Posynomial Models of Inductors for Optimization
of Power Electronic Systems by Geometric
Programming
Andrija Stupar*, Josh A. Taylor**, Aleksandar Prodic*
*- Laboratory for Integrated SMPS; ** - Institute for Sustainable Energy
University of Toronto
Toronto, Canada
a.stupar@utoronto.ca
Abstract- When optimizing power electronic converters for
multiple objectives, such as power density and efficiency, the
optimization of the magnetic components is often the most
challenging and time-consuming task. In order to perform
optimization quickly and efficiently, it would be advantageous
to formulate converter optimization as a geometric program,
a proven convex optimization method. In order to optimize
for losses and volume via a geometric program however, all
loss and volume models of the various components must be
in the form of posynomials. While some loss models, such as
those of semiconductors, are naturally in posynomial form or
easily transformed, this is not the case for inductors. This paper
presents a derivation of posynomial loss, volume, temperature,
and saturation models for families of inductive components, based
both on simulation and on adapting familiar analytical models
into approximate posynomial form. The terms of the derived
posynomial models are the inductor design variables, such as the
number of turns, the air gap, and so forth. This allows inductors
to be optimized for multiple design objectives as a geometric
program.
I. INTRODUCTION
Inductors account for a major part of the volume and losses
of power electronic converters [1], [2], [3], [4], [5], and special
care must be paid to their design when optimizing convert-
ers for efficiency and power density. In fact, the magnetic
component optimization sub-problem can be said to be the
most complex and challenging part of converter optimization
[2], [3], [6]. Meaningful optimization of magnetic components
depends on accurate calculation of their losses under diverse
operating conditions. The accurate calculation of core losses is
especially challenging, and the entire problem is very complex
[7] as the effects of all electric and magnetic operating
conditions (including DC pre-magnetization) must be taken
into account together with core and winding temperature, and
the thermal cross-coupling between the core and the windings.
A diverse literature on inductor optimization exists, focusing
either on analytically-derived models, such as in [1], [3], [8],
[4], measurement-based simulation models [7], [6], and finite-
element simulation models [9], [10]. [9] and [10] are notable
for their use of genetic algorithms for optimization, while the
remainder utilize mostly exhaustive search.
978-1-5090-1815-4/16/$31.00 ©2016
To the authors' best knowledge, no publications have at-
tempted to frame inductor optimization for power electronics
as a convex optimization problem. Convex optimization is
an extremely powerful framework for the solving of diverse
optimization problems [11], and is a field which has been
extensively studied. Powerful software solvers for convex
optimization problems exist, such as [12], which allow the
computation of optimal solutions efficiently and quickly. Al-
though many real-world engineering problems are inherently
non-convex, they can often be closely approximated by convex
problems or divided into convex sub-problems [11].
Geometric programming, which has been used to to solve
circuit-sizing problems [13], [14], [15] is a promising convex
optimization approach for use in power electronics [16].
However, geometric programming operates on special types
of polynomial functions, called posynomials. Loss models
for inductors are not inherently posynomial, and therefore
the purpose of this paper is to explore and demonstrate the
derivation of posynomial models for inductor losses, inductor
volume, and inductor temperature.
In Section II, geometric programming is briefly presented.
In Section III, appropriate posynomial functions that can be
used to model inductors are discussed and a general fitting
procedure is presented. In Section IV, posynomial models of
inductors are derived from accurate and comprehensive simu-
lations. An alternative approach, in which posynomial models
are derived from standard well-known analytical equations, is
given in Section V. Examples of posynomial models fitted to
data and the accuracy of the fits are given in both sections.
Section VI presents conclusions.
II. GEOMETRIC PROGRAMMING
A geometric program (GP) has the following form
n
minimize fo(x) = L (aolxl kOLlXlol2 ...X
n
k
O ln
) (1)
1=1
n
subject to fi(X) = L(allxlkill ...xnkiln)::; l,i = 1, ... ,m
1=1
hj(x) = ajXlkjlx2kj2 ...xnkjn = l,j = 1, ... ,p