Académique Documents
Professionnel Documents
Culture Documents
I. INTRODUCTION
rejection, constraining the maximum overshoot of transient response, etc.). The resulting problem can be solved numerically
using nonlinear programming techniques.
In power electronics, especially in the design of power
supplies, there is often as much freedom in the selection of
the plant (power circuit) parameters as there is in the controller. Moreover, both the plant and controller are designed
by the same person, which means that it is quite possible
to design them simultaneously. Thus, CSO is well-suited
to power-supply design. This type of optimal simultaneous
design has not been previously reported in power electronics
literature.
In this paper, the feasibility of applying CSO to the design
of a power supply and its control is explored. An overview of
the method is given first, and then, the method is demonstrated
on the design of a buck converter.
II. CSO
References [3][6] give different approaches to CSO, with
emphasis on mechanical systems. Since CSO is a multidomain
optimization of the system, there are generally at least two
objective functions: one is related with the structure, and
the other is related with the controller. In most cases, the
well-known quadratic cost function of the linear quadratic
regulator (LQR) theory is used as the control cost, and the total
structural mass is used as the structural cost. The optimization
parameters are cross-sectional areas of truss members and
the
and
matrices used in the optimal controller. A
number of complex constraints are defined for the multidomain optimization problem, for example, minimum buckling
strength, closed-loop damping, and robustness bounds, which
are also functions of structural and control design parameters.
Finally, the design is optimized using nonlinear programming
methods.
The LQR-based method described above has some disadvantages. First, it requires full state feedback. Second, there
is no general way to choose the weighting matrices to satisfy
both the control and structural objectives/constraints.
Another approach to the CSO problem uses parameter
optimization to select the elements of the feedback gain
matrix or coefficients of the numerator and denominator of
the compensator transfer function. A parameter-optimization
method developed by Layton and Peterson [7][9] is the
most promising for power-converter design. Laytons method
is based on covariance control theory (CCT), which was
developed by Skelton and his colleagues [10][15]. CCT is a
method for the parameterization of all-stabilizing controllers in
GEZGIN et al.: DESIGN OF POWER STAGE AND CONTROLLER FOR SWITCHING POWER SUPPLIES
559
where
is the th initial condition and
denotes the
intensity of the disturbance.
, defined in (9), is shown to be the solution of the
following Lyapunov equation [12]:
(10)
where
diag
A. An Overview of CCT
In this section, a brief introduction to CCT is presented. Its
relevance to the CSO problem will be established in Section IIB. This discussion is limited to dynamic compensators, which
are widely used for the control of power converters. Other
control schemes, like static output control or state-feedback,
are a special form of the more general dynamic compensator
case. Let the plant and compensator be described as follows:
(1)
(2)
(3)
(4)
(5)
diag
If
is a controllable pair, then
is asymptotically
stable if
solves (10) [12]. Conversely, if
is a stability matrix, i.e., all eigenvalues of
are
stable, then (10) has a unique solution for all
[10], [16].
Note that (10) has two unknowns, the covariance matrix
and the compensator matrix
The strategy is to select a
and then solve for
To be feasible,
must be
feasible
symmetric and positive definite, and it must be assignable to
(6) by some
The necessary and sufficient conditions under
which
is assignable are given in the following theorem
from [11][13].
Theorem 1: A specified
is assignable to the system
if and only if
(6) by
(11)
(12)
(13)
denotes
the
(14)
(6)
(7)
(15)
(16)
(8)
(17)
The one at-a-time covariance of the closed-loop system is
defined as
(18)
(19)
(9)
diag
diag
(20)
where
denotes the state response of (6) when only the
th excitation is applied (while others are held at zero) from a
total of
excitation events and
is the complexconjugate transpose of
These excitations are either
initial conditions on plant and controller states or impulsive
that will
There may be an infinite number of controllers
assign a given
to the closed-loop system (6). The paramthat assign
is
eterization of all covariance controllers
given in the following theorem.
560
Theorem 2: Suppose
is assignable. Then, all the controllers, which assign
as a state covariance to the system
(6), are given by
(21)
(31)
is an
subject to:
over the design space and elements of
1)
;
2)
;
3)
;
4)
.
The first three constraints for the controller are the
assignability conditions of Section II-A. They guarantee the
existence of a feedback matrix , which will assign
to
the closed-loop system. Additional structural constraints on
the plant parameters, such as upper or lower bounds, may be
included in the optimization.
The constrained nonlinear optimization problem expressed
above must be reformulated in order to apply standard nonlinear programming techniques. The matrix
must be symmet, there are actually
ric, so with
free parameters in
Also, the constraint
has to be recast in a more standard form. Let
be a Cholesky decomposition for
The matrix is an upper
triangular with
nonzero elements; the
constraint
is equivalent to requiring that the diagonal
elements of
be positive. Thus, the design parameters are
taken to be the plant parameters and the nonzero elements
of
and the constraints are the assignability constraints and
constraints
for
where
is an arbitrary skew-symmetric matrix,
arbitrary matrix of appropriate dimension, and
(22)
(23)
(24)
(25)
Proofs of these theorems are given in [13].
B. Application to CSO
When CSO, based on covariance matrix parameterization,
is used, the design variables are composed of both plant
parameters, , and controller parameters, which are the elements of the covariance matrix
One benefit of covariance
parameterization is that the control cost can be expressed as an
explicit function of
For example, the following quadratic
function is related to the control energy:
(26)
(27)
(28)
Similarly
(29)
A. The Plant and Controller
(30)
is related to the system performance [12], [15]. The vector
contains the error variables that are important for
the designer, and
is a positive definite weighting matrix.
The smaller the error variables, the better the performance. In
subsequent sections, (28) will be called control cost, and (30)
will be called performance cost.
In the general CSO problem, the objective function is
composed of both structural objectives (weight, price, etc.),
which are denoted as here, and controller-related objectives,
(28) and (30). The overall cost can be defined as the weighted
sum of these objectives. The boundaries of the feasible region
are determined by equality and inequality constraints. Usually,
the three assignability constraints of Theorem 1 are the only
equality constraints. Various structural objectives and closedloop stability are represented as inequality constraints.
B. CSO Formulation
For the simultaneous optimal design of the plant and controller, different objective functions can be used. For example,
losses in the power stage are a good candidate for
structural objective. Another choice is the efficiency. Both
of these objectives can also be used as side constraints. For
example, during the design process, losses can be constrained
to be below a certain level. Equivalently, efficiency can be
GEZGIN et al.: DESIGN OF POWER STAGE AND CONTROLLER FOR SWITCHING POWER SUPPLIES
561
which are given in (28) and (30); and are the weighting
constants. Since and are scalars, the weight matrices in
(28) and (30) are set to
For the CSO problem, the
constraints are imposed on the structural variables,
and
Duty
Duty
Duty
and
562
C. Design Procedure
is
(32)
Duty
For the buck converter, the CSO problem now can be cast
as follows.
Minimize
(33)
(38)
Duty
and
(35)
The minimum value for
then becomes
Duty
(36)
Duty
Duty
where and
are given in MOSFET data sheets.
The total losses in the circuit become
(37)
and constraint becomes
appropriately.
, where
Duty
Duty
is chosen
optimization toolbox.
GEZGIN et al.: DESIGN OF POWER STAGE AND CONTROLLER FOR SWITCHING POWER SUPPLIES
563
(a)
(b)
Fig. 2. Comparison of benchmark design and CSO design. (a) Bode plot for the linearized benchmark circuit. (b) Bode plot for the linearized CSO circuit.
564
EFFECT
TABLE I
INCREASING
OF
OF
TABLE II
EIGENVALUES FOR TWO CASES
(a)
(b)
Fig. 3. Comparison of benchmark design and CSO design. (a) Impulse
response of the linearized benchmark plant. (b) Impulse response of the
linearized CSO plant. (c) Closed-loop impulse response of the linearized
benchmark circuit. (d) Closed-loop impulse response of the linearized CSO
circuit.
GEZGIN et al.: DESIGN OF POWER STAGE AND CONTROLLER FOR SWITCHING POWER SUPPLIES
565
(a)
(b)
(c)
(d)
CSO design). (a) Output-voltage transient during startup and disturbance. (b) Inductor current
566
REFERENCES
[1] J. Cheng, G. Ianculescu, C. S. Kenney, A. J. Laub, J. Ly, and P. M.
Papadopoulos, Control-structure interaction, IEEE Control Syst. Mag.,
vol. 12, pp. 413, Oct. 1992.
[2] Y.-P. Harn, G. M. Kabuli, and R. L. Kosut, Optimal simultaneous
control structure design, Proc. American Control Conf., vol. 2, 1991,
pp. 14421447.
[3] M. Milman, M. Salaman, R. E. Scheid, R. Bruno, and J. S. Gibson,
Combined control-structure optimization, Computational Mechanics,
vol. 8, pp. 18, Aug. 1991.
[4] P. G. Maghami, S. M. Joshi, and D. B. Price, An integrated controlstructures design methodology for a flexible spacecraft, Advances
Astronautical Sci., vol. 78, pp. 369385, 1992.
[5] M. Salama, J. Garba, L. Demsetz, and F. Udwadia, Simultaneous
optimization of controlled structures, Computational Mechanics, vol.
3, pp. 275282, Mar. 1988.
[6] I. Kajiwara, K. Tsujioka, and A. Nagamatsu, Approach for simultaneous optimization of a structure and control system, AIAA J., vol. 32,
pp. 866873, Apr. 1994.
[7] L. D. Peterson and J. B. Layton, Multiobjective structural control design by optimizing the closed-loop covariance, AIAA/ASME Structures,
Structural Dynamics and Materials Conf., vol. 1, 1991, pp. 534541.
[8] J. B. Layton and L. D. Peterson, Benefits of integrated control/structure
design using covariance controlA comparison of two approaches,
AIAA/ASME Structures, Structural Dynamics and Materials Conf., 1993,
pp. 31263136.
[9] J. B. Layton, Control/structure covariance optimization including performance robustness, Proc. SPIEThe Int. Society for Optical Eng.,
vol. 2442, 1995, pp. 7285.
[10] A. Hotz and R. E. Skelton, Covariance control theory, Int. J. Control,
vol. 46, 1987, pp. 1332.
[11] R. E. Skelton, Control of state and input covariances for dynamic
systems, Proc. IEEE Conf. Decision and Control, vol. 3, 1988, pp.
19021907.
[12] R. E. Skelton and M. Ikeda, Covariance controllers for linear
continuous-time systems, Int. J. Control, vol. 49, 1989, pp. 17731785.
[13] K. Yasuda and R. E. Skelton, Covariance controllers: A new parameterization of the class of all stabilizing controllers, Proc. American
Control Conf., vol. 1, 1990, pp. 824829.