1894 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO.

6, JUNE 2009

Model Predictive Direct Torque Control—Part I:

Concept, Algorithm, and Analysis
Tobias Geyer, Member, IEEE, Georgios Papafotiou, Member, IEEE, and Manfred Morari, Fellow, IEEE

Abstract—This paper focuses on direct torque control (DTC) for need to be kept within prespecified hysteresis bounds. In high-
three-phase ac electric drives. A novel model predictive control power applications, where often neutral-point-clamped (three-
scheme is proposed that keeps the motor torque, the stator flux, level) inverters with integrated gate commutated thyristors are
and (if present) the inverter’s neutral point potential within given
hysteresis bounds while minimizing the switching frequency of used, the control objectives are extended to the inverter and
the inverter. Based on an internal model of the drive, the con- include the balancing of the inverter’s neutral point.
troller predicts several future switch transitions, extrapolates the In this paper, we present a DTC scheme that yields a signif-
output trajectories, and chooses the sequence of inverter switch icant performance improvement with respect to the state of the
positions (voltage vectors) that minimizes the switching frequency. art without requiring excessive computational power to enable
The advantages of the proposed controller are twofold. First,
as underlined by the experimental results in the second part of the implementation on (already existing) DTC hardware. Such
this paper, it yields a superior performance with respect to the a control scheme can be derived by adopting the principles
industrial state of the art. Specifically, the switching frequency is of constrained optimal control with a receding horizon pol-
reduced by up to 50% while the torque and flux are kept more icy, i.e., model predictive control (MPC) [5]. Specifically, the
accurately within their bounds. Moreover, the fast dynamic torque hysteresis bounds are inherited from standard DTC, whereas
response is inherited from standard DTC. Second, the scheme is
applicable to a large class of (three-phase) ac electric machines the DTC switching table is replaced by the following online
driven by inverters. optimization. Over a short switching horizon, all admissible
switching sequences are considered. Based on the measured (or
Index Terms—AC motor drives, direct torque control (DTC),
model predictive control (MPC), power electronics. estimated) machine quantities, for each switching sequence, the
evolution of the torque, stator flux, and neutral point potential is
predicted, using a nonlinear discrete-time model of the drive. To
I. I NTRODUCTION emulate a long output horizon, the predicted trajectories of the
torque, flux, and neutral point potential are extrapolated, and the
I N ADJUSTABLE-SPEED ac drives, dc–ac inverters are
used to drive three-phase ac machines as variable-frequency
voltage or current sources. One of the various methods used
number of time steps is determined for which these quantities
are kept within their hysteresis bounds. For each switching
sequence, an approximation of the average switching frequency
for controlling the machine’s torque and speed is direct torque
is computed that is given by the total number of switch tran-
control (DTC) [1]–[4]. Exploiting the motor’s fast stator flux
sitions in the sequence divided by the time duration of the
dynamics, DTC directly manipulates the stator flux vector
extrapolated trajectory. Minimizing the switching frequency
such that the desired torque (and magnitude of the stator flux)
over all switching sequences yields the optimal sequence of
is achieved by choosing an inverter switch combination that
switch transitions. Of this sequence, only the first step is applied
provides the appropriate phase voltages to the motor windings.
to the drive. At the next sampling instant, this procedure is
In state of the art drives, this choice is made, e.g., every Ts =
repeated with new measurements thus establishing a receding
25 μs using a predesigned switching table that—depending on
horizon policy, as this method is commonly referred to.
the particularities of the application—addresses a number of
This control approach carries several important advantages.
different control objectives. These primarily concern the motor.
Introducing additional control objectives (like the balancing of
More specifically, the electromagnetic torque and the stator flux
the switching power losses) is straightforward. As all computa-
tions are performed online, all quantities may be time varying
Manuscript received April 15, 2008; revised September 9, 2008. First pub- including model parameters, set points, and hysteresis bounds.
lished October 31, 2008; current version published June 3, 2009. This research
was performed at the Automatic Control Laboratory, ETH Zurich, Switzerland.
Even more importantly, the controller can be directly applied
This work was supported in part by ABB Switzerland Ltd. and in part by to a large class of three-phase ac drive. More specifically, in-
the two European Commission research projects IST-2001-33520 [Control and duction machines (both squirrel-cage and ring-rotor type), syn-
Computation (CC)] and FP6-IST-511368 [Hybrid Control (HYCON)].
T. Geyer is currently with the Department of Electrical and Computer
chronous and permanent-magnet machines can be addressed,
Engineering, The University of Auckland, Auckland 1020, New Zealand as well as inverter topologies such as two-, three-, or five-
(e-mail: [email protected]). level inverter. Yet, to simplify the exposition of the new control
G. Papafotiou is now with the ABB Corporate Research Center, 5405
Baden-Dättwil, Switzerland (e-mail: [email protected]). scheme, we focus in this paper on a specific application, where
M. Morari is with the Automatic Control Laboratory, ETH Zurich, 8092 a neutral-point-clamped (three-level) voltage-source inverter
Zurich, Switzerland (e-mail: [email protected]). drives a squirrel-cage induction motor (IM).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. The major benefit of this controller, however, is its superior
Digital Object Identifier 10.1109/TIE.2008.2007030 performance in terms of the switching frequency. For ABB’s

0278-0046/$25.00 © 2009 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1895

well-established three-level voltage-source inverter ACS 6000 their references. There are no hysteresis bounds on these
[6] with a squirrel-cage IM, the proposed model predictive DTC variables.
(MPDTC) scheme reduces the switching frequency over the 3) Except for [16], [17], and [20]–[22], a modulator is in-
whole range of operating points by up to 50%, with an average cluded in the control loop greatly simplifying the control
reduction of 25%, while better respecting the torque and flux problem. Thus, unlike the DTC philosophy, the switch
hysteresis bounds. Notably, this result is independent from the positions are not directly chosen by the controller.
power rating of the machine, which was varied from a few 4) Only two-level voltage-source inverters are considered
kilovoltamperes to several megavoltamperes. All performance with [12] and [22] being an exception.
evaluations were carried out by applying the control scheme 5) The cost function does not emulate the switching fre-
to a very accurate and detailed Matlab/Simulink model of the quency of the inverter. In combination with reference
drive, which was provided by ABB to very closely resemble tracking and the limited number of voltage vectors (for
a physical drive and ensure a simulation setup that is as re- a two-level inverter), a high switching frequency is to be
alistic as possible. Experimental results shown in the second expected. Particularly for the high-power applications in
part of this paper confirm these simulation results to be very the megavoltampere range considered in this paper, such
accurate. an approach does not seem to be applicable. Reference
However, MPC schemes are inherently computationally de- [22] is an exception to this by also aiming at minimizing
manding as an underlying optimization problem needs to be the switching frequency.
solved. Yet, as will be illustrated, the proposed control ap- 6) Linear (or locally linearized) models are used as predic-
proach requires a computational effort that is greatly reduced tion models. In particular, for three-level inverters that
with respect to standard MPC approaches. Given the strong include a neutral point potential with its nonlinear dynam-
performance and design advantages, such an approach can ical behavior, linear models tend to be inaccurate. More-
be considered to be cost effective particularly in the case of over, in [15], [19], and [20], the cross couplings between
large drives operating in the Megawatt region. In particular, the d- and the q-axis are neglected in the machine model.
a reduction of the switching frequency directly translates into Similarly, [21] and [22] consider only a resistive inductive
reduced losses and, therefore, into energy and cost savings (in load with a back EMF rather than an electric motor.
terms of operation and installation), which are significant in
Here, due to the space limitation, we have provided only
high-power applications.
an overview of the literature on predictive control of motor
The proposed control scheme can be considered as a combi-
drives that is most related to our approach. A coverage of the
nation of the two DTC concepts [7], [8] we have proposed ear-
academic contributions to the DTC problem can be found in
lier. Specifically, the notion of optimality and the approximation
[23, Section 5.6] and [4].
of the average switching frequency by the number of switch
This paper is structured as follows. Section II summarizes the
transitions over a short prediction horizon were introduced in
physical model of the DTC drive, while Section III reformulates
[7]; the concept of the evaluation forward in time is inherited
this model such that it can be used as prediction model in the
from [8]. The key features of this new control scheme are
MPC scheme. After stating the control problem in Section IV,
the extrapolation, the fact that all computations are performed
the MPDTC scheme is detailed in Section V. This controller is
online (in [7] and [8], we have precomputed the control law
available in two forms with control (switching) horizons N > 1
offline and stored it in a lookup table), the admissible switching
and N = 1, which differ mostly in the switching sequences’
sequence, and the use of a nonlinear (rather than a piecewise
degree of freedom and the handling of the switching con-
affine [7], [8]) prediction model for the controller synthesis.
straints. The control algorithms and the related computations
As shown in [9], the research community has recently started
are discussed in Section VI. Simulation results using ABB’s
to consider MPC schemes as a way of introducing performance
simulation setup are presented in Section VII, and conclusions
improvements in electrical drives. Even though DTC itself is
are drawn in Section VIII.
widely interpreted as a predictive control strategy [10]–[14], it
predicts only one step (one switch transition) ahead, and it lacks
an internal model, a cost function, and the notion of optimality, II. P HYSICAL M ODEL
which are fundamental elements of an MPC scheme. Some Throughout the paper, we will use the normalized time
of these elements are present in the more recent approaches scale t with one time unit corresponding to 1/ωb seconds,
[15]–[22]. Yet, these schemes differ in several significant as- where ωb is the base angular velocity. Additionally, we will use
pects from the control scheme proposed in this paper. ξ(t), t ∈ R, to denote continuous-time variables, and ξ(k), k ∈
N, to denote discrete-time variables with the sampling interval
Ts = 25 μs.
1) Except for [19] and possibly also [20], the prediction All variables ξabc = [ξa ξb ξc ]T in the three-phase system
horizon is restricted to one and the usage of a larger (con- (abc) are transformed to ξαβ0 = [ξα ξβ ξ0 ]T in the orthogonal
trol) horizon combined with a receding horizon policy is αβ0 stator reference frame through
not mentioned.
ξαβ0 = P ξabc . (1)
2) The DTC problem is formulated as a reference tracking
problem, namely, the formulated control problem tries to Using a stator reference frame and aligning the α-axis
minimize the deviation of the torque and stator flux from with the a-axis, the following transformation matrix is
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 1896 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 6, JUNE 2009

Fig. 1. Equivalent representation of a three-level voltage-source inverter

driving an IM.

obtained:
⎡ ⎤
1 − 12 − 21
2⎣ √ √
P = 0 3
− 23 ⎦ . (2)
3 1 2
1 1
2 2 2

A. Physical Model of the Three-Level Inverter Fig. 2. Voltage vectors produced by a three-level inverter on the αβ plane, the
corresponding values of the switch positions uabc (where “+” refers to “1” and
The equivalent representation of a three-level voltage source “−” to “−1”), and the admissible switch transitions (courtesy of ABB ATDD,
inverter driving an IM is shown in Fig. 1. At each phase leg, Switzerland).
the inverter can produce the three different voltages −Vdc /2,
0, Vdc /2, where Vdc refers to the total dc-link voltage. Let about the nature of the neutral point potential and methods
the integer variables ua , ub , uc ∈ {−1, 0, 1} denote the switch employed to tackle the related balancing problem, the reader
position in each phase leg, i.e., the phase state, where the values is referred to [24] and [25].
−1, 0, 1 correspond to the phase voltages −Vdc /2, 0, Vdc /2, In the inverter considered here—due to the fact that only
respectively. The 27 vectors of the form uabc = [ua ub uc ]T one di/dt snubber is available in the upper and the lower half,
are transformed into the stator reference frame using (1). respectively—not all switch transitions are possible, as shown
The resulting vectors of the form uαβ0 = [uα uβ u0 ]T are in Fig. 2. As shown, from [1 1 1]T , for example, switching
shown in Fig. 2, where they are mapped into the αβ plane. The is only possible to [0 1 1]T , [1 0 1]T or [1 1 0]T (and
vectors in the αβ plane are commonly referred to as voltage not to any of the other 23 switch positions).
vectors, whereas we will refer to ua , ub , uc as the switch
positions. The actual voltages applied to the machine terminals
B. Physical Model of the IM
are calculated from
The dynamics of the squirrel-cage IM are modeled in the
Vdc
vαβ0 = P uabc . (3) stator αβ0 reference frame. The α- and β-components of the
2 stator and the rotor flux linkages per second ψsα , ψsβ , ψrα , and
ψrβ , respectively, and the rotor’s rotational speed ωr are used
As shown in Fig. 1, the neutral point potential υn depends as state variables. The input voltages vα and vβ are the stator
on the state of charge of the two dc-link capacitors and is only voltages in the stator reference frame. The model parameters
affected when current is drawn directly from it, i.e., when one are the base angular velocity ωb , the stator and rotor resistances
of the switch positions is zero. Introducing υn as a state, the rs and rr , the stator, rotor, and mutual reactances xls , xlr , and
neutral point potential is described by xm , respectively, the inertia J, and the mechanical load torque
T . Note that throughout this paper, if not otherwise stated, we
dυn 1
=− ((1 − |ua |) isa + (1 − |ub |) isb + (1 − |uc |) isc ) are using normalized quantities, and the rotor quantities are
dt 2xC referred to the stator circuit. The state equations are [26]
(4)

with the stator phase currents isa , isb , isc and one of the two dψsα xrr xm
= −rs ψsα + rs ψrα + vα (6a)
symmetric capacitors xC of the dc-link. Taking into account dt D D
dψsβ xrr xm
that isa + isb + isc = 0, it is straightforward to derive = −rs ψsβ + rs ψrβ + vβ (6b)
dt D D
dυn 1 dψrα xm xss
= |uabc |T P −1 is,αβ0 (5) = rr ψsα − rr ψrα − ωr ψrβ (6c)
dt 2xC dt D D
dψrβ xm xss
= rr ψsβ + ωr ψrα − rr ψrβ (6d)
where is,αβ0 is the stator current expressed in the stator refer- dt D D
ence frame, and |uabc | = [|ua ||ub ||uc |]T is the componentwise dωr 1
= (Te − T ) (6e)
absolute value of the inverter switch positions. For more details dt J
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1897

with approximation approach, the following discrete-time model of

the DTC drive is derived:
xss = xls + xm (7a)  
A 0
xrr = xlr + xm (7b) x(k+1) = I + T x(k)+
0 0 s
D = xss xrr − x2m (7c)  
B1 0
+ Ts u(k)+ T |u(k)| (14a)
and the electromagnetic torque 0 B2 (x(k)) s
xm
Te = (ψsβ ψrα − ψsα ψrβ ). (8) y(k) = g (x(k)) . (14b)
D
The length of the stator flux vector is given by In this model, I denotes the identity matrix, and Ts = 25 μs is
 the sampling interval. The definitions of the matrices A, B1 and
Ψs = ψsα 2 + ψ2 . (9) B2 , and the vector g(k) can be found in the Appendix. Note
sβ
that the zeros in (14) are vectors and matrices of appropriate
Equations (6)–(9) represent the standard dynamical model of dimensions. In (14), the first two terms capture the motor
an IM, where the saturation of the machine’s magnetic material, equations, while the third expression captures the dynamic of
the changes of the rotor resistance due to the skin effect, and the the neutral point potential.
temperature changes of the stator resistance are neglected.
IV. C ONTROL P ROBLEM
III. I NTERNAL M ODEL OF THE C ONTROLLER The DTC control objectives are to keep the three-output
In this section, we derive a discrete-time model of the drive variables, namely, the torque, the length (or magnitude) of
that is suitable to serve as an internal prediction model for the stator flux, and the neutral point potential, within given
the model predictive controller proposed in the next section. (hysteresis) bounds. At the same time, the average switching
The purpose of this model is to predict the trajectory of the frequency of the inverter
electromagnetic torque, the stator flux, and the inverter neutral N
point potential over several sampling intervals in an open-loop 1
f = lim u() − u( − 1) (15)
fashion. N →∞ N Ts
=0
As the time-constant of the rotor speed dynamic exceeds the
length of the prediction interval by several orders of magnitude, needs to be minimized, where  ·  denotes the 1-norm.
the rotor speed dynamics are neglected, and the speed is as- The control problem is complicated by the fact that the
sumed to remain constant within the prediction horizon. This control objectives comprise phenomena of very different time
allows us to treat the speed as a model parameter rather than as scales. Specifically, the control objectives relevant to the motor
a state thus removing (6e) from the motor model. depend on the very fast dynamics of the stator flux, which is af-
The model of the inverter has one state, namely, the neutral fected by the applied stator voltage within a few microseconds.
point potential, whose dynamic is described by (5) as a function On the other hand, the average inverter switching frequency
of |uabc | and is,αβ0 . The α and β-components of is,αβ0 are needs to be evaluated over a time frame of several 100 ms.
linear combinations of the stator and rotor flux components This is particularly the case in high-power applications, where
(see, e.g., [26] for details), and the 0-component is always zero1 the switching frequencies are in the range of 200–400 Hz. This
implies that every semiconductor switch is turned on roughly
1
is,αβ0 = [xrr ψsα − xm ψrα xrr ψsβ − xm ψrβ 0]T . (10) every 3–5 ms. Since the switch positions can be altered every
D Ts = 25 μs, a very long horizon (a large N ) of several hundred
We define the overall state vector of the drive as steps is necessary to capture the average switching frequency,
thus leading to an intractable MPC scheme.
x = [ψsα ψsβ ψrα ψrβ υn ]T (11) Apart from this, additional restrictions on the inverter switch
transitions may be present resulting from the construction of the
the switch positions ua , ub , and uc as the input vector inverter (as mentioned in Section II-A). We will refer to them
as switching constraints.
u = uabc = [ua ub uc ]T ∈ {−1, 0, 1}3 (12)

and the electromagnetic torque, the length of the stator flux, and V. M ODEL P REDICTIVE D IRECT T ORQUE C ONTROL
the neutral point potential as the output vector
Adopting the principles of MPC, we present a novel con-
y = [Te Ψs T
υn ] . (13) trol methodology that considers all (admissible) switching se-
quences over a rather short switching horizon N , which is
Combining the motor model (6a)–(6d), (8), and (9) with the referred to as the control horizon in the MPC community. A
model of the inverter (5) and (10), and using forward Euler switching sequence is defined as a sequence of semiconductor
switch positions uabc over the time-interval of length N from
1 This follows from (1), taking into account that isa + isb + isc = 0. time-step 0 to time-step N − 1. In a next step, based on the
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 1898 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 6, JUNE 2009

Fig. 3. Block diagram of the MPDTC scheme, where U denotes a switching sequence, and the initial state x0 is composed of x(k) and u(k − 1).

nonlinear discrete-time prediction model (14), the MPC scheme 2) For these switching sequences, compute the drive re-
computes for each switching sequence the drive’s response, i.e., sponse, i.e., compute all torque, stator flux, and neu-
the evolution of the output variables over the switching horizon tral point potential trajectories starting from x(k)
N . To emulate a long output horizon, the “promising” output over the switching horizon N given by Y i (k) =
trajectories are extrapolated, and the number of time-steps is [y i (k), . . . , y i (k + N )].
determined when the first output variable hits a hysteresis 3) Determine those switching sequences that yield output
bound. The cost associated with each switching sequence is trajectories that are either feasible at the end of the
determined by dividing the total number of switch transitions switching horizon, or point in the proper direction at all
in the sequence by the length of the extrapolated trajectory. time-steps within the switching horizon. We refer to these
Minimizing this penalty yields the optimal switching sequence switching sequences as candidate sequences U i (k) with
and the next optimal switch position to be applied to the i ∈ Ic ⊆ I. Feasibility means that the output variable lies
inverter. within its corresponding bounds at time-step k + N ; to
The MPDTC scheme is available in two forms with N > 1 point in the proper direction refers to the case in which an
and N = 1 differing mostly in the degree of freedom for output variable is not necessarily feasible, but the degree
the switching sequences and the handling of the switching of the bounds’ violation decreases at every time-step
constraints and, thus, in the computational burden and the within the switching horizon. For the case N = 2, Fig. 4
performance. shows several example output trajectories that visualize
these properties. The aforementioned condition needs to
hold componentwise, i.e., for all three-output variables.2
A. Horizon N > 1 4) For the candidate sequences, extrapolate the output tra-
For MPDTC with a switching horizon larger than one, we jectories from time-instant k + N on linearly3 using the
consider only switching sequences of length N that meet the samples at the time-instants k + N − 1 and k + N . De-
switching constraints imposed by the physics of the inverter rive the number of time-steps after which the first of the
(see Section II-A). As shown in Fig. 3, given the current state three-output variables leaves the feasible region defined
x(k), the last switch position u(k − 1), the bounds on the by the corresponding upper and lower bound.4 This yields
output variables, and using the nonlinear discrete-time predic- the number of time-steps ni , i ∈ Ic this switching se-
tion model (14) of the DTC drive, the controller computes at quence can be applied before switching is predicted to
time-instant k the next switch position u(k) according to the
following procedure. 2 As an example, consider the case where the torque is feasible, the stator flux
points in the proper direction, and the neutral point potential is feasible.
1) Given the last switch position u(k − 1) and taking into 3 Particularly during high-speed operation, it is advantageous to extrapolate

account the constraints on the switch transitions in- the stator flux quadratically using also the flux sample at k + N + 1. The latter
duced by the inverter topology, determine all switching is computed by applying the switch position u(k + N ) = u(k + N − 1).
4 Note that we determine when the first output variable leaves the feasible
sequences U i (k) = [ui (k), . . . , ui (k + N − 1)] over the region rather than when it hits a bound. This is done to account for situations in
switching horizon N , where i ∈ I and I is an index set. which an output variable lies outside its bounds but steers toward one of them.
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1899

Fig. 4. Example torque trajectories for N = 2 (the same concept applies to

the stator flux and the neutral point potential). The feasible region between the
hysteresis bounds is hatched. (a) Trajectories that are either (left) feasible or
(right) pointing in the proper direction. (b) Trajectories that are neither (left)
feasible nor (right) pointing in the proper direction.

be required again. Thus, ni refers to the total length of

the (extrapolated) output sequence.
5) Compute for each candidate sequence i ∈ Ic the cost
ci = si /ni , where

k+N −1
si = ui () − ui ( − 1)1 (16)
=k

is the total number of switch transitions in the switching

sequence U i (k), and ni is the corresponding sequence
length. The cost ci is an approximation of the average
switching frequency, and ni can be interpreted as a time-
varying output horizon.
6) Choose the switching sequence U ∗ = U i (k) with the Fig. 5. Torque and stator flux trajectories and the switch positions of
minimal cost, where i is given by Example 1 starting at time instant k for the three switching sequences U 1 (k),
U 2 (k), and U 3 (k). The trajectories within the switching horizon N = 2 are
solid, their extrapolations are dashed lines. The numbers refer to the indices
i = arg min ci . (17) of the switching sequences. The regions between the upper and lower (hys-
i∈Ic
teresis) bounds are hatched. (a) Torque trajectories. (b) Stator flux trajectories.
(c) Switch positions.
7) Apply only the first switch position u∗ = ui (k) of this
sequence.
At the next time instant, repeat the procedure. U 2 (k) as the optimum. Note that this solution requires two
Example 1: To visualize the control concept, consider the switch transitions (one at time-instant k, the second one at
example shown in Fig. 5. Assume that there are the three time-instant k + 1), but this investment pays out due to the
switching sequences U i (k), i ∈ I = {1, 2, 3} over the switch- longer length of the output trajectory. Without extrapolation,
ing horizon N = 2. According to the definition, U 1 (k) and the controller would have selected U 1 (k) as the optimum, since
U 2 (k) are candidate sequences, whereas U 3 (k) is not. Extrapo- the corresponding cost expressions would be 1/2 and 1 for
lating the torque and the stator flux trajectories and determining U 1 (k) and U 2 (k), respectively. In the long run, however, this
when they leave the feasible region leads to the results sum- choice would be inferior compared with U 2 (k), thus motivating
marized in Table I. Minimizing the cost yields the sequence the concept of extrapolation. In this example, for the sake
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 1900 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 6, JUNE 2009

TABLE I the shortest path (in terms of time steps). 2) Choose the switch
CHARACTERISTICS OF THE THREE SWITCHING SEQUENCES IN EXAMPLE 1
position that yields the least number of switch transitions at
time-instant k. 3) Choose the switch position that provides
the most alternatives at time instant k + 1. The admissible
switching sequence can be stored in a lookup table, which
holds only the first switch position in the sequence according
to the receding horizon policy. This leads to a lookup table
of dimension 27 × 27. Exploiting the 2π/3 symmetry of the
of simplicity, we neglect the neutral point potential, which is voltage vectors in the αβ plane, the dimension of this lookup
treated in exactly the same way as the torque and the stator flux. table can be reduced to 11 × 27 as detailed in [23].
Generally, the worst case computation time of the algorithm Setting the switching horizon to N = 1, the control algo-
has to be accommodated in the given sampling interval. Yet, in a rithm is the same as described in Section V-A with the following
real-world drive application, where other tasks like supervision differences, where the numbering corresponds to the one in
and thermal protection are executed, the average computation Section V-A.
time should also be kept at a minimum. When computing the 1) Given the last switch position u(k − 1) and ignoring
next switch position, this can be achieved by first evaluating the constraints on the switch transitions induced by the
whether switching can be avoided altogether, i.e., whether the inverter topology, 27 switching sequences of length one
output variables are at time-step k + N within their respective result.
bounds when reapplying the last switch position for N time- 2) From the lookup table read out of the admissible switch-
steps. Only if this simple test fails, the aforementioned out- ing sequence from u(k − 1) to u(k) the first switch
lined computations need to be performed. Furthermore, bound position u∗ (k), and apply it to the inverter.
techniques can be added to prune suboptimal branches, thus
avoiding the computation of the whole switching tree over
N steps. VI. D ISCUSSION AND C OMPUTATIONAL E FFORT
The notion of the candidate switching sequences associated
with output trajectories that are elementwise feasible at the end
B. Horizon N = 1
of the switching horizon or point toward the bounds, leads to
The computational burden imposed by the MPDTC scheme the following properties. First, step changes in the bounds (e.g.,
with a switching horizon N > 1 might exceed the capabilities large steps in the torque reference) can be straightforwardly
of some of the existing DTC control hardware. A further reduc- handled. In such cases, a switching sequence may not exist that
tion of the computation time can be achieved by restricting the moves the output variables within N time-steps back inside
switching horizon to N = 1. However, given the constraints on the bounds. Considering also sequences that move the output
the switch transitions, such a short switching horizon imposes variables only toward—rather than inside—the bounds allows
restrictions on the set of reachable voltage vectors and may lead the control scheme to easily address this issue.
to an infeasible control problem. Second, excessive switching is avoided. As shown from
To reduce the computation time while dealing with the issue Figs. 4(a) and 5, the bounds on the output variables are in
of infeasibility, we propose in this section a modified scheme general not strictly imposed by the MPDTC scheme, and output
that uses a switching horizon of N = 1 and initially ignores trajectories of candidate sequences may violate the bounds
the switching constraints. As a result, the proposed algorithm similar to standard DTC. As a result, one or more output
yields an optimal switch position u∗ (k) that may not be directly variable might slightly violate a bound before a new switch
reachable from the last switch position u(k − 1), since it was position is selected. Hence, the MPDTC approach refrains from
calculated by ignoring the switching constraints. In a last step, unnecessary switching when the bounds are slightly shifted,
the constraints are reintroduced, and an admissible sequence of or when measurement noise and model uncertainties affect the
switch positions is calculated that leads to the optimal switch predictions in an adverse way.
position within several time steps. We will refer to this sequence Since the proposed MPDTC scheme is intended to capture
as an admissible switching sequence. and minimize the average switching frequency, a long predic-
More specifically, given the two switch positions μ and ν, tion interval is beneficial. To avoid an explosion of the related
with μ, ν ∈ {−1, 0, 1}3 , the admissible switching sequence computational complexity, a short switching (i.e., control) hori-
connects μ with ν via intermediate switch positions while zon N (usually two or three steps), but a long output horizon
taking into account the restrictions on the allowed switch (up to 100 steps) is used. This is achieved by extrapolating
transitions. From the fact that switching in one component (one the output trajectories from the end of the switching horizon
stack of the inverter) by one switch transition at a time step is onwards until the time instant where the first output variable
always possible, it follows directly that an admissible switching hits one of its bounds. Linear extrapolation is straightforward
sequence always exists that additional switch transitions are not to implement and computationally inexpensive. Note that this
required and that the cost is not increased. approach is closely related to blocking control moves in the
As the admissible switching sequence connecting two switch control literature and the multiple rate prediction model ap-
positions (with the lowest cost) is, in general, not unique, we proach, which we introduced in [7]. As the simulation results
use the following rules to narrow down the choices. 1) Follow will show, the concept of a short switching horizon combined
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1901

with extrapolation allows us to greatly increase the length of TABLE II

MOTOR RATINGS OF THE THREE CASE STUDIES
the prediction interval thus enhancing the performance of the
controller while keeping the computation times short.
Most importantly, the proposed control scheme can easily be
adapted to different drives with different motors and inverter
topologies (like two- or five-level inverters), as only the motor
and inverter models need to be updated, and because there
are no tuning parameters. Furthermore, possible constraints on switching sequence is applied to the inverter. This is done in
the allowed switch transitions of the inverter can easily be accordance with the receding horizon policy.
incorporated by storing them in a lookup table. Apart from that, bounds on the output variables are handled
differently. In cases where the chosen switch position meets
the switching constraints, i.e., the desired switch position can
A. Horizon N > 1
be directly applied and the admissible switching sequence is
Nevertheless, when constraints on the allowed switch tran- of length one, the bounds are strictly respected. However, if
sitions are present, short switching horizons restrict the set the admissible switching sequence comprises more than one
of switch positions that can be reached within the switching element, the bounds are not guaranteed to be strictly respected
horizon. More specifically, up to three consecutive steps are (as it is the case for the scheme with N > 1).
necessary to switch from one switch position to another (e.g., Considering again the three-level inverter as above, the
from [−1 − 1 1]T to [1 1 − 1]T and vice versa). Thus, computational effort for N = 1 is reduced by a factor of five
a switching horizon of N = 2 is too short to ensure that any compared to the N = 2 case. Specifically, the upper bound is
arbitrary switch position can be reached within the switching 1859 operations.
horizon. For N = 2, this occurs only very rarely, because For a detailed analysis of the computational burden for the
those unreachable switch positions involve multiple switch MPDTC algorithms with N > 1 and N = 1, the reader is
transitions and are thus very expensive in terms of the cost referred to [23, Sections 7.4.1 and 7.4.2], respectively. These
expression. These issues are analyzed, visualized, and solved sections also provide the pseudocode of the algorithms.
in the second part of this paper [27] and in [23, Section 7.5].
Next, we briefly analyze the computational burden of the
VII. P ERFORMANCE E VALUATION
control algorithm. For this, we assume that all basic operations
such as additions, multiplications, divisions, and comparisons This section compares the performance of the proposed
require one computation cycle as well as evaluating a lookup MPDTC schemes with ABB’s ACS 6000 drive [6]. This implies
table. Possible operations for the loading or storing of variables that we will use ABB’s DTC as a reference, to which we refer
and the execution of loops are neglected. as standard DTC. The performance comparison is done through
To simplify the exposition, the switching horizon is fixed to simulations that are based on a very accurate and detailed
N = 2 hereafter. We assume that in the first step the inverter Matlab/Simulink model of the drive, which was provided by
can switch from the last switch position u(k − 1) to s1 voltage ABB to ensure an as realistic as possible simulation setup.
vectors. Subsequently, by branching on each of these s1 switch This model includes a state estimator for the motor fluxes, and
positions we obtain a total of s2 feasible switch positions at an outer (speed) control loop that adjusts the torque reference
the second step. Then, as shown in detail in [23, Section 7.4.1], and accordingly the (time-varying) bounds on the torque. For
the total number of operations required to execute the control MPDTC, the lookup table with ABB’s DTC strategy is replaced
algorithm is 16 + 39s1 + 79s2 . For ABB’s three-level inverter by a function that solves at each sampling-instant the optimal
with the particular constraints on the admissible switch tran- control problem according to Fig. 3 and Section V.
sitions, s1 is upper bounded by 13 and s2 is upper bounded Three case studies are considered comprising two medium-
by 121. Therefore, the upper bound on the total number of voltage drives and a low-voltage drive. Table II provides a
operations per control cycle is given by 10 082. Note that the rough overview of these case studies, while the detailed rat-
computationally most expensive parts of the algorithm, which ings and parameters can be found in [23, Section 7.6]. Two
requires 70% of the total computation power, can easily be aspects are compared. First, the torque, flux, and neutral point
parallelized. potential are compared when applying a torque step. Second,
the average switching frequency is compared over the whole
range of operating points, and the root mean squared violation
B. Horizon N = 1
of the torque and stator flux bounds is evaluated, too. For
When limiting the switching horizon to N = 1, the same both control schemes, the same torque and flux bounds are
properties described above are induced, except for the following used. For the neutral point potential, the bounds are chosen to
difference. Relaxing the constraints on the switch transitions reflect the behavior of ABB’s control scheme, thus ensuring the
and using the notion of the admissible switching sequence comparison to be meaningful.
avoids the restrictions on the set of switch positions that can The evaluations are performed for the whole operating range
be reached within the switching horizon. More precisely, any by gridding the speed ωr and the load torque T at 0.1, 0.2, . . .
switch position can be chosen by the optimization step, yet 1.0 p.u. The case of very high speed (0.9 and 1.0 p.u.) was left
only the first switch position of the corresponding admissible out for Case Studies I and II as the (total) dc-link voltage of
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 1902 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 6, JUNE 2009

Fig. 6. MPDTC with N = 2: Torque step from 0.1 to 0.9 p.u. torque at t = 100 ms and 50% speed. The torque, flux, and neutral point potential trajectories are
plotted together with their bounds over the time axis.

Fig. 7. Standard DTC: Torque step from 0.1 to 0.9 p.u. torque at t = 100 ms and 50% speed. The torque, flux, and neutral point potential trajectories are plotted
together with their bounds over the time axis.

4294 V is too low to allow for an operation at high speed. At average switching frequency. Averaging the data in Fig. 8(b)
each operating point, the behavior of the drive was simulated over all grid points yields an average switching frequency
over 2 s. reduction of 25% (or 60 Hz), while the maximum improvement
amounts to 42% (or 120 Hz). Over the whole range of operating
points, the bounds on the torque, stator flux, and neutral point
A. Torque Step for Horizon N > 1 potential are at least as well respected as by the standard DTC
For Case Study I and at 50% speed a torque step is applied scheme, as shown in Fig. 9 (the definition of the root-mean-
from 0.1 to 0.9 p.u. Fig. 6 shows the resulting closed-loop squared violation is given in the appendix). Nevertheless, as
behavior of the torque, the stator flux, and the neutral point described in Section VI-A, also the MPDTC scheme allows for
potential with MPDTC, whereas Fig. 7 shows the correspond- slight violations of the bounds.
ing trajectories resulting from ABB’s DTC strategy. As shown, It is interesting to note that the largest performance improve-
the MPDTC scheme preserves the rapid dynamic response ment is achieved around a modulation index of 0.5. Here, at the
achieved by the standard DTC approach of less than 2 ms at this transition between the inner and the outer hexagon in the tree-
operating point, while the bounds imposed on the torque, stator level inverter, the “density” of voltage vectors and, thus, the
flux, and neutral point potential are slightly better respected. degrees of freedom is at its maximum. This allows the MPDTC
scheme to choose among several different switching strategies
the one that meets the given performance objective best. On the
B. Average Switching Frequency for Horizon N > 1 other hand, it seems that the standard DTC scheme struggles to
First, we consider the MPDTC scheme proposed in take advantage of this.
Section V-A with a switching horizon N > 1. We set N = 2 The figures for Case Studies II and III are very similar and
and, additionally, limit the number of switching sequences by thus omitted here. As shown in Table III, the average and
imposing an upper bound of three on the total number of switch the maximal reductions are similar for all three case studies
transitions within a sequence. This bound removes a priori indicating that the proposed control scheme works equally well
switching sequences with very high cost and generally does not for drives with very different characteristics and ratings.
affect the performance.
For Case Study I, Fig. 8(a) compares the average switching
C. Average Switching Frequency for Horizon N = 1
frequencies of standard DTC with our proposed scheme. This is
done over the aforementioned defined grid of operating points. For the case of MPDTC with N = 1 as introduced in
Fig. 8(b) shows the percentagewise (relative) reduction of the Section V-B, we consider only Case Study I as an example.
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1903

Fig. 8. Comparison of the average switching frequency f of standard DTC with respect to MPDTC with N = 2 over the grid of operating points (speed and
torque). (a) Average switching frequency of standard DTC (upper surface) compared with MPDTC (lower surface). (b) Relative (percentagewise) reduction of the
average switching frequency.

Fig. 9. Percentagewise root mean squared violation of the torque and the stator flux (hysteresis) bounds for standard DTC (upper surface) compared with MPDTC
with N = 2 (lower surface) over the grid of operating points (speed and torque). (a) Torque violation of the bounds. (b) Stator flux violation of the bounds.

TABLE III Compared to the performance results for N = 2, the per-

REDUCTION OF THE AVERAGE SWITCHING FREQUENCY f OF MPDTC
WITH N = 2 RELATIVE TO THE STANDARD DTC formance improvement for N = 1 is smaller by one third.
FOR THE T HREE C ASE S TUDIES However, at least in the absence of noise, the MPDTC scheme
with N = 1 keeps the controlled variables tightly within their
bounds, whereas small violations of the bounds are tolerated
for N > 1.

VIII. C ONCLUSION
For Case Study 1, the average reduction of the switching In this paper, we have presented the new drive control
frequency over all grid points is given by 16% (or 40 Hz), concept MPDTC that is based on an internal controller model,
while the maximum improvement amounts to 38% (or 91 Hz). switching sequences comprising multiple time-steps, a con-
The bounds on the torque, stator flux, and neutral point po- troller objective function, and an optimization stage. The latter
tential are almost perfectly respected over the whole range of minimizes the objective function with respect to the internal
operating points. The corresponding figures are omitted here model dynamic, the bounds on the torque, stator flux, and
due to space limitations. The interested reader is referred to neutral point potential, and the admissible switch transitions.
[23, Section 7.4.4]. This optimization is performed over a multiple step prediction
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 1904 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 6, JUNE 2009

⎡ xrr ⎤
horizon. The control scheme is available in two forms—with a D 0 0
xrr
1 ⎢
switching horizon larger than one and with a horizon equal to 0 0⎥
⎢ xm ⎥ D
one. For illustration purposes, we have focused in this paper B2 (x(k)) = xT (k) ⎢− D 0 0 ⎥ P −T (20)
2xC ⎣ ⎦
on a three-level voltage source inverter driving a squirrel- 0 − xDm 0
cage IM. 0 0 0
Compared to the state of the art in drive control, the proposed ⎡ xm ⎤
D (x2 (k)x (k) − x4 (k)x1 (k))
control scheme offers two major advantages.  3
g(k) = ⎣ x21 (k) + x22 (k) ⎦ (21)
1) Performance: MPDTC inherits the very fast dy-
x5 (k)
namic torque response of standard DTC that is inherent in
hysteresis-based control concepts. By penalizing the short-term where xi denotes the ith component of the vector x.
switching frequency in the objective function, the average The percentagewise root mean squared violation of the
switching frequency is reduced—compared to ABB’s ACS torque over the discrete-time axis k = 1, 2, . . . , kmax is
6000 scheme—by up to 50%. In average over the whole range defined as
of operating points, the switching frequency is reduced by a 
quarter. At the same time, the imposed bounds on the torque, 
 1 kmax
flux, and neutral point potential are more strictly respected. In 100% ( T (k))2 (22)
the second part of this paper [27], these simulation results are kmax
k=1
accurately confirmed by experimental results. Moreover, the
performance improvement is practically independent from the with
rating and the specific drive characteristics. ⎧
⎨ Te (k) − Tmax , if Te (k) > Tmax
2) Flexibility and Simplicity: The MPDTC scheme is highly − Te (k), if Te (k) < Tmin
T (k) = T (23)
flexible. It is straightforward to incorporate additional or differ- ⎩ min
0, otherwise
ent performance and control objectives by simply modifying the
cost function. Most importantly, the controller can be directly where Te(k) and the bounds are given in per unit. The violations
applied to a large class of three-phase ac drives—since the for the stator flux and the neutral point potential are defined
controller is based on an internal prediction model, only this accordingly.
internal model needs to be adapted. This adaptation can be
done online as a parameter adaptation to account for a varying
stator resistance, for example, or it can be done offline as a ACKNOWLEDGMENT
model structure change to make the controller applicable to an- The authors would like to thank Chr. Stulz, P. Jörg, and
other drive with a different inverter topology and/or a different P. Schroderus of ABB ATDD, Turgi, Switzerland, and
electrical machine. The fact that the internal model is based A. Poncet of ABB Corporate Research, Baden-Dättwil,
on first principle dynamic equations simplifies the controller Switzerland, for their continuous advice. This work was per-
design. Except for the width of the hysteresis bounds, tuning formed at the Automatic Control Laboratory, ETH Zurich,
parameters (like controller gains in field oriented control) are Switzerland.
not required.
However, these advantages impose a significant computa- R EFERENCES
tional burden inherent to MPC. It is our opinion that the
[1] I. Takahashi and T. Noguchi, “A new quick response and high effi-
computational burden associated with the algorithms is close ciency control strategy for the induction motor,” IEEE Trans. Ind. Appl.,
to the lowest possible that is achievable for a MPC approach. vol. IA-22, no. 2, pp. 820–827, Sep./Oct. 1986.
Moreover, ABB’s successful implementation of the MPDTC [2] I. Takahashi and Y. Ohmori, “High-performance direct torque control of
an induction motor,” IEEE Trans. Ind. Appl., vol. 25, no. 2, pp. 257–264,
approach shown in the second part of this paper [27] proves that Mar./Apr. 1989.
such algorithms can be run on the already existing hardware in [3] P. Pohjalainen, P. Tiitinen, and J. Lulu, “The next generation motor con-
the time scales required. trol method, DTC direct torque control,” in Proc. Eur. Power Electron.
Chapter Symp., Lausanne, Switzerland, 1994, vol. 1, pp. 115–120.
[4] G. S. Buja and M. P. Kazmierkowski, “Direct torque control of PWM
A PPENDIX inverter-fed AC motors—A survey,” IEEE Trans. Ind. Electron., vol. 51,
no. 4, pp. 744–757, Aug. 2004.
The matrices and vectors of the discrete-time drive model [5] J. Maciejowski, Predictive Control. Englewood Cliffs, NJ: Prentice-
Hall, 2002.
(14) are given as follows: [6] ABB Asea Brown Boveri Ltd, Product Webpage of ACS 6000. online
document. [Online]. Available: www.abb.com/motors&drives
⎡ ⎤
−rs xDrr 0 rs xDm 0 [7] G. Papafotiou, T. Geyer, and M. Morari, “A hybrid model predic-
⎢ 0 −rs Dxrr
0 rs xDm ⎥ tive control approach to the direct torque control problem of induction
A = ⎣ xm ⎦ (18) motors (invited paper),” Int. J. Robust Nonlinear Control, vol. 17, no. 17,
rr D 0 −rr xDss −ωr pp. 1572–1589, Nov. 2007.
0 rr xDm ωr −rr xDss [8] T. Geyer and G. Papafotiou, “Direct torque control for induction motor
⎡ ⎤ drives: A model predictive control approach based on feasibility,” in
1 0 0 Hybrid Systems: Computation and Control, vol. 3414, M. Morari and
Vdc ⎢ 0 1 0⎥ L. Thiele, Eds. Berlin, Germany: Springer-Verlag, 2005, pp. 274–290.
B1 = ⎣ ⎦P (19) [9] R. Kennel and A. Linder, “Predictive control of inverter supplied electrical
2 0 0 0 drives,” in Proc. 31st IEEE Power Electron. Spec. Conf., Galway, Ireland,
0 0 0 2000, vol. 2, pp. 761–766.
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.
 GEYER et al.: MODEL PREDICTIVE DIRECT TORQUE CONTROL—PART I: CONCEPT, ALGORITHM, AND ANALYSIS 1905

[10] M. Pacas and J. Weber, “Predictive direct torque control for the PM syn- Tobias Geyer (M’08) received the Dipl.-Ing. and
chronous machine,” IEEE Trans. Ind. Electron., vol. 52, no. 5, pp. 1350– Ph.D. degrees in electrical engineering from the
1356, Oct. 2005. Swiss Federal Institute of Technology (ETH) Zurich,
[11] G. Abad, M. A. Rodriguez, and J. Poza, “Predictive direct torque control Zurich, Switzerland, in 2000 and 2005, respectively.
of the doubly fed induction machine with reduced torque and flux ripples From 2006 to 2008, he was with General Electric’s
at low constant switching frequency,” in Proc. Annu. Conf. IEEE Ind. Global Research Center, Munich, Germany, where
Electron. Soc., 2006, pp. 1000–1005. he was a Project Leader working on control and mod-
[12] P. Correa, M. Pacas, and J. Rodríguez, “Predictive torque control for ulation schemes for large electrical drives. In 2008,
inverter-fed induction machine,” IEEE Trans. Ind. Electron., vol. 54, no. 2, he joined the Power Electronics Group, Department
pp. 1073–1079, Apr. 2007. of Electrical and Computer Engineering, The Uni-
[13] R. Morales-Caporal and M. Pacas, “A predictive torque control for the versity of Auckland, Auckland, New Zealand, as a
synchronous reluctance machine taking into account the magnetic cross Research Fellow. His research interest is at the intersection of the fields of
saturation,” IEEE Trans. Ind. Electron., vol. 54, no. 2, pp. 1161–1167, control, optimization, and power electronics, which includes model predictive
Apr. 2007. control, hybrid systems, electrical drives, and dc-dc conversion.
[14] M. Nemec, D. Nedeljković, and V. Ambroz̆ic̆, “Predictive torque control
of induction machines using immediate flux control,” IEEE Trans. Ind.
Electron., vol. 54, no. 4, pp. 2009–2017, Aug. 2007.
[15] R. Kennel, A. Linder, and M. Linke, “Generalized Predictive Control
(GPC)—Ready for use in drive applications?” in Proc. 32nd IEEE
Power Electron. Spec. Conf., Vancouver, BC, Canada, 2001, vol. 4,
Georgios Papafotiou (M’03) received the Diploma
pp. 1839–1844.
and Ph.D. degrees in electrical engineering from
[16] G. Escobar, A. M. Stanković, E. Galvan, J. M. Carrasco, and R. Ortega,
Aristotle University of Thessaloniki, Thessaloniki,
“A family of switching control strategies for the reduction of torque ripple
Greece, in 1997 and 2002, respectively.
in DTC,” IEEE Trans. Control Syst. Technol., vol. 11, no. 6, pp. 933–939,
From 2003 to 2006, he was with the Auto-
Sep. 2003.
matic Control Laboratory, ETH Zurich, Zurich,
[17] J. Rodríguez, J. Pontt, C. Silva, P. Cortes, U. Amman, and S. Rees,
Switzerland, where he worked on model predictive
“Predictive direct torque control of an induction machine,” in Proc. 11th
control for hybrid systems, with a special focus on
IEEE Power Electron. Motion Control Conf., Riga, Latvia, 2004.
dc–dc converters and induction motor drives. Since
[18] J. M. Retif, X. Lin-Shi, A. M. Llor, and F. Morand, “New hybrid direct-
2006, he has been with the ABB Corporate Research
torque control for a winding rotor synchronous machine,” in Proc. 35th
Center, Baden-Dättwil, Switzerland, where he is cur-
IEEE Power Electron. Spec. Conf., Aachen, Germany, 2004, vol. 1,
rently working on the development of modern control and estimation methods
pp. 1438–1442.
for power electronics applications.
[19] A. Linder and R. Kennel, “Model predictive control for electrical drives,”
in Proc. 36st IEEE Power Electron. Spec. Conf., Recife, Brazil, 2005,
pp. 1793–1799.
[20] A. Linder and R. Kennel, “Direct model predictive control—A new direct
predictive control strategy for electrical drives,” in Proc. Eur. Conf. Power
Electron. Appl., Dresden, Germany, 2005.
[21] J. Rodríguez, J. Pontt, C. A. Silva, P. Correa, P. Lezana, P. Cortés, and Manfred Morari (F’05) received the Diploma from
U. Ammann, “Predictive current control of a voltage source inverter,” the Swiss Federal Institute of Technology (ETH)
IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 495–503, Feb. 2007. Zurich, Zurich, Switzerland, and the Ph.D. degree
[22] R. Vargas, P. Cortés, U. Ammann, J. Rodríguez, and J. Pontt, “Predictive from the University of Minnesota, both in chemical
control of a three-phase neutral-point-clamped inverter,” IEEE Trans. Ind. engineering.
Electron., vol. 54, no. 5, pp. 2697–2705, Oct. 2007. He was appointed Head of the Automatic Control
[23] T. Geyer, “Low complexity model predictive control in power electronics Laboratory at ETH Zurich in 1994. Prior to that, he
and power systems,” M.S. thesis, Autom. Control Lab., ETH Zurich, was the McCollum-Corcoran Professor of Chemical
Zurich, Switzerland, 2005. Engineering and Executive Officer for Control and
[24] H. du Toit Mouton, “Natural balancing of three-level neutral-point- Dynamical Systems at the California Institute of
clamped PWM inverters,” IEEE Trans. Ind. Electron., vol. 49, no. 5, Technology, Pasadena. His interests are in hybrid
pp. 1017–1025, Oct. 2002. systems and the control of biomedical systems. He has held appointments with
[25] N. Celanovic and D. Boroyevich, “A comprehensive study of neutral-point Exxon and ICI plc and serves on the technical advisory boards of several major
voltage balancing problem in three-level neutral-point-clamped voltage corporations.
source PWM inverters,” IEEE Trans. Power Electron., vol. 15, no. 2, Dr. Morari has received numerous awards in recognition of his research con-
pp. 242–249, Mar. 2000. tributions, including the Donald P. Eckman Award and the John R. Ragazzini
[26] P. C. Krause, Analysis of Electric Machinery. New York: McGraw-Hill, Award of the Automatic Control Council, the Allan P. Colburn Award and the
1986. Professional Progress Award of the AIChE, the Curtis W. McGraw Research
[27] G. Papafotiou, J. Kley, K. Papadopoulos, P. Bohren, and M. Morari, Model Award of the ASEE, Doctor Honoris Causa from Babes-Bolyai University,
predictive direct torque control—Part II: Implementation and experi- Fellow of IFAC, the IEEE Control Systems Field Award, and he was elected
mental evaluation, vol. 56, no. 6, Jun. 2009. to the National Academy of Engineering (U.S.).

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY KURUKSHETRA. Downloaded on December 30,2023 at 06:58:26 UTC from IEEE Xplore. Restrictions apply.