This action might not be possible to undo. Are you sure you want to continue?
VoltageSource PWM Inverter
1.1 Introduction
In contrast to grid connected ac motor drives, hardly variable in speed, power
electronic devices (e.g. inverter), providing voltage supply variable in both
frequency and magnitude, are used to operate ac motors at frequencies other than the
supply frequency. Developments in this direction have taken place long ago, but a
technoeconomical solution could not be found until the late 1980s because of
stringent space requirements, nonavailability of high power devices and prohibitive
cost of electronic devices and components.
Rapid developments in the field of power electronics (inverter grade thyristor, GTO
thyristor, IGBT etc.) and miniaturization/mass production of control electronics
(development of VLSI technology and microprocessor based digital control
systems) have reached such a stage that variable ac inverter drives are becoming
increasingly popular in today’s motor drives. Presently, inverter drives meet not
only weight and space constraints, but also are economically viable.
In general, two basic types of inverters exist: Voltagesource inverter (VSI),
employing a dc link capacitor and providing a switched voltage waveform, and
currentsource inverter (CSI), employing a dc link inductance and providing a
switched current waveform at the motor terminals. CSinverters are robust in
operation and reliable due to the insensitivity to short circuits and noisy
environment. VSinverters are more common compared to CSinverter since the use
of Pulse Width Modulation (PWM) allows efficient and smooth operation, free from
torque pulsations and cogging [Bose 97]. Furthermore, the frequency range of VSI is
higher and they are usually more inexpensive when compared to CSI drives of the
same rating [Dub 89].
In this chapter, only voltagesource inverters are considered. Although the power
flow through the device is reversible, it is called an inverter because the predominant
power flow is from the dc bus to the threephase ac motor load. Bidirectional power
flow is an important feature for motor drives as it allows regenerative breaking, i.e.
the kinetic energy of the motor and its load is recovered and returned to the grid
when the motor slows down. In electric vehicle application, the dc bus energy is
supplied directly from primary energy sources, e.g. batteries.
2 Chapter 1
In ac grid connected motor drives, a rectifier, usually a common diode bridge
providing a pulsed dc voltage from the mains, is required. Alternatively, a second
actodc converter, acting as a rectifier during the motoring mode and an inverter
during the breaking mode, is used between drive and utility grid. An additional
benefit of the active front end is enabling unity power factor, (sinusoidal) current
flows to or from the grid.
Although the basic circuit for an inverter may seem simple, accurately switching
these devices provides a number of challenges for the power electronic engineer.
The most common switching technique is called Pulse Width Modulation (PWM).
PWM is a powerful technique for controlling analog circuits with a processor’s
digital outputs. PWM is employed in a wide variety of applications, ranging from
measurement and communications to power control and conversion. In ac motor
drives, PWM inverters make it possible to control both frequency and magnitude of
the voltage and current applied to a motor. As a result, PWM inverterpowered
motor drives are more variable and offer in a wide range better efficiency and higher
performance when compared to fixed frequency motor drives. The energy, which is
delivered by the PWM inverter to the ac motor, is controlled by PWM signals
applied to the gates of the power switches at different times for varying durations to
produce the desired output waveform.
There are several PWM modulation techniques. It is beyond the scope of this book
to describe them all in detail. The following illustration describes the basic three
phase inverter topology and typical pulse width modulation methods. Furthermore,
issues of phase voltage distortion/identification due to the inverter nonlinearity are
discussed in detail.
1.2 VoltageSource PWM Inverter
A typical voltagesource PWM converter performs the ac to ac conversion in two
stages: ac to dc and dc to variable frequency ac. The basic converter design is shown
in figure 1.1. The grid voltage is rectified by the line rectifier usually consisting of a
diode bridge. Presently, attention paid to power quality and improved power factor
has shifted the interest to more supply friendly actodc converters, e.g. PWM
rectifier. This allows simultaneously active filtering of the line current as well as
regenerative motor braking schemes transferring power back to the mains.
The dc voltage is filtered and smoothed by the capacitor C in the dc bus (figure 1.1).
The capacitor is of appreciable size (220 mF) and therefore a major cost item
[Bose 97]. Alternatively, the inverter can be supplied from a fixed dc voltage. The
filtered dc voltage is usually measured for control purpose. Because of the nearly
constant dc bus voltage, a number of PWM inverters with their associated motor
drives can be supplied from one common diode bridge. The inductive reactance L
between rectifier and ac supply is used to reduce commutation dips produced by the
rectifier, to limit fault current and to soften voltages spikes of the mains.
VoltageSource PWM Inverter 3
T1
T4
D4
D1
C
T3
T6
D6
D3
T5
T2
D2
D5
power
supply
AC
motor
Udc
Rectifier DC bus Inverter
Switching logic
L
Figure 1.1: Basic threephase voltagesource converter circuit.
Neglecting the voltage drop of the inductances (current depending) and diodes
(U
d
≈ 1V if i > 0), the positive potential of the dc bus voltage equals the highest
potential of the three phases and the negative potential equals the lowest potential of
the three phases. Since each phase owns one negative and one positive maximum
potential during one period of the net frequency, the rectifier input voltage equals
the maximum of the positive and negative line voltages, respectively. Thus, the
rectifier input voltage traces six pulses as shown in figure 1.2 by the thick line.
0 5 10 15 20
400
500
600
t [ms]
U
d
c
[
V
]
0 5 10 15 20
0
0.5
1
t [ms]
i
B
6
[
A
]
0 5 10 15 20
1
0
1
t [ms]
i
a
[
A
]
0 5 10 15 20
400
500
600
t [ms]
U
d
c
[
V
]
0 5 10 15 20
0
20
40
t [ms]
i
B
6
[
A
]
0 5 10 15 20
40
20
0
20
40
t [ms]
i
a
[
A
]
U
dc
U
dc
u
bc
u
ab
u
ca
u
bc
u
ab
u
ca
u
bc
u
ab
u
ca
u
bc
u
ab
u
ca
Figure 1.2: Line voltages (u
ab
, u
bc
, u
ca
), dc bus voltage U
dc
, line current
of the first phase i
a
and output current i
B6
of a B6diode bridge.
Left: No inverter output power (inverter losses ≈ 10 W).
Right: Inverter output power P
out
≈ 5,5 kW.
Figure 1.2 presents typical voltage and current waveforms of a B6diode bridge
supplied by a stiff grid. As indicated by the dashed lines, the rectifier current i
B6
increases, if the absolute value of a line voltage is higher than the dc voltage.
Consequently, the dc voltage increases slightly. A dc voltage higher than the current
voltage supply causes a reduction of the rectifier input current until the current
4 Chapter 1
equals zero and the diode bridge blocks the supply voltage. The rectifier current i
B6
is identically reflected by the line currents. The sign of each line current depends on
the two nonblocking diodes each conducting the positive and negative rectifier
current, respectively.
During the conducting period, the difference of line and dc voltage is active as
voltage drop over the line inductances and resistances. The higher the line
inductances, the smaller the line current peaks. However, the value of the line
inductances is limited due to economic and efficiency reasons. Furthermore, the
average dc voltage depends on the line inductances and the inverter output power.
The maximum dc voltage (no load) is equal to the maximum amplitude of the line
voltages. Due to voltage drops of line inductances, resistances and rectifier diodes,
the dc voltage slightly decreases with increasing load. For more details concerning
the rectifier, see [Bose 97], [Dub 89] et al.
According to figure 1.1, the dc voltage is switched in a threephase PWM inverter
by six semiconductor switches in order to obtain pulses, forming threephase ac
voltage with the required frequency and amplitude for motor supply. The switching
devices must be capable of being turned “on” as well as turned “off”. During the last
years, major progress has been made in the development of new power
semiconductor devices. The simpler requirement driving the power switches and the
higher maximum switchingfrequency, enabling higher operating frequencies
(higher motor speed), provide continually rising output power. The new generation
of switching devices is capable of conducting more current and blocking higher
voltages. The alternatives at present are gate turnoff thyristor (GTO), MOS
controlled thyristor (MCT), bipolar junction transistor (BJT), MOS field effect
transistor (MOSFET) and insulated gate bipolar transistor (IGBT).
The IGBT is a combination of power MOSFET and bipolar transistor technology
and combines the advantages of both. In the same way as a MOSFET, the gate of the
IGBT is isolated and its driving power is very low. However, the conducting voltage
is similar to that of a bipolar transistor. Presently, IGBTs dominate the medium
power range of variable speed drives. Since the maximal current rating of IGBT
modules is around 1 kA and the voltage rating is approximately 3 kV, they will
gradually replace GTOs at higher power levels [Vas 99].
Parallel to the power switches, reverse recovery diodes are placed conducting the
current depending on the switching states and current sign. These diodes are
required, since switching off an inductive load current generates high voltage peaks
probably destroying the power switch. Exemplary for one inverter leg, figure 1.3
presents the basic configuration and the inverter output voltage depending on the
switching state and current sign. The basic configuration of one inverter output
phase consists of upper and lower power devices T1 and T4, and reverse recovery
diodes D1 and D4.
When transistor T1 is on, a voltage ½ U
dc
is applied to the load. Considering an
inductive load, the current increases subsequently. If the load draws positive current,
VoltageSource PWM Inverter 5
it will flow through T1 and supply energy to the load. To the contrary, if the load
current i
a
is negative, the current flows back through D1 and returns energy to the dc
source.
u
a0
T
1
T
4
D
4
D
1
i
a
< 0
C/2
C/2
ωt
i
a
u
a0
ωt
½ Udc
½ Udc
T
1
on
T
4
on
τ
dead
D1
drop
D4
drop
T1
drop
T4
drop
0
0
u
a0
T
1
T
4
C/2
½ U
dc
D
4
D
1
i
a
> 0
C/2
½ U
dc
½ U
dc
½ U
dc
T4 off
T1 off
T4 on
T1 on
ua0
Similarly if T4 is on, which is equal to T1 off, a voltage ½ U
dc
is applied to the load
and the current decreases. If i
a
is positive, the current flows through D4 returning
energy to the dc source. A negative current yields T4 conducting and supplying
energy to the load.
Figure 1.3: Basic configuration of a halfbridge inverter and centertapped inverter output voltage.
Left: Switching states and current direction. Right: Output voltage and line current.
According to figure 1.3, with T1 on and drawing positive load current i
a
, the output
voltage u
a0
will be less than ½ U
dc
by the onstate voltage drop of T1. When the load
current reverses, the output voltage will be higher than ½ U
dc
by the voltage drop
across D1. Similarly, the output voltage is slightly changed by the voltage drop of
the lower devices T4 and D4.
Normally, the onstate voltage and diode drops (≈1 V) are ignored and the center
tapped inverter is represented as generating the voltage ½ U
dc
and ½ U
dc
,
respectively. Neglecting additionally the deadtime interval τ
dead
, the behavior of the
power devices together with the reverse recovery diode is equally described by ideal
twoposition switches.
6 Chapter 1
1.3 Pulse Width Modulation
Usually, the on and offstates of the power switches in one inverter leg are always
opposite. Therefore, the inverter circuit can be simplified into three 2position
switches. Either the positive or the negative dc bus voltage is applied to one of the
motor phases for a short time. Pulse width modulation (PWM) is a method whereby
the switched voltage pulses are produced for different output frequencies and
voltages. A typical modulator produces an average voltage value, equal to the
reference voltage within each PWM period. Considering a very short PWM period,
the reference voltage is reflected by the fundamental of the switched pulse pattern.
Apart from the fundamental wave, the voltage spectrum at the motor terminals
consists of many higher harmonics. The interaction between the fundamental motor
flux wave and the 5
th
and 7
th
harmonic currents produces a pulsating torque at six
times of the fundamental supply frequency. Similarly, 11
th
and 13
th
harmonics
produce a pulsating torque at twelve times the fundamental supply frequency
[Dub 89]. Furthermore, harmonic currents and skin effect increase copper losses
leading to motor derating. However, the motor reactance acts as a lowpass filter and
substantially reduces highfrequency current harmonics. Therefore, the motor flux
(IM & PMSM) is in good approximation sinusoidal and the contribution of
harmonics to the developed torque is negligible. To minimize the effect of
harmonics on the motor performance, the PWM frequency should be as high as
possible. However, the PWM frequency is restricted by the control unit (resolution)
and the switching device capabilities, e.g. due to switching losses and dead time
distorting the output voltage.
There are various PWM schemes. Wellknown among these are sinusoidal PWM,
hysteresis PWM, space vector modulation (SVM) and “optimal” PWM techniques
based on the optimization of certain performance criteria, e.g. selective harmonic
elimination, increasing efficiency, and minimization of torque pulsation [Jen 95].
While the sinusoidal pulsewidth modulation and the hysteresis PWM can be
implemented using analog techniques, the remaining PWM techniques require the
use of a microprocessor.
A modulation scheme especially developed for drives is the direct flux and torque
control (DTC). A twolevel hysteresis controller is used to define the error of the
stator flux. The torque is compared to its reference value and is fed into a threelevel
hysteresis comparator. The phase angle of the instantaneous stator flux linkage space
phasor together with the torque and flux error state is used in a switching table for
the selection of an appropriate voltage state applied to the motor [Dam 97],
[Vas 97]. Usually, there is no fixed pattern modulation in process or fixed voltage to
frequency relation in the DTC. The DTC approach is similar to the FOC with
hysteresis PWM. However, it takes the interaction between the three phases into
account. In the following subsections, hysteresis PWM, sinusoidal PWM and SVM
are discussed in more detail.
VoltageSource PWM Inverter 7
1.3.1 Hysteresis PWM Current Control
Hysteresis current control is a PWM technique, very simple to implement and taking
care directly for the current control. The switching logic is realized by three
hysteresis controllers, one for each phase (figure 1.4). The hysteresis PWM current
control, also known as bangbang control, is done in the three phases separately.
Each controller determines the switchingstate of one inverter halfbridge in such a
way that the corresponding current is maintained within a hysteresis band ∆i.
1/2 Udc
0
ωt
ωt
Output
voltage
0
Hysteresis band ∆i
Current reference
Real current
1/2 Udc
ia
ua0
Switching
logic
ib
*
∆i
ib
ic
*
∆i
ic
ia
*
∆i
ia
Figure 1.4: Hysteresis PWM, current control and switching logic.
To increase a phase current, the affiliated phase to neutral voltage is equal to the half
dc bus voltage until the upper bandrange is reached. Then, the negative dc bus
voltage ½ U
dc
applied as long as the lower limit is reached &c. More complicated
hysteresis PWM current control techniques also exist in practice, e.g. adaptive
hysteresis current vector control is based on controlling the current phasor in a α/β
reference frame. These modified techniques take care especially for the interaction
of the three phases [Jen 95].
Obviously, the dynamic performance of such an approach is excellent since the
maximum voltage is applied until the current error is within predetermined
boundaries (bangbang control). Due to the elimination of an additional current
controller, the motor parameter dependence is vastly reduced. However, there are
some inherent drawbacks [Brod 85]:
• No fixed PWM frequency: The hysteresis controller generates involuntary
lower subharmonics.
• The current error is not strictly limited. The signal may leave the hysteresis
band caused by the voltage of the other two phases.
• Usually, there is no interaction between the three phases: No strategy to
generate zerovoltage phasors.
• Increased switching frequency (losses) especially at lower modulation or
motor speed.
• Phase lag of the fundamental current (increasing with the frequency).
8 Chapter 1
Hysteresis current control is used for operation at higher switching frequency, as this
compensates for their inferior quality of modulation. The switching losses restrict its
application to lower power levels. Due to the independence of motor parameters,
hysteresis current control is often preferred for stepper motors and other variable
reluctance motors.
A carrierbased modulation technique, as described in the next subsection,
eliminates the basic shortcomings of the hysteresis PWM controller [Bose 97].
However, when being compared to the hysteresis PWM, an additional current
control loop, calculating the reference voltages, is required when subsequent
modulation schemes are applied to highperformance motion control systems.
1.3.2 Sinusoidal Pulse Width Modulation
Threephase reference voltages of variable amplitude and frequency are compared in
three separate comparators with a common triangular carrier wave of fixed
amplitude and frequency (figure 1.51.6). Each comparator output forms the
switchingstate of the corresponding inverter leg [Dub 89], [Leo 85]. In torque
controlled ac motor drives using sinusoidal PWM, the reference voltages (u
*
a
, u
*
b
, u
*
c
)
are usually calculated by an additional current control loop (FOC).
Switching
logic
ub
*
uc
*
ua
*
comparator
Current
controller
id
*
id
d,q
a,b,c
ud
*
Current
controller
iq
*
iq
uq
*
Carrier wave
comparator
comparator
Figure 1.5: Sinusoidal PWM, current control and switching logic.
As shown in figure 1.6, a sawtooth or triangularshaped carrier wave, determining
the fixed PWM frequency, is simultaneously used for all three phases. This
modulation technique, also known as PWM with natural sampling, is called
sinusoidal PWM because the pulse width is a sinusoidal function of the angular
position in the reference signal.
VoltageSource PWM Inverter 9
u
ab
u
a0
u
b0
u
c0
Phase b Phase a Phase c
ωt
ωt
ωt
ωt
ωt
Carrier wave
Uref
U
dc
/2
U
dc
/2
Upper switch “on”
Lower switch “on”
Figure 1.6: Principle of sinusoidal PWM generation.
Since the PWM frequency, equal to the frequency of the carrier wave, is usually
much higher than the frequency of the reference voltage, the reference voltage is
nearly constant during one PWM period T
PWM
. This approximation is especially true
considering the sampled data structure within a digital control system. Depending on
the switching states, the positive or negative half dc bus voltage is applied to each
phase. At the modulation stage, the reference voltage is multiplied by the inverse
half dc bus voltage compensating the final inverter amplification of the switching
logic into real power supply.
According to figure 1.7, the mean value of the output voltage, resulting from a
reference voltage being constant within one PWMperiod, depends on the on and
offstates of the affiliated switch:
( )
2
1 1
2 1 0
dc
PWM T
a
PWM
ao
U
t t
T
dt u
T
u
PWM
∆ − ∆ = =
∫
(1.1)
10 Chapter 1
1
u
a0
∆t
1
∆t
2
T
PWM
1
0
t
t
Udc /2
Udc /2
u
a0
*
0 0 a a
u u =
1
u
a0
∆t
1
/2 ∆t
2
T
PWM
1
0
∆t
1
/2
t
t
Udc /2
Udc /2
u
a0
*
0 0 a a
u u =
2
*
0
dc
a
U
u
2
*
0
dc
a
U
u
Sawtooth
carrier wave
Triangular
carrier wave
Figure 1.7: Sinusoidal modulation at constant or sampled reference voltage for one phase.
Left: Sawtooth shaped carrier wave. Right: Triangularshaped carrier wave.
The switch on and offtimes (∆t
1
and ∆t
2
) are calculated according to figure 1.7 by
setting the carrier wave equal to the reference voltage related to the dc bus voltage:
2
2
1
*
0
!
1
dc
a
PWM
U
u
t
T
= ∆ + − (1.2)
⇒


.

\

+ = ∆
2
1
2
*
0
1
dc
a PWM
U
u T
t (1.3)


.

\

− = ∆ − = ∆
2
1
2
*
0
1 2
dc
a PWM
PWM
U
u T
t T t (1.4)
Applying (1.3)(1.4) on (1.1) shows the mean value of the output voltage u
a0
being
equal to the reference voltage u
*
a0
:


.

\



.

\

− −


.

\

+ =
2
1
2 2
1
2 2
1
*
0
*
0
dc
a PWM
dc
a PWM dc
PWM
ao
U
u T
U
u T U
T
u (1.5)
⇒
*
0 a ao
u = u (1.6)
Apart overmodulation, this modulation technique produces an average voltage
value, equal to the reference voltage within each PWM period. Therefore, the
VoltageSource PWM Inverter 11
fundamental of the switched pulse pattern equals the corresponding reference
voltage.
The modulation technique using a sawtooth shaped carrier wave always sets the
output to a high level at the beginning of each PWM period, resulting in
asymmetrical PWM pulses. The pulses of an asymmetric edgealigned PWM signal
always have the same side aligned with one end of each PWM period. On the
contrary, the pulses of a symmetrical PWM signal, e.g. obtained by using a
triangularshaped carrier wave, are always symmetric with respect to the center of
each PWM period. The symmetrical PWM is often preferred, since it generates less
current and voltage harmonics [Bose 97], [Dub 89].
The sinusoidal PWM is easy to realize in hardware by using analog integrators and
comparators for the generation of the carrier and switching states [Ter 96]. However,
due to the variation of the reference values during a PWM period, the relation
between reference and carrier wave is not fixed. This introduces subharmonics of the
reference voltage causing undesired lowfrequency torque and speed pulsations. In
contrary, software implementation provides sampled data during a PWM period
(uniform/ regular sampling) and hence, the pulse widths are proportional to the
reference at uniformly spaced sampling times. Compared to the analog
implementation, the modulation with uniform sampling has lower lowfrequency
harmonics. Since the phase relation between reference and carrier wave is fixed,
even for the asynchronous mode, the subharmonics and the associated frequency
beats are not present [Dub 89].
The ratio of the reference magnitude to that of the carrier wave is called modulation
index m. Considering the mean output voltage equal to the reference phase voltage
(1.6) in the linear range (m ≤ 1), the fundamental component of the line voltage is:
m
U
U
U
U
dc
phase
dc
line
2 2
3
ˆ
2
3
= = , m ≤ 1 (1.7)
The boundary of the sinusoidal modulation is reached at the modulation index m = 1
(figure 1.8). For m > 1, the number of pulses becomes less and the modulation
ceases to be sinusoidal PWM. The modulation is still working, but the output
voltages are no longer sinusoidal: they correspond to the reference values with
limitation to the half dc bus voltage. The fundamental component of the line voltage
then is [Jen 95]:

.

\

− + 
.

\

⋅ =
2
1
1
1
arcsin
2 π
3
m m
m
U
U
dc
line
, m > 1 (1.8)
12 Chapter 1
0
1 2 3 4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
m [ ]
U
l
i
n
e
/
U
d
c
[
]
2 2
3
π
6
Figure 1.8: Line voltage (rms) in function of the modulation index.
When m is made sufficiently large, the phase voltage becomes a square wave and the
line voltage becomes a 6step waveform.
u
a0
ωt
u
*
2
dc
U
−
2
4
dc
U
π
u
a0
2
dc
U
−
u
b0
2
*
0
dc
a
U
u
2
*
0
dc
b
U
u
u
dc
U −
dc
U
2
dc
U
ωt
ωt
1
1
carrier wave
u
ab
fundamental
Figure 1.9: Strong overmodulation and squarewave shaped output voltage with affiliated fundamental.
Top: Reference voltages (u
*
a0
, u
*
b0
) and carrier wave. Middle: Phasetoneutral output voltage u
a0
and
affiliated fundamental. Bottom: Phasetoneutral output voltage u
b0
and line voltage u
ab
.
The square wave of the phase voltage expressed in Fouriercoefficients is:

∑
∞
−
−
=
n
dc
a
t n
n
U
u ω
π
) 1 2 ( sin
1 2
1
2
4
0
 (1.9)
Using sinusoidal PWM generation, the maximum fundamental phase voltage is
limited by the dc bus voltage:
VoltageSource PWM Inverter 13
dc phase
U U
π
2
ˆ
max ,
= (1.10)
However, this maximum voltage should not be exploited since overmodulation
results in a strong increased spectrum of lower voltage and current harmonics
especially for the 5
th
, 7
th
and 11
th
harmonics. In figure 1.10, the current of an
induction motor (scalar control) in the linear range (m = 1) and at overmodulation
(m = 1,33) is presented to illustrate the involuntary current distortion.
0 0.01 0.02 0.03 0.04 0.05
4
2
0
2
4
t [s]
i
a
[
A
]
m = 1
m = 1,33
Figure 1.10: Measured current at different modulation indexes.
(induction motor in open loop: u
ref
= 200 V sin(ωt); U
dc
= 400 V and U
dc
= 300 V resp.)
Basic drawbacks of the sinusoidal PWM are the not ideal use of the dc bus voltage
and the nonexistent interaction between the three phases resulting in superfluous
changes of switching states, increasing semiconductor losses and introducing a
higher harmonic content at the motor terminals.
1.3.2.1 Injection of a Third Harmonic
According to (1.9), also multiple of third harmonics are present in the voltage
spectrum. However, the third harmonics are eliminated and not existent in the
current spectrum since the sum of the phase current of a threephase ac machine
equals zero. As shown in figure 1.11, the range of the sinusoidal PWM can be
increased by adding third harmonics to the reference voltages. The same third
harmonic is added to each of the three reference voltages. Adding third harmonics
agrees with a simultaneous variation of the potential in all phases, thus not
recognized at the terminals of an ac motor with isolated neutral point:
) ( ) (
0 0
!
0 0 third b third a b a ab
u u u u u u u + − + = − = (1.11)
Therefore, the introduction of a third harmonic does not distort the line voltages
since third harmonic components in the phase voltages are cancelled.
14 Chapter 1
A geometrical calculation yields the maximum possible increase of the linear area
with the harmonic amplitude being 1/6 of the reference voltage amplitude. Such an
injection of a third harmonic results in a 15,5% higher maximum output voltage
without overmodulation. According to [Jen 95], the harmonic content of the
resulting current spectrum of ac motor drives is minimal at injection of a third
harmonic with the amplitude being 1/4 of the reference voltage amplitude, still
increasing the maximum output voltage without overmodulation by 12%.
u
a0
t
2
*
0
dc
a
U
u
u
*
a0
third harmonic
2
dc
U
−
2
dc
U
t
t
reference plus
third harmonic
1
1
carrier wave
Figure 1.11: Injection of a third harmonic and modulation.
Obviously, also multiple of third harmonics do not disturb the current spectrum and
are suitable injection signals. As can be shown [Jen 95], the subsequently described
space vector modulation is equal to the sinusoidal PWM with injection of a suitable
triangularshaped signal containing all existing multiple of third harmonics.
1.3.3 Space Vector Modulation (SVM)
Spacevector pulse width modulation has become a popular PWM technique for
threephase voltagesource inverters in applications such as control of induction and
permanent magnet synchronous motors. The mentioned drawbacks of the sinusoidal
PWM are vastly reduced by this technique. Instead of using a separate modulator for
each of the three phases, the complex reference voltage phasor is processed as a
whole. Therefore, the interaction between the three motor phases is exploited. It has
been shown, that SVM generates less harmonic distortion in both output voltage and
current applied to the phases of an ac motor and provides a more efficient use of the
supply voltage in comparison with direct sinusoidal modulation techniques [Jen 95].
VoltageSource PWM Inverter 15
As shown in table 1.1, there are eight possible combinations of on and off patterns
for the three upper electronic switches feeding the threephase power inverter
(figure 1.1). Notice that the on and off states of the lower power switches are
opposite to the upper ones and so completely determined once the states of the upper
power electronic switches are known. The phase voltages corresponding to the eight
combinations of switching patterns can be mapped into the α/β frame through α/β
transformations [Hen 92]. This transformation results in six nonzero voltage vectors
and two zero vectors. The nonzero vectors form the axes of a hexagonal containing
six sectors (S
1
− S
6
) as shown in figure 1.12. The angle between any adjacent two
nonzero vectors is 60 electrical degrees. The zero vectors are at the origin and apply
a zero voltage vector to the motor. The derived α/β voltages in terms of the dc bus
voltage U
dc
are summarized in table 1.1.
Table 1.1: Switching table and α/β transformation of affiliated state voltage vectors.
3
dc
U 3 2
dc
U
3
dc
U
3
dc
U
3 2
dc
U
3
dc
U 3
dc
U 3 2
dc
U
3 2
dc
U − 3 2
dc
U
3
dc
U − 3
dc
U − 3 2
dc
U
3
dc
U
3
dc
U −
3 2
dc
U
Switch no. α/βtransformation of the states
State S1 S3 S5 U
x,α
U
x,β
U
x

000 OFF OFF OFF 0 0 0
100 ON OFF OFF 2 0
110 ON ON OFF
010 OFF ON OFF
−
011 OFF ON ON 0
001 OFF OFF ON
101 ON OFF ON
111 ON ON ON 0 0 0
100
110
101 001
011
010
U
β
000
111
U
α
S6
S2
S1
S5
S4
S3
Figure 1.12: Hexagon, formed by the basic space vectors and sector definition (S
1
− S
6
).
Using the transformation of the three phase voltages to the α/β reference frame, the
voltage phasor U
ref
represents the spatial phasor sum of the three phase voltages.
When the desired output voltages are threephase sinusoidal voltages with 120°
16 Chapter 1
phase shift, U
ref
becomes a revolving phasor with the same frequency and a
magnitude equal to the corresponding linetoline rms voltage.
The objective of the space vector PWM technique is to approximate the reference
voltage phasor U
ref
by a combination of the eight switching patterns. Practically,
only the two adjacent states (U
x
and U
x+60
) of the reference voltage phasor and the
zero states should be used [Jen 95] as demonstrated by the example in figure 1.13.
The reference voltage U
ref
can be approximated by having the inverter in switching
states U
x
and U
x+60
for t
1
and t
2
duration of time respectively.
(
60 2 1
1
+
+ =
x x
PWM
ref
U t U t
T
U ) (1.12)
Of course, the affiliated sector must be known first. The sector identification and the
calculation of t
1
and t
2
are presented in the next subsection. Since the sum of t
1
and t
2
should be less than or equal to T
PWM
, the inverter has to stay in zero state for the rest
of the period. The remaining time t
0
is assigned to one or both zerovoltage phasors.
2 1 0
t t T t
PWM
− − = (1.13)
Applying only one of the two zerovoltage states during a PWM period, results in an
asymmetric edgealigned PWM signal. This is often undesired (higher harmonics)
but reduces the required switching number by 33% since one inverter leg does not
switch during that particular PWM period. Here, the remaining time t
0
is equally
assigned to both states. As illustrated in figure 1.13, all state changes are obtained in
each case by switching only one inverter leg.
40% ‘100’
50% ‘110’
U
ref
= U e
jωt
100 110 110 100
111 000 000
T
PWM
u
a0
u
b0
u
c0
5% ‘000’
5% ‘111’
Figure 1.13: Example of dutycycle generation.
As mentioned above, the reference voltage is actually equal to the desired three
phase output voltages mapped to the α/β frame. The envelope of the hexagonal
formed by the basic space vectors, as shown in figure 1.12, is the locus of the
maximum output voltage. In order to avoid overmodulation, the magnitude of U
ref
must be limited to the shortest radius of this envelope. This gives a maximum rms
value of the linetoline and phase output voltages of
VoltageSource PWM Inverter 17
2
ˆ
2
3
max , max ,
dc
phase line
U
U U = = (1.14)
being approximately 15% higher when compared to the original sinusoidal PWM.
1.3.3.1 RealTime Implementation of the SVM
Presently in industry, the SVM is often applied as inverter control strategy because
of its advantages when compared to other PWM techniques: SVM provides efficient
use of the supply voltage and low harmonic distortion in both output voltage and
current. Furthermore, it can easily be implemented with modern DSPbased control
systems. Even recent developments of the DTCalgorithm are modified in regard to
exploit the advantages of the SVM.
As shown in table 1.1, the reference voltage U
ref
, usually represented by its α/β
components U
α
*
and U
β
*
, can be approximated easily by a linear combination of the
two adjacent states and the zero states, i.e. no trigonometric functions are required to
calculate the duty cycles. First, the sector must be identified to determine the
appropriate states. This is performed, as illustrated in figure 1.14, by a comparison
of the α/β components specifying the position in the α/βplane. For instance, if the
reference voltage U
β
*
is positive, the sector of the reference voltage is in the upper
half of figure 1.12 (sector S1, S2 or S3). Otherwise, the sector is in the lower half.
Further sector splitting/identification is done by comparison (geometrical
calculation) of the α and βcomponents. The applied normalization at the beginning
eliminates the dc bus voltage dependence of the output voltages. The resulting duty
ratios (a
*
, b
*
and c
*
), as required for PWM generation using e.g. TI’s TMSM320P14
DSP, are calculated according to following flowchart. A duty ratio a
*
= 1 indicates a
continuously closed upper switch of the first inverter leg. At a duty ratio a
*
= 0, the
turnon time during each PWM period is equally distributed to the lower and upper
switch and the resulting mean value of the phase voltage u
a0
is zero. At a duty ratio
a
*
= 1, the lower switch is continuously closed, etc.
The relation between the duty cycles of the three phases in percent (relation of the
switchon to the switchoff times of the three inverter legs within one PWM period)
and the given duty ratios a
*
, b
*
and c
*
is:
% 100
2
1
;
2
1
;
2
1
* * *
; ;
+ + +
=
c b a
cycles duty
c b a
(1.15)
Usually, the presented algorithm is easily incorporated into the initialization part of
the realtime program, e.g. by including handwritten Ccode. Then, the duty ratios
are directly mapped by a DSP into signals driving the inverter switching logic. As
illustrated in figure 1.14, a final data processing and transmission is required, when
18 Chapter 1
additionally a slave DSP generating the PWM pulses, e.g. TI’s TMSM320P14, is
used. To avoid overflow of the fixedpoint slave DSP, all duty ratios must be limited
to ± 1. Since the P14 DSP uses 16bit compare registers for the PWM generation,
the calculated values are adjusted by the given multiplication before they are finally
transmitted to the slave DSP generating the PWM signals. As illustrated in a
subsequent chapter (e.g. figure 3.2), each two PWM channels are employed to
generate the correct pulses for the inverter.
Sector 2 & 5: Sector 1 & 4: Sector 3 & 6:
Sector 1 Sector 2 Sector 3 Sector 4 Sector 5 Sector 6
Yes
No
No
No
No
No
P14
DSP
u ≤ 1
Overflow
protection
16 bit
compare
register
normalization
PWM 1−6
U
α
*
, U
β
*
2
15
1
dc
U
1
2
3
0
*
≥
β
u
* *
3
1
β α
u u ≥
* *
3
1
β α
u u
−
≥
* *
3
1
β α
u u ≥
*
3
1
β α
u u
−
≥
* * *
3
1
β α
u u a + =
* * *
3
3
β α
u u b + − =
* *
a c − =
* *
2
α
u a =
* *
3
2
β
u b =
* *
b c − =
* * *
3
1
β α
u u a − =
* *
a b − =
* * *
3
3
β α
u u c − − =
duty ratios:
(a
*
; b
*
; c
*
)
Voltage
reference
Figure 1.14: Flowchart of SVM and data transmission to a TMSM320P14 DSP.
The turnon times t
0
, t
1
and t
2
of the applied switching states during each PWM
period, as introduced in (1.12)(1.13) for illustration purpose, are not required for
implementation of the SVM. However, they are easily calculated by the duty cycles
of the three phases. For instance, the zero states ‘000’ and ‘111’ are each equal to
the minimum of the duty cycles given in (1.15) multiplied by the PWM period T
PWM
.
VoltageSource PWM Inverter 19
1.4 DeadTime Effect & Voltage Distortion
For voltagesource PWM inverters, a deadtime interval is required to prevent the
“shootthrough” effect of a halfbridge during a change of the switching states. All
semiconductorswitching devices react delayed to the turnoff signals owing to the
storage time. During this storage time, depending on the operating point, the switch
is not able to block the dc link voltage. Therefore, to avoid a short circuit of the half
bridge, a deadtime interval must be introduced between the turnoff signal of a
switch and the turnon signal controlling the opposite switch. The dead time τ
dead
is
usually constant and determined as the maximum value of storage time τ
st
plus an
additional safety margin. The dead times of common IGBTinverters used in
industry vary between τ
dead
≈ 15 us.
Although the dead time is short, it causes deviations from the desired fundamental
inverter output voltage. The effects of the dead time on the output voltage will be
described from one halfbridge of the PWM inverter according to figure 1.15. The
basic configuration consists of upper and lower power devices T1 and T4, and
reverse recovery diodes D1 and D4.
½ Udc
½ Udc
UT1
UD4
u
a
0
Ideal gating
pulse pattern
T1
T4
pulse pattern
with dead time
T1
T4
τdead
½ Udc
½ Udc
UD1
UT4
u
a
0
T1
T4
T1
T4
τdead
Ideal gating
pulse pattern
pulse pattern
with dead time
τdead fPWM Udc
τdead fPWM Udc
u
a0
T
1
T
4
C/2
½ U
dc
D
4
D
1
i
a
> 0
C/2
½ U
dc
T
1
off
T1 on
u
a0
T
1
T
4
D
4
D
1
i
a
< 0
C/2
C/2
½ U
dc
½ U
dc
T
4
off
T
4
on
Figure 1.15: Error voltage due to the deadtime effect.
Left: Positive load current. Right: Negative load current.
Considering the noload case, the storage time of the semiconductors is very small
when compared to the dead time: Switching off a power device, the current
20 Chapter 1
commutates directly to the diodes. This condition results in the desired voltage,
which is applied to the motor terminals. In contrast to this, switching on a power
device is delayed by the dead time. During the deadtime interval, the diode
continues conducting until the dead time elapses and the opposite power device is
switched on. This condition results in a loss of voltage at the motor terminals
indicated by the gray marked area in figure 1.15. With a positive current, the duty
cycles are shorter and with negative currents longer than required. Hence, the actual
duty cycle of a bridge is always different from the one of the reference voltage. It is
either increased or decreased, depending on the load current polarity. Furthermore,
the voltage drops of the power switches U
T
, respectively the voltage drop of the
reverse recovery diode U
D
, are considered. Summarizing, the voltage distortion can
be described by an error voltage ∆U
dc PWM dead
D T
U f
U U
U ⋅ ⋅ +
+
≈ ∆ τ
2
(1.16)
depending on the dead time τ
dead
, the dc bus voltage U
dc
, the PWM frequency f
PWM
and the voltage drops U
D
and U
T
of IGBT and diode [Bose 97]. This error voltage
and the resistances R
T
and R
D
of the switch changes the inverter output voltage u
a0
from its intended value u
ref
to:
) sign(
2
0
i U
R R
i u u
D T
ref a
⋅ ∆ −
+
⋅ − ≈ , (1.17)
The dead time reduces the effective turnon time and produces the undesired fifth
and seventhorder harmonics in the inverter output voltage [Dod 90]. Furthermore, it
generates subharmonics, resulting in torque pulsation and possible instability at
lowspeed and lightload operation [Leg 97], [Mur 92]. The resulting speed
oscillation and the voltage distortion are illustrated in figure 1.16 showing the dead
time effect on a 1,5 kW induction motor drive in open loop (scalar) control at low
speed and lightload operation. Considering the given drive setup (τ
dead
= 2,5 us;
f
PWM
= 10 kHz) and according to (1.16), the error voltage amounts to ∆U = 12,5 V
(equal to 15,3 V in the alpha/beta reference frame). A reduction of the average
voltages occurs according to (1.17) when one of the phase currents changes its sign.
The motor currents have the tendency to maintain their values after a zero crossing.
In generator mode, the behavior of the motor current is contrary resulting in a
steeper rise of the current after zero crossing. In any case, the motor torque is
influenced as it can be observed by speed oscillations at six times of the fundamental
frequency.
The deadtime problem is more serious in highpower gateturnoff thyristor (GTO)
inverter systems than in the case of IGBT or MOSFET inverters, since the GTO
requires a longer dead time. However, the use of fast switching devices using high
carrier frequencies (520 kHz) with lower deadtime values (15 us), will not free
the system of the described distortion. Higher PWM frequencies improve the
VoltageSource PWM Inverter 21
waveform quality by raising the order of theoretical harmonics, but low frequency
subharmonics persist due to the dead time. Furthermore, to avoid unnecessary
switching losses and shortterm overheating of a switching device, minimum time
duration in the switching states must be forced. If the commanded voltage value is
less than the required minimum, the affiliated switching state must be either
extended in time or skipped. This causes additional distortion of the inverter output
voltages. Therefore, a compromise must be made by choosing a suitable PWM
frequency: a high PWM frequency improves the theoretical quality of the waveform,
but may increase simultaneously the voltage distortion due to the deadtime effect.
0 0.25 0.5 0.75 1
40
20
0
20
40
t [s]
U
α
[
V
]
0 0.25 0.5 0.75 1
4
2
0
2
4
t [s]
I
α
[
A
]
0 0.25 0.5 0.75 1
35
40
45
50
55
60
t [s]
n
[
r
p
m
]
40 20 0 20 40
40
20
0
20
40
U
α
[V]
U
β
[
V
]
u
ref
u
β
(u
α
)
u
α
u
ref
Figure 1.16: Openloop control of an induction motor & deadtime effect (U
dc
=500 V, no load).
Left: Measured current I
α
, measured voltage U
α
and reference voltage U
ref
.
Right: Measured/reference speed and voltage trajectories.
1.4.1 DeadTime Compensation
Remarkable efforts have been made to compensate the voltage distortion due to the
deadtime effect [Choi 96], [Leg 97], [Sep 94]. Most deadtime compensation
methods are based on an average value theory: the lost voltage is averaged over an
operating cycle and added vectorially to the command voltage [Mur 87], [Jeo 91].
Deadtime compensation can be implemented in hardware or software.
The hardware compensator operates by closed loop control [Mur 87]. Previous
commutation times are measured and used to control the next duty cycles. However,
a potentialfree measurement of the inverter output voltages is required. Software
compensators are mostly designed in feedforward mode. Depending on the sign of
the respective phase current, a fixed time delay is either added to or subtracted from
the command voltage.
22 Chapter 1
However, a complete compensation of the deadtime effect may not be achieved
since the actual storage delay is not exactly known. Furthermore, the PWM
generation is a part of a superimposed highbandwidth current control loop
compensating the involuntary torque/speed distortions to a certain extent. This may
eliminate the need for a separate deadtime compensator. Figure 1.17 illustrates the
deadtime effect on an induction motor drive in fieldoriented speed control mode at
low speed and lightload operation. Except the control mode, the conditions are the
same as in figure 1.16.
0 0.25 0.5 0.75 1
40
20
0
20
40
t [s]
U
α
[
V
]
0 0.25 0.5 0.75 1
4
2
0
2
4
t [s]
I
α
[
A
]
0 0.25 0.5 0.75 1
35
40
45
50
55
60
t [s]
n
[
r
p
m
]
40 20 0 20 40
40
20
0
20
40
U
α
[V]
U
β
[
V
]
u
ref
u
β
(u
α
)
u
α
u
ref
Figure 1.17: Fieldoriented control of an induction motor & deadtime effect (U
dc
=500 V, no load).
Left: Measured current I
α
, measured voltage U
α
and reference voltage U
ref
.
Right: Measured/reference speed and voltage trajectories.
The influence of the dead time on the current/torque is vastly reduced by the speed
and current control loop. Of course, the falsification of the motor terminal voltages
is the same, but the harmonic distortion of the fundamental voltage is transmitted to
the reference voltages. Due to the arguable compensation by the current controller,
common industrial drives are not always equipped with an additional deadtime
compensation.
Note, that permanent magnet synchronous motor drives behave more sensitive to the
deadtime effect than induction motor drives: Due to the absence of a magnetizing
component in the stator current and the low main reactance, they tend to operate
partly in discontinuous current mode at light load. These machines require an
advanced compensation scheme when applied to highperformance motion control
systems or, alternatively, an additional daxis current to bridge the discontinuous
current time intervals [Bose 97].
VoltageSource PWM Inverter 23
1.4.2 DeadTime Generation
The switching transitions of real switches, especially the transition from current
conducting to voltage blocking, are not infinitely fast. After conducting, a finite time
is required, mainly to remove the space charge, before a semiconductor switch is
able to block the supply voltage. Switching off a power device, the current
commutates to the opposite recovery diode (constant current direction) and the
power switch starts to block the dc voltage. If a switch of one inverter leg is turned
on before the opposite switch blocks the dc bus voltage, the whole dc bus voltage is
shorten across this leg (figure 1.1) resulting in a very high shortcircuit current only
limited by the resistances of the power switches. Obviously, such a high shortcircuit
current may destroy the power switches as well as the drive system and the dc link
capacitor.
To avoid such shortcircuit conditions, a deadtime interval is added between the
turnoff signal of a switch and the turnon signal controlling the opposite switch.
Dead time control prevents any crossconduction or shootthrough current from
flowing through the main power switches during switching transitions by controlling
the turnon times of the semiconductor drivers. The highside driver is not allowed
to turn on until the voltage at the junction of the opposite power switch is low and
vice versa. During the deadtime interval, recovery diodes continue conducting until
the dead time elapses and the opposite power device is switched on.
In modern DSP systems, the dead time generation is usually programmable, e.g.
added as extra time in a compare register/timer. Considering analog circuits, the
fixed deadtime generation of one halfbridge is easily generated by a RCcircuit
coupled to two optocouplers, each controlling the opposite switches of one inverter
halfbridge as described in figure 1.18. Additionally, such a hardware realization
takes care for galvanic isolation of the digital control system and the power
electronics. The resistance R is calculated by the resistance voltage drop divided by
the operating current of the optocoupler I
P
:
P
d s
I
U U
R
−
= (1.18)
According to figure 1.18, changing the switching signal U
s
from a positive to a
negative voltage (e.g.: U
s
= ±12V) results in a discharging of the capacitor
depending on the photodiode operating voltage (U
d
≈ 1V if i > 0). While the
photodiode P1 directly blocks, the deadtime τ
dead
passes before the capacitor
voltage equals the voltage U
d
, equal to the on state of photodiode P2 driving the
opposite switch of the inverter leg:
d d
C R
d s dead c
U U e U U t U
dead
− = +


.

\

− + = =
− !
1 ) ( ) (
τ
τ (1.19)
24 Chapter 1
Thus, a minimum capacitor value is required to guarantee the deadtime interval
τ
dead
:
⇒


.

\

+
−
≥
d s
d
dead
U U
U
R
2
1 ln
C
τ
(1.20)
Optocoupler 1
PWM logic:
US = ±12V
R
C
IP2
IP1
UC
US
IP
IP2 IP1 IP1
τdead
t
Switching logic
Ud
UC
Ud
Us
UC
t
t
τdead
Optocoupler 2
Figure 1.18: Analog deadtime generation. Left: Exemplary hardware circuit for one inverter leg.
Right: Switching logic, voltage and affiliated current of an optocoupler driving the power switch.
1.5 PWM Inverter Drives and Motor Insulation
Variable speed ac drives are used in everincreasing numbers because of their well
known benefits for energy efficiency and for flexible control of processes and
machinery using lowcost readily available maintenancefree ac motors. While the
connection of a motor to an inverter supply is straightforward, some basic
considerations are necessary to ensure trouble free longterm operation. Insulation
performance is one of the considerations required in engineering variable speed
drive solutions. Following summary provides basic information to enable the correct
matching of low voltage ac motors and PWM inverters with respect to motor
insulation:
Motor winding insulation experiences higher voltage stresses when used
with an inverter than when connected directly to the ac mains supply.
The higher stresses are dependent on the motor cable length and are caused
by the fast rising voltage pulses of the drive and transmission line effects in
the cable.
For supply voltages less than 500V ac, most standard motors are immune to
these higher stresses.
VoltageSource PWM Inverter 25
For supply voltages over 500V ac, a motor with an enhanced winding
insulation system is required. Alternatively, additional components can be
added to limit the voltage stresses to acceptable levels.
Where the drive spends a large part of its operating time in braking mode,
the effect is similar to increasing the supply voltage by up to 20%.
For drives with PWM active front ends (regenerative and/or unity power
factor), the effective supply voltage is increased by around 15%.
1.6 Conclusions
Controlled power supply for electric drives is obtained usually by converting the
mains ac supply. A typical converter consists of power electronic circuits,
employing switching devices such as thyristors, transistors, GTOs, MOSFETSs,
IGBTs and diodes as well as a host of associated control and interfacing circuits.
The conversion process allows fast control of voltage, current or power to the motor
via the gate circuits of the converter switches. In this way, the required dynamic
response requirements of highperformance ac motor drives can be met.
This chapter provides a detailed survey of voltagesource PWM inverter drives with
emphasis on the modulators and control methods. The most common threephase
inverter topology is that of a switch mode voltage source inverter. VSinverters
consist of two main sections, a controller to set the operating frequency and a three
phase inverter to generate the required sinusoidal threephase voltage from a dc bus
voltage.
The basic concepts of pulse width modulation are illustrated. PWM is the process of
modifying the width of the pulses in a pulse train in direct proportion to a small
control signal. The greater the control voltage, the wider the resulting pulses
become. By using a sinusoid of the desired frequency as control voltage for a PWM
circuit, it is possible to produce a highpower waveform whose average voltage
varies sinusoidally in a manner suitable for driving ac motors. Due to the significant
flexibility in controlling the inverter switches, a large number of switching
algorithms were introduced and some of these have gained wide acceptance and are
fully developed.
Usually, the behavior of the power devices together with the reverse recovery diode
is described by ideal twoposition switches. In practice, a deadtime interval is
required to prevent the “shootthrough” effect of a halfbridge during a change of the
switching states. Although the dead time is short, it causes deviations from the
desired fundamental inverter output voltage. Issues of the resulting phase voltage
distortion due to the inverter nonlinearity as well as compensation methods are
discussed in detail.
2. Regenerative Braking and Ride
Through at Power Interruptions
2.1 Introduction
Voltage dips and sags of short duration constitute a serious problem for many
applications and especially for variable speed drives (frequency converters) in
industry. Early types of frequency converter for motor drives were notoriously
sensitive to supply disturbances and often had to perform a full stop and restart to
resume operation. The economic impact, of what actually is a mere incident,
therefore could turn out to be quite substantial.
Usually, voltage source PWM inverter drives are equipped with an undervoltage
protection mechanism, causing the system to shut down within a few milliseconds
after a power interruption in the regular grid. This shut down mechanism can be
associated with a total loss of system control since the control electronics are usually
powered by the (in this case discharged) dclink capacitor. Particularly in multi
motor drives, a loss of mutual synchronization may be critical. This may entail
damage or loss of material in sensitive applications as the production of textile
fibers, paper mills, or extrusion drives. Generally, it is required to wait until the
machine has come to a complete standstill to enable restarting [Baa 89]. Braking to
zero speed and restarting obviously is not an adequate solution. Many continuous
production processes in industry are sensitive to a larger variation in speed or losing
control at worst. In addition, time and additional workload required to get a plant
ready for restart may be considerable.
This chapter discusses a design concept avoiding the standstill/restarting interval at
power interruptions. The proposed solution to the problem is to recover some of the
mechanical energy stored in the rotating masses by kinetic buffering. When the
power supply is interrupted, a dc link voltage control is applied to force an
immediate transition into the regenerative mode. During the interruption interval, the
drive system continues to operate at almost zero electromagnetic torque, just
regenerating a minor amount of power to cover the electrical losses in the inverter.
This maintains the dc link capacitor well charged, keeping the electronic control
circuits active, since they are supplied from the dc link through a switched mode
converter. In this way, the drive remains controllable even at power interruptions of
several seconds. Of course, the (still controlled) braking of the drive depends on the
28 Chapter 2
actual load torque. Since drive control is never lost, the voltage control scheme can
be applied to multi motor drives as well.
The implemented regenerative braking scheme allows the inverter to keep its dc bus
voltage at a predetermined minimum level as long as possible, expanding the time in
which supply voltage can be reapplied without the timeconsuming dclink capacitor
recharging cycle. The temporary speed dip is generally tolerable, since the most
frequent power interruptions last only for a few milliseconds. The implemented
voltage control scheme is derived from a torque controlled dc bus voltage.
Considering realistic conditions, the ridethrough capability at shorttime power
interruptions is discussed. Measured results are presented and evaluated to
demonstrate the performance and the stability of the system.
2.2 Voltage Dips
A voltage dip is a shortduration reduction in the supply voltage, in many cases due
to network faults somewhere in the energy distribution system. During a voltage dip,
the voltages in the three phases are no longer the same, causing a number of
problems. A major fault more than 100 km away from a customer may still yield a
significant voltage dip. Mains voltage dips and short interruptions are caused by a
wide variety of phenomena. They can be caused by nearby events, such as a faulty
load on an adjacent branch circuit causing a circuit breaker to operate, or perhaps by
a large motor or other large load on the same circuit being switched on. They can
also be caused by far away events such as lightning strokes or downed power lines.
In case of a fault in the power distribution grid, an automatic circuit recloser may
cycle open and close several times within a short period attempting to clear the fault,
thus resulting in a sequence of short interruptions noticed by downstream loads. In
any case, the voltage changes produced can affect the operation of or even damage
nearby electrical equipment as e.g. drives. Therefore, immunity for these types of
events should be available to ensure safe and reliable product operation.
Voltage dips are probably the power quality disturbance with the highest impact on
customers. The voltage drop yields tripping of process control equipment such as
adjustablespeed drives, process computers and switchgears. This in turn leads to
production halts, lasting much longer than the dip itself. Voltage dips of 100 ms
duration can lead to production halts of 24 hours or more. The economic impact per
event may be less than for regular interruptions, but the annual impact is in many
cases higher.
An ac motor directly connected to the regular grid may slow down during such a
power failure. An airgap flux wave may be still in existence, but its magnitude,
phase angle and speed changes. Then, a return of the voltage with inadequate values
necessarily produces large current/torque transients. As has been reported by
industrial users, these transients generated by the motor may even cause a break of
the drive shaft. However, this problem can be overcome using a simple relay as a
Regenerative Braking and RideThrough at Power Interruptions 29
watchdog or overcurrent protection. Nevertheless, a timeconsuming restart or other
special mechanisms may be required.
Concerning motor drives supplied by voltage source inverters, a dip on all three
phases leads to an instantaneous decrease of the dc link voltage, whereas a single
phase dip may allow continued operating of the drive, albeit at higher rectifier stress.
Rectifier bridges must be properly designed to withstand these high peak currents.
Due to advances in semiconductor technology, modern variable speed drives can
tolerate the high peak currents occurring when the power supply is restored after a
short disturbance. Furthermore, powerful digital signal processors enable drive
manufacturers to implement regenerative braking schemes allowing the inverter to
keep its dclink voltage at a required minimum level.
The availability of electrical power from the public supply as a function of the down
time at interruptions (in Germany) is given in [Sch 85] indicating that a power
interruption of more than 10 ms is likely to occur every 200 h, on average. Against
this, the mean times between failures due to long time power interruptions are of the
order of several 10 000 h. Short time interruptions of the power supply are therefore
the most frequent cause for inverter failure. A ridethrough scheme at these short
time power interruptions is presented in the next subsection.
2.3 RideThrough Scheme
A relatively large electrolytic capacitor (1001000 uF / kW) is usually inserted in the
dc link to stiffen the dc bus voltage and provide a path for the rapidly changing
currents drawn by the inverter. However, the amount of energy stored in the dc link
capacitor is normally insufficient to maintain the inverter active during a short
power failure interval. When a power interruption occurs, the dclink energy is
absorbed by the motor within a few milliseconds. Since the electronic control
system loses power as well, the inverter shuts down commanded by an under
voltage protection in order to avoid possible damage to the electronic or drive
equipment [Baa 89]. It is then required to wait until the machine has come to a
complete standstill to enable restarting. However, timeintensive restarting is
obviously not an adequate solution.
One approach to avoid the standstill interval following a power interruption is
described in [Sei 92]. The control scheme is applicable to generalpurpose inverters
with scalar motor control. Although this scheme can catch a running machine, the
time required for synchronization (up to 6 s) is too long for many critical
applications. It becomes even more severe with multi motor drives. Here, a solution
is presented using the high dynamic performance of a fieldoriented motor control.
The dc link capacitor is a major cost item in the drive system and an increase of the
capacitor is therefore economically not feasible [Bose 97]. In contrast, the kinetic
30 Chapter 2
energy of the moving masses of motor and driven system is substantially higher.
This reservoir can be tapped for bridging the time interval of power interruptions.
Energy is fed back from the rotating masses to the dc link circuit to maintain the dc
link voltage at a predetermined level. This is possible also in the presence of
additional loads connected to the dc link.
The principle of forcing a fast reversal of power flow at a breakdown of the supply
voltage is explained by the trace of the dc bus voltage according to figure 2.1.
Normally, the dc voltage changes within certain limits as indicated by the (shaded)
regular voltage band. The lower limit allows for voltage sags due to load variations,
fluctuations of the supply voltage or singlephase voltage dips. The upper voltage
limit may be reached at fast decelerations of the drive. Normally, a rising dc voltage
forms no problem since the generated kinetic energy can be conducted using a
brakeresistance within the dc link, a common dc bus or a twoway PWM inverter.
Nevertheless, the proposed ridethrough scheme can be adopted allowing a
controlled deceleration within a maximum predetermined dc link voltage. This can
be used to save energy rather than a fast deceleration with power dissipation of, e.g.,
a brakeresistance. First, only a low voltage ridethrough scheme bridging the time
of a threephase power interruption is considered. The latter approach is presented at
the end of this chapter.
With reference to figure 2.1, the power supply is interrupted at t = t
1
. The power
interruption is detected at t
2
when the dc bus voltage reaches the predetermined level
U
KB
causing the system to switch automatically to voltage control mode. Thereafter,
the voltage is maintained by a closed loop control forcing the drive system to
operate at almost zero electromagnetic torque. The motor regenerates just a minor
amount of power by kinetic buffering to cover the electrical losses in the inverter
and motor until the return of the power supply at t
4
. The return of the power supply
results in a fast rise of the dc link voltage. This reactivates the regular speed control
of the drive at t
5
and the motor accelerates to the set value.
U
dc,N
U
KB
U
min
t t
1
t
2
t
4
Undervoltage
protection
t
5
t
Regular voltage band
0
1
D
i
p
d
e
t
e
c
t
i
o
n
t
3
U
dc
Figure 2.1: Controlled dc bus voltage during power interruptions.
The lower trace of figure 2.1 shows a logic signal indicating the detected event of a
power interruption. This signal is used in order to switch between voltage and speed
Regenerative Braking and RideThrough at Power Interruptions 31
control mode. If the inverter control did not react on this signal, the dc bus voltage
continues falling as indicated by the dashed line. The inverter would shut down at t
3
by the undervoltage protection at the voltage level U
min
. Without kinetic buffering,
the maximum acceptable duration of a power interruption can be determined by
( ) (
∫
+ = −
3
1
2
min
2
,
2
1
t
t
load loss N dc
dt T P U U C ω ) (2.1)
where P
loss
is the power dissipation of motor and inverter. In speed control mode, the
motor speed and consequently the losses as well as the load torque are usually
constant. Typical values of this time interval, mainly depending on the prevailing
mechanical power at the motor shaft, are of the order of 150 ms. Of course, the
voltage control must become active before this time has been elapsed. The
maximum time interval ∆t
max
of bridging power interruptions by kinetic buffering
can be appraised by solving:
( ) (
∫
∆
+ = + −
max
2 2
min
2
,
2
1
2
1
t
load loss ref N dc
dt T P J U U C ω ω ) (2.2)
In contrast to (2.1), losses and load torque are now speed dependent. The stored
kinetic energy is obtained by the inertia of the moving masses and the actual speed
at power interruption, normally equal to the reference speed ω
ref
. Using kinetic
buffering, a maintained and controlled operation of several seconds is possible.
2.4 DC Bus Voltage Control
Primarily, the proposed voltage control scheme at power interruptions has been
developed for a PVpowered water pump system [Ter 02]. There, the voltage control
is designed to withstand abrupt power interruptions, occurring at an instantaneous
decrease of the irradiance intensity (e.g. passing clouds). The total power failure
considered here can be regarded as a worstcase situation.
The most important control loop for the stability of the entire system is the dc bus
voltage control. The system has been set up to work independently in island
operation. All control and measurement units are supplied by the dc bus. A dc
voltage beyond given limits leads inevitably to a crash of the entire system. The
voltage reference is calculated by an overlaid MPPTracking and controlled directly
or indirectly by the speed of the motor.
Due to the lack of a major storage element in the dc bus, the power of the PV array
must be used immediately to accelerate the PMSM. As irradiance increases,
resulting in a higher output power of the PV array, the input power of the dc bus is
32 Chapter 2
higher than the output. The voltage control must immediately accelerate the PMSM
to stay in the MPP of the PV array. With decreasing irradiance, the power of the PV
array is smaller than the output in the dc bus. The difference comes from the
capacitor, being discharged. This is the most critical condition. The dc bus collapses,
if this condition remains resulting in a voltage drop beyond given limits. Hence, the
inverter must slow down the PMSM to a new stable operating point. Therefore, the
voltage controller has to accelerate/decelerate the motor very quickly guaranteeing a
balanced input/output power ratio in the dc bus. Figure 2.2 shows the energy flow
within the system without loss considerations.
I
PV
I
Inv
Solar
generator
U
dc
I
dc
motorpump
system
P
kinetic
P
pump
P
PV
Figure 2.2: Energy flow of the PVpowered water pump systems.
The energy generated by the PV array is used to drive the motor/pump system.
Depending on the difference between energy generation and consumption, the dc
bus capacitor is charged or discharged:
dt I
C
U
dc dc
∫
=
1
(2.3)
The dynamic behavior of the voltage control is determined by energy equations. The
electromagnetic power developed by the motor can be divided in kinetic power
P
kinetic
accelerating the motorpump system and pumping power P
pump
. Only the
kinetic power can be used to feed back energy to the dc bus and to control the
voltage.
load el
T
dt
d
J T + =
ω
(2.4)
load pump kinetic el el
T
dt
d
J P P T P ω
ω
ω ω + = + = = (2.5)
Subsequently, the drive efficiency is not taken into account, because of the opposite
influence at acceleration and braking. The losses are small compared to the
mechanical energy consumption. Furthermore, the loss fluctuation is almost as
slowly as the variation of the pumping power. Therefore, they are as being a part of
Regenerative Braking and RideThrough at Power Interruptions 33
the load. Without considering the drive efficiency, the input power of the inverter
matches the electromagnetic output power generated by the motor.
(
dc PV dc Inv dc el
I I U I U P ) − = ≈ (2.6)
In steady state, the voltage U
dc
and motor speed ω are constant. The energy
generated by the PV array is completely used to pump water:
• ⇒ const =
dc
U 0 =
dc
I (2.7)
• const = ω ⇒ (2.8)
¹
´
¦
≈
=
PV dc pump
kinetic
I U P
P 0
2.4.1 Speed controlled dc bus voltage
Normally, the motor speed of a conventional drive supplied by a regular grid via a
diode rectifier is completely independent of the dc bus voltage. Here, a PV array is
the source and a water pump acts as load. A relation between motor speed and dc
bus voltage can be obtained by linearization of the dynamic behavior. The
electromagnetic torque of the motor can be controlled very fast given the bandwidth
of the current control loop (960 Hz), whereas the load torque varies slowly with
speed. The speed can be controlled beyond current/torque limitation with a
bandwidth of approximately 26 Hz. Therefore, also the kinetic power P
kinetic
can be
varied faster than the pumping power P
pump
. Due to similar considerations, the dc
current I
dc
can be controlled faster than the dc bus voltage U
dc
. Therefore, the
following equation is valid during transients:
dt
d
d
dT
dt
d
J
load
ω
ω
ω
>>
2
2
(2.9)
Using (2.5)(2.6) and assuming constant pumping power and constant current I
PV
of
the PV array for a short time, the linearized relation between dc voltage and motor
speed ω is described by:
load dc dc PV dc
T
dt
d
J I U I U ω
ω
ω + = − (2.10)
const I U T P
PV dc load pump
≈ ≈ = ω (2.11)
dc dc kinetic
I U
dt
d
J P − ≈ =
ω
ω (2.12)
34 Chapter 2
dt
d
J
dt
dU
C U
dc
dc
ω
ω − = ⇒ (2.13)
With the transfer function of the closed loop speed control beyond current/torque
limitations
1
1
) (
) (
*
+
=
speed
s s
s
τ ω
ω
(2.14)
and using (2.13), the resulting linearized transfer function with the reference speed
ω
*
as input and the dc bus voltage U
dc
as output can be written as
1
1
1
1
) (
) (
*
+
⋅
+
⋅ − =
Vf speed
dc
s s C
J
s
s U
τ τ ω
, (2.15)
where τ
speed
is the equivalent time constant of the speed control loop and τ
Vf
the time
constant of the voltage measurement including all other smaller time constants.
In fact, the loop to be controlled covers a dominant time constant and a smaller time
constant. Using a PI controller, the dominant time constant can be equalized. The
cutoff frequency of the control loop is calculated by setting the time constant of the
PI controller equal to the largest open loop time constant and choosing a phase
margin guaranteeing a stable system:
speed u
τ τ = (2.16)
( )
2
) ( arctan
π
ω τ π ω ϕ − − =
c Vf c R
(2.17)
Vf R c
τ ϕ
π
ω / )
2
tan( − = ⇒ (2.18)
The gain of the PI controller K
pu
is determined by setting the brokenloop
amplification at the cutoff frequency A(ω
c
) to zero:
( ) 0 1 log 20 log 20 ) (
!
2
=

.

\

+ −


.

\

− − =
= c Vf c
pu
u
K J
C
A
c
ω τ ω
τ
ω
ω ω
(2.19)
( ) 1
2
+ − = ⇒
Vf c c u pu
J
C
K τ ω ω τ (2.20)
Regenerative Braking and RideThrough at Power Interruptions 35
During practical investigations, the best results have been obtained using a common
PI controller for the voltage control and choosing a phase margin ϕ
R
= 85°. The
input of this inner control loop is the voltage error, calculated from the measured
and filtered dc bus voltage and a reference voltage given by the main control loop.
The PI controller used is equipped with an antiwindup system limiting the
maximum allowed speed of the drive (figure 2.3).
Kpu
ω
*
U
dc
*
U
dc
Ts /τu
z
1 ω < ω
max
Figure 2.3: PI controller with antiwindup.
The dc bus voltage controlled by the speed of the motor has significant drawbacks.
Choosing a phase margin ϕ
R
= 85°, the voltage control loop has a very low
bandwidth f
B
= 14 Hz. Decreasing the phase margin leads to involuntary speed
oscillations. By no means, the voltage can be controlled faster than the underlying
speed, if such a cascaded structure is proposed. The speed control loop has a
bandwidth f
B
≈ 26 Hz. Some approaches described in literature suffer also from such
oscillation effects [Mul 97].
Subsequently, the described MPPT is performed by varying the dc voltage
triangularly. However, applying a ramp (∆U/s
2
) as a reference voltage and using
(2.15) results in a steady state voltage error U
error
:
J
C
K
U
U
pu
u
error
τ ∆
− = (2.21)
The implemented speed based voltage control turned out to malfunction at very
quickly changing irradiance power. However, no undesired crash of the entire
system due to a completely discharged capacitor has been detected during the
practical tests. Nevertheless, the voltage error between optimum and measured
voltage amounts to 10% (~20 V) during such power transients (e.g. passing clouds),
what is absolutely not acceptable for a good working MPPT and for the claim to
pump as much water as possible. Therefore, the voltage has to be controlled in
another way as described in the next subsection.
2.4.2 Torque controlled dc bus voltage
The electromagnetic torque developed by the motor is proportional to the qaxis
current and can be controlled very fast with the equivalent time constant τ
eq,i
of the
current control loop.
36 Chapter 2
1
1
) (
) (
,
*
+
=
i eq el
el
s s T
s T
τ
(2.22)
Neglecting the load torque, the following relation between motor speed and
electromagnetic torque is valid:
s J s T
s
el
1
) (
) (
=
ω
, T
load
= 0 (2.23)
In fact, the load torque is presently handled as a system disturbance, being true
considering pumping and PV power to be equal in steady state.
Replacing the speed ω in (2.14)(2.15) by the electromagnetic torque T
el
defined in
(2.22)(2.23), results in a linearized transfer function with the reference torque T
el
*
as
input and the dc bus voltage U
dc
as output:
1
1
1
1 1 1
) (
) (
,
*
+
⋅
+
⋅ ⋅ − =
Vf i eq el
dc
s s s
C J s T
s U
τ τ
, (2.24)
The voltage can be controlled directly by the electromagnetic torque of the motor. A
PI controller equipped with an antiwindup system limiting the maximum allowed
torque/current is used to calculate the reference torque. The parameters of the PI
controller are determined by choosing the time constant τ
u
larger than the sum of the
two open loop time constants and setting the gain K
pu
in order to get a maximum
possible phase margin ϕ
R
, guaranteeing a stable system:
( )
σ
τ τ τ k k T
Vf i eq u
= + =
,
, with: k > 1 (2.25)
u
pu
C J k
K
τ
− = (2.26)
( ) )
1
( arctan ) ( arctan
k
k
c R
− = ⇒ ω ϕ (2.27)
Best results are obtained by choosing 10 < k <40, corresponding to a phase margin
of 55° < ϕ
R
< 72°.
The practically implemented torque controlled voltage loop has a bandwidth
f
B
≈ 235 Hz, being 16 times larger than the other approach. The calculation of the
bandwidth takes no current/torque limitation into account.
Regenerative Braking and RideThrough at Power Interruptions 37
Applying a ramp (∆U/s
2
) as reference voltage, results in a zero steadystate voltage
error, being obviously due to the integrating term in (2.24). This property is very
advantageously for the implementation of a MPPT.
Compared to the other approach, the dc bus voltage directly controlled by the torque
has many advantages regarding speed of response, steadystate error and robustness.
Thus, all following experiments are made based on this approach. The voltage
controller (figure 6.5) switches only in speed control mode, if the maximum speed is
reached and the PV array generates sufficient power or if the motor/pump system
pumps too much water for the storage capacity.
In spite of controlling the dc bus voltage by the electromagnetic torque, the
speed/position estimation within this PV powered water pump system is not
superfluous. Both, PMSM and induction motor, being part of a high performance
drive, require information of the field position.
The structure of the dc bus voltage control integrated into the speed control system
is presented in figure 2.4. The proposed control algorithm requires a fast torque
control scheme. The wellknown principle of field orientation [Leo 85] is employed
here. The high performance speed/torque control of the given ac motor drives are
described in chapter 7 as well as the calculation of the controller parameter. The
mechanical position sensor is replaced by an observer requiring no additional
measurements. Only measurements of motor current and dc bus voltage are
necessary.
Whenever the logic signal (‘dip logic’, figure 2.4) indicates a power interruption, the
torque reference is temporarily switched from the regular speed controller to the
voltage controller. The ‘dip logic’ is obtained using a simple (digital) relay with,
considering the given installation, a switch on point U
KB
= 340 V and a switch off
point at 360 V. The predetermined reference voltage U
dc
*
should be within these
boundaries to prevent involuntary torque transients or oscillations of the logic
signal: Ideal is the switch on point. The integrator is used for both speed and voltage
control. Of course, the integrator time constant is automatically tuned. This prevents
involuntary torque transients and saves computation time. Note the negative sign of
the voltage controller gain as well as the multiplication by the sign of the motor
speed. During power interruptions, the dc bus voltage can be maintained only when
negative electromagnetic power is generated by the motor. A positive power
decreases the dc voltage. Considering a fourquadrant drive, negative
electromagnetic power is generated by inverse signs of torque and speed:
el el
T P ω = (2.28)
The multiplication with the sign of the motor speed is dropped in the PVpowered
control system since the pump is driven only in one (positive) direction.
38 Chapter 2
Enable
voltage
control
Torque
Control
&
EKF
SVM
Inverter
i
a
u
b
*
u
c
*
AC
motor
i
b
load
u
a
*
ω
ω
*
T
el
*
U
dc
U
dc
*
U
dc
PI voltage control
Power
supply
z
1
T < Tmax
Ts /τu
Ts /τn
Kpu
Kpn
Speed
reference
PI speed control
sign(ω) ω
Dip logic
Figure 2.4: Block diagram of the dc bus voltage control integrated into the speed control system.
The entire system is controlled by a digital signal processor (DSP) realizing dc bus
voltage control, speed/torque control of the drive and startup and shut down
automatism's. All control and measurement units are supplied by the dc bus. A dc
voltage beyond given limits leads inevitably to an undesired crash of the entire
system. In particular, the supply of power to the electronic control circuits of the
inverter must continue without interruption to maintain the system in operation.
2.5 Experimental Examples at Power Interruptions
A 3 kW permanent magnet synchronous machine (PMSM) supplied by a voltage
source PWM inverter has been used to verify the proposed approach. Instead of
using a PMSM, the implemented regenerative voltage control is also suitable for an
induction motor driving the load. The dynamic performance of induction motor and
PMSM are similar, only the efficiency of the former is lower especially at partial
load. The used load machine is a dc motor drive with constant electrical excitation
coupled with a variable resistor bench. The performance of the ridethrough at
power interruptions has been tested using a regular grid as power supply and
applying an abrupt power interruption on all three phases for a short time
(approximately 2 s). Figure 2.5 presents a measurement of dc bus voltage and motor
speed without applying the proposed voltage control. The speed reference amounts
to ω
*
= 1000 rpm and no load is applied. The dc link capacitor is discharged to a
critical level within ∆t ≈ 0,1 s. The undervoltage protection switches on and the
drive is out of control. This time is much shorter with applied load torque.
Regenerative Braking and RideThrough at Power Interruptions 39
0 2 4 6 8 10 12 14
0
100
200
300
400
500
t [s]
U
d
c
[
V
]
0 2 4 6 8 10 12 14
0
200
400
600
800
1000
t [s]
n
[
r
p
m
]
Undervoltage protection
Out of control
Figure 2.5: Power interruption without voltage control (no load).
Top: Voltage of the dc link. Bottom: Motor speed.
Figure 2.6 shows the experimental result of the voltage control enabled when a short
time threephase power interruption is applied. After detecting the voltage dip, the
voltage controller has to decelerate the motor very quickly guaranteeing a balanced
input/output power ratio in the dc bus. Otherwise, the dc bus would be discharged
and the system collapses. However, the implemented regenerative braking scheme
allows the inverter to keep its dc bus voltage at the predetermined minimum level,
expanding the time in which supply voltage can be reapplied without the time
consuming dclink capacitor recharging cycle.
Initially, the drive system is in speed control mode with a reference speed
n
*
= 1000 rpm. If the capacitor is discharged to a level lower than U
KB
=340 V, a
voltage dip is detected and the system switches automatically to the voltage control
mode with a predetermined voltage reference of U
dc
*
= 340V. Choosing the reference
voltage U
dc
*
lower than the ‘dip logic’ switch on point U
KB
results in a current/torque
peak at the beginning of the voltage control mode: As can be seen in [Ter 00b], the
controller starts then with an involuntary acceleration of the drive. The system
returns to the speed control mode at a voltage level higher than U
dc
= 360 V or if the
motor speed is higher than the speed reference. The deceleration of the motor during
the power interruption is small, because no load is applied. Figure 2.7 shows the
experimental results of a comparable power interruption but with a load torque
applied to the motor. The applied load amounts to 75% of the rated torque. Due to
the load, the deceleration is much faster. However, the power needed to keep the
voltage at a minimum level is the same, as can be seen at the small negative qaxis
current during the interruption interval. In fact, this power (~20 W) generated by the
kinetic energy of the drive system is nearly constant and almost completely used to
compensate the inverter losses.
40 Chapter 2
1 0.5 0 0.5 1 1.5 2 2.5 3 3.5 4
320
340
360
380
400
420
t [s]
U
d
c
[
V
]
1 0 1 2 3 4
200
400
600
800
1000
1200
t [s]
n
[
r
p
m
]
1 0 1 2 3 4
5
0
5
10
15
20
t [s]
i
q
[
A
]
voltage dip
Figure 2.6: Ridethrough at power interruption without load torque.
Top: Voltage of the dc link. Bottom: Motor speed and qaxis current.
1 0.5 0 0.5 1 1.5 2 2.5 3 3.5 4
320
340
360
380
400
420
t [s]
U
d
c
[
V
]
1 0 1 2 3 4
200
400
600
800
1000
1200
t [s]
n
[
r
p
m
]
1 0 1 2 3 4
5
0
5
10
15
20
t [s]
i
q
[
A
]
voltage dip
Figure 2.7: Ridethrough at power interruption with load torque (75% rated torque).
Top: Voltage of the dc link. Bottom: Motor speed and qaxis current.
Without the voltage dip control, the capacitor is completely discharged, considering
the given experiment and according to
ω
ω
load
load
T
U C
t
t T U C
2
2
2
1
2
1
= ∆ ⇒
∆ ≈
(2.29)
Regenerative Braking and RideThrough at Power Interruptions 41
in ∆t = 29 ms. A critical voltage level according to (2.1) is reached after ∆t = 10 ms
requiring a timeintensive restart of the converter. With the implemented voltage
control, a kinetic buffering during a total energy drop is possible for several seconds.
The span of time depends on the drive moment of inertia and the actual speed at the
moment of the voltage dip. Figure 2.8 presents a measurement at sustained power
failure. Before the drive gets uncontrolled by the undervoltage protection, the motor
has come to a complete standstill. However, the motor stays controllable during
braking, being important especially for critical applications as multimotor drives.
1 0 1 2 3 4 5 6 7 8 9
0
100
200
300
400
t [s]
U
d
c
[
V
]
1 0 1 2 3 4 5 6 7 8 9
0
200
400
600
800
1000
1200
t [s]
n
[
r
p
m
]
Undervoltage protection
Figure 2.8: Ridethrough at everlasting power failure (no load).
Top: Voltage of the dc link. Bottom: Motor speed.
2.6 Special Drive Deceleration Tool
The proposed ridethrough scheme at power interruptions can be easily transformed
into a special drivebraking tool. The upper voltage limit of the inverter may be
reached at fast braking of the drive. Regularly, the surplus generated kinetic energy
is handled using a brakingresistance within the dc link, a common dc bus or a two
way PWM inverter (figure 2.9). Nevertheless, the voltage control can be adopted
allowing a controlled braking with a maximum predetermined dc link voltage and
without redirecting the kinetic energy. The kinetic energy is usefully conducted to
the load mainly responsible for the braking during the voltage control mode. Thus,
this deceleration tool can be used, when saving energy is preferable to a fast braking
with power dissipation in, e.g., a brakeresistance. If the dynamic performance is not
crucial, the installation of a brakeresistance, power switch and cooler may be
eliminated. Especially in small motor drives, the economic gain is considerable.
42 Chapter 2
Power
supply
Common dc bus PWM converter
Motor 1
Inverter 1
Load 1
Motor 2
Inverter 2
Load 2
Braking
resistance
S1
Figure 2.9: Active front end, brakingresistance and common dc bus.
With reference to the given drive, the system switches automatically to voltage
control mode at a preset level higher than U
dc
> 600 V. The block diagram of the dc
bus voltage control integrated into the speed control system is equal to the earlier
described structure (figure 2.4). Once in voltage control mode, the system continues
to operate with the predetermined
voltage reference U
dc
*
= 600V. Finally,
the system turns back to speed control
mode when the motor speed reaches
the reference speed or at a voltage level
lower than U
dc
= 590 V. The ‘switching
logic’ is obtained using a simple
(digital) relay linked to the required
speed information (figure 2.10).
 ω
*
 ≤  ω 
U
dc
Switching
logic
Udc > 600V ⇒ 1
Udc < 590V ⇒ 0
ω
*
ω
Figure 2.10: Switching logic of the braking tool.
The design constraint of the absolute value of reference speed being lower than the
absolute value of real speed is very important. Otherwise, the motor would, once in
voltage control mode, brake to zero and wait until the capacitor is discharged by the
inverter losses to the level of U
dc
= 590 V. Thereafter, the motor accelerates to the
reference speed.
Figure 2.11 shows the experimental result of the implemented drive deceleration
tool using the 3 kW PMSM. Initially, the drive system is in speed control mode with
a reference speed of n
*
= 1500 rpm. At t = 0 s, the speed reference changes to
n
*
= 500 rpm and the switch on point (U
dc
= 600V) of the voltage control is reached
60 ms later. The implemented control scheme enables the inverter to keep its dc bus
voltage at the predetermined level. The braking of the drive is mainly caused by the
load torque. Reaching the reference speed, the system returns to the speed control
mode.
Using a brakeresistance within the dc link, the kinetic energy ∆W
kin
(
2
2
2
1
2
1
ω ω − = ∆ J W
kin
) (2.30)
Regenerative Braking and RideThrough at Power Interruptions 43
is almost completely dissipated in the resistance. With the proposed deceleration
tool, this energy is directed to the load. Considering the load torque as useful, energy
is saved.
0.5 0 0.5 1 1.5 2 2.5
400
500
600
t [s]
U
d
c
[
V
]
0.5 0 0.5 1 1.5 2 2.5
500
1000
1500
t [s]
n
[
r
p
m
]
0.5 0 0.5 1 1.5 2 2.5
20
10
0
10
20
t [s]
i
q
[
A
]
Speed
control
Speed
control
voltage control
Figure 2.11: Voltage dip with load torque (speed dependent load). Top: Motor speed.
Middle: Voltage of the dc link. Bottom: Electromagnetic torque producing qaxis current.
2.7 Conclusions
Voltage dips and sags of short duration constitute a serious problem for electrical
drives in industry. Initiated by their undervoltage protection, in general, voltage
source PWM inverter drives shut down even at short interruptions of the power
supply. The resulting shut down of critical applications as a production line may
entail loss or damage of material. Especially multimotor drives lose mutual
synchronization. Usually, time and additional workload to get a plant ready for
restart is then required. However, this and the resulting economic losses can be
avoided by using the proposed ridethrough scheme.
Here, the time interval of the power interruption is bridged by kinetic buffering. A
fast reversal of the machine operation from motor to generator mode is commanded
at the event of a power failure. Energy is fed back from the rotating masses to the dc
44 Chapter 2
link circuit to maintain the dc link voltage at a predetermined level. This is possible
also in the presence of additional loads connected to the dc link. Due to advances in
semiconductor technology, modern electric drives can withstand the high peak
currents occurring when the power supply is restored after a short disturbance.
Keeping the capacitor well charged has the additional advantage of the control
electronics being powered over a longer time span, avoiding a timeconsuming
restart of the drive.
In voltage control mode, the dc bus voltage is directly controlled by the
electromagnetic torque of the motor. The proposed voltage control scheme was
primarily developed for the PVpowered water pump system. The drive continues
operating even after a quite long power interruption of several seconds. The
temporary speed dip is generally tolerable, since the most frequent power
interruptions last only for a few milliseconds. Since drive control is never lost, the
voltage control scheme can be applied to multimotor drives as well.
Finally, the proposed ridethrough scheme at power interruptions is transformed into
a special drive deceleration tool for saving energy and simplifying the inverter setup.
Measured results are presented and evaluated to demonstrate the performance and
the system stability. Powerful digital signal processing is used to implement the
proposed regenerative braking schemes, expanding the time in which supply voltage
can be reapplied without the timeconsuming dclink capacitor recharging cycle.
3. DSPbased Drive Control and
Measurements
3.1 Introduction
Classically, motor control was designed with analog components as they are easy to
design and can be implemented with relatively inexpensive components.
Nevertheless, there are several drawbacks with analog systems including aging,
temperature drift and reliability due to EMC problems. Regular adjustment is
required in those cases. Furthermore, any upgrade is difficult, as the design is
hardwired. Digital systems, on the other hand, offer improvement over analog
circuits. The mentioned drawbacks as drift and external influences are eliminated
since most functions are performed digitally. DSP technology allows both, a high
level of performance and cost reduction. Upgrades can easily be made in software.
DSP’s have the capabilities to concurrently control a system and simultaneously
monitor it. A dynamic control algorithm adapts itself in real time to variations in
system behavior. Furthermore, implementation of complex control approaches is
possible and the drive system reliability can be improved.
For development purpose, a commercially available DSP based environment is used.
The heart of the controller board is a TMS320C31 digital signal processor. A slave
processor is employed to perform the digital input and output and generate the PWM
signals. The controller board can be directly programmed using
MATLAB
©
/SIMULINK
©
. A standard VSPWM inverter with IGBTs is adapted to be
commanded by the DSP controller board.
This chapter presents the mutual interactions between control design and realtime
implementation. The DSP controller board, code generation, experiment
management and hardware interface including required measurements are explained.
Issues of phase voltage distortion/identification due to the inverter nonlinearity are
discussed in detail. Finally, the used inverter and different PWM generation schemes
are evaluated. Optimizing (slimming down) a working control algorithm regarding
required computational effort, code optimization and implementation on a more
inexpensive hardware for the final product is something, to be considered in the final
stage of the development process.
46 Chapter 1
3.2 Controller board, Programming and Experiment
Management
The motor control is implemented using a DSP based controller board with
additional I/O features and an encoder interface. The DS1102 singleboard system
from dSpace™ (Germany) employs a TMS320C31 digital signal processor
operating at 60 MHz for the main program and a slave subsystem with a
TMSM320P14 fixedpoint DSP for the I/O subsystems and PWM generation. In
addition, the used development platform contains a comprehensive selection of I/O
interfaces that meet typical requirements for rapid motor control prototyping:
• 4 analogtodigital converters
• 4 digitaltoanalog converters
• 16 bitselectable digital I/O lines
• PWM generation on up to 6 channels
• 2 incremental encoder interfaces
A connector panel provides easy access to all input and output signals: Analog
signals via BNC connectors, all digital signals via SubD connectors. The single
board hardware (appendix A) is integrated on a standard 16bit PC/AT card slotted
straight into a PC using the ISA bus as a backplane.
The main DSP of the controller board can be directly programmed using
MATLAB
©
/SIMULINK
©
by The MathWorks™. This application software is defacto
standard in the control community, so no further explanation is given. SIMULINK
©
is
a graphical user interface integrated in MATLAB
®
for modeling and constructing
block diagrams via drag & drop operations. Its large block library is enhanced by
specific dSpaceblocks and own userdefined libraries simplifying automatic code
generation and experiment setup including initialization of the I/O subsystems and
PWM generation. RealTime Workshop is the code generation extension provided
by The MathWorks™. It generates Ccode automatically from block diagrams and
stateflow systems. For flexibility, the user can introduce own Ccode into the block
diagram by computationtime extensive Sfunctions or alternatively by special user
codes implying a change of the support software. The own Ccode should be
preferred, whenever a part of the control algorithm contains many ifloops (e.g.
space vector modulation) or in very extensive programs, e.g. sensorless speed
control with Kalman filtering. Addressing the TI compiler and automatic download
to the DSP is done via the RealTime Interface (RTI). For more information on
programming and implementation software, it is referred to the appropriate manuals.
The Total Development Environment (dSpace’ TDE) bundles a set of tools
supporting seamless transition from theory to simulation of new control algorithms
to realtime implementations. Software tools, such as CONTROLDESK, allow
parameter tuning/changing and data recording during the experiment in realtime
mode. CONTROLDESK is the comprehensive experimental environment software
providing management, control and automation of experiments. This user interface
VoltageSource PWM Inverter 47
enables access to every variable of the original block diagram. Figure 3.1 gives an
idea on how the experiment management looks like. Controller parameters can be
changed online (e.g.: speed reference) while the response is observed/recorded
simultaneously.
Figure 3.1: Screen plot during ac motor control experiment with the DS1102.
Usually, no code for the TMSM320P14 fixedpoint DSP is generated, but the
appropriate I/O functions are automatically included by the slaveDSP’s EPROM.
However, the support software has been changed in order to implement different
PWM strategies as well as variable PWM frequencies. This is extremely valuable
during the development of highperformance motor control using PWM outputs in
order to drive power switches. The modification of the support software has been
made in assembler code, since no Ccompiler for the slaveDSP exists. Compared to
the original threephase PWM generation performed in 73 us, the computation time
has been significantly reduced to 17 us. This new code is automatically included at
every compilation of the main program. The implemented modifications are
summarized in a separate manual.
3.3 Controller Interface
The laboratory test drive consists of a host PC for the controller board, an IGBT
inverter, and ac motor drive with variable load, current/voltage sensors and an
incremental encoder. Figure 3.2 shows the control setup with the DS1102 controller
board. Photos of the experimental setup are presented in appendix A.
48 Chapter 1
High performance motor control requires accurate information on motor currents
and dc bus voltage. Here, the motor currents are measured in two phases using LEM
sensors. The dc bus voltage is measured via a galvanically isolated potentiometer.
These signals are fed to the interface connected to the inputs of the A/D converters.
Due to involuntary parasitic disturbances (EMCproblems), the measured signals
should be filtered in either an analog or digital way. In general, digital filtering is
preferred. Phase shifts introduced by filtering can be corrected (if necessary) by the
transformation angle from the stator to the rotor reference frame. Only lowweighted
analog first order filters with a cutoff frequency 5 kHz are added between
voltage/current measurements and the A/D converters of the controller board. The
rotor position is measured using an incremental encoder and directly fed to the
encoder interface of the controller board.
Interface
PWM
Inverter
i
a
u
b
*
u
c
*
AC
motor
i
b
u
a
*
U
dc
Power
supply
MPP
Tracking
Current [A]
Voltage [V]
control
prototyping
Load
EXOR
EXOR
EXOR
PWM 2
PWM 1
PWM 3
PWM 4
PWM 5
PWM 6
enable
P14
DS1102Processor Board
C31
Incremental
encoder signal
Θ
Figure 3.2: Control setup with DS1102.
The PWM generation scheme implemented in the slave processor is based on phase
voltage reference values. PWM generation on up to 6 channels is possible. Both
subharmonic PWM generation and space vector modulation have been implemented.
The inverter used is a modified standard VSPWM inverter with IGBTs. An
interface provides a galvanic isolation between controller board and inverter. The
PWM switching signals are fed directly from the slave processor to the inverter
using a high performance optical link, allowing to keep both inverter and drive
several meters from the PC with the controller board. Therefore, the signal
transmission is unaffected by EMCproblems. An enable signal, using one of the
digital I/O lines together with the same high performance optical link, supervises
both entire control system and inverter.
The TMSM320P14 slave DSP generates duty cycles with 40 ns edge resolution and
160 ns PWM period resolution. In this high precision mode, the P14 always sets the
output to a high level at the beginning of each PWM period, resulting in
VoltageSource PWM Inverter 49
asymmetrical PWM pulses. The pulses of an asymmetric edgealigned PWM signal
always have the same side aligned with one end of each PWM period. On the
contrary, the pulses of a symmetrical PWM signal are always symmetric with
respect to the center of each PWM period. The symmetrical PWM is often preferred,
since it generates less current and voltage harmonics [Bose 97], [Dub 89].
In order to overcome the problem of asymmetrical PWM generated by the P14, each
of the two PWM channels are employed to generate the pulses for one phase as
shown in figure 3.2 and 3.3. By means of an EXOR gate, pulses symmetrical to the
center of the PWM period can be achieved if the switching times of each two
channels, depending on the required duty cycle, are calculated according to the
example reflecting the calculation for the first motor phase:
2
1
a phase
cycle duty
1 PWM
−
= (3.1)
2
1
a phase
cycle duty
2 PWM
+
= (3.2)
TPWM
PWM 1
PWM 2
ua
*
TPWM
E
X
O
R
Figure 3.3: Principle of symmetrical PWM generation with DS1102.
The presented algorithm is incorporated into the ‘usercode’ of the realtime
program. The ‘usercode’ offers the inclusion of handwritten Ccode into the
initialization part and the timerdriven task running with the base sample time and is
preferable compared to the use of Simulink Ccoded Sfunction since it saves
computation time. Every Sfunction block used in a Simulink model introduces an
execution time overhead of about 9 us in the realtime program due to the associated
function calls. Considering the given development platform (DS1102 60MHz), the
computation requirement of the implemented SVM and the data transmission to the
slave DSP amounts to 22 us.
50 Chapter 1
3.4 Measurements
Phase current and dc bus voltage measurements, as described in the following
subsections, are required for most highperformance motion control systems. The
measurement of the motor speed/position may be eliminated by estimation
techniques. The rotor position is measured here for control purpose or for
comparison with sensorless drive schemes. An incremental encoder with 1024 lines
is used. This signal is directly fed to the encoder interface of the controller board.
More details on position measurement and the transformation to a speed signal are
given in chapter 4.
3.4.1 Phase Current Measurement
Accurate measurement of the phase current is a key element in obtaining optimum
highperformance motor control. Measurement accuracy and bandwidth influence
directly the current control loop as well as all overlaid loops. Filtering a feedback
signal additionally decreases the dynamic response time of the loop [Leo 85].
Current is typically measured by one of two methods: voltage drop across a resistor
or magnetic transducer. Resistive shunt sensing has the advantage of a relatively
lowcost sensor. A drawback is the tradeoff between sensitivity and power
dissipated in the resistor. Since the actual motor current is the desired value, the
sensing resistor is usually placed in series with the motor phase. This complicates
the measurement, because the signal of interest is a millivolt differential value
across the resistor, but the commonmode voltage of the motor phase is typically
hundreds of volts switching at high frequency with rapid du/dt.
Magnet sensors, on the other hand, are isolated by their very nature. This means that
the motor current can directly be measured without the commonmode voltage
problems mentioned before. They use a ringtype magnetic core with a Halleffect
semiconductor element placed in an air gap to measure the magnetic flux resulting
from the primary current i
p
through the center of the core (figure 3.4). In “closed
loop” Hall effect current sensors, a canceling coil of e.g. 1000 turns is wound around
the magnetic core. A builtin feedback amplifier drives current through the canceling
coil in such a way that the flux, measured by the Halleffect sensor, is always forced
to be zero. Therefore, dc current can be measured. The output of the current
transducer is the canceling current, equal to the measured current scaleddown by
the ratio of coil turns. The overall bandwidth, accuracy and temperature
independence of these transducers has proven to be sufficient for motor drive
applications.
VoltageSource PWM Inverter 51
Figure 3.4: Principle of current measurement via “closedloop” Hall effect current sensor.
In this work, the motor currents are measured by LEMmodules. The bandwidth of
the used magnet sensor devices is 150 kHz and the response time is smaller than
1 us. The secondary (canceling) current i
s
is transformed into a voltage signal u
i
by
measuring the voltage drop across the sensing resistor R
M
(figure 3.4). This signal is
fed to an antialiasing filter connected to the inputs of the A/D converters.
Subsequently, the measured signals may be filtered by a digital lowpass filter.
However, rather than additionally filtering the current signals, observerbased
techniques can be used in order to reduce phase lags.
3.4.2 Measurement of DC Bus Voltage
In most highperformance motor control applications, the measurement of the dc bus
voltage is required for the exact transformation of the reference voltages into the
duty cycles for the inverter PWM. Even when the inverter is supplied by a constant
voltage (regular grid), the dc voltage varies due to load variations. In some
applications described later (e.g. power interruptions, braking schemes, PV
systems), the dc bus voltage is the main control variable. Furthermore, knowledge of
the dc bus voltage makes a more complicated measurement of the phase voltages
superfluous.
The measurement of the dc bus voltage is not as crucial as the current measurement
since the dc voltage is filtered and smoothed by a capacitor of appreciable size
present in the dc bus. Usually, one side of the dc bus is grounded eliminating the
commonmode problem already described at the current measurement. In the
applications mentioned, the dc voltage is measured via a resistive potentiometer, a
highperformance galvanic isolation and a firstorder analog filter (cutoff frequency
5 kHz) connected to one A/Dconverter of the controller board.
3.5 Phase Voltages
The fieldoriented control of ac motor drives, e.g. induction motor and PMSM,
demands the measurement of the motor current in two phases and the knowledge of
the dc link voltage. This makes a more complicated measurement of the phase
voltages superfluous. However, in some applications, such as sensorless field
52 Chapter 1
oriented control and exact flux estimation, the inverter output voltages are required
to calculate desired state values. The output voltage can be measured or, by using
the information of the dc link voltage, estimated by means of the reference voltages.
However, the inverter output voltages are much distorted when compared to the
reference voltages and the use of the estimation is therefore not obvious.
3.5.1 Phase Voltage Measurement
A phase voltage measurement is difficult since the inverter output voltages are
composed of discrete highvoltage/highfrequency pulses. Therefore, a potentialfree
measurement is required. A possible measurement setup and affiliated problems are
described in [Maes 01]. Beyond overmodulation, the frequency spectrum of the
output voltages generated by SVM consists of a fundamental frequency and many
higher harmonics around the PWM frequency. Only the fundamental voltage wave
contains useful information for the digital motion control. Thus, all highfrequency
components should be eliminated by a lowpass filter. Due to the lowpass filter, the
measured voltages suffer from phase delay and are not adequate for use in control
purposes [Choi 96].
Particularly at lowspeed and lightload operation, where the undesired phase delay
is negligible, problems due to the accuracy of measurement may arise: The
fundamental phase voltage is very small in these operating points and only a fraction
of the measured pulses with a magnitude equal to the dc link voltage. Nowadays, a
measurement of the phase voltages is seldom used. This is mainly caused by the
complexity and extra costs of the additional measuring devices.
The development platform used here enables a simultaneous measurement of only
four signals, limited by the number of available analogtodigital converters.
However, three A/D converters are already reserved for the measurement of dc bus
voltage and motor current in two phases. Thus, considering the given control setup,
the motor voltages must be calculated considering the inverter’s nonlinearity.
In a voltagesource PWM inverter several causes distorting the output voltages can
be found. The reasons for this originate from the inherent characteristics of the
power switches such as voltage drop, voltage transition slope, turn on/off time and
delay of the control signals. However, this delay distortion is small when compared
to the deadtime effect [Bose 97] and is therefore usually disregarded.
3.5.2 Phase Voltage Estimation
For some subsequent described applications, such as sensorless speed control and
flux estimation, the exact inverter output voltages are required to calculate desired
state values. However, they are not measured due to the lack of sufficient analogto
digital converters, but calculated by means of the reference voltages with
VoltageSource PWM Inverter 53
consideration of the inverter nonlinearity and the homopolar component generated
by the SVM. A compensation of the deadtime effect is not implemented since the
actual storage delay, varying depending on the operating point, is not exactly
known. Practical investigations have shown even a deterioration of the observer
performance by using an inadequate compensation approach: If the compensation is
not perfect, a duplication of the deadtime effect at zero crossings of the current may
occur.
Due to the delayed reaction of almost all semiconductor switches at turnon and
turnoff, the phase voltages strongly deviate from the reference voltages. The
voltage distortion does not depend on the magnitude of the reference voltages and
hence its relative influence is very strong in the lower speed range where the
reference voltage is small. Actually, the deadtime error is one of the major reasons
limiting the performance of sensorless control in low speed operation [Choi 94],
[Lee 96]. Disregarding this distortion yields in the subsequently described
speed/flux observer to large position and speed errors, especially at low motor
speed, where the error voltage becomes a multiple of the reference voltage.
Therefore, the deadtime effect is considered at the estimation of the phase voltages.
Figure 3.5 presents a comparison of the error voltage calculated by (1.17) and the
measured falsification of the fundamental voltages.
10 8 6 4 2 0 2 4 6 8 10
15
10
5
0
5
10
15
I
1
[A]
U
r
e
f

U
1
0
[
V
]
Measurement
Equation (3.11)
Figure 3.5: Deadtime effect: Measured and estimated error voltage
(U
dc
=400 V, f
PWM
= 10 kHz, τ
d
= ± 2,5 us).
The speedcontrolled ac motor is supplied by a voltagesource PWM inverter. The
PWM generation is performed by space vector modulation. SVM provides a more
efficient use of the supply voltage in comparison with sinusoidal modulation
methods by imposing a homopolar system u
0
in all three phases (multiple of third
harmonics).
(
c b a
u u u u + + =
3
1
0
) (3.3)
54 Chapter 1
However, this homopolar system reflected in the lineto neutral voltages, must be
considered in the Park transformation:
−
− −
=
c
b
a
u
u
u
u
u
2
3
2
3
0
2
1
2
1
1
3
2
β
α
(3.4)
In the case of an ideal inverter, the fundamental voltages at the motor terminals
assume the shape of the reference voltage. The reference voltages U
ref
are equal to
the duty ratios x
ref
= (a
*
; b
*
; c
*
) calculated by the SVM (figure 1.14) multiplied by
the half dc bus voltage:
ref dc ref
x U U
2
1
= , a
*
 ≤ 1; b
*
 ≤ 1; c
*
 ≤ 1 (3.5)
All together, using these reference voltages, the required alpha/beta voltages u
α
, u
β
are calculated according to the block diagram in figure 3.6 considering both the non
linearity of the inverter and the homopolar system of the SVM.
SVM
uα
Eq.(3.10)
i1
i2
sign
i3
Udc
1/2
Equation
(3.13)
uβ
xref = (a
*
; b
*
; c
*
)
Figure 3.6: Block diagram of voltage estimation.
Note that the calculation of the alpha/beta voltages described above is only valid
without strong overmodulation. The voltage spectrum in normal operation consists
approximately of one fundamental and many higher harmonics around the PWM
frequency. Overmodulation yields a voltage spectrum consisting of all uneven
harmonics. In fact, a current controller with a special antiwindup system (see
subsection 2.6.3) has been implemented, avoiding these operating points as well as
the low harmonics.
3.6 Safety Issues & Enable Subsystem
A computeraided control system is used as a development platform monitoring and
recording the experimental data. Furthermore, the safetyrelated monitoring and the
startup and shut down automatisms are implemented in software on the main DSP
VoltageSource PWM Inverter 55
board. The implemented safetyrelated monitoring (figure 3.7) consists of detecting
overcurrent and overspeed, both depending on the drive system, and a pre
determined voltage window. The minimum and maximum admissible dc bus voltage
mainly depends on the inverter used. An inadmissible failure disables the entire
system, requiring a manual reset. The reset signal is activated only by the rising edge
of a manual reset protecting the drive/inverter from an everlasting reset while an
error may be still active. In addition, fuses are integrated in the motor current circuit
as well as in the dc bus voltage measurement.
error
logic
Udc
Iα,β
enable
signal
n
enable
z
1
reset u > 0
z
1
1
error ≡ 0
no error ≡ 1
0 ⇔ 1
0 ⇔ 1
Figure 3.7: Safetyrelated monitoring & enable logic of the drive system.
The enable signal controls both the entire control system and the inverter
(figure 3.2). All gating pulses of the power switches are set to zero in case of an
error. However, special care has to be paid when a PMSM with high motor speed is
operated in flux weakening mode. A disabled inverter causes the return of the
unrestrained permanent magnet flux linkage and the dc bus voltage may reach
unacceptable (dangerous) high values, if no additional power dissipation is
connected in the dc link.
The offset of the current measurement is seldom exact equal to zero, which causes a
summation by the integrators of the controller even when the drive is disabled.
Therefore, all integrator values within the control scheme are multiplied by the
enable signal. This feature resets all registers at a restart and prevents an unwanted
overflow of integrator registers.
3.7 Conclusions
All subsequently described motor control algorithms are implemented using the
DSPbased development platform DS1102 from dSpace™. In this chapter, an
overview of the given controller board and the hardware interface between DSP and
drive system has been presented. The main DSP of the controller board can be
directly programmed using MATLAB
©
/SIMULINK
©
. For flexibility, the user can
introduce own Ccode into the block diagrams. Software tools allow parameter
tuning/changing and data recording during the experiment in realtime mode. The
laboratory test setup consists of a host PC for the controller board, an IGBT
inverter, and ac motor drives with variable load, current/voltage sensors and an
56 Chapter 1
incremental encoder. A hardware interface providing symmetric PWM signals and
transforming required measurements has been added to the experimental setup.
Support software has been changed in order to implement different PWM strategies
as well as variable PWM frequencies on the TMSM320P14 slaveDSP. Different
PWM generation schemes are evaluated. A standard VSPWM inverter with IGBTs
is adapted to be commanded by the DSP controller board. The PWM switching
signals are fed directly from the slave processor to the inverter using a high
performance optical link. Furthermore, the inverter is supervised by an enable
subsystem.
Issues of phase voltage distortion/identification due to the inverter nonlinearity are
discussed in detail. The required inverter output voltages are not measured but
calculated by means of the reference voltages with consideration of the inverter non
linearity and the homopolar component generated by the SVM.
4. Sensorless Speed Control of
Induction Motor Drives
4.1 Introduction
Induction motors are relatively cheap and rugged machines. Much attention has been
given to induction motor control for starting, braking, speed reversal, speed change,
etc. When the drive requirements include fast dynamic response and accurate speed
or torque control, it is necessary to operate the motor in a closed loop mode with
feedback of the motor speed. Only a closed loop control of the motor meets the
requirements including fast dynamic response, accurate speed and torque control or
even a higher efficiency by means of flux optimization.
However, the speed sensor has several disadvantages from the viewpoint of drive
cost, reliability and signal noise immunity. Therefore, it is necessary to achieve
precise motor control without using position or speed sensors. This chapter deals
with the speed control of induction motor drives without a shaft sensor. The field
oriented control (FOC) technique is used, together with an estimation of the motor
speed. Both rotor field magnitude and position are estimated by summation of rotor
speed and slip frequency. The structure of the implemented sensorless control is
based on the Extended Kalman Filter theory (EKF).
There are many models of sensorless speed controllers described in literature
dealing with the Extended Kalman Filter theory. They are mostly based on the
models of [Bru 90] or [Vas 94]. Brunsbach estimates four states in a rotorfixed
reference frame. The model of Vas, using the motor equations in a statorfixed
reference frame, has shown a more stable behavior, but its disadvantage is its higher
order (5 states are observed). This is a drawback when the EKF algorithm has to be
implemented in realtime. However, the model is much simpler than the first one,
since it does not contain conversions between the stator and field coordinate system,
resulting in comparable execution times for both. This approach has become
commonplace. However, this model also causes some problems, especially at low
motor speed and speed reversals. The estimated states are timedependent resulting
in an error driven nature of the observer even at steady state. Furthermore, the
estimated speed is lagging the real speed during transients, because the speed is
assumed to be constant during the sampling period.
58 Chapter 5
Here, a new model for speed estimation is proposed. This approach is shown to offer
a significant improvement of the drive performance. Along with the speed, also rotor
flux, flux position and acceleration of the drive are estimated. The speed estimation
does not lag the actual motor speed, both in steady state and during periods of
acceleration or braking.
The discussion starts by selecting a suitable motor model. Two marginal different
models are given; their advantages and drawbacks are briefly discussed. Then, the
design and implementation of the observer are explained in detail. A 1,5 kW
induction motor experimental system has been built to verify this approach. Results
are presented to demonstrate the performance of the system. The discussion ends by
evaluating the influence of motor parameter variations and designing a parameter
adaptation scheme in realtime to track these variations.
4.2 Model of the Induction Motor in Discrete Time
As mentioned above, a motor model is required for the implementation of speed
estimation via the Kalman filter approach. Choosing a stator flux reference frame
causes timedependent states resulting in an error driven nature of the observer even
at steady state. Signal lags are inevitably increased. Significant problems arise
especially due to the zero crossing of the states at low motor speed and speed
reversal. Here, the system model of the induction motor used is based on the motor
equations in a rotor flux reference frame [Bla 72], [Hen 92]. The angle of the
transformation from the stator to the rotor reference frame coincides with the rotor
flux angle γ rotating at synchronous speed ω
u
. Thus, the rotor flux lies entirely in the
daxis. At steady state, all values, apart from the flux angle, are constant. The
electrical properties of the induction motor in continuous time are completely
described by two voltage equations of the stator, two rotor equations and a torque
equation:
dt
di
i
R
u
i
dt
di
q
s
d
d
d
u
u
τ σ ω στ στ
1 1 1
) 1 ( − − + = + (4.1)
u u u
ω τ σ ω στ στ i i
R
u
i
dt
di
d
s
q
q
q
1 1 1
) 1 ( − − − = + (4.2)
d
i i
dt
di
= +
u
u
τ
2
(4.3)
u
u
τ
ω ω
i
i
q
r
2
+ = (4.4)
q
r
h
q rd
r
h
el
i i
L
L
p i
L
L
p T
u
ψ
2
1 1
= = (4.5)
Sensorless Speed Control of Induction Motor Drives 59
The torque equation (4.5) clearly shows the required torque control property of
providing a torque proportional to the torque command current i
q
. The mechanical
equation of the drive is:
dt
d
p
J
dt
d
J T T
r
load el
ω ω
= = − (4.6)
The choice of input and output vector of the model has been determined by the
structure of the electrical equivalent circuit. The induction motor is supplied by a
voltage source PWM inverter. The voltages are not necessary measured, but can be
calculated by means of the reference voltages. The current has to be measured for
the implementation of the fieldoriented control.
According to (4.4), the flux speed ω
u
can be written as a function of the electrical
rotor speed, qaxis and magnetizing current. This property is neglected in many
speed observers assuming the speed of the rotor flux to be constant during the small
sample time interval T
s
[Bru 91]. [Lut 93] uses this approximation even for the
discrete state space control of the induction motor. However, such an approximation
can be the origin of a poor estimation during transients. In fact, the speed of the rotor
flux, illustrated in figure 4.1, changes directly with and as fast as the qaxis current,
i.e. the electromagnetic torque.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
50
100
150
200
250
t [s]
ω
[
r
a
d
/
s
]
ω
r
ω
u
ω
slip
Load
step
Figure 4.1: Flux, slip and rotor speed during transients.
(simulation of a 0,8 kW induction motor drive)
Furthermore, the derivative of the magnetizing current is often disregarded [Bru 91].
Neglecting a change of the magnetizing current in (4.1) may be an acceptable
approximation of the d/qaxis current equations, but yields no significant advantage
with regard to the computing effort. Thus, the substitutions in the model matrices
should be made by using (4.3)(4.4). This eliminates both flux speed and flux
derivative in the stator voltage equations (4.1)(4.2):
60 Chapter 5
( )
s
d
q
q
r d
s
d
d q
q
r
d d
R
u
i i
i
i
i
R
u
i i i
i
i
i
dt
di
1 2 2 2 1
1 2 2 1
) 1 ( ) 1 ( 1
) 1 (
στ στ
σ
τ
ω
στ
σ
στ
στ στ
σ
τ
ω
στ
u
u
u
u
+
−
+


.

\

+ +


.

\
 −
−
−
=
+ −
−
−


.

\

+ +
−
=
(4.7)
s
q
r d
q
r q
s
q q
r d
q
r
q q
R
u
i i
i
i
i
R
u i
i i
i
i i
dt
di
1 2 2 1
1 2 2 1
) 1 ( ) 1 ( 1
) 1 (
στ
ω
σ
σ
τ
ω
στ
σ
στ
στ τ
ω
σ
σ
τ
ω
στ
u
u
u
u
+
−
−


.

\

+ −


.

\
 −
−
−
=
+


.

\

+
−
−


.

\

+ −
−
=
(4.8)
Assuming a very small sample time T
s
, the transformation from continuous to the
discrete time state space causes a negligible error. This discretization error is usually
disregarded, but might be considered later as a part of the noise covariance matrix.
Consequently, the error is compensated by the filter feedback matrix.
s
d
s s q
q
r s d s k d
R
u
T i T i
i
i
T i T i
1 2 2 2 1
1 ,
) 1 ( ) 1 ( 1
1
στ στ
σ
τ
ω
στ
σ
στ
u
u
+
−
+


.

\

+ +


.

\



.

\
 −
+ − ≈
+
(4.9)
s
q
s r s d
q
r s q s k q
R
u
T i T i
i
i
T i T i
1 2 2 1
1 ,
) 1 ( ) 1 ( 1
1
στ
ω
σ
σ
τ
ω
στ
σ
στ
u
u
+
−
−


.

\

+ −


.

\



.

\
 −
+ − ≈
+
(4.10)
Equations (4.3)(4.4) lead directly to an expression of the magnetizing current
respectively the flux position in the discrete time domain:
u u
τ τ
i
T
i
T
i
s
d
s
k


.

\

− + ≈
+
2 2
1 ,
1 (4.11)
u
u
τ
ω γ ω γ γ
i
i
T T T
q
s r s s k
2
1
+ + = + ≈
+
(4.12)
The flux angle is limited to γ  < π avoiding an overflow of a register at rotation of
the rotor in one direction over a long time span. This nonlinearity reflects no
negative influence on the EKF.
The electrical behavior of the induction motor is completely described by these
equations in discrete time and with the rotor speed as a variable. The speed must be
estimated by the filter. Thus, a suitable state equation is required. Because usually
neither the load torque nor its time variation is known, a simplification of the
mechanical equation is necessary.
Sensorless Speed Control of Induction Motor Drives 61
4.2.1 System model without load torque estimation
In a first approach, the electrical rotor speed ω
r
is assumed to be constant in the
small time interval (sampling time T
s
). Information on drive inertia is not required.
Mechanical and electrical model are fully decoupled. Nevertheless, this model
causes some problems. As will be shown, the estimated speed is lagging the real
speed during transients.
noise model
2 2
1 ,
+ = − + ≈
+ r load s q
r
h
s r k r
T
J
p
T i i
L J
L p
T ω ω ω
u
(4.13)
The known electromagnetic torque must not be used as part of the speed calculation
in (4.13) when also the load is disregarded. This would lead to a steady state speed
error since the Kalman algorithm assumes a zero mean value of the disturbances,
which is not correct, except at noload, considering only the load torque as a
disturbance. Thus, both electromagnetic and load torque must be handled as system
disturbances while the speed is treated as a constant.
The selection of the first motor model in discrete time is completed by choosing d
and qaxis current i
d
, i
q
, rotor flux i
u
, flux position γ and the electrical rotor speed ω
r
as state variable x
k
and the fundamental voltage as input u
k
. The resulting system
model and its Kalman filter are referred in following discussions as “Model 1”:
k k k
u x x
k k
B A + =
+1
;
k
s
s
k
U
U
u


.

\

=
β
α
;
k
r
q
d
k
i
i
i
x






.

\

=
γ
u
ω
; (4.14)












.

\

−

.

\
 −
−


.

\
 −
+ −


.

\

+ −


.

\
 −


.

\

+


.

\
 −
+ −
=
1 0 0
0 1 0 0 0
0 0 1 0
0 0
1 1 1
1
0 0
1 1 1
1
2
2 2
2 1 2
2 2 2 1
s
µ
s
s s
r s s
µ
q
r s
s
µ
q
r s s
T
i
T
T T
ω
σ
σ
T
σ
σ
σ
T
i
i
ω T
σ
σ
T
i
i
ω T
σ
σ
σ
T
τ
τ τ
τ τ τ
τ τ τ τ
k
A (4.15)
• L
s
, L
r
, L
h
, Stator, rotor, main inductance
• R
s
, R
r
Stator and rotor resistance
• σ = 1L
h
2
/(L
s
L
r
) Blondel coefficient
62 Chapter 5
• τ
1
= L
s
/R
s
Stator time constant
• τ
2
= L
r
/R
r
, Rotor time constant
The input matrix B
k
describes the weighted transformation from a statorfixed to the
rotor flux reference frame.
( ) ( )
( ) ( )






.

\

−
=
0 0
0 0
0 0
cos sin
sin cos
γ γ
γ γ
σ
s
s
L
T
k
B , (4.16)
The resulting output vector y
k
consists of the estimated motor current in a stator
fixed reference frame (α/βsystem, indices:
‘s’
). To avoid double calculations, the
sin/costerms of the flux angle should be calculated only once and used in both input
and output matrix.
( ) ( )
( ) ( )
k k s
s
k
x x
I
I
y


.

\
 −
= =


.

\

=
0 0 0 cos sin
0 0 0 sin cos
ˆ
ˆ
γ γ
γ γ
β
α
k
C (4.17)
A block diagram of the discrete motor model together with the feedback matrix of
the observer is shown in figure 4.2. The motor speed as well as all other states are
considered as both, state and parameter. The model matrices B
k
and C
k
depend on
the position of the rotor flux γ, the matrix A
k
on qaxis current i
q
, rotor flux i
u
and
rotor speed ω
r
.
B
k
xk+1 ⇒ xk
A
k
+
+
EKF
C
k
+
U
α
S
U
β
S
xk+1
γk+1
γk
Measurement:
I
α
s
I
β
s
yk
+
+

∆yk
∆x
z
1
Figure 4.2: Block diagram of the discrete motor model and EKF.
4.2.2 System model with load torque estimation
The second motor model, in future referred to as “Model 2”, uses additional
information on the electromagnetic torque generated by the motor. Additionally to
Sensorless Speed Control of Induction Motor Drives 63
the given states, also the acceleration due to the load torque is estimated. The model
presented in this subsection does not assume the velocity ω
r
but the load torque T
load
to be constant in a small time interval (sampling time T
s
). This results in an
improved performance during transients of the motor speed. Iron and friction losses
of the induction motor are also part of the estimated load torque. Using the torque,
rather than the speed gives a better handle on the mechanical behavior, as in this
way acceleration is controlled, being the input to the speed variations.
The acceleration of the drive equals the difference between electromagnetic T
el
and
load torque T
load
related to the drive inertia J. The load torque is generally unknown,
but constant at steady state. It creates a disturbance of the speed control loop, which
is compensated by the controller. In steady state, the acceleration of the drive is zero
by definition. Thus, the differential equation of the acceleration due to the load
torque is:
0 ≈ 
.

\

=
load
l
T
J
p
dt
d
dt
dα
(4.18)
load l k l
T
J
p
= ≈ ⇒
+
α α
1 ,
(4.19)
Now, the known electromagnetic torque is used as part of the speed calculation,
improving the accuracy of the speed:
l s q
r
h
s r k r
T i i
L J
L p
T α ω ω
u
− + ≈
+
2 2
1 ,
(4.20)
In contrast to the remarks concerning the load torque in (4.13), the inaccuracy of
(4.18)(4.20) has a zero mean value, being a precondition of the Kalman algorithm.
Only a variation of the load is handled as model inaccuracy. This inaccuracy is
neglected here, but will be taken into account afterwards at the evaluation of the
noise covariance matrix. In addition, erroneous electromagnetic torque calculation
and inertia identification are handled as model noise. However, the influence of both
an incorrect estimation of the electromagnetic torque due to electrical parameter
variations and an incorrect identification of the drive inertia are small compared to a
potential load variations. The discrete form of the second model is:
k k k
u x x
k k
B A + =
+1
;
k
s
s
k
U
U
u


.

\

=
β
α
;
k
l
r
q
d
k
i
i
i
x








.

\

=
α
γ
ω
u
(4.21)
64 Chapter 5















.

\

−
−

.

\
 −
−


.

\
 −
+ −


.

\

+ −


.

\
 −


.

\

+


.

\
 −
+ −
=
1 0 0 0 0 0
0 1 0 0
0 1 0 0
0 0 0 1 0
0 0 0
1 1 1
1
0 0 0
1 1 1
1
2
2 2
2 2
2 1 2
2 2 2 1
s
s
s
r
h
s
s s
r s s
q
r s
s
q
r s s
T
i
T
T i
L J
L p
T
T T
T T
i
i
T
T
i
i
T T
u
u
u
u
τ
τ τ
ω
σ
σ
στ
σ
στ τ
ω
στ
σ
τ
ω
στ
σ
στ
k
A
(4.22)
( ) ( )
( ) ( )








.

\

−
=
0 0
0 0
0 0
0 0
cos sin
sin cos
γ γ
γ γ
σ
s
s
L
T
k
B (4.23)
( ) ( )
( ) ( )
k k s
s
k
x x
I
I
y


.

\
 −
= =


.

\

=
0 0 0 0 cos sin
0 0 0 0 sin cos
ˆ
ˆ
γ γ
γ γ
β
α
k
C (4.24)
This model has a disadvantage: its order is higher. This is a drawback when the EKF
algorithm has to be implemented in realtime. However, one major advantage of this
model is that it does not assume the speed to be constant during the sample time.
The involuntary lag of the speed signal is avoided by the additional estimation of the
load acceleration. In fact, the estimation of the acceleration is insignificantly lagging
at a continuous load torque variation. Nevertheless, the acceleration is, apart from an
initial change, nearly constant during both changing the speed reference and
applying load torque. This special drive property is caused by the current/torque
limitation within the speed control loop. Thus, the acceleration is almost constant
and can be estimated accurately.
The other advantage originates from the higher accuracy of the speed specification.
This accuracy is considered at the calculation of the noise covariance matrix. A
lower value indicates a more accurate estimation and accordingly results in a
smother speed signal.
Obviously, the performance of the system increases as the information of the known
electromagnetic torque is used. Only the load torque is handled as if it were an
unknown system disturbance, being true for many motor drives. If the loadspeed
relation is known, this information can be used for further improvement of the
Sensorless Speed Control of Induction Motor Drives 65
estimation performance. In that case, the equation of the load acceleration α
l
is
determined employing equations (4.5), (4.6) and (4.19):
J
T T
p
d
dT
dt
d
d
dT
dt
dT
load el
r
load r
r
load load
−
⋅ = =
ω
ω
ω
(4.25)
⇒
r
load
k l q
r
h
s k l k l
d
dT
i i
L J
L p
J
p
T
ω
α α α
u


.

\

− + ≈
+ ,
2 2
, 1 ,
(4.26)
If the known loadspeed relation is applied to the algorithm, also the system model
inaccuracy is lower. The noise covariance Q can be reduced, resulting in a very
smooth steady state speed signal and almost no lag during acceleration or braking
periods.
Both with and without applying the loadspeed relation, the performance of the
estimator is only slightly affected by a precise knowledge of the inertia. If the inertia
J is set to infinite, the behavior of the algorithm is like the one neglecting the torque
command inputs and assuming the speed to be constant in a small time interval.
Naturally, the inertia must not be set to zero to guarantee a stable functioning. In all
other cases, a mismatch of the inertia is handled by the EKF as system noise. The
steady state estimation of the load torque becomes erroneous but the speed
estimation remains correct.
4.3 Extended Kalman Filter Algorithm
The induction motor torque depends on both airgap flux and speed, but neither
torque versus flux nor torque versus speed relations are linear. This complicates the
design of control systems and speed estimation for induction machines. Due to the
lack of a system with linear equations, also the state model of the induction motor
used is nonlinear. The mechanical speed and position of the flux are considered as
both, state and parameter. The model matrices B
k
and C
k
depend on the position of
the rotor flux, the matrix A
k
on qaxis current i
q
, rotor flux i
u
and rotor speed ω
r
.
Therefore, the extended Kalman filter (EKF) has to be used to estimate the
parameters of the model matrices, as well. The EKF performs a relinearization of
the nonlinear state model for each new estimation step, as it becomes available.
Furthermore, the EKF provides a solution that directly cares for the effects of
measurement or system noise. The errors concerning the parameters of the system
model are also handled as system noise.
A more complete introduction to the general idea of the Kalman filter can be found
in literature [Bram 94], [May 79], Bro 92]. Here, only the basic equations of the
EKF are repeated. The EKF algorithm used is based on [Bram 94]. The Kalman
filter estimates a process by using a form of feedback control. The signal flow of the
EKF in a recursive manner is shown in figure 4.3.
66 Chapter 5
1/z
)
`
¹
q
d
u
u
k k
x
1 +
x ∂
Φ ∂
k k
x ∆
1 − k k
x
x
h
∂
∂
1/z
Predictor
Filter
Pkk1
Pk+1k
Pkk
Kk
Kk ∆Yk
k measured
k
y y Y − = ∆
Figure 4.3: Block diagram of the extended Kalman filter.
The Kalman algorithm distinguishes between filter and predictor equations. The
predictor equations are responsible for projecting the state to obtain the “a priori”
estimation of the next time step. The filter equations, also called measurement
update, are responsible for the feedback to obtain an improved “a posterior”
estimate. The predicted value of the state vector x
k+1k
is corrected by adding the
product of filter gain and the difference between estimated and measured output
vector y
k
to the state vector x
kk
. In addition still the equation for the corrected
covariance matrix P
kk
is required.
( ) ( ) k x h y x x
k k
k
k k k k
,
1  1   − −
− + =
k
K (4.27)
1 k  k k 1 k  k k  k
P K P P
− = −
−
∂
∂
− =
1 

k k
x x
x
h
(4.28)
The matrix K
k
is the feedback matrix of the extended Kalman filter. This matrix
determines how the state vector x
kk
is modified after the output of the model y
k
is
compared to the measured output of the system. The filter gain matrix is defined by:
1
1  1  1 
  
−
= − = = −


.

\

+
∂
∂
∂
∂
∂
∂
=
− − −
R P P K
1 k  k 1 k  k k
k k k k k k
x x
T
x x x x
T
x
h
x
h
x
h
(4.29)
in which R is based on the covariance matrix of the measurement signal noise.
Based on the calculated state vector x
kk
, a new value of the state vector can be
predicted. The same applies to the error covariance matrix. The prediction is:
( )
k k k k k
u x k k x , , , 1
1   1 − +
+ Φ = (4.30)
T
k k x x
T
x x
k k k k
x x
Γ Γ +
∂
Φ ∂
∂
Φ ∂
=
= = +
Q P P
k  k k  1 k
 
  (4.31)
Sensorless Speed Control of Induction Motor Drives 67
with the covariance matrix Q reflecting the system noise.
All equations of the EKF algorithm can be written as a function of a system vector
Φ and an output vector h describing the relinearized model of the induction motor.
The system and output vector respectively can be derived from the model equations
of the induction motor.
( ) ( ) ( )
k k k k k k k k k k k
u x x x u x k k
    1 
, , , 1
k k
B A + = + Φ
−
(4.32)
( ) ( )
1  1  1 
,
− − −
=
k k k k k k
x x k x h
k
C (4.33)
In addition, the derivatives of system and output vector are required for the EKF
algorithm. The derivative of the system vector of Model 2 results in:
68 Chapter 5
+















.

\



.

\

+
−
+ −


.

\

+ −


.

\

+


.

\
 −
+ −
=
∂
Φ ∂
0 0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0
1 1
1
0 0 0 0 2
1 1
1
2
2 2
2
2 2 1 2
2 2 1
u
u
u u
u
τ
τ
τ στ
σ
στ τ
ω
τ
ω
στ
σ
στ
i
T
i
L J
L p
T
T
i
i
T
i
i
T
i
i
T T
x
s
r
h
s
s
d
s
q
r s
q
r s s
(4.34)
















.

\

−
−
−
− 
.

\
 −
+ −


.

\

−
−


.

\

−
−
+
1 0 0 0 0 0
0 1 0 0
0 1 0 0
0 0 0 1 0 0
0
1 1
0 0
0
1
0 0
2
2
2 2
2
2
2
2
2
2
2
s
q
s
s q
r
h
s
s
d
s
s
d s r
d q
s
q
s
s
q s
q
s
T
i
i
T
T i
L J
L p
T
T
u
L
T
i i T
i
i i
T
u
L
T
i T
i
i
T
u
u
u
u
τ
τ
σ σ
σ
ω
σ
σ
τ
σ τ στ
σ
where u
q
and u
d
are voltages in a rotorflux reference frame, already calculated by
the product of α/βvoltages and input matrix B
k
. Thus, the result can be used to save
computing time. Note that the qaxis voltage influences the linearized specification
of daxis current and viceversa.
The corresponding derivative of the system vector of Model 1 is obtained by
dropping the last column as well as the last row in (4.34) and setting the elements
} 2 , 4 {
x ∂
Φ ∂
and } 3 , 4 {
x ∂
Φ ∂
to zero. In this way, the influence of electromagnetic and
load torque on the speed is canceled.
The derivative of the output vector of Model 2 is:
Sensorless Speed Control of Induction Motor Drives 69
( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( )


.

\

− −
=


.

\

−
− − −
=
∂
∂
0
ˆ
0 0 cos sin
0
ˆ
0 0 sin cos
0 sin cos 0 0 cos sin
0 cos sin 0 0 sin cos
α
β
γ γ
γ γ
γ γ γ γ
γ γ γ γ
i
i
i i
i i
x
h
q d
q d
(4.35)
The calculation of the estimated α/βcurrent is already executed by the output matrix
C of the system model and should be used in (4.35) to avoid double calculations.
The corresponding derivative for Model 1 is obtained by dropping the last column in
(4.35).
The remaining variables of the algorithm are the noise covariance matrices Q and R
and an initial matrix P
00
representing the covariance of the known initial conditions.
They consist only of diagonal elements.
4.4 RealTime Implementation of the EKF
4.4.1 Measurement & system noise
One critical step towards the implementation of the extended Kalman filter
algorithm is the search for the best covariance matrices. They have to be setup
based on the stochastic properties of the corresponding noise. The noise covariance
R accounts for the measurement noise introduced by the current sensors and the
quantization errors of the A/D converters. Increasing R reflects a stronger
disturbance of the current. The noise is weighted less by the filter, causing a more
filtered current but also a slower transient performance of the system. The noise
covariance Q describes the system model inaccuracy, the errors of the parameters
and the noise introduced by the voltage estimation. Q has to be increased at stronger
noise levels driving the system, entailing a more heavily weighting of the measured
current and a faster transient performance. Thus, changing the covariance matrices R
and Q affects both the transient duration and the steady state operation of the filter.
An initial matrix P
00
represents the matrix of the covariance in knowledge of the
initial conditions. Varying P
00
affects neither the transient performance nor the
steady state conditions of the system and can be chosen at random.
The covariance matrices R, Q and P
00
are assumed to be diagonal due to the lack of
sufficient statistical information to evaluate their offdiagonal terms. Furthermore,
the diagonal characteristic holds the possibility of saving a lot of computing time as
shown in the next subsection.
In general, the entries of the covariance matrices are unknown and cannot be
calculated. They are often set to the unity matrix. In order to achieve the optimal
70 Chapter 5
filter performance, the filter parameters R and Q can be obtained by tuning based on
experimental investigations. This describes an iterative process of searching the best
values. It is almost impossible to find a plausible evaluation of these parameters in
literature with regard to the sensorless control of motor drives.
However, it is preferable to have a rational basis for choosing the required
parameters. In either case, whether or not a superior filter performance can be
obtained by an additional tuning process, an initial guess of the values is welcome.
As shown, the value of the different parameters differs a lot. Furthermore, the filter
performance may change dramatically by varying only one value. Without any
previous knowledge and considering the high dimension of the matrices, tuning is
very arduously or can even lead to an unstable behavior of the observer. For
instance, changing the sample time requires a new tuning process. A design equation
has the additional advantage of being independent of the given installation, and it
can easily be assigned to other drive installations without an expert tuning the
parameters.
The measurement covariance R can be measured easily in advanced. Measuring is
generally possible because the current measurement is needed anyway while
operating the filter. Some offline sample measurements are taken in order to
determine the variance of the measurement error. This is done by applying a
constant linetoline voltage across two phases containing a current sensor and
measuring the resulting dc current. It must be noted, that the measured current
should not be supplementary filtered, apart from an antialiasing filter of course. The
noise on the raw measurements will possibly be nonlinearly transformed resulting
in second order terms, which may be significant. The Kalman approach handles
white and uncorrelated measurement noise and produces the minimum variance
estimate. Therefore, this is already an optimal filter. A current measurement
respecting the given installation yields a measurement noise covariance matrix,
being almost proportional to the dc bus voltage within a voltage range
300V < U
dc
< 600V:
2
4
4 2
6
6
A
10 5 , 1 0
0 10 5 , 1
V
A
10 8 , 1 0
0 10 8 , 1


.

\

⋅
⋅
−


.

\

⋅
⋅
=
−
−
−
−
dc
U R (4.36)
In case of the system covariance, the calculation is less deterministic. Nevertheless,
an estimation of the matrix elements is possible using some simplifications.
Furthermore, the given assumptions have been examined experimentally. All given
values are calculated using the parameters of the 1,5 kW induction motor drive and a
sample time T
s
= 200 us. The calculated values are valid for both system models;
Considering (4.9)(4.10), the inaccuracy of the current calculation is mainly affected
by the accuracy of the voltage identification being the input of the system.
Sensorless Speed Control of Induction Motor Drives 71
inaccuracy voltage inaccuracy model current ⋅ ≈
s
s
L
T
σ
(4.37)
The voltage can be either measured or calculated by means of the reference voltages
being the output of the entire control loop. Here, the phase voltages are calculated.
Therefore, the accuracy is only affected by the nonlinearity of the converter. This
nonlinearity has its origin in the delayed reaction of the switches at turnon and 
off, also called deadtime effect [Bose 97]. Therefore, the accuracy of the voltage
calculation, described by an error voltage ∆U, is simplified dependent on the dead
time τ
dead
, the dc bus voltage U
dc
and the PWMfrequency f
PWM
of the inverter:
dc PWM dead
U f U τ ≈ ∆ (4.38)
The influence of parameter variation is marginal compared to this deadtime effect.
So, the covariance of the current model inaccuracy can be estimated by:
2
3
1
) 1 , 1 (


.

\
 ∆
≈
s
s
L
U T
Q
σ
(4.39)
) 1 , 1 ( ) 2 , 2 ( Q Q = (4.40)
For the given drive, using a sample time T
s
= 200 us, a deadtime τ
dead
= 2 us and a
PWMfrequency f
PWM
= 10 kHz, the covariance amounts to:
2
A 0018 , 0 ) 2 , 2 ( ) 1 , 1 ( = = Q Q (4.41)
The model estimation inaccuracy of magnetizing current i
u
and flux position γ is
only caused by the discretization of the continuous equations. In contrast to
Model 1, the speed specification within Model 2 is very accurate. The inaccuracy is
much lower and mainly caused by the discretization error. Considering a very small
sample time, these errors are negligible. Nevertheless, the worst of all
approximations is to set the model inaccuracy to zero. White noise is a much better
approximation than zero. Thus, this discretization error is considered by a very small
value in the noise covariance matrix Q. The maximum discretization error of the
magnetizing current is dependent on the maximum motor current and the rotor time
constant.


.

\

−
−
−
=
∑
∫
∞
=
+
0
) 1 (
2
, ,
2
, ,
var ) 3 , 3 (
k
T k
kT
k k d k k d
s
s
s
dt
i i i i
T Q
τ τ
u u
(4.42)
72 Chapter 5
2 5
2
2
2 2
max
2
max
A 10 6 , 2
3
var ) 3 , 3 (
−
⋅ ≈ =


.

\

< ⇒
τ τ
s s
T i T i
Q (4.43)
Assuming a maximum acceleration of α
max
= 1000 s
2
and a sample time T
s
= 200 us,
the variance of the flux position is estimated by:


.

\

− <


.

\

− =
∫
∑
∫
+
∞
=
+
s
s
s
s
T k
kT
s s
k
T k
kT
k s
dt t T k T dt t T Q
) 1 (
max max
0
) 1 (
var ) ( var ) 5 , 5 ( α α ω ω (4.44)
11
4 2
max
2
max
10 3 , 3
48 2
var ) 5 , 5 (
−
⋅ ≈ =


.

\

< ⇒
s s
T T
Q
α α
(4.45)
For Model 2, the inaccuracy of the motor speed calculation is also caused by the
discretization error dependent on the maximum acceleration and its time variation.
Based on common bandwidth of torque control loops, a maximum variation time
constant τ
torque
= 1 ms is chosen.


.

\

− =
∑
∫
∞
=
+
0
) 1 (
2 Model
) ( var ) 4 , 4 (
k
T k
kT
k s
s
s
dt t T Q α α (4.46)
2 2
4 2
max
2
max
2 Model
1
0033 , 0
48 2
var ) 4 , 4 (
s
T T
Q
torque
s
torque
s
≈ =


.

\

< ⇒
τ
α
τ
α
(4.47)
These values are very small resulting in smooth signal shapes. They could be set to
the maximum in order to achieve maximum dynamic performance of the drive.
For Model 1, the inaccuracy of the motor speed calculation is higher due to the
simplified specification and can be found from the maximum inertia related torque
variation:
(
max 1 Model
2 var ) 4 , 4 ( ) α
s
T Q =< (4.48)
2
2
max
2
1 Model
s
1
013 , 0
3
) 4 , 4 ( ≈ < ⇒
α
s
T
Q (4.49)
In contrast to Model 1, the speed inaccuracy for Model 2 is transferred to the
acceleration equation. The variance of the acceleration in system Model 2 equals the
variance of the inertia related load. The load torque is generally unknown. However,
an estimation of the variance toplimit can be obtained by assuming a maximum
torqueinertia relation of the drive. The calculation is based on the considerations
made in chapter 4: A constant relation of T
max
/J = 1000 s
2
is assumed. Considering
Sensorless Speed Control of Induction Motor Drives 73
the number of pole pairs p and a maximum torque variation time constant
τ
load
= 2 ms, the toplimit of the acceleration inaccuracy and of the process variance
is:
torque
s k load k load
J
T T p
J
T T
p
τ
max
max
1 , ,
inaccuracy model ≈


.

\
 −
<
−
(4.50)
4
2
max
1
1482
12
1
) 6 , 6 (
s J
T T p
Q
torque
s
≈


.

\

< ⇒
τ
(4.51)
The dynamic and smoothness of both speed and acceleration estimation is tuned by
Q(6,6). This parameter should be smaller, if the load torque is known very well or a
smooth speed signal is more important than the loop dynamic. A high value
increases the dynamic performance, but also the noise of the estimated signals. With
respect to the given drive setup, Q(6,6) is set to 5% of the value given in (4.51) in
order to obtain a good compromise between dynamic performance and smooth
torque command response. All other coefficients of the system covariance matrix are
set to the given values.
It should be noted, that the calculation of all process covariance matrices is
proportional to the square of the sample time. Thus, the given constants should be
adapted accordingly, if a different sample time is chosen. However, it is a major
advantage of the proposed model, that estimation accuracy and stability of the entire
control system are much less sensitive to tuning the covariance matrices compared
to other models.
4.4.2 Computing requirements
The speed estimation and the entire control of the induction motor are implemented
on a TMS320C31 DSP with 128 K × 32bit RAM. The implemented algorithm
estimates five states for Model 1 and six for Model 2. The computing demand grows
almost with the third power of the state dimension. Furthermore, they contain
conversions between stator and field coordinate system and a computation time
intensive matrix inversion. The algorithm can be implemented with relatively few
instructions using matrix calculation. However, without any modification, the
resulting algorithm leads to a program that is not suitable for realtime
implementation, since it is very complex especially due to the matrix inversion. The
execution time would be higher than 400 us, respectively 700 us, using the given
DSP. In consequence, also the bandwidth of the current/torque controller would be
very small. Furthermore, the performance of the EKF decreases as the sample time
increases.
74 Chapter 5
The turnaround time of the final control system, using Model 2, amounts to 187 us.
Only a few extra calculations are necessary compared to the speed observer based on
Model 1 requiring a turnaround time of 167 us. The used sample time is set to
T
s
=220 us. However, the execution time is not that meaningful. The DSP power is
simultaneously used for monitoring and recording the experimental data. Due to
developing reasons of the installation, the remaining fieldoriented control is not
optimized regarding the computation requirement: e.g., the implemented FOC
contains three different speed controllers for performance comparison. An overview
of the computing requirement considering the different approaches is summarized in
table 4.1.
Table 4.1: Computation requirement of different EKF approaches.
Number of
Summations [ ]
Number of
Multiplications [ ]
Turnaround time of
the entire control [us]
Model 1
Matrix calc.
546 662 >400
Model 2
Matrix calc.
881 1026 >700
Model 1
optimized
207 254 167
Model 2
optimized
255 320 187
Keeping the size of the program limited is achieved by optimizing the model with
hand calculations and exploiting matrix symmetry. The covariance matrices Q, R
and initial matrix P
00
are set to be symmetrically. In consequence, also the matrix
P
kk1
becomes symmetrical which can be exploited avoiding double calculations and
higher memory demand. The implemented EKF covers no superfluous
multiplications by zero. Several matrix calculations of the EKF algorithms are the
same and can be used in different equations, e.g. in (4.28)(4.29). Furthermore, the
sine and cosine of the flux angle is calculated only once and used in the EKF as well
as in the Clarketransformations [Bose 97] of currents and voltages.
4.4.3 Model comparison
Two models are closely examined. The first one is based on an approach, that has
become commonplace in almost every speed observer. They do not recognize the
actual torque command inputs to the system and assume the velocity ω to be
constant in a small time interval. In effect, such techniques treat the known torque
command input as if it were an unknown disturbance torque. Thus, they generally
lag the actual motor speed during periods of acceleration or braking.
The second motor model uses the additional information on the electromagnetic
motor torque. Additionally, the acceleration due to the load torque is estimated. This
model does not assume the velocity but the load torque to be constant in a small time
interval (sampling time T
s
). The inaccuracy of the speed calculation is transferred to
the load, which is usually unknown anyway. This results in an improved
Sensorless Speed Control of Induction Motor Drives 75
performance during transients (figure 4.4) presenting the response to a step of the
speed reference and to a load step. Figure 4.5 shows some important details of
figure 4.4.
0.1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
500
1000
1500
t [s]
n
[
r
p
m
]
0.1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
5
0
5
10
15
t [s]
T
[
N
m
]
Figure 4.4: Step of the speed reference and response to a load step. Top: Real and estimated speed with
and without load torque estimation. Bottom: Estimation of electromagnetic and load torque.
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
0
200
400
600
t [s]
n
[
r
p
m
]
0.05 0.1 0.15 0.2 0.25 0.3
900
950
1000
1050
t [s]
n
[
r
p
m
]
1
With α
l
estimation
Without α
l
estimation
T
el
T
load
2
Figure 4.5: Details of figure 4.4. Top: Details indicated by Box 1.
Bottom: Details indicated by Box 2.
Simulations are performed to compare both algorithms using the improved
algorithm as feedback to control the motor. With the other algorithm in the control
loop, the obtained results are almost the same but with a higher overshoot due to the
delayed response of speed and current controller. A simultaneous realtime
implementation of both algorithms requires a faster DSP. Nevertheless, the
simulation has the advantage of calculating the real motor speed without any delay
in contrast to a realtime implementation using a filter for the measured speed signal
(see also figure 5.7). The estimated speed signal requires no additional filter. The
smoothness and the transient performance of the signals are adjustable by the noise
covariance matrix Q of the EKF algorithm [Ter 01].
76 Chapter 5
From now on, only the second model is considered. Model 2 offers a large
improvement of the performance. The price to be paid is only marginally extra
computing effort.
4.4.4 Observer integrated into the fieldoriented control
The proposed algorithm can be implemented in software with an arguable
requirement of computation time. Figure 4.6 shows the closed loop observer
integrated into a simplified fieldoriented speed control loop of an induction motor
drive. The voltages required as input for the EKF are either measured or obtained
from the reference voltages. Here, they are calculated regarding the inverter non
linearity as explained in chapter 3. The current is measured in two phases. As
mentioned earlier, the current should not be additionally filtered. Prefiltering
decreases the performance of the proposed observer. The current measurement
should be offsetfree as the Kalman filter assumes a zero mean value of the error. An
offset generates erroneous estimations, especially at very low motor speed.
ia
ib
iα
iβ
Voltage
calculation
3 ⇒ 2
iα
iβ
sin(γ)
cos(γ)
ωr
id
iq
EKF
ud
*
uq
*
id,error
iq,error
sin(γ)
cos(γ)
uα
*
uβ
*
SVM
current
control
inverse
Park Trans.
PWM
generation
Speed
reference
id
*
iq
*
Speed
&
Flux
Control
iu
ud
*
uq
*
∆ud
*
∆uq
*
decoupling
SVM
Inverter
ub
*
uc
*
AC
motor
ua
*
Udc
Load
uα
uβ
Udc
Power
supply
Digital motion control
2
Udc
Figure 4.6: Velocity observer integrated into a fieldoriented speed control loop.
The presented observer achieves the objective of eliminating lag of the estimated
motor speed by additional load estimation. It should be noted, that the velocity
estimation described here can easily be extended to allow for further improvement
of the entire drive performance especially at load torque variation by adding
acceleration feedback. The information of load acceleration can be used directly by
compensating for the load torque. Rejecting load disturbances improves the dynamic
stiffness of the drive. Therefore, this feedback causes the disturbance to perceive a
more robust system less sensitive to disturbances.
Sensorless Speed Control of Induction Motor Drives 77
4.5 Experimental Results
A 1,5 kW induction motor has been used to verify the applied approach. The inertia
of the whole drive system (motor and load machine) is about 0,008 kgm
2
. The used
load machine is a dc motor drive with constant excitation. At high motor speed, the
dc motor is coupled with a resistor bench. The experiments at low motor speed are
done with the dc motor supplied by a thyristor converter. The load machine can be
controlled in either torque or speed control mode.
All presented results are obtained with the second model in the loop. Figure 4.7
shows the experimental results of a speed reversal using the estimation of speed,
rotor flux and flux angle as feedback to control the motor. Additionally, the real
speed is measured and compared. It can be seen that there is a very good accordance
between real and estimated speed, without any steady state error. During transients,
the estimation of the speed is even faster thus better than the measured one, because
a filter is used for the speed measurement causing a delay of the signal (incremental
encoder with 1024 lines, cutoff frequency of the used speed filter ≈ 1 kHz).
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1000
500
0
500
1000
t [s]
n
[
r
p
m
]
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
50
25
0
25
50
t [s]
∆
n
[
r
p
m
]
∆n = n
est
 n
m
Figure 4.7: Speed reversal test. Top: Reference, measured and estimated speed.
Bottom: Difference between estimation n
est
and measurement n
m
.
The current controller, using also the estimated values of d and qaxis current, has a
bandwidth of 847 Hz. Figure 4.8 presents the response of the induction motor to a
load step at a motor speed of 1500 rpm. The applied load amounts to 65% of the
rated value.
The behavior at low motor speed is shown in figure 4.9. First, the response to a
square wave shaped speed reference is given. With a sinusoidal speed reference,
there is almost no difference between estimation, measurement and reference.
78 Chapter 5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1400
1450
1500
1550
1600
t [s]
n
[
r
p
m
]
0 0.25 0.5 0.75 1
0
2
4
6
t [s]
i
q
[
A
]
0 0.25 0.5 0.75 1
0
5
10
15
t [s]
T
l
o
a
d
[
N
m
]
n
m
n
est
Figure 4.8: Response to a load step. Top: Measured speed n
m
and
estimated speed n
est
. Bottom: qaxis current and load torque estimation.
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
50
25
0
25
50
t [s]
n
[
r
p
m
]
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
25
0
25
t [s]
n
[
r
p
m
]
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
50
25
0
25
50
t [s]
n
[
r
p
m
]
n
est
n
m
n
est
n
m
n
ref
Figure 4.9: Behavior at low motor speed. Top: Square wave speed reference (20 Hz).
Middle: Sinusoidal speed reference (20 Hz). Bottom: Sinusoidal speed reference (100 Hz).
Even at low motor speed and standstill the proposed control scheme is able to
manage the load torque (figure 4.104.11). Load is applied using a dc motor
operating in torque control mode. The arising torque ripple components are typical
of a thyristor converter and are returned to the signals of speed and load torque
estimation respectively.
Sensorless Speed Control of Induction Motor Drives 79
0 0.5 1 1.5 2 2.5 3
2
0
2
4
6
8
t [s]
I
d
c
[
A
]
0
0.5 1 1.5 2 2.5 3
60
40
20
0
20
40
60
t [s]
n
[
r
p
m
]
n
est
n
m
Figure 4.10: Response to a load step at standstill. Top: Armature current of the load machine.
Bottom: Measured speed n
m
and estimated speed n
est
.
0 0.5 1 1.5 2 2.5 3 3.5 4
2
0
2
4
6
8
t [s]
I
d
c
[
A
]
0 0.5 1 1.5 2 2.5 3 3.5 4
0
60
120
t [s]
n
[
r
p
m
]
0 0.5 1 1.5 2 2.5 3 3.5 4
0
5
10
15
t [s]
T
l
o
a
d
[
N
m
]
Figure 4.11: Response to a load step at low motor speed (n
ref
= 60 rpm). Top: Armature current of the
load machine. Middle: Measured and estimated speed. Bottom: Estimation of the load torque.
Also at high motor speed and flux weakening, a good performance of the EKF can
be obtained. Figure 4.12 demonstrates the behavior of the speed and flux estimation
in this speed range. The flux is inversely proportional to the motor speed. In
consequence, the applied fundamental motor voltage remains nearly constant. Only
a small transient is needed to adapt the required d and qaxis current.
This feature of the EKF makes the proposed system also suitable for applications
with flux optimization increasing the drive efficiency. However, a minimum flux is
required to guarantee a stable operation of the EKF. The minimum flux for the given
induction motor drive amounts to approximately 10% of the rated value.
80 Chapter 5
0 0.05 1 0.15 0.2 0.25
0
500
1000
1500
2000
2500
3000
t [s]
n
[
r
p
m
]
0.
0 0.05 0.1 0.15 0.2 0.25
1
1.5
2
2.5
3
3.5
t [s]
i
u
[
A
]
0 0.05 0.1 0.15 0.2 0.25
0
50
100
150
200
250
300
t [s]
+

[
V
]
n
m
n
est
n
ref
Flux
weakening

U
α
U
β
300 200 100 0 100 200 300
300
200
100
0
100
200
300
u
α
[V]
u
β
[
V
]
Flux
weakening
Figure 4.12: Behavior at high motor speed and flux weakening. Top: Speed response
and applied voltage amplitude. Bottom: Magnetizing current and applied voltage.
4.6 Motor Parameter Sensitivity and Adaptation
The used motor model, as well as the implemented EKF, contains four electrical
motor parameters: stator inductance L
s
, stator resistance R
s
, rotor time constant τ
2
and leakage (Blondel) coefficient σ. All other parameters, as e.g. the stator time
constant τ
1
, are linked to these parameters. Obviously, the quality of the speed
estimation in the observer depends on the accuracy with which the motor parameters
are known. Inaccurate model parameters lead to misalignment of the fieldoriented
coordinate system, impairing the dynamic performance of the drive. Possibly more
important is the steadystate accuracy of the speed control, being poor with detuned
model parameters.
If the machine operates under noload conditions, the relevant parameters are stator
resistance and stator selfinductance. Particularly at low motor speed, the speed
estimation is sensitive to an inaccurate stator resistance value in the observer model.
Also the leakage inductance value, being the decisive parameter at high motor
speed, should be properly tuned to the actual leakage inductance of the machine. The
implemented realtime adaptation of these parameters is based on monitoring of
magnetizing and daxis current in steady state. Assuming a constant rotor flux,
equation (4.3) can be simplified:
0
2
= = − =
dt
di
i i
d
u
u
τ ε (4.52)
Equation (4.52) is valid in every steadystate operating condition, guaranteed by the
flux controller of the fieldoriented control system. However, the error ε becomes
nonzero at a mismatch of stator resistance and inductance respectively. As can be
Sensorless Speed Control of Induction Motor Drives 81
derived from (4.1)(4.2), a resistance detuning yields an error of the daxis current
estimation. Thus, the error is positive, if the resistance is underrated. The same
applies for a positive detuned inductance.
The error value ε introduced by an inaccurate estimation of the stator resistance
decreases as the supply frequency increases. At high motor speed, stator resistance
detuning causes a negligible speed estimation error [Wang 99]. Therefore, knowing
that there is a finite precision in measurements of stator voltages and currents,
rational stator resistance adaptation is possible only at low motor speed. Due to
similar considerations, stator inductance detuning affects the speed estimation only
at higher motor speed. Practically, the error ε is used as a feedback signal to adapt
accurately stator resistance at low supply frequencies (ω
u
 < 5 Hz) and inductance at
high motor speed and supply frequencies (ω
u
 ≥ 5 Hz) respectively. Beyond these
boundaries, they are kept constant.
Figure 4.13 presents the experimental result of the stator resistance adaptation at
standstill. Starting with an initial error of 60 %, the adaptation is enabled at t = 0,5 s.
After a short period, magnetizing current matches the daxis current, confirming the
welltuned resistance value.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
2.8
3
3.2
3.4
3.6
3.8
t [s]
i
[
A
]
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
2
3
4
5
6
t [s]
R
s
[
Ω
]
i
d
i
u
Real R
s
Estimated R
s
Figure 4.13: Stator resistance adaptation at low speed (n = 0 rpm). Top: Magnetizing
and daxis current. Bottom: Estimated and real stator resistance value.
Figure 4.14 illustrates the stator inductance adaptation at n = 1000 rpm. A starting
error of the inductance (±44%) has been introduced resulting in a poor estimation of
the motor speed. The parameter adaptation scheme, switched on at t = 0,2 s, detects
the steadystate deviation of magnetizing and daxis current and tunes the
inductance.
It should be noted, that erroneous inductance estimation, directly reflected in both
incorrect torquecurrent mapping and load estimation, is not compensated by the
speed controller. The implemented speed controller with load torque rejection
consists of a proportional gain and contains no integralacting part. Thus, parameter
82 Chapter 5
mismatch yields a steadystate error of the speed control loop. However, this steady
state error can be used as an adaptation watchdog or, alternatively to the proposed
algorithm, as a parametercorrecting feedback signal.
0 0.25 0.5 0.75 1 1.25 1.5
3
4
5
t [s]
i
[
A
]
0 0.25 0.5 0.75 1 1.25 1.5
350
400
450
500
550
t [s]
L
s
[
m
H
]
0 0.25 0.5 0.75 1 1.25 1.5
975
1000
1025
1050
t [s]
n
[
r
p
m
]
0 0.25 0.5 0.75 1 1.25 1.5
2
2.5
3
3.5
t [s]
i
[
A
]
0 0.25 0.5 0.75 1 1.25 1.5
200
250
300
350
t [s]
L
s
[
m
H
]
0 0.25 0.5 0.75 1 1.25 1.5
994
996
998
1000
1002
t [s]
n
[
r
p
m
]
i
u
i
d
i
u
i
d
Parameter adaptation
switched ON
Parameter adaptation
switched ON
n
m
n
est
n
ref n
m
Figure 4.14: Adaptation of the stator inductance. Top: Magnetizing and daxis current. Middle: Stator
inductance. Bottom: Estimated, real and reference speed. Left: Initial inductance value 44 % overrated.
Right: Initial inductance value 44 % underrated.
Figure 4.15 shows the experimental result of the realtime inductance adaptation at
variable rotor flux. As can be seen, the inductance is clearly dependent on the
saturation level of the machine. Actual and estimated speed are in excellent
agreement, confirming the welltuned inductance value. Applying load torque, the
obtained results are equivalent. Usually, a flux controller keeps the rotor flux
constant. Rotor flux variations due to both load and electromagnetic torque changes
are small and do not significantly affect the observer performance. However,
according to (4.52), the adaptation must be disabled during fast rotor flux transients
at e.g. initial start, flux weakening and flux optimization.
With respect to the given drive setup, the influence of the leakage coefficient σ on
the observer performance is very low and hardly measurably. Furthermore, a
mismatch is partially compensated by the inductance adaptation. Therefore, an
adaptation of σ is not implemented.
No major problem exists in determining the stator frequency ω
u
. To the contrary, the
estimation error of the rotor frequency ω
slip
, directly reflected in the accuracy of the
rotor speed estimation ω
r
, depends on the rotor time constant. This error increases
proportionally to the qaxis current and load respectively. The rotor time constant
varies in a fairly wide range during operation. Variations of the rotor inductance are
caused by changes of magnetization. Furthermore, the rotor time constant changes
with the machine temperature. Assuming an equivalent influence of the saturation
level on both the stator and rotor inductance, the stator inductance adaptation
Sensorless Speed Control of Induction Motor Drives 83
scheme is used to compensate these variations. However, a realtime adaptation
scheme of the rotor time constant compensating the temperature variations has not
yet been realized. A simulation scheme based on the feedback of the observer state
error ∆x (figure 4.2) turned out to malfunction in practice, since the obtained signals
are too low and noisy to carry suitable information.
A promising solution of tuning the rotor time constant in realtime is based on the
evaluation of rotor slot harmonics [Jia 97]. This method permits high speed
accuracy in steady state and allows even position control. However, detection of
rotor slot harmonics should not be used as a standalone solution for speed
estimation, since the dynamic performance of such systems is very poor [Ish 82],
[Kre 92]. To the contrary, the dynamic performance of the proposed observer is
excellent. Together with the detection of rotor slot harmonics, compensating slow
temperature variations, a system with both high dynamic performance and high
steadystate accuracy is obtained.
0 10 20 30 40 50
1
2
3
4
5
t [s]
i
[
A
]
0 10 20 30 40 50
250
300
350
400
450
t [s]
L
s
[
m
H
]
0 10 20 30 40 50
990
995
1000
1005
1010
t [s]
n
[
r
p
m
]
1 2 3 4 5
250
300
350
400
450
iu [A]
L
s
[
m
H
]
n
m
n
est
Figure 4.15: Adaptation of the stator inductance. Top: Magnetizing current, daxis current and stator
inductance. Bottom: Estimated, real, reference speed and flux dependence of the stator inductance.
4.7 Conclusions
This chapter presents the design and the implementation of a fieldoriented high
performance motor drive with speed, flux and torque estimation. The speed
controlled induction motor drive requires no shaft sensor measuring speed or
position. The price to be paid is a more extensive and complicated control algorithm.
However, no additional measurements are required. The structure of the
implemented sensorless control is based on the extended Kalman filter theory.
Results of the dynamic and steadystate behavior of a sensorless speed control of an
induction motor are given. After the correct system model is chosen for the
Extended Kalman Filter, the results are satisfactory. Both at very low and at high
84 Chapter 5
motor speed with flux weakening, the proposed control scheme is working very
well. The described control system is a solution without mechanical sensors for a
wide range of applications where good steady state and dynamic properties are
required. Keeping the size of the program reasonable and still reaching a very good
performance is achieved by optimizing the model by hand calculations and
exploiting matrix symmetry.
The performance of the system increases as the information of the known
electromagnetic torque is used. Only the load torque is handled as if it were an
unknown system disturbance. If the loadspeed relation is known, this information
can be used for a further improvement of the performance. The speed estimation
does not lag the actual motor speed, both in steady state and during periods of
acceleration or braking. Steadystate errors are used for parameter adaptation.
5. Sensorless Speed Control of PMSM
5.1 Introduction
With the introduction of permanent magnets with a high flux density as well as a
high coercivity in the late eighties, synchronous motors with permanent magnets
became an attractive alternative for applications in high performance variable speed
drives. Significant advantages arise from the simplification in construction, the
reduction in losses and the improvement in efficiency. One of the most active areas
of control development during recent years involving these motor types has been the
evolution of new techniques for eliminating the position and speed sensor.
Elimination of the shaftmounted sensor is required in many applications since this
device is often one of the most expensive and fragile components in the entire drive
system.
The approaches to sensorless drives vary depending on the rotor flux distribution. A
motor with a trapezoidal rotor flux distribution (BLDM, brushless dc motor)
provides an attractive candidate: Two out of three stator windings are excited at the
same time and the unexcited winding can be used as a sensor. The control scheme as
well as the position detection is relatively simple. The rotor speed and position can
be determined by the electromagnetic field induced in the unexcited winding
[Erd 84], [Mat 90]. This is usually done either by a zero crossing approach of the
backEMF or by a phaselocked loop technique to lock on to the backEMF
waveform in the unexcited winding. It is enough to detect the rotor position every
60° to obtain a proper switching sequence.
On the contrary, the permanent magnet synchronous motor (PMSM), having a
sinusoidal flux distribution, excites all three windings at the same time. Both the
control algorithm and the speed/position estimation become more complicated. The
information on the rotor position is required continuously. However, the PMSM is
applicable for fine torque control where a very low level of torque pulsations is
required. Several schemes for position sensorless operation of PMSM have been
reported in literature and are reviewed e.g. in [Raj 94]. The position detection
methods are mainly based on Kalman filtering or Model Reference Adaptive
Systems using the motor parameters and measurements of motor currents and
voltages.
86 Chapter 6
The drive system studied in this chapter is a sensorless control of a PMSM based on
the extended Kalman filter theory using only the measurements of motor current and
dc bus voltage for the estimation of speed and rotor position. On top of the speed,
also the acceleration of the drive is estimated offering a significant improvement of
the drive performance. The applied approach is mainly a transfer of the earlier
described sensorless control of the induction motor to the motor equations of the
PMSM. Therefore, this chapter explains only differences in detail while many still
valid statements of previous chapters are not repeated. Theoretical analyses based on
the physical viewpoint are presented and the associated experimental results are
shown. This chapter also describes the influence on the control design reflected by
the feedback of the estimated values. A torque that at average differs from zero, is
only produced if the excitation is precisely synchronized with the rotor speed and
instantaneous position. The controller has to ensure that the motor never experiences
loss of synchronization. Due to rotor asymmetry, the PMSM is also suitable for
position control. A 3 kW, 4 kW and 45 kW PMSM have been used to verify this
approach. The discussion ends by evaluating a parameter adaptation scheme in real
time to track motor parameter variations.
5.2 Model of the PMSM in Discrete Time
The system model considered is a PMSM having permanent magnets mounted on
the rotor. The resulting backEMF voltage induced in each stator phase winding
during rotation can be modeled quite accurately as a sinusoidal waveform. A
mathematical model describing the PMSM motor dynamics in a rotor flux reference
frame is well known [Jah 86], [Hen 92]. The electrical properties of the PMSM in
continuous time are completely described by two stator voltage equations:
q q r
d
d d s d
i L
dt
di
L i R u ω − + = (5.1)
Md r d d r
q
q q s q
i L
dt
di
L i R u Ψ + + + = ω ω (5.2)
In a PMSM with surfacemounted magnets, torque control can be achieved very
simply, since the instantaneous electromagnetic torque can be expressed similarly to
that of the dc machine as the product of qaxis current i
q
and magnet flux Ψ
Md
. In
case of interior permanent magnets, the additional reluctance torque can be
exploited:
( ) ( )
d d q Md q el
i L L i p T − − Ψ = (5.3)
In contrary to the induction machine, the flux angle γ rotates synchronously with the
rotor speed. With the same simplifications as introduced in the induction motor
study, the mechanical equation is:
Sensorless Speed Control of PMSM 87
dt
d
p
J
dt
d
J T T
r
load el
ω ω
= = − (5.4)
r
dt
d
ω
γ
= (5.5)
Tolerating a small discretization error, the transformation from continuous time to
the discrete time state space is equivalent to:
d
d
s q
d
q
r s d
d
s
s k d
L
u
T i
L
L
T i
L
R
T i + +


.

\

− ≈
+
ω 1
1 ,
(5.6)
q
q
s
q
Md
s d
q
d
r s q
q
s
s k q
L
u
T
L
T i
L
L
T i
L
R
T i +
Ψ
− −


.

\

− ≈
+
ω 1
1 ,
(5.7)
( ) ( )
load s d d q Md q s r k r
T
J
p
T i L L i
J
p
T − − − Ψ + ≈
+
2
1 ,
ω ω (5.8)
r s k
T ω γ γ + ≈
+1
(5.9)
From the control viewpoint, the PMSM has four electrical parameters: stator
resistance R
s
, d and qaxis inductance L
d
and L
q
, and permanent magnet flux linkage
Ψ
Md
. The inductances are considered to be constant, which is verified by
measurements [Van 98] and numerical calculations of the given PMSM [Pah 98].
The influence of parameter variations is compensated by realtime adaptation of the
flux linkage Ψ
Md
.
The known electromagnetic torque is used as part of the speed calculation, vastly
increasing the accuracy of the speed specification and the dynamics of the drive.
However, this approach requires information on the load, since the Kalman
algorithm assumes a zero mean value of the disturbances. As mentioned in the
previous chapter, an additional estimation of the load torque increases the observer
performance as well as the performance of the speed control loop. The price to be
paid is a minor extra computing time. Therefore, the acceleration due to the load
torque is estimated additionally:
0 ≈ 
.

\

=
load
l
T
J
p
dt
d
dt
dα
(5.10)
load l k l
T
J
p
= ≈ ⇒
+
α α
1 ,
(5.11)
88 Chapter 6
The dynamic model for the PMSM, choosing d and qaxis current i
d
, i
q
, the
electrical rotor speed ω
r
, rotor position γ and the acceleration α
l
as state variable x
k
and the fundamental voltage as input u
k
, is described by following equations. The
output vector y
k
consists of the estimated motor current in a statorfixed reference
frame.
k k k
u x x
k k
B A + =
+1
;
k
s
s
k
U
U
u


.

\

=
β
α
;
k
l
r
q
d
k
i
i
x






.

\

=
α
γ
ω ; (5.12)
( ) ( )










.

\

− − − Ψ
Ψ
− − −
−
=
1 0 0 0 0
0 1 0 0
0 1 0
0 0 1
0 0 0 1
2
s
s d d q Md
s
q
Md
s
q
s
s
q
d
s r
d
q
s r
d
s
s
T
T i L L
J
p T
L
T
L
T
R
L
L
T
L
L
T
L
T
R
ω
ω
k
A (5.13)
( ) ( )
( ) ( )









.

\

−
=
0 0
0 0
0 0
cos sin
sin cos
γ γ
γ γ
q
s
q
s
d
s
d
s
L
T
L
T
L
T
L
T
k
B (5.14)
At each time step, using the previously predicted position and current information,
the current is estimated in two stages to correct the predicted states by the Kalman
feedback matrix.
( ) ( )
( ) ( )
k k s
s
k
x x
I
I
y


.

\
 −
= =


.

\

=
0 0 0 cos sin
0 0 0 sin cos
ˆ
ˆ
γ γ
γ γ
β
α
k
C (5.15)
The model matrices B
k
and C
k
depend on the position of the rotor γ, the matrix A
k
on daxis current i
d
, flux linkage Ψ
Md
and rotor speed ω
r
. The block diagram of the
discrete motor model together with the feedback matrix of the observer is equal to
the one shown in chapter 5. In speed control mode, the flux angle is limited to
Sensorless Speed Control of PMSM 89
γ  < π. In contrary to the induction motor drive, the studied PMSM is also suitable
for position control since the rotor asymmetry can be exploited. Therefore, the
overflow protection has to be disregarded in position control mode. However, a loss
of exact position information is not admissible in any case.
5.3 RealTime Implementation
According to the EKF algorithm described in chapter 5, all equations can be written
as a function of a system vector Φ and an output vector h describing the re
linearized model of the PMSM. The derivatives of the output and the transposed
system vector are:
( )
( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( )


.

\

− −
=


.

\

−
− − −
=
∂
∂
=
∂
∂
0
ˆ
0 cos sin
0
ˆ
0 sin cos
0 sin cos 0 cos sin
0 cos sin 0 sin cos
α
β
γ γ
γ γ
γ γ γ γ
γ γ γ γ
i
i
i i
i i
x
x
x
h
q d
q d
k
k
k
k
C
(5.16)
( )
( )
( ) ( )
( )












.

\

−
−
Ψ + −
− − Ψ −
−
−
− −
=


.

\

∂
+ ∂
=


.

\

∂
Φ ∂
1 0 0 0
0 1 0
0 1
0 0 1
0 0 1
2
2
s
d
q
s
q
d
s
s Md d d
q
s
q
d
q
s
d d q Md
s
q
s
s
d
q
s r
q d q
s
q
d
s r
d
s
s
T
k
k k
T
k
k
T
u
L
T
u
L
T
T i L
L
T
i
L
L
T
i L L
J
p T
L
T
R
L
L
T
i L L
J
p T
L
L
T
L
T
R
x
u x
x
ω
ω
k k
B A
(5.17)
The noise covariance matrices Q and R and an initial matrix P
00
are evaluated
corresponding to the remarks on the induction motor drive. The matrices R, Q and
P
00
are diagonal due to the lack of sufficient statistical information to evaluate their
offdiagonal terms. Furthermore, their diagonal nature saves a lot of computing time.
P
00
affects neither the transient performance nor the steady state conditions of the
system and can be chosen at random. An offline current measurement, referring to
the 3 kW PMSM drive installation, yields a measurement noise covariance matrix,
which is almost proportional to the dc bus voltage within a voltage range
300V < U
dc
< 600V:
90 Chapter 6
2 3
2
5
A 10
7 , 1 0
0 7 , 1
V
A
10
15 , 1 0
0 15 , 1
− −


.

\

−


.

\

=
dc
U R (5.18)
The measurement of the noise largely exhibits independence of the motor current.
The values a larger compared to the measurement noise of the induction machine
supplied by the same inverter. This is mainly due to the smaller inductances of the
PMSM smoothing the PWM pulses. For an external field (armature reaction) the
magnets behave as air, introducing a large reluctance and thus a low main
inductance. The coefficients of the system covariance matrix are calculated
according to subsection 4.4.1 taking a sample time T
s
= 200 us and the parameter of
the 3 kW PMSM into account:
2
2
A 0,021
3
1
) 1 , 1 ( =


.

\
 ∆
=
d
s
L
u T
Q (5.19)
2
2
A 0,0059
3
1
) 2 , 2 ( =


.

\

∆
=
q
s
L
u T
Q (5.20)
2
5 6 10 4
s
1
10 2 , 3 s 10 2 ) 3 , 3 (
− −
⋅ = ⋅ =
s
T Q (5.21)
11 4 4 4
10 2 , 3 s 10 2 ) 4 , 4 (
− −
⋅ = ⋅ =
s
T Q (5.22)
4 6
8 2 2
s
1
2 , 43
s
1
10 2 , 1 ) 5 , 5 ( ⋅ = ⋅ = k T p k Q
s
, with: k ≤ 1 (5.23)
The transient performance of the observer is tuned by the factor k in (5.23). All other
coefficients of the system covariance matrix are set to the given values. A high
tuning factor k increases the dynamic performance, but also the noise of the
estimated signals.
Within the implemented speed control, the information on load acceleration is used
as input of the speed controller directly compensating the load torque (figure 5.1).
Rejecting load disturbances inproves the dynamic stiffness of the drive and is
superior compared to common PI controller [Lor 99]. However, a rough torque
command results in increased torque ripples and motor heating by current
harmonics. In order to obtain a good compromise between dynamic performance
and a smooth torque command response and with respect to the given installation,
the tuning factor k is set to 10%. Figure 5.1 shows the structure of the implemented
position and speed controller with load torque rejection. The calculated reference
torque T
el
*
is mapped into reference commands for d and qaxis current. The current
commands i
d
*
and i
q
*
are extracted according the constraint of maximum torqueper
ampere operation, being nearly equivalent to maximum drive efficiency [Bose 97].
Sensorless Speed Control of PMSM 91
ω
*
T
el
*
Θ
Position
controller
Θ
*
ω ⇔ Θ
Position
reference
K
pp
pn s
K T
z
/
1
1 −
−
ω
*
ω
*
ω
T
el
<T
max
l load
p
J
T αˆ
ˆ
=
Proportional
gain
K
n
Speed controller with
load torque rejection
Control mode
Figure 5.1: Position and speed controller with load torque rejection.
Figure 5.2 shows the block diagram of the entire control system with the proposed
observer integrated in the digital motion control loop. The inputs of the control
system are measured motor current in two phases and the dc bus voltage. The
voltages required as input for the EKF are obtained from the reference voltages,
available at the output of the system. Due to the nonlinearity of the inverter, a phase
voltage calculation block is added compensating for this nonlinearity. The
homopolar component of the phase voltages arising due to the SVM [Leo 85] is also
considered within this block. The measurements should not be additionally filtered
since the Kalman filter handles with white and uncorrelated measurement noise and
produces the minimum variance estimate. Therefore, this is already an optimal filter.
Generally, the estimated states are used as feedback signals of the controller,
because they are less disturbed compared to measured values. Furthermore, the
smoothness of the state signals can be tuned by the system and measurement
covariance matrices.
ia
ib
iα
iβ
AC voltage
calculation
3 ⇒ 2
iα
iβ
sin(γ)
cos(γ)
ωr
id
iq
EKF &
model
ud
*
uq
*
id,error
iq,error
sin(γ)
cos(γ)
uα
*
uβ
*
current
control
inverse
Park Trans.
PWM
generation
Speed
reference
id
*
iq
*
Speed
Control
ud
*
uq
*
∆ud
*
∆uq
*
decoupling
ub
*
uc
*
ua
*
uα
uβ
Udc
ia
ib
A/D
αl
SVM
2
Udc
Figure 5.2: Velocity observer integrated into the digital motion control loop.
The entire speed control system consists of a speed and two current controllers. The
torque of the PMSM is controlled by a reference current, calculated by the speed
controller. Due to the higher qaxis inductance, a negative daxis current is
impressed to benefit from the reluctance torque. In position control mode, the speed
reference is given by an overlaid position controller (figure 5.1), using the estimated
rotor position as input.
The program code of the EKF is optimized according to the remarks specified in
chapter 5. The computing requirement of the final algorithm takes up 236
multiplications and 178 summations. Compared to the induction machine, both the
EKF program code and the control algorithm are less extensive. The realtime
92 Chapter 6
implementation of the Kalman filter integrated into the motion controller is carried
out using a TMS320C31 DSP in which the turnaround time of the entire control
system amounts to 153 us. Therefore, the filter can operate in a system having a
maximum sampling frequency of 6,5 kHz, or a theoretical system bandwidth of
3,25 kHz. This high bandwidth allows the EKF to be used in highperformance real
time motion systems.
5.4 Experimental Results
The proposed speed sensorless control scheme has been tested using a 3 kW, 4 kW
and 45 kW PMSM (data are given in appendix B). However, to keep the presented
results clear, all experimental results presented in this chapter are measurements
using the 3 kW prototype motor. Some additional experiments regarding the 4 kW
motor, specially designed for PVpowered water pump systems, are presented in
chapter 7.
Since the DSP power is simultaneously used for monitoring purposes and recording
experimental data, the sample time used is fixed to T
s
=200 us. A dc generator with
constant excitation coupled to a variable resistor bench is used to load the PMSM.
Via a power switch a load step can be applied. Additionally, the load torque is
measured by a torque transducer. All experimental results and measurements are
carried out using the estimated states as feedback to a speed controller with load
torque rejection (figures 5.1 and 5.2). The bandwidth of both current controllers
using also the estimated values of d and qaxis current is about 950 Hz. The
bandwidth of the current loop is not decreased by using the EKF instead of the field
oriented control with position measurement.
Figure 5.3 shows the experimental results of a speed reversal using the estimated
speed and position as feedback. Additionally, the real speed and position are
measured for comparison. There is a very good agreement between real and
estimated speed and position respectively. Using the information on generated
electromagnetic torque and drive acceleration, the noise as well as the lag in the
estimated speed signal is even lower than the measured and filtered speed signals
during transients.
Furthermore, figure 5.3 exhibits the influence of signal lag due to data transmission.
Without any delay, the required phase voltages, calculated by means of the voltage
references controlling the inverter, are directly available in the control loop. To the
contrary, the affiliated current response is measured not before the next sample
period of the digital control system. Neglecting the current signal lag causes a poor
estimation of the position angle. Therefore, an extra sample delay is added in the
loop of the estimated phase voltages used in the observer algorithm.
The observer presented achieves the objective of eliminating the lag of the estimated
motor speed by additional estimation of the load. Figure 5.4 presents the response of
Sensorless Speed Control of PMSM 93
the PMSM to a load step (75% rated torque) at a motor speed of 1000 RPM. The
information on load acceleration is directly used to compensate for the load torque.
Rejecting load disturbances increases the dynamic stiffness of the drive. Therefore,
this feedback causes the disturbance to perceive a more robust system responding
less to disturbances. Compared to a common PI speed controller, the overshoot at
steps of both speed reference and load torque is vastly decreased or even vanishes
since the speed controller used (figure 5.1) contains of no integralacting part.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
500
250
0
250
500
t [s]
n
[
r
p
m
]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
5
0
5
10
15
20
t [s]
i
q
[
A
]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
20
10
0
10
20
t [s]
∆
n
[
r
p
m
]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
0.2
0.1
0
0.1
0.2
t [s]
∆
γ
[
r
a
d
]
n
ref
(a)
i
q
*
i
q
(b)
∆n = n
est
 n
m
(c)
Without extra delay With extra delay
(d)
Figure 5.3: Speed reversal test. Top: Speed reference n
ref
, estimated speed n
est
and measured speed n
m
.
Middle: Estimated qaxis current (b) and difference between estimated and measured speed (c).
Bottom: Error of the angle estimation and influence of current/voltage signal lag.
94 Chapter 6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
980
1000
1020
t [s]
n
[
r
p
m
]
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
5
0
5
10
15
20
t [s]
i
q
[
A
]
0 0.2
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
5
0
5
10
15
t [s]
T
[
N
m
]
T
load
n
m
n
est
T
el
T
m
T
el
Figure 5.4: Response to a load step (75% rated torque). Top: Measured and estimated speed. Middle:
Estimated qaxis current. Bottom: Estimated load T
est
, measured load T
m
and electromagnetic torque T
el
.
In steady state, the estimated load in figure 5.4 equals to the measured load. This is
not obvious since iron and friction losses of the PMSM are part of the estimated but
not of the measured torque. Furthermore, erroneous electromagnetic torque
calculations and inertia identification are directly reflected at the load calculation.
However, this influence is small compared to a potential load variation. In fact, the
torque calculation might be incorrect, but the real load is compensated by the real
torque and the steady state speed error sticks to zero. Only the dynamics of the load
estimation are important for exact speed calculation without any delay. As can be
seen, the delay between estimated and measured load is insignificant. Therefore, the
proposed speed control offers a vast improvement of the drive performance also if
the load is not absolutely known.
At low motor speed (n ⇒ 0), the equations of the PMSM are simplified, as the
voltage induced by the magnets is very small. Therefore, no prediction can be made
on the position of the magnets and the EKF fails. Since at standstill only dcvalues
are given, the necessary flux variation must be forced by impressing a test signal
into the system. A signal, easy to implement, is an additional sinusoidal reference
current in the daxis of the motor, using the d/q axissymmetry of the rotor to
estimate the real position. In all experimental results presented the following daxis
reference current is used:


.

\

− ⋅ + =
+ =
rpm 300
1 )
s
1
100 2 sin( A 3
*
*
,
n
t i
i i i
d
test d ref d
π
, with: n ≤ 300 rpm (5.24)
whereby the reference daxis current i
d
*
results from the speed controller calculating
the required torque motor. The amplitude and frequency of the test signal is
Sensorless Speed Control of PMSM 95
experimentally chosen regarding observer stability and low acoustic noise level.
Nevertheless, further investigations on optimal shape, frequency and magnitude of
the additional daxis current have to be made. Figure 5.5 presents the response of the
d and qaxis current to a step of the speed reference from standstill to 1000 rpm.
The corresponding speed signal is shown in figure 5.6. The unwanted reluctance
torque, generated by the test signal in the daxis, is compensated by an appropriate
qaxis current. The modification of the qaxis current i
q
*
, calculated by the speed
controller, is obtained by the demand for a constant electromagnetic torque, not
disturbed by the impressed test signal.
( ) ( ) ( ) ( )
ref d d q Md ref q d d q Md q el
i L L i p i L L i p T
, ,
!
* * *
− − Ψ = − − Ψ = (5.25)
( )
( )





.

\

−
−
Ψ
+ =
− − Ψ
− − Ψ
= ⇒
ref d
d q
Md
test
q
ref d d q Md
d d q Md
q ref q
i
L L
i
i
i L L
i L L
i i
,
*
,
*
*
,
1 (5.26)
0 0.2 0.4 0.6 0.8 1 1.2 1.4
0
5
10
15
20
t [s]
i
q
[
A
]
0 0.2 0.4 0.6 0.8 1 1.2 1.4
10
5
0
5
t [s]
i
d
[
A
]
i
d
n = 300 rpm
n = 1000 rpm
Figure 5.5: Current at a speed step (figure 5.6). Top: Reference i
q
*
and qaxis current i
q
.
Bottom: Reference i
d
*
and daxis current i
d
.
The developed torque remains nearly constant as can be seen on figure 5.6, showing
the corresponding speed response, marked optimum torque control. The optimal
control of the motor takes advantage of the reluctance torque by introducing a
negative (L
d
< L
q
) direct axis current component. In the same figure, a comparison is
given of motor control with feedback of the estimated speed and position, optimum
daxis current and no daxis current respectively. In spite of identical maximum
current amplitude, the maximum torque using optimum torque control is higher,
yielding a faster acceleration of the motor. The bandwidth of the speed control with
the EKF is comparable to the common control with speed measurement due to the
omission of the filter for speed measurement.
96 Chapter 6
0 0.2 0.4 0.6 0.8 1 1.2 1.4
0
200
400
600
800
1000
1200
t [s]
n
[
r
p
m
]
n
ref
Optimum torque
control
i
d
= 0
Figure 5.6: Speed step with feedback of the estimated speed and position (EKF). Comparison of torque
control with optimum daxis current and no daxis current respectively.
5.5 Position Estimation and Startup
Using a position sensor as feedback device, a special startup strategy is required to
find the absolute rotor position as indicated by an encoder index pulse. This startup
procedure is necessary in both position control and speed or torque control mode. A
torque is only produced if the excitation is precisely synchronized with the rotor
speed and instantaneous position. A startup strategy is executed by impressing a
(assumed) qaxis current and slowly increasing the initial assumption of the rotor
angle until the motor rotates and the index pulse is found. However, the motor has to
rotate up to one mechanical revolution. Once the index is found, all registers are
reset and the drive is ready for normal operation.
In drive systems without a position sensor, it is generally difficult to estimate the
initial rotor position. If the rotor position can not be exactly estimated, the starting
torque of the motor decreases and the motor may temporarily rotate in the wrong
direction after start. A starting strategy often proposed is based on energizing two
windings by a large armature current (about rated current) and expecting the rotor to
align with a certain definite position. This method yields the direction of the magnet
axis but cannot distinguish between North and South Pole.
To the contrary, the presented sensorless control scheme is selfstarting. Here, the
variation of the inductance as a function of the rotor position is used to obtain the
position. Due to the low permeability of the magnet material, the inductance along
the qaxis of the PMSM with interior permanent magnets is larger than the
inductance along the daxis. Impressing the test signal (5.24), the difference is
detected by the Kalman algorithm and the estimated position converges
automatically to the real position. Figure 5.7 presents the initial startup of the
digital control system. Aligning the initial value of the estimated position to the
Sensorless Speed Control of PMSM 97
magnet position, the resulting convergence is very smooth. If the initial value of the
estimated position is opposite to the rotor position, the motor temporarily rotates in
the wrong direction. This effect can be avoided by operating the drive in openloop
control and impressing a test signal in one motor phase. Once the position is
detected, the drive returns to the closedloop control.
The controller has to ensure that the motor never experiences loss of
synchronization. However, the rotor asymmetry makes the PMSM also suitable for
position control (figure 5.8). For the drive setup with the 3 kW PMSM, the steady
state error of the electrical position angle is smaller than 2,3°. This error, as well as
the performance of the algorithm, mainly depends on the quality and accuracy of
voltage and current measurement. It should be remarked, that the proposed
algorithm only identifies the electrical position. The absolute mechanical rotor
position is not detectable.
0 0.5 1 1.5
4
2
0
2
4
t [s]
γ
[
r
a
d
]
0 0.5 1 1.5
100
50
0
50
100
t [s]
n
[
r
p
m
]
0 0.5 1 1.5
0.4
0.2
0
0.2
t [s]
γ
[
r
a
d
]
0 0.5 1 1.5
10
5
0
5
10
t [s]
n
[
r
p
m
]
γ
m
γ
m
γ
est
γ
est
n
m
n
est
n
m
n
est
Figure 5.7: Startup of the sensorless speed control. Top: Estimated position γ
est
and measured positionγ
m
.
Bottom: Estimated and measured speed. Left: Initial value of the position estimation almost opposite to
the rotor position. Right: Initial value of the position estimation almost aligned to the rotor position.
98 Chapter 6
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
50
0
50
100
150
200
t [s]
n
[
r
p
m
]
0 0.05 0.1
0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
1
2
3
t [s]
γ
[
r
a
d
]
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0.1
0.05
0
0.05
0.1
t [s]
∆
γ
[
r
a
d
]
∆γ = γ
est
 γ
m
γ
ref
n
est
n
m
n
ref
Figure 5.8: Position control of the 3 kW PMSM. Top: Position reference γ
ref
, estimated position γ
est
and
measured position γ
m
. Middle: Difference between estimated and measured position. Bottom: Speed
reference n
ref
, estimated speed n
est
and measured speed n
m
.
Sensorless Speed Control of PMSM 99
5.6 Motor Parameter Adaptation
The motor model of the PMSM as well as the implemented EKF contains four
electrical motor parameters: d/qaxis inductances L
d
and L
q
, stator resistance R
s
and
permanent magnet flux linkage Ψ
Md
. For the given PMSM, the ohmic voltage drop
is very small. Thus, the influence of a stator resistance variation is very low and
hardly measurably. The key mechanical drive parameter is the moment of inertia J.
A mismatch of the inertia affects the observer performance only during transients
and causes no steadystate error. Therefore, an adaptation of R
s
and J is not
considered.
The most influential motor parameter, affecting the steadystate error and the
observer performance, is the permanent magnet flux linkage Ψ
Md
. Applying flux
adaptation, the torquecurrent mapping via a lookup table according to figure 2.12
is no longer suitable. Assuming exact knowledge of the motor parameters and using
the daxis current, the torque reference T
el
*
, determined by the speed controller, is
transformed to a qaxis current reference i
q
*
:
) ) ( (
*
*
d d q Md
el
q
i L L p
T
i
− − Ψ
= (5.27)
According to (2.50), the optimum torque control of the PMSM yields two solutions
for the daxis current reference i
d
*
:
2
*
2
*
) ( 2 ) ( 2
q
d q
Md
d q
Md
d
i
L L L L
i +


.

\

−
Ψ
±
−
Ψ
= (5.28)
The positive sign is valid for PMSM with L
d
> L
q
. Here, only the case L
d
≤ L
q
(PMSM with inset magnets) is considered. Thus, the negative sign in (5.28) must be
used.
Erroneous flux estimation yields incorrect speed estimation. Furthermore, it is
directly reflected in the torquecurrent mapping. An incorrect torquecurrent
mapping is not compensated by the speed controller since the implemented speed
controller with load torque rejection consists of a proportional gain and contains no
integralacting part. Thus, erroneous flux estimation results in a steadystate error of
the speed control loop. However, this error is used for flux adaptation. The structure
of the implemented flux adaptation, the speed controller with load torque rejection
and the modified torquecurrent mapping is shown in figure (5.9). The
electromagnetic torque is almost proportional to the flux linkage. According to (5.3)
and (5.8), increasing the estimated flux linkage results in a higher absolute value of
100 Chapter 6
the estimated electromagnetic torque and speed respectively. The presented
adaptation must be disabled at steps of the speed reference to avoid erroneous flux
calculation during transients.
sign(ω
*
)
Tload
id
*
p ΨMd
LqLd 2
iq
*
x1
x2
2
2
2
1
x x +
Kn
Tel
*
Tel < Tmax
ω
*
ω
α
J/p
∆ΨMd
id
∆ΨMd K
Ψ
∆ΨMd = ∫ K
Ψ
∆ω dt
∆ω
Current mapping
Speed control and
flux adaptation
Initial value
Initial value
Flux
error
Figure 5.9: Realtime adaptation of the flux linkage, speed control
with load torque rejection and modified current mapping (L
d
≤ L
q
).
Figure 5.10 presents experimental results of the proposed flux adaptation. An initial
error of the flux linkage (±20 %) has been introduced resulting in poor motor speed
estimation. The speed estimation as well as the steadystate error is affected by a
parameter mismatch. The flux adaptation detects the steadystate error and corrects
the initial flux linkage. After a short period, the estimated speed matches the
measured speed, indicating the correct estimation of the flux linkage.
0 0.5 1 1.5 2 2.5 3
0.25
0.26
0.27
0.28
0.29
0.3
0.31
t [s]
Ψ
M
d
[
V
s
]
0 0.5 1 1.5 2 2.5 3
10
5
0
5
10
t [s]
∆
n
[
r
p
m
]
0 0.5 1 1.5 2 2.5 3
0.2
0.21
0.22
0.23
0.24
0.25
0.26
t [s]
Ψ
M
d
[
V
s
]
0 0.5 1 1.5 2 2.5 3
10
5
0
5
10
t [s]
∆
n
[
r
p
m
]
estimated
flux linkage
real flux
linkage
Parameter adaptation
switched ON
Parameter adaptation
switched ON
∆n = n
est
 n
ref ∆n = n
est
 n
m
Figure 5.10: Adaptation of the flux linkage. Top: Estimated and real flux linkage. Bottom: Difference
between reference speed (n
ref
= 1000 rpm), estimated speed n
est
and measured speed n
m
. Left: Initial flux
linkage 20 % overrated. Right: Initial flux linkage 20 % underrated.
Sensorless Speed Control of PMSM 101
Considering PMSM’s with magnet placing of the insettype, the daxis inductance
L
d
is generally independent of the load state [Cha 85]. A slight L
d
mismatch and
variations due to different saturation levels are completely compensated by an
appropriate variation of the flux linkage Ψ
Md
[Van 98]. Figure 5.11 presents the flux
adaptation, based on the structure shown in figure 5.9, at variable daxis current, no
load and a motor speed n = 1000 rpm. The coincidence of estimated and measured
speed verifies the proposed approach.
0 5 10 15 20 25 30 35 40 45 50 55 60
15
10
5
0
5
10
15
t [s]
i
d
[
A
]
0 5 10 15 20 25 30 35 40 45 50 55 60
0.24
0.25
0.26
0.27
t [s]
Ψ
M
d
[
V
s
]
0 5 10 15 20 25 30 35 40 45 50 55 60
4
2
0
2
4
t [s]
∆
n
[
r
p
m
]
∆n = n
est
 n
m
∆n = n
est
 n
ref
Figure 5.11: Flux adaptation at variable daxis current (no load). Top: daxis current.
Middle: Estimated flux linkage. Bottom: Difference between reference
speed (n
ref
= 1000 rpm),estimated speed n
est
and measured speed n
m
.
Furthermore, experimental investigations have shown the capability of the flux
adaptation to compensate also for a slight mismatch of the qaxis inductance L
q
as
well as for a loaddependent variation/saturation. Figure 5.12 demonstrates the flux
adaptation at variable load torque, i
d
*
= 0 A and a motor speed n = 1000 rpm. Again,
the coincidence of estimated and measured speed shows the validity of the approach.
Thus, all motor parameters, except for the flux linkage, are set constant. The
influence of parameter variations is compensated by flux adaptation. The approach
of setting the inductances of the given PMSM constant is also verified by numerical
calculations [Pah 98] and measurements [Van 98]. However, a mismatch of motor
parameters is not arbitrary. Approximate values, guaranteeing the stable operation of
the observer, are also required for exact tuning of the current controller.
102 Chapter 6
0 5 10 15 20 25 30 35 40 45 50 55 60
0
5
10
15
t [s]
i
q
[
A
]
0 5 10 15 20 25 30 35 40 45 50 55 60
0.25
0.255
0.26
0.265
t [s]
Ψ
M
d
[
V
s
]
0 5 10 15 20 25 30 35 40 45 50 55 60
4
2
0
2
4
t [s]
∆
n
[
r
p
m
]
∆n = n
est
 n
m
∆n = n
est
 n
ref
Figure 5.12: Flux adaptation at variable load (i
d
*
= 0 A). Top: qaxis current.
Middle: Estimated flux linkage. Bottom: Difference between reference
speed (n
ref
= 1000 rpm), estimated speed n
est
and measured speed n
m
.
5.7 Conclusions
This chapter presents the design and the implementation of sensorless speed control
of permanent magnet synchronous motor drives. The algorithm used is based on the
extended Kalman filter theory. A systematic and analytic approach for developing
the algorithm is given. The discrete extended Kalman filter is well suited to speed
and rotor position estimation of a PMSM. The proposed approach has been validated
by means of realtime experiments using a TMS320C31 DSP. The high bandwidth
allows the EKF to be used in highperformance realtime motion systems.
The known electromagnetic torque is used as part of the speed estimation, vastly
increasing the accuracy and dynamic performance of the drive. The implemented
speed controller with load torque rejection contains no integralacting part,
providing a system with extremely high stiffness to disturbance inputs. The speed
estimation does not lag the actual motor speed, both in steady state and during
periods of acceleration or braking. A negative daxis current is impressed to benefit
from the reluctance torque.
The presented sensorless control scheme is selfstarting. At low motor speed, the
required flux variation is forced by impressing a test signal in the daxis. The
unwanted reluctance torque is compensated by a complementary qaxis current. Due
to the rotor asymmetry, the PMSM is also suitable for position control.
Sensorless Speed Control of PMSM 103
Mismatch of motor parameters yields incorrect speed estimation and erroneous
torquecurrent mapping. Therefore, a realtime flux adaptation scheme, tracking
motor parameter variations, has been implemented. All proposed control approaches
are verified by experimental results.
6. PVPowered Water Pump Systems
6.1 Introduction
The use of photovoltaic (PV) energy sources for water pumping and irrigation
applications, especially in remote or rural areas in developing countries, is receiving
considerable attention. Large urban populations in developing countries do not have
access to safe drinking water sources (standpipes or boreholes) or to sanitary
services (sewers, septic tanks or wet latrines). According to statistics of the World
Health Organization, the number of people without access to safe water in 1990 was
1,1 billion [WHO 96]. Human health depends on an adequate supply of potable
water. Therefore, PVpowered water pump systems can improve peoples living
conditions, where power from a utility is not available or too expensive to install.
Furthermore, it is not economically viable to connect such remote areas to the
national electric grid.
While many of the references for residential applications are available in technical
details, it is difficult to locate technical references for the interaction between PV
arrays and an electric machine, especially in water pumping without battery storage
[Mul 97]. This chapter briefly reviews present technology and applications of PV
powered water pump systems and exhibits an extensive description of a new control
approach.
The two basic design approaches of PV arrays for water pumping system
applications are the use of battery, for a backup of the pumping system, and the
other is to pump directly from the PV power without battery. There are advantages
and drawbacks associated with each design. With a battery module, the system
energy generated by the sun can be stored in the battery. With the second approach,
the motor/pump subsystem can be powered either by directly connecting to the PV
array, or by using a maximum power point tracker (MPPT), a dcdc converter and an
inverter interfacing motor and PV array. In this work, a system is designed, not
requiring a battery. The fact that no battery is required is a key element in the
design. Batteries tend to be very unreliable in the overall framework and
furthermore, they are of “interest” to people living there for other purposes, too
(read: they are often stolen).
The system analyzed here is a PV powered water pumping system avoiding the use
of the additional dcdc converter, a battery and its losses. Several new approaches
106 Chapter 8
are developed, analyzed and tested, as the algorithms described in literature
[Mul 97], [Dus 92] for this kind of systems turned out to malfunction, when tested
under realistic conditions. Due to the lack of storage in the dc bus, the power of the
PV array must be used immediately to accelerate an ac motor. To optimize the
energy captured by the PV array and to pump as much water as possible, the output
power should always be at its maximum power point. Therefore, a novel MPPT
algorithm, realized by feeding back the dc voltage and current to a controller, has
been implemented. The entire system is controlled by a digital signal processor
(DSP) based developing platform realizing MPPT, dc bus voltage control,
speed/torque control of the drive and startup and shut down automatism's.
Considering realistic conditions, advantages and drawbacks of the different control
units are discussed. Measured results are presented and evaluated to demonstrate the
performance and the stability of the system developed here.
6.2 Pilot Installation
The system, experimentally installed both at the K.U. Leuven and in industry, is a
PV powered water pump system, consisting of a PV array, a low cost inverter, a
permanent magnet synchronous motor (PMSM) and a water pump with water
storage. The PV array has a peak power of 4,32 kW. To avoid a power supply by an
electric grid, the system has been set up to work independently in island operation.
All control and measurement units are supplied by the dc bus between inverter and
PV array. A block diagram of the pilot installation is shown in figure 6.1. The
inverter operates as a variable frequency source (PWM) for the PMSM driving the
pump. Since a PMSM in open loop is unstable (see section 7.4.1), a fieldoriented
control with feedback of speed and position is proposed. However, a mechanical
speed/position sensor has several drawbacks from the viewpoint of drive cost,
reliability and signal noise immunity. Especially in submergedmotor/pump systems,
an installation of the additionally sensor is problematic or even impossible. Here,
speed and field position are estimated by an extended Kalman filter described in
chapter 6.
water
storage pump inverter
Solar
generator
PMSM
MPP
Tracking
Current [A]
Voltage [V]
Phase current
for EKF
DC bus voltage/current
measurement for MPPT
control
prototyping
PWM
Figure 6.1: Block diagram of the PVpowered water pump systems.
RideThrough at Power Interruptions 107
Furthermore, the control system is equipped with a MPPT and a voltage control
guaranteeing a balanced input/output power ratio in the dc bus. The MPPT and
voltage control are realized by feeding back the dc bus voltage and current to the
controller. Additionally, measurements of motor currents in two phases are required
for speed estimation and torque control of the PMSM. The motor voltages required
for the EKF are not measured, but calculated from the reference voltages
determining the PWM output of the controller.
A computeraided control system is used as a developing platform monitoring and
recording the experimental data. Provisionally, the entire control algorithm, safety
related monitoring and the startup and shut down automatism’s are implemented on
a TMS320C31 DSP. The I/O subsystems and the PWM generation are based on
TMS320P14 working as a slaveDSP. However, the final algorithms are intended to
be implemented in a simple microcontroller ensuring an overall low cost system.
6.3 PV Array
The voltagecurrent characteristic of one PV element at constant cell temperature
and with the irradiance of the sunlight as a parameter is shown in figure 6.2. In the
same figure, also the output power of the PV element is drawn as a function of the
module voltage. The power P
PV
is calculated by the product of dc bus voltage U
dc
and current I
PV
. The Maximum Power Point (MPP) is characterized by the voltage,
where the PV array generates maximum output power.
0 5 10 15 20 25
0
1
2
3
4
5
6
7
8
9
U
dc
[V]
I
P
V
[
A
]
15
30
45
60
75
90
105
120
135
P
P
V
[
W
]
current
MPP
power
Figure 6.2: Characteristic of one PV element at constant cell temperature.
The PV element characteristics are a function of the irradiance of the sunlight and
the cell temperature. In figure6.2, six different levels of insolation are illustrated.
High current curves correspond to high insolation levels while low current curves
correspond to lower insolation levels. With increasing irradiance, the MPP moves
108 Chapter 8
along the marked line. In order to stay at the point of maximum power at rising
irradiance, the current in the dc bus must be increased, while the dc voltage remains
nearly constant.
The voltage at the MPP changes with the array temperature while the current is
almost unaffected. At lower cell temperature, the MPP characteristic is situated in a
higher voltage range. The voltage temperature coefficient of the PV elements used
amounts to –82 mV/°C. Therefore, connecting 12 modules in series and a
temperature variation of 10 K results in an optimum voltage shift of 9,84 V on a
rated voltage of 180 V, i.e. ±5 %. Thus, the optimum output voltage of the PV array
is not constant and moves as condition varies.
The practically studied PV array consists of 36 modules with a total peak power of
4,32 kW. All wires of the single PV elements are assembled in a modular way using
a switchboard panel, connected in series or parallel. The different experimental
connections are presented in Table 6.1. The experimental results obtained are similar
demonstrating the high flexibility of both the previously and later described control
algorithms. However, to match the requirements of the final inverter and to keep the
presented results clear, all experimental results presented in this chapter are
measurements using 12 modules in series and 3 modules in parallel.
Table 6.1: Various connections of the PV array with a peak power of 4,32 kW.
Number of
modules in
series
Number of
modules in
parallel
I
max
[A]
U
max
[V]
6 6 44,7 129
9 4 29,8 193,5
12 3 22,35 258
18 2 14,9 387
6.4 Motor/Pump Subsystem
Surface applications for irrigation systems are mostly driven by dc machines while
for installations in the drilling holes submersible induction motor/pump systems are
used. Commutator motors have very desirable control characteristics, but they are
not applicable for submersible installations. Furthermore, their use is limited by a
number of factors [Bose 97]:
• Need for regular maintenance of the commutator;
• Relatively heavy rotor with a high inertia;
• Difficulty in producing a totally enclosed motor as required for some
hazardous (e.g. submersible motor/pump system) applications;
• Relatively high cost;
A pumping system based on an ac motor drive is an attractive alternative where
reliability and maintenancefree operation is important [Bhat 87]. However, small
RideThrough at Power Interruptions 109
induction motors have, when compared to permanent magnet motors, a lower
efficiency especially at partial load. Thus, motor selection and design theory
[Hen 96] were limited to a permanent magnet synchronous motor (PMSM) coupled
to centrifugal or submersible pumps.
Pumping pure drinking water is mainly done by a submersible combination of motor
and pump. Due to this hazardous application, placing additionally sensors, e.g.
encoder for speed and position measurement, is costly and problematic or even
impossible. Therefore, the first approach of the PV powered water pump system was
an open loop control of a PMSM with a damper cage (figure 6.3). Rotor bars have
been implemented in order start up the motor and to balance disturbances.
Copper bars
magnets
Figure 6.3: Cross sectional view of the 4pole PMSM rotor geometry.
A centrifugal pump commonly requires a single quadrant drive. The load torque of
the centrifugal pump expressed as a function of speed is
2
ω
ω
K T
load
= (6.1)
where K
ω
is the constant of the hydraulic system. Thus, to vary the output of the
water pump, the speed of the motor driving the pump must be varied. The properties
of the PMSM used are summarized in appendix B.3.
6.4.1 Open loop control of a PMSM with damper cage
The open loop control of a PMSM is based on a constant voltagefrequency ratio.
This U/f ratio is predetermined for every (steady state) motor speed, choosing a
voltage level corresponding to the lowest motor current. The voltagecurrent
characteristic at different load torque levels is presented for the studied PMSM in
figure 6.4. The calculations are made at a frequency of 10 Hz corresponding to a
motor speed of n = 300 rpm. If all motor parameters and the load characteristic are
known, the optimal U/f ratio can be calculated for every motor speed. The current is
settled automatically depending on the difference between induced (EMK) and
supply voltage. However, a direct control of the current is impossible proposing this
approach.
Considering figure 6.4, a slight variation of the voltage can easily lead to a very high
overcurrent of the motor with the risk of demagnetization. In fact, an erroneous
110 Chapter 8
calculation of the optimum voltage cannot be avoided due to parameter and load
torque variations as well as measurement errors and the nonlinearity of the inverter.
10 11 12 13 14 15 16 17
0
5
10
15
20
25
U [V]
I
[
A
]
Torque
[0>10Nm]
Figure 6.4: Voltagecurrent characteristic at variable torque (f = 10 Hz).
Due to deadtime effects, the error between reference and output voltage of the
inverter used amounts to ∆U ≈ 15 V resulting unaccompaniedly in a current
variation of 22 A for the studied PMSM at a speed of 300 rpm. According to (1.16)
(1.17), the voltage error ∆U depends on PWM frequency, dc bus voltage, dead time
and current direction.
This nonlinear effect creates a distortion of motor current and torque. Without a
damper cage, the PMSM in open loop is an undamped, oscillating system [Mel 91],
[Hen 91]. Slight variations of the electrical angle
ϑ ϑ ϑ ∆ + =
0
(6.2)
result in a selfexited oscillation with an undamped natural frequency f
e
:
0 cos
0 max 2
2
= ∆ +
∆
ϑ ϑ
ϑ
T
dt
d
p
J
(6.3)
) 2 sin( t f
e
π ϑ = ∆ (6.4)
p J
T
f
e
0 max
cos
2
1 ϑ
π
= ⇒ (6.5)
The frequency f
e
of the PMSM used varies between 0 Hz < f
e
<10 Hz for maximum
load (ϑ
0
= 90°) and no load (ϑ
0
= 0°) respectively. This oscillation creates a
RideThrough at Power Interruptions 111
distortion of current and torque generated by the motor, leading to an
electromagnetic instability of the drive. In order to soften the system, the PMSM
should be equipped with a damper cage [Hen 91]. The resulting damper time
constant T
damp
can be calculated using:
2
1
'
2 2
1
2
1
2
) 1 (
3
2
σ
ω
+ = R
U p
J
T
damp
(6.6)
The modified PMSM used has a damper time constant of T
damp
≈ 60 ms. Considering
the natural frequency, this value is far too large. For a sufficiently damped system,
the time constant and thus the rotor resistance must be smaller. Therefore, a stable
and dynamic openloop control of the given PMSM is impossible.
A modification of the PMSM geometry, including a completely closed damper cage
[Con 86] and inset magnets, could improve the performance. Nevertheless, it is
much better to consider another control approach. As explained in the next sections,
a high dynamic drive is indispensably for a correct operation of the water pump
system. In fact, this has been the starting motivation for the development of the
earlier described highperformance motor drive with speed, flux and torque
estimation.
6.5 Control of a PVPowered Water Pump System
Most PV powered water pump systems consist of two different control units. The
first is a dc unit with or without a battery as energy storage. In this unit, the MPPT is
controlled by varying the duty ratio of a dcdc converter. Using this converter leads
to a less complicated control algorithm for the MPPT. Varying the dc bus voltage
can be done more quickly and without changing the power or frequency of the
motor. The influence of a changing irradiance level during a searching procedure is
reduced. On the other hand, this converter introduces many losses, amongst others:
• Switching losses
• Valve losses
• Copper and iron losses in the filter coil
In control systems without a battery, the dc bus may collapse when an unbalanced
input/output power ratio occurs at the dc bus. Therefore, systems without battery
require a more complex and complicated control algorithm. The second unit controls
the speed of motor and pump.
Here, a novel control approach is proposed avoiding the use of the additional dcdc
converter, a battery and its losses. The overall control of the PVpowered water
pump system consists of a current/torque controller, a speed controller without a
shaft sensor and a main control consisting of a dc bus voltage controller and the
112 Chapter 8
MPPT. The voltage control varies the speed/torque of the PMSM in order to stay
within the calculated optimum voltage given by the MPPT. The structure of the
overall control system is shown in figure 6.5.
STARTUP
&
MPPT
1
Torque
Control
&
EKF
SVM
Inverter
i
a
u
b
*
u
c
*
AC
motor
i
b
pump
u
a
*
ω
ω
*
T
el
*
T
el
*
ω<ωmax
T
el
*
U
dc
U
dc
*
I
PV
U
dc
I
PV
PI with
anti windup
PID with
anti windup
PI with
anti windup
PV power
supply
Voltage
control
Figure 6.5: Block diagram of the entire control system.
In contrast to the very quickly and frequently changing irradiance intensity, the cell
temperature of the PV array and thus the dc voltage in the MPP varies very slowly.
Therefore, the control of the PVpowered water pump system is performed by
varying the dc voltage in a small range, searching the MPP and controlling the
speed/torque characteristic of the motor in order to stay within the calculated
optimal voltage corresponding to the highest efficiency of the system. The adaptive
MPPT algorithm is described in detail in sections 7.7.
The dc bus voltage can be controlled either by the speed of the motor requiring an
additional speed control loop or directly by the electromagnetic torque affecting the
motor speed derivation. The different control approaches are indicated by the switch
in figure 6.5. Advantages and drawbacks of both control approaches are explained in
the next section. The first approach turned out to malfunction, when tested under
extreme but realistic conditions. However, many algorithms described in literature
are based on the variation of the speed reference; e.g.: [Mul 97] varies the power by
changing the frequency output of the inverter stepwise, being even slower than using
an extra speed control loop.
The highperformance speed/torque control of the PMSM is described in previous
chapters. The mechanical position sensor is replaced by an observer requiring no
additional measurements. Only measurements of motor current and dc bus voltage
are necessary. Figure 6.6 shows the experimental results of a speed step with a
centrifugal pump as load and using a regular grid as power supply (U
dc
= 220V).
With sufficient power generated by a PV array, the results obtained would be the
same. Otherwise, the dc bus would be discharged and the system collapses. The
applied load at n = 2000 rpm amounts 85% of the rated motor torque. The
current/torque controller has a bandwidth of 960 Hz. According to subsection 2.5.1,
RideThrough at Power Interruptions 113
the optimal torque control of the motor takes advantage of the reluctance torque by
introducing a negative (L
d
< L
q
) directaxis current component increasing the
efficiency of the drive.
0 0.25 0.5 0.75 1 1.25 1.5
0
500
1000
1500
2000
2500
t [s]
n
[
r
p
m
]
Measurement
Reference
0 0.25 0.5 0.75 1 1.25 1.5
10
0
10
20
30
t [s]
i
[
A
]
iq
id
Figure 6.6: Speed step with a pump load using a regular grid as power supply.
Top: Measured and reference speed. Bottom: d and qaxis current.
The startup and shut down logic is based on the motor speed and the opencircuit
voltage. Below a speed of n ≈ 180 rpm, the centrifugal pump used is not able to
pump water. A nonproductive idle run is not conducive for the durability of the
pump and all other wear. Therefore, the system goes in standby modus and all PWM
pulses are disabled if the PV power supply is definitively too low. The energy
consumption of all control and measurement units and the inverter in standby modus
amounts to 20 W. A startup procedure requires a predetermined minimum open
circuit voltage guaranteeing a productive motor speed, being higher than the
minimum speed of the shut down automatism.
The implemented safetyrelated monitoring consists of detecting overcurrent and
overspeed, both depending on the motor/pump system, and a predetermined
voltage window mainly defined by the PV array coupled to the inverter. An
inadmissible failure disables the entire system, requiring a manual reset by an
expert. However, no such failure has been detected during weeks of testing.
6.6 DC Bus Voltage Control
114 Chapter 8
6.6.1 Experimental results of the voltage control
In figure 6.7 the response of the voltage, qaxis current and speed for a step of the
voltage reference from 225 to 125 V and back to 225 V is shown. It can be seen, that
the voltage and the torque producing qaxis current are controlled very fast. They are
already in steady state, while the speed still varies. The speed of the motor changes
indirectly, controlled by the electromagnetic torque until the reference voltage is
reached.
0 0.2 0.4 0.6 0.8 1
100
150
200
250
t [s]
U
d
c
[
V
]
0 0.2 0.4 0.6 0.8 1
30
15
0
15
30
t [s]
i
q
[
A
]
0 0.2 0.4 0.6 0.8 1
600
700
800
900
t [s]
n
[
r
p
m
]
Reference
Measurement
Figure 6.7: Step of the voltage reference from 225V to 125V and back to225V.
Top: dc bus voltage. Middle: qaxis current. Bottom: Motor speed.
The MPP of the PV array used is situated between 185 and 195 V. Without voltage
control the voltage area below the MPP (U
dc
< 185 V) is unstable. Slightly
increasing the motor speed in this unstable area, results in a very quick discharge of
the capacitor leading inevitably to a crash of the entire system. The averaged voltage
error of this inner control loop is smaller than 0.1%. Even at the starting procedure
and under very quickly changing irradiance, the error reaches a maximum of 0.5%.
The averaged voltage error delivers the minimum search range for the later
described main control (MPPT) providing the reference voltage.
An important feature of the system is the independence of the pipe characteristic.
The pumping head or water pressure can be varied by a throttle lever
increasing/decreasing both water pressure and reversely water flow. Thus, also the
pipe characteristic changes. Figure 6.8 demonstrates this independence by varying
the pumping head from a ½ m (=½ bar) to 10 m (=10 bar) and back to ½ m. The
RideThrough at Power Interruptions 115
response of the voltage control is plotted above, below the speed. The reaction time
of the voltage control is quite slow due to the time intensive valve closure. The
experiments are made with a reference voltage of U
dc
= 191V approximately
agreeing with the MPP of the PV array.
0 1 2 3 4 5
180
185
190
195
200
t [s]
U
d
c
[
V
]
0 1 2 3 4 5
1100
1200
1300
1400
1500
t [s]
n
[
r
p
m
]
0 1 2 3 4 5
180
185
190
195
200
t [s]
U
d
c
[
V
]
0 1 2 3 4 5
1100
1200
1300
1400
1500
t [s]
n
[
r
p
m
]
Reference
Figure 6.8: Variation of the pumping head.
Left: ½ m ⇒ 10 m. Right: 10 m ⇒ ½ m.
One of the most important features of the voltage control is its robustness during
power interruptions, occurring at instantaneous decrease of irradiance (e.g. passing
clouds). This property has been tested using a regular grid as power supply and
applying a complete power interruption on all three phases for a short time agreeing
with the worstcase condition of the system. These tests were done with a smaller,
3 kW prototype PMSM. The implemented regenerative braking scheme allows the
inverter to keep its dc bus voltage at the predetermined minimum level, expanding
the time in which supply voltage can be reapplied without the timeconsuming dc
link capacitor recharging cycle. The experimental results of such short time three
phase power interruptions are shown in section 8.5.
6.7 Maximum Power Point Tracking (MPPT)
The voltage reference of the dc bus voltage control is calculated by an overlaid
maximum power point tracking (MPPT). Compared to a common voltage tracking,
the efficiency of the output power of a PV array can be increased about 2 % by
MPPTracking. The MPP is characterized by the voltage, where the PV array
generates maximum output power.
The main problems of matching the MPP with a PV array as power supply are
related to the nonlinear, solar irradiance and cell temperaturedependent voltage and
current characteristics of the PV array. The characteristics are affected by the
contamination, the sunlight irradiance and the cell temperature. To reach the MPP at
rising irradiance level, the current in the dc bus must be increased while the dc bus
116 Chapter 8
voltage remains nearly constant. The voltage at the MPP changes with array
temperature and current is almost unaffected. At lower cell temperature, the MPP
characteristic is situated in a higher voltage range. Thus, the optimum output voltage
of the PV array is not constant and moves as conditions vary.
The MPPT is the main control loop, calculating the MPP and the search range of the
dc bus voltage, and delivers a reference quantity to the voltage control loop. First, a
default voltage and search range is given. After the default voltage is reached, it is
varied slightly around this point. The quantity of this variation is given by the search
range. During this variation, the power generated by the PV array is measured and
the voltage linked to the maximum power is stored during the respective searching
procedure. The new optimum voltage and the new search range are calculated from
these actual measurements and in its stored values by an adaptive control algorithm.
With these new quantities, the controller starts again a searching procedure to find
the MPP. Figure 6.9 shows the flow chart of the MPPTracking.
Uref < Uopt  ∆U
Decrease Uref
Calculate Uopt
• Optimum voltage: Uopt
• Search range: ∆U
• Search speed: dU/dt
Measurement
Reference
voltage U
ref
Startup with
initial values
NO
NO
YES
YES
Uref
Increase Uref
Uref > Uopt + ∆U
Calculate ∆U
Uref
Uref
t
Figure 6.9: Flow chart of the MPPTracking.
The previous values are very important for the calculation of the new optimum
voltage and search range. If, e.g., the new calculated optimum voltage during a
searching procedure with rising voltage is situated higher than the last optimum
voltage, the MPPvoltage seems to change. However, this can also indicate an
increasing irradiance. If the second condition occurs, the controller should not
change the new optimum voltage. Otherwise, the calculated voltage drifts from the
MPP. The same considerations are also valid for decreasing irradiance. Thus, the
adaptive control must be able to distinguish between a changing MPP and changing
conditions. A new optimum voltage is only calculated, when a tendency is indicated
by a searching procedure with both rising and falling voltage reference.
RideThrough at Power Interruptions 117
Pictorially expressed, the shape of the reference voltage looks like the movement
playing an accordion. Both, search range and speed depends on the variation of the
calculated optimum voltage. If a new calculated MPP is situated in the half of the
past voltage range, the search range and the search speed are reduced, otherwise
both are increased. If the irradiance power changes very often and too fast to track
the real MPP, the MPPT algorithm behaves as a common constant voltage tracking.
6.8 Experimental Results
Figures 6.106.12 demonstrate the startup procedure and the automatic operation
mode of the entire control system consisting of MPPT, voltage control and torque
control. Figure 6.10 shows the power generated by the PV array and the speed of the
motor driving the pump during 5 min of MPPT, while figure 6.11 exemplifies the dc
bus voltage for the same span of time. The characteristic of figure 6.12 indicated by
“Start 1” shows the mentioned power as a function of the dc voltage.
0 50 100 150 200 250 300
0
1
2
3
4
t [s]
P
P
V
[
k
W
]
0 50 100 150 200 250 300
0
500
1000
1500
2000
t [s]
n
[
r
p
m
]
Figure 6.10: Startup procedure and MPPTracking.
Top: PV power. Bottom: Motor/pump speed.
Starting with the opencircuit dc voltage, in which the power generated by the PV
array supplies only the control and measurement electronics (~20 W), the voltage
decreases in voltage control mode to a predetermined reference value. The MPPT is
switched on after reaching this operation point and searches subsequently for the
optimum voltage, where the PV array generates maximum power. The MPP is
reached after approximately 15 s and the voltage is varied from now on slightly
around this point.
During the MPPT, the power generated by the PV array is measured and the voltage,
linked to the maximum power, is calculated. The implemented system tracks
automatically the present conditions; e.g. with increasing insolation, the optimum
voltage is situated in a higher voltage range. As can be seen from the details of
figure 6.11, the variation of the voltage depends on external influences as insolation
118 Chapter 8
or temperature variations and is adjusted automatically. In steady state, the variation
of reference voltage is very small and almost constant.
0 50 100 150 200 250 300
180
200
220
240
t [s]
U
d
c
[
V
]
265 270 275 280 285 290 295
190
191
192
193
194
195
t [s]
U
d
c
[
V
]
Box 1
U
dc
U
dc,ref
Figure 6.11: Startup procedure and MPPTracking.
Top: dc bus voltage. Bottom: Details indicated by Box 1.
Due to the lack of storage element in the dc bus, the power of the PV array must be
used immediately to accelerate the PMSM. Measurements during a sunny day of the
implemented MPPT are plotted in figure 6.12 showing the power of the PV array as
a function of the dc voltage U
dc
.
Three different starting conditions are shown to demonstrate the ability of
reproduction of the MPPT. The direction of the searching procedure is indicated by
the arrows. The characteristic indicated by ‘Start 1’ refers to the time exposure of
figures 6.106.11. The artificial starting point exhibits the MPPT starting in an
unstable area, where slightly increasing the motor speed results in a very quick
discharge of the capacitor. This starting point is reached by first using the voltage
control mode with a reference U
dc
*
= 120V and then switching over to the MPPT
control mode. However, this artificial starting point is never reached during regular
operation. The measured results demonstrate the ability of reproduction as well as
the stability of the entire system.
RideThrough at Power Interruptions 119
100 120 140 160 180 200 220 240
0
0.5
1
1.5
2
2.5
3
3.5
P
P
V
[
k
W
]
U
dc
[V]
MPP
Start 2
Artificial
starting point
Start 1
Figure 6.12: Measured results of the MPPTracking using a PV array with a peak
power of 4,32 kW and a PMSM driving a centrifugal pump (sunny day).
Instead of a permanent magnet synchronous motor, the implemented MPPT and
voltage control are also suitable for an induction motor driving the pump. The
performance of induction motor and PMSM are similar, only the induction motor
efficiency is lower especially at partial load. Here, in contrast to all other earlier
presented results, a PV array with a peak power of 1,2 kW and a 1,5 kW induction
motor driving the centrifugal pump was used. The MPPT during a cloudy day, with
this second installation, is presented in figure 6.13. Four different starting conditions
(ad) are shown. Characteristic ‘c’ and ‘d’ starts in an artificial operation point. The
characteristic ‘a’ is situated in a higher voltage range, because it shows the first
searching procedure at a lower cell temperature.
120 130 140 150 160 170 180 190 200 210
0
50
100
150
200
250
P
P
V
[
W
]
U
dc
[V]
MPP
d
c
b
a
Figure 6.13: Measured results of the MPPTracking using a PV array with a peak
power of 1,2 kW and an induction motor driving the pump (6 hours of a cloudy day).
120 Chapter 8
6.9 Conclusions
This chapter presents the design and the implementation for PVpowered water
pump systems using a PMSM without a shaft sensor. In a first approach, the
performance of a PMSM with damper cage in open loop control is evaluated. Due to
the insufficient damper cage, a stable and dynamic openloop control of the given
PMSM is impossible. Furthermore, a simple U/f control is absolutely inferior
compared to a fieldoriented high performance motor drive with speed, flux and
torque estimation. The price to be paid is a more extensive and more complicated
control algorithm. However, no additionally measurements are required.
Additionally, an increased efficiency of the entire system is achieved.
Due to the lack of storage in the dc bus, the power of the PV array is used
immediately to accelerate the motor. Practical investigations are done to
demonstrate the stability of the dc bus voltage control, the independence of the
pump characteristic and its robustness during power interruptions. To optimize the
energy captured by the PV array and to pump as much water as possible, the output
power should always be at its maximum power point. Increasing the efficiency of
the system is very advantageously considering the costintensive PV array
installation amounting to 70% of the total system costs. Therefore, a novel MPPT
algorithm, realized by feeding back the dc voltage and current to a controller, has
been developed and implemented. The measured results of the MPPT exhibit the
ability of reproduction and the stability of the entire control system.
The development of a second prototype PMSM for submersible applications has
been stopped for practical and economic reasons. The design constraints due to the
mechanical construction of the motor housing and stator iron, together with the
filling of the motor interior with water yield a very unusual mechanical rotor
construction. The classical rotor design of a permanent magnet synchronous
machine with surface mounted magnets would inherit the fixing of the magnets with
glue and a polymer bandaging. The longterm stability for use directly in water of
both permanent magnets and bandage cannot be guaranteed by the manufacturer.
This applies even for coated magnets. The production costs of such a machine are
enormous. Due to the design constraints, only a marginal efficiency increase can be
expected by replacing the induction motor with the PMSM. However, the
implemented MPPT and voltage control are suitable for both a PMSM and an
induction motor driving the pump.
In the meantime, a request for the installation of the presented PV powered water
pump system has been received from four different countries: Mali, Mauritania,
Senegal and Chad.
7. Conclusions
7.1 Summary & Conclusions
This thesis concerned highperformance motion control systems with a voltage
source inverter supplying both squirrelcage induction motors and permanent
magnet synchronous motors. Basic control techniques, that allow dynamic torque
and flux control in a decoupled way, are direct torque control (DTC) and field
oriented control (FOC). Summarizing, the DTC provides a better dynamic torque
response whereas the FOC provides a better steadystate behavior. With respect to
the planned applications where the motor speed is the main control variable, the
FOC has been chosen as final control scheme.
Considering fieldoriented control, it is possible to control separately the flux and
torque producing components of the supply currents. In particular, the concept of
field orientation and the resulting ability to directly control the electromagnetic
torque were discussed in chapter 2. Torque control, which constitutes the most basic
motor control function, maps very directly into current control because of the close
association between current and torque generation in any PMSM and induction
motor drive.
There are many excellent books on the topics of electrical machines and drives.
However, it is believed that the present thesis is novel in many respects. The basic
FOCscheme is refined systematically adding additional features step by step. Flux
weakening is widely known in literature. Less appreciated is the ability to operate
the induction motor above the nominal flux at low speed to enhance the torque per
ampere relation and thus better utilize the available power supply current. The
approach has been further refined by flux optimization. Contrary to the assertions in
literature, this feature makes the induction motor superior compared to the PMSM in
a wide operation range when efficiency is considered especially in the range where
iron losses are dominant. The choice of a suitable flux control strategy depends on
the respective application. In the realized implementations, it can be switched over
easily and in realtime to different strategies. Considering flux weakening of the
PMSM, the algorithms presented in literature are based on precalculations
postulating a constant dc bus voltage. In this work, the dc bus voltage is variable
over a wide range requiring an alternative approach. Therefore, an automatic flux
adaptation scheme has been implemented.
122 Chapter 9
Whereas antiwindup systems are well known in literature, an antiwindup system
within the current controller is neglected since the maximum voltage is limited by
the power inverter itself. However, the presented current control with antiwindup is
essential considering the dc bus voltage variable over a wide range.
A commercially available DSP based environment is used for development purpose.
However, the support software has been changed in order to implement different
PWM strategies as well as variable PWM frequencies. This is extremely valuable
during the development of highperformance motor control using PWM outputs in
order to drive power switches. Chapter 3 presents the collaboration between control
design and realtime implementation. The DSP controller board, code generation,
experiment management and hardware interface including required measurements
are explained. Issues of measurement distortion/identification due to the inverter
nonlinearity are discussed in detail.
Chapter 4 exhibits a new approach of speed estimation employing an incremental
encoder as measurement device. Among the speed, also rotor position and the
acceleration of the drive are estimated. The implemented algorithm is based on a
linear Kalman Filter. The discussion extends to the implementation of an advanced
speed control loop. It has been shown that this approach offers a significant
improvement of the entire drive performance. This chapter can be also regarded as a
smart introduction into observer theory. Advanced observer theory has been applied
to approaches eliminating the need of position/speed measurement.
The sensorless speed control of both permanent magnet synchronous motor and
squirrelcage induction motor drives, which is nowadays the most attractive research
area of electrical motor drives, is presented in chapter 5 and 6. New models for
speed estimation are proposed. The structures of the implemented sensorless control
schemes are based on the extended Kalman filter theory. The approach requires no
additional measurements. The terminal voltages are not measured; they are
reconstructed by using the monitored dc bus voltage and the switching functions of
the inverter considering nonlinearities due to the dead time of the power switches.
Among the speed, also rotor flux, flux position and the acceleration of the drive are
estimated. The speed estimation does not lag the actual motor speed, both in steady
state and during acceleration/braking. Compared to sensorless control schemes
described in literature, the experimental results have shown to offer a significant
improvement of the drive performance.
Within this thesis, different control approaches considering respective applications
were developed and implemented. Special care has been taken for the viability of the
realtime implementation: A comprehensive and clear description of controller
design and affiliated parameter calculation is given for all treated applications.
Furthermore, all proposed control schemes were verified by experimental results.
The implementation of a PVpowered water pump system using a PMSM without a
shaft sensor is described in chapter 7. This system reflects one application
employing many of the drive features, designed and implemented in this work. New
Conclusions 123
approaches were developed because the algorithms described in literature for this
kind of systems turned out to malfunction. A novel maximum power point tracking
optimizing the energy captured by the PV array has been designed and implemented.
The PVpowered water pump system consists, among other control loops, of a high
performance dc bus voltage control, which constitutes the most important control
function guaranteeing the stability of the drive.
As discussed in chapter 8, the realized dc bus voltage control is also applicable for
ridethrough schemes at power interruptions considering invertercontrolled drives
supplied by a regular grid. The proposed solution to the problem is to recover some
of the mechanical energy stored in the rotating masses by kinetic buffering. This
maintains the dc link capacitor well charged keeping the electronic control circuits
active. The temporary speed dip is generally tolerable, since the most frequent
power interruptions last only for a few milliseconds. Furthermore, the proposed ride
through scheme at power interruptions has been transformed into a special drive
braking tool saving energy and simplifying the inverter setup: the installation of
brakeresistance, power switch and cooler may be eliminated.
7.2 Further Research
Using the TMS320C31 DSP providing 60 MFlops, the proposed algorithms have
been realized only by means of costly code optimizations. The limit of the possible
code and memory size has been reached. The calculation of the closed loop current
transfer function has shown the large influence of delays within the loop as e.g.:
measurement filter, sample time, PWM frequency and signal lag of data
transmission. Further increasing the program size will lead to execution times,
which are no longer suitable for highperformance motion control.
In particular, it is interesting to implement the proposed algorithms on faster DSPs.
It is expected that the dynamic performance, especially of the torque control loop,
can be vastly increased. Presently, a new DSP development platform based on TI’s
most recent processorgeneration, the TMS320C6711 DSP providing 1000 MFlops,
is under construction within the ELECTA group. This new development platform
provides a system, which will be capable of implementing even more extended and
computation time intensive algorithms.
Considering a more powerful control system, there are many applications possible,
e.g.: The noise covariance matrices within the mentioned sensorless speed control
system can be adapted in realtime dependent on the given operating point. This can
be done by e.g. another extended Kalman filter or artificial intelligence.
Various reluctance motors will have an increased role in the future. An expansion of
the proposed sensorless control schemes to these motor types forms surely an
interesting task.
124 Chapter 9
The proposed observer together with advanced control techniques can be applied to
the active filter (active frontend) design, which forms nowadays an interesting field
in the area of power quality. Especially a disturbance (current harmonics) rejection
approach, similar to the proposed load torque rejection approach within the
speed/current control loop of drives, promises a vastly increased performance.
Classical control theory suffers from some limitations due to nonlinearity, time
invariance etc. of the controlled system. These problems can be overcome by using
artificialintelligencebased control techniques. In literature, e.g. [Vas 99], it is
expected that intelligent sensorless instantaneous torquecontrolled drives
incorporating some form of intelligence will become the standard in the future.
These drives will not require machine or controller parameters, and all the control
and estimation tasks are performed by a single artificialintelligencebased system.