Modern Electrical Machine Design Optimization: Techniques, Trends, and Best Practices
Modern Electrical Machine Design Optimization: Techniques, Trends, and Best Practices
Modern Electrical Machine Design Optimization: Techniques, Trends, and Best Practices
net/publication/322947351
CITATIONS READS
13 6,490
4 authors, including:
J. Pyrhönen
Lappeenranta – Lahti University of Technology LUT
437 PUBLICATIONS 5,292 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by J. Pyrhönen on 02 April 2018.
S OLUTION
II. T ECHNIQUES FOR EVALUATING MACHINE DESIGNS
Fig. 1: Typical structure of an optimization process comprising main
Selecting a dedicated technique for modeling and evaluation stages (colored boxes) and exemplary tasks (white boxes); dashed
is crucial for the investigation and comparison of different lines give loops for possibly required iteration(s).
machine design candidates.
In the early beginning, analytical equations were applied
III. T ECHNIQUES FOR OPTIMIZING MACHINE DESIGNS
to characterize the performance of machine configurations by
relating electric, magnetic, and mechanical variables. Many The range of problems in terms of selected objectives
modeling strategies based on analytical equations were con- and parameters and applied techniques for design candidate
tinuously adjusted and improved based on experimental results modeling, evaluation, and optimization is very widespread.
such that the therewith predicted machine performance often Citing all different scenarios thus would go beyond the scope
is of high accuracy. On the other hand, those models usually of this article. Instead, the authors try to give the reader
focus on a single machine topology. Thus, the developed an overview about integral components and most popular
techniques lack universality. In addition, emphasis is laid on methods for defining and solving such kind of problems in
deriving accurate performance criteria at the price of not the field of electric machine design.
providing the opportunity to study local effects. Nevertheless, In Fig. 1 a typical structure of an optimization process is
those techniques are still very useful, especially if there is illustrated. To take respect of the diversity of the optimization
need for a very rapid evaluation of a huge number of different problems considered in the field, an abstract overall setup is
machine designs, e.g., [7]. presented. The typical work flow is indicated by arrows on
The nonlinearity of the magnetic circuit makes even the bold black lines and the four relevant steps are introduced as
electromagnetic optimizing tasks complicated and it is often
• definition of objectives and constraints,
not possible to analytically calculate the torque or losses in
• definition of the search space,
motors with sufficient accuracy for optimization. In order
to allow for investigating machine designs in more detail, • exploration of the solution space and the
finite element analyses (FEA) are adopted. Due to limits in • evaluation and interpretation of the results.
computational power, most often two-dimensional scenarios Practicing engineers know that very often solving an opti-
are applied. Thanks to dramatically increased performance mization problem is an iterative process instead of a purely
of computers, nowadays electric machines are often analyzed sequential procedure. Due to obtained findings, the problem
in 3D comprising the simultaneous investigation of different might be reconfigured, e.g., by reformulating the objectives
physical aspects at once, e.g., the combination of electro- and constraints, redefining the search space for machine design
magnetic and thermal aspects as well as mechanical stress candidates, or modifying the applied technique(s) for the ex-
[8]. Moreover, the analysis often comprises the entire drive ploration of the solution space. Most often, the reconfiguration
to consider interactions of control, power electronics and the is done by engineers themselves. However, at least some
machine design to find the overall best configuration [9]. Using reconfiguration parts can also be automatized, as for instance
FEA makes it possible to obtain sufficient accuracy with some a dynamic change of the design parameter discretization.
restrictions. For example, assessing iron losses accurately in Possible iterative loops are illustrated using thin dashed lines
FEA still builds a challenge. with arrows in Fig. 1.
In the following, crucial methods and techniques used for 1
solving optimization problems are itemized and explained.
This comprises basic mathematical definitions and equations. 0.9
Engineers often avoid studying pure mathematical explana-
efficiency
0.8
tions and definitions as they prefer gaining hands-on experi-
ence. Nevertheless, understanding the basic ideas of typically 0.7
applied techniques is important for feasibly configuring the
optimization process. To accomplish this balancing act, the 0.6 Pareto points /
fundamental mathematical principles are briefly described pro- non−dominated points
dominated points
viding references where they are adopted in the field of electric 0.5
0 0.1 0.2 0.3 0.4 0.5
machine design. This should enable the interested readers to relative cost
get a good overview about this particular field.
Fig. 2: Pareto front for an optimization problem with two objectives
relative cost and efficiency.
A. Definition of Objectives and Constraints
To efficiently run and solve an optimization problem and
definition is that standard optimization techniques developed
to obtain reliable results, it is crucial to correctly define
for single-objective scenarios can be applied. Moreover, design
the problem. This comprises the variables (continuous or
candidates are characterized by a single measure allowing for
discrete), the objectives to be maximized (or minimized) and
an easily comparison. On the other hand, the definition of
the conditions (constraints) under which the problem has
the weighting factors has a crucial impact on the results and
to be solved. The relationship of variables, objectives, and
objectives are very different, e.g., efficiency and cost. Hence,
constraints further allows for classifying problems.
defining these factors is a difficult task as each weighting
1) Single objective: For sake of completeness and as an results in different optimum and it requires wide experience
introduction, a general definition of a single-objective problem about the considered application field.
is presented. This comprises an objective function F (x) and Instead, most of the engineers today are choosing the right-
vectors for equality and inequality constraints, g(x) and h(x): hand definition presented above, i.e. a vector-based approach is
applied. This requires more complex optimization algorithms,
min F (x) subject to g(x) ≤ 0 and h(x) = 0 , which will be presented in section III-C, and normally the
x
runtime is increased. Typically some objectives are contradic-
where x is the design vector which specifies the design under tory in electric machine design. Thus, Pareto fronts of optimal
optimization. In electric machine design, typical objectives are solutions are obtained. A 2D-example is given in Fig. 2. As a
cost, efficiency, torque or power density, Back-EMF waveform, common optimization problem is about maximizing efficiency
and reliability regarding mechanical stress, thermal aspects, while minimizing cost, the upper and more left a point is
etc. Examples can for instance be found in [10], [11]. While in located, the better. The red crosses give the Pareto-optimal
the past a single operational point was investigated, nowadays points which often are called non-dominated points. The latter
analyses evermore comprise multiple load points, e.g., for designation implies that there is no solution that is better in
electric vehicles with particularly defined driving cycles or all objectives compared to a Pareto-optimal point. From this
other applications with operating cycles [12]–[20]. it follows that all the Pareto points are optimal and it is up to
All the listed quality indexes can also be considered as the engineer’s choice to select one of those designs by trading
constraints, e.g., motor efficiency must not be less than y % or cost vs. efficiency. By contrast, the blue crosses in Fig. 2 give
cost should not be more than z USD. Usually, some constraints dominated points, which are not Pareto-optimal.
are determined by the construction space, the maximum Even though some extra effort is required when using
operational temperature in specific parts of the motor (e.g., the vector-based multi-objective optimization approach, it
permanent magnets), the minimum value of the magnetic field gives beneficial ancillary information about the relationship
in the permanent magnets to avoid demagnetization, etc. of different objectives. The engineer can investigate these
It comes as no real surprise that optimizing electric coherences and can specify the weighting based on these find-
machine designs typically comprises the consideration of ings. The additional expense is especially justified if different
several performance measures. Thus, the multiple objectives weightings need to be analyzed. While the first definition of
approach must be adopted. the multi-objective problem necessitates a separate run per
each weighting factor set, using the second definition only
2) Multiple objectives: A multi-objective problem is de-
post-processing of the results needs to be run multiple times.
fined as
min qT F(x) or min F(x) ,
x x B. Definition of the search space
while the same definition of constraints as for the single- Besides selecting appropriate objectives, the definition of
objective case can be applied. the search space provides the basis for the results that can be
Using the definition on the left, introducing a vector of obtained. When optimizing electric machine designs, the def-
weighting factors q is required. This vector is of same size inition of the respective geometry is crucial. In the following,
as F which comprises all objectives. The advantage of this the two most prominent formulations are presented.
π ple of synchronous reluctance machines typically works best
pz
with a particular number of barriers and respective shapes
(rectangular or round) depending on the number of stator
slots, the stator winding configuration and the number of rotor
dr
e
poles. By contrast, maximizing the mean torque and thus the
er output power often requires a different number of barriers
d ri hPM and barrier shape. If the engineer fixes those two barrier-
related parameters in advance, the obtained solution space
might not comprise the best possible solutions with regard to
both performance parameters torque ripple and mean torque.
r
h
x3
0
soft magnetic materials in electric machines comprising soft
magnetic composites and cobalt iron. We also have novel −1 1
−1 0 x2
heat tolerant Samarium-Cobalt permanent magnet materials 0 1 −1
x1
(e.g., RECOMA [33]) that can operate at high temperatures
around 300◦ C and maintain high remanence. Besides better Fig. 5: Box-Behnken based selection of designs investigated in the
magnetic circuit materials, conductive materials much better sampling phase - blue circles give designs considered, while red
than copper seem to be within reach for the future development crosses denote parameter combinations that are ignored.
step of electric machinery [34]. For instance, carbon nanotube
materials at the moment feature the specific conductivity at the
of experiments. Originally, the development of these tech-
same level as aluminum. The big future promise in carbon-
niques was inspired from experimental practices. The max-
nanotube-based conductors is based on the fact that individual
imum number of experiments is always bounded to obtain
best nanotubes demonstrate conductivity levels higher than
results at reasonable effort and within acceptable time. Even
100 MS/m corresponding to resistivity levels of 1×10−8 Ωm.
though the simulation of a device is much faster and cheaper
These kind of conductors enable no practical space for eddy
than manufacturing and measuring a prototype, limitations
currents. As a consequence, a new machine design paradigm
regarding computational power and analysis time still apply.
is needed when they will be available [35]–[37].
Thus, DOE techniques are also utilized in electric machine
Cooling of the machine is always a central aspect that affects
design for simultaneously varying multiple design parameters
machine performance and material selection [38]–[40]. There
and retrieving the impact of each parameter on the objective
is no globally optimal cooling means for a machine. For in-
functions. If a suitable technique is applied, it allows for sig-
stance, the totally-enclosed fan-cooled squirrel-cage induction
nificantly reducing the number of required design evaluations
motor does not offer the highest possible efficiency, it is heavy
compared to a full grid analysis while providing comparable
and its power factor is low at least for multiple-pole machines.
informative content. Thus, the overall effort and runtime for
But the machine is cheap and satisfies many of the boundary
solving optimization problems can be minimized. Famous
conditions set by manufacturers and customers. Hence, present
DOE techniques are Latin-Hypercube-Sampling, Fractional or
solutions in practice typically constitute optimal designs with
Full Factorial tests, Central Composite Design or the Box-
regard to particularly set cooling options.
Behnken sampling [46], [47].
The definition of the materials involves a computationally-
Fig. 5 gives an example for the Box-Behnken approach.
efficient and highly-accurate formulation and evaluation of
A grid of three design parameters with three levels each is
their properties, including the impact of manufacturing [41]–
considered, thus analyzing all possible combinations would
[43] and the evaluation of associated losses [44].
require 33 = 27 evaluations. By contrast, with the Box-
Behnken approach less calculations are required, i.e. the
C. Techniques for efficiently exploring the search space / combinations defined by red crosses in Fig. 5. Any other result
obtaining the solution space is then calculated based on the obtained results and, e.g., a
consequently created surrogate model. Fig. 6 gives an example
Due to limitations in computational power, the number of
of Latin-Hypercube-Sampling. The design parameter ranges
evaluated candidates for a typical machine design process was
of x1 and x2 are discretized and every parameter section
rather limited in former times. Hence, usually a grid search of
is used once (or multiple times, but usually same frequency
a few most relevant parameters was applied for finding the best
is applied). The crosses in the figure denote the (randomly)
machine design. The amount of data was easily manageable
defined combinations of the sections.
and engineers normally did not care about topics like efficient
exploration of the design space.
Nowadays, this topic is highly relevant in machine design. 1.0
0.9
Due to (economic) competition, engineers try to find optimal 0.8
solutions for particular applications and try to consider 0.7
0.6
multiple design parameters, physical domains, and objectives
x2
0.5
and constraints. Thus, optimization problems are very 0.4
0.3
complex and techniques for exploring the parameter space 0.2
efficiently are developed by considering the minimization of 0.1
Fig. 9: Flow chart of the optimization process for the motor for a natural gas compressor.
IM (4poles)
200
PMSM 12 08 Cost1
175 PMSM 12 08 Cost2
PMSM 06 04 Cost1
costs [Euro] 150 PMSM 06 04 Cost2
125 IE2 IE3 IE4
100
75
Fig. 10: Prototype of the optimized machine design.
50
25
Embedded.NdFeB.Co
prefabricated windings and open stator slots. Semimagnetic 150 Embedded.NdFeB.Alu
SyncRM.Ferrite.Co
wedges offer a solution to mitigate permeance harmonics’ 125
SyncRM.Ferrite.Alu
SyncRM.NoMagnet.Co
effects on the solid rotor surface and, thus, minimizing the 100
SyncRM.NoMagnet.Alu