PWM optimisation for three-level voltage inverter

based on clonal selection algorithm

H. Lou, C. Mao, D. Wang and J. Lu

Abstract: A clonal selection algorithm (CSA) is first applied to determine the switching time for a
three-level voltage pulse width modulation (PWM) inverter. To obtain an optimal PWM switching
pattern, CSA should calculate a set of nonlinear transcendental equations with nonlinear equality
and inequality constraints. Unlike the general CSA by penalising infeasible solutions, this new
CSA deals with constraints and guides the search to the true optimum solution by an infeasibility
degree with a random disturbance (IFDD) selection operation. The proposed method minimise the
total harmonic distortion and generates a waveform with adjustable amplitude of the fundamental
component. Calculation results verify that the proposed method is efficient to obtain the optimal
PWM switching pattern and is superior to natural sampling and genetic algorithm-based PWM
switching pattern.

1 Introduction optimum. However, the GA has some limits in practical

application, such as slow speed of convergence [12, 13].
Pulse width modulation (PWM) technology is widely used Clonal selection algorithm (CSA) [12, 13] is a colony
to control high-power converters, in which the switching search strategy; both parallelity and randomicity in search
frequency is limited [1 – 4]. The problem of the optimal mechanism make it difficult to be trapped into local
design of PWN waveforms is a constrained nonlinear pro- optimum. Compared with GA optimisation, the CSA optim-
gramming problem, which otherwise requires iterative/ isation facilitates the acquisition of the global optimum. The
approximate solutions. The key problem in the iteration CSA can thus converge into the global optimum with a
process is the computing time required for a set of nonlinear higher probability and at a higher speed.
transcendental equations with nonlinear equality and In this paper, CSA is first used to determine PWM
inequality constraints, which involves many local optima. switching time to minimise the total voltage harmonic dis-
Because of this inherent nonlinear nature, an accurate, tortion and to generate a desired waveform with adjustable
reliable and real-time solution of these transcendental amplitude of the fundamental component, which is con-
equations is impractical [5]. verted into an optimisation problem with equality and
Owing to different control objectives, such as selective inequality constraints. The general way of dealing with con-
harmonic elimination [6], minimum total harmonic distor- straints is by penalising infeasible solutions. However,
tion (THD) [7] and distortion factor [8], different cost func- because it is difficult to set the penalty parameters appropri-
tions are taken to obtain the optimum PWM switching ately and that the penalty term distorts the characteristic of
pattern. However, the optimisation methods are the same. the objective function of the problem, the performance of
Recently, Newton –Raphson (NR) method [9], conjugate penalty function methods is not always satisfactory. In
gradient descent method and genetic algorithm (GA) order to handle constraints and guide the search direction,
[6 – 8], [10, 11] have been proposed to achieve certain the infeasibility degree with a random disturbance (IFDD)
optimisation for the output voltage of the power electronic selection operation is designed. Different from the infeasi-
converters, especially for high-power converters. bility degree (IFD) selection [14], which is implemented
Newton’s method is an iterative algorithm that provides a by judging whether the IFD of a solution is greater than
local solution which requires that the cost function be twice the threshold, this method generates a random disturbance
differentiable. In addition, the final solution depends highly to the IFD each time it is calculated. The threshold is the
on the initial point. This step could even lead to a non- product of a coefficient and the average IFDDs in the
optimum solution for some cases. current population, and the coefficient decreases in every
GA is a global optimising algorithm, mimicking the bio- generation. Compared with IFD-based CSA (IFD-CSA)
logical evolutionary processes. There are examples of GA and penalty function-based CSA (PCSA), the IFDD-based
applications in power electronics in the literature [6 – 8, CSA (IFDD-CSA) shows more advantages of searching
10, 11]. Theoretically, GA can reach the real global the optimum solution, with certain constraints.
The validity of the proposed method has been verified by
numerical experiment studies on a PWM-controlled,
# The Institution of Engineering and Technology 2007
medium-voltage, variable-frequency drive with various
doi:10.1049/iet-epa:20060428
output frequencies and carrier frequencies. In the whole
Paper first received 28th October 2006 and in revised form 24th April 2007
output frequency range, THD is reduced effectively and
high-quality output waveforms are obtained.
The authors are with the College of Electrical and Electronic Engineering,
Huazhong University of Science and Technology (HUST), Wuhan 430074, The rest of this paper is organised as follow. Section 2
People’s Republic of China presents the description of the three-level medium voltage
E-mail: cxmao@[Link] variable frequency drive system and the transcendental
870 IET Electr. Power Appl., 2007, 1, (6), pp. 870 –878

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
equations for minimising THD. Section 3 describes the The Fourier representation of the phase voltage vAN given in
basic idea of CSA. Section 4 presents the main procedure Fig. 2 can be easily determined to be
of IFDD-CSA. In Section 5, proposed approach is applied
to minimising THD. In Section 6, calculation results are X
1 pffiffiffi
vAN (d) ¼ VANk 2  sin k d (2)
presented in order to verify that the proposed method is effi- k¼1,3,5, 
cient to obtain the optimal PWM switching pattern and
N
superior to natural sampling (NS) and GA-based PWM 4V dc X 2
 
switching pattern. VANk ¼ pffiffiffi (1)i1  cos k di (3)
2 k p i¼1

2 System and formulation where VANk is the line to neutral RMS value of the harmonic
components including the fundamental voltage of the inver-
2.1 System ter. The neutral is shown in Fig.1a.
Let d ¼ vt
A typical medium-voltage, variable-frequency system
X
1 pffiffiffi
based on three-level neutral point clamped inverter is vAN (t) ¼ VANk 2  sin k vt
shown in Fig. 1a. Equivalent circuit of the medium k¼1,3,5,
voltage variable frequency system with filters is shown
N
in Fig. 1b. Where, Vdc is dc link voltage; VAN is three- 4V dc X 2
 
level inverter voltage and Vm is load voltage. An equival- VANk ¼ pffiffiffi (1)i1  cos k vti (4)
ent circuit of the motor model can be seen. Where, R1 is 2 k p i¼1
stator resistance; L1s is stator leakage inductance; R 20 is As the interest here is a three-phase drive, the triplen harmo-
rotor resistance referred to the reference frame of the nics in the phase voltages will be cancelled in the line – line
stator; L 20 s is rotor leakage inductance referred to the voltages
reference frame of the stator; Rm is magnetising resist-
ance; Lm is magnetising inductance and s is slip. When pffiffiffi N2
4 3V dc X  
the output frequency f, rated voltage V, active power P VANLk ¼ pffiffiffi (1)i1  cos k vti (5)
and power factor cos w are given, the simplified represen- 2 k p i¼1
tation of the equivalent circuit of the motor Rd and Xd can
be calculated. where k ¼ 1, 6m + 1, m ¼ 1, 2, 3, . . . integer.
The phase output voltage vAN of the three-level By analysing the circuit shown in Fig. 1 b, the line to
inverter is shown in Fig. 2. To ensure the absence of neutral RMS value of the motor voltage V mk can be
even-order harmonics, each waveform is assumed to be obtained, if the inverter output voltage VANk is given
quarter-period symmetrical and half-period inversely V mk ¼ VANk  H ðk vÞ (6)
symmetrical.
A symmetrical PWM waveform with a given structure where k ¼ 1, 3, 5, . . . odd and H (kv) is the amplitude ratio
will be uniquely determined if the N2 pulses in the first of V mk to VANk . By using ohm’s law, Kirchhoff’s current
half period are specified. Evidently, these angles must and voltage laws, we can obtain
satisfy the basic constraint  
V_ m
p H(k v) ¼  
0
d1
d2
d3
   dN2
(1) V_ 
AN
2
0 11=2
B R2d þ k 2 v2 L2d C
B C
¼B 2 2 2 2 2 C
@ (Rd þ R  k v Ld RC  k v Rd LC) A
3 3 2
þ (k vLd þ k vL þ k vRd RC  k v Ld LC)
(7)
Substituting (5) into (6), we can obtain motor voltage
V mk . And the motor line–line voltage of the harmonic com-
ponents including the fundamental voltage is as following
pffiffiffi N2
4 3V dc X  
V mLk ¼ pffiffiffi (1)i1  cos k vti  H ðk vÞ (8)
2 k p i¼1

where k ¼ 1, 6m + 1, m ¼ 1, 2, 3, . . . integer.

Fig. 1 Medium voltage variable frequency system with filters

a Three-phase three-level circuit
b Equivalent circuit Fig. 2 Inverter phase output voltage vAN

IET Electr. Power Appl., Vol. 1, No. 6, November 2007 871

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
Fig. 3 Process of the clonal strategies

2.2 Equations for minimising THD

Fig. 4 Process of the new CSA
The objective is to find the switching time to make the funda-
mental output voltage VmL1 ¼ Vref and minimise the THD, This is converted to an optimisation problem and is stated
where Vref is the reference output voltage. The total harmonic as follows.
voltage distortion defines the relationship between the total For N2 number of pulses per half cycle, let F (t) be the
harmonic voltage and the fundamental voltage. Thus, if the objective function; then, mathematically, we can write the
fundamental ac line-to-line voltage is VmL1 and the total problem as
line-to-line harmonic voltage is VmLk , then
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  
uk¼50 1
u X V m2 minimise: F(t) ¼
THD ¼ t Lk
(9) V mL1
V m vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k¼2 L1 u P  
uX
An upper summation limit of k ¼ 50 is chosen for t V m2L(6 m1) þ V m2L(6 mþ1) (10)
calculation purposes [15]. It gives good practical results. m¼2

Fig. 5 Flowchart of THD minimisation based on IFDD-CSA

a Main flowchart
b Flowchart of IFDD selection

872 IET Electr. Power Appl., Vol. 1, No. 6, November 2007

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
 explored in the field of artificial intelligence, are used in
the CSA. Just like the GA, the CSA is shown to be an evol-
utionary strategy capable of solving complex machine
learning tasks, like numerical optimisation by adopting
the clonal operator and also the problem of global optimis-
ation with high convergence speed [12]. The process of
CSA is shown in Fig. 3.
Clone: after generating randomly the antibodies set A[n],
each antigen A[i] subjected to certain mechanism is
copied to obtain a new antibody population. The clonal
size subjected to antibody affinity (fitness) function is
relating to an integer coefficient.
Clone mutation: in order to save the information of the
original population, the clonal mutation is not performed
on original antibody population, but only on the cloned anti-
body population.
Clone selection: find the highest affinity antigen among the
original antigen A[i] and each A 0 [i] in the mutated antigens
to replace the antigen A[i] in the original population.
The algorithm will be halted when the restricted iterative
number is met or when the best solution in current gener-
ation satisfies the accuracy condition.

4 New CSA

Unlike the general CSA, which penalises infeasible sol-

utions, this new CSA deals with constraints and guides
the search direction to the true optimum solution by IFDD
selection operation. The other operations, such as clone,
clonal mutate and clonal select are the same as the
general CSA.
For constrained optimisation problem, its feasible region
is less than the solution space. Because CSA is a stochastic
algorithm and its search space is the whole solution space, it
will generate many infeasible solutions during the
implement. In [14], IFD is used to measure the IFD of a can-
didate solution quantitatively. Solutions obtained by IFD
satisfy the constraints well; however, sometimes they are
just feasible solutions but not optimum ones. Here we intro-
duce a random disturbance in the IFD (IFDD) and we let
this disturbance equal zero during the last several gener-
ations, which not only enlarges the search space, but also
lets the solution satisfy the constraints.
The IFDD f of a candidate solution ti is defined as
follows
X
J

Fig. 6 Switching time obtained with different methods f(t i ) ¼ [hj (t i )  M  rand  m]2 (13)
j¼1
a IFDD-CSA
b NS (natural sampling)
c GA where hj(ti ) ¼ M is equality constraints, J is the number of
equality constraints, m denotes the weighted coefficient
which is two orders of magnitude less than M, rand
p denotes a random real number between 21 and 1. In
subjected to: 0
t1
t2
t3
  
tN2
(11)
2v order to obtain the feasible solution, we make m zero
V mL1 ¼ Vref , V mL5 ¼ 0, V mL7 ¼ 0, during last several generations, that is
V mL11 ¼ 0, V mL13 ¼ 0 (12) X
J
f(t i ) ¼ [hj (t i )  M]2 (IFD)
j¼1
In this paper, the THD is computed throughout the 50th,
so P ¼ 8. The even and triplen harmonics are not computed The IFD of the solution can be regarded as the distance
in THD because they do not appear in the line –line between the solution and the feasible region. For feasible
voltages. solution, its IFD is zero but for infeasible solution it is
greater than zero. The greater the distance of the solution
3 Basic idea of CSA from the feasible region, the larger the IFD of the solution
is. The IFDD represents a random disturbance added to
On the basis of the clonal selection theory, the main mech- IFD, which can enlarge the search space during searching
anisms of clone in immune system, which have been the optimum. The pressure of rejection on infeasible
IET Electr. Power Appl., Vol. 1, No. 6, November 2007 873

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
Fig. 7 Motor voltage waveforms and its harmonic analysis at 50 Hz
a IFDD-CSA
b NS (natural sampling)
c GA

solutions should be increased as iteration. It is implemented where temperature T ¼ Tstart þ (Tend  Tstart )  k=kmax is a
by a threshold value fcrit [14], which is designed by two coefficient which increases from Tstart to Tend along kmax
parts: the coefficient decreased with the iteration and the iterations to control the acceptable bound of the infeasible
average IFDD of the population solutions. POPSIZE is the size of population, fcrit is a
threshold value to determine whether an infeasible solution
POPSIZE  is accepted or rejected. An infeasible solution is accepted if
P
f(ti ) its IFDD is less than or equal to fcrit , else it is rejected. In
1 i¼1 order to keep the size of the population fixed, the rejected
fcrit ¼  (14)
T POPSIZE solutions are replaced by the most feasible solutions in the
874 IET Electr. Power Appl., Vol. 1, No. 6, November 2007

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
 chosen as the affinity function JF
1
JF(t) ¼ (15)
F(t)
Step 4: Clone.
Select each set of switching time ti to clone; the number of
clones Qi is proportional to the affinity. Generally, Qi is
given by
0 1
B JF(t ) C
B C
Qi ¼ (int)B Nc  POPSIZE i C (16)
@ P A
JF(t i )
j¼0
Fig. 8 Variation of THD value with CSA and GA
where POPSIZE is the number of the initial solutions and Nc
is a given integer Nc . POPSIZE. In this paper, we measure
current population. After the IFDD selection operation, the Nc with POPSIZE, for example Nc ¼ 5  POPSIZE
new antibody population is generated and takes part in the (POPSIZE ¼ 500). If the POPSIZE itself is very large,
next clonal operation. such as 5000, then we may let Nc a little larger than
The function of fcrit is to reduce the search space and POPSIZE, for example 2  POPSIZE. And int(x) rounds
close the feasible region gradually. But the fitness function the elements of x to the least integer bigger than x.
is the objective function of the problem, making the search Step 5: Clonal mutation.
directed to the global optimum solution in the whole sol- Compared with the cross-over and mutation operator in
ution space. The combined force from the two aspects GA, the clonal algorithm mainly searches the neighbour-
guides the solution convergence to the global feasible hood region decided by the mutation operator. It is
optimum solution. obvious that it improves the parallel searching ability and
The main procedure of IFDD-CSA is shown in Fig. 4. expands the search area. Generally, cross-over and mutation
probabilities in GA are set beforehand (for example to 0.9
5 THD minimisation based on IFDD-CSA and 0.05, respectively). The probability of clonal mutation
is also set at a larger number 0.3.
5.1 Optimisation with an assigned output Step 6: Clonal selection.
frequency
t 0j , JF(t i ) , JF(t 0j )
ti ¼ (17)
Step 1: Initialisation. t i , else
Create an initial set of solutions t randomly. The number of P Pi
variables in each solution ti is the number of the switching where, i [ (0, POPSIZE  1), j [ ( i1 r¼0 Qr , r¼0 Qr ).
0
time N2 in a quarter period shown in Fig. 2, that is The superscript ‘ ’ indicates that it is the solution after clone
t i ¼ [t1 , t2 , ::: , tN2 ]. and clonal mutation operation.
Step 2: IFDD selection. Step 7: Replace the current solutions t with the new t.
IFDD of each solution is calculated from equation (13). In Step 8: Go to step 3 until termination criterion is reached.
this application, (13) and IFD should be changed to The flowchart of this problem is shown in Fig. 5.
f(ti ) ¼ [V mL1 (ti )  Vref  rand  m]2 þ V m2L5 þ V m2L7 þ
V m2L11 þ V m2L13 and f(t i ) ¼ [V mL1 (t i )  Vref ]2 þ V m2L5 5.2 Optimisation in a frequency range
þV m2L7 þ V m2L11 þ V m2L13 , respectively. Then, the IFDD
f of each solution is compared with the threshold value The output harmonic intensity of the inverter may always be
fcrit calculated from (14) and the more feasible solutions randomly changed with various random sources and these
are selected. If the iteration number reaches kmax , turn to random harmonic distributions may be different at every
step 3, otherwise, increase T and back to step 2. sinusoidal cycle. To solve this issue, this paper proposes
Step 3: Evaluation. using IFD-CSA to find the switching time, which produces
The degree of ‘goodness’ of a solution is qualified by a minimum THD in a frequency range.
assigning a value to it. This is done by defining a proper affi- The parameters in Fig. 1 are given: L ¼ 3 mH, C ¼ 50 mF,
nity function for the problem. In the present problem, THD Vrated ¼ 6 kV, P ¼ 2.5MW, cos w ¼ 0.85. The equivalent
and the difference between the fundamental component motor is calculated Rd ¼ 10.404 V, Ld ¼ 20.5mH.
VmL1 and the reference output voltage Vref is to be mini- Considering the main operating frequency of a medium
mised. The reciprocal of the objective function (10) is voltage variable frequency device, let the output frequency

Table 1: Conditions in a frequency range (Udc 5 1.35)

f, Hz 30 35 40 45 50
N (N2) 26 (13) 22 (11) 18 (9) 14 (7) 14 (7)
f  N, Hz 780 770 720 630 700
Uref 0.6 0.7 0.8 0.9 1.0
Termination criterion, % 25 25 20 15 15
THD (NS), % 25.72 38.14 33.29 32.06 22.00
THD (IFDD-CSA), % 14.00 13.35 13.7 10.50 5.26

IET Electr. Power Appl., Vol. 1, No. 6, November 2007 875

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
Fig. 9 Motor voltage waveforms and its harmonic analysis at different output frequency
a 40 Hz (IFDD-CSA)
b 40 Hz (NS)
c 30 Hz (IFDD-CSA)
d 30 Hz (NS)
876 IET Electr. Power Appl., Vol. 1, No. 6, November 2007

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
interval be [30 Hz, 50 Hz]. The fundamental output voltage: 6.2 Optimisation in a frequency range
VmL1 ¼ Vref ¼ 1.0  f/50 p.u., Udc ¼ 1.35 p.u.
The initial parameters are L ¼ 3 mH, C ¼ 50 mF, As shown in Table 1, in the whole output frequency range
f ¼ fmin . The optimisation process is as follows. 30– 50 Hz, N pulses per cycle change with the output fre-
quency f, and make sure that f  N
800 Hz. The funda-
Step 1: Calculate the fundamental voltage Vr and the mental output voltage changes to VmL1 ¼ Vref  f /50.
corresponding equivalent load Rd þ jvLd with the output Different termination criterions are employed with various
frequency f. output frequencies. Let fstep ¼ 5 Hz, a set of switching
Optimisation calculation: Calculate the switching time time at each frequency can be calculated. The output wave-
with the output frequency f. forms at 40 and 30 Hz are shown in Fig. 9. The THDs
Then, we can obtain a set of optimised parameters: obtained with different methods are also shown in Table 1.
Switching time t1 – tN2 . The authors also use the PCSA and IFD-CSA to obtain
Step 2: Set f ¼ f þ fstep , where fstep is the iteration step of the the optimum PWM switching time. If the output frequency
output frquency, if fmax is not exceeded. is larger than 45 Hz, PCSA, IFD-CSA and IFDD-CSA can
Step 3: Stop when fmax is exceeded and save each set of find the optimum, and the speed of PCSA is faster.
switching time at each frequency. However, at the output frequency below 45 Hz, PSCA
and IFD-CSA cannot find the optimum for an acceptable
period of time. The results of PCSA always violate the con-
straints. The results of IFD-CSA satisfy the constraints well,
6 Results but lacks the ability to search the optimum solution and also
the THDs are larger than the NS method.
In this paper, the real-valued IFDD-CSA is performed on a PC PCSA has the largest search space. But the penalty term
(CPU: Celeron 2.53 GHz, RAM: 512 MB, OS: Microsoft XP) distorts the characteristic of the objective function of the
with Microsoft Visual Cþþ 6.0 software. The parameters of problem. Thus, the penalty parameters must be set appropri-
IFDD-CSA program influence the dynamics and quality ate to guide the search direction convergence to the feasible
of convergence of IFDD-CSA. The proper determination of global optimum solution. This is very difficult when the
these parameters for a particular problem is still an open ques- output frequency is below 45 Hz.
tion. In this work, the authors vary the parameters and study The search space of IFD-CSA is shrunk, which leads to a
the impact. The following IFDD-CSA program parameters lack of ability to find the optimum.
may yield satisfactory results: IFDD-CSA has the advantages of both IFD-CSA and
Initial population size – 500; maximum number of genera- PCSA. Its searching space is larger than IFD-CSA. And it
tions – 50; probability of clonal mutation – 0.3; Tstart – 0.8; reserves the ability of searching the optimum of CSA as
Tend – 1.2; kmax – 20; Nc – 5  POPSIZE; m – 0.04 (and much as possible. IFDD-CSA can find the feasible
equal 0 during last 10 generations). optimum solution in the whole frequency range. It performs
In practice, number 10 is decided like this: we first let it more universally and efficiently. Moreover, if the frequency
equal half of MAXGEN. According to the results obtained is not in the range of 30– 50 Hz, the number of switching
with IFDD-CSA, if IFDD-CSA can easily have feasible time N2 , that is the variable number of the problem, will
optimum solutions, we may take a larger number. be recalculated. Then, by using the proposed method, we
Otherwise, like in this paper, a smaller number is taken, can also obtain the best solutions.
that is one-fifth of MAXGEN.

7 Conclusion

6.1 Optimisation with different methods In this paper, a new artificial immune system algorithm,
CSA is first used to find the switching time of a
The NS method and GA are also employed to solve the PWM-controlled, medium-voltage, variable-frequency
problem Optimisation with an Assigned Output frequency. drive. The IFDD is proposed to solve constrained optimis-
The switching time and motor voltage waveforms obtained ation. A comparison is made among NS-based PWM,
with IFDD-CSA and other methods at the same output fre- GA-based PWM method and IFDD-CSA-based PWM.
quency f ¼ 50 Hz are shown in Figs. 6 and 7, respectively. The study shows that the IFDD-CSA method is superior
In order to make the switch frequency fswitch
800 Hz to conventional methods and GA on several counts.
for high power converters, we let N2 ¼ 7, that is N ¼14 Meanwhile, experimental studies have also demonstrated
pulses per cycle; fswitch ¼ f  N ¼ 700 Hz. that the IFDD-CSA-optimised PWM technique is a promis-
As shown, the switching time via NS method generates ing method to converge to the optimum solution in a fre-
larger THD than other methods. quency range.
The GA is a search algorithm employing multiple concur-
rent search points and is well fitted to optimisation pro-
blems. However, with the same accuracy termination 8 Acknowledgments
criterion, the GA needs more iterations than IFDD-CSA,
so it takes more calculation time. It takes 35.921 s to This project was supported by the Program for New Century
obtain the same termination criterion, and the results are Excellent Talents in University and by the Excellent Young
averaged over 30 runs. IFDD-CSA spent, on the average, Teachers Program of MOE, People’s Republic China.
almost 3.705 s to estimate the parameters. The variation
of THD value at different iterative steps is shown in 9 References
Fig. 8 for the proposed method and GA. The Fig. 8 indicates
that the CSA-based method not only has a higher conver- 1 Brogan, P., and Yacamini, R.: ‘Harmonic control using an active
gence speed, but also converges to the global optimum drive’, IEE Proc., Electric Power Appl., 2003, 150, (1), pp. 14– 20
2 Narayanan, G., and Ranganathan, V.T.: ‘Synchronised PWM
with larger probability than the GA-based method. strategies based on space vector approach. II. Performance

IET Electr. Power Appl., Vol. 1, No. 6, November 2007 877

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.
 assessment and application to V/f drives’, IEE Proc., Electric Power 8 Shi, K.L., and Li, H.: ‘Optimized PWM strategy based on genetic
Appl., 1999, 146, (3), pp. 276–281 algorithms’, IEEE Trans. Ind. Electron., 2005, 52, (5), pp. 1458– 1461
3 Wang, D., Mao, C., Lu, J., Fan, S., and Chen, L.: ‘The research on 9 Sun, J., Beineke, S., and Grotstollen, H.: ‘Optimal PWM based on
characteristics of electronic power transformer for distribution real-time solution of harmonic elimination equations’, IEEE Trans.
system’. Proc. IEEE Power Engineering Society Transmission and Power Electron., 1996, 11, (4), pp. 612 –621
Distribution Conf. Exhibition, 2005, pp. 1– 5 10 Sundareswaran, K., and Kumar, A.P.: ‘Voltage harmonic elimination
4 Singh, B.N., Chandra, A., and AI-Haddad, K.: ‘DSP-based in PWM A.C. chooper using genetic algorithm’, IEE Proc., Electric
indirect-current-controlled STATCOM Part 2: Multifunctional Power Appl., 2004, 151, (1), pp. 26–31
capabilities’, IEE Proc., Electric Power Appl., 2000, 147, (2), 11 Ozpineci, B., Tolbert, L.M., and Chiasson, J.N.: ‘Harmonic
pp. 113–118 optimisation of multilevel converters using genetic algorithms’,
5 Sun, J., and Grotstollen, H.: ‘Pulsewidth modulation based on IEEE Power Electron. Lett., 2005, 3, (3), pp. 92–95
real-time solution of algebraic harmonic elimination equations’. 12 Du, H., Jiao, L., and Wang, S.A.: ‘Clonal operator and antibody clone
Proc. 20th Int. Conf. Industrial Electronics, Control and algorithms’. Proc. 1st Int. Conf. Machine Learning and Cybernetics,
Instrumentation, IECON, 1994, vol. 1, pp. 79– 84 2002, pp. 506–510
6 Maswood, A.I., and Wei, S.: ‘Genetic-algorithm-based solution in 13 Liu, F., and Yang, H.: ‘A clone based multicast algorithm with
PWM converter switching’, IEE Proc., Electric Power Appl., 2005, adjustable parameter’, J. Softw., 2005, 16, (1), pp. 145– 150
152, (3), pp. 473–478 14 Wang, Y., Wu, C., Hu, X., Mu, S., and Liu, L.: ‘Annealing-genetic
7 Hasanzadeh, A., Zolghadri, M.R., Kaboli, S., and Homaifar, A.: ‘A algorithm for constrained optimisation problems’, High Technol.
genetic algorithm based programmed PWM optimum switching Lett., 2004, pp. 10– 14
pattern calculation’. Proc. 5th Int. Conf. on Power Electronics and 15 IEEE Recommended practices and requirements for Harmonic
Drive Systems, 2003, pp. 1081–1085 Control in Electrical Power Systems, IEEE standard 519 –1992

878 IET Electr. Power Appl., Vol. 1, No. 6, November 2007

Authorized licensed use limited to: PONDICHERRY ENGG COLLEGE. Downloaded on April 14,2010 at 16:29:38 UTC from IEEE Xplore. Restrictions apply.