Anda di halaman 1dari 22

Dear Author,

Here are the proofs of your article.

• You can submit your corrections online, via e-mail or by fax.

• For online submission please insert your corrections in the online correction form. Always
indicate the line number to which the correction refers.

• You can also insert your corrections in the proof PDF and email the annotated PDF.

• For fax submission, please ensure that your corrections are clearly legible. Use a fine black
pen and write the correction in the margin, not too close to the edge of the page.

• Remember to note the journal title, article number, and your name when sending your
response via e-mail or fax.

• Check the metadata sheet to make sure that the header information, especially author names
and the corresponding affiliations are correctly shown.

• Check the questions that may have arisen during copy editing and insert your answers/
corrections.

• Check that the text is complete and that all figures, tables and their legends are included. Also
check the accuracy of special characters, equations, and electronic supplementary material if
applicable. If necessary refer to the Edited manuscript.

• The publication of inaccurate data such as dosages and units can have serious consequences.
Please take particular care that all such details are correct.

• Please do not make changes that involve only matters of style. We have generally introduced
forms that follow the journal’s style.
Substantial changes in content, e.g., new results, corrected values, title and authorship are not
allowed without the approval of the responsible editor. In such a case, please contact the
Editorial Office and return his/her consent together with the proof.

• If we do not receive your corrections within 48 hours, we will send you a reminder.

• Your article will be published Online First approximately one week after receipt of your
corrected proofs. This is the official first publication citable with the DOI. Further changes
are, therefore, not possible.

• The printed version will follow in a forthcoming issue.

Please note
After online publication, subscribers (personal/institutional) to this journal will have access to the
complete article via the DOI using the URL: http://dx.doi.org/[DOI].
If you would like to know when your article has been published online, take advantage of our free
alert service. For registration and further information go to: http://www.link.springer.com.
Due to the electronic nature of the procedure, the manuscript and the original figures will only be
returned to you on special request. When you return your corrections, please inform us if you would
like to have these documents returned.
Metadata of the article that will be visualized in
OnlineFirst

Please note: Images will appear in color online but will be printed in black and white.
ArticleTitle Chaos-assisted multiobjective evolutionary algorithm to the design of transformer
Article Sub-Title
Article CopyRight Springer-Verlag Berlin Heidelberg
(This will be the copyright line in the final PDF)
Journal Name Soft Computing
Corresponding Author Family Name Tamilselvi
Particle
Given Name S.
Suffix
Division Department of Electrical and Electronics Engineering
Organization Sri Venkateswara College of Engineering
Address Chennai, Tamil Nadu, India
Email tamilselvi.manjuraj@gmail.com
ORCID

Author Family Name Baskar


Particle
Given Name S.
Suffix
Division Department of Electrical and Electronics Engineering
Organization Thiagarajar College of Engineering
Address Madurai, Tamil Nadu, India
Email sbeee@tce.edu
ORCID
Author Family Name Anandapadmanaban
Particle
Given Name L.
Suffix
Division Department of EEE
Organization Kingston College of Engineering
Address Vellore, Tamil Nadu, India
Email chari_master@yahoo.co.in
ORCID
Author Family Name Kadhar
Particle
Given Name K. Mohaideen Abdul
Suffix
Division Department of EEE
Organization M.V.J. College of Engineering
Address Bangalore, Karnataka, India
Email mohaideen9683@gmail.com
ORCID
Author Family Name Varshini
Particle
Given Name P. R.
Suffix
Division Department of Electrical and Electronics Engineering
Organization Thiagarajar College of Engineering
Address Madurai, Tamil Nadu, India
Email varshini@tce.edu
ORCID

Received
Schedule Revised
Accepted
Abstract In this paper, multiobjective transformer design (TD) optimization is carried out using multiobjective
evolutionary algorithm (MOEA) based on decomposition with dynamical resource allocation (MOEA/D-
DRA) for four sets of conflicting TD bi-objectives such as (i) purchase cost and total loss, (ii) purchase
cost and total lifetime cost (TLTC), (iii) total mass and total loss and (iv) total mass and TLTC, subjected
to 14 various practical constraints. Significant decision variables with enlarged search space are employed
for obtaining reliable and efficient TD with minimum losses and TLTC. TD is accompanied by 3D-finite
element method assessment to validate the designed no-load loss calculated from analytical equations. To
improve the searching ability of MOEA/D-DRA (MDRA) in solving this complex multimodal TD
optimization problem (TDOP), this paper proposes integration of chaos with MDRA, enabling chaotic
variation in the crossover rate and mutation scaling factor. To prove the effectiveness of chaos-assisted
MOEA, logistic chaotic map-assisted MDRA, and iterative chaotic map with infinite collapses- (ICMIC)
assisted MDRA (ICMDRA) have been successfully applied to multiobjective TDOP and their TD results
are compared with those of MDRA, knee point-driven evolutionary multiobjective optimization algorithm
(KnEA), and non-dominated sorting genetic algorithm (NSGA) II. This paper identifies which chaotic map
can assist MDRA and solve TDOP by comparative analysis of performance of logistic and ICMIC chaotic
maps. Efficient TD results and two MOEA performance indicators confirm the superiority of ICMDRA
over all the other MOEAs in terms of diversity and convergence in solving TDOP.
Keywords (separated by '-') Multiobjective transformer design optimization - MOEA/D-DRA - Chaos - NSGA II - KnEA
Footnote Information Communicated by V. Loia.
Soft Comput
DOI 10.1007/s00500-016-2145-7

METHODOLOGIES AND APPLICATION

Chaos-assisted multiobjective evolutionary algorithm to the design

of
of transformer
S. Tamilselvi1 · S. Baskar2 · L. Anandapadmanaban3 ·
K. Mohaideen Abdul Kadhar4 · P. R. Varshini2

pro
Author Proof

© Springer-Verlag Berlin Heidelberg 2016

1 Abstract In this paper, multiobjective transformer design ity of MOEA/D-DRA (MDRA) in solving this complex 14

(TD) optimization is carried out using multiobjective evo- multimodal TD optimization problem (TDOP), this paper
2

6
lutionary algorithm (MOEA) based on decomposition with
dynamical resource allocation (MOEA/D-DRA) for four sets
of conflicting TD bi-objectives such as (i) purchase cost and
total loss, (ii) purchase cost and total lifetime cost (TLTC),
cted proposes integration of chaos with MDRA, enabling chaotic
variation in the crossover rate and mutation scaling factor.
To prove the effectiveness of chaos-assisted MOEA, logis-
tic chaotic map-assisted MDRA, and iterative chaotic map
15

16

17

18

19

7 (iii) total mass and total loss and (iv) total mass and TLTC, with infinite collapses- (ICMIC) assisted MDRA (ICMDRA) 20

1 8 subjected to 14 various practical constraints. Significant deci- have been successfully applied to multiobjective TDOP and 21

9 sion variables with enlarged search space are employed for their TD results are compared with those of MDRA, knee 22

10 obtaining reliable and efficient TD with minimum losses point-driven evolutionary multiobjective optimization algo- 23

11 and TLTC. TD is accompanied by 3D-finite element method rithm (KnEA), and non-dominated sorting genetic algorithm 24
orre
12 assessment to validate the designed no-load loss calculated (NSGA) II. This paper identifies which chaotic map can assist 25

13 from analytical equations. To improve the searching abil- MDRA and solve TDOP by comparative analysis of perfor- 26

mance of logistic and ICMIC chaotic maps. Efficient TD 27


Communicated by V. Loia. results and two MOEA performance indicators confirm the 28

superiority of ICMDRA over all the other MOEAs in terms


B S. Tamilselvi
29

tamilselvi.manjuraj@gmail.com; tamilselvis@svce.ac.in of diversity and convergence in solving TDOP. 30

S. Baskar
sbeee@tce.edu Keywords Multiobjective transformer design optimiza- 31

L. Anandapadmanaban
tion · MOEA/D-DRA · Chaos · NSGA II · KnEA 32

chari_master@yahoo.co.in
unc

K. Mohaideen Abdul Kadhar


mohaideen9683@gmail.com 1 Introduction 33

P. R. Varshini
varshini@tce.edu Distribution transformers (DTs) are electric machines that 34

enable both the transmission and the distribution of elec- 35


1 Department of Electrical and Electronics Engineering, Sri tric energy in a cost-effective way. Since the transformer life 36
Venkateswara College of Engineering, Chennai, Tamil Nadu,
India
expectancy deteriorates, a wide research to produce an effi- 37

2
cient transformer design (TD) is crucial. Williams et al., in the 38
Department of Electrical and Electronics Engineering,
Thiagarajar College of Engineering, Madurai, Tamil Nadu,
year 1955, had dimensioned various parts of the transformer 39

India and brought out the TD in computer. After the 1970s, more 40

3 Department of EEE, Kingston College of Engineering,


focus was given to either minimize transformer weight, active 41

Vellore, Tamil Nadu, India part cost (APC) (sum of core and winding cost), load loss or 42

4 Department of EEE, M.V.J. College of Engineering,


maximize efficiency, VA output using several mathematical 43

Bangalore, Karnataka, India programming techniques (MPT) such as conventional TD 44

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

45 procedure, calculus of variation, geometric programming, cost and calculated present value of future lifetime losses as 98

46 mixed integer programming (MIP), mathematica software, well. But, some commercial customers and contractors who 99

47 etc,. (Judd and Kressler 1977; Wu and Lee 1980; Pham et al. are not conscious about this fact purchase transformers at 100

48 1996; Jabr 2005; Amoiralis et al. 2008a; Masood et al. 2012). the lowest cost. Therefore, it falls within the duties of TM to 101

49 After the 1970s, MPT such as calculus of variation, geomet- satisfy the contradictory needs of these customers. 102

ric programming, MIP, and mathematica software became Thus customer specification request, standards, TD prob-

of
50 103

51 the focal point. Hence, Judd and Kressler (1977), Wu and lem nature, and market globalization have taken the TM to 104

52 Lee (1980), Pham et al. (1996), Jabr (2005), Amoiralis et al. face a competitive environment for maintaining sales figures, 105

53 (2008a), and Masood et al. (2012) had concentrated all their and necessitate expressing the TD problem as multiobjec- 106

54 attention on the needs of minimizing transformer weight, tive optimization problem (MOP). Such formulation can 107

APC (sum of core and winding cost), load loss or maximizing only aid finding a trade-off between contradictory objec-

pro
55 108

56 efficiency, VA output, etc, using MPT. After 2005, a combi- tives. But research activities tracked in TD optimization 109

57 nation of numerical technique with MPT such as boundary (TDO) field reveal that efforts are focused only on single 110

element (BE) method, knowledge-based systems (KBS), par- objective optimization (SOO). In contrast to SOO, MOP
Author Proof

58 111

59 allel MIP, mixed integer non-linear programming (MINLP) finds such a solution that satisfies all the objectives that 112

60 with branch bound algorithm for either minimizing power leads to a set of solutions termed as Pareto optimal set. 113

61 loss, main materials cost, total life time cost (sum of initial Conventionally, MOPs are solved by converting them into 114

62 cost and cost due to losses throughout the transformer life- SOO problem, resulting in single solution using weighted 115

63 time) (TLTC) or maximizing efficiency had started appearing sum method, lexicographic method, goal programming, 116

in the literature (Tsili et al. 2005; Tsivgouli et al. 2007; Her- e-constraint method, etc. But for getting many trade-off
64

65

66

67

68
nandez and Arjona 2007; Amoiralis et al. 2008b, 2009).
cted
After 1999, TD research activities were geared up, due to the
various applications of artificial intelligence (AI) techniques
such as iron loss reduction based on neural network (Georgi-
solutions, these approaches need to be executed many times
with appropriate weight coefficients. Evolutionary algo-
rithms (EAs) are naturally suitable for solving MOPs as they
use population-based strategy for achieving an approxima-
117

118

119

120

121

69 lakis et al. 1999, 2001), APC minimization using improved tion of Pareto set in a single run and can handle discrete, 122

70 genetic algorithm (GA), simulated annealing technique, and multimodal, discontinuous problems with mixed variables. 123

71 bacterial foraging algorithm (BFA), (Hui et al. 2001; Padma The concept of solving MOP using EA was first suggested 124

72 et al. 2006; Subramanian and Padma 2011a), manufacturing by Rosenberg in the late 1960s. In 1985, Schaffer developed 125

73 cost (MC) minimization on the usage of the heuristic solution vector evaluated GA, evaluating one objective function at 126
orre
74 (Georgilakis et al. 2007a), winding material selection with a time for a subproblem. But it resulted in a weak Pareto 127

75 decision tree method (Georgilakis et al. 2007b), and TLTC front. Then, non-dominated sorting algorithms such as non- 128

76 minimization using covariance matrix adaptation evolution dominated sorting GA (NSGA), Niched Pareto GA (NPGA), 129

77 strategy (CMA-ES) (Tamilselvi and Baskar 2014). To find the NSGA-II, Strength Pareto EA (SPEA), Pareto Archived 130

78 optimal TD, after 2007, researches had taken place on fusion Evolutionary Strategy (PAES), and SPEA II were intro- 131

79 of AI technique and FEM such as neural network combined duced in the 1990s, for problems with two objectives. For 132

80 with hybrid FEM-BE, recursive GA combined with FEM for problems with three or more objectives, non-dominated sort- 133

2 81 the minimization of MC (Amoirals et al. 2007; Georgilakis ing and diversity maintenance operator are computationally 134

82 2009a), and main material cost (Amoiralis et al. 2012). From expensive. Efficient, reference point-based EAs such as mul- 135

83 2013, there was a move toward the mixture of MPT and sto- tiobjective EA based on decomposition (MOEA/D) (Zhang 136
unc

84 chastic methods for minimizing TLTC (Cheema et al. 2013). and Li 2007), MOEA/D with dynamic resource allocation 137

85 The TD problem is a constrained, multidimensional, (MOEA/D-DRA) approach (Zhang et al. 2009), NSGA-III, 138

86 multiobjective, mixed variable, non-convex, non-linear opti- and knee point-driven evolutionary multiobjective optimiza- 139

87 mization problem. Since TD stands as a complex iterative tion algorithm (KnEA) (Zhang et al. 2015) have been evolved 140

88 procedure, experienced designers approximate the design for many objective problems recently. 141

89 process to reduce the computation burden. Customers who However, so far only a very few works with solution to TD 142

90 have realized this fact specify limits for transformer parame- MOP including (i) BFA with sum of weighted objectives for 143

91 ters to transformer manufacturers (TM) for improving the purchase cost minimization and efficiency maximization of 144

92 reliability. Manufacturers are so forced to adopt a safety single-phase transformer using very few decision variables 145

93 margin to correct these parameters within the desired lim- (Subramanian and Padma 2011b), (ii) multiobjective differ- 146

94 its specified by international standards and specifications. ential evolution (DE) for minimization of mass and losses 147

95 Moreover, some electric utilities and industrial users, those of 0.4 kVA single phase transformer (Coelho and Mariani 148

96 who are aware that losses amount to economic cost, prefer 2013; Coelho et al. 2014), and (iii) multiobjective particle 149

97 efficient minimum TLTC transformers. TLTC considers first swarm optimization minimizing core and copper volume for 150

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

151 power transformers design (Adly and Abd-El-Hafiz 2014) are the cost of the remaining material and labor for building 193

152 have been advanced. Overall, literature survey reveals the core with ‘m’ steps, respectively in $. 194

153 following: (i) mostly MPTs with less number of decision


154 variables and constraints, which cannot promise efficient TD 2.1.2 Minimization of total mass ( f 2 ) 195

155 are employed; (ii) no single literature in MOP has minimized


TLTC, although it renders efficient TD. (iii) Nevertheless, It minimizes the mass of the transformer materials with

of
156 196

157 literature with MOEAs also has not concentrated on recent


158 MOEA, and overall TD aspects. Hence, this paper addresses Twt = {G FE + G LV,l + G HV,l + G INS 197

159 all these drawbacks by solving multiobjective TDO prob- + G OIL + G RAD + G SS }(kg). (3) 198

160 lem using recent and efficient reference point-based MOEA,


without compromising multiple contradictory bi-objectives,

pro
161 2.1.3 Minimization of total loss ( f 3 ) 199

162 3D FEM, enlarged search space, practical constraints, and


163 minimum TLTC. This objective minimizes the transformer total loss by 200
Author Proof

Tlos = {PNLL + PLL }(W). (4) 201

164 2 Multiobjective transformer design optimization


165 problem formulation 2.1.4 Minimization of TLTC ( f 4 ) 202

166 Multiobjective TDO of three phase, oil immersed, multi- This objective minimizes TLTC using 203

stepped, core-type transformers is carried out in this paper.


167

168

169

170
The proposed methodology optimizes two conflicting objec-
tives simultaneously, subject to 14 constraints using eight
decision variables. Mathematically, MOP can be defined as
cted CTLT =



(CTM )
(1 − Sm )

′ ′

+ A .PNLL + B .PLL ($),

where A , B , Sm , PNLL , and PLL are no-load loss cost rate


(5) 204

205

171 Minimize F(x) = { f i : ∀i = 1, . . . , M} in $/W, load loss cost rate in $/W, sales margin, designed 206

172 Subject to h j (x) ≤ d j , for j = 1, 2, . . . , q no-load loss in watts, and designed load loss in watts, respec- 207

173 xt ≥ 0, for t = 1, 2, . . . , n tively. 208

174 xt,min ≤ xt ≤ xt,max , for t = 1, 2, . . . , n (1)


2.2 Transformer design constraints 209
orre

175 where f i is the ith objective function; ‘M’ is the number of


176 objectives; ‘h j ’ is the jth inequality constraint; ‘xt ’ is the tth No−Load loss constraint : PNLL ≤ (1 + t1 ).PGNLL (6) 210

177 decision variable; ‘n’ is the dimension; ‘d j ’ are constants of Load loss constraint PLL ≤ (1 + t1 ).PGLL (7) 211

178 constraints; q is the number of constraints; xt,min and xt,max Total loss constraint : Tlos ≤ (1 + t2 )(PGLL + PGNLL ) (8) 212
179 are lower and upper limit on ‘xt ’, respectively.
Impedance constraint : (1 − t2 )Us ≤ Uk ≤ (1 + t2 )Us (9) 213

180 2.1 Transformer design objectives Efficiency constraint : η ≥ ηmin (10) 214

Magnetizing current constraint : Im ≤ Im,max (11) 215

181 2.1.1 Minimization of purchase cost ( f 1 ) 216


unc

182 It minimizes the first cost of the transformer using Strip size constraint : 217

(Rs + 2) ≤ As ≤ (6Rs )
183 CTM = {CFE,k .G FE + CCO,l .(G LV,l + G HV,l ) (12), (13) 218
(R p + 2) ≤ A p ≤ (6R p )
184 + CINS .G INS + COIL .G OIL + CRAD .G RAD 219

185 + CSS .G SS + Crem + Clab,m }($), (2)


Window constraint : 220

186 where G FE , G LV,l , G HV,l , G INS , G OIL , G RAD , and G SS are ⎨ 2d ≤ x2 ≤ 3d, if x7 = 1
187 the mass of the magnetic material, LV conductor, HV con- 3d ≤ x2 ≤ 4d, if x7 = 2 (14) − (16) 221

ductor, insulating paper, oil, radiator, and sheet steel in kilo- (0.3x2 + d) ≤ D ≤ (0.3x2 + d + 10)

188

189 gram respectively; CFE,k=1,2,...,10 , CCO,l=1,2 , CINS , COIL , 222

190 CRAD , and CSS are the unit cost of the ‘kth’ magnetic mate-
191 rial, ‘lth’ conductor material, insulating paper, oil, radiator, Thermal constraints : 223

192 and sheet steel in $/kg respectively; Crem , and Clab,m=1,2,...,4 G LV,D ≤ GraLV,max ; G HV,D ≤ GraHV,max (17), (18) 224

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

4
x 10
225
226 Boundary constraints :
2.66
227 xt,min ≤ xt ≤ xt,max , for t = 1, 2, . . . , n, (19)
2.64
where PGLL , PGNLL are guaranteed load loss and no-load

TLTC ($)
228

229 loss in watts respectively;t1 is the tolerance allowed for 2.62

of
230 designed no-load loss and load loss in %; t2 is the toler-
2.6
231 ance allowed for designed total loss and impedance in %; d
232 is the diameter of the core in mm; As , Rs , A p , and R p are
2.58
233 breadth and thickness of the rectangular strip of LV and HV 8
234 windings, respectively, in mm; D is the distance between 6

pro
235 the core centers in mm; Uk , Us are designed and guaran- 4
100
teed impedances, respectively, in %; G LV,D , G HV,D , and 2 80
236 60
40
0 20
237 GraLV,max , GraHV,max are the designed temperature gradi- width of core leg 0
height of window
Author Proof

238 ents of LV, HV windings, and maximum allowed temperature


239 gradients of LV and HV windings in deg, respectively; Im , Fig. 1 Surface plot of TLTC vs window height vs width of core leg
240 and Im,max are designed and maximum allowed magnetiz- core leg
241 ing currents in % respectively; η and ηmin are designed and 4
x 10
242 minimum required efficiency in %.
2.63

243

244

245

246
2.3 Decision vector and search space

window height (x2 ) in mm, maximum flux density (x3 ) in


tesla, current density of HV winding (x4 ) in A/mm2 , current
cted
Decision variables, xt = 1,2,...,8 are volt per turn (x1 ) in Volts,
TLTC ($) 2.62

2.61

2.6
247 density of LV winding (x5 ) in A/mm2 , TMM (x6 ), TWM
248 (x7 ), and NCS (x8 ). Variables x1 , x3 , x4 and x5 are discrete 2.59

249 variables. x1 varies at discrete values of the interval 10–25


2.58 100
250 with 0.01 V step. x3 is varied within 1.6–1.8 in the step of 0.01 100
80
T. x4 and x5 vary in the step of 0.01 A/mm2 within the range 60 50
orre
251 40
20
of 2.5–3.6 for Cu and 1.5–2.6 for Al. Variables x2 , x6 , and 0 0 width of core leg
252 volt per turn
253 x8 are integers. x2 takes a value between 200 and 1000. x6 is
254 varied in the interval 1–10 for indicating ten magnetic mate- Fig. 2 Surface plot of TLTC vs volt per turn vs width of width of core
leg
255 rials, {M6 0.35, M5 0.30, M4 0.27, M3 0.23, 23ZDKH90,
256 23ZDMH85, 23ZH90, H2 0.3, H1 0.27, H0 0.23}, respec- Galván et al. 2010). Eight variables are used as decision 273
257 tively. x8 takes 1–4 for identifying {‘5 step’, ‘7 step’, ‘9 step’, variables. Considering only two basic decision variables, 274
258 ‘11 step’}, respectively. x7 is a binary variable that assumes the surface plot of the objective function (TLTC) has been 275
259 [0, 1], for representing {‘Cu’, ‘Al’}, respectively. analyzed and found to be multimodal as in Figs. 1 and 2. 276

An efficient MOEA, which can handle complicated eight- 277


unc

260 3 Implementation of MOEA for TD MOP dimensional TD MOP with its strong searching ability in 278

terms of finding a diverse set of solutions and converging 279

261 In general, TDO problem (TDOP) has only two conflict- to the near optimal Pareto front is needed. On compari- 280

262 ing objectives. TD MOP is thus formulated with four sets son with NSGA II, MOEA/D algorithm is promising in 281

263 of conflicting bi-objectives such as minimization of (i) pur- dealing with most of the continuous multiobjective optimiza- 282

264 chase cost ( f 1 ) and total loss ( f 3 ); (ii) purchase cost ( f 1 ) tion test instances with complicated Pareto set shapes (Li 283

265 and TLTC ( f 4 ); (iii) total mass ( f 2 ) and total loss ( f 3 ); and and Zhang 2009). MOEA/D with DRA strategy has outper- 284

266 (iv) total mass ( f 2 ) and TLTC ( f 4 ), subjected to constraints formed its other similar algorithms in MOEA competition in 285

267 (6)–(19). The transformer considered has a rating of 1500 CEC2009 (Zhang et al. 2009). But MOEA/D-DRA (MDRA) 286

268 kVA, 3, oil immersed, 13.2/0.22 kV, 60 Hz (Hernandez and applies simply DE operators with fixed ‘F’ and ‘CR’ for 287

269 Arjona 2007). The design specifications are PGNLL = 2500 recombination and crossover. Nelder and Mead’s simplex 288

270 W, PGLL = 12500 W, Us = 5 %, GraLV,max = GraHV,max = optimization technique lent an idea to DE, which is the dif- 289

271 20◦ , Im,max = 1 %, ηmin = 98 %, t1 = 15 %, and t2 = ference between two randomly chosen vectors. DE mutates 290
′ ′
272 10 %. Factors A = 8.16 $/W, B = 4.02 $/W (Olivares- these vectors by adding weighted random vector differentials 291

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

292 to them. Genetic operators of DE have two control parame- behavior dependent on initial value and control parameter 318

293 ters crossover rate, ‘CR’, and mutation scaling factor, ‘F’, (Chen et al. 2014). The simplest one-dimensional dynamic 319

294 for convergence. These parameters do not get changed during system evidencing chaotic behavior is the logistic map. 320

295 the search procedure and cannot guarantee the optimization Different chaotic maps augmented over MOEA contribute 321

296 ergodicity in the search phase completely and lead to pre- different degrees of performance improvement for any prob- 322

mature convergence (Coelho and Mariani 2006). Tuning of lem. To identify which chaotic map can give the best

of
297 323

298 these parameters is essential during the evolutionary search assistance for MDRA algorithm and TD MOP, comparative 324

299 and this complex multiobjective TDOP mandates suitable analysis of logistic map and iterative chaotic map with infi- 325

300 methodology to control these parameters in different stages nite collapses (ICMIC) map has been done, by integrating 326

301 of the search process. Recently, the application of chaotic them with MDRA in this paper. 327

sequences instead of random sequences in DE is a powerful

pro
302

303 strategy to diversify the population and improve the DE’s per- 3.1.1 Logistic map 328

304 formance in preventing premature convergence (Coelho and


Mariani 2006). Literature such as chaotic NSGA II (Chen In fact, logistic map made an excellent point of entry to con-
Author Proof

305 329

306 et al. 2014) and chaotic PSO (Altinoz et al. 2010) show that sider chaos concept into MOEA by its simplicity. In logistic 330

307 EA embedded with chaos is an efficient approach for improv- map, the chaotic behavior can arise from very simple non- 331

308 ing the searching ability of the algorithm. In this direction, for linear dynamical equation, 332

309 the first time, this paper proposes chaos-assisted MOEA/D-


310 DE-DRA and provide comparison of performance of five cxk = µ1 .cxk−1 (1 − cxk−1 ). (21) 333

various MOEAs in solving multiobjective TDOP.


311

cted
_________________________________________________________________________________________________________
334

Framework of MDRA
Require: M (number of objectives), N (number of subproblems), T (number of neighborhoods), max_gen (maximum number of
generations), δ (probability that parent solutions are selected from the neighborhood), nr (maximum number of replacements), F (mutation
scaling factor), CR (crossover rate), ηm (polynomial mutation index), pm (probability to perform polynomial mutation), stopping criteria.
1: Initial population generate N subproblems
2: Evaluate fitness for the initial population
orre

3: Neighborhood Generate N weight vectors and identify T closest weight vector for each subproblem
4: Initialize reference utopia point and set utility
5: while stopping criterion not fulfilled do
6: Mating Parents Select {N/5-M +2} parents (population, 10-tournament selection, and utility)
7: for all parents do
8: Generate Child variation (DE crossover and polynomial mutation)
9: Evaluate fitness for the offspring
10: Improve of utopia point
11: Apply replacement strategy (Child, neighborhood, constraint handling, nr, Tchebycheff approach)
12: end for
unc

13: Update utility


14: Increment generation count
15: end while (Objective values of the populationgen = max_gen forms the Pareto set
16: return population for the next generation
335

336

where cxk−1 means the parent chaos variable, cxk means the 337
312 3.1 Chaos-assisted MOEA/D-DE-DRA algorithm
children chaos variables, ‘k’ denotes the current iteration, 338

‘k−1’ is the previous iteration and ‘µ1 ’ is the control para- 339
313 Chaos has attracted much attention in research, due to its
meter. The chaotic sequence generated by logistic map using 340
314 characteristics such as randomicity, ergodicity, regularity,
(21) is dependent on (i) the initial value of chaotic variable 341
315 and sensibility. The description of chaos is that chaotic
cx and (ii) µ1 . Chaotic sequences display an unpredictable 342
316 systems exhibit a great sensitivity to the initial conditions.
long-term behavior, due to their sensitiveness to initial con- 343
317 Chaos is a kind of characteristic of non-linear dynamic sys-
ditions. Small difference in the initial value of chaos variable 344
tem which exhibits bounded dynamic unstable and ergodic
cause large differences in its longtime behavior. The source 345

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

(a) 1 (b) (c) 1

0.9 0.9

1
0.8 0.8

0.7 0.8 0.7

of
0.6 0.6 0.6

cx(k+1)
cx(k)

cx(k)
0.5 0.4 0.5

0.4 0.4
0.2

pro
0.3 0.3
0
1
0.2 0.2
1
0.1 0.1
Author Proof

0.5
0.5
0 0
0 0.2 0.4 0.6 0.8 1 cx(k) 0 0 0 200 400 600 800 1000
cx(k-1)
cx(k-1) Iteration (k)

Fig. 3 Quadratic curve of the logistic map

346

347

348

349
of such sensitivity to initial conditions is that the map rep-
resents a repeated folding and stretching of the space on
which it is defined. The control parameter µ1 decides the
behavior of the logistic map to show whether stabilized or
cted ted between the parent chaotic variables and children chaotic
variables in 2D Fig. 3a and 3D in Fig. 3b.

3.1.2 ICMIC map


379

380

381

350 oscillatory or chaotic. Different values of control parame-


351 ter µ1 and initial values of cx have been tried to learn how Lyapunov exponent, an important indicator used to show the 382

352 and when chaotic behavior can arise from the non-linear existence of chaos, can also reflect the sensitivity to initial 383

353 Eq. (21). This logistic map behaves chaotically in an unpre- values of chaotic maps. Some 1-D iterative chaotic maps such 384

354 dictable manner only when µ1 = 4 and the initial value as the Logistic map and Chebyshev map have only finite col- 385

of cxk ∄ {0, 0.25, 0.5, 0.75, 1}, although the value of cxk is lapses in the iterative regions. These finite collapses limit 386
orre
355

356 distributed in the range, [cxlb cxub ] = [01]. Different ini- their Lyapunov exponents to some level. But another one- 387

357 tial values for cxk−1 have been analyzed and it is set as dimensional ICMIC map, which has infinite fixed points in 388

358 cxk−1 = 0.01 finally. Once the initial position has been the iterative region [−1, 0) U (0, +1], has higher Lyapunov 389

359 obtained, the next sequence is generated with the current exponents and stronger chaotic characteristics (He et al. 390

360 position by the map. One approach is to generate the long 2001). Hence, this paper has attempted to embed ICMIC map 391

361 chaos sequence and stock up the sequence during the evo- also to improve the searching ability of the MDRA algorithm. 392

362 lutionary search process initialization. The pointer is made For the ICMIC map, the chaotic behavior can arise from 393

363 to remain in the actually used value in the memory. If the


364 entire sequence is used up, the new one is generated from cxk = sin(µ2 /cxk−1 ). (22) 394
unc

365 the last known one as the new start position. The second way
366 is that the chaotic map is not reinitialized and no long data In this ICMIC map, cxk is distributed in the range 395

367 series is stored, whereas the current state of the map is kept [cxlb cx ub ] = [−11]. The control parameter µ2 = 2. The 396

368 in memory to obtain the next value. This paper applies the initial value of cxk = 0. Figure 4 illustrates the chaotic char- 397

369 former approach for the chaotic sequence generation. Figure acteristics of (22) on the interval, [−1 1] for 1000 iterations at 398

370 3. illustrates the characteristics of quadratic difference equa- µ2 = 2 and initial value of cxk = 0.01. This ICMIC chaotic 399

371 tion of logistic map (21) on the interval, [0 1] for µ1 = 4. sequence is drawn against parent and children chaotic vari- 400

372 (Logistic Map 2001, online document). cxk is computed for ables in 2D in Fig. 4a and 3D in Fig. 4b. Point plot of the 401

373 successive values of k, starting with a small value for cxk−1 ICMIC map for 1000 iterations is shown in Fig. 4c. 402

374 as 0.01. The chaotic sequence so generated in 1000 itera-


375 tions from the dynamic equation of the logistic map is not 3.1.3 Chaos-assisted DE in MDRA 403

376 symmetric as shown in Fig. 3c and this in fact facilitates the


377 chaos search and prevents entrapment of the population in a In this paper, the chaos variables are Fk and CRk , instead 404

378 local optima. This sequence for logistic chaos map is plot- of traditional ‘F’ and ‘CR’. In traditional DE, the ‘F’ value 405

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

(a) 1 (b) (c) 1

0.8 0.8

1
0.6 0.6

0.4 0.5 0.4

of
0.2 0.2

cx(k+1)
0
cx(k)

cx(k)
0 0

-0.2 -0.5 -0.2

pro
-0.4 -0.4
-1
1
-0.6 -0.6
1
-0.8 0.5 -0.8
Author Proof

0
0
-1 -0.5 -1
-1 -0.5 0 0.5 1 cx(k) -1 -1 cx(k-1) 0 200 400 600 800 1000
cx(k-1) Iteration (k)

Fig. 4 Diagram of ICMIC map

406

407

408

409
can vary in the range [F lb F ub ], which is [0 2] and ‘CR’
is suggested in the interval [CRlb CRub ] = [01]. In MDRA
(Zhang et al. 2009), DE operators are fixed as F = 0.5 and
CR = 1. In DE strategy, reproduction is performed with
cted u(j) =

D(j), if rcc(j) < CRk
r1 (j), otherwise
; for j = 1, 2, . . . , n (28)

where, CRk =µ1 .CRk−1 (1−CRk−1 ) for logistic mapping


(29)
426

427

428

µ2
410 D = r1 + F.(r2 − r3 ). (23) CRk = sin when ICMIC map is used. (30) 429
CRk−1

411 Crossover in DE strategy is done using


3.1.4 Chaos search implementation 430
orre

D(j), if rcc(j) < CR
412 u(j) = ; for j = 1, 2, . . . , n, (24) a. Adopt one chaos map. 431
r1 (j), otherwise
b. Assume a random initial value for the chaos variable 432

413 where, ‘rcc’ is a uniform random vector of size 1 × n. cxk−1 for k = 1, in the specified range of the map, 433

[cxlb cxub ]. 434

c. Calculate the chaotic variable, cxk , for the subsequent 435


414 • Chaos-assisted DE mutation operator generations/iterations using (21) or (22) and get the 436

chaotic iterative sequence, varying in the range of the 437

415 In this paper, chaos-assisted mutation operator is introduced respective map. 438

416 by ‘Fk ’, for controlling ‘F’ in each generation ‘k’. So, the d. Perform inverse variable mapping to convert the chaos 439
unc

417 equation (23) is replaced with variable range to the required range of variables, F and 440

CR. For each generation ‘k’, find Fk and CRk from its 441

418 D = r1 + Fk (r2 − r3 ) (25) corresponding cxk , using 442

419 where, Fk = µ1 .Fk−1 (1 − Fk−1 ) for logistic mapping


(26) (cxk − cxlb )
∗ (F ub − F lb ) + F lb ,
420
Fk = (31) 443
(cxub − cxlb )

µ2
421 Fk = sin , for ICMIC chaos map. (27)
C Rk−1
444

422 • Chaos-assisted DE crossover operator (cxk − cxlb )


CRk = ∗ (CRub − CRlb ) + CRlb . (32) 445
(cxub − cxlb )
423 Similarly, chaos-assisted crossover operator is used for con-
424 trolling ‘CR’ in each generation ‘k’, by replacing the Eq. (24) e. Apply the values of Fk and C Rk in (25) and (28) for 446

425 by generating new solutions. 447

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

448 3.2 Updating NPS pool; (iii) pull out all the solutions in the first non-dominated 475

Pareto front to form a combined Pareto set (CPS); (iv) the 476

449 A common procedure is followed for all the MOEAs, in number of solutions in CPS is identified by CPS TOT ; (v) 477

450 updating New Pareto Set (NPS) to weigh against MOEAs calculate the number of solutions contributed by a particular 478

451 implemented for solving TD MOP. For each algorithm, 20 MOEA (CPSMOEA ) to CPSTOT . The advantage of this pro- 479

independent runs are carried out. Pareto pool is formed by cedure is that any number of algorithms can be compared

of
452 480

453 combining Pareto solutions of each run. After 20 runs, non- simultaneously. However at this point, only two algorithms 481

454 dominated sorting (Deb et al. 2002) is performed on the are combined like C-metric measure for comparing the infer- 482

455 Pareto pool. NPS is formed by filtering all the solutions in ence of two measures conveniently. 483

456 the first non-dominated Pareto front.

pro
457 3.3 C-metric evaluation
4 Results and a comparative study 484

This set coverage metric (C-metric), (Zitzler 1999, Ph.D.


Author Proof

458
According to ‘no free lunch theorem’, no algorithm can 485
459 thesis) investigates mainly which MOEA is well suited for
solve all the types of problems. Five different MOEAs such 486
460 the MOP by measuring the level of dominance of a Pareto
as MDRA, ICMIC Chaos-assisted MOEA/D-DRA algo- 487
461 front over another, in the interval [0 1]. Let A and B be two
rithm (ICMDRA), Logistic Chaos-assisted MOEA/D-DRA 488
462 Pareto sets of MOP. C(A, B) is defined as percentage of the
algorithm (LCMDRA), NSGA II, and KnEA have been 489
463 solutions in B that are dominated by at least one solution in
successfully applied with the parameter settings as shown 490
A. C(A, B) = 1
464

465

466
C(A, B) =
|{b ∈ B| ∃ a ∈ A:a ≥ b}|
|B|
cted
indicates that all solutions in B are dominated by some solu-
(33)
in Table 1 and a comparative study has been carried out
between them, with the aim of identifying which MOEA
best suits the solving of the multiobjective TDOP. These
algorithms as such do not have sophisticated constraint-
handling techniques. To solve the complex constrained
491

492

493

494

495

467 tions in A and C(A, B) = 0 implies that none of the solutions TDOP, constrained-domination principle (CDP) (Deb et al. 496

468 in B are dominated by any solution in A. Since C-metric is 2002) has been incorporated to handle constraints, during 497

469 asymmetric, both C(A, B) and C(B, A) are examined. neighborhood sharing of replacement operation in MDRA, 498

ICMDRA, and LCMDRA algorithms. For NSGA II and 499

470 3.4 Quality measure with CPS KnEA algorithms, infeasible solutions are handled at the 500
orre

stage of ranking in non-domination sorting procedure. Tests 501

471 Quality of the non-dominated solutions produced by vari- are carried out on AMD Phenom-II N640 dual core processor 502

472 ous MOEAs is examined using quality measure also (Abido Laptop operating at 2.9 GHz, with 2GB RAM, using MAT- 503

473 2006). The procedure is: (i) combine NPS of two MOEAs LAB 7.4. For each TD bi-objective, 20 independent runs are 504

474 to form a pool; (ii) perform non-dominated sorting on the performed with each algorithm. For all the algorithms, stop- 505

Table 1 Parameter settings of


five MOEAs MDRA, ICMDRA, LCMDRA N T nr  ηm pm
100 0.1 N 0.01 N 0.9 20 1/n
unc

KnEA N ηc ηm pm Tk
200 20 20 1/n 0.6
MDRA F CR
0.5 1
LCMDRA µ1 cxlb cxub Initial CRk Initial Fk
4 0 1 0.48 0.48
ICMDRA µ2 cxlb cxub Initial CRk Initial Fk
2 −1 1 −0.48 −0.2
ICMDRA, LCMDRA F lb F ub CRlb CRub
0 2 0 1
NSGA II N ηc ηm pm
100 5 20 1/n

ηc and ηm distribution index for crossover and mutation, T k rate of knee points

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

Table 2 Comparison of NPS


ND count ICMDRA LCMDRA NSGA II MDRA KnEA
TOT of various MOEAs
Purchase cost vs total loss bi-objective
NPSTOT 1234 1041 734 524 3763
NPSUR 118 119 139 111 130
Purchase cost vs TLTC bi-objective

of
NPSTOT 310 196 181 560 1373
NPSUR 32 37 32 102 27
Total weight vs total loss bi-objective
NPSTOT 1933 778 461 908 3911

pro
NPSUR 36 35 38 37 11
Total weight vs TLTC bi-objective
NPSTOT 1950 2000 148 674 4000
Author Proof

NPSUR 25 24 20 25 10

506 ping criterion is set based on max_gen, which is fixed as 1000 and 243 non-dominated solutions when combining ICM- 538

507 in every run. For MOP, multiple criteria decision making is DRA with (i) LCMDRA, (ii) NSGA II, (iii) MDRA, and 539

508 necessary. This paper utilizes fuzzy-based set theory (Abido KnEA, respectively. ICMDRA contributes the majority of 540

509

510

511

512

513
2006) to extract the best compromise TD over the Pareto
front of NPS. This method uses membership function and
extracts a compromise solution that has maximum member-
ship value. Table 2 gives the non-domination (ND) count of
NPS obtained by each MOEA for all the four bi-objective
cted 51 and 67 %, (CPSMOEA ) solutions to CPSTOT , when paired
with LCMDRA and MDRA, respectively, whereas slightly
lesser solutions of 48 % each to CPS TOT when paired with
NSGA II and KnEA, respectively. So, it necessitates the study
of MOEAs in Utopia point matching with SOO result from
541

542

543

544

545

514 cases. NPS TOT gives the count of number of non-dominated Table 4. The minimum solution of CMA-ES and Utopia point 546

515 solutions present in the NPS. NPSUR is the count of filtered (z1,min , z3,min ) obtained by LCMDRA, NSGA II, KnEA, and 547

516 unrepeated non-dominated solutions of NPS TOT . MDRA algorithms are identical for objective f 3 ; but there 548

was mismatch for f 1 , and thus could not prove their diversity 549

4.1 Behavior of MOEAs in multiobjective TDOP characteristics. However, Utopia point of ICMDRA algo- 550
orre
517
rithm gets matched with that of SOO result for objectives, 551

518 Pareto fronts of five different MOEAs have been plotted f 1 as well as f 3 . This z3,min mismatch of NSGA II appears 552

519 with their respective NPS UR solutions, for all the four bi- to be a very slight deviation from the actual one. But, more 553

520 objectives. All the fronts overlap each other and it is difficult significant non-dominated solutions are found missing in the 554

521 to judge the potential of solutions graphically by inspection mid of its Pareto as in Fig 6b, in comparison to that of ICM- 555

522 alone. Hence, (i) comparison of Utopia point of MOEA and DRA given in Fig. 6a, although NSGA II contributes more 556

523 minimum solution obtained from SOO is required to prove of 52 % solutions to CPSTOT . Similarly, even if there is a 557

524 the diversity characteristics of MOEA; and (ii) two perfor- more share of 130 NPSUR solutions and 52 % CPSMOEA 558

525 mance indicators such as quality and C-metric measures are to CPSTOT , by KnEA. Fig 7 shows that there is a Utopia 559
unc

526 also needed to confirm the superiority of MOEA in terms mismatch and missing of more solutions with KnEA, in 560

527 of convergence. Pareto fronts of four MOEAs are paired comparison with ICMDRA. Furthermore, for concluding, C- 561

528 with the proposed ICMDRA algorithm and their performance metric measure proves that ICMDRA dominates 4.2, 7.9, 48, 562

529 measures such as quality measure and C-metric measure are and 3 % NPSUR solutions of LCMDRA, NSGA II, MDRA, 563

530 noted in Table 3. and KnEA, respectively. Moreover, C (MDRA, ICMDRA) 564

is zero, by which it is seen that non-dominated solutions of 565

MDRA do not cover any of the ICMDRA solutions. Also, 566


531 4.1.1 Purchase cost vs total loss bi-objective
NSGA II, KnEA, and LCMDRA have covered only 0.85 % 567

solutions of ICMDRA, which is a relatively very low percent- 568


532 In this case study, the purchase cost ( f 1 ) and total loss ( f 3 )
age when compared to the measure in the reverse direction. 569
533 are the competing objectives. Figure 5 compares the Pareto
The conclusion that can be drawn from this C-metric mea- 570
534 fronts of different MOEAs. Table 4 gives (i) the best com-
sure, quality measure, and Utopia point matching with the 571
535 promise solution of the trade-off curve, (ii) the results of the
SOO result is that ICMDRA is efficient for solving this TD 572
536 SOO of objectives f 1 and f 3 , and (iii) Utopia point obtained
bi-objective. 573
537 by each MOEA. CPSTOT gives a total of 231, 245, 175,

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Author Proof

123
Table 3 Comparison of quality measure and C-metric measure of various MOEAs

Journal: 500 MS: 2145


TD bi-objective ND count/metrics C(ICMDRAA , LCMDRAB ) C(ICMDRAA , NSGA IIC ) C(ICMDRAA , MDRAD ) C(ICMDRAA , KnEAE )
C(LCMDRAB , ICMDRAA ) C(NSGA IIC , ICMDRAA ) C(MDRAD , ICMDRAA ) C(KnEAE , ICMDRAA )

C(A, B) C(B, A) C(A, C) C(C, A) C(A, D) C(D, A) C(A, E) C(E, A)

TYPESET
Purchase cost vs total loss NPSUR 118A 119B 118A 139C 118A 111D 118A 130E
CPSTOT 231 245 175 243
CPSMOEA 117A 114B 117A 128C 118A 57D 117A 126E

DISK
CPSMOEA (%) 51 49 48 52 67 33 48 52
C-metric 0.042 0.0085 0.0791 0.0085 0.4865 0 0.0308 0.0085

LE
Purchase cost vs TLTC 32 37 32 32 32 102 32 27
unc
NPSUR
CPSTOT 47 62 33 57
CPSMOEA 32A 15B 32A 30C 32A 1D 32A 25E
CPSMOEA (%) 68 32 52 48 97 3 56 44
C-metric 0.5946 0 0.0625 0 0.9902 0 0.0741 0
orre
Total weight vs total loss NPSUR 36 35 36 38 36 37 36 11
CPSTOT 71 74 72 47
CPSMOEA 36A 35B 36A 38C 36A 36D 36A 11E
CPSMOEA (%) 51 49 49 51 50 50 77 23
cted
C-metric 0 0 0 0 0.027 0 0 0
Total weight vs TLTC NPSUR 25 24 25 20 25 25 25 10
CPSTOT 49 31 44 35
25A 24B 25A 6C 25A 19D 25A 10E

CP Disp.:2016/4/5 Pages: 18 Layout: Large


CPSMOEA
pro
CPSMOEA (%) 51 49 81 19 57 43 71 29
C-metric 0 0 0.7 0 0.24 0 0 0
of
S. Tamilselvi et al.
Chaos-assisted multiobjective evolutionary algorithm...

4
1.8
x 10 the winding material is optimized as copper, core material as 579

ICMDRA 23ZH90, and step size as ‘11’. Such transformer is an expen- 580
LCMDRA sive one with high CTM of 15918 $, which is 3853 $ more 581
1.7 NSGA II
MDRA than the one, designed out of SOO of f 1 alone. However such 582

KnEA cheap transformer, which is only 12065 $ CTM , due to the 583
Purchase cost ($)

1.6
selection of aluminium conductor, has higher Twt , Tlos , CTLT ,

of
584

PNLL , and PLL than that of compromise TD. Thus, MOP for- 585
1.5 mulation is effective since their TD results are superior over 586

SOO, in terms of less loss and cost-effectiveness, rather than 587

1.4 single TD, which can satisfy only one customer. A 3D finite 588

element model built is shown in Fig. 8 for the TD obtained

pro
589

1.3 from ICMDRA algorithm for this bi-objective. No load loss 590

of the FEM model is estimated to be 1802 W and it is cal- 591

culated from the proposed methodology involving analytical


Author Proof

1.2 592
1.22 1.24 1.26 1.28 1.3 1.32 1.34 1.36 1.38 1.4 1.42
Total loss (W)
4 equation as 1833 W. The difference between FEM estimated 593
x 10
no load loss and calculated no load loss is −1.69 %, which 594

Fig. 5 Comparison of NPS of five different MOEAs for purchase cost is well within the tolerable limit of 6 % error. Hence, the 595

vs total loss bi-objective proposed methodology can be utilized for designing trans- 596

formers. 597

3
574

575

576

577

578
To discuss the benefits of MOP formulation, consider
two different TDs such as best compromise design of NPS,
obtained by ICMDRA and the design obtained from the min-
imization of single objectives f 1 and f 3 individually. In Table
4, it is seen from the compromise solution of ICMDRA that
cted 4.1.2 Purchase cost vs TLTC bi-objective

In this instance, conflicting bi-objectives, purchase cost ( f 1 ),


and TLTC ( f 4 ) of the transformer are chosen for simulta-
598

599

600

Table 4 Best compromise TD


Variables CMA-ES Best compromise TD for f 1 vs f 3
for purchase cost vs total loss
bi-objective Min f 1 Min f 3 ICMDRA LCMDRA NSGA II KnEA MDRA
orre

x1 (V) 18.15 18.17 18.15 18.15 18.15 18.17 18.15


x2 (mm) 719 613 613 613 613 613 613
x3 (T) 1.8 1.8 1.80 1.80 1.80 1.8 1.80
x4 (A/mm2 ) 1.54 2.5 2.50 2.50 2.50 2.5 2.50
x5 (A/mm2 ) 1.6 2.5 2.53 2.53 2.53 2.5 2.53
x6 8 6 7 7 7 7 8
x7 2 1 1 1 1 1 1
x8 2 4 4 4 4 4 4
PNLL (W) 2302 1682 1833 1833 1833 1836 2090
unc

PLL (W) 11,575 10,674 10,687 10,687 10,687 10,674 10,687


Uk (%) 4.951 4.875 4.853 4.853 4.853 4.875 4.85
Im , D (%) 0.68 0.7337 0.33 0.33 0.33 0.33 0.61
η (%) 98.85 98.98 98.96 98.96 98.96 98.96 98.94
GraLV,D (◦ C) 17.03 19.03 19.10 19.10 19.10 19.03 19.10
GraHV,D (◦ C) 13.21 15.59 15.63 15.63 15.63 15.59 15.63
CTM ($) 12,064.95 17,972.55 15,918 15,918 15,918 15,961 14,990
Tlos (W) 13,877 12,355 12,520 12,520 12,520 12,510 12,777
Twt (kg) 3429.15 3341.899 3332 3332 3332 3341 3332
CTLT ($) 82,580.82 82,353.6 80,698 80,698 80,698 80,732 81,468
Utopia point
z 3,min 12,355 12,355 12,355 12,355 12,355 12,355
z 1,min 12,065 12,065 12,069 12,066 12,103 13,958

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

4 X: 1.236e+004 4 X: 1.236e+004

(a) x 10 Y: 1.798e+004
(b) x 10 Y: 1.798e+004
1.8 1.8
ICMDRA NSGA II

1.7 1.7

Purchase Cost ($)


1.6
Purchase cost ($)

1.6

of
1.5
1.5 Missing
Solutions

Additional Non 1.4


dominated
1.4 solutions of Utopia mismatch

pro
ICMDRA

1.3
1.3
X: 1.381e+004
Y: 1.207e+004
X: 1.388e+004
Author Proof

Y: 1.206e+004
1.2
1.22 1.24 1.26 1.28 1.3 1.32 1.34 1.36 1.38 1.4
1.2 4
1.22 1.24 1.26 1.28 1.3 1.32 1.34 1.36 1.38 1.4 Total Loss (W) x 10
4
Total Loss (W) x 10

Fig. 6 a Pareto front of ICMDRA for f 1 vs f 3 . b Pareto front of NSGA II for f 1 vs f 3


4 4
1.8 x 10 x 10
9

1.7
ICMDRA
KnEA

cted 8.9

8.8
ICMDRA
LCMDRA
NSGA II
MDRA
Purchase cost ($)

1.6 8.7 KnEA

8.6
TLTC ($)

1.5
Missing
Solutions of
8.5
KnEA
1.4 8.4

Utopia
8.3
1.3 Mismatch
8.2
orre

1.2 8.1
1.22 1.24 1.26 1.28 1.3 1.32 1.34 1.36 1.38 1.4
4
Total Loss (W) x 10 8
1.2 1.25 1.3 1.35 1.4 1.45 1.5 1.55 1.6 1.65 1.7
4
Fig. 7 Comparison of Pareto fronts of KnEA and ICMDRA for f 1 vs Purchase cost ($) x 10
f3
Fig. 9 Comparison of NPS of five different MOEAs for purchase cost
vs TLTC bi-objective

neous optimization by MOEAs. Figure 9 gives the Pareto 601

front of every MOEA, obtained when attempted to mini- 602


unc

mize this bi-objective. Table 5 provides the best compromise 603

TD solution among NPS UR non-dominated solutions for 604

all MOEAs. By observing the CMA-ES results, it is found 605

that ICMDRA and LCMDRA have provided the minimum 606

Utopia points (z1,min , z4,min ) for both objectives f 1 and f 4 , 607

whereas, NSGA II, KnEA, and MDRA could not provide for 608

z4,min and MDRA failed for z1,min too. Although NPSUR , of 609

ICMDRA is lesser than LCMDRA and MDRA as depicted 610

in Table 2, 68, 52, 97 and 56 % of CPSTOT is contributed 611

only by ICMDRA, (CPSMOEA ), when paired with LCM- 612

DRA, NSGA II, KnEA, and MDRA respectively, as in Table 613

3. Figure 9 may give the impression that performance of 614

Fig. 8 3D FEM model NSGA-II is as similar as ICMDRA. But, Fig. 10b clarifies 615

that, in the initial portion and end region of NSGA II Pareto, 616

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

Table 5 Best compromise TD for purchase cost vs TLTC bi-objective


dv/ pv/ov CMA-ES Best compromise TD for f 1 vs f 4

Min f 1 Min f 4 ICMDRA LCMDRA NSGA II KnEA MDRA

x1 (V) 18.15 18.17 18.15 18.15 18.15 18.15 18.15

of
x2 (mm) 719 613 728 725 728 725 610
x3 (T) 1.8 1.69 1.72 1.80 1.72 1.8 1.80
x4 (A/mm2 ) 1.54 2.5 1.5000 1.50 1.5 1.54 2.51
x5 (A/mm2 ) 1.6 2.51 1.5000 1.54 1.5 1.5 2.62
x6 8 7 7 7 7 8 8

pro
x7 2 1 2 2 2 2 1
x8 2 4 4 2 4 2 2
PNLL (W) 2302 1577 1847 2033 1847 2317 2067
Author Proof

PLL (W) 11,575 10,933 11,430 11,372 11,430 11,372 10,884


Uk (%) 4.951 4.980 5.01 4.98 5.01 4.98 4.89
Im,D (%) 0.68 0.286 0.32 0.36 0.32 0.68 0.61
η (%) 98.85 98.96 98.9 98.89 98.9 98.87 98.93
GraLV,D (◦ C) 17.03 19.05 16.35 16.57 16.35 16.57 19.57
GraHV,D (◦ C) 13.21 15.59 12.73 12.78 12.73 12.78 15.78
CTM ($)
CTLT ($)
Twt (kg)
Tlos (W)
12,064.95
82,580.82
3429.15
13,877
16,632.00
80,619.98
3486.22
12,510
cted
13,790
80,754
3522
13,277
13,131
81,096
3417
13,405
13,790
80,754
3522
13,277
12,104
81,943
3417
13,689
14,760
81,741
3328
12,951
Utopia point
z 1,min 12,065 12,065 12,065 12,065 12,065 13,877
z 4,min 80,620 80,620 80,620 80,698 80,962 80,635

617 more non-dominated solutions are missing with respect to 4.1.3 Total weight vs total loss bi-objective 641
orre

618 ICMDRA Pareto. Similarly, Fig. 10a shows very obviously


619 that KnEA struggles and could not reach the ICMDRA In this case, total weight ( f 2 ) and total loss ( f 3 ) of the 642

620 Pareto. Moreover, C-metric measure confirms that solutions transformer are chosen as bi-objective. Figure 11 gives the 643

621 of ICMDRA cover LCMDRA, NSGA II, KnEA, and MDRA comparison of the Pareto front of unrepeated new Pareto set 644

622 solutions with a maximum coverage of 59.46, 6.25, 7.41, (NPSUR ) given by every MOEA. Table 6 provides the best 645

623 and 99.02 %, respectively. Also, unlike previous bi-objective compromise TD solution of NPS UR Pareto solutions of all the 646

624 case, C-metric measure of all the other MOEAs with ICM- MOEAs. Unlike the previous two bi-objective cases, except 647

625 DRA, i.e., C(XX, ICMDRA) is zero for this case, by which it KnEA, all the MOEAs including MDRA have provided the 648

626 is evident that solutions obtained by all the MOEAs including same best minimum, since Utopia points (z2,min , z3,min ) are 649
unc

627 LCMDRA do not dominate even one solution of ICMDRA. being matched with the actual, for objectives f 2 as well as 650

628 Again in this bi-objective, ICMDRA has proven its improved f 3 . In this case, all the MOEAs are equally efficient in find- 651

629 searching ability. In Table 5; yet again, the solution of TD ing the true Pareto front and the same compromise solution, 652

630 MOP by ICMDRA has yielded efficient TD, since it has got except KnEA, which failed even for the doubled population 653

631 1826 $ lower TLTC and 600 W lower total loss, than those size, i.e., 200, as seen in Table 1. 654

632 of SOO of f 1 . The best part of this MOP is that ICMDRA Here, C-metric measure being zero in either direction, 655

633 design could derive such efficient transformer even with alu- none of the algorithm is dominated by ICMDRA, except 656

634 minum conductor, rather than the one from MDRA, which MDRA. But for the case of KnEA, it has been misinterpreted 657

635 possesses 987 $ higher TLTC with copper. Even though the that KnEA is on par with ICMDRA, since half the portion 658

636 same fuzzy mechanism is applied for finding the best com- of the entire KnEA Pareto is not committed. But actually, 659

637 promise solution in their NPS for both ICMDRA and MDRA, more significant solutions in the KnEA Pareto are not spotted, 660

638 TLTC is lesser in ICMDRA-based design only. Thus in this when compared with ICMDRA as seen from Fig. 12, which 661

639 bi-objective case also, ICMDRA has proved its searching proves the performance of ICMDRA clearly. 662

640 ability.

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

4
(a) x 10
4 Purchase cost Vs TLTC bi-objective (b) 8.24 x 10 X: 1.207e+004
Y: 8.233e+004
8.26 ICMDRA
ICMDRA NSGA II
8.24 8.22
KnEA

8.22 8.2

8.2 8.18
Missing
solutions of

of
TLTC ($)

TLTC ($)
NSGA II
8.18
8.16
8.16
8.14
8.14
8.12
8.12
Utopia
8.1 Additional

pro
8.1 mismatch
solutions
of NSGA II
of ICMDRA
8.08 8.08
8.06
1.2 1.25 1.3 1.35 1.4 1.45 1.5 1.55 1.6 1.65 8.06
Author Proof

4 1.2 1.25 1.3 1.35 1.4 1.45 1.5 1.55 1.6 1.65
Purchase cost ($) x 10
x 10
4
Purchase cost ($)

Fig. 10 a Comparison of Pareto fronts of KnEA and ICMDRA for f 1 vs f 4 . b Comparison of Pareto fronts of NSGA II and ICMDRA for f 1 vs
f4
3350
ICMDRA
matches with that of ICMDRA, the quality measure says that 675

3340

3330
LCMDRA
NSGA II
MDRA
KnEA
cted NSGA II and MDRA give relatively less share of only 19 %
and 43 %, respectively, to CPS TOT, when paired with ICM-
DRA. Similarly, KnEA also gives lesser solutions of 29 % to
676

677

678
Total Weight (kg)

3320 CPS TOT, when paired with ICMDRA, with different com- 679

3310 promise solution where the TLTC is high comparatively. Set 680

coverage metric measure in Table 3 also confirms the same 681


3300
inference of quality measure that ICMDRA is superior to 682

3290 other MOEAs, since in this case C-metric measure of all the 683

3280 MOEAs with respect to ICMDRA is zero, meaning that all 684

the solutions of other MOEAs could not dominate any sin- 685
orre
3270
gle solution of ICMDRA. NSGA II is not found suitable for 686

3260
1.235 1.24 1.245 1.25 1.255 1.26 1.265 1.27 1.275 1.28 1.285
solving this bi-objective TDOP, since a lot of non-dominated 687

Total loss (W) x 10


4 solutions by NSGA II from Fig. 14b and domination of 70 % 688

of its solutions by ICMDRA from Table 3 are observed to 689

Fig. 11 Comparison of NPS of five different MOEAs for total weight be missing. In addition to the previous bi-objective, LCM- 690
vs total loss bi-objective DRA is on par with ICMDRA in this case also, as none of 691

its solutions are covered by any solution of ICMDRA. Like 692

663 4.1.4 Total weight vs TLTC bi-objective the previous bi-objective case, here also the reverse C-metric 693

measure could not interpret the fact that KnEA is inferior to 694

ICMDRA, since KnEA has not provided Pareto for initial and
unc

695
664 In this case, total weight ( f 2 ) and TLTC ( f 4 ) of the trans-
665 former are the conflicting objectives. Figure 13 gives the end region completely, whereas Fig. 14a rules out the miscon- 696

666 comparison of Pareto fronts of all the five MOEAs. Table ception. It is clear from the investigations that chaos-assisted 697

667 7 shows that the best compromise design among NPS UR MDRA is more suitable for solving this TD bi-objective 698

668 non-dominated solutions of different MOEAs are identical than the other three MOEAs. As far as efficient design is 699

669 designs, except the one given by KnEA. But, it is seen from concerned, MOP formulation of the TDOP has proved its 700

670 the optimization results that ICMDRA and LCMDRA are potential in this case also. On observing the results of SOO 701

671 only able to get their Utopia point (z2,min , z4,min ) same as that of objective f 2 , it is found that the overall weight of the 702

672 of the CMA-ES result of SOO of objectives f 2 , and f 4 . Figure transformer is absolutely less in comparison with the com- 703

673 13 and Table 7 prove that NSGA II, KnEA, and MDRA could promise design. However, TLTC, Tlos , and CTM are more 704

674 not match their Utopia values for both objectives. Moreover, than ICMDRA-based MOP design. 705

even when the compromise solution of NSGA II and MDRA

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

Table 6 Best compromise TD


dv/ pv/ov CMA-ES Best compromise TD for f 2 vs f 3
for total weight vs total loss
bi-objective Min f 2 Min f 3 ICMDRA LCMDRA NSGA II KnEA MDRA

x1 (V) 18.19 18.17 18.15 18.15 18.15 18.15 18.15


x2 (mm) 603 613 607 607 607 613 607

of
x3 (T) 1.8 1.8 1.80 1.80 1.80 1.8 1.80
x4 (A/mm2 ) 2.69 2.5 2.50 2.50 2.50 2.5 2.50
x5 (A/mm2 ) 2.60 2.5 2.69 2.69 2.69 2.53 2.69
x6 6 6 6 6 6 6 6
x7 1 1 1 1 1 1 1

pro
x8 4 4 4 4 4 4 4
PNLL (W) 1665 1682 1669 1669 1669 1679 1669
PLL (W) 11,159 10,674 10,830 10,830 10,830 10,687 10,830
Author Proof

Uk (%) 4.871 4.875 4.85 4.85 4.85 4.85 4.85


Im , D (%) 0.726 0.7337 0.72 0.72 0.72 0.73 0.72
η (%) 98.942 98.98 98.96 98.96 98.96 98.98 98.96
GraLV,D (◦ C) 19.866 19.03 19.98 19.98 19.98 19.10 19.98
GraHV,D (◦ C) 16.99 15.59 15.82 15.82 15.82 15.63 15.82
Twt (W) 3267.16 3341.899 3294 3294 3294 3333 3294
Tlos (W)
CTM ($)
CTLT ($)
12,824
17,599.95
83,631.58
cted 12,355
17,972.55
82,353.6
12,499
17,717
82,509
Utopia point
12,499
17,717
82,509
12,499
17,717
82,509
12,366
17,926
82,315
12,499
17,717
82,509

z 2,min 3267.16 3267 3267 3267 3291 3267


z 3,min 12,355 12,355 12,355 12,355 12,356 12,355

3350 3500
ICMDRA ICMDRA
orre
Utopia mismatch
3340 KnEA of NSGA II NSGA II

3330 3450
Missing solutions of
Total Weight (kg)

3320 KnEA
Total Weight (kg)

3310 3400

3300

3290 3350
Missing solutions
3280 of NSGA II
Utopia mismatch
unc

of NSGA II
3270 3300

3260
1.235 1.24 1.245 1.25 1.255 1.26 1.265 1.27 1.275 1.28 1.285
4
Total loss (W) x 10 3250
8.05 8.1 8.15 8.2 8.25 8.3 8.35 8.4
4
Fig. 12 Comparison of Pareto fronts of KnEA and ICMDRA for f 2 TLTC ($) x 10
vs f 3
Fig. 13 Comparison of NPS of five different MOEAs for total weight
vs TLTC bi-objective

706 5 Conclusion
improving reliability, design, efficiency of the transformer 711

707 This multiobjective problem formulation is a powerful tool and in turn its life expectancy. To satisfy all the needs of 712

708 for decision making as this approach performs to find all various types of customers, manufacturers, utilities, etc, TD 713

709 the trade-off between conflicting TD criteria. Multiobjec- MOP has considered minimization of four sets of competing 714

710 tive formulation for TDOP presented in this paper aims at bi-objectives, taking in to account performance and ther- 715

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

Table 7 Best compromise TD for total weight vs TLTC bi-objective


dv/ pv/ov CMA-ES Best compromise TD for f 2 vs f 4
Min f 2 Min f 4 ICMDRA LCMDRA NSGA II KnEA MDRA

x1 (V) 18.19 18.17 18.15 18.15 18.15 18.15 18.15


x2 (mm) 603 613 607 607 607 613 607

of
x3 (T) 1.8 1.69 1.80 1.80 1.80 1.8 1.80
x4 (A/mm2 ) 2.69 2.50 2.50 2.50 2.50 2.53 2.50
x5 (A/mm2 ) 2.60 2.51 2.69 2.69 2.69 2.5 2.69
x6 6 7 7 7 7 7 7

pro
x7 1 1 1 1 1 1 1
x8 4 4 4 4 4 4 4
PNLL (W) 1665 1577 1822 1822 1822 1833 1822
Author Proof

PLL (W) 11,159 10,933 10,830 10,830 10,830 10,687 10,830


Uk (%) 4.871 4.980 4.85 4.85 4.85 4.85 4.85
Im,D (%) 0.726 0.286 0.32 0.32 0.32 0.33 0.32
η (%) 98.942 98.96 98.95 98.95 98.95 98.96 98.95
GraLV,D (◦ C) 19.866 19.05 19.98 19.98 19.98 19.10 19.98
GraHV,D (◦ C) 16.99 15.59 15.82 15.82 15.82 15.63 15.82
Twt (W)
CTLT ($)
CTM ($)
Tlos (W)
3267.16
83,631.58
17,599.95
12,824
3486.22
80,619.98
16,632.00
12,510
cted
3294
80,901
15,720
12,652
Utopia point
3294
80,901
15,720
12,652
3294
80,901
15,720
12,652
3292
80,984
15,917
12,520
3294
80,901
15,720
12,652

z 2,min 3267.16 3267 3267 3279 3270 3269


z 4,min 80,620 80,620 80,620 80,675 80,698 80,639

(a) 3500 Total Weight Vs TLTC Bi-objective (b) 3500 ICMDRA


orre
Utopia mismatch
ICMDRA of NSGA II NSGA II

KnEA
3450
3450
Total Weight (kg)
Total Weight (kg)

3400 3400

3350 3350
Missing solutions
of NSGA II
Utopia mismatch
of NSGA II

3300 3300
unc

3250 3250
8.05 8.1 8.15 8.2 8.25 8.3 8.35 8.4 8.05 8.1 8.15 8.2 8.25 8.3 8.35 8.4
4 4
TLTC ($) x 10 TLTC ($) x 10

Fig. 14 a Comparison of Pareto fronts of KnEA and ICMDRA for f 2 vs f 4 . b Comparison of Pareto fronts of NSGA II and ICMDRA for f 2 vs
f4
716 mal and practical manufacturing constraints. The selection proposed methodology has been proved to be effective for 723

717 of significant decision variables and enlarged search space designing transformers. Comparative analysis has been done 724

718 has guaranteed efficient and reliable TD. The validity of the against five MOEAs in solving all the four TD bi-objectives. 725

719 proposed methodology is illustrated by its investigation on This paper has proposed chaos-assisted MOEA/D-DE-DRA 726

720 a sample 1500 kVA transformer. The core losses of the TD algorithm for TD MOP by embedding chaos-assisted DE 727

721 obtained using analytical equation show acceptable deviation crossover operator and DE mutation operator in the MDRA 728

722 from its numerical simulation core loss result and hence the algorithm. LCMDRA and ICMDRA have been success- 729

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Chaos-assisted multiobjective evolutionary algorithm...

730 fully applied to TDOP, and the improved searching ability Cheema MAM, Fletcher JE, Dorrell D (2013) A practical approach for 783

731 of chaos-embedded MDRA has been proved in all the bi- the global optimization of electromagnetic design of 3-phase core 784

type distribution transformer allowing for capitalization of losses. 785


732 objective cases. Implementation of ICMIC chaos in MDRA IEEE Trans Magn 49(5):2117–2120 786
733 has provided the best non-dominated solutions which are not Coelho LDS, Mariani VCM (2006) Combining of chaotic differen- 787

734 being covered by any of the solutions of MDRA in all the tial evolution and quadratic programming for economic dispatch 788

four bi-objective cases. However, 48.65, 99.02, 2.7, and 24 % optimization with valve-point effect. IEEE Trans Power Syst 789

of
735
21(2):989–996 790
736 solutions of MDRA are dominated by solutions of ICMDRA Coelho LDS, Mariani VC, Ferreira da Luz MV, Leite JV (2013) Novel 791
737 in the respective four cases. Except for one bi-objective case, gamma differential evolution approach for multiobjective trans- 792

738 purchase cost vs total loss, none of the solutions of NSGA II, former design optimization. IEEE Trans Magn 49(5):2121–2124 793

739 KnEA, and LCMDRA dominate ICMDRA solutions. With Coelho LDS, Mariani VC, Guerra FA, Ferreira da Luz MV, Leite JV 794

(2014) Multiobjective optimization of transformer design using a 795


all the four bi-objective case studies, Utopia comparison of

pro
740
chaotic evolutionary approach. IEEE Trans Magn 50(2):669–672 796
741 MOEAs results with SOO and two performance measures Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and elitist mul- 797

742 such as C-metric measure and quality measure has proved tiobjective genetic algorithm: NSGA-II. IEEE Trans Evol Comput 798

that ICMDRA is more efficient, when compared to all the 6(2):182–197 799
Author Proof

743
Georgilakis PS, Hatziargyriou N, Paparigas D (1999) AI helps reduce 800
744 other four MOEAs with its good convergence and diversity
transformer iron losses. Comput Appl Power IEEE 12(4):41–46 801
745 characteristics in solving multiobjective TDOP. Hence, it is Georgilakis PS, Doulamis N, Doulamis A, Hatziargyriou N, Kollias 802

746 concluded that ICMDRA can solve the MOP formulation of S (2001) A novel iron loss reduction technique for distribution 803

747 the TD problem effectively. transformers based on a combined genetic algorithm-neural net- 804
work approach. IEEE Trans System Man Cybern C 31:16–34 805

Georgilakis PS, Tsili M, Souflaris A (2007a) A heuristic solution to 806


748 Compliance with ethical standards
749

750

751

752
cted
Funding This research work was not funded by any company/agency.

Conflict of interest The authors declare that they have no conflict of


interest.
the transformer manufacturing cost optimization problem. J Mater
Process Technol 181:260–266
Georgilakis PS, Gioulekas A, Souflaris A (2007b) A decision tree
method for the selection of winding material in power transform-
ers. J Mater Process Technol 181:281–285
Georgilakis PS (2009a) Recursive genetic algorithm-finite element
807

808

809

810

811

812

method technique for the solution of transformer manufacturing 813

753 Ethical approval This article does not contain any studies with human cost minimization problem. IET Electr Power Appl 3(6):514–519 814

754 participants or animals performed by any of the authors. Hansen N (2006) The CMA evolution strategy: a comparing review. 815

Towards a new evolutionary computation. Springer, Berlin, pp 75– 816

102 817

He D, He C, Jiang L, Zhu H, Hu G (2001) Chaotic characteristics of a 818

one-dimensional iterative map with infinite collapses. IEEE Trans


orre
819
755 References Circ Syst I: Fundam Theory Appl 48(7) 820 5
Hui L, Li H, Bei H, Shunchang Y (2001) Application research based on 821

756 Abido MA (2006) Multi objective evolutionary algorithms for electric improved genetic algorithm for optimum design of power trans- 822

757 power dispatch problem. IEEE Trans Evol Comput 10(3):315–329 formers. In: Proceedings of the fifth IEEE international conference 823

758 Adly AA, Abd-El-Hafiz SK (2014) A performance-oriented power on electrical machines and systems, ICEMS, vol 1, pp 242–245 824

759 transformer design methodology using multi-objective evolution- Hernandez C, Arjona MA (2007) Design of distribution transformers 825

4 760 ary optimization. J Adv Res based on a knowledge-based system and 2D finite elements. Finite 826

761 Altinoz OT Yilmaz AE, Ha GWW (2010) Chaos Particle swarm Elem Anal Des 43:659–665 827

762 optimized PID controller for the inverted pendulum system. In: Jabr RA (2005) Application of geometric programming to transformer 828

763 Proceedings of second international conference on engineering design. IEEE Trans Magn 41(11):4261–4269 829

764 optimization, Portugal Judd FF, Kressler DR (1977) Design optimization of small low- 830
unc

765 Amoiralis EI, Georgilakis PS, Tsili M (2008a) Design optimization frequency power transformers. IEEE Trans Magn 13(4):1058– 831

766 of distribution transformers based on mixed integer programming 1069 832

767 methodology. J Optoelectron Adv Mater 10(5):1178–1183 Li H, Zhang Q (2009) Multiobjective optimization problems with com- 833

768 Amoiralis EI, Tsili M, Georgilakis PS, Kladas A, Souflaris A (2008b) A plicated Pareto sets, MOEA/D and NSGA-II. IEEE Trans Evol 834

769 parallel mixed integer programming-finite element method tech- Comput 12(2):284–302 835

770 nique for global design optimization of power transformers. IEEE Logistic Map (2001) Online document, available from http://en. 836

771 Trans Magn 44(6):1022–1025 wikipedia.org/wiki/Logistic_map 837

772 Amoiralis EI, Georgilakis PS, Tsili M, Kladas A (2009) Global Masood MA, Jabbar RA, Masoum MAS, Junaid M, Ammar M (2012) 838

773 transformer optimization method using evolutionary design and An innovative technique for design optimization of Core type 3- 839

774 numerical field computation. IEEE Trans Magn 45(3):1720–1723 phase distribution transformer using Mathematica. Glob J Technol 840

775 Amoiralis EI Tsili M, Kladas A (2012) Global transformer design opti- Optim 3:2012 841

776 mization using deterministic and nondeterministic algorithms. In: Olivares-Galván JC, de León F, Georgilakis P, Escarela-Pérez R (2010) 842

777 Proceedings of the twentieth International Conference on Electri- Selection of copper against aluminium windings for distribution 843

778 cal Machines (ICEM2012), France, pp 2012 transformers. Inst Eng Technol Electr Power Appl 4(6):474–485 844

779 Chen Z, Yuan X, Ji B, Wang P, Tian H (2014) Design of a fractional Padma S, Bhuvaneswari R, Subramanian S (2006) Optimal design of 845

780 order PID controller for hydraulic turbine regulating system using power transformer using simulated annealing technique. In: Pro- 846

781 chaotic non-dominated sorting genetic algorithm II. Energy Con- ceedings of IEEE conference on industrial technology (ICIT), 847

782 vers Manage 84:390–404 India, pp 1015–1019 848

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
S. Tamilselvi et al.

849 Pham TH, Salon SJ, Hoole SRH (1996) Shape optimization of windings Williams SB, Abetti PA, Magnesson EF (1955) How digital computers 867

850 for minimum losses. IEEE Trans Magn 32(5):4287–4289 aid transformer designer. Gen Electr Rev 58:24–25 868

851 Promotion Centre and European Copper Institute Wu CJ, Lee FC (1980) Minimum weight EI core and pot core inductor 869

852 Subramanian S, Padma S (2011a) Optimization of transformer design and transformer designs. IEEE Trans Magn 16(5):755–757 870

853 using bacterial foraging algorithm. Int J Comput Appl 19(3):52–57 Zhang X, Tian Y, Jin Y (2015) A knee point driven evolutionary algo- 871

854 Subramanian S, Padma S (2011b) Bacterial foraging algorithm based rithm for many-objective optimization. IEEE Trans Evol Comput. 872

855 multi objective optimal design of single phase transformer. J Com- Accepted for publication 873

of
856 put Sci Eng 6(2):1–6 Zhang Q, Li H (2007) MOEA/D: a multi-objective evolutionary 874

857 Tamilselvi S, Baskar S (2014) Modified parameter optimization of dis- algorithm based on decomposition. IEEE Trans Evol Comput 875

858 tribution transformer design using covariance matrix adaptation 11(6):712–731 876

859 evolution strategy. Int J Electr Power Energy Syst 61:208–218 Zhang Q, Liu W, Li H (2009) The performance of a new version of 877

860 Tsivgouli AJ, Tsili MA, Kladas A, Georgilakis P, Souflaris A, Skarla- MOEA/D on CEC09 Unconstrained MOP Test Instances. CEC 878

861 tini A (2007) Geometry optimization of electric shielding in power 2009 879

pro
862 transformers based on finite element method. J Mater Process Zitzler E (1999) Online Document, Evolutionary algorithms for multi- 880

863 Technol 181:159–164 objective optimization: methods and applications. Ph.D. Thesis, 881

864 Tsili M, Kladas A, Georgilakis P, Souflaris A, Paparigas D (2005) Swiss Federal Institute of Technology (ETH) (Dissertation ETH 882

865 Numerical techniques for design and modelling of distribution No. 13398) 883
Author Proof

866 transformers. J Mater Process Technol 161:320–326

cted
orre
unc

123
Journal: 500 MS: 2145 TYPESET DISK LE CP Disp.:2016/4/5 Pages: 18 Layout: Large
Journal: 500
Article: 2145

Author Query Form

of
Please ensure you fill out your response to the queries raised below
and return this form along with your corrections

Dear Author

pro
During the process of typesetting your article, the following queries have arisen. Please check your typeset proof carefully
against the queries listed below and mark the necessary changes either directly on the proof/online grid or in the ‘Author’s
Author Proof

response’ area provided below

Query Details required Author’s response


1. Please check the edit to the article title.
2. As reference ‘Amoirals et al. 2007’ is cited in text, but

3.
cted
given in the reference list. Please provide in the refer-
ence list or delete this citation.
Please provide a definition for the significance of [bold]
in the tables 4, 5, 6 and 7.
4. Please provide volume id, page range for the reference
Adly and Abd-El-Hafiz (2014).
5. Please provide page range for the reference He et al.
(2001).
orre
unc

Anda mungkin juga menyukai