830 83'IEL TRANSACTIONS ON INDUSTRY APPLICAT'IONS, VOL. IA-16, NO. 6.

NOVEMBI R/L)ECI-mBfLR 1980

Optimum Output Waveforms in PWM Inverters

GIUSEPPE S. BUJA, MEMBER. 1-ELE

Abstract-A general method is considered for optimizing the output gested to utilize the conjugate gradient descent (CGD) imieth-
waveform of pulsewidth-modulated (PWM) inverters, whatever load od, because it exlhibits a very fast convergenice for probleins
may be supplied. It consists of the minimization of a cost function of
operation of the load fed by the inverter. The mathematical tools and
of this kind [91. The use of thle CGD method requires the
the numerical techniques to calculate the optimum switching patterns knowledge of the gradienit of the functioni to be minimized.
in the space of PWM inputs are derived using the state variable To this purpose an analytical expression of it has been derived
description of the dynamics of the system. Nonlinear state equations for PWM systemiis.
are assumed. Some examples and comparisons are included to show Finally, an exanmple of the synthiesis of an optimumi wave-
the features of the proposed PWM methods. form has beeni carried out on the digital computer witlh refer-
ence to an R load supplied by a single-phase lhalf-bridge
1. INTRODUCTION inverter whlichl includes an LC filter at its output.
pULSEWIDTH-modulated (PWM) inverters lhave been used
for many years not only for controllinig the amplitude of 11. GENERAL OPTIMUM PROBLENI
thie output voltage but also for affecting its harmonic conitent The output wavefore of a PWM inverter is assumiied to be
[11, [21. Several techniquLes of miiodulation have beeni pro- periodic anid with quarter- anid half-wave symminletry as slhown
posed for this purpose. Accordinig to the miianner in which the in Fig. l. The amlplitude is taken equal to ± 1, and the switchI-
techniques operate, they can be divided into analogic and digi- ing angles in the first quarter of the period are called y, with
tal teclhniques. In the formier techlniques the switching angles i=1, 2, ---, M, and
occur at the crossings of two waves, thus giving to the PWM 7T
wave a lharmonic content according to the rules of the sub- Mo<eoe It<is2<a-YAs <t
---

oscillation tlheory [31. In thie latter teclhniques the switchiing

angles are chosen to directly affect the harmiionic content: this
is done by elimliinating somiie hlarmonics in [41 anid [51 and by Moreover, it is assunlied that
minimizing a figure of current distortion in the load in [6] Tl
and [71. -YO =0, (2)l+
( Z)
=

These PWM digital teclhniques are able to significatively

improve the performances of the inverter-load systenms. The wave of Fig. I can be considered a series of step
However, their implementation requires the hanidling of a functions of amplitude equal to 2 but positive and negative
large quantity of data, and hence it has becomne practical one after the other. In the initerval [0, 7r they are placed at
onily recenitly, because of the availability of conmplete proc- 0, 71s, 72, -YMT M, "'T y-2, and 7 71. Thius
essinig unities in a single board, as in microcomputers [7], the equationi of the wave is of the formii
[8]. The experimental results obtained from early realiza-
tionis hiave allowed us to verify the advantages of using the U-= U(a)
PWM digital teclhniques, thlus puslhing to a tlmore general I ,~ ~ ~ ~ t
analysis of the possibilities offered by these techniques. U7r(a.) = ls(- 1 Yll -1 + 26(at) + 2 ,i(-1)i
This paper presents a general approach to the optimiza-
tion problemii of thie wavefornm at the output of PWM in-
verters. It refers to a cost functioni that is arbitrarily se- * [6(U i) 1(U -7T + 7i)l O < a < 7T ,

lected, but it is related to the dynaimiic behavior of the

state variables of the load supplied by the PWM inverter U2,r(a) = -i-uIT(a ir),
=
i.(a < 21T,
anid determiiines the switchiing anigles whlichl inimi-nize the
function. This very broad approachl covers all the cases of (3)
interest in PWM inverter-load systemiis aind includes the digital
techniques mentioned above. where 6 is the unit step function and s is equal to +I(- I)
To coinpute the minimiium of the cost funiction it is sug- if the output waveform is positive (niegative) in the interval
[yzl (iT/2) 1. In the following s(- 1 )M is indicated by a.
Paper SPCC 79-31, approved by the Static Power Converter Comi- The peak value of the fundamental coimiponent of the wave
mittee ot the ILLE Industry Applications Society for publication in of Fig. I is
this TRANSACTIONS. Manuscript released tfor publication July 22,
1980.
The author is with the Istituto di Elettrotecnica e di Elettronica, Ul= aL-[ 1 2 i (-I )i Cos yui (4)
Universita di Padova, Padova 35100, Italy.

0093-9994/80/1 100-0830S00.75 © 1980 IEEE

BUJA: OPTIMUM OUTPUT WAVEFORMS 831

+1

- 1
Y3 Y4 YM
Fig. 1. Pulsewidth-modulated waveform .
U1 is assumed to be always positive; to do this a shift of ir the optimization problem for the load with PWM input is to
of the wave of Fig. 1 must be done if (4) results in a negative. find the -y E S that minimizes (9). Of course y can be subject
The shift changes only the sign of a and hence that of Ul. As to additional constraints other than (1).
a consequence of this assumption the waveform of Fig. 1 is
completely specified only by means of yl, y2, * , yM- SEARCH OF THE MINIMUM
III.

Thus each possible PWM wave corresponds to a value of the The tool which will be used for finding the minimum of the
following vector: function of interest is the CGD method. This method con-
verges in M steps for quadratic function with M-dimensional
OY =[^Y 1, 72, , ZM] (5) domain and has a rate of convergence that is faster than other
methods. Unfortunately the function J defined in (9) is not
yeGSCRM necessarily quadratic, but since in the neighborhood of the
where S is the subset of RM (M-dimensional real space) for minimum J is expandable into a Taylor series with the second
which the constraint (1) holds. This can be expressed by writ- derivative positive and bounded, the CGD method applied to
ing (3) in the form J retains good properties of convergence [91 .

The purpose of the CGD method is to obtain the direction

u = U(a, y). (6) of the descent sequence by means of the gradient of the func-
tion. Therefore some expression of the gradient of J is needed,
Let preferably an analytical one. By calling this expression g, it is
defined by
d
-x =f(x, u); x(O) =
x0(Y) (7) d d
dcu j(p , Y)= de J:) = W(o,
de
eC

6=0
-Y), z), (10)
be the equation of load dynamics where the scalar input is
the function u given by (6). In (7), x is the N-dimensional state where (', ^) is the notation for the scalar product. On the
vector of the load and x(0) is the initial state. As it appears other hand, from (9) it results that
from (7), the fact that the initial state for loads operating in
steady state depends on -y is considered. In Appendix A the d/ d \
expression xo(y) for linear loads has been derived. The state deJ(0, -y) = hx [0, y)I, - X(:,,y)) (1 1)
of the system (7) forced by u from 0 to is given by de \ ~~~de
where
X(:3, y) = xo(y) + f(x, u) da. (8) a
hx[x(I, t)] =- h[x(3, y)I, k = 1, 2, ---, N, (12)
k aXk
It is assumed that the state is unique for each input.
If a cost criterion of operation of the load dynamics (7)
from 0 to ,B is assigned of the form dX(3, Y) = dedX(:,
de
eZ) (13)
J =
h[x(g,,y)] J(O,,y),
=
(9) By (11) (d/de)x is needed to evaluate the gradient. From
8-32 832 LLE FRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. IA-16, NO. 6. NOVEMBER/DFCEMBER 1980

(8) anid (6) it is withi f3 = 7. taking inito accouit ( 16)-(19),

x( 'y + CZ) = xo(y + CZ)
-d 7T'Y.r'7) = S 7FT, °) a vO0z) '2a(- I
+ 0 f[lX(ce -t + cz), u(a, -Y + cz)] da,

(14) lf'r Yi)AX[, U(,Yi y)]

and hence

r r
- 0(7T, 7r - yi)fu[x, U(ir - -yi, 7)] Z1. (21)
d d
- x(l3, 7)- dexo (y) + LJxf( LI:,
t)e,
de
de I
By substituting (21 ) into ( 1 1 ) the expression
d
+ f1,(x, u) -u(ac, 7) da, (15)
de j
T
de
J(r, 7)'Y 1i
h x
[X(, 7)] j (7T, 0) -xo(Y)
where

d a
(16)
- 2a (-)i{(n7. yj)f'J,[x, u('yi, 7)]
-xo (Y) = 2:
dE Lay37 xo(7)j zi. (22)
- (er, 7-i)fJ[x, u(iT-ti' 7)1 }zi
a
(fJ(X, U)k, )- -ax1 .fk(X, U), (17)
is obtained. Equation (22) shows that the gradient g is an M
vector whose ith component is
a
f(X U)k) = a fk(X, u),
- k-= 1 Al-, 1VJ,N (18)
au
gi =q (°) X(z 2a(-l)'Iq T(,Yi)fu[X, U(,Yi, Y)] _

-
U(c, 7)
de (23)
qT(7r -,Yi)fu[x, UOT -

yi, Y)l 1,
d ~~~~~Al where
- u7r(a) )=-2a i (-l )i
de I
q(,yi) OT (7T, yi)hx[X(7r, 'Y)l -
(24)
* [ 5 I(a - i) - I(a -'- 77 + 'yi)] ,

(19)
It is easy to prove that q(cv) satisfies the adjoint system of
(I 5):
d d1
U2 -Tu, v-7r(i-r), 7r <m a

dc de d T
-q (a)iI=-fx (x, u)q(a); q(7r) =
hx[x(7r, 7)1, (25)
da
6 is the uniit imilpulse funiction. Equation (1 5) is the integral
of the state equation of a load having (l/dLe)x as the state so that the construction of the gradient is even more simpli-
vector anid where the initial state vector, the systeim mnatrix, fied.
the input distribution vector, and the scalar input are given If the yi are subject to the constraint that U1 is equal to a
by (16)-(19), respectively. If 0(0, 0) is thie state transition given value ranging from 0 to 4/7T, the dimension of y be-
matrix of this system, the solution for (L/Ldeux is comes M -1 since yM is function of the other yi as follows:

d d yA arc cos b
de
X(3, y) -
00, 0) de
-xo(7) =

b = (-1)M (U 4-1(-I
)- cos7 (26)
±J] 0((, cv)fJ(x, U)-u(cv 7) dcv, (20)
0 dE
As a consequence of (26) the equations (16)-(19) and (21)-
wlhichl allows one to obtain the gradienit of J. For instance, (23) change. In particular, the ith component of the gradient
 BUJA: OPTIMUM OUTPUT WAVEFORMS 833

g results in the following: L

gi q
T
(0)
Faa
xo (y) (y)
+-X0yo
a7M
La-i
=

a-37 a7l v

-2a b(- I )qT(7yj)fu[x, U((y, y)]

Fig. 2. The circuit considered.
+( 1) q T(,M)fu [X U(YM, a)]a7iM|
This algorithm has been implemented on the digital coIml-
puter to search for the optimumn PWM waveform with refer-
- f-1)iqT(Tr - yi)fu[x, u(r yi, y)1 ence to the circuit of Fig. 2. The circuit consists of an R load
and of an LC filter; the imlpedance values at the work fre-
quency are R 1 Q2, XL = 0.66 Q, and Xc 2.64 Q. The
= =

+ (-JqQT -7M)fu[x, u(r-yM, y)] aI| voltage v is the output voltage of a single-phase PWM inverter
and has the waveform of Fig. 1, with s = I aiid M = 2. As cost
i=1, 2, *--,M -1. (27) function of the steady-state operation of the circuit, the
following is taken
From (26) the partial derivative of yM with respect yj is equal
to J( 2r 1 U da
=: C2
+ XLiL)d
LiL
)d2

a-M_ _ (_1)M+i I- sin

yi, i =
1, 2, ---,M 1.
a,yj-%,l-1
VC 2+ XL iL
b2 2
-T1
(X dce
(28)
The gradient (23) or (27) has been obtained using for u the (jXCis2 + XLiL dc
equation of the PWM waveform of Fig. 1. A voltage with this
)
,
(32)
waveforn is available at the output of a single-phase half-
bridge inverter, and therefore the previous analytical expres- being vC1 -iR. J is equal to the kilovolt ampere ratings of
=
sion of the gradient holds for this case. In Appendix B the the filter. Moreover, it is also bound to the rms value of the
expression of the gradient for the tlhree-plhase half-bridge current into the load. Since for a given VI the fundamental
case is derived. coinponent of the current into the load is given too, the
minimization of J implies the minimization of the rating of
IV. NUMERICAL SOLUTION AND RESULTS the filter and, at the samne time, a drastic reduction of the
The steps involved in the numerical computation of 'y,
current distortion into the load.
whichl minimizes J are as follows: The dynamics of the circuit of Fig. 2 as a function of a

are
1) guess a value of 'ycalled y1;
2) integrate (7) to get x'; d 1 1
3) hlavingx1(r), calculate q(ir) and hence solve (25); dIoL da VC XV
+ XL
XL
4) construct the gradient g' = g(yl );
5) test to see if g' is zero or sufficiently near to zero; d Xc
6) repeat 1)-5) until 5) is satisfied to the desired degree - VC = XCiL VC (33)
of accuracy, assumning for the next value of y cia R

d XL 2 1
2
yn+1 zn + C ;n
=
n= 1, 2, ---. (29) dca 3 VC2.
According to the CGD method, cn is chosen to iminimnize The state variable X3 has been introduced to express the cost
J(7n + cn1X), while >,n is set equal to function as in (9). As a matter of fact, it is
XI =-g1
(30) J(7) = X3(7T, -Y). (34)
at the first iteration and to
The initial values of the electrical state variables are evalu-
Xn n=
gn +
(g'gl n- ated from (41), while X3(0) iS taken to be zero. Moreover, h.
(31) is
given by

at the other iterations. hx [X(7r, Y) j [ 0, 0, I j T.

=
(35)
834 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. IA-16, NO. 6, NOVEMBER/DECEMBER 1980

lo~~~~~~~~~~

101

I0.5 1.0 VI
Fig. 3. JH versus VI for differenit modulations with s = 1 and M = I.

The additional constraint of a given value of VI is also puted. The results are shown in Fig. 5 by the continuous
assumed. curve. The dotted curves, on the contrary, give the values of
The results of the computation are shown in Fig. 3 by the JH at different iterations, having taken an initial guess that
continuous curve. Instead of the minimum of J, the minimum y = 0. The curves suggest that only three iterations must be
of carried out to reach the minimum with a good approximation.

1 Vc)-XL
( II) V. CONCLUSION
JH =J-- (36)
xc A modulation method suitable for use in PWM inverters
has been suggested, and the computation procedure of the
is shown as a function of V1 . In (36), Vc1 and IL1 are the switching patterns has been developed. The inethod exhibits
peak values of the fundamnental component of the capacitor the feature of being of extremnely broad application since it
voltage and of the inductance current. Also shown in Fig. 3 consists of the minimization of every cost function definable
are the values of JH for the harmonic elimination technique to the applicative conditions of the system fed by the inverter.
in the case of eliminating the third harmonic (dotted curve) Thus the load current distortion, the kilovolt ampere ratings
and the values of JH for the suboscillation modulation tech- of the filter or other significant parameters as well as a func-
nique (dashed curve). To accomnplish a significant comparison tion of one or of more of these parameters can be minimized.
it has been taken the same number of switching angles for Hence the optimum waveform modulation inethod can be
period for the three kinds of modulations. From the results considered as a tool for the design of the switching patternis
one can see the convenience of the optimum modulation, for PWM inverters.
especially for high values of VI. The switching pattern for The main limitation of the optimum modulation method
the optimum modulation is given in Fig. 4. concerns its practical implementation, which requires the
It can also be pointed out that the modulation with the buildup of a microcomputer apparatus for the inverter con-
third harmonic elimination is a particular case of the method trol, but the growing diffusion of the microcomputers smooths
of the optimum modulation and corresponds to the following the incidence of this factor.
choice of J:
J(3 APPENDIX A
A(Y) =V2r J
| (vV3 sin 30t)2 daK, (37) Let
d
- x =Ax+ Bu
where V3 is the peak value of the third harmonic of v. dO- (38)
Always with reference to the circuit of Fig. 2, the mini-
mum JH in the cases of s = 1 and of M = 3 has been com- be the dynamics of a linear system with the scaler input u
BUJA: OPTIMUM OUTPUT WAVEFORMS 835

Fig. 4. Switching angles versus V1 for optimum modulation of Fig. 3.

4 *10'40-
31 1 i T I 1
0 0.5 1.0
Fig. 5. JH versus V1 at different iterations of computation of optimum modulation with s = 1 and M = 3.

given by (3). For steady-state operations the state of the From (39) and (40) one may obtain
system at aO= 0 depends on y as follows. From the equations
M+1
x(,yl) = eA'Ylx(O) -a[I- e'4yl]A -'B, x(0) = -a(I + eA7r) (1 )i[eA (7-yi) + eAy-i ]
x(y2)=eA 72x(yl) + a[[-eA (Y2 --(1l)] A 1B,
-
(I - eA(,yi-,yi- O)]A -1
B. ( (41)
x(r) eAY 1xx(rT-1y) + a [I-eA'Y 1 A 1 B,
=
(39)
Equation (41) shows the dependencce of x(0) on y. From (36)
it is possible to get x(nT) as a function of x(O) only by means it also follows that
of substitutions.
The periodicity and the half-wave symmetry of u give a
x(Qr) = -x(O). (40) a(,-
A7i +eA4(7T-i)i]B. (42)
836 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. IA-16, NO. 6, NOVEMBER/DECEMBER 1980

APPENDIX B forr < jand

The line-to-line output of a three-phase inverter is equal to
a
U12 = ul 2(°a, T) a7i x0(y) 2a(-1)'
qT(0)T - -

= Ul (a, 7y)-U2 (a, )

(a, 27r , Y) , { ( 3 ))[( 3Y)]
=
ul (a"Y) -
uI --

(43)
3
+ qT( 27r3 _ 7) f[X ( 7, )]
where uI is given by (3). If yj is such that
1j ir
< ;
3,j 7Yi+i >- (44) -vq T(yi)fu [x, u(yi, y)] q(n-yi)f [x, U(T -yi7 v()
+
3
u2 can be put in the form (47)
for i >j.
U2 = U2(a, 7)
REFERENCES
2,1(<t, Y) = a |(ljl+ 2Ss ( 27- [1] B. Mokrytzki, "'Pulse width modulated inverters for AC motor
drives," IEEE Trans. Ind. Appl., vol. lGA-3, pp. 493-503, Nov./
Dec. 1967.
[2] R. Chauprade. "Inverters for uninterruptable power supplies,"
+ 2 zi (-1)'16(ct + 3--Yi) IEEE Trans. Ind. Appl., vol. IA- 13, pp. 281-297, July/Aug. 1977.
[3] A. Schonung and H. Stemmler, "'Reglage d'un motor triphase
reversible a l'aide d'un convertisseur statique de frequence com-
mande suivant le procede' de la sous-oscillation,- Rev. Brown
(a6 O ±+/i) -2(-1)2 [41
Boveri, vol. 51, pp. 555-576, Aoti,./Sept. 1973.
H. S. Patel and R. G. Hoft, "Generalized techniques of harmonic
elimination and voltage control in thyristor inverters: Part I-
Harmonic elimination," IEEE Trans. Ind. Appl., vol. IA-9, pp.
310-317, May/June 1973.
[5] -, "Generalized techniques of harmonic elimination and
*[8( 3- T) 6 (- +3i)3 voltage control in thyristor inverters: Part I1-Voltage control
techniques," IEEE Trans. Ind. Appl., vol. IA-10, pp. 666-673,
0 a<c<r, (45) Sept./Oct. 1974.
[6] G. S. Buja and G. B. Indri, "Optimal pulsewidth modulation for
U227r(0 7)")
= U2,r,(a -
7T, 7), feeding ac motors," IEEE Trans. Ind. Appl., vol. IA-13, pp.
38-44, Jan./Feb. 1977.
Ir<a <2 7r. [7] J. B. Casteel and R. G. Hoft, "Optimum PWM waveforms of a
microprocessor controlled inverter," in Conf. Rec. 1978 IEEE
By substituting the derivative of (45) in (20) one obtains Power Electronic Specialists Conf., June'1978.
[8] H. Le-Huy, "A microprocessor-controlled pulsewidth modulated
the gradient of J for the three-phase case. The ith component inverter," in Conf. Proc. 1978 fECI Industrial Applications of
of it is Microprocessors Conf., Mar. 1978.

2a71 +)
[9] L. Hasdorff, Gradient Optimization and Nonlinear control. New
York: Wiley, 1976.
gi= T(°) a X(,)
-
f-
(
27

Giuseppe S. Buja (M'78) was born in Padua,

Italy, on March 5, 1946. He received the Doctor
+(i,7)] + qT( _ degree in electronic engineering from the Uni-
* fu [X,
'Yi) versity of Padua, Padua, Italy, in 1970.
He joined the Institute of Electrical and
Electronic Engineering of the University of
Padua, Padua, Italy, in 1971. Since 1974 he has
'fu [X, U(ir_t Y)] -q i (7i)fu[x,u(7l, 7-)I been an Associate Professor at that Institute. His
major fields of interest are power electronics and
industrial ac drives. He is presently engaged in
the microprocessor control of inverters for UPS
+ qT (T -
i)fu [X, u(r -
ij, 'y)] tk'+)
{A A
and._for drives.
Dr. Buja is a member of AEI.