformer design
optimization problem can be stated in the form as,
Find x = {x1,x2,. . ..,xn}, which
Minimize zðxÞ
Subject to hj ðxÞ 6 dj ; for j ¼ 1; 2; . . . ; q
ð1Þ
xt 6 0; for t ¼ 1; 2; . . . ; n
xt;min 6 xt 6 xt;max ; for t ¼ 1; 2; . . . ; n
where ‘z’, is the objective function to be minimized and hj, jth
inequality constraint are functions of ‘n’ decision variables ‘x’; dj,
are constants of constraints; and, xt,min and xt,max are lower and
upper limit on ‘x’, respectively.
(i) Minimum purchase cost design-Objective function z1: This
objective minimizes the total cost of transformer main materials
(CTM) including core, LV conductor, HV conductor, insulating paper,
duct strips, oil, corrugated panel, and sheet steel using [15,29]:
Min;z1 ¼ C TM
¼ fC FE;i  GFE;i þ C COND  GLV þ C COND  GHV þ C INS  GINS
þ C DS  GDS þ C OIL  GOIL þ C CORR  GCORR þ C SS  GSS g ð2Þ
where GFE, GCU, GINS, GDS, GOIL, GCORR, GSS and GSS are weight of
magnetic material, LV conductor, HV conductor, insulating paper,
duct strips, oil, corrugated panel, and sheet steel in kg respectively;
CFE,i=1,2. . .10, CCOND, CSS, COIL, CINS, CDS, and CCORR are the unit cost of
magnetic material, LV/HV conductor, sheet steel, oil, insulating
paper, duct strips, and corrugated panel respectively.
(ii) Minimum TLTC design-Objective function z2: The cost optimal
design of transformer has to minimize the sum of transformer cost
and running cost. In this objective, transformer’s selling price and
losses are considered as transformer price. So material saving
and energy saving are the two important aspects in minimization
of TLTC. This objective minimizes TLTC of transformer (CTLT), (using
[15,29]:
 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218
 
ðC TM þ C rem þ C lab Þ
Min;z2 ¼ C TLT ¼ þ AA  PNLL þ BB  P LL
ð1  Sm Þ
where, CTLT, AA, BB, Crem, Clab, Sm, PNLL and PLL are total life time cost
in euro, no-load loss cost rate in euro per watts, load loss cost rate in
euro per watts, remaining materials cost in euro, labor cost in euro,
sales margin, designed no-load loss in watts and designed load loss
in watts respectively.
(iii) Minimum mass design-Objective function z3: This objective
minimizes the total mass of transformer materials (Twt) (like core,
LV conductor, HV conductor, insulating paper, duct strips, oil,
corrugated panel, and sheet steel using [15,29]:
Min;z3 ¼ T wt
¼ fGFE þ GLV þ GHV þ GINS þ GDS þ GOIL þ GCORR þ GSS g
(iv) Minimum total loss design-Objective function z4: This objec-
tive minimizes the total loss of the transformer (Tlos) using [13]:
Min;z4 ¼ T los ¼ fPNLL þ PLL g
(v) Design vector dv: Design variables are collectively called as
design vector. Modified Design Variables (MDV) are described in
the following three sets of design vectors (dv) to realize the TD case
study, detailed in section ‘Computational results’.
dv1 ¼ fx1; x2; . . . :x6 g
dv2 ¼ fx2; x3; . . . :x7 g
dv3 ¼ fx2; x3; . . . :x8 g
where x1 (integer variable) is LV turns; x2 (integer variable) is width
of the core leg in mm; x3 (integer variable) is height of the core win-
dow in mm; x4 (discrete variable) is maximum magnetic flux den-
sity in Tesla; x5 (discrete variable) is current density in LV
winding in A/mm2; x6 (discrete variable) is current density in HV
winding in A/mm2; x7 (discrete variable) is voltage per turn in volts;
x8 (integer variable) is TMM.
(vi) Constraints of performance indices: Transformer performance
must meet certain standards and rules, specified in IEC 60076-1,
include,
 No-load loss constraint:
PNLL 6 P GNLL ð1 þ e1 Þ
 Load loss constraint:
PLL 6 PGLL ð1 þ e1 Þ
 Total loss constraint:
ðP NLL þ PLL Þ 6 ðPGNLL þ PGLL Þð1 þ e2 Þ
 Impedance voltage constraint:
U ks ð1  e2 Þ 6 U k 6 U ks ð1 þ e2 Þ
(vii) Constraints of manufacturing process indices: These con-
straints are based on the transformer manufacturer specifications.
 Ratio of width of the core leg to the height of the core window
should be less than or equal to unity;
x 2 6 x3
 2-Multiplied core leg thickness (Tcl) must be between a mini-
mum of half core leg width and a maximum of 90% core leg
width;
0:5x2 6 2T cl 6 0:9x2
211
(viii) Constraints of material performance:
ð3Þ
 Heat transfer constraint: Total heat produced by total loss of
transformer must be smaller than the total heat dissipated by
convection and radiation (Hdiss) through cooling arrangement;
P NLL þ PLL 6 Hdiss ð15Þ
 Magnetic flux density constraint;
x4;min 6 x4 6 x4;max ð16Þ
CMA-ES algorithm an overview
ð4Þ
Evolution strategies (ES) are stochastic, derivative free methods
for numerical optimization of non-linear problems. CMA-ES is an
efficient ES for problems for which derivative based methods are
unsuccessful, due to rugged search space with multiple discontinu-
ð5Þ ities, sharp bends, and local optima. This algorithm is analogous to
gradient based quasi-Newton method. CMA-ES has emerged as a
very competitive real-parameter optimizer for continuous search
spaces. It adapts two unique principles; maximum likelihood prin-
ciple and two evolution paths and thus distinct from other ES.
It is a continuous EA that generates new population members
ð6Þ by sampling from a multi-variate normal distribution N(m, C) con-
structed by its mean value, m e Rn and its symmetric positive def-
ð7Þ inite covariance matrix, C e Rnxn during the optimization process.
‘m’ of the distribution determines the translation displacement
ð8Þ and gets updated such that the likelihood of previous successful
candidate solutions is maximized. ‘C’ has geometrical interpreta-
tion, can be uniquely identified with the iso-density ellipsoid
[31]. ‘C’ determines the shape of the distribution ellipsoid, whose
principal axes are eigen vectors of ‘C’ and squared axes lengths
are eigen values [31,32]. This algorithm exploits two adaptation
mechanisms; Covariance Matrix Adaptation (CMA) and step size
(r) adaptation.
CMA learns all pairwise dependencies between the variables
and increases the probability to repeat the successful steps. One
evolution path, pc enhances CMA procedure in place of single suc-
cessful search steps and facilitates possibly much faster increase of
favorable directions. CMA identifies the function landscape which
is convex-quadratic one with the concept of Hessian matrix (H).
ð9Þ The CMA-ES estimates the inverse Hessian matrix (H1) in the form
of a covariance matrix of the search distribution within an iterative
procedure. Setting C = H1 on convex quadratic function is to
ð10Þ rescale the ellipsoid projection of the multi-variate normal distri-
bution into a spherical one. The aim of CMA is to approximate
the matrix, ‘H’ and to closely suit the search direction to the
ð11Þ contour lines of the objective function to be optimized [28,32].
The use of covariance matrix adaptation is clearly depicted in
Fig. 1, generated by CMA-ES [32], for the linear function,
ð12Þ P
flinear ¼  2i¼1 xi to be minimized, population size, k = 150, number
of selected best individuals in the population, l = 50, and weights,
wi=1,. . .,l = 1/l. This Fig. 1 illustrates that there is an increase in the
expected variance in gradient direction, for all l < k=2. Left image
of Fig. 1 shows the sample of k = 150 distributed points. Middle
image of Fig. 1 demonstrates the l selected vectors and the esti-
mated covariance matrix ellipsoid. Right image of Fig. 1 shows
the search distribution of the next generation. Thus for a particular
ð13Þ function landscape if it becomes qualified to convert it into a
spherical projection by correspondingly adapting the eigen-
decomposed matrix, then the algorithm converges fully to the
global optimum.
The CMA influences the scale of the distribution. Nevertheless,
ð14Þ additional step-size, r control is necessary [28]. So, the step size
update is also introduced to enhance the scaling adaptation and
212 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218
Fig. 1. Estimation of the covariance matrix on flinear ¼ 
in particular to facilitate the increase of distribution spread, which
is very difficult with CMA only. Step size adaptation aims to make
consecutive movements of the distribution mean orthogonal in
expectation and prevents the premature convergence. The method
used for step size control is path length control (cumulative step
size adaptation). To control the step size, other evolution path, pr
is utilized [31]. (a) Whenever pr is short, single steps cancel each
other and are anti-correlated. Hence, r should be decreased. (b)
Whenever the evolution path, pr is long, single steps are pointing
to similar directions and they are correlated causing r to be
increased. (c) Submitting in the desired situation the steps are
approximately perpendicular in expectation and therefore uncor-
related [31,28]. To decide if pr is long or short, the length of the
path is compared with its expected length under random selection.
In ideal situation, selection does not bias the length of the evolu-
tion path and the length equals its expected length under random
selection. Thus pr controls the step-size.
CMA-ES based TD optimization
CMA-ES minimizes objective functions z1 to z4, Eqs. (2)–(5)
individually, subject to the constraints, Eqs. (9)–(16). Set the TD
objective function, and design vector dv. Assume suitable values
for the transformer variables in section ‘Preliminary input for
TD’, lower and upper bounds of ‘n’ transformer design variables xt,-
min, xt,max and constants of the constraints in section ‘Mathematical
model’.
Modifications on CMA-ES algorithm
Our contributions on the optimization algorithm CMA-ES for
handling the TD optimization problem are constraint handling
segment and tuning of CMA-ES parameters namely, population
size (k), and maximum number of function evaluations (Feval).
 Penalty parameter less constraint handling
Penalty parameter less constraint handling is more suitable
than the other constraint handling methods and it is incorporated
for treating the complex constraints of this TD optimization prob-
lem [33]. This method is preferred in such a way that the different
magnitude violation of every constraint is handled properly and
the algorithm can get away from the local optima. In this handling
method the infeasible solutions are compared based on their
overall constraint violation, whereas the feasible solutions are
compared only based on their objective function values and the
overall normalized constraint violation is calculated using the
following steps.
P2
i¼1 xi .
(i) Transform all the constraints Eqs. (9)–(15) into inequality
form, hj ðxÞ 6 dj ; for j ¼ 1; 2; . . . ; q, Eq. (1).
(ii) Since the constraint violation, (hj(x)  dj) for each constraint
‘j’ is of different range, calculate the normalized constraint
violation using, errj(x) = (hj(x)/dj)1. Positive value of errj(x)
denotes the violation of constraint and negative value of
errj(x) shows no violation.
(iii) Calculate the overall normalized constraint violation as
Pq
j¼1 max ð0; err j ðxÞÞ and used in fitness function evalua-
tion using penalty parameter less constraint handling
scheme.
 Tuning of CMA-ES Parameters k and Feval
Initially, TD optimization problem is solved using the default
parameter setting of CMA-ES with population size (k) as
4 + floor(3logn) and maximum number of function evaluations
(Feval) as 3e3  ½ðn þ 5Þ2 þ k. But it is trapped in local optima.
Hence, a sensitivity analysis has been conducted to identify the
appropriate values for k and Feval. Twenty-five independent runs
are carried out for each objective with dv3 (n = 7), to check the con-
sistency of obtaining optimal solutions, with different settings of k,
and Feval, and statistical results are recorded. Algorithm did not
converge to global value in all the trials, until the maximum num-
ber of objective function evaluations, Feval was increased to 4000
and population size, k to 20. The aim of the analysis is not only
to obtain the best global value, but also minimum Standard Devi-
ation (SD). Only when k is high (80 or above) with the minimum
of 6000 function evaluations, minimum SD and lesser computa-
tional time are observed. k and Feval are finally fixed at 100 and
10,000 respectively for all the objectives. Increased population size
rather than default is encouraged for such very difficult problem, as
it improves the performance of CMA-ES on multi-modal functions
[34]. In this case, the stopping criterion is not based on Feval alone.
Stop flag is set, when any one of the following conditions get sat-
isfied. (i) if maximum number of iterations is exceeded (ii) if num-
ber of function evaluations reaches maximum value (iii) if ‘x’
values changes smaller (iv) if fitness value changes smaller and
(v) if fitness values are equal.
Single objective TD optimization by CMA-ES
Step 1: As CMA-ES is a quasi-parameter free algorithm, set
default values [32] for all CMA-ES parameters except for k, Feval,
gmax, and initial step size r(0). Choose r(0) = 0.25(xt,max  xt,min),
k = 100, Feval = 10,000 and gmax = 100.
Step 2: Initialize generation count, g = 0.
Step 3: If stopping criterion is met: i.e., if stop flag is set, go to
step 11. Else go to step 4.
 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218
Step 4: Generate k candidate solutions by sampling from a
multi-variate Gaussian distribution, N(m, C) with mean, covariance
matrix and standard deviation [32].
Step 5: Determine the fitness function, (FF(x)), using penalty
parameter-less constraint handling [33],
8 9
> zðxÞ; if feasible >
< =
FFðxÞ ¼ X
q
ð17Þ
max
:z þ
> max ð0; errj ðxÞÞ; otherwise >
;
j¼1
where, FF(x) is the fitness function; z(x) is the objective function
value; q = no. of constraints, excluding the boundary constraints
of design variables; errj(x) is the normalized constant violation for
each constraint; zmax is the objective function value of the worst fea-
sible solution in the current population. The fitness of a feasible
solution is always fixed and is equal to its objective function value
z(x). However the fitness of the infeasible solutions not only
depends on the overall constraint violation, but also on the fitness
of the worst feasible solution, zmax. If there is no feasible solution
in the current population zmax, is zero. The infeasible solutions are
compared based on their overall normalized constraint violation.
Step 6: The (k) sampled points are sorted in order of ascending
fitness and (l) best are selected. The new mean (m(g+1)) of all cur-
rent population vectors which is a weighted average of (l) selected
ðgþ1Þ
vectors from the samples X 1 ; . . . ; X k ðg þ 1Þ with weight param-
eter being updated using weighted recombination of the selected
points [32].
ðgþ1Þ ðgþ1Þ
Step 7: Update evolution paths pc and pr [32].
(g+1) ðgþ1Þ
Step 8: Update the covariance matrix C by pc [32].
ðgþ1Þ
Step 9: Update global step size r(g+1) using pr [32].
Step 10: Increment generation count, g = g + 1; Go to step 3.
Step 11: Stop the optimization process.
Multi-objective TD optimization
So far we have discussed the application of CMA-ES for distribu-
tion TD optimization, minimizing four objectives; purchase cost,
total life-time cost, total mass and total loss, individually that
resulted in four independent transformer designs. But the decision
maker may be interested in the tradeoffs between each of the con-
flicting criteria. For example, compromising a design that would
provide low weight transformer without imposing too much of
losses. Therefore it is necessary to look at this TD problem as a
multiple criterion decision making problem that is concerned with
more than one objective function to be optimized simultaneously.
Classical optimization methods have limitation in handling
MOP as they convert the MOP into a single objective optimization
problem to generate only one Pareto optimal solution using
weighted-sum of objectives. To generate multiple Pareto solutions,
the classical optimization model has to be used several times.
Nowadays several Multi-Objective EAs (MOEAs) have gained pop-
ularity with the advantage of generating the whole Pareto-optimal
solutions in a single simulation. The main principles of MOEA are
to push the entire population toward Pareto Optimal Front (POF)
and to maintain diversity among the solutions in the POF. In this
paper, we have applied an elitist non-dominated sorting genetic
algorithm, NSGA-II, one of the state-of-the art multi-objective opti-
mization algorithms for solving the TD MOP, since Deb [35] sug-
gested that NSGA-II outperforms two other contemporary
MOEAs: Pareto-archived ES (PAES), Strength Pareto EA (SPEA), in
terms of finding a diverse set of solutions and in converging near
the true Pareto-optimal set.
Multi-objective problem description
A general MOP can be stated as
213
Minimize FðxÞ ¼ ffi : 8i ¼ 1; . . . ; Mg
x ¼ fxt : 8t ¼ 1; . . . ; ng
Subject to hj ðxÞ 6 dj ; for j ¼ 1; 2; . . . ; q ð18Þ
xt P 0; for t ¼ 1; 2; . . . ; n
xt;min 6 xt 6 xt;max ; for t ¼ 1; 2; . . . ; n
where, fi=1,. . .,M is the objective function values of ‘M’ objectives.
NSGA-II
The steps involved in the working of NSGA-II are given below.
Step 1: Initially generate a random parent population, Pt of size
Np at generation ‘t’.
Step 2: Calculate all the objective function values separately for
the entire population. Sort the population based on non-domina-
tion levels.
Step 3: Assign a rank for each solution based on its non-domina-
tion level.
Step 4: Apply the binary tournament selection, Simulated Binary
Crossover (SBX) crossover, and polynomial mutation operators on
Pt and create an offspring population Qt of size Np.
Step 5: Combine parent and offspring populations to implement
elitism and form an intermediate population, Rt of size 2Np.
Step 6: Evaluate the fitness for 2Np population solutions using
multiple objective functions.
Step 7: Perform non-dominated sorting over 2Np population to
rank and divide the individuals into different POFs [35]. Assign
rank for the POFs in such a way that the first non-dominating
POF with rank one and so on.
Step 8: Calculate crowding distance of all the solutions [35].
Higher the value of crowding distance better is the probability of
the solution to be selected for the next generation.
Step 9: Select Np member for the new population Pt+1 from the
combined population 2Np, on the basis of ranking and crowding
distance. i.e., choose individuals of the best low ranked POFs first
followed by the next best POF and so on. In case there is space only
for a portion of a POF to accommodate in the new population,
choose individuals among those in the front that are from the least
crowded regions (determined by a crowded-distance operator
[35]) and fill up the required number in the new population, Pt+1.
Step 10: Form the parent population for the next generation Pt+1.
Step 11: Check if termination criteria is met. If yes, get the Par-
eto optimal solution. Else go to step 4.
Computational results
To demonstrate the effectiveness of the proposed MDV, a design
example of 400 KVA, 50 Hz, 20/0.4 kV, 3 phase, shell type, wound
core transformer with vector group, Dyn11 has been considered.
The upper and lower bounds of the design variables, AA, BB, unit
price of transformer materials Crem, Clab, Sm, e1 = 15%, e2 = 10% are
derived from [15]. As reference transformer, the transformer with
loss category AB0 according to CENELEC [36] is selected, which
means that PGNLL = 750 W, PGLL = 4600 W, and Uks = 4%. Coding for
TD optimization is developed using MATLAB 7.4 on Intel core, i3
processor Laptop, operating at 3.2 GHZ, with 3 GB RAM. Suitable
modifications are incorporated for handling TD constraints in the
coding of CMAES algorithm [37]. The population size and maxi-
mum number of function evaluations are fixed at 100 and 10,000
respectively.
Case study 1
This case study is done particularly to prove the efficiency of
CMA-ES over deterministic method, BBA-MINLP [15] and heuristic
214 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218

Table 1
Optimization Results of CMA-ES with dv1.

dv1/ov z1 z2
CTDP [33] BBA–MINLP [19] CMA-ES CTDP [33] BBA-MINLP [19] CMA-ES
x1 17 18 18 17 20 17
x2 (mm) 220 239 212 220 231 242
x3 (mm) 245 248 241 245 299 242
x4 (Tesla) 1.7 1.8 1.8 1.7 1.6 1.6
x5 (A/mm2) 3.0 3.0 3.6 3.0 3.0 3.0
x6 (A/mm2) 3.0 3.0 3.6 3.0 3.0 3.0
CTM (Euro) 4475 4203 3863 4475 4428 4617
PLL (Watts) 4148 4288 5012 4148 4613 4265
PNLL (Watts) 818 859 856 818 719 685
CTLT (Euro) 27199 27500 28790 27199 27467 26588

Table 2
Optimization Results of CMA-ES.

dv/ov z1 z2 z3 z4
With dv1 With dv2 With dv3 With dv1 With dv2 With dv3 With dv1 With dv2 With dv3 With dv1 With dv2 With dv3
k k k k k k k k
x1 18 NA NA 17 NA NA 19 NA NA 16 NA NA
x2 (mm) 212 224 224 242 245 245 248 244 232 232 220 219
x3 (mm) 241 227 227 242 245 245 270 275 285 232 220 219
x4 (Tesla) 1.8 1.8 1.8 1.6 1.6 1.7 1.8 1.8 1.8 1.6 1.6 1.75
x5 (A/mm2) 3.6 3.6 3.6 3.0 3.0 3.0 3.6 3.6 3.0 3.0 3.0 3.0
x6 (A/mm2) 3.6 3.6 3.6 3.0 3.0 3.0 3.0 3.0 3.45 3.0 3.0 3.0
k k k k
x7 (Volts) NA 13.22 13.22 NA 13.20 13.20 NA 11.85 11.27 NA 14.90 14.91
k k k k k k k k
x8 NA NA 4 NA NA 5 NA NA 10 NA NA 9
CTM (Euro) 3863 3817 3817 4617 4543 4629 3993 3970 4820 4754 4837 5680
PLL (Watts) 5012 4886 4886 4265 4297 4215 4771 4781 4981 4092 3958 3827
PNLL (Watts) 856 854 854 685 662 613 818 804 563 737 765 664
CTLT (Euro) 28790 28394 28394 26588 26371 25882 28060 27936 27648 26786 26794 26844
Twt (kg) 1271 1260 1260 1460 1428 1360 1238 1229 1216 1533 1587 1470
Tlos (Watts) 5868 5740 5740 4950 4959 4828 5589 5585 5544 4829 4723 4491
k
NA – Not applicable.

Table 3
Comparison of TD Optimization results of CMA-ES, SaDE, RGA for z1.

dv/ov dv1 dv2 dv3

CMA-ES SaDE RGA CMA-ES SaDE RGA CMA-ES SaDE RGA
k k k k k k
x1 18 18 18 NA NA NA NA NA NA
x2 (mm) 212 212 212 224 223 226 224 227 224
x3 (mm) 241 241 241 227 228 236 227 237 227
x4 (Tesla) 1.8 1.8 1.8 1.8 1.8 1.8 1.8 1.8 1.8
x5 (A/mm2) 3.6 3.6 3.6 3.6 3.6 3.6 3.6 3.6 3.6
x6 (A/mm2) 3.6 3.6 3.6 3.6 3.6 3.6 3.6 3.6 3.6
k k k
x7 (Volts) NA NA NA 13.22 13.23 13.20 13.22 13.21 13.22
k k k k k k
x8 NA NA NA NA NA NA 4 4 4
CTM (Euro) 3863 3863 3863 3817 3824 3817 3817 3820 3817
PLL (Watts) 5012 5012 5012 4886 4880 4865 4886 4864 4886
PNLL (Watts) 856 856 856 854 858 861 854 862 854
CTLT (Euro) 28790 28790 28790 28394 28422 28400 28394 28411 28394
Twt (kg) 1271 1271 1271 1260 1263 1255 1260 1256 1260
Tlos (Watts) 5868 5868 5868 5740 5738 5726 5740 5726 5740
k
NA – Not applicable.

search technique, CTDP [29], in solving this complex TD optimiza-
tion problem. Here, CMA-ES minimize objective functions z1, and
z2, Eqs. (2) and (3), subject to the constraints Eqs. (9)–(16), by find-
ing optimum values for the design vector, dv1 [15]. Transformer
output variables (ov) like CTM, CTLT, PLL, PNLL, Twt, Tlos are given in
Table 1. As can be seen from Table 1, three techniques converged
to three different solutions for both the objectives. In particular,
the CMA-ES algorithm converges to the best result with the cost
savings of about 8% and 15%, for z1 and, 3% and 2%, for z2, when
compared with BBA-MINLP method and CTDP.
Case study 2
The CMA-ES optimizes the design vector dv2, minimizing the
objective functions z1 to z4, Eqs. (2)–(5) subject to the constraints,
Eqs. (9)–(16), and simulation results are tabulated in Table 2. This
case study discusses the effectiveness of the addition of proposed
TD variable x7, voltage per turn in place of x1, LV turns in the exist-
ing design vector dv1 taken from [15]. Variable x7 is varied at dis-
crete levels in the interval [0 15] with step 0.01. In this case, an
additional constraint is included that variable x1, LV turns must
 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218 215

Table 4
Comparison of TD Optimization results of CMA-ES, SaDE, RGA for z2.

dv/ov dv1 dv2 dv3

CMA-ES SaDE RGA CMA-ES SaDE RGA CMA-ES SaDE RGA
k k k k k k
x1 17 17 17 NA NA NA NA NA NA
x2 (mm) 242 242 242 245 245 245 245 245 245
x3 (mm) 242 242 242 245 245 245 245 245 245
x4 (Tesla) 1.6 1.6 1.6 1.6 1.6 1.6 1.7 1.7 1.7
x5 (A/mm2) 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0
x6 (A/mm2) 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0
k k k
x7 (Volts) NA NA NA 13.20 13.20 13.20 13.20 13.20 13.20
k k k k k k
x8 NA NA NA NA NA NA 5 5 5
CTM (Euro) 4617 4617 4617 4543 4543 4543 4629 4629 4629
PLL (Watts) 4265 4265 4265 4297 4297 4297 4215 4215 4215
PNLL (Watts) 685 685 685 662 662 662 613 613 613
CTLT (Euro) 26588 26588 26588 26371 26371 26371 25882 25882 25882
Twt (kg) 1460 1460 1460 1428 1428 1428 1360 1360 1360
Tlos (Watts) 4950 4950 4950 4959 4959 4959 4828 4828 4828
k
NA – Not applicable.

Table 5
Comparison of TD Optimization results of CMA-ES, SaDE, RGA for z3.

dv/ov dv1 dv2 dv3

CMA-ES SaDE RGA CMA-ES SaDE RGA CMA-ES SaDE RGA
k k k k k k
x1 19 19 19 NA NA NA NA NA NA
x2 (mm) 248 248 248 244 245 244 232 213 232
x3 (mm) 270 270 270 275 273 273 285 277 285
x4 (Tesla) 1.8 1.8 1.8 1.8 1.8 1.8 1.8 1.8 1.8
x5 (A/mm2) 3.6 3.6 3.6 3.6 3.6 3.6 3.0 3.4 3.0
x6 (A/mm2) 3.0 3.0 3.0 3.0 3.0 3.0 3.45 3.45 3.45
k k k
x7 (Volts) NA NA NA 11.85 11.91 11.90 11.27 11.27 11.27
k k k k k k
x8 NA NA NA NA NA NA 10 9 10
CTM (Euro) 3993 3993 3993 3970 3971 3970 4820 4781 4820
PLL (Watts) 4771 4771 4771 4781 4779 4779 4981 5157 4981
PNLL (Watts) 818 818 818 804 805 805 563 552 563
CTLT (Euro) 28060 28060 28060 27936 27941 27939 27648 27939 27648
Twt (kg) 1238 1238 1238 1229 1229 1229 1216 1220 1216
Tlos (Watts) 5589 5589 5589 5585 5584 5584 5544 5709 5544
k
NA – Not applicable.

Table 6
Comparison of TD Optimization results of CMA-ES, SaDE, RGA for z4.

dv/ov dv1 dv2 dv3

CMA-ES SaDE RGA CMA-ES SaDE RGA CMA-ES SaDE RGA
k k k k k k
x1 16 16 16 NA NA NA NA NA NA
x2 (mm) 232 232 232 220 220 230 219 219 219
x3 (mm) 232 232 232 220 220 231 219 219 219
x4 (Tesla) 1.6 1.6 1.6 1.6 1.6 1.6 1.75 1.75 1.75
x5 (A/mm2) 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0
x6 (A/mm2) 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0 3.0
k k k
x7 (Volts) NA NA NA 14.90 14.93 14.55 14.91 14.91 14.91
k k k k k k
x8 NA NA NA NA NA NA 9 9 9
CTM (Euro) 4754 4754 4754 4837 4837 4750 5680 5680 5680
PLL (Watts) 4092 4092 4092 3958 3958 4072 3827 3827 3827
PNLL (Watts) 737 737 737 765 767 740 664 664 664
CTLT (Euro) 26786 26786 26786 26794 26820 26756 26844 26844 26844
Twt (kg) 1533 1533 1533 1587 1589 1530 1470 1470 1470
Tlos (Watts) 4829 4829 4829 4723 4725 4812 4491 4491 4491
k
NA – Not applicable.

be an integer. MM used for building the core is fixed as MOH-0.27
like the previous case study 1 [15]. Objective values optimized after
the inclusion of variable x7 are found lesser than the results of the
case before adding x7, for all the objectives. Optimization results
of CMA-ES with dv2 have yielded performance improvement of
1%, 0.8%, 0.7%, and 2%, respectively, for all the objective functions
z1 to z4, on comparison with the results of CMA-ES with dv1.
Case study 3
In this case study, CMA-ES optimizes the design vector dv3, for
the objective functions z1 to z4, Eqs. (2)–(5), subject to the con-
straints from Eqs. (9)–(16), and the results are depicted in Table 2.
In the TD optimization problem, soundness of the TD variable x8,
TMM is analyzed by adding it with the design vector dv2. Here,
216 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218

MM used is not common for all the objective functions, whereas Table 7
MM is also optimized, during the run. The proposed variable x8 is Design Optimization Results of 400 KVA, 20/0.4 kV, 50 Hz transformer by CMA-ES
with dv3 for z2.
used, such that the no-load losses and cost due to no load losses
for different MM can be evaluated during the design phase, and Design variables Width of core leg 245 mm
the appropriate core material, out of 10 materials can be optimized Height of core window 245 mm
Maximum flux density 1.70 Tesla
for each design objective considered. Variable x8 is defined as an Current density in LV winding 3.00 A/sq.mm
integer, such that it varies from 0 to 10 for representing the 10 Current density in HV winding 3.00 A/sq.mm
magnetic materials. In this case, an additional constraint is Voltage per turn 13.20 V
included that variable x1 LV turns must be an integer. From Table 2, Magnetic material 5 (23ZDKH90)
it is explicit that the optimization results of dv3 by CMA-ES are Core design Thickness of core leg 74 mm
superior to the solutions of dv2 for the objective functions z2, z3 Window width of small core 71 mm
Window width of large core 142 mm
and z4. The design variable x8 can impart appreciable change in
Weight of small core 72 kg
the objective values, only when the core loss calculation is used Weight of large core 142 kg
during the fitness function evaluation. Since z1 design does not rely
Tank design Total width of transformer 585 mm
on losses computation, TMM optimized in this case study 3 Total height of transformer 757 mm
(material no: 4) is same as that of the MM fixed in case study 2, Total length of transformer 1056 mm
i.e. MOH-0.27. Complete design results of 400 KVA, 20/0.4 kV, LV winding design Current in LV winding per phase 577.35 Amps
50 Hz, shell type, wound core distribution transformer by CMA- Area of LV conductor 192.45 sq.mm
ES with dv3 for the function z2 are given in Table 7. From Table 2 Total number of LV Turns 17
it is evident that CMA-ES with dv3 has given cost savings of about Width of LV conductor 226 mm
Total radial thickness of LV winding 29.58 mm
1%, 2%, for z2, material savings of about 1%, each for z3, and loss Inner diameter of LV winding 258.68 mm
reduction of about 5%, 7% for z4, on comparison with CMA-ES Outer diameter of LV winding 318.08 mm
dv2, CMA-ES dv1 respectively. From the simulation results, it is LMT of LV winding 905.50 mm
evident that CMA-ES with proposed MDV is able to find the opti- Resistance of LV winding 0.001677 ohms
Weight of LV 82.64 kg
mal solution for TD optimization problem, irrespective of the
objective function. HV winding design Total Number of HV Turns(20 KV) 1515
Max number of HV turns (+5% tap) 1591
Min number of HV turns (5% tap) 1439
Statistical analysis of CMA-ES results Area of HV conductor 2.11 sq.mm
Diameter of HV conductor 1.6 mm
Current in HV winding 6.349206 Amps
Due to the stochastic nature of CMA-ES algorithm, 25 indepen- Number of layers in HV 17
dent runs are performed to prove the consistency in obtaining the Number of HV turns per layer 91
optimal solutions and their statistical results such as mean, and SD, Thickness of HV winding 35.25 mm
Total radial thickness of HV winding 41.89 mm
apart from the best, and worst solutions are reported in Table 8 for
Total height of HV winding 161 mm
all the four objective functions z1 to z4, using dv1, dv2, and dv3. The Inner diameter of HV winding 331.92 mm
performance of CMA-ES is shown with respect to solution Outer diameter of HV winding 424.18 mm
accuracy, and computation time (CT) to reach the optimum in LMT of HV winding 1187.09 mm
Table 8. For all the objective functions employed, the numerical Resistance of HV winding 17.81 ohms
Weight of HV 132.72 kg
results of CMA-ES are found more satisfactory, in terms of
performance, consistency, and faster convergence for the TD opti- Specific Total copper loss 4215 W
characteristics Total core loss 613 W
mization problem.
Total loss 4828 W
Percentage resistance 1.05%
Percentage reactance 4.08%
Comparison of CMA-ES results with other EAs Percentage impedance 4.21%
Percentage regulation 3.29%
In order to provide a comparative study of CMA-ES results with Full load efficiency 98.81%
other meta heuristics, two other EAs namely, Self Adaptive Differ- Average oil temperature rise 50 °C
Total heat transferred by convection 6086 W
ential Evolution (SaDE), and Real coded Genetic Algorithm (RGA)
and radiation
are also employed to optimize the TD optimization problem and
Weight of main Weight of LV winding 82.64 kg
their simulation results are tabulated in Tables 3–6, for all the
materials Weight of HV winding 132.72 kg
objective functions z1 to z4 respectively. It is seen from Table 4 Total weight of copper 215.37 kg
that all the three algorithms converge to the same global opti- Total weight of core 497.85 kg
mum for z2 objective alone. Table 3 shows that CMA-ES and Weight of insulating paper 18.24 kg
RGA are able to attain the global solution in all the TD case stud- Weight of duct strips 9.92 kg
Weight of tank 156.52 kg
ies, whereas SaDE gets struck in local minima for dv2 and dv3 Weight of corrugated panel 111.93 kg
cases. Table 5 depicts that RGA and SaDE, find difficult to reach Weight of oil in tank 350.11 kg
the optimum weight of the transformer, when optimized with Total weight 1360 kg
the design vector dv3, on comparison with CMA-ES. Table 6 Cost of Purchase cost 4629 Euros
explains that RGA and SaDE are unable to reach the optimum loss transformer Bid price 10293 Euros
of the transformer, when compared with CM-ES for the case dv2. Operating cost 15589.38 Euros
Total cost of transformer 25882 Euros
In order to compare the efficiency of the three different EAs with
respect to consistency, 50 independent runs have been conducted
in obtaining the global solution for the objective ‘minimization of
TLTC’. On comparative study as well as statistical study, CMA-ES is Multi-objective results of NSGA-II
found suitable for TD optimization, as it is giving satisfactory glo-
bal results consistently, irrespective of the objective functions and NSGA-II has been applied to solve this TD MOP. We have
case studies. adopted the following control parameters: population size Np is
 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218 217

Table 8
Statistical analysis of CMA-ES results.

z1 (Euro) z2 (Euro) z3 (kg) z4 (Watts)

With dv1 With dv2 With dv3 With dv1 With dv2 With dv3 With dv1 With dv2 With dv3 With dv1 With dv2 With dv3
Best value 3863 3817 3817 26588 26371 25882 1238.3 1228.6 1216 4829 4723 4491
Worst value 3869 3834 3826 26588 26371 25882 1238.6 1236 1216.5 4852 4741 4491
Mean 3865 3834 3819.8 26588 26371 25882 1238 1229 1216.22 4831 4733 4491
SD 8.91 3.5 3.82 0 0 0 0.7882 1.3 0.00865 12.3 7.6 0
CT (sec) 95 88 96 92 89 86 98 102 92 93 97 98

Fuzzy set theory is then employed to find the best compromise

solution over the tradeoff curve [38] and is given in Table 9. TD
obtained from the best compromise solution shoots up the pur-
chase cost to 4227 Euros, which is 410 Euros higher than the global
minimum purchase cost of 3817 Euros obtained from single objec-
tive TD optimization by CMA-ES. But at the same time, this best
compromise solution saves 1785 Euros of TLTC, which is found to
be the effective transformer design objective.

Conclusion

In this paper, CMA-ES algorithm is employed for the optimum

design of three phase distribution transformer. The work proposed
aiming at contributing a TD that minimize the objective(s) such as
purchase cost, total life-time cost, total mass, and total loss of the
transformer using proposed MDV, taking into account the con-
straints imposed by the international standards, transformer spec-
ifications and customer needs. The validity of the CMA-ES
algorithm for solving TD optimization problem is illustrated by
its application to a 400 KVA distribution transformer design and
Fig. 2. Pareto optimal front of NSGA-II.
comparison of its simulation results with CTDP, BBA-MINLP
method, SaDE, and RGA. Case studies 1–3 have clearly investigated
Table 9
the significance of MDV for all the TD objective functions and have
Best compromise solution of TD MOP by fuzzy set theory. proven that MDV is efficient for the TD optimization problem. The
proposed MDV are not only capable of producing optimum design,
dv3/ov Single objective optimization by CMA-ES Compromise solution
by fuzzy set theory
but also render considerable cost savings, material savings and loss
z1 z2 reduction. Statistical analysis has clearly demonstrated the effec-
x1 k
NA k
NA k
NA
tiveness of CMA-ES with respect to its global searching, solution
x2 (mm) 224 245 245 precision, consistency in obtaining solutions, and faster conver-
x3 (mm) 227 245 245 gence. Thus it is evident that CMA-ES with dv3 is able to give least
x4 (Tesla) 1.8 1.7 1.75 purchase cost (z1), TLTC (z2) with cost savings of about 10%, 6% and,
x5 (A/mm2) 3.6 3.0 3.2
17%, 5% respectively on comparison with BBA-MINLP method [15]
x6 (A/mm2) 3.6 3.0 3.0
x7 (Volts) 13.22 13.20 13.2 and CDTP [29]. This paper has also dealt the solution of TD MOP
using NSGA-II with the tradeoff between purchase cost and TLTC.
x8 4 5 3
CTM (Euro) 3817 4629 4227 The best compromise solution obtained by fuzzy set theory for
PLL (Watts) 4886 4215 4274 the multi-objective TD optimization gives much better transformer
PNLL (Watts) 854 613 752 design than the one obtained by single objective TD optimization
CTLT (Euro) 28394 25882 26609 using CMA-ES, by saving 1785 Euros of TLTC, which is accepted
Twt (kg) 1260 1360 1320
Tlos (Watts) 5740 4828 5026
by SEEDT as an energy efficient TD objective.

Acknowledgements

200, stopping criterion is 20,000 function evaluations, crossover
probability is 0.9, and mutation probability rate is 0.14. Two TD
optimization objectives: the purchase cost (z1), and the TLTC (z2)
are minimized simultaneously by NSGA-II and a trade-off is
achieved between these competing objectives in POF. Considering
the complexity of the TD problem, 30 independent runs have been
carried out. From all the best optimization runs, non-dominated
solutions of the first POF are filtered and combined to form a
new Pareto set. Non-dominated sorting is again performed on this
augmented new Pareto set to extract the true POF as shown in
Fig. 2. The tradeoff curve so obtained in Fig. 2 has better diversity
characteristics than the POF obtained from single run of NSGA-II.
The authors would like to acknowledge the efforts of Marina A.
Tsili, and Eleftherios I. Amoiralis for their valuable assistance in
supplying the transformer design specifications.
References
[1] Amoiralis Eleftherios I, Tsili Marina A, Kladas Antonios G. Transformer design
and optimization: a literature survey. IEEE Trans Power Delivery
2009;24(4):1999–2024.
[2] Georgilakis PS, Doulamis ND, Doulamis AD, Hatziargyriou ND, Kollias SD. A
novel iron loss reduction technique for distribution transformers based on a
combined genetic algorithm-neural network approach. IEEE Trans Syst Man
Cybern 2001;31:16–34.
218 S. Tamilselvi, S. Baskar / Electrical Power and Energy Systems 61 (2014) 208–218
[3] Georgilakis P, Hatziargyriou N, Paparigas D. AI helps reduce transformer iron
losses. IEEE Comput Al Power 1999;12(4):41–6.
[4] Tsivgouli Anastassia J, Tsili Marina A, Kladas Antonios G, Georgilakis Pavlos S,
Souflaris Athanassios T, Skarlatini Annie D. Geometry optimization of electric
shielding in power transformers based on finite element method. J Mater
Process Technol 2007;181:159–64.
[5] Subramanian S, Padma S. Optimization of transformer design using bacterial
foraging algorithm. Int J Comput Appl 2011;19(3):52–7.
[6] Subramanian S, Padma S. Bacterial foraging algorithm based multi objective
optimal design of single phase transformer. J Comput Sci Eng 2011;6(2):1–6.
[7] Hernandez C, Arjona MA, Dong Shi-Hai. Object-oriented knowledge based
system for distribution transformer design. IEEE Trans Magn 2008;44(10).
[8] Arjona MA, Hernandez C, Cisneros-Gonzalez M. Hybrid optimum design of a
distribution transformer based on 2-D FE and a manufacturer design
methodology. IEEE Trans Magn 2010;46(8):2864–7.
[9] Jabr Rabih A. Application of geometric programming to transformer design.
IEEE Trans Magn 2005;41(11):4261–9.
[10] Amoiralis Eleftherios I, Georgilakis Pavlos S, Tsilia Marina A. Design
optimization of distribution transformers based on mixed integer
programming methodology. J Optoelectron Adv Mater 2008;10(5):1178–83.
[11] Amoiralis Eleftherios I, Tsili Marina A, Georgilakis Pavlos S, Kladas Antonios G,
Souflaris Athanassios T. A parallel mixed integer programming-finite element
method technique for global design optimization of power transformers. IEEE
Trans Magn 2008;44(6):1022–5.
[12] Padma S, Bhuvaneswari R, Subramanian S. Optimal design of power
transformer using simulated annealing technique. In: IEEE conference on
industrial technology (ICIT), Mumbai, India, December 15-17, 2006, p. 1015–
19.
[13] Georgilakis Pavlos S, Tsili Marina A, Souflaris Athanassios T. A heuristic
solution to the transformer manufacturing cost optimization problem. J Mater
Process Technol 2007;181:260–6.
[14] Georgilakis PS. Recursive genetic algorithm-finite element method technique
for the solution of transformer manufacturing cost minimization problem. IET
Electr Power Al 2009;3(6):514–9.
[15] Amoiralis Eleftherios I, Georgilakis Pavlos S, Tsili Marina A, Kladas Antonios G.
Global transformer optimization method using evolutionary design and
numerical field computation. IEEE Trans Magn 2009;45(3):1720–3.
[16] Li Hui, Han Li, He Bei, Yang Shunchang. Application research based on
improved genetic algorithm for optimum design of power transformers.
Electrical Machines and Systems, ICEMS, Proceedings of the fifth international
conference 2001;1:242–5.
[17] Zhang Shuang, Hu Qinghe, Wang Xingwei, Zhu Zhiliang. Application of chaos
genetic algorithm to transformer optimal design. In: IEEE computer society,
2009, International workshop on chaos-fractals theories and applications,
2009.
[18] Ramezani M, Falaghi H, ParsaMoghaddam M, Haghifam MR. Genetic based
algorithm for optimal placement of distribution transformers. IEEE 2006:1–5.
[19] Kefalas Themistoklis, Tsili Marina A, Kladas Antonios. Unification of anisotropy
and FEM-BE models for distribution transformer optimization. IEEE Trans
Magn 2008;10(5):1143–8.
[20] Sridhara G, Sastry VV, Venkata Rao. Comparison of non linear programming
techniques for the optimal design of transformers. Proc IEE
1977;124(12):1225–6.
[21] Amoiralis Eleftherios I, Georgilakis Pavlos S, Kefalas Themistoklis D, Tsili
Marina A, Kladas Antonios G. Artificial intelligence combined with hybrid fem-
be techniques for global transformer optimization. IEEE Trans Magn
2007;43(4):1633–6.
[22] Amoiralis Eleftherios I, Tsili Marina A, Georgilakis Pavlos S. The state of the art
in engineering methods for transformer design and optimization: a survey. J
Optoelectron Adv Mater 2008;10(5):1149–58.
[23] Cheema MAM, Fletcher JE, Dorrell David. A practical approach for the global
optimization of electromagnetic design of 3-phase core type distribution
transformer allowing for capitalization of losses. IEEE Trans Magn 2013;49(5).
[24] Coelho Leandro dos Santos, Mariani Viviana Cocco, Ferreira da Luz Mauricio V,
Leite Jean Vianei. Novel gamma differential evolution approach for
multiobjective transformer design optimization. IEEE Trans Magn
2013;49(5):2121–4.
[25] Hansen N, Ostermeier A. Completely derandomized self-adaptation in
evolution strategies. Evol Comput 2001;9(2):159–95.
[26] González Fernández Yasser, Ortiz Marta Rosa Soto, Benchmark for the CEC
2005 Special Session on Real-Parameter Optimization, Aug 29 2013.
[27] Suganthan PN, Hansen N, Liang JJ, Deb K, Chen Y-P, Auger A et al. Problem,
definitions and evaluation Criteria for the CEC 2005. In: Special session on real-
parameter optimization, technical report, Nanyang Technological University,
Singapore, May 2005.
[28] Hansen N, Auger A, Ros R, Finck S, Pošík P. Comparing results of 31 algorithms
from the black-box optimization benchmarking BBOB-2009. In: Proceedings of
the 12th annual conference companion on genetic and evolutionary
computation, GECCO 2011, ACM, July 2010, p. 1689–96.
[29] Georgilakis PavlosS. Spotlight on modern transformer design. Springer-Verlag
London Limited; 2009.
[30] http://www.nsc.co.jp/en/product/kind/sheet/product-list.html.
[31] Nikolaus Hansen, The covariance matrix adaptation evolution strategy: a
tutorial –: April 2008.
[32] Hansen N. The CMA evolution strategy: a comparing review. In: Lozano JA,
Larrañga P, Inza I, Bengoetxea E, editors. Towards a new evolutionary
computation. Advances in estimation of distribution algorithms. Springer;
2006. p. 75–102.
[33] Deb Kalyanmoy. An efficient constraint handling method for genetic
algorithms. Comput Methods Appl Mech Eng 2000;186(2/4):311–38.
[34] Hansen N, Kern S. Evaluating the CMA evolution strategy on multimodal test
functions. In: Yao Xin et al., editors. Parallel problem solving from nature –
PPSN VIII, LNCS, vol. 3242. Springer; 2004. p. 282–91.
[35] Deb Kalyanmoy, Pratap Amrit, Agarwal Sameer, Meyarivan T. A Fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE Trans Evol Comput 2002;6(2).
[36] CENELEC, Harmonization Document HD428: 1 S1, 1992.
[37] Hansen N. The CMA evolution strategy: a tutorial, June 28, 2011.
[38] Abido MA. Multi objective evolutionary algorithms for electric power dispatch
problem. IEEE Trans Evol Comput 2006;10(3):315–29.