Available online at www.sciencedirect.

com

Energy Conversion and Management 49 (2008) 89–95

www.elsevier.com/locate/enconman

Selective harmonic elimination of new family of multilevel

inverters using genetic algorithms
Khaled El-Naggar, Tamer H. Abdelhamid *

College of Technological Studies, P.O. Box 42325, Shuwaikh 70654, Kuwait

Received 4 October 2006; accepted 20 May 2007

Available online 19 July 2007

Abstract

Selective harmonic elimination has been a widely researched alternative to traditional PWM techniques. This paper presents the selec-
tive harmonic elimination of a new family of multilevel inverters using genetic algorithms. The new topology has the advantage of a
reduced number of devices compared to traditional configurations and can be extended to any number of levels. The presented system
reduces both the implementation complexity and the computational time. The GA technique finds the optimal solution set of switching
angles, if it exists, for each required harmonic profile. Both simulation results and experimental verification of the proposed inverter
topology for different numbers of levels and different harmonic profiles are presented.
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Multilevel inverters; Selective harmonic elimination; Programmed PWM; Genetic algorithms

1. Introduction the classic carrier based sinusoidal pulse width modulation

(SPWM) that uses the phase shifting technique to reduce
In recent years, multilevel inverters have received more the harmonics in the load voltage [2]. Another interesting
and more attention because of their capability of high volt- alternative is the space vector modulation (SVM) strategy
age operation, high efficiency and low electromagnetic [3]. Methods that work with low switching frequencies gen-
interference (EMI) [1]. The desired output of a multilevel erally perform one or two commutations of the power
inverter is synthesized by several sources of dc voltages. semiconductors during one cycle of the output voltages,
With an increasing number of dc voltage-sources, the generating a staircase waveform. Representatives of this
inverter voltage output waveform approaches a nearly family are multilevel selective harmonic elimination [4]
sinusoidal waveform while using a fundamental frequency and space vector control (SVC) [5]. Selective harmonic
switching scheme. This results in low switching losses, elimination depends on solving a series of trigonometric
and because of several dc sources, the switches experience equations developed from the Fourier expansion of the
lower voltage stresses. inverter output voltage. The solution is obviously the
The modulation methods used in multilevel inverters switching angles that will put specific low order frequency
can be classified according to switching frequency. Meth- components equal to zero while controlling the fundamen-
ods that work with high switching frequencies have many tal frequency component. According to the Fourier expan-
switchings in one period of the fundamental output volt- sion formula of the multilevel inverter output voltage, to
age. A very popular method in industrial applications is keep the number of eliminated harmonics at a constant
level, all switching angles must be less than p/2. However,
if the switching angles do not satisfy the condition, this
*
Corresponding author. scheme no longer exists. The quarter wave symmetry
E-mail address: [email protected] (T.H. Abdelhamid). assumption guarantees that the even harmonics will be zero

0196-8904/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enconman.2007.05.014
90 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95

and the odd harmonics will either be in phase or anti-phase number of levels. The simplest member of this new family
with the fundamental. is the 5 level inverter shown in Fig. 1. It has a main H
Available techniques for selective harmonic elimination bridge inverter Q1–Q4, two auxiliary switches Q5 and Q6
include iterative methods such as the Newton–Raphson and two dc sources. The function of the auxiliary switches
method [6] and elimination by the theory of resultant [7]. is to control the connection of the dc sources so as to con-
Iterative methods mainly depend on the initial guess and struct the staircase output voltage. The output voltage and
divergence problems are likely to occur, especially for high the required gating signals are shown in Fig. 2. This config-
numbers of inverter levels, and no optimum solution is uration can be operated in three different modes; powering,
guaranteed in terms of the total harmonic distortion free-wheeling and regenerating, according to the polarity of
(THD) as most of the higher order harmonics may strongly load voltage and current. In a 7 level inverter, another two
exist. In contrast to iterative methods, which produce only auxiliary switches and a dc source are added as shown in
one set of solutions, elimination by the theory of resultant, Fig. 3. This process is repeated for each extra level as
which depends on solving polynomials of the 22nd degree, shown in Fig. 4 for a 9 level inverter. The number of saved
produces all possible solutions, so that the one that pro- switches compared to the traditional cascaded H bridge
duces the lowest THD can be selected. Both techniques inverter is n3, where n is the number of levels. The draw-
are complicated and time consuming. back of this new family is the high rating of the main H
Since most existing multilevel inverter topologies belong bridge switches as they have to withstand the whole dc
to either a diode clamped configuration or a cascaded H bus voltage. Therefore, this new family of multilevel invert-
bridge configuration, they both have the drawback of their ers is recommended for medium power applications.
high number of switching devices. Selective harmonic elim-
ination applied to these configurations using traditional
techniques will add to the overall system complexity.
Genetic algorithms (GAs) are stochastic optimization
Main H-bridge inverter
techniques. They are simple, powerful, general purpose,
derivative free, stochastic global search algorithms inspired
by the laws of natural selection and genetics. They follow Vdc D1 D2
Q1 Q2
Darwin’s theory of evolution, where fitter individuals are
likely to survive in a competing environment. These algo-
rithms are derivative free in the sense that they do not need D5 D6 Load IL
Q5 Q6
functional derivative information to search for a set solu-
tion that minimizes (or maximizes) a given objective func-
tion. The properties of GAs reduce the computational D3 D4
Vdc Q3 Q4
burden and search time and also enable them to solve com-
plex objective functions [8].
This paper presents the selective harmonic elimination
of a new family of multilevel inverters using genetic algo- Fig. 1. The 5 level inverter of the new family.
rithms. The various steps involved in a GA based approach
for harmonic elimination in the PWM (pulse width modu-
lation) single phase inverter are explained, describing how Vo
the GA is designed and applied to the present problem. 2Vdc
Various components of GAs such as chromosomes, fitness Vdc
function, reproduction, crossover and mutation are illus-
trated as applied to the present work. The new family θ1 θ 2 π/2 π 3π /2 2π
has the advantage of a reduced number of switching
devices compared to other multilevel configurations having IL
the same number of levels. Therefore, both circuit complex-
ity and computational time are reduced. Problem formula-
tion, system simulation and experimental verification are
presented. Q1, Q4

2. Multilevel inverter new family Q2, Q3

The main purpose of the new family of multilevel invert- Q5

ers is to reduce the number of switching devices used with-
out changing the staircase nature of the output voltage. Q6
Therefore, it should have the same number of dc sources
as a conventional cascaded H bridge inverter for the same Fig. 2. Waveforms of the new 5 level inverter.
 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95 91

Main H-bridge inverter For equal dc sources, Vdc, the Fourier series expansion
of the output voltage waveform using the fundamental fre-
D5 D7 D1 D2 quency switching scheme shown in Fig. 5 is given by:
Q5 Q7 Q1 Q2
1
X 4V dc
V 0 ðxtÞ ¼ ðcosðnh1 Þ þ cosðnh2 Þ þ   
Vdc
n¼1;3;5;:::
np
Load IL
þ cosðnhs ÞÞ sinðnxtÞ ð1Þ
Vdc
Ideally, given a desired fundamental voltage V1, it is
Vdc possible to determine the switching angles h1, h2,. . ., hs so
D6 D8 D3 D4 that V0(xt)=V1sin(xt), and specific higher harmonics are
Q6 Q8 Q3 Q4 equal to zero. It has been proved that in order to control
the fundamental output voltage and eliminate n harmonics,
n + 1 equations are needed. Therefore, a 7 level inverter,
Fig. 3. The 7 level inverter of the new family. for example, can provide control of the fundamental com-
ponent besides the ability to eliminate or control the ampli-
tudes of two harmonics, not necessarily being consecutive
Main H-bridge inverter [6].
D5 D7 The switching angles can be found by solving the follow-
Q5 Q7 ing equations:
D1 D2
9
Vdc Q1 Q2 cosðh1 Þ þ cosðh2 Þ þ    þ cosðhs Þ ¼ m >>
>
Vdc cosð3h1 Þ þ cosð3h2 Þ þ    þ cosð3hs Þ ¼ 0 >
>
>
>
=
Load IL
cosð5h1 Þ þ cosð5h2 Þ þ    þ cosð5hs Þ ¼ 0 ð2Þ
D10 D9
Q10 Q9 .. >
>
.
>
>
>
>
>
cosðnh1 Þ þ cosðnh2 Þ þ    þ cosðnhs Þ ¼ 0
;
Vdc
D3 D4
Vdc Q3 Q4 where m = V1/(4Vdc/p) and the modulation index ma is gi-
D6
ven by ma = m/s.
Q6 Q8 D8 One approach to solving the set of nonlinear transcen-
dental Eq. (2) is to use an iterative method such as the
Newton–Raphson method [6]. In contrast to iterative
Fig. 4. The 9 level inverter of the new family.
methods, the transcendental equations characterizing the
harmonic content can be converted into polynomial equa-
3. Mathematical method of switching tions and solved using the theory of resultants to find the
solutions when they exist [7]. These sets of solutions are
The control of the new family of multilevel inverters is then examined for its corresponding total harmonic distor-
to choose a series of switching angles to synthesize a tion in order to select the set that generates the lowest har-
desired sinusoidal voltage waveform. Fig. 5 shows a gener- monic distortion (mostly due to the 11th and 13th
alized quarter wave, symmetric, stepped voltage waveform harmonics). The computed THD in % is defined by:
synthesized by a 2s + 1 level inverter, where s is the number sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
of switching angles, which also equals the number of dc V 23 þ V 25 þ V 27 þ    þ V 219
%THD ¼  100 ð3Þ
sources. V 21
The iterative methods give only one specific solution,
Vo which depends on the initial guess, and there is no guaran-
tee that it is the optimum solution. On the other hand, the
sVdc
solution using resultant theory produces all possible solu-
tions, but it is complicated and time consuming. It is, there-
2Vdc
fore, worth considering more techniques and simple
Vdc
techniques such as GAs.
θ1 θ2 θs π /2 π 3π/2 2π
4. Solution using genetic algorithms

Since the beginning of the dawn of the era of comput-

ers, scientists have been mimicking biological processes
Fig. 5. Generalized stepped voltage of multilevel inverter. with computer processes. In the 1950s and 1960s, computer
92 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95

scientists were toying with the idea that evolution could be summation of the individual absolute errors (AE). Thus,
used as a model for optimization in engineering problems. we can write AE as:
The main idea was to evolve a solution using operators
1
inspired by genetic variation and natural selection. These AE ¼ ð4Þ
N
evolution inspired algorithms were not given much notice P
jei j þ D
until John Holland (1960s) invented genetic algorithms [9]. i¼1
He got a bit more biological with these GAs. The idea was
where D is a small constant (D = 0.00001 in this work) to
to move from one population of ‘‘chromosomes’’ (or bit
avoid overflow problems if Fsum goes to zero. It is impor-
strings of 1’s and 0’s) to another population using such
tant to mention here that this formulation assumes that
operators as selection, crossover and mutation. The earlier
the error of Eq. (4) is a Gaussian white sequence with
scientists just used mutation to reach their desired end.
known covariance as,
Today, many other variations of this method have come
to be subsumed under the name of genetic algorithms. 0; i 6¼ j

T
There is not a strict definition; rather, it is a main idea, EfeðiÞeðjÞ g ¼ ð5Þ
RðiÞ; i ¼ j
which allows scientists to experiment in new and interest-
ing ways. A GA for optimization is different from ‘‘classi- In addition to coding and fitness evaluation, the simple GA
cal’’ optimization methods in several ways: random versus is composed of another three basic operations; reproduc-
deterministic operation, population versus single best solu- tion, crossover and mutation. Each string of the old popu-
tion and selecting solutions via ‘‘survival of the fittest’’ lation goes through these three steps before a new
[10]. population is generated [11].

4.1. Encoding of a chromosome

4.3. Selection
A simple GA starts with random generation of a popu-
Inspired by the role of natural selection in evolution, the
lation. A population consists of a set of chromosomes
GA performs a selection process in which the ‘‘most fit’’
(strings). Usually, the string size ranges from 50–1000.
members of the population survive, and the ‘‘least fit’’
The population may be of any size according to the accu-
members are eliminated. In a constrained optimization
racy required. The population size remains constant
problem, the notion of ‘‘fitness’’ depends partly on whether
throughout the whole process. A chromosome should, in
a solution is feasible (i.e. whether it satisfies all of the con-
some way, contain information about the solution that it
straints), and partly on its objective function value. The
represents. The problem variables are coded using a suit-
selection process is the step that guides the evolutionary
able coding system. The most used way of encoding is a
algorithm towards ever better solutions. The parent selec-
binary string. Indeed, there are many other ways of encod-
tion can accomplished in a variety of ways. In this study,
ing. The encoding depends mainly on the problem consid-
the process is executed using the roulette wheel technique
ered. For example, one can directly encode integers or real
[10].
numbers; sometimes it is useful to encode in special ways
and so on. In this study, a binary coding system is used.
A string then could look like this (in the binary case): 4.4. Crossover
String 1 1101100100110110
After reproduction, we can proceed to the crossover
String 2 1101111000011110
operation. Crossover operates on selected genes from the
A string in GAs may be divided into a number of sub- parent chromosomes and creates new offspring. The sim-
strings. The number of sub-strings, usually, equals the plest way to do that is to choose, randomly, some crossover
number of problem variables. Therefore, the above strings point, copy everything before this point from the first par-
could be a representation of one parameter (16 bit) or 2 ent and then copy everything after the crossover point from
parameters (8 bits) or any other combination. the other parent. Crossover can be illustrated as follows:
(j is the crossover point):
Chromosome 1 11011j00100110110
4.2. Fitness function
Chromosome 2 11011j11000011110
Offspring 1 11011j11000011110
Once the random population is generated, the solution
Offspring 2 11011j00100110110
represented by each string should be evaluated. The fitness
function is the function responsible for evaluation of the There are other ways to implement crossover, for exam-
solution at each step. The objective here is to minimize ple, we can choose more crossover points. Crossover can be
the estimation error (e) in determining the switching angles quite complicated and depends mainly on the encoding of
such that some selected harmonics are minimized or put chromosomes. Specific crossover made for a specific prob-
equal to zero. The fitness function is set to minimize the lem can improve the performance of the GA.
 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95 93

4.5. Mutation 100

Elimination of 3rd and 5th harmonics
90

After performing crossover, mutation takes place. 80

Mutation is used to prevent all the solutions in the popula- 70

tion falling into a local optimum of the solved problem. 60

Degrees
θ3
The mutation operation randomly changes the offspring 50

40
that resulted from crossover. In case of binary encoding,
30
we can switch a few randomly chosen bits from 1 to 0 or θ2
20
from 0 to 1. Mutation can be illustrated as follows:
10
θ1
Original offspring 1 1101111000011110 0
0.5 1 1.5 2 2.5 3
Original offspring 2 1101100100110110 m
Mutated offspring 1 1100111000011110 Fig. 7. Solutions for 3 angles versus m for 7 level inverter (3rd and 5th
Mutated offspring 2 1101101100110110 harmonics eliminated).
The technique of mutation (as well as crossover)
depends mainly on the encoding of the chromosomes.
For example, when we are encoding permutations, muta-
tion could be performed as an exchange of two genes. 100
A GA toolbox package for Matlab, ‘‘FlexGA version 1- Elimination of 3rd, 5th and 7th harmonics
90
CynapSys.LLC’’. is used to perform the study This code 80

initializes a random sample of individuals with different 70

parameters to be optimized using the GA algorithm Degrees 60 θ4

approach. Based on experience and trial and error, the 50

GA operator probabilities and sizes are selected [11]. A 40

θ3
population size of 100 was found to enough for our prob- 30
θ2
lem. The crossover probability used was 0.6, and mutation 20

probability was 0.02. The coding system used in all tests 10

θ1
was the binary system. 0
1 1.5 2 2.5 3 3.5 4
m
5. Simulation results Fig. 8. Solutions for 4 angles versus m for 9 level inverter (3rd, 5th and 7th
harmonics eliminated).
The generalized transcendental equations of the multi-
level inverter Eq. (2) are solved using the described GA
for different values of inverter levels. The solutions for
switching angles versus m for different levels are shown in Another set of waveforms for a 7 level inverter are shown
Figs. 6–9, where consecutive harmonics can be eliminated in Fig. 11 (s = 3, ma = 0.6, m = 1.8, V1 = (4 mVdc/p) =
starting from the 3rd harmonic. A great number of simula- 45.8 V). Usually, the described GA technique finds more
tion results are obtained, however, selected results for a 5 than one set of solutions for each particular harmonic pro-
level inverter are shown in Fig. 10 for Vdc = 20 V, and file. The set of switching angles that produces the smallest
ma = 0.6 (s = 2, m = 1.2, V1 = (4 mVdc/p) = 30.5 V). THD is chosen.

90 100
Elimination of 3rd harmonic Elimination of 3rd, 5th, 7th and 9th harmonics
80 90

70 80

60 70
θ5
Degrees

60
Degrees

50

θ2 50
40
40 θ4
30
30 θ3
20
θ1 20
10 θ2
10
0 θ1
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 0
1 1.5 2 2.5 3 3.5 4 4.5
m m

Fig. 6. Solutions for 2 angles versus m for 5 level inverter (3rd harmonic Fig. 9. Solutions for 5 angles versus m for 11 level inverter (3rd, 5th, 7th
eliminated). and 9th harmonics eliminated).
94 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95

a 40
b 30
30
25
20
Output Voltage (V)

Output Voltage (V)

20
10

0 15

-10
10
-20
5
-30

-40 0
0.02 0.022 0.024 0.026 0.028 0.03 0.032 0.034 0.036 0.038 0.04 0 100 200 300 400 500 600 700 800 900 1000

Time (sec) Frequency (Hz)

Fig. 10. (a) Output voltage, (b) corresponding FFT of 5 level inverter for 3rd harmonic elimination at Vdc = 20 V, and ma = 0.6.

a 60
b 50

45
40
40
Output Voltage (V)

Output Voltage (V) 35

20
30

0 25

20
-20
15

10
-40
5

-60 0
0.02 0.022 0.024 0.026 0.028 0.03 0.032 0.034 0.036 0.038 0.04 0 100 200 300 400 500 600 700 800 900 1000

Time (sec) Frequency (Hz)

Fig. 11. (a) Output voltage, (b) corresponding FFT of 7 level inverter for 3rd and 5th harmonics elimination at Vdc = 20 V, and ma = 0.6.

6. Implementation consideration using microprocessors or DSPs depending on the desired

level of performance and flexibility. A real time controller
Although the theory presented offers alternative switch- based on the available MCB-1A Hampden microprocessor
ing waveforms, the generalized selective harmonic control kit is used to implement the harmonic elimination PWM
technique can still be classified as a programmed PWM method. A prototype multilevel inverter has been built
method. As a result, hardware implementation is similar using IRF620 (200V-5A) MOSFETs as switching devices.
to other programmed PWM techniques. The basic princi- The switching angles obtained with the GA technique are
ple relies on storing pre-computed switching angles in converted into time intervals and stored in the form of look
memory and generating the PWM output by cycling up tables for different values of modulation indices. Then,
through the table. This architecture is normally achieved these switching angles are converted into switching pat-

Fig. 12. (a) Output voltage, (b) corresponding FFT of 5 level inverter for 3rd harmonic elimination at Vdc = 20 V, and ma = 0.6.
 K. El-Naggar, T.H. Abdelhamid / Energy Conversion and Management 49 (2008) 89–95 95

Fig. 13. (a) Output voltage, (b) corresponding FFT of 7 level inverter for 3rd and 5th harmonics elimination at Vdc = 20 V, and ma = 0.6.

terns using a down counter and some logic operations and show that the algorithm can be effectively used for selec-
then stored in an in house EPROM. The switching patterns tive harmonic elimination of the new family of multilevel
obtained from the controller are interfaced to the inverter inverters and results in a dramatic decrease in the output
power switches through optocoupler isolators. To verify voltage THD.
the presented idea, the hardware implementation is only
developed for 5 level and 7 level inverters, and it can be
extended to any number of levels and any desired harmonic References
profile. Experimental verification of the presented simula-
[1] Rodriguez J, Lai J, Peng FZ. Multilevel converters: a survey of
tion results are shown in Figs. 12 and 13, where they are
topologies, controls, and applications. IEEE Trans Ind Appl
in close agreement with each other. The THD of the exper- 2002;49(4):724–38. Aug.
iments was found to be a little higher than that of the sim- [2] Liang Y, Nwankpa CO. A new type of STATCOM based on
ulation because the control resolution is limited to 8 ls, cascading voltage-source inverters with phase-shifted unipolar
and the switches are not ideal. The effective switching fre- SPWM. IEEE Trans Ind Appl 1999;35:1118–23.
[3] Rodriguez J, Correa P, Moran L. A vector control technique for
quency of the main four switches is 50 Hz, while it is
medium voltage multilevel inverters. In: Proc. IEEE APEC, Anaheim,
100 Hz for the auxiliary switches. CA, March 2001. p. 173–8.
[4] Sirisukprasert S, Lai JS, Liu TH. Optimum harmonic reduction with
7. Conclusions a wide range of modulation indexes for multilevel converters. In:
Conf. rec. IEEE-IAS annu. meeting, Rome, Italy, October 2000. p.
2094–9.
The selective harmonic elimination of a new family of
[5] Rodrı́guez J, Morán L, Silva C, Correa P. A high performance vector
multilevel inverters using GA has been presented. The control of a 11-level inverter. In: Proceedings of the 3rd international
new configuration has the advantage of a reduced num- power electronics and motion control conference, Beijing, China,
ber of switching devices compared to traditional configu- August 2000. p. 1116–21.
rations of the same number of levels. However, it has the [6] Patel HS, Hoft RG. Generalized techniques of harmonic elimination
and voltage control in thyristor inverters: part I – harmonic
drawback of the high rating of the four main switches as
elimination. IEEE Trans Ind Appl 1973;3:310–7.
they are subjected to the whole dc bus voltage. The [7] Chiasson JN, Tolbert LM, McKenzie KJ, Du Z. Control of a
remaining devices are rated according to the number of multilevel converter using resultant theory. IEEE Trans Cont Syst
dc sources they connect. The GA technique proved its Theory 2003;11(3):345–54. May.
effectiveness, and almost a continuous range of switching [8] Ovaska SJ, Bose T, Vainio O. Genetic algorithm-assisted design of
adaptive predictive filters for 50/60 Hz power systems instrumenta-
angles are produced. Some of these sets of solutions
tion. IEEE Trans Instrum Meas 2005;54(5):2041–8.
couldn’t be obtained before when using traditional tech- [9] Holland JH. Adaptation in natural and artificial systems. University
niques. The GA technique usually produces more than of Michigan Press; 1975.
one possible solution set for each harmonic profile and [10] Ramerirez-Rosado JI, Bernal-Agustin JL. Genetic algorithms applied
a given specific modulation index. For multiple solutions, to the design of large power distribution systems. IEEE Trans Power
Syst 1998;13(2):696–703.
the solution that gives the lowest THD is selected.
[11] Ozdemir A, Lim JY, Singh C. Voltage and reactive power distribution
Experimental implementation proved the effectiveness of factors calculated by genetic algorithms. In: Proceedings of IEEE
both the new inverter topology and the presented GA power engineering society transmission and distribution conference,
technique. Both simulation and experimental results 2002. p. 1276–81.