Você está na página 1de 6

# Proceedings of the 2006 IEEE WeC04.

4
International Symposium on Intelligent Control
Munich, Germany, October 4-6, 2006

## Optimization of THD and Suppressing Certain Order Harmonics in

PWM inverters using Genetic Algorithms
A. Sayyah, Student Member, IEEE, M. Aflaki, and A.R.Rezazade

Abstract In this paper the aim is to minimize the total These transcendental equations are solved using iterative
harmonic distortion (THD) in PWM inverters while numerical techniques, where the initially provided
suppressing chosen harmonics concurrently and maintaining approximate solutions must be sufficiently close to the exact
the fundamental component of the output voltage at a required solutions to ensure the convergence of the solutions, and this
level. This is reformed as an optimization task and the optimal is the most difficult task associated with these methods .
pulse patterns are accomplished using genetic algorithm (GA) Some attempts have been made to simplify harmonic
optimization technique to minimize a predefined fitness
elimination equations by applying Walsh functions -.
function. The complete solutions that suppress the 5th and 7th
harmonics and optimize the THD are given. A practical It was believed that the system of transcendental nonlinear
limitation on the determination of sequential switching angles harmonic elimination equations resulting from Fourier
has been considered as a set of constraints in derivation of the analysis of PWM waveform could be replaced by a system
results. of linear algebraic equations when Walsh analysis is
applied. However, it has been illustrated that Walsh
I. INTRODUCTION functions approach is neither theoretically nor practically

## A MONG all power converters used in industry, the pulse-

applicable . In some other studies , the complete
solutions to the harmonic elimination problem have been
width modulated(PWM) have the most widespread
presented using the theory of resultants from elimination
applications such as: uninterruptible power supply (UPS),
theory , . These expressions are difficult and time
phase controlled rectifier, and adjustable speed servo drives.
consuming to derive. An increase in the number of
The main features that have made them candidates in many switching angles can exacerbate the problem due to its call
industrial applications are: their capabilities of controlling for exorbitant calculations. This paper has a three-fold
the output voltage and frequency simultaneously, and objective. First, minimization of the THD in the case of
generating output waveform with low harmonic distortion. three switching angles, second, elimination of the 5th and 7th
The output voltage of an ideal inverter is sinusoidal. harmonics and presenting complete solutions as in  but
However, the waveforms of practical inverters are non- this time considering a practical limitation on choosing the
sinusoidal and contain certain harmonics. In low- and sequential switching angles, that makes the utilization of
medium-power applications, square-wave or quasi-square- switching instants possible in low modulation indices, and
wave voltages may be acceptable, but in high-power third, optimization of an objective function formed by
applications, sinusoidal waveforms are desired to be less combining the first and second terms discussed above. All
distorted . In this case, direct optimization of the the steps enumerated above, have been carried out using
waveform based on specification of an optimal value for GA. Genetic Algorithm (GA) is a randomized parallel
each switching instant seems indispensable . search method modeled on natural selection. GA can be
This optimization can be achieved via various approaches. applied to a variety of problems and is becoming a
Elimination of the lower frequency dominant harmonics and significant tool in machine learning and function
Minimization of the total harmonic distortion are the two optimization -.
widely-used methods in choosing the switching instants. The
more studied of the two techniques is the harmonic II. THREE-PHASE INVERTERS AND PWM WAVEFORM
elimination method in which the switching angles are chosen Three-phase inverters are normally used for high-power
(programmed) so as to eliminate specific harmonics. These applications. Three single-phase half (or full)-bridge
switching patterns can be determined by solving some inverters can be connected in parallel as shown in Fig. 1 to
systems of nonlinear and transcendental equations -. form the configuration of three-phase inverters.

## A. Sayyah is Senior in Electrical Engineering in Shahid Beheshti Q1 D1 Q3 D3 Q5 D5

University, Tehran, 19839-63113 (phone: +98-21-44080421; e-mail: VS
arash.sayyah@gmail.com).
a
M. Aflaki, has graduated from Shahid Beheshti University in Electrical b
Engineering, Tehran, 19839-63113. She is now doing her graduate studies c
in Sharif University of Technology, Tehran, 11365-8639 Iran (e-mail: Q4 D4 Q6 D6 Q2 D2
mitra.aflaki@gmail.com). VS
A. R. Rezazade is the faculty of Electrical and Computer Engineering
Department, Shahid Beheshti University , Tehran ,19839-63113 (e-mail:
alireza.rezazade@gmail.com). Fig. 1. Three-phase voltage source inverter (GTOs as power switches).

## 0-7803-9798-3/06/\$20.00 2006 IEEE 874

In this study, the switching angles B1, B2 and B3 shown functions. Each switching angle is called a gene, which can
in Fig. 2, are chosen in a way to minimize certain defined be coded as a binary or floating-point string of length. In this
functions. study, a binary string is used. A chromosome consists of all
f (R) B1 genes.
1 Some brief comparisons between the performance of GA
and other approaches applied to harmonic elimination
Q Q 3Q 2Q Xt problem (for instance) and also advantages of GA over
2 2 conventional optimization methods are considered to be:
1 1) An initial estimate of the solution is not required. This is
B2 B3 in contrast with iterative numerical techniques where
Fig. 2. A symmetrical two-level normalized PWM waveform. the initially provided approximate solution has a pivotal
Quarter-wave symmetry (i.e. f ( Q2 R)  f ( Q2  R)) and role for the convergence of the algorithm. On the other
half-wave inverse symmetry (i.e. f (Q R)  f (Q  R)) hand, due to the stochastic nature of the search
considered in the Fig. 2, ensure the elimination of even- mechanism in GA's, they have the capability to explore
order harmonics in the output of the inverter. the entire search space with more likelihood of finding
One important point to be considered in selection of all solutions (as in the harmonic elimination problem
switching instances is that there must be a minimum delay where in the case of 3 switching angles a totally new,
between the moment a thyristor goes out and the moment the and not previously reported , , , set of solutions
next thyristor fires. Otherwise, the two thyristors would was found).
form a short-circuit. To avoid this condition, we assume a 2) Computational complexity is reduced in this approach
minimum difference of 100Ns between the two adjacent in comparison with other applied approaches. Derived
switching instances. The corresponding angle would be analytical expressions that use the mathematical
0.0314 rad (the fundamental frequency is assumed 50 Hz ). Resultants theory to compute the optimum switching
Hence in 3-switch case we have: angles for the harmonic elimination problem can be
(i) B1  R (ii) B2  B1  R considered for comparison . The expressions are
(iii) B3  B2  R (iv) Q/2  B3  R/2 (1) polynomials of 22nd degree, which are difficult and
Satisfaction of these constraints turns optimization, from a time-consuming to derive .
pure mathematical procedure into a more practical one. The 3) Probability of getting trapped in locally optimal
calculated switching angles are stored in a read-only solutions is significantly decreased. Industrial
memory (ROM) and are served as a look-up table. During engineering problems usually contain nonlinearities and
real-time operation, the required fundamental amplitude is complex interactions among problem variables.
used for addressing the corresponding switching angles in Application of classical point-to-point algorithms (direct
the look-up table which are read out for controlling the or gradient-based) may result in suboptimal solutions.
inverter. To avoid the use of large memory components and The possibility of this occurrence is greatly reduced due
to improve the flexibility and maintenance of the system, a to the random process blended with GA.
real-time generation of switching instants can be performed 4) Particularly apt for addressing constrained optimization
from simple functions that approximate off-line calculated problems. A vast majority of industrial engineering
solution trajectories . The accuracy of the on-line optimization problems are constrained problems. The
generated switching instants has a great reliance on the presence of constraints significantly affects the
proximity of functions and could be increased, at least performance of every optimization algorithm .
theoretically, to any extent by using high-order polynomials. Classic linear programming and nonlinear programming
methods are often either unsuitable or empirical when
III. GENETIC ALGORITHM applied to some constrained problems . Difficulties
arise because either the amount of computation required
Genetic algorithms are defined as stochastic optimization
quickly becomes unmanageable as the size of the
techniques relying on the genetic natural evolution
problem increases, or the constraints violate the
mechanisms of creative beings ,  .These algorithms required assumptions, such as differentiability or
encode a potential solution to a specific problem on a simple convexity.
chromosome-like data structure and apply recombination Several methods have been developed to handle
operators to these structures so as to preserve critical constraints in evolutionary computation techniques,
information. Although genetic algorithms are often viewed which can be classified into categories such as: methods
as function optimizers, the range of problems to which based on penalty functions (which is the most common
genetic algorithms can be applied is quite broad. approach in genetic algorithms community) , ,
In the problem discussed in this paper, the solution is to the assumption of superiority of feasible solutions over
choose 3 switching instants (angles) in an electrical cycle infeasible ones , and repair algorithms .
(Fig. 2) for turning the switches on and off in a full bridge In this study, we have used the method of rejection of
inverter (Fig. 1) so as to produce a desired fundamental infeasible individuals. This is a popular option in many
amplitude while optimizing the predefined objective evolutionary techniques. Rejection of infeasible

875
individuals offers a few simplifications of the algorithm; It should be stated that to increase the precision of the
there is no need to evaluate infeasible solutions and to optimal solution accomplished by the algorithm, we used a
compare them with feasible ones. Details of applying local search function.
this procedure are given in the sequel.
GA procedure that is used herein is as follows: IV. PROBLEM FORMULATION AND TOTAL HARMONIC
Step 1: Initialization of population: An initial population DISTORTION MINIMIZATION
of individuals is chosen. An n q 2 random matrix is As mentioned earlier, the PWM waveform illustrated in
generated. The first and second dimensions of this matrix Fig. 2 does not contain even-order harmonics due to its
represent the number of individuals and the number of symmetrical structure. Thus it can be expanded using
independent variables that lie in [0, Q/2] , respectively. An Fourier series as:
increase in number of individuals can enhance the d

performance and result in faster convergence of the f (R)  V2k 1 sin[(2k 1)R ] (2)
k 0
algorithm significantly. The value of n is considered 100 . th
In which V2k 1 is the amplitude of the 2k 1
harmonic
It has been shown in section IV that the third switching
voltage k  0,1,2, !
which can be calculated as:
angle is dependent on the first and second switching angles Q/2
and the modulation index (m ) . The rows of the matrix are V2k 1 (4/Q) f (R) sin[(2k 1)R ]
0
sorted in ascending order. The feasibility of solutions, based (3)
3
on satisfaction of constraints stated in previous section, can  (4/Q 2k 1
) 1 2 1
cos[(2k 1)Bi ]
i

## be checked in this stage. Then, feasible solutions undergo i 1

the next steps of the algorithm. A practical situation can be considered that the inverter is
Step 2: Evaluation: The fitness of each individual is used to supply ac power of constant frequency to a general
evaluated. In the sequel, objective functions for optimization three-phase ac load, where a filter is typically installed in the
are defined in proper positions. output terminal. In this case high-order harmonics are
Step 3: Selection: After the fitness of each individual is attenuated by the filter. The cut-off frequency of the filter is
computed and evaluated, parent chromosomes are selected considered as 1 kHz . Hence, elimination of harmonics larger
for reproduction. This process imitates the survival of the than 20th is assumed to be done (fundamental frequency is
fittest mechanism in nature. Roulette wheel selection  is assumed 50 Hz ). It is intuitively obvious that an increase in
the most common and easy-to-implement selection the cut-off frequency of the filter, leads to a significant
mechanism and is used in this studys approach. reduction of filter size. Also, system efficiency tends to
Step 4: Mating and Crossover: This is a crucial step to the increase. Based on this discussion, in this section, the aim is
formation of the next generation. The individual to minimize the THD while maintaining the fundamental
corresponding to an angle that is generated randomly in component (V1 ) at a constant level. That is:
Roulette wheel selection undergoes crossover and mutation 1
to produce offspring for the next generation. In this work, THD
V1
h Sh Vh2 subject to V1  m (4)
multipoint crossover is adopted for increasing efficiency.
Step 5: Mutation: The genes of the offspring are randomly In which S h  {5, 7,11,13,17,19} .
perturbed, usually with a small probability ( 0.01 in this
mentioned in (4). The reason of nonappearance of triple
study). This step avoids the algorithm from getting stuck in
harmonics in S h , is that as a three phase system is of interest,
local minima.
it can be shown  that, the triple harmonics cancel each
Step 6: Elitism: In order to preserve and use the previous
other in line-to-line voltages and thus can be ignored when
best solutions in subsequent generation, an elite-preserving
defining the optimal waveforms. Also noticing the constraint
operator is often recommended. In a simple implementation,
V1  m (that has occasionally appeared in sections of this
the best F% of the population from the current population is study), we observe that one decision variable B3
, can be
directly copied to the next generation. Number of elite count eliminated using this equality:
considered here is 2 . V1  m
Step 7: Termination criteria: The algorithm is repeated (5)
until a predetermined number of generations is reached. In  (4/Q)(1 2 cos B1  2 cos B2 2 cos B3 )
this study the termination criteria is reaching 500th Equation (5) can be rewritten as:
generation.
The values of the parameters of the algorithm greatly
B3  cos1 (1/2) m Q/4 1  2 cos B1 2 cos B2

(6)
determine whether the algorithm will find a near-optimum This way, the equality constraint is automatically satisfied
solution and whether it will find such a solution efficiently. in every step of the optimization process. The optimal
Choosing right parameter values, however, is a time- switching angles and corresponding THD are shown in Figs.
consuming task and considerable effort has gone into 3 (a, b).
developing good heuristics for it. This is due to the fact that As Fig. 3 (a) shows, step changes occur in the trajectory
parameters often interact in a complex way. of switching angles in m  0.1 and m  0.365 .

876
S Considering the handling of V1  m
in previous
section, we can define the objective function here as the

summation of these two harmonics normalized to the

D fundamental:
 V V
D
w B1, B2
5 7 . (8)

D
V1 V1
 Genetic Algorithm program has been run for several times
and for specified values of modulation index. Two sets of

solutions have been accomplished. These two sets are shown
 in Fig. 5.
 S
       

## Optimal Switching Angles (in radians)

Modulation Index (m)

(a)
 

##   set 1

Total Harmonic Distortion (THD)

##   set 2

 

 

 






       
 Modulation Index (m)
Fig. 5. Two sets of switching angles in harmonic elimination problem.

      
Modulation Index (m) In m [0, 0.08] subinterval, B1 and B2 of set1 are constant
(b) at 0.4415 rad and 0.4729 rad , respectively. In m [0, 0.1] ,
Fig. 3. (a) Optimal switching angles and (b) corresponding THD. the third switch of set2 is constant at 1.555 rad and keeps its
The frequency content of the waveform resulted from the distance from Q/2 at R/2  0.0157 rad . This is due to the
minimized THD is illustrated in Fig. 4 versus m and constraints applied to the algorithm. The corresponding
harmonic number. THD's of set1 and set2 are plotted versus m
in Fig. 6.


Harmonic Content Magnitude




Total Harmonic Distortion (%)




 set 1

 
set 2
 
 
 

  
 
Harmonic Number   Modulation Index (m)

Fig. 4. Harmonic content magnitude vs. m and harmonic number. 
      
Modulation Index (m)
V. HARMONIC ELIMINATION Fig. 6. Comparison of THD of two sets of switching angles.
Harmonic elimination aims to completely remove N  1
low-order harmonics ( N is the number of switches in the It is observed in Fig. 6 that the THD value is smaller in
first-quarter period) from a PWM waveform and, at the set2 in comparison with set1 in the considered interval of
same time, to regulate the fundamental amplitude V1 at a modulation index. The great deviation is seen at
pre-specified level (i.e. m ). m  0.99 with the error percentage of 38.8% . Thus in
In this case, we intend to eliminate the 5th and 7th applying the set of solutions, set2 is preferable. Normalized
harmonics. That is: FFT of the two sets of solutions at m  0.2, 0.4, 0.6, and 0.8
i)V1  m, ii)V5  0, iii)V7  0. (7) are brought in Fig. 6 for comparison.

877
Multi-Objective Estimate a
set 1 Optimization Problem relative
Minimize f1
Normalized FFT (V h /V1)

Higher-level importance
 Minimize f2 Information
 ! vector
Minimize fm
 Subject to constraints M1, M2,!, Mm


set 2


 Single-Objective
One Optimum Single-Objective Optimization Problem

 m=0.8 Solution Optimizer F  M1 f1 M2 f2 ! Mm fm

 m=0.6 a composite function
 m=0.4
 Fig. 7. A schematic diagram of converting multi-objective optimization to
 m=0.2
Harmonic Number Modulation Index (m) single-objective optimization.
GA was run for the minimization of (11), for different
Fig. 6. Normalized FFT Vk /V1
for m  0.2, 0.4, 0.6, and 0.8 .
values of M M  0.1, 0.2, 0.8, 2, 5,10
. It is important to
realize that the solution achieved by using the preference-
VI. THD OPTIMIZATION AND SUPPRESSING CERTAIN based strategy is greatly dependent on the relative
ORDER HARMONICS preference vector used in forming the composite function.
Here, the definition of the problem is: Besides, it is intuitive to realize that finding a relative
1 preference vector itself is highly subjective, and requires an
Minimize w1 (B1, B2 )
V1
h Sh Vh2 analysis of the qualitative and experience-driven information
to find a quantitative relative preference vector. The results
V shown in the sequel, elaborate this decision-making.
w2 (B1, B2 )  5 (9) The optimal switching angles for the two cases of
V1
M  0.1 and M  10 are shown in Fig. 8.
V S
w 3 (B1, B2 )  7 subject to V1  m 
V1



## As can be observed in the definition above, this 

optimization problem deals with more than one objective 
function. Hence, it is called a multi-objective optimization 
O
problem (MOOP) . In fact there exist a number of 
fundamental differences between the working principles of 
single and multi-objective optimization algorithms. But in 
this study to reduce the complexity, we have transformed O

multiple objectives into a single objective function by using

some predefined parameters. In other words, a composite
objective function is formed as the weighted sum of the 
      
objectives, where a weight for an objective is proportional to Modulation Index (m)
the preference factor assigned to that particular objective. Fig. 8. Optimization results for M  0.1 and M  10 .
This procedure is called preference-based multi-objective In the case of M  10 , the switching scheme changes at
optimization. A schematic of this procedure is shown in Fig. m  0.31 due to the change in the value of global optimal
1-3 in  (or Fig. 7). solution from (B1, B2 )  (0.7657, 0.8485) at m  0.31 to
Based on discussion above, the definition of the problem (B1, B2 )  (0.05278,1.1257) at m  0.315 . Also, changes in
can be considered as: the trajectory of switching angles in the case of M  10
1 M M occur at m  0.04 and m  0.15 .
F (B1, B2 )
V1
h Sh h
V1

V 2 12 V52 22 V72
V1 (10)

We can define Harmonic Elimination (HE) parameter as:
V V
subject to V1  m HE  5 7 . (12)
If an equal importance is assumed for suppressing the 5th V1 V1
and 7th harmonics (i.e. M1  M2  M ), then: This parameter has been shown in Fig. 9 versus m and the
1 M values of M . As Fig. 9 shows, the value of HE, as we were
F (B1, B2 )  2 h S Vh2 2 V52 V72

V1 h V1
(11)

expecting, falls with increase of M . Indeed, increase of M is
assigning a larger weight to the harmonic elimination term
subject to V1  m in (11).

878
 H. S. Patel and R. G. Hoft, Generalized harmonic elimination and
voltage control in thryristor inverters: part Iharmonic elimination,
IEEE Trans. Ind. Applicat., vol. 9, pp. 310317, May/June 1973.
  _____, Generalized harmonic elimination and voltage control in
Harmonic Elimination (HE)

##  thryristor inverters: part IIvoltage control technique, IEEE Trans.

Ind. Applicat., vol. 10, pp. 666673, Sept./Oct. 1974.
  J. Sun and I. Grotstollen, Pulse width modulation based on real-time
 solution of algebraic harmonic elimination equations, in Proc. 20th
Int.Conf. Ind. Electron., Contr. Instrum. IECON, vol. 1, 1994, pp. 79
 84.
  J. Sun, Optimal Pulse width modulation Techniques for High-Power

Voltage-Source Inverters, VDI VERLAG, 1995.
   J. A. Asumadu, and R. G. Hoft, Microprocessor-based sinusoidal
  
  waveform synthesis using Walsh and related orthogonal functions,


  O IEEE Trans.Power Electron., vol. 4, no. 2, pp. 234-241, 1989.
Modulation Index (m)  
 M. H. Park, D. S. Ahn, C. Y. Won, H. K. Lee, T. H. Kim,
Fig. 9. Harmonic Elimination parameter vs. m and M . Microprocessor-based new harmonic elimination method using
Walsh-Fourier transformation, in Proc. Int. Power Electron. Conf.
The corresponding THD's of different values of M have (IPEC90), pp. 799-808, 1990.
been plotted in Fig. 10.  M. H. Park, D. S. Ahn, C. Y. Won, H. K. Lee, T. H. Kim, A study on
It should be reemphasized that selection of M depends on harmonic elimination method of PWM inverter fed induction motor
higher-level information that originates from practical system using Walsh series, IFAC Symposia Series, no. 8, pp.417
423, 1990.
applications. One may find THD at an accepted level and  T. Kailath, Linear Systems. Englewood Cliffs, NJ: Prentice-Hall,
want to increase the weight of harmonic elimination portion 1980.
of (11), to suppress the harmful harmonics further. While in  C. Chen, Linear Systems Theory and Design. 3rd ed. London, U.K.:
Oxford University Press, 1999.
another application, to specify a low weight to harmonic
 J. Chiasson, L. M. Tolbert, K. McKenzie, and Z. Du, A complete
elimination is desired and its tried to minimize the THD. solution to the harmonic elimination problem, IEEE Trans. Power
 
M Electron. vol. 19, no. 2, Mar. 2004.
  M  D. E. Goldberg, Genetic algorithms in search, optimization and
M machine learning. 1st ed. Addison-Wesley, 1989.
Total Harmonic Distortion (%)

## M  M. Mitchell, An introduction to genetic algorithms. 1st ed. Prentice-


M Hall, India, 1996.
 M
 D. B. Fogel, What is evolutionary computation?, IEEE Spectr., no.
37, pp. 2632, 2000.
  K. F. Man, K. S. Tang, and S. Kwong, Genetic algorithms: concepts
 and design, IEEE Ind. Electron. News letter, vol. 45, no. 4, 1998.
 K. Deb, Multi-Objective Optimization Using Evolutionary
 Algorithms., John Wiley & Sons, Chichester, England, 2001.
  M. H. Rashid, Power Electronics: Circuits, Devices, and
Applications, Prentice Hall, 2nd ed., 1993.
   J. Sun, S. Beineke, and, H. Grotstollen, Optimal PWM Based on
         
Real-Time Solution of Harmonic Elimination Equations, IEEE
Modulation Index (m) Modulation Index (m)
Trans. Power Electronics, vol. 11, no. 4, July 1996, pp. 612-621.
Fig. 10. Total Harmonic Distortion for different values of M vs. m .  L. Davis, Ed., Handbook of Genetic Algorithms. New York: Van
Nostrand Reinhold, 1991.
VII. CONCLUSION  Z. Michalewicz, D. Dasgupta, R.G. Le Riche, and M. Schoenauer,
Evolutionary algorithms for constrained engineering problems,
Minimization of Total Harmonic Distortion (THD) is Computers & Industrial Engineering J., vol.30, no.2, pp. 851-870,
done first. Then 5th and 7th harmonics are eliminated from Sep. 1996.
 J. H. Kim and H. Myung, Evolutionary programming techniques for
the inverter output and the complete solutions are presented. constrained optimization problems, IEEE Trans. Evol. Comput., vol.
In the sequel, a composite function has been defined and 1, no. 2, pp. 129140, July 1997.
optimized to suppress 5th and 7th harmonics and to minimize  A. S. Homaifar, H.-Y. Lai, and X. Qi, Constrained optimization via
the THD concurrently. The fundamental component is genetic algorithms, Simulation, vol. 62, pp. 242254, 1994.
 J. A. Joines and C. R. Houck, On the use of nonstationary penalty
maintained at a constant level (i.e. modulation index) functions to solve nonlinear constrained optimization problems with
throughout the procedure, and a practical limitation has been gas, in Proc. IEEE Conf. Evolutionary Computation, June 1994, pp.
considered on determination of adjacent switching angles. 579584.
Optimal pulse patterns are accomplished using genetic  D. Powell and M. M. Skolnick, Using genetic algorithms in
engineering design optimization with nonlinear constraints, in Proc.
algorithm (GA) optimization method.
5th Int. Conf. Genetic Algorithms S. Forrest, Ed. Los Altos, CA:
Morgan Kaufmann, 1993, pp. 424430.
REFERENCES  Z. Michalewicz and G. Nazhiyath, Genocop III: A co-evolutionary
algorithm for numerical optimization problems with nonlinear
 N. Mohan, T. M. Undeland, and W. P. Robbins, Power Electronics:
constraints, in Proc. IEEE Int. Conf. Evolutionary Computation.
Converters, Applications, and Design. 2nd ed. New York: Wiley,
Piscataway, NJ: IEEE Press, 1995, pp. 647651.
1995.
 B. Ozpineci, L. M. Tolbert, and J. N. Chiasson, Harmonic
 P. N. Enjeti, P. D. Ziogas, and J. F. Lindsay, Programmed PWM
optimization of multilevel converters using genetic algorithms, in
techniques to eliminate harmonics: a critical evaluation, IEEE Trans.
Proc. 35th Annual IEEE Power Electronics Specialists Conf., Aachen,
Ind. Applicat., vol. 26, pp. 302316, Mar./Apr. 1990.
Germany, 2004.

879