INTRODUCTION
S. RAHMAN, Member, IEEE F.C. LEE, Member, IEEE Virginia Polytechnic Institute and State University
The developmeiit of mathematical models suitable for minimum weight boost and buckboost converter designs are presented. The facility of an augumented Lagrangian (ALAG) multiplierbased nonlinear programming technique is demonstrated for minimum weight design optimizations of boost and buckboost power converters. ALAGbased computer simulation results for those two minimum weight designs are discussed. Certain important features of ALAG are presented in the framework of a comprehensive design example for boost and buckboost power converter design optimization. The study provides refreshing design insight of power converters and presents such information as weight and loss profiles of various semiconductor components and magnetics as a function of the switching frequency.
Manuscript received January 20, 1982. This work was supported by Subcontract G82313 CH8M to Virginia Polytechnic Institute and State University from TRW Defense and Space System Groups under NASA Lewis Research Center Prime Contract NAS32105 1. Authors' address: Electrical Engineering Department, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061.
00189251/82/09000598 $00.75
1982 IEEE
The advantages of a comprehensive power converter design approach were demonstrated previously on a buck converter and a halfbridge converter [1, 2]. The design allows one to identify a set of power converter design parameters which satisfies all design requirements and concurrently minimizes the converter weight and/or loss. This paper presents an extension of the previous work to design optimizations of a boost and buckboost converter. First, mathematical models are formulated for the design and operation of these two power converters. Based on these mathematical models, an effective nonlinear programming technique, augmented Lagrange (ALAG) penalty function algorithm, is selected to search for the optimal set of converter design parameters. The facility of ALAG nonlinear programming technique is demonstrated successfully for the boost and buckboost design. In addition, six variable scaling options are designed and implemented in the ALAG program to facilitate programming convergence and to expedite the rate of convergence for the complex and multidimensional nonlinear optimization problems. At the beginning the mathematical background of the ALAGbased nonlinear programming algorithm is presented briefly. This is followed by a discussion of boost and buckboost converter power circuits. A more general buckboost converter circuit is examined in detail. Various design requirements and physical operating characteristics of this converter are summarized in the form of equality and inequality constraints. The minimum weight design requirement is formulated as the objective function. For a detailed study of a boost converter see Lee et al. [3]. Following the presentation of the mathematical model, the programming requirements for using the ALAGbased algorithm are discussed. Special characteristics of an exterior point based optimization algorithm (like ALAG) are discussed here. The influence of scaling on the rate of convergence is also discussed. Impacts of variable scaling versus constraint scaling are examined. This results in a compromise selection of both scaling techniques in the algorithm. Finally, detailed results are presented for two sample problems of minimum weight design optimizations of a boost and a buckboost converter. The sample designs provide detailed values for each component (e.g., R, L, C, crosssectional area and number of turns of L, etc.) for various switching frequencies. These show the tradeoffs between weight and power loss as a function of switching frequency.
A. Power Converter OptimizationMathematical Model
The utility of a design optimization is to pinpoint the detailed converter design which meets given performance
VOL. AES18, NO.5
598
SEPTEMBER 1982
specifications and to achieve concurrently the minimization of a certain converter characteristic defined by the designer. The task may be represented as a mathematical model as follows:
P1: minimize f(X, Y), Subject to
,P(x, r)
f(x) + +[g(x), r]
XEE+
i= 1,2,...,p
j= 1,2,...,q.
Here X is an ndimensional vector representing power and controlcircuit parameters to be optimally selected. The components of X are values of R, L, C, the switching frequency, and the design details of the magnetic components such as core area, mean core length, permeability, wire size, number of turns, and turns ratio of multiplewinding magnetics. The E+ is the positive orthant of ndimensional Euclidean space E". The YE El represents the vector of constants related to component characteristics. These constants are known to designers. Examples include winding and core densities, winding resistivity, windowfill factor of the core, windingpitch factor (i.e., the ratio of the mean length of oneturn winding to the core circumference), transistor and diode conduction and switching characteristics, coreloss parameters, intended maximum operating flux of given magnetics, and equivalent series resistance (ESR) as well as energystorage characteristics of filter capacitors. The Z E E+ represents the vector of performance requirements to be met by the optimum design. Controlindependent requirements include inputoutput voltages, output power, maximum weight, minimum efficiency, source EMI, and maximum output ripple. Controlindependent requirements include regulator stability, minimum audiosusceptibility rejection, maximum output impedance, transient response subjected to a step change of the input voltage or the load. Further details of design optimization approach can be found in Rahman and Lee [4].
II. ALAGBASED OPTIMIZATION TECHNIQUE The augmented Lagrangian penalty function (or ALAG) technique is essentially a method for solving a constrained optimization problem using an unconstrained optimization method with the help of some transformation. In other words, this is a concept of minimizingf(x) (see P1) with an unconstrained optimization method while maintaining implicit control over the constraint violations by penalizing the augmented f(x) at points where the constraints are violated.
One approach to implementing the above is as follows. Choose 4) and a sequence rk such that x(rk) is determined some way and x(rk) > x* as k  oo. x* is the value of the vector x for which f(x) is minimized.
B. Penalty Function Technique
problem.
where *( ) is the augmented f(x), r, in general, is a vector of controlling parameters, and 4) is a realvalued function which imposes the penalty. However, the action of imposing the penalty is controlled by r. It may be observed here that with a suitable choice of 4. and its control r, one can use an efficient unconstrained algorithm (e.g., Broyden's quasiNewton method, see [5]) in order to completely solve the constrained
As mentioned earlier, the penalty function transformation technique imposes an increasing penalty on the augmented objective function as constraint violation increases. Various penalty functions and their modifications have been proposed in connection with the penalty function based transformation. Penalty functions utilized in the algorithms for our study are discussed in the fol
lowing.
equality constraints to alleviate some of the computational difficulties. The modified penalty functions contained controlling parameters 0 E Ep+q and o E E+P+q whereas the traditional penalty function included only one controlling parameter r E E+.. Powell [7] and Haarhoff and Buys [8] proposed algorithms in which the vector of parameters a is changed only when the rate of convergence is not satisfactory and e is changed every iteration so as to enforce constraint satisfaction. The important feature of this approach is that lI/ri is not required to tend to zero for the convergence of the algorithm. Rockafellar [9] proposed a suitable modification of the PowellHestenes augmented Lagrangian (ALAG) penalty function to solve the inequality constrained problem. Fletcher [10] proposed algorithms for solving the inequality constrained problem using the ALAG penalty function technique. The augmented Lagrangian penalty function for P1 is obtained by combining the PowellHestenes penalty function and the Rockafellar penalty function.
at about the same time, proposed very similar modifications to the traditional penalty function technique for
I

hj
A. Outline of a Transformation gi = min[(gi  A5/oi), 01, ai=AJ/,, VL For reasons of simplicity let us consider only the set of design variables X and the inequality constraints from P1. In the preceding function Ai/ori represents a penalizing
RAHMAN/LEE: COMPUTER SIMULATIONS OF OPTIMUM BOOST AND BUCKBOOST CONVERTERS
599
? 2 (ii) z
threshold for ith inequality constraint. Increases in :ri to enforce faster convergence reduce the penalty threshold level and lead to closer constraint satisfaction. For inequality constraints, when gi > 0, Ai (or Oi) is relaxed to zero. Otherwise it is changed so as to make the corresponding constraint active at the current solution X. The ALAG penalty function algorithms are based on the following duality results. Let X (A) = X (A, cr) be the unconstrained minimizer of (x, A, o) for specified A and Cr. Then the dual function at X(A) may be represented as d(A) = k[X(A), A, a]. The duality results may be summarized as follows:
A. Design Variables
There are 24 unknown variables representing the details of buckboost converter magnetic design. This circuit contains a twostage inputfilter, a twowinding energy storage inductor, a power transistor, a diode, and an output filter. The design variables are
RL, R2, Rp dc winding resistances of inductors LI, L2, Lp, respectively; note: Np = N, is
L1, L,
R,
ad/aAj
hj,
1, 2, ..., q
Lp
A2d [NTGlN
I3 I
0?]
ZA, Z,, Zp
N1, N,2, Np
where G is the Hessian of W with respect to X. S is a diagonal matrix with elements oi corresponding to inactive inequalities gi >, 0, and the columns of the matrix N are the gradients of equalities and violated inequalities gi < 0, at X(A). These duality results indicate that A should be varied so as to maximize d(A). For details regarding iterative methods for varying A and associated penalty function algorithm see [11].
used in the design example input filter damping resistor input filter inductors primary inductance of twowinding inductor filter capacitors core crosssectional area of inductors LI, L2, and Lp, respectively mean magnetic path length of inductors L,, L,, and Lp, respectively number of turns on inductors LI, L,, and Lp, respectively; note: Np = N, is used in this design example winding areas per turn for Ll, L,, and Lp, respectively transistor switching frequency overall operating efficiency.
It should be mentioned here that in order to separate the resonant frequencies of the two input filter stages the following relationship between LI and L2 is used: L1I/L2 = PE2; see [13] for details. III. BOOST AND BUCKBOOST POWER In order to perform the tradeoff study between CONVERTERS switching frequency, and power loss and weight, the In this section the circuit schematics (see Figs. 1 and variable F is held constant for each run of the optimiza2) for boost and buckboost converters are presented tion program. Therefore there is a net total of 22 along with a detailed examination of a more general variables in the design optimization program. buckboost converter circuit. The purpose is to define this converter design problem in such a way that the applicability of ALAGbased optimization technique B. Design Constants can be demonstrated. The inputoutput relationship and These design constants are obtained either through derivations of the constraints are given in [12]. The results of design optimization are presented in a later manufacturer's specifications or the designers' own experiences; numerical values are given in the parenthesis, section. all units are in MKS system:
RL
Fig. 1. Schematic of boost converter.
Eo
RL
I n
LE
VBE
BI DK VST
D, DC
TSR TSF
= (mean length per turn) (core circumference) (1.9) core window fill factor (0.4) conductor resistivity (0. 172 x 10 7 Q m) core density (7800) conductor density (8900) maximum operating flux density (0.4) weight per farad (DK,, DK4, DKS/210, 1100, 72) transistor saturation voltage drop (0.25 V) transistor emittertobase voltage drop (0.8 V) transistor turnon rise time (0.15 Ls) transistor tumoff fall time (0.2 ,us)

SEPTEMBER 1982
VD diode conduction voltage drop (0.9 V) diode turnon rise time (0.03 us) diode turnoff fall time (0.05 pis) diode turnoff recovery time (0.03 ,us) heat sink weight density (15.4 W/kg)
PIF
[P0/(effEl)J2 (RI
R2).
2. PQ = transistor saturation loss + base drive loss + transistor turn on loss + transistor turn off loss,
PQ
(POVST)/(eff El)
C. Power Converter Performance Requirements The following power converter performance requirements are specified. These performance requirements are employed in the next section to formulate design constraints.
E1EO/(2Lp(Eo
nE1)F))]
P0 output power (70 W) S frequencydependent source conducted interference (0.1 A); this specification limits the maximum percentage of the switching current being reflected back to the source to ensure that the source is not disturbed much by the switching action downstream VR output ripple factor (1 percent) which is defined as output ripple factor (percent) = (peaktopeak output ripple voltage)/(nominal dc output voltage) PEI inputfilter resonant peaking limit (2); the input filter peaking at its resonant frequency should be limited in order not to degrade the stability and the audiosusceptibility of the converter.
[E1EO/(2Lp(Eo + nE1)F)J}.
{[Po(Eo
+
nE1)/(eff nE1E4)J
[EIEOI(2nLp(Eo + nE1)F)]}
losses,
AC2N2fA2 +
2Acj,,NP0VXp)
[E52Ej/(12L2(EO
WC (capacitor weight) = DK3C3 + DK4C4 + DR5C, WS (source weight) = P0/(eff KJ) WH (heat sink weight) = Po(l  eff)/(eff K.,)
D. Design Constraints
+ [nE,/(Eo +
+ nE,)2F2)] +
(Eo + nE1)2P^/
+
(eff2n2E#E2)}n2Rp

[EIEo/((Eo + nEj)Np);
flux density, magnetic winding resistance, and some other converter properties. These are modeled in the form of constraints presented in the following: 1. Loss constraint C(l)
=
Highlights of the buckboost converter characteristics include power loss, core window area, core 5. PCAP
PCAP = [nE1R5/(EO +
nEi)1([E2Ej/(12n2Lp2 (Eo
nE1)Po(nE1E0 eff)]
0:

+ nE,)2F2)J {[(EO +

C(1)
PQ
PD  POF
(pdlEO)}2
Pj/(nE,EO)).
LI, L2, T, C(2) = C(3) =
601
C(12) = 0,
In constraint C(15), the sum of R, and R2 are limited to within a certain predetermined value RT (= 0.1 Q in the present examples). IV. SPECIAL FEATURES OF ALAG
C(9)
0,
C(9)
NpAp
(L1/BSp){[Po(Eo
+ nE1)/(eff EIEo)]
+ E1Eo/(2Lp(Eo +
nE1)F)}
C(7), C(8) =
(NiAci/Fw)05
i = 1, 2.
 Z1/2Tr + \/A7/2;
C(10) =
(2N0Acp/Fw)05
[P0(E0
+
 Z42T +
A2 .
C(11)
VR
+

nE1)/(effEIE6n)
nE1)C5F).
Ej1(2Lp(Eo + nE1)nF)]R5
Po/(2EO(Eo
UNCONSTRAINED MINIMA
I) (lID)
FEASIBLE REGION
(C4/C3)(2nF V7 3I )2]f
Fig. 3. Minimize xx2 subject to x + x2
0,
xI + x2
+ I
0.
D = R3(C/L1 )5.
12. Other inequality constraints C(14), C(15), C(16), C(17) > 0. The following constraints are employed to confine the design variables in certain reasonable ranges in order to facilitate fast convergence:
602
Penalty function methods can be computationally ineffective if certain constraints dominate others. Due to the wide ranging values of design variables and the complicated nature of the constraints such computational difficulties are not uncommon in the use of ALAG technique. Fortunately, the use of appropriate scaling can greatly alleviate this problem. Two such scaling methods (variable and constraint) are discussed below.
VOL. AES18, NO.5 SEPTEMBER 1982
Variable Scaling. In a switching power converter design, the values of design variables are scattered over a wide range. For example, the capacitance may be on the order of 10A and the switching frequency on the order of 105. These widely scattered values of variables are causes of convergence difficulties. Therefore the variable scaling technique is provided in the computer program to scale all the variables. For example, for xl = 0.5 x 10,x2 = 0.8 x 103, onecanuse VSCAL(1) = 107, VSCAL (2) = 102, SO that
START
CALL SCALES
That is, the largest difference between two values (from consecutive iterations) of any variable must be less than the tolerance required. It should also be mentioned that the program can also exit via a constraint convergence criterion discussed in the following. Constraint Scaling. It is very unlikely that the initial guess of starting point can satisfy all the constraints to the extent that each equality constraint residual is relatively smaller than the constraint tolerance and each inequality constraint is also satisfied. If that is the case, then we have already solved the problem without using the computer. In reality, based on the set of initial guesses, the constraint values can vary over a wide range. It is desirable to scale each constraint by a factor so that the effect of violating a given constraint is of the same order of magnitude as the effect of violating any other constraint. Unfortunately, there are so far no universal guidelines for selecting the constraint scaling parameters. It has been observed that faster convergence can be achieved by the proper selection of these parameters. However, improper use of constraint scaling can cause divergence problems. Details on the available scaling options are given in the next section. Whenever the maximum scaled constraint violation AKK(k) is less than AKMIN (constraint tolerance), program convergence is reached. The ALAGbased optimization program can be run by using variable scaling alone. However, judicious selection of constraint scale factor has been found to aid rapid convergence. Available Scaling Options. There are six different userselectable scaling options programmed in the
optimization model and simulation package. In each case, with the exception of first two, the subroutine SCALES (see flow chart in Fig. 4) determines the scale factors on the basis of the input information. The options are as follows:
1) No scaling; the variable and function values are used unmodified in the main program. 2) All scale factors are set by the user. 3) Scaled variable values lie within a user selectable range. No scaling for constraint and the objective functions. 4) Variable scaling is same as in option 3. However, scaled values of constraints and objective functions lie between 1.0 and 1.0. 5) Scaled variable values and scaled moduli of the constraint and objective functions lie within a userselectable range. 6) Variable and objective function scalings are same as in option 5. However, the squares of the scaled constraint functions lie within a user selectable range. The designer chooses the scaling option on the basis of input information. He must pay particular attention to the diversity of variable and constraint function values. Also, this option allows him to place varying amounts of emphasis on different quantities (e.g., variable, constraint, and objective functions). Fig. 4 presents a flowchart summarizing the major operations within the ALAGbased optimization algorithm used for this research project. Note that the optimization is carried out internal to the main subroutine ALAGA which in turn uses some other subroutines to perform the optimization.
603
20K
30K
40K
50K
100K
110K
0.934x10 51.336
0.272x10
120K
Al
Ni
0.324x10
50.613
0.288x10
44.098
0.254x10
41.895
0.226x10
0.202x10
40.050
0.174x10
43.077 0.323x10
0.938xl0
61.340 0.321x10
0.942x10 57.774
0.954x10 53.123
0.911x10
48.837
40.892
0.366x10
ACI 0.543x10
0.444x10
0.398x10
0.335x10
0.309x10
0.284x10
0.251x10
Zl
RI LI
A2
0.473xlOI 0.417xlO1 0.387x101 0.366xlO 0.346xlOI 0.340x10oI 0.345x10I 0.333xlO 0.315xlO1 0.306x10 0.291x0_I
0.699x10
0.698x10
0.167x10
0.695x10
0.141x10
0.694x10
0.702x10
0.729x10(
0.766x10
0.752x10
0.756xl0 0.712x10
0.754x10
0.768x10
0.217x10
0.123x10
0.108x10
0.103x10
0.805x10
0.763x10
0.671x10
0.619x10
0.203x104 0.166xlU4 0.145x104 0.128x104 0.125x104 0.167x104 0.121x104 0.113x104 0.103x104 0.966x105 0.0964x105
27.lb7
25.535 24.420
N2
AC2
Z2
24.121
21.543
14.872
15.952
16.099
0.29
16.440 6 0.284xl0
16.539
15.377
0.534xlO6 0.452xlO6
0.355xlO1
0.400xlO 6
0.370xlO6 0.336xlO
0.262xlO 0.297xlO 0.358xlO
0.295xlO
0.246xlO
0.342xl()
I
0.310xlU
0.234xlO 0.234xlO
7xl0
0.274x106 0.271x106
I
0.228xlOI
0.238xlO
0.222xlO 0.243xlO
R2
L2
AP
NP
0.300xlO1 d.72,xlO
0.327xlO 68.957
0.302xlO 1
0.304xlO1
0.305xlO
4 0.409xlO
0.271x1lOI
0.14Ox10
50.082
0.245xlO 0.224xlO
4
0.268xlO 4 0.254xlO
0.237xlO
4
0.206xlO
4
0.180x10 55.911
0.165xlO
54 317
.
0.160xlO
2 7 32 5
.
0.149xlO
27.883
0.146xlO4
29,712
0 131xlO
30.358
0.886xlO
44.22
5
ACP 0.172xlO ZP RP LP C3
0.452xlO1I
0.135xlO6 0.179xlO6
O.324xlO
0.175xlO
0.188xlO
0.181xlO
0.175xlO
0.301xlO
0.297xlOl
n.807xlO
0.156xlO
0.307xl0
0.299xlO1 0.314xlO0.384xlO
0.796xlO
0.148xlO
0.172x10)
0.179xlO4
0.931
484l
0.103x103 0.779x104 0.621x104 0.519x104 0.453x104 0.389x104 0.438x104 0.353x104 0.128x104 0 115xl(14 0.774x105
1.128
1.106
R3
C4
1.136
1.166
1.228
1.129
1.330
1.408
0.974
0.162xlO4 0.112x104 0.898x105 0.760xlU5 0.672x105 0.563>;10 5 0.836c105 0.788x1l)5 0.720xlO5 0.605x1)5 0.399x105
0.148x10 3 0.121x10 3 0.107xlO 0.8208 0.8238 0.8271
C5
E
0.985x10 018295
2 1187
907xO1
0.892xlO
0.8319 2 3252
2 1994
0.0567
2.4924
0.112xlO 0.8751
0.lUlxlO
0.8761 2.9771
0.881xlO 0.8783
3,1171
2.1182
0.868xl()
0.8751 3.3447
2.1492
0.819xlt)
0.8691
3.5515
2.2235
PQ 1.4679
PD
2.8014
1.9662
2.0944
2.1650
2.1445
2.0273
0 3188
PCAp 0.1958
PMAG 11.5281
PT 15.2862 WS
0.2637
0.2761
6.7959 11.7030
2 6529
0.3397
4.8831
9.9904
0.2964
4.1676
9 6993 2 5876 0 6298
0.3016
4 1959
0.3018
4.4609 10.5377
10.9421
14.9736
9.8383
14.3857
4.5803
9.9035
2 5943
14.6346
2 7479
0 9503
9.9915
2.5971
0.6488
2.7689
0.02992
0.02419 0.0501
2.7589
0.9723
0.02114
2.7398
2.7320 0.9186
0.01214
2.5971
2.6149
WHI 0.9925
WI WW
WC
0.9341
0.01394
0.01068
0.7603
0,01138
0.6487
0.847xlO
0.6431
0 793x10
0.6843
0 01678
0.762xlO
0.691x10
0.583xlO
0.01639 0.03744
0.03753
0.01275
0.03065
0.02634
0.2462 3.7249
0.1>2073
0.02646
0.01638
0.02336
0.01528
3.2759
0.01694
0.01467
3.2491
0.0153
0.01344 3.2747
0.0119
0.01234
WMAG. 0.0534
WT
0.02952 3.7583
0(.01995
3. 4 !3
3.8650
3.8062
3.6954
3.2886
3. 3234
V. OPTIMIZATION RESULTS OF BOOST AND BUCKBOOST CONVERTERS By treating the switching frequency as a constant in each optimization run, a set of converter parameter data
is
To facilitate comparison of the optimal converter designs between the boost converter and the buckboost
obtained.
converter
This set of data represents the optimum design under the specified switching frequenruns are
inststhewloss are weighT breakdo agarnst the switching freuwenght breakdowns are plotted
breakdowns
an boos.thconvereres ade bintionseutiedl
are shown
used.
con
cy. A number of
quency
between 20 kHz to 130 kHz in 10 kHz steps. Detailed optimization results for buckboost and boost converters are presented in Tables I and II, respectively.
604
below:
SEPTEMBER 1982
TABLE II
20K
0.469x10
30K
40K
Boost
50K
70K
80K
90K
100K
110K 0.175x10
120K
Ai
N1
0.438x10
79.345
0.436x10
79.098
0.358x10
69.350
0.243x10
0.459x10
27.441
0.279x10
26.907
0.229x10
27.109
0.236x10 19.767
0.208x10
0.174x10
22.771
94.297
22.550
22.798
ACI
0.334x106
0.405xlO1
0.697xl10
0.322x10
60.319x106
0.119xlO 6 0.856x1O
70.783x10
70.555x10
0.104xlO1
Zl
0.349xlO1
0.137xlO1
0.129x101
0.111X10 1.112xlO1
0.673x10
0.105xlO1
RI
0.676x10 I0.678x1O
0.711x10
0.710x10
O.562x10
Li
A2 N2
0.665x104
0.527xI0O4
46.461
0.523xI0
0.376xlO14
44.602
0.189xlO 4 0.IlJxlU
40.931xI0
50.203x10
10.216
0.306x10 5 0.250x10
48.282
50.247x10
46.648
50.102xl0
12.945
11.584
12.280
12.934
AC2 0.369xlO 6 Z2
0.351x106
0.274x0l 0.176x10
0.291xlO 1 0.276x10
R2
0.300x10
0.222x10 0.178x10 34.332
I0.275x1O I0.274x1O
0.106x10
60.952x10 10.100xlo
0.878x102
0.878x10 2 0.829x10
I
70.642x10 20.829x10
0.204x1O 1 0.200xI10
0.243xlO 1 0.256x10
0.267x10
I0.267xl0
50.187x10
0.318x10 76.534 0.285x10
0.139xlO 0.626
5
L2
AS
N5
0.174xlO4
34.494
50.310x10
45.681
0.127x104 0.108x10
37.577
0.902xl0
35.774
0.886x105 0.64xlO
35.111 45.280
0.592x10 5 0.433x10
74.055
0.440x10 5 0.331x10
70.519 76.611
AC5 0.398x10 6
Z5
0.329x106 0.295x106
0.309xlO1 0.282xlO1
70.816x10
0.185xlO1
0.1786
0.340xlO
0.199xlO
0.1353
.194x10
0.1648
0.161xlO1
0.510
0.154x10
0.588 0.317xlO
0.141x10
0.626
R5
L5 C3
0.477xI0
0.531x10
0.273x10
4
0.504x10 0.229x10
5
0.541x10
0.255xl0
0.220x10
5 5
0.245x10
0.246x10
0.133xlO
0.322xlO
0.162x10
0.318x10l 0.201x10.
0.162x10
5
0.239x10
0.320x10
0.200x10

0.244x10
205x
5
0.239x10
59x1 0
0.200xl
0.522x10
5
R3
C4
0.163xlO1
0.162x10
0.524x10
0.524x10
0.524x10
0.524x10
0.523x10 2 0.523x10
0.655x10 5 0.161x10
50.159x10
0.678x10
0.IlOlxlO
0.595x10
0.9462 0.5729
5 0.10OxlO 5 0.lOQxlO
50.lOOxlO
0.442x10
0.9336 0.7581 1.8622
50.100x10
0.378x10
C6
E
0.109x10 3 0.768x10
0.9397 0.3883 0.9465
0.4427
0.511x10
0.9389 0.6319 1.8243
0.468x10
0.9355 0.6945 1.8457
0.390x10 4 0.376x10
0.389x10
(
0.9467 0.5106
1.7776
0.8991
0.3362
1.9635
0.8929
0.9056 1.9918
8R21
n( RRli
1.0570 2.0354
PQ
PD
0.9916 2.020
1.7682
1.7725
1.7904
PCAP I.30x10
0.175x10
1.7262
0.198x10
1.6362
0.226x10
1.5969
0.277x10
2.0648
0.310x10
2.2546
0.333x10
2.3244
0.479x10
5.009
7.8568
2.5278
0.518x10
5.4482
0.538x10
6.2944
0.556x10
6.2469
PMAG2.3232
PT WS WH WI
WW
WC WMAG WT
4.4928
2.4186
0.2917
3.9589
2.4013
0.2571
3.9442
2.4007
0.2561
3.9829
2.4020 0.2586
4.5489
2.4204 0.2954
4.8259
4.9780
2.4343 0.3232
8.3979
2.5454 0.5453
9.30nn
2.5766 0.6078
9.3050
2.4;294
0.3133
2.5773
0.6101
0.5101
0.691x10 2 0.479x10
0.113x10
20.409x10 2
0.757x10
20.124x10 20.925x10
0.217x10
2.7644
0.843x10
3 0.603x10
0.799x10
0.554x10
0.573x10
0.554x10 3
0.834x10
3.445x1O
0.434x10
0.424xlO
0.179xlO 1 0.798x10
0.730x10
0.489x102 U.470xj0
0.256x10
2.7502
0.4J2x12 0.422xl2 0
0.145x10
3.0437
2436xlO
0.182x10
2.7465
0.131x10
2.6794
0.117x10
2.6759
0.385x10
2.7249
0.135x10
3.0962
0.102x10
3.1898
2.989x10
3.1930
PQ PD PCAP PMAG PT WS WH WI
total power dissipation in the transistor power dissipation in the diode power dissipation in the output filter capacitor total magnetic loss (core loss + winding loss) total loss source weight packaging weight magnetic core weight
WC capacitor weight WW winding (copper conductor) weight WMAG WW + WI total weight. WT The difference in simulation results between the boost and buckboost converter designs are summarized in the
following:
605
10 '
PT
8.
F
U)
10
U)6
cn
0
U) 8 U) 0
J4
F(KHZ)
40K
60K
F(KHZ)
80K
lOOK
120 K
3
Ws
cr
(9
2.0 _
3.5
WT
V.
(9
1.5 Z
"I
1.0
25
Ir
WH
05
<[ 2.0
WMAG
i
_1.5_
20K
40K
60K
80K
IOOK
120K
F(KHZ)
4) Magnetic losses PMAG (core loss + winding loss) for the buckboost converter are dominated by the 0.5 winding loss in low frequencies. The PMAG loss WMAG WC characteristic falls rapidly as the switching frequency inThe high magnetic losses in low frequencies creases. 120K 80K IOOK 60K 40K 20K cause severe weight penalty. It is clearly demonstrated in F(KHZ) 6 that in order to minimize the converter 5 and Figs. Fig. 6. Weight breakdown for optimal buckboost converter design. weight/loss, it is desirable to operate the converter at 1) The buckboost converter is heavier than the frequencies in the range of about 80 kHz to 100 kHz. boost converter. In order to have the minimum weight For the minimum weight/loss boost converter design, design the buckboost converter has to operate at a however, the optimal frequency lies in the range of 40 kHz to 60 kHz as shown in Figs. 7 and 8. higher frequency than the boost power converter. 2) Switching losses of semiconductor devices are higher for the buckboost converter. This is logical since VI. CONCLUSIONS the switching current amplitude is considerably higher Nonlinear programming techniques have been sucthan that of the boost converter due the same input and cessfully employed to generate computer simulations of output voltage and the same power level. minimum weight designs of switching power conthe con3) Magnetic components for the buckboost For power converter optimization, the ALAG verters. in and heavier in size weight. verter are generally larger
WH
10
606
SEPTEMBER 1982
package has been found to be quite effective in terms of the computation time, ease of coding, and rate of convergence. Adopting the ALAG routine, a cost effective computeraided design approach is presented which provides a minimumweight converter design down to the details of component level while meeting all powercircuit performance requirements. This computeraided design approach provides important design insights which help to assess the following important design concerns:
1) tradeoffs between weight and loss as the switching frequency is increased 2) optimum converter design down to the details of component levels 3) optimum component designs as a function of the switching frequency and their relationships to the overall system optimization
4) significance of the Ushape curves representing totalweight and totalloss versus frequency are observed in Figs. 5 through 8; this allows the designer to easily identify the optimum switching frequency or a range of frequencies over which the total weight/loss is minimum in the practical sense 5) impact of various critical component characteristics (viz. magnetic losses and switching losses of semiconductor devices) on the overall system 6) optimal converter topology for a given application. Employing the nonlinear program based optimization technique, the power converter designer can conceive the overall optimum system design. This is done by taking into consideration the powercircuit related performance requirements with the design objective of either minimizing weight, loss, or any other physically realizable quantity. It thus sets the stage for a more scientific design approach instead of relying on subjective bruteforce, trial and error, piecemeal design.
607
REFERENCES
[1]
Wu, C.J., Lee, F.C., Balachandran, S., and Goin, H.L. (1980) Design optimization for halfbridge dcdc converter. IEEE Power Electronics Specialists Conference Record, 1980, Atlanta, Ga. Yu, Y., Lee, F.C., and Triner, J.E. (1979) Power converter design optimization. IEEE Transactions on Aerospace and Electronic Systems, May 1979, AES15, 344355. Lee, F.C., Rahman, S., Wu, C.J., and Kolecki, J. (1981) A new approach to the minimum weight/loss design of switching power converters. Proceedings of POWERCON, Apr. 1981, Dallas, Tex. Rahman, S., and Lee, F.C. (1981) Nonlinear program based optimization of boost and buckboost converter designs. IEEE Power Electronics Specialists Conference Record, 1981, 180191.
[81
[2]
[91
[3]
[10]
[4]
[11]
[5]
[6]
[7]
Problems. New York: Academic Press, 1969, 143164. Powell, M.J.D. (1969) A method for nonlinear constraints in minimization problems. In R. Fletcher (Ed.), Optimization. New York: Academic Press, 1969, 283293.
timization. New York: Academic Press, 1972, 87106. Hestenes, M.R. (1969) Multiplier and gradient methods. In Zadeh et al. (Eds.), Computing Methods in Optimization
[12]
[13]
Haarhoff, P.C., and Buys, J.D. (1970) A new method for the optimization of a nonlinear function subject to nonlinear constraints. Computer Journal, May 1970, 13, 178184. Rockafeller, R.T. (1973) A dual approach to solving nonlinear programming problems by unconstrained optimization. Mathematical Programming, 1973, 5, 354373. Fletcher, R. (1974) Methods related to Lagrangian functions. In P.E. Gill and W. Murray (Eds.), Numerical Methodsfor Constrained Optimization. New York: Academic Press, 1974, 219240. Balachandran, S., and Lee, F.C. (1981) Algorithm for power converter design optimization. IEEE Transactions on Aerospace and Electronic Systems, May 1981, AES17, 422432. Lee, F.C., Rahman, S., Carter, R.A., Wu, C.J., Yu, Y., and Chen, R. (1980) Modeling and analysis of power processingphase III. Prepared by TRW Defense and Space Systems Groups and the Virginia Polytechnic Institute and State University, NASA Report, NASA CR16558, NASA LeRC, Dec. 1980. Yu, Y., Buchmann, M., Lee, F.C., and Triner, J.E. (1976) Formulation of methodology for power circuit design optimization. IEEE Power Electronics Specialists Conference Record, 1976.
Saifur Rahman (S'74M'78) was born in Dacca, Bangladesh. He graduated from the Bangladesh University of Engineering and Technology in 1973 with a B.Sc. degree in electrical engineering. He obtained the M.S. degree in electrical sciences from the State University of New York at Stony Brook in 1975 and the Ph.D. degree in electrical engineering from the Virginia Polytechnic Institute and State University in 1978. He has taught in the Department of Electrical Engineering, Bangladesh University of Engineering and Technology, State University of New York at Stony Brook, Texas A&M University, and Virginia Polytechnic Institute and State University, where he is currently an Assistant Professor. His industrial experience includes work at Brookhaven National Laboratory and the Carolina Power and Light Company. His areas of interest are large scale optimization, power system analysis, and alternative energy systems. Dr. Rahman is a member of the American Section of the International Solar Energy Society.
Fred C.Y. Lee received the B.S. degree in electrical engineering in 1968 from ChengKung University, Taiwan, Republic of China, and the M.S. and Ph.D. degrees in electrical engineering, both from Duke University, in 1971 and 1974, respectively. From 1974 to 1977, he was employed as a member of the technical staff of the Control and Power Processing Department, TRW Systems, where he was engaged in design, simulation, and analysis for spacecraft power processing equipment. In 1977 he joined the Virginia Polytechnic Institute and State University as an Assistant Professor in the Department of Electrical Engineering. He is presently an Associate Professor in that department. His research interests include power electronics, electric vehicle propulsion, nonlinear modeling and analysis, and design optimization of power processing components and systems. Dr. Lee has published over 60 articles in the areas of power processing modeling, analysis, and design.
608 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES18, NO.5
SEPTEMBER 1982