You are on page 1of 37

Chapter 14

OPTIMIZATION DESIGN OF SURFACE PMSM


Contents
14.1. The design theme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
14.2. Electric and magnetic loadings . . . . . . . . . . . . . . . . . . . . . . . . 2
14.3. Choosing a few dimensioning factors . . . . . . . . . . . . . . . . . . . . . 3
14.4. A few technological constraints . . . . . . . . . . . . . . . . . . . . . . . . 4
14.5. Choosing the magnetic materials . . . . . . . . . . . . . . . . . . . . . . . 4
14.6. Dimensioning methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 6
14.7. Optimal design with genetic algorithms . . . . . . . . . . . . . . . . . . . 18
14.8. Optimal design of synchronous permanent magnet motor using Hooke
Jeeves method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
14.1. The design theme
In general, the following specications make up the design theme of a variable speed
PMSM:
- Base continuous power P
b
- Base speed n
b
.
- Maximum voltage V
n
- Overload factor k
l
.
- Maximum speed n
max
.
- Power at maximum speed P
max
.
- Number of phases m
Also a few constrains are added:
- efciency at P
b
and n
b
14-2 Electric Machines
- material insulation class (allow temperature)
- protection degree against alien bodies
- total initial cost of active materials
Flux weakening (n
max
/n
b
> 1.5) is not very feasible with surface PMSM, (unless its
synchronous inductance L
s
is large in p.u., such as for tooth-wound stator windings) and
thus n
max
= n
b
.
14.2. Electric and magnetic loadings
Choosing the electric and magnetic loadings consist the rst stage in general or optimization
design of electric machines. For SPMS they are:
a. Specic electric loading J
l
(A/m) represents the total effective value of the am-
pere turns in stator slots per stator periphery length. J
l
relates to thermal loading
and torque density. Large values of J
l
lead to large torque density, and consequently
to smaller machine size, which may end up in machine over heating and lower ef-
ciency. When high torque density is target in the design, the specic tangential force,
ftp, concept may be preferred. It is measured in N/cm
2
and varies from 0.1N/cm
2
in micro-motors to 10N/cm
2
in larger torque density. J
l
and ftp are complementary
concepts.
b. PM air-gap ux density: Bag (T), varies from 0.2T in micro-motors to 1T in large
torque density design. Together Bag, J
l
, ftp, determine the volume of the machine
for given base torque (Teb).
c. Stator tooth ux density: Bst(T) determine the magnetic saturation degree in the
machine; it varies from 1.2T to 1.8T in general, for silicon laminated stator cores. As
the machine magnetic air-gap (which includes the surface PM thickness) is large the
teeth magnetic saturation inuence is rather small and thus Bst may be chosen rather
large; in contrast, if B
st
is large (and the fundamental frequency f
1b
is large) the core
losses would be large. Smaller values of B
st
lead to wider teeth and thus thinner
slots, that is larger design current densities and consequently, larger copper losses.
To prevent this, deeper slots are adapted, but then the machine external diameter
(and volume) is increased. Moreover the copper losses may not decrease notably by
deepening the slot as the average diameter (and length) of end connection of coils
increase. So the design should start with moderate to large values for B
st
and then
reduce or increase them according to obtained performance in comparison to the
design theme.
d. The stator yoke ux density B
st
(T) is again chosen as a compromise between
magnetic saturation level and core losses limitations. Small values of B
sy
may lead
to a larger machine size and weight, especially if the number of poles is small (2p
=2,4).
e. Current density J
s
(A/mm
2
) determines the copper losses and the copper volume.
Small values of J
s
(2 3.5A/mm
2
) corroborated thin and deep slots may lead to
Optimization design of surface PMSM 14-3
high leakage inductance, and machine volume (and weight). On the other hand, high
J
s
values (> 8A/mm
2
) in general not only imply forced cooling but also lead to
lower efciency, while reducing the machine volume.
f. Rotor yoke ux density B
ry
(T) is imported in machine with a large number of
poles (and large diameter) when the PMs are not any more placed directly on the
shaft.
14.3. Choosing a few dimensioning factors
a. Machine shape factor
c
: the ratio between the axial (stack) length of the machine
lstack and the pole pitch (
c
= l
stack
/, usual between 0.15 and 3)
b. PM pole span factor
PM
: ratio between PM width and pole pitch. This factor
inuences the PMtotal weight, the emf space harmonics content, and, to some extent,
the cogging torque.
c. Number of slots per pole per phase q
1
;
d. Number of winding layers n
L
(n
L
= 1 for single layers and n
L
= 2 for two layers
windings);
e. Stator winding current path count a
1
, is always a divisor of number of poles for
two layers winding and of the number of pole pairs for single layers windings. To
avoid circulation current between current paths, due to inherent machine symmetry
imperfection, a
1
= 1 whenever possible, with the exception of low voltage large
current (automotive) applications.
f. Elementary conductors in parallel: for large currents (or fundamental frequency)
the coil turns are made of multiple elementary conductors in parallel with same de-
gree of transposition to reduce skin effects.
g. Coil span x
s
is the distance between forward and return sides of coil; it may be
measured not only in mm but also in number of slot pitches: between 2q
1
and 3q
1
for
3 phase distributed windings, but below 0.5 for tooth wounded windings.
h. Slot opening: s
o
(g. 14.4), Its minimum value is limited by the possibility to
introduce the coils, turn by turn, in the slot and by the increase of slot leakage induc-
tance and PM ux fringing its maximum value is limited by the PM ux reduction,
cogging torque increase and torque ripple.
i. Tooth top height h
s4
is limited down by technological (and magnetic saturation)
constraints and upper by the increase slot leakage inductance. (in high speed PMSM,
with f
1b
> 1kHz, h
s4
may be increased, to increase machine inductances L
s
and
thus decrease current ripple).
14-4 Electric Machines
14.4. A few technological constraints
a. The laminated core lling factor, k
stk
= 0.8 0.95, is the ratio of laminations
height to total laminated core length (laminations insulation coating means k
stk
< 1).
b. Slot copper lling factor, k
sf
= 0.33 0.7; lower values corresponds to semi-
closed slots with coils introduced turn by turn in the slot while larger values corre-
sponds to open slot and pre-made coils made of rectangular cross section conductors.
c. Minimum air-gap gmin from mechanical considerations in super high speed ma-
chines gmin is increased to reduce PM eddy currents losses due to stator mmf space
harmonic; a resin coating of PMs is also added for mechanical rigidity;
d. Overload demagnetization safety factor k
sPM
: the max overload stator mmf that
avoids PM demagnetization in worst load scenario, is limited by this factor;
e. Slot geometry type: if more than one slot shape is used, design expression for cho-
sen shape should be provided and the computer code calls for the specic geometrical
and specic magnetic permeance expression on a menu basis.
f. Angle of slot wedge,
w
;
g. Slot insulation thickness;
h. The straight turn end connection, l
f1
, that exits the slot;
i. Difference between PM, l
PM
, and stator stack length, l
stack
, (l
PM
l
stack
=
(1 2)(g +h
pm
), where h
pm
is the PM thickness (along magnetization direction),in
order to avoid axial forces on the bearings;
j. Minimum shaft diameter, based on the maximum envisaged torque.
14.5. Choosing the magnetic materials
We refer here to the permanent magnets and the magnetic cores. For the PMs we choose
the remnant ux density B
r
, coercive eld H
c
, and the maximum allowed rotor (PM)
temperature T
r
. The choice of B
r
depends on the air-gap ux density B
ag
< B
r
for
SPMSM. The coercive eld (force) H
c
depends on B
r
and on the recoil p.u. permeability:

rec
(1.051.3) for NeFeB, SmCo5, and ferrite PMs (B
r
= (0.30.4)T),
rec
1.05,
B
r
= 0.6 0.8T for bounded PMs, and B
r
= 1.1 1.37T for sintered PM at 20

C.
H
c

B
r

rec(pu)

0
(A/m) (14.1)
The PM characteristics are obtained from the manufacturers catalogues and they in-
clude:
B
r
at 20

C,
H
c
at 20

C,
Optimization design of surface PMSM 14-5
B
r
and H
c
temperature coefcients,
Energy density,
Interior coercive elds (force),
PM magnetization eld (1.8-3)H
c
, (1.8-2.5)B
r
,
Maximum operation temperature,
Electric resistivity of PM material, to calculate the PM eddy current losses due to
stator mmf space and time harmonics: much higher with bonded NeFeB than for
sintered NeFeB,
Mass density (kg/m
3
).
For up to 150Hz maximum fundamental frequencies general silicon laminated cores made
of 0.5mm (0.65mm for increased productivity) un-oriented grain laminations are used. For
fundamental frequencies above 150Hz, thinner laminations (0.2mm or less but with stack-
ing factor k
sk
> 0.8) are to be used to reduce eddy currents and losses. The magnetization
curve characterizes the silicon lamination (Fig. 14.1).
Figure 14.1. Magnetization curves for a few typical soft magnetic materials
Using soft magnetic materials with higher saturation ux density (B
sat
= 2.35T at
H
sat
= 10
4
A/m for frequencies 50Hz) leads to more compact PMSM at slightly higher
initial cost (hyperco 50 is more expensive than M19 silicon steel).
Other important characteristics of soft magnetic materials are the core losses which depend
on ux density amplitude and frequency saturation. The core losses are segregated into
eddy current and hysteresis losses. Some manufactures offer core losses for a few frequen-
cies (Fig. 14.2)
M19 and SURA are two widely used trade names for silicon laminations. Hyperco50 is a
14-6 Electric Machines
trade name for a high saturation ux density laminated material with up 50% cobalt and
moderate losses at frequencies below 500Hz. Somaloy is the trade name of a soft material
composite which is for f
B
> 500Hz (in general) and (or) where three dimensional a.c.
magnetic eld lines have to be accommodated.
Thin lamination such as SURA007 (0.18mm) shown acceptable losses at 2500Hz (Fig
14.2). Magnoval is a sintered material with low relative magnetic permeability (
rec
=
(3 5)
0
) used for slot wedges to decrease PM ux density pulsations cogging and total
torque pulsation, at the cost of larger leakage and synchronous inductance. Magnoval is
good also to increase the machine inductance in high frequency f
1b
> 1kHz machines to
reduce current ripple.
Figure 14.2. Iron losses versus ux density
A good approximation of core power losses (in W/kg) for the practical frequency and
ux density is, still, the Steinmetz formula:
dP
tot
= k
h
B
2
m
f +
2
d
2
6
(B
m
f)
2
+ 8.67k
e
(B
m
f)
2/3
(14.2)
14.6. Dimensioning methodology
First we calculate the machine size-constant C
0
. and stator bore diameter D
si
, based on
some already chosen parameters.
C
0
=

2

2
B
ag
J
l
k
w
(14.3)
Where:
B
ag
-PM air-gap ux density (T),
J
l
- linear electric specication loading (A
turns
/m),
Optimization design of surface PMSM 14-7
k
w
- winding factor: it includes zone factor k
ws
and chording factor k
chs
,
y
s
- coil span in slot pitch count.
k
w
= k
ws
k
chs
; k
ws
=
sin
_

6
_
q
1
sin
_

q
1
6
_; k
chs
= sin
_
y
1
q
1
m


2
_
(14.4)
Equation (14.3) are strictly valid for integer q
1
. D
si
in mm is:
D
si
= 1000
3
_
60p
s
n
n

P
n

c
C
0
(14.5)
p
s
- number of rotor poles (p
s
= 2p
1
): it is either given or calculated for the base speed
and chosen frequency, P
b
is base/rated electromagnetic power.
Equation (14.5) is safety to use mainly if the reaction air-gap ux density of stator mmf is
small (less than 25%) with respect the PM air-gap ux density B
ag
.
An alternative equation uses the electromagnetic torque T
eb
and the tangential specic force
f
tsp
(in N/cm
2
):
D
si
(mm) = 100
3

2p
s
10
2

T
eb

c
f
tsp
(14.6)
Fig.14.3 shows the PMSM main geometrical parameters.
Figure 14.3. The geometry main dimensions of PMSM
14-8 Electric Machines
Once D
si
is established the pole pitch and the core length stack, l
c
, are calculated:

p
=
D
si
p
s
(14.7)
l
c
=
c

p
(14.8)
Based on given B
ag
(0.2 1T) and
PM
= 0.6 0.9 the PM ux per pole
PM
is:

PM
=
2

B
ag

p
l
c
sin
_

PM

2
_
(14.9)
Now the stator and rotor yoke thickness h
sy
, h
ry
may be calculated:
h
sy
=
B
ag
B
sy

(14.10)
h
ry
=
B
ag
B
ry

(14.11)
The number of stator slots N
ss
, the stator slot pitch
ss
and teeth width w
st
are:
N
ss
= q
q
m p
s
(14.12)

ss
=
D
si
N
ss
(14.13)
w
st
=
ss
B
ag
B
st
(14.14)
The geometrical stator parameters in Fig. 14.4 may all be calculated. But before that the
slot geometrical angle
s
, and angle of tooth slice
st
are:

s
=
2
N
ss
;

s
=

N
ss
(14.15)

s0
= 2asin
_
s
0
D
si
_
(14.16)

st
=
s

s0
(14.17)
We verify now if the tooth top (slice) is larger than tooth width w
st
(14.12); if not the slot
opening, s
0
, is reduced accordingly. At the limit the slot may remain open when the central
angle for a tooth is:

stmin
= 2asin
_
w
st
D
si
_
(14.18)
For open slots:
s
0max
= D
si
sin(

stmin
/2) (14.19)
The main slot geometrical parameters (Fig.14.4) are:
Optimization design of surface PMSM 14-9
Figure 14.4. Stator slots geometry
h
s3
=
s
0max
s
0
2
+h
s4
tg
_

s
_
1
tg(w)
tg (

s
)
(14.20)
w
s3
= s
0max
+ 2(h
s4
+h
s3
)tg
_

s
_
(14.21)
w
s2
= w
s3
+ 2h
s2
tg
_

s
_
(14.22)
Where h
s2
is the slot classic insulation thickness.
The copper area in a slot S
cu
and the required slot area, S
slot
, are:
S
Cu
=
I
ts
J
s
(14.23)
S
slot
=
S
Cu
K
sf
(14.24)
Where I
ts
, is total RMS ampere-turn/slot; to be calculated from either (14.23) or (14.24 ):
I
ts
=
D
si
J
l
N
ss
(14.25)
I
ts
=
T
em

2p
1

p
N
ss
(14.26)
14-10 Electric Machines
Knowing the active slot area S
slot
its main dimensions are:
h
s1
=
w
s2
+
_
w
2
s2
+ 4S
slot
tan(

s
)
2tan(

s
)
(14.27)
w
s1
= w
s2
+ 2h
s1
tan(

s
) (14.28)
h
s
= h
s4
+h
s3
+h
s2
+h
s1
(14.29)
Then the outer stator diameter D
so
is approximated to a integer number in mm.
D
s0
= round(D
si
+ 2h
s
+ 2h
sy
) (14.30)
Now the stator yoke thickness h
sy
is:
h
sy
=
D
s0
D
si
2
h
s
(14.31)
The end connection coil length l
f
and total half turn length l
mc
and the axial length of stator
between end connections l
ff
are:
l
f
=

2

p
y
1
q
1
m
_
1 +
h
s
D
si
_
+ 2l
f1
(14.32)
l
mc
= l
c
+l
f
(14.33)
l
ff
= l
c
+ 2l
f1
+
p
y
1
q
1
m
_
1 +
h
s
D
si
_
(14.34)
The l
mc
length offer information about frame length.
a. Rotor sizing
The PMthickness h
PM1
is supposed to produce certain PMux density in the air-gap B
ag0
,
while also avoid demagnetization under short overload: .
h
PM1
=
B
ag0

min
k
s

0
H
c
_
1
B
ag0
Br
k
s
_ (14.35)
h
PM2
= 1000
I
ts
N
ss
p
1

2
k
1
k
PM
H
c
(14.36)
Where k
s
is saturation factor.
In both cases h
PM
is in mm and the largest of h
PM1
and h
PM2
is chosen and rounded to
tens of mm. If h
PM2
is larger than h
PM1
, the air-gap g, has to be recalculated also.
g =
0
H
c
h
PM
1
B
ag0
Br
k
s
B
ag0
k
s
(14.37)
Optimization design of surface PMSM 14-11
The gap g is rounded to an multiple of 0.05mm.
g = 0.05 round(20g) (14.38)
The rotor external and interior diameters, D
ro
, D
ri
, and core length l
cr
are:
D
r0
= D
si
2 (14.39)
D
ri
= D
r0
2(h
PM
+h
ry
) (14.40)
l
cr
= l
c
+
sr
(14.41)
The interior rotor diameter D
ri
is rounder to an integer number of mm, and thus the rotor
yoke thickness is slightly adjusted. If the D
ri
is smaller than the shaft diameter either the
dimensioning methodology is redone for a larger D
si
(by reducing B
ag
, J
l
, f
tsp
) or the
shape factor
c
, or the PMs will be placed on the shaft (massive rotor yoke).
b. PM ux computation
Failing to realize the required PM ux implies large current (and copper losses) for the base
speed (power) torque. For the so far preliminary design we added a saturation factor k
sat
and presupposed that it is feasible. This is needed to recalculate the PM ux with better
precision, as the PMs have a linear demagnetization curve with
rec
= (1.05 1.1)
0
.
Consequently, the Carter coefcient has to be calculated for the equivalent air-gap g
e
:
g
e
= g +
B
r

0
H
c
h
PM
(14.42)
k
c
=

s

s
s
0
(14.43)
Where:

s
=
_
s
0
ge
_
2
5 +
s
0
ge
(14.44)
Due to the nonlinearity of magnetic core materials the PM air-gap ux density has to be
calculated iteratively after we still use k
s
for saturation from preliminary design.
B
ag0
=
o
H
c
h
PM
g
e
k
C
k
s
(14.45)
The PM ux per pole is done with (14.9) and then we can recalculate B
st
, B
sy
, B
ry
with
(14.46-14.48).
B
st
= B
ag

s
w
st
(14.46)
B
sy
=

p
2l
c
h
sy
10
6
(14.47)
B
ry
=

p
2l
c
h
ry
10
6
(14.48)
14-12 Electric Machines
Making use of the core magnetization curve trough interpolation, the magnetic eld in the
stator H
st
, yoke H
sy
and rotor yoke H
ry
are calculated. Based on this the magnetisation
voltage V
mst
, V
msy
, V
mry
in the stator teeth, yoke and rotor yoke are calculated:
V
mst
= 0.001H
st
h
s
(14.49)
V
msy
= 0.001H
sy
x
p
s
C
x
(14.50)
V
mry
= 0.001H
ry
y
p
s
C
y
(14.51)
Where:
D
x
, and D
y
are respectively, the average diameters around which the average ux
path close in the two yokes
C
x
, and C
y
are, respectively two path length reduction coefcients that account ap-
proximately for the fact that non-uniform local saturation takes place in the stator and
rotor yokes.
D
x
= D
s0

4
3
h
sy
(14.52)
D
y
= D
ro
2h
PM

2
3
h
ry
(14.53)
C
x
, and C
y
depend on the ux density distribution and on the number of poles of the
rotor. They may be calculated through interpolation from the curves C in g.14.5, obtained
empirical or, better, by FEM.
Figure 14.5. Flux path length reduction coefcient C
Optimization design of surface PMSM 14-13
The magnetic voltage along air-gap V
mg
is:
v
mg
= 0.001g
e
k
C
B
ag

0
(14.54)
So the total mmf V
m
required to produce B
ag0
in the air-gap is:
v
m
= v
mst
+v
msy
+v
mry
+v
mg
(14.55)
The new value of the saturation factor k
sat
is:
k
s
=
V
m
V
mg
(14.56)
Based on this now value of k
s
the computation cycle of B
ag0
is redone until an error of less
than 1% between two successive values of k
s
is obtained. High precision is required in the
optimal design but then a mix analytical and a FEM calculation has to be converged. To
speed up the convergence of k
s
(or B
ag
) a relaxation coefcient may be used.
B
ag,k
= (1 r)B
ag,k1
+r B
ag,new
(14.57)
Where:
B
ag,k1
is the old value of air-gap ux density,
B
ag,new
is the value calculated with the just calculated ksat,
B
ag,k
the value of the B
ag
used in the next iteration cycle.
A value of r=0.2-0.3 should produce good convergence for saturated machines but would
slow convergence for less saturated machines. Convergence problem could ocurr for very
low saturation degree, but then the trouble process could be avoided by assigning ks a
reasonable (safe) value from start (k
s
= 1.05 1.1). After B
ag
was calculated with the
desired precision the PM ux per pole
p
is recalculated and, based on this, the PM ux
linkage for a single turn coil
PM1
is computed:

PM1
= q
1
n
L
p
1
2a
1

p
k
w
(14.58)
Now the equivalent current trough the one turn coil winding is:
l
q1
=
2T
n
3p
1

PM1
(14.59)
The stator phase resistance for the one turn coil winding R
s1
is:
R
s1
= 1000
n
cs
l
mc
a
1
S
cu
(14.60)
Where is the copper electrical resistively at the working temperature T
w1
, (14.61) and n
cs
is the number of coils in series.
=
20
(1 + (T
w1
20)
cu
) (14.61)
14-14 Electric Machines
Where
20
is the copper electrical resistively at 20

C,
cu
is the resistively coefcient with
temperature.
The number of conductors in series per phase for a single turn coils is:
n
cs
=
2q
1
p
1
n
L
a
1
(14.62)
Where a
1
is the number of current path and n
L
the number of winding layers (n
L
= 1, 2).
The cyclic magnetization inductance of the single turn coil winding is:
L
m1
= 2m
0
_
ncs
2
f
w
_
l
c

p
p
1
k
C
k
s
g
e
(14.63)
The saturation factor k
s
has been calculated on no load. For heavily saturated machine
k
s
has to be calculated including the inuence of stator mmf. Equation 14.63 is valid for
q
1
> 1. The leakage inductance is calculated using geometrical slot and end connection
permeances
ss
,
s0
.
k
2
=
_

_
3
_
2
y
1
mq
1
_
+1
4
if
y
1
mq
1
> 1
1+3
y
1
mq
1
4
if
2
3
<
y
1
mq
1
< 1
6
y
1
mq
1
1
4
if
y
1
mq
1
<
2
3
(14.64)
k
1
=
1 + 3k
2
4
(14.65)

ss
= k
1
h
s1
b
1
+k
2
_
h
s2
b
2
+
h
s3
b
3
+
h
s4
s
0
_
(14.66)
Where b
1
is the active slot average width and b
2
, b
3
the width corresponding to the trape-
zoidal shape zone free of conductors.
b
1
=
(w
s1
+w
s2
)
2
0.25(3w
s1
+w
s2
) + 0.5w
2
s1
w
s2
3w
s1
(w
s1
w
s2
)
2
+
w
4
s1
(w
s1
w
s2
)
3
log
_
w
s1
w
s2
_ (14.67)
b
2
=
w
s2
w
s3
log
_
w
s2
w
s3
_ (14.68)
b
3
=
w
s2
s
0
log
_
w
s3
s
0
_ (14.69)
The end connection geometrical permeance
s0
analytical computation is difcult but there
are quite a few empirical approximate expressions that are widely used.

s0
= 0.34q
1
l
f
0.64(D
si
+h
s
)
y
1
2p
1
mq
1
l
c
(14.70)
Optimization design of surface PMSM 14-15
This expression is mainly valid for distributed windings (q
1
> 1); for the tooth wound coil

s0
is larger as the conductors are closer to the core but the end connections are shorter.
Finally the leakage inductance L
1
per phase for single turn coil is:
L
1
= 2
0
_
n
cs
2
_
2

ss
+
s0
p
1
q
1

l
c
1000
(14.71)
The total cyclic (all three currents are present) inductance L
s1
is:
L
s1
= L
m1
+L
1
(14.72)
Considering the total current in axis q and the parameters of the single turn case we may
calculate the number of turns per coil s
b1
for operation at given voltage and speed (fre-
quency):
s
b1
=
V
n
_
(
1

PM1
+R
s1
I
q1
)
2
+ (
1
L
s1
I
q1
)
2
(14.73)
The number of turns per coil s
b1
has to be an integer; so it has to be rounded to s
b
. Conse-
quently the machine length has to be recalculated:
l
c
=
s
b1
s
b
l

c
(14.74)
Where l

c
is the old machine length (14.8). The l
c
value is rounded also to an integer number
of mm. The machine parameters, dependent on the machine length, such as R
s
,
PM1
, and
L
s1
, are recomputed. However, at least for n
s1
> 20 turns, use of (14.74) only, is feasible.
Now we calculate the conductor area:
q
cu
=
S
cu
S
b
(14.75)
To reduce skin effect it might be needed to use n
ce
elementary conductors of d
ce
diameter
in parallel:
d
ce
= 2
_
q
cu
n
ce
(14.76)
After norming d
ce
(according to standards) the total nal conductor area is recalculated. All
machine parameters, N
1
(turns parameter),
PM
, R
s
, L
m
, L

are recalculated for the nal


values of s
b
, l
c
calculated above:
N
1
= s
b
q
1
n
L
p
1
a
1
(14.77)

PM
= N
1

p
k
w
(14.78)
I
n
=
2

2
m

T
nem
p
1

PM
(14.79)
R
s
= 1000
2N
1
l
mc
a
1
q
cu
(14.80)
L
m
= 2m
0
(N
1
k
w
)
2
l
c

2
p
1
a
1
k
C
k
s
g
e
(14.81)
14-16 Electric Machines
L
1
= 2
0
N
2
1
p
1
a
1
q
1
(
ss
+
s0
)
l
c
1000
(14.82)
c. The active materials weight
The weight of windings:
m
cu
= 2ma
1
N
1
l
mc
q
cu

cu
10
9
(14.83)
The PM weight:
m
PM
=
PM
(D
r0
h
PM
)h
PM
l
cr

Fe
10
9
(14.84)
The rotor yoke weight:
m
rFe
= (D
ri
+h
ry
)h
ry
l
cr

Fe
10
9
(14.85)
The stator yoke weight:
m
sy
= (D
s0
h
sy
)h
sy
l
c
k
stk

Fe
(14.86)
The stator teeth weight:
m
sth
= N
ss
w
st
h
s
l
c
k
stk

Fe
10
9
(14.87)
The stator core weight:
m
sFe
= m
sy
+m
sth
(14.88)
The active material weight:
m
tot
= m
Fe
+m
PM
+m
sFe
+m
Cu
(14.89)
As the laminations are stamped from squares their weight mFe is in fact:
m
uFe
= D
2
s0
l
c
k
stk

Fe
10
9
(14.90)
This weight has to be used when the critical cost of the machine materials is calculated.
The rotor inertia J
s
is:
J
s
= m
PM
D
2
r0
+D
2
r1
8
10
6
+m
ry
D
2
r1
+D
2
ri
8
10
6
(14.91)
The shaft inertia is added to value from (14.91).
d. Losses
Winding losses
P
Cu
= mR
s
I
2
n
(14.92)
Core losses
The core losses are calculated separately for the stator teeth and yoke. Based on available
Optimization design of surface PMSM 14-17
core loss data we may calculate either separately hysteresis and eddy current losses or
together[4]:
p
ht
=
1

Fe
k
h
f
efh
n
B
eBh
st
(14.93)
p
hy
=
1

Fe
k
h
f
efh
n
B
eBh
sy
(14.94)
p
et
=

2
6
Fe

Fe
f
efe
n
B
eBe
st
g
egt
t
(14.95)
p
ey
=

2
6
Fe

Fe
f
efe
n
B
eBe
sy
g
egt
t
(14.96)
p
Fet
= p
ht
+p
et
(14.97)
p
Fey
= p
hy
+p
ey
(14.98)
p
Fet
= p
Fet
m
sth
(14.99)
p
Fey
= p
Fey
m
sy
(14.100)
p
Fe
= p
Fet
+p
Fey
(14.101)
Where:
- efh - frequency exponent for hysteresis losses;
- eBh - ux density exponent for hysteresis losses;
- efe - frequency exponent for eddy losses;
- eBe - ux density exponent for eddy losses;
- egt - thickness exponent.
The segregation of yoke p
Fey
and teeth p
Fet
core losses allows to follow their evolution and
thus facilitates corrective measures to reduce them in the critical zone.
e. Thermal verication
To roughly verify the machine winding over temperature rst the total frame area A
fr
for
heat transfer is calculated:
A
fr
= 10
6
(k
f
D
s0
l
f
+

2
2
D
2
s0
) (14.102)
where k
f
is the cooling surface increase ratio due to ns.
Then adding an equivalent heat transfer coefcient
t
(in W/m

C), say from 14 for no


ventilated frames to 100 for water cooled frame jackets, the winding over-temperature T
w1
is:
T
w
=
P
Cu
+P
Fe

t
A
fr
(14.103)
If the over-temperature is high, while
t
has a realistic value in accordance with the adopted
cooling system, then the machine design is redone from start with smaller J
l
(or f
tsp
) and
or longer
c
(longer core length).
14-18 Electric Machines
f. Machine characteristics
In general for a PMSM the characteristics include:
base speed nb, maximum speed nmax,
continuous torque at base speed, and base (maximum) voltage,
peak torque at base speed,
torque at maximum speed,
emf at maximum speed,
efciency and power factor (or stator current) versus load and speed for given inverter
maximum fundamental voltage V
1
.
14.7. Optimal design with genetic algorithms
For the optimization design of SPMSM with genetic algorithms (GA) we introduce the
following optimization variables:
linear electric loading J
l
(A/m),
air-gap ux density, B
ag
(T),
stator teeth ux density, B
st
(T),
stator yoke ux density, B
sy
(T),
rotor yoke ux density, B
ry
(T),
stator current density, J
s
(A/mm
2
),
machine shape factor,
c
= l
c
/
p
,
slot per pole per phase, q
1
,
slot opening, s
o
(mm),
tooth top height, h
s4
(mm),
coil span, y
1
(mm),
PM pole span coefcient,
PM
.
These 12 variables are grouped in a vector X
0
. The variable vector X is not directly ini-
tialized by the designer, but he (she) will attribute minimum and maximum values X
min
,
X
max
together with corresponding resolution X. Also all geometrical variables minimum
values that are imposed technologically or for mechanical reasons are grouped in G
dmin
:
outer rotor diameter,
Optimization design of surface PMSM 14-19
interior rotor diameter,
minimum active slot height,
minimum total slot height,
minimum slot width,
minimum active slot area.
The chosen GA operates binary coding of optimization variables. As some of the opti-
mization variables such as the number of stator slots per pole and phase q
1
, coil span in
slot pitches are integer while others are rational numbers with extended dynamics, they are
represented with different number of bits. Now the variable domain X
max
X
min
and the
representative resolution the required number of bits N
bit
is:
N
bit
= 1 +int
_
log
2
_
X
max
X
min
X
__
(14.104)
To set the GA (described in detail in chapter 13) we give: population size, n
p
, number of
individuals, in a generation), number of generations, n
g
, elitism factor, k
elit
, mutation rate,
r
m
, an occupation factor k
ex
of individuals which participate to produce genetic material
for the next generation.
a. The objective (tting) function
A complex objective (tting) function we have chosen here: the motor initial cost plus the
loss cost for the machine active life plus additional framing and transportation costs due to
its weight. The initial cost C
i
, includes only active materials costs as the total manufacturing
and selling costs depend tightly on the fabrication technology and cost management of the
particular manufacturer.
C
i
= m
cu
p
w
+m
Feu
p
lam
+m
PM
p
PM
(14.105)
Where p
w
, p
lam
, p
PM
are the copper, lamination, PM unitary price, say in USD(or EU)/kg.
The cost of losses C
E
is:
C
E
= P
N
_
1
1

N
_
n
hy
n
y
p
E
(14.106)
Where n
hy
is the annual operation hours, n
y
operational years; p
E
- energy cost
(USD(EU)/kWh).
As very rarely, a motor operates all the time at rated power and speed, we may introduce an
ideal operation cycle characterized by probability
i
to operate at different power loads P
i
and efciency
i
.
C
E
= n
hy
n
y
p
E
n

i=1

i
P
i
_
1
1

i
_
(14.107)
Where:

i
= 1, i = 1, n
14-20 Electric Machines
Though C
E
is in general larger than C
i
, the value of C
E
is diminished in the design as
many buyers can not afford increased initial cost; in such cases, in fact, a kind of optimal
efciency in terms of initial/total cost is adopted. Now the cost of losses C
E
is:
C
E
= n
hy
n
y
p
E
n

i=1

i
P
i
(14.108)
Where p
i
=
_
P
i
_
1

oi

i
_
if
i
<
oi
0 if
i

oi
In aircraft applications the low weight is a critical performance index. A similar situation
occurs in wind generators while low cost but poor performance materials lead to larger
machine weight which implies a heavy nacelle and tower and their extra costs may offset the
reduction of generator initial costs. To force the algorithm to reduce the motor (generator)
weight we may add some additional costs C
m
proportional to motor weight m
t
.
C
m
= m
t
p
m
(14.109)
Also, surpassing the maximum allowable winding (and core and PM) temperature, leads
to immature aging of the equipment. To avoid this with the optimization algorithm we
introduce a penalty cost for over-temperature C
temp
in the objective function. The penalty
cost C
temp
may vary liniarly or exponentialy with over temperature.
C
temp
=
_
k
T
(T T
max
)C
i
if T > T
max
0 if T
max
(14.110)
We may add penalty costs for any other technical constraints. For example, we may add a
penalty for lower power factor, which is related to the converter kVA rating costs, also for
the outer diameter or for machine axial length, etc. The larger the specic costs of these
constraints, the higher the probability that the optimizations design will observe them.
Finally the total cost C
t
is:
C
t
= C
i
+C
E
+C
m
+C
p
(14.111)
b. PMSM optimization design using genetic algorithms: a case study
An exercise design of a three phase, four pole synchronous motor with surface permanent
magnet rotor for rated power of 2.2kW and base frequency is presented. The maximum and
minimum values of optimization variables are presented in table 14.1. The technological
constraints used for this exercise design could be found in the input le (pmm1.m) of the
PMSM optimal design MATLAB code which is stored on the CD in the PMSMdesign
folder. The total cost C
t
represents the objective function.
The algorithm evolution is illustrated in the following gure. Every generation has a mem-
ber that is best tted for the required optimum criterion which minimizes the cost function.
The generation itself is characterized by the average cost function. For each generation
Optimization design of surface PMSM 14-21
Table 1. Optimization variable bounds
Optimization Minimum Maximum Units Comments
Variable Values Values
J
1
15 30 kA/M Electric specic load specications
Bagsp 0.45 0.75 T Air-gap ux density
Bst 1 2 T Stator tooth ux density
Bsy 0.9 1.9 T Stator yoke ux density
Bry 0.9 2.1 T Rotor yoke ux density
J
s
3 8 A/mm
2
Stator current density
lcpertau 0.5 3 Core stack length per pole pitch
q
1
2 4 Stator slots per pole per phase
so 1 5 mm Stator slot opening width
sh4 0.5 2 mm Stator tooth pole tip height
cSpan 0.66 1 Coil span per pole pitch
Alpm 0.5 1 Permanent magnet angle per pole angle
there is also the least adapted member. Following the evolution of these three values we
may appreciate the genetic algorithm convergence. At the beginning there is a large dif-
ference between the cost function associated to the best member and the average value of
the cost function. The value of the cost function of the worsted adapted member is more
than ten times larger than the average value. A large gap between minimum and max-
imum cost function over a generation members shows a population with good evolution
perspective. The statistics values (minimum, average and maximum) of the cost function
tend to decrease during evolution. The minimum value of the cost function is decreasing
monotonically along the generations while the average and maximum values sometimes
are increasing especially, for the rst generation, g. 14.6, where a large variation of the
maximum cost could be observed. In time, the population members start to contain similar
codes and the variation of maximum and minimum cost function from a generation to the
next diminish. Gradually the minimum value of the objective function remains unchanged
for many generations and the cost variation is small when it even exists.
14-22 Electric Machines
Figure 14.6. Minimum, maximum and average cost function evolution
When the members were chosen randomly, in pairs, to generate offspring it was imposed to
have a minimum difference, in order to avoid a premature evolution. A premature evolution
has produced numerous twins in the population and maximum and average costs collapsed
to the minimum cost. The main method, crossover, used to generate offspring is not able
to produce any improvement from this stage. Then a new genetic code is produced only
by mutation, but if the new members have the cost function higher than the majority of the
population members, which are very similar, the new genetic code will be eliminated in a
few generations.
The main geometric dimensions and the slot geometry evolution of the best member along
generations are presented in g. 14.7. It could be observed that geometric dimensions take
discrete values and the same dimension could arrived again after several generations.
Figure 14.7. Main geometric evolution of the best member
Optimization design of surface PMSM 14-23
The power losses and energy efciency evolution of the best member is shown in g. 14.8.
The mechanical power loss is set constant as 0.5% of the rated power in the input power.
A small efciency improvement (around 1%) is observed along to the rst ten generations.
The best candidate solution from the rst generation has already a good efciency because
it represents the best solution of a population with 150 members.
Figure 14.8. Power losses and efciency evolution of the best member
The machine component weights and their cost evolution are presented in g. 14.9. The
laminations are punched from rectangular sheets and the minimum rectangular (square)
surface that encloses the stator lamination is considered when computing the weight of
used iron. When the rotor core is made of laminations, then it is punched from the same
sheet as the stator lamination at no extra cost for the material. Solid rotor core is used in
many cases for small PMSM, and in this case the cost of rotor core material is added.
Figure 14.9. The components weight and their cost for the best candidate solutions
The initial cost and energy cost evolution for the best candidate solution is shown in g.
14.10. We can observe that their values are about the same which is a well known optimiza-
tion thumb rule. The sum of initial cost and energy cost is equal to total cost, which is
14-24 Electric Machines
objective function plotted again in g. 14.10 in a better scale. Usually, the over temperature
penalty cost is equal to zero for the best candidate solution.
Figure 14.10. Evolution of the cost (objective) function of the best candidate solutions
The optimization variable evolutions for the best candidate solution are presented in the
following pictures. The maximum and minimum values of optimization variables as well
as their resolution are given in the input le. The number of discrete levels of each variable
is usually not equal with powers of two. The required number of bits, (14.102) necessary to
cover the discrete levels of the optimization variable, could code a larger number of levels
that it is required from the initial le specication. Using the described crossover method
it is possible to produce offspring which has one of more optimization variable larger than
upper bounds. One solution is to discard all offspring that has the optimization variable
outside the bounds and try again to generate viable offspring. Other solution is to let the
optimization variable to be larger than upper limits and let the optimization algorithm to
eliminate non-feasible candidates. This method is applied when an optimization variable
larger than the upper bounds has physical meaning such as for electric and magnetic loads.
Other method to solve this problem is to decrease the resolution in such way that all binary
code is used to represent variables between given bonds.
When variable values out of their bounds are nonsense, then a mathematically transforma-
tion was applied to force the optimization variable back between its bounds. For example
the ratio between permanent magnet width and pole pitch larger than one is a nonsense.
The following mathematical transformation was applied, to avoid this nonsense.

PM
=
_

PM
if
PM
1
2
PM
if
PM
> 1
(14.112)
Some features could have two distinct codes. For example the code which is producing the
number 1.1, is forced to produce a relative PM width equal to 0.9 and there exist also the
Optimization design of surface PMSM 14-25
code which produces directly 0.9.
The optimal value of linear electric load is larger than 40kA/m, g. 14.11, while the
maximum value in the input le was set at 30kA/m. The optimal value of the air-gap ux
also tends to its upper bound, g.14.12. Larger limits of these optimization variables will
be set if there is no other restriction.
Figure 14.11. Specic electric load evolution of the best candidate solutions
Figure 14.12. Flux density evolutions of the best candidate solutions
14-26 Electric Machines
The upper limit of the current density is larger enough in our design and thus there is no
optimization value of the best candidate solution larger than the upper limit, g.14.13. The
codication method allows for larger current density because the required domain needs 50
samples to be covered and 6 bit is enough to code 64 samples.
Figure 14.13. Current density evolution of the best candidate solutions
The ratio of core length to pole pitch as well as ratio of coil span to pole pitch and ratio
of permanent magnet width to pole pitch are presented in g. 14.14 for the best candidate
solution from each generation. The ratio of coil span to pole pitch given by genetic algo-
rithm is a round to the nearest rational number which has the denominator equal to number
of slots per pole.
Optimization design of surface PMSM 14-27
Figure 14.14. Dimensions ratio evolutions of the best candidate solutions: core length per
pole pitch, lcpertau, coil span per pole pitch, coil span, PM width per pole pitch, (all in pu).
The number of stator slots per pole per phase is always an integer number and it is shown
in g. 14.15 for the best candidate solution from each generation.
Figure 14.15. Slots per pole per phase evolution of the best candidate solutions
14-28 Electric Machines
Figure 14.16. The cost function over the last generation
All members of the last generation are obtained from a long selection process and there
would be interesting to analyze the optimization variables distribution. At rst sight the
members look close to the objective function, g.14.16. We may observe that not all possi-
ble level are taken, g.14.17-14.20, and , that there is no unambiguous rule on what values
could minimize the optimization function.
We are not able to assert that always the value of linear electric load which characterizes
the best solution (40.4kA/m) will produce good results, g.14.17. We can state that prob-
ably linear electric load smaller than 25kA/m is too small to get an acceptable motor cost.
Good solutions are possible with an air-gap ux density around 0.75T, g.14.18, respec-
tively, with the stator yoke and stator tooth ux density around 1.55T. The rotor yoke ux
density has small inuence on the objective function, g. 14.18. There is not possible,
probably, to minimize the objective function when current densities become smaller than
4A/mm
2
. About the ratio-type optimization variables we can state that a coil span per
pole pitch equal to 0.83 and the permanent magnet width to pole pitch equal around 0.91
could produce good solutions, g.14.20. Ratio of core length to pole pitch larger than 2.5
probably produce always bad results, g.14.20.
Optimization design of surface PMSM 14-29
Figure 14.17. The specic electrical load over the last generation
Figure 14.18. Flux densities over the last generation
14-30 Electric Machines
Figure 14.19. Current density along the last generation
Figure 14.20. Dimensions ratio over last generation
There are only two values selected for the number of slots per pole per phase and we can
remark that in the last generation 2 slots per pole per phase have higher occurrence fre-
quencies than 3 slots per pole per phase, especially for high performance members, g.
14.21. The best candidate solution for many populations has 2 slots per pole per phase, so,
probably, this value will minimize the objective function.
Optimization design of surface PMSM 14-31
Figure 14.21. The slots per pole per phase over last generation
Further improving of the genetic algorithm is possible by reducing search area and even set
some variables as constant and then eliminate these from the optimization variable vector.
14.8. Optimal design of synchronous permanent magnet motor using Hooke
Jeeves method
The same optimization variable vector and objective function, as that presented for genetic
algorithm method, is used also to implement the Hooke Jeeves optimization algorithm pre-
sented also in detail in chapter 13. The geometric dimensions, as well as the optimization
variables and cost function modications look now as continuous functions, compared with
genetic algorithm functions, where these functions look rather discontinuous. The evolu-
tion of the main geometric and slot dimensions is presented in g. 14.22. The power losses
and efciency evolution is shown in g. 14.23 while the components weight and cost are
visible in g. 14.24.
14-32 Electric Machines
Figure 14.22. Geometric dimensions evolution
Figure 14.23. Power losses and efciency evolution
Figure 14.24. Components weight and their cost evolution
Optimization design of surface PMSM 14-33
Figure 14.25. Cost (objective) function evolution
Figure 14.26. Specic load evolutions
14-34 Electric Machines
Figure 14.27. Flux density evolutions
Figure 14.28. Current density evolutions
The ratio of permanent magnet width to pole pitch and ratio of coil span to pole pitch are
going to the same values as in genetic algorithm, g. 14.29, while the core length per pole
pitch ratio is going to smaller value. The algorithm starts with 3 slots per pole per phase and
after three steps it change to 2 slots per pole per phase and remains at that value, g.14.30.
Optimization design of surface PMSM 14-35
Figure 14.29. Dimensions ratios evolution
Figure 14.30. Slot per pole per phase evolution
Comparative results of optimal design using genetic algorithms, respectively, Hooke Jeeves
method are presented in table 14.2. The total cost is about the same for the two methods
but the simulation time is about 75 times larger for genetic algorithms than for the Hooke-
Jeeves algorithm.
The genetic algorithm chooses for some optimization variables, values outside the bounds
14-36 Electric Machines
Table 2. Comparisons of optimization results
Parameter GA HG Units Comments
sDo 139 149 mm Stator external diameter
sDi 73 86 mm Stator inner diameter
sh4 0.9 0.6 mm Stator tooth pole tip height
so 1 2.5 mm Stator slot opening width
sh3 0.9 0.9 mm Height of wedge for stator slots
shOA 24.2 20.7 mm Total height of stator slot
sh1 21.7 18.5 mm Main (coil) height of the stator slots
shy 8.8 10.8 mm Stator yoke height
swt 4.7 5.1 mm Stator teeth width
N
1
336 312 Numbers of turns per phase
rDo 69.6 82.9 mm Rotor outer diameter
rDi 43 57 mm Rotor inner diameter
hpm 3.7 3.2 mm Permanent magnet height (thickness)
hag 1.7 1.55 mm Air-gap
J
l
40.4 30 kA/m Linear electric load
Bagsp 0.76 0.75 T Air-gap ux density
sBtp 1.55 1.67 T Stator tooth ux density
sByp 1.56 1.49 T Stator yoke ux density
rByp 1.44 1.65 T Rotor yoke ux density
J
s
5.3 4.94 A/mm
2
Stator current density
Lcpertau 1.9 1.6 Ratio of core stack length to pole pitvh
q
1
2 2 Stator slots per pole per phase
cSpan 0.83 0.83 Ratio of coil span to pole pitch
alpm 0.91 0.91 Ratio of PM width to pole pitch
WeightStCu 3.365 3.255 kg Coil weight
WeightIronUsed 15.268 16.961 kg Used iron weight
WeightStIron 5.003 5.509 kg Stator lamination weight
WeightPM 0.595 0.603 kg PM weight
WeightRtIron 1.372 1.712 kg Rotor iron weight
WeightMot 10.336 11.079 kg Motor total weight
Iqn 3.9 3.81 A Rated current
Vfn 220 220 V Rated voltage
Pcu 97.39 88.15 W Cooper losses
Pfe 38.76 41.2 W Iron losses
Pmec 11 11 W Mechanical losses
Etan 93.73 94 USD Efciency
cu
c
33.649 32.547 USD Copper cost
lam
c
76.339 84.806 USD Used laminations cost
PM
c
29.772 30.173 USD Permanent magnet cost
rotIron
c
6.8634 8.5575 USD Rotor solid iron cost
pmw
c
51.679 55.395 USD Penalty cost of active material weight
i
cost
198.30 211.48 USD Initial material cost
energy
c
220.73 210.52 USD Loss energy cost during 10 years, 1500h/year
t
cost
419.03 422.00 USD Total cost (objective function)
sim
time
340.86 4.516 s Simulation time on Pentium IV-2.6GHz
Optimization design of surface PMSM 14-37
while the Hooke Jeves optimization variables remain strictly within the specied bounds.
The motor efciency designed by Hooke-Jeves algorithm is larger than the motor efciency
designed with genetic algorithm but the initial cost is smaller for genetic algorithm and
probably this is a consequence of larger linear electric load allowed for the genetic algo-
rithm. The rst variation step for each variable in the research movement of Hooke-Jeeves
algorithm was 10% of the entire domain. Such large initial step may help the algorithm to
examine a large area and thus increase the chance to pass over local minima.
Conclusion
This chapter may be considered a solid introduction to the analytical optimization design of
PMSM. The results of the full analytical model should help in grasping better the principles
and some subtleties.
Though magnetic saturation was considered in the analytical model, cogging torque was
not.
Total torque pulsations due, mainly, to magnetic saturation, or due to static or dynamic
eccentricity were not considered either. It should be that, tying up analytical optimization
design methodologies as in this chapter integrating them into nite element software, for
validation and exploration of cogging torque, total torque ripple, etc is, perhaps, the next
step. The interested reader is invited to follow it up on his (her) own.
References
[1] Daune C. Hanselman, Brusless Permanent Magnet Motor Design , The Writers Colec-
tive; 2
nd
edition (March 2003);
[2] J.R.Hendershot, T.J. Miller, Design of Brushless Permanent-Magnet M0tors, Oxford
University Press, 1995;
[3] Wen Ouyang; Zarko, D. Lipo, T.A.; Permanent Magnet Machine Design Practice and
Optimization; Industry Applications Conference, 2006. 41st IAS Annual Meeting.
Conference Record of the 2006 IEEE Vol. 4, 8-12 Oct. 2006 Page(s):1905 - 1911;
[4] Claus B. Rasmusen, Modeling and Simulation of Surface mounted PM Motors,
Ph.D. Thesis, Institute of Energy Technology, Aalborg University, Denmark, 1996;
[5] IEC/EN 60034;
[6] Thin non-oriented Electrical Steel - SURA 007, Products Information, Surahammars
Bruk, Sweden.