8164 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO.

10, OCTOBER 2020

Discrete Space Vector Modulation-Based

Model Predictive Torque Control
With No Suboptimization
Ilham Osman , Student Member, IEEE, Dan Xiao , Member, IEEE, Kazi S. Alam , Member, IEEE,
Sayed Mohammad Showybul Islam Shakib , Student Member, IEEE,
Md. Parvez Akter , Student Member, IEEE, and Muhammed F. Rahman , Life Fellow, IEEE

Abstract—This article presents a simplified discrete nonlinearities in the control objectives [1], [2]. the objective
space vector modulation (DSVM)-based predictive torque of model predictive torque control (MPTC) for motor drives
control (PTC) scheme in order to improve the performance is to optimize the switching state that will be applied to the
of a two-level inverter-fed induction motor drive. DSVM
technique creates a number of virtual vectors which are voltage source inverter (VSI) to minimize the torque and flux
evaluated in the conventional all vector-based discrete deviations. [3], [4]. The behaviors of the controlled variables
space vector modulation-based model predictive torque are predicted (typically one step ahead in time) based on the
control (DSVM-MPTC) method. The high number of admis- linear mathematical model of electric machines [5]. Until now,
sible vectors increases the computational burden of DSVM- several modified MPTC methods proposed in the literature
MPTC, significantly. In this article, an efficient optimal volt-
age vector selection method is proposed to reduce the generally focus on two main aspects: Complexity reduction and
computational load of DSVM-MPTC from 37 to 13 enumera- performance improvement of FCS-MPTC [6]–[12].
tions. The vector selected from the reduced set of admissi- Apart from the abovementioned switching-table-based algo-
ble voltage vectors produces the same cost function value rithms, researchers proposed multiple-vector-based algorithms
as that of all vector-based DSVM-MPTC in the entire range to improve the steady-state performance of FCS-MPTC. Duty-
of operation of induction motor (IM) drives. The proposed
method reduces the computational burden effectively with- cycle control concept was introduced to FCS-MPC to reduce
out causing any suboptimization issues in both transients the ripples of torque and stator flux linkage by using one ac-
and steady states. Experimental results verify the effec- tive and one null vector [6], [13], [14]. The duty-cycle-based
tiveness of the proposed algorithm and its superior perfor- FCS-MPC was extended to optimize the tuning of the weighting
mance compared to the switching-table-based DSVM-MPTC
factor by using the derivative of torque ripples [6], [13]. Two
and the classic finite-control-set model-predictive-control
which only utilizes the real voltage vectors. vector-based modulated MPTC was proposed in [15], where
the duty cycles for two adjacent active vectors were used in
Index Terms—Cost function, discrete space vector each sampling period in order to reduce the common-mode
modulation (DSVM), flux vector, induction machine, model
predictive torque control (MPTC).
voltage in a three-phase voltage inverter [15]. A generalized
two-vector-based PTC was proposed for further performance
I. INTRODUCTION improvement in [7], where a combination of one active and one
zero vectors or two adjacent active vectors was applied during
N RECENT years, finite-control-set model-predictive-
I control (FCS-MPC) has been widely researched and suc-
cessfully applied to the motor drive applications due to its
each sampling period.
Despite the good features, FCS-MPTC has some challenges.
The conventional MPTC is computationally intensive due to all
intuitive features, such as fast dynamic response, multivariable
voltage vectors evaluation in the prediction stage [11]. When
control, easy implementation, and inclusion of constraints and
performance improvement is a major concern, the methods with
a single switching state in each sampling period results in high
Manuscript received November 17, 2018; revised March 22, 2019,
June 27, 2019, and August 9, 2019; accepted September 15, 2019. Date
torque and flux ripple [16]. This is due to the exploitation of the
of publication October 29, 2019; date of current version June 3, 2020. limited number of voltage vectors provided by the converters
This work was supported by the Energy Systems Research Group, [17], [18]. Calculation load becomes a more severe issue in
School of Electrical Engineering and Telecommunications, University
of New South Wales, Sydney, Australia. (Corresponding Author: Ilham
double-vector-based MPTC schemes due to the additional calcu-
Osman.) lation of vector duty ratios. Particularly for multilevel converters,
The authors are with the Energy Systems Research Group, the computational burden rises exponentially [19], [20].
University of New South Wales, Sydney, NSW 2052, Australia
(e-mail: [email protected]; [email protected]; shehjadr@gmail.
A discrete space vector modulation (DSVM) technique was
com; [email protected]; [email protected]; f.rahman@ proposed in [21] and [22] to enhance the performance of DTC
unsw.edu.au). of induction motor (IM) drives. In the DSVM technique, a
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org.
large number of virtual voltage vectors are synthesized within
Digital Object Identifier 10.1109/TIE.2019.2946559 one sampling period. DSVM integrated with FCS-MPC for a
0278-0046 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
OSMAN et al.: DSVM-MPTC WITH NO SUBOPTIMIZATION 8165

power converter was proposed in [23]. The results show that to produce the same cost function values as the conventional
the DSVM-MPC has the same performance as conventional FS- DSVM-MPTC without knowledge of the location and ampli-
MPC operates lower sampling frequency. Besides, DSVM-MPC tude of the reference voltage vector. In order to verify the
offers improved performance by creating a number of virtual effectiveness and performance improvement of the proposed
voltage vectors [24]. However, the calculation load increases preselection PTC method, two MPTC schemes are also imple-
rapidly with a large number of virtual vectors. Thus, the method mented, namely, the real vector-based basic MPTC (Method-1)
is not practical for a cost-effective drive system. It should be and the preselection-based DSVM-MPTC (Method-2). From
noted that a large number of subdivisions contribute to the com- comparing the cost function values of the proposed and con-
plexity of the switching table. A solution to this issue is proposed ventional DSVM-MPC methods, the zero suboptimality of the
in [21] and it was found that the subdivision of sampling period proposed method is verified in wide operating conditions of the
by a factor of three was optimal. drive. Experimental results show that the proposed method has
In order to overcome the drawbacks of the high computational performance superior to those of Method-1 and Method-2.
burden of the DSVM-MPC technique, several researchers have
tried to reduce the number of admissible voltage vectors. Wang II. TWO-LEVEL INVERTER AND INDUCTION MOTOR
et al. [25] proposed a DSVM-PTC combined with deadbeat
In this article, a two-level three phase standard voltage-source
control for computation load reduction of conventional DSVM
inverter (VSI) is used. Eight real voltage vectors are generated by
PTC. However, the desired deadbeat voltage vector was calcu-
the inverter, as shown in Fig. 1(a) and described in the following
lated by neglecting the stator resistance. This may cause sub-
equation:
optimality in the voltage selection for PTC as the approximated
dq-components on the flux linkage circle are not accurate. As 2
Vj = udc (S1 + aS2 + a2 S3 ) (1)
pointed in [26], the controller proposed in [25] is not param- 3
eter robust, because it considers a narrow region formed by where j = 0,....,7 representing the number of each available
only three voltage vectors for cost function minimization. A voltage vector, Sx (x = 1, 2, 3) denote the switching states of
duty-ratio-based algorithm is proposed in [24] which is based the three inverter legs, and a = ei2π/3 . Two zero vectors are of
on the locus of reference voltage vector determining the optimal the same magnitude. In this article, only one zero vector V(000)
sector. Only the extended control set (ECS) vector located in this is considered.
optimal sector was considered for cost function minimization. The equations of a squirrel-cage induction machine can be
The selection of optimal ECS vector must be followed by a presented in the stator reference frame (α, β) [29].
complex duty ratio calculation in order to reduce the torque and
dψs
flux ripples. A DTC-inspired preselection-based method was υs = Rsis + (2)
reported in [27] and [28], where the admissible voltage vectors dt
are selected based on torque error and stator flux error. The dψr
0 = Rrir + r
− jωe ψ (3)
method in [28] considered 12 out of 37 voltage vectors and the dt
preselected vectors are grouped in a complex three-dimensional
(3-D) table in terms of sector number and torque and flux errors. where υs (= [vαs vβs ]T ) is the stator voltage vector,
The discrete space vector modulation-based model predictive is (= [iαs iβs ]T ) is the stator current vector, ir (= [iαr iβr ]T )
torque control (DSVM-MPTC) in [28] does not always select is the rotor current vector, ψ s (= [ψαs ψβs ]T ) is the stator flux
the optimal vector which leads to performance degradation. T
r (= [ψαr ψβr ] ) is the rotor flux vector, Rs is stator
vector, ψ
An effective preselection method should preserve the per- resistance, and Rr is rotor resistance referred to the stator side.
formance of all-vector-based conventional DSVM-MPC while The flux linkage equations and torque equation in terms of ψ s
achieving the computational burden reduction. Limiting the 
and is are presented in (4)–(6).
feasible set (i.e., the candidate voltage vectors) in preselection-  
based FCS-MPC should be only used if the preselection method s = Lsis + Lmir = Ls ψ
ψ r − σLrir (4)
selects the same voltage vector that the conventional method Lm
 
would select. The cost function values of the conventional r = Lmis + Lrir = Lr ψ
ψ s − σLsis (5)
method and preselection-based method can be used to evaluate Lm
the suboptimization of the preselection method for MPC. If 3  
the same cost function values are generated by a preselection Te = p ψ s × is
 (6)
2
method, it implies the same performance as all-vector-based
where Ls is the stator inductance, Lr is the rotor winding induc-
method will be achieved. Suboptimal solutions will degrade the
tance, Lm is the mutual inductance, Te is the electromagnetic
system performance in terms of the control objectives.
torque, σ = 1 − (L2m /Ls Lr ) is the leakage coefficient, and p is
This article presents an improved DSVM-MPTC algorithm
the number of pole pairs.
which is computationally efficient with zero suboptimality and
yet very simple to implement. The proposed method reduces the
number of admissible voltage vectors from 37 to 13 by using III. DISCRETE SPACE VECTOR MODULATION TECHNIQUE
a novel preselection method based on geometrical analysis of The limited number of the voltage vectors of 2L-VSI results
vectors generated by DSVM. The proposed algorithm is able in high torque ripples and stator current harmonics. To avoid

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
8166 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO. 10, OCTOBER 2020

Fig. 1. (a) Real voltage vectors. (b) DSVM technique with three equal time intervals. (c) 30 Virtual voltage vectors and eight real voltage vectors.

these disadvantages in motor drive, the DSVM technique was variables: Torque and flux with an appropriate weighting factor
proposed as a promising solution [23]. According to the DSVM [28], [30]. The cost function is expressed as
technique, virtual voltage vectors can be synthesized in one
sampling period for prefixed time intervals [21]. For a two-level 2
 2
VSI, if one sampling period is subdivided into equal N number g = T ref − T p (k + 2) + λf |ψ s
 p (k + 2) |
 ref |−| ψ
s
of intervals, the virtual voltage vectors v vir can be expressed by (12)
(7). The total number of the feasible voltage vectors, including where T ref is the torque reference generated by the speed con-
real voltage vectors and the virtual voltage vectors, is calculated  ref | is the module of the flux reference [30]. This
troller and | ψ s
as in (8) weighting factor, λf , is eliminated in the proposed algorithm
 using the reference stator flux vector proposed in [31]. As
v vir = tj Vjreal (7) shown in Fig. 2, the stator flux vector reference ψ  ref is con-
s
j=1,2,...,N structed from the reference stator flux vector calculation block.
where Vjreal ∈ {V0, V2, . . . V7 } and tj = Ts A proportional-integral (PI) controller generates the required
N
angular slip frequency ωsl from the torque error. The change
ntotal = 3N 2 + 3N + 1. (8) in estimated stator flux’s position θest is calculated using ωsl
and rotor speed ωm . Finally, the stator flux vector, ψsref , is
Seven basic voltage vectors are shown in Fig. 1(a). If each
reconstructed with the magnitude of stator flux reference |ψsref |
sampling period is subdivided by three equal time intervals, 30
and present position of stator flux vector θsref . The torque and
virtual voltage vectors will be synthesized, as shown in Fig. 1(b)
stator flux control are implemented by regulating the stator flux
and (c).
vector. In this way, the weighting factor is eliminated, which
simplifies the cost function tuning and prediction.
IV. FLUX VECTOR-BASED PREDICTIVE TORQUE CONTROL
Weighting factor is one of the reasons leading to suboptimality
A. Measurement and Estimation in the preselection-based MPTC schemes. With two prime vari-
s (k) is estimated from the ables with different units, a reduced set of control vectors will
In the MPTC method, stator flux ψ
r (k) is derived from the dynamic model
r (k) and ψ choose suboptimal solutions. The value of the weighting factor
rotor flux ψ
influences the degree of suboptimality of the algorithm. When
of induction motor.
the torque control is optimized with the reduced set of control
r (k) = ψ
ψ r (k − 1) vectors, the optimal vector for stator flux amplitude control may
  be excluded from these reduced control sets. This results in the
Lm  Rr r (k − 1)
+ Ts Rr is (k)− −jωe (k) ψ suboptimization of stator flux control. Therefore, cost function
Lr Lr simplification as well as weighting factor elimination is required
(9) by the proposed algorithm to make optimal decisions in the entire
→
− range of operating conditions.
s (k) = Lm ψ̂ r (k) + σLsis .
ψ (10)
Lr
The estimated torque is then obtained as C. Prediction and Cost Function Minimization
3  
The predicted values of the stator current ips (k + 1) and stator
Te (k) = p ψ s (k) × is (k) .
 (11)
2  p (k + 1) can be obtained by the following equations:
flux ψ s

B. Weighting Factor Elimination

 p (k + 1) = ψ
ψ s (k) + Tsvs (k) − Ts Rsis (k) . (13)
In conventional MPTC for 2L-VSI, the cost function for s

selecting optimal voltage vector usually consists of two control

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
OSMAN et al.: DSVM-MPTC WITH NO SUBOPTIMIZATION 8167

Fig. 2. Control diagram for proposed DSVM-MPTC.

And the predicted current


ips (k + 1) = 1 − Ts .is (k)
τσ

Ts 1 kr r (k) + vs (k)
+ . . − j.kr ωe (k) ψ .
τσ Rσ τr
(14)
Finally, a cost function is constructed to calculate the deviation
between |ψ p (k + 2)| and | ψ ref |
s s
2
g= ψ  p (k + 2) .
 ref − ψ (15)
s s

It should be noted that the predicted value at (k + 2)th instant

is used instead of (k + 1)th in (15) for prediction of the stator
flux and current to compensate for the effect of one-step delay Fig. 3. (a) Space vector diagram for proposed DSVM-MPTC. (b) Can-
in digital control [5]. v opt1 = V
didate vectors in stage-2 when   3.

V. PROPOSED DSVM–MPTC
The control block diagram of the proposed DSVM-MPTC
IM drive is shown in Fig. 2. To reduce the number of voltage
vectors in an efficient way, a simple voltage scanning method is
proposed which can reduce the total number of enumerations in
the prediction loop from 37 to 13 with no suboptimization.

A. Voltage Vectors Selection Method

The full voltage control set is presented in Fig. 3(a) for
realizing the proposed DSVM-MPTC algorithm. The voltage
control set consists of 7 real voltage vectors V n (n ∈ 0, 1 . . . 6)
and 30 virtual voltage vectors vm (m = 8, 9 . . . 37). From (2) Fig. 4. Experimental setup.
1  s (k + 1) + Ts Rsis (k + 1)

vs (k + 1) = ψs (k + 2) − ψ
Ts
(16) To obtain the minimum cost function value (g = ∼ 0), the de-
where vs (k + 1) is the voltage vector applied in (k+1)th instant. sired voltage vector vs (k + 1) ∼= Δψ err /Ts should be applied
s0
Substituting (16) into (15) the cost function can be expressed as [7]. DSVM-PTC is a type of FCS-PTCs and thus the desired
  2 voltage vector is not always of the finite control set vectors.
g = {ψ  ref − ψ
s (k + 1)}−Ts vs (k + 1)−Rsis (k + 1)
s Among the FCS vectors, the voltage vector which is closest to
2 this desired voltage vector vs (k + 1) should be selected as the
 err − vs (k + 1) Ts
= Δψ (17) optimal vopt to minimize the cost function g.
s0
Fig. 3(a) presents a two-level space vector diagram divided
where Δψ  ref − ψ
 err = ψ s (k + 1) + Ts Rsis (k + 1). into six sectors denoted as I to VI. The 30 virtual vectors are
s0 s

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
8168 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO. 10, OCTOBER 2020

TABLE I TABLE II
REARRANGED VECTOR CONTROL SET CANDIDATE VOLTAGE VECTORS IN STAGE 1 AND 2

classified into four categories as given in Table I. Each of the

sectors contains one real active vector, one zero vector, two large
virtual vectors, one medium virtual vector, two second-shortest
prediction loop in the proposed algorithm is reduced from 37 to
virtual vectors, and one shortest vector as shown in Fig. 3(b).
13 compared to all-vector-based DSVM-MPTC. The candidate
The voltage and flux vector lengths are denoted as “d” in Fig. 3.
virtual vectors in each optimal sector are presented in Table II.
The vopt selection algorithm has the following stages:
1) Stage 1. Scanning of the Real Active Voltage Vectors
V1 − V6 : In this stage, the proposed algorithm evaluates (14) for VI. EXPERIMENTAL STUDIES
all real active voltage vectors V1 ∼ V6 in Fig. 3(a). This stage To demonstrate the zero suboptimality of the proposed
(1) (1)
finds one preliminary optimal vector as vopt that gives gopt . method, the experimental comparison between the proposed
(1)
Based on vopt , the optimal sector is determined. Vectors within method and all-vector-based DSVM-MPTC (conventional al-
this optimal sector have shorter distances to vref compared to all gorithm I) is presented in this section. Conventional algorithm
the other vectors. One of the candidate voltage vectors within the I and the proposed algorithm use the same weighting-factor-
optimal sector gives the minimum value of cost function without free cost function (15) for a fair comparison. An experimental
any suboptimization. The optimal vector is finalized in Stages 2 comparison is also conducted between Method-2 [28] and the
and 3. all-vector-based DSVM-MPTC (conventional algorithm II) to
It should be noted that the scanning in the first stage can be investigate the suboptimality issue. Conventional algorithm II
started from any voltage vectors aligned with the real vectors. and Method-2 use the same cost function (12). The performance
For example, instead of scanning the real vectors, the medium improvement by integrating DSVM into the proposed algorithm
or shortest virtual vectors can also be scanned in the first stage is also confirmed by comparing the experimental results with
(1) basic MPTC with seven real vectors (Method-1).
to determine vopt and the optimal sector. However, if the scan-
All of the MPTC schemes are experimentally tested on the
ning starts with the second-shortest virtual vectors, a different
same 2L-VSI driven 415 V, 3-Ø, 50 HZ IM drive. A dSPACE
sector-division should be used. This results in a higher number
1103 controller board is used to implement the control algorithm.
of candidate voltage vectors allocated in the second stage.
For Method-1, Method-2 and conventional algorithm II, the
2) Stage 2. Scanning of Candidate Virtual Vectors and
weighting factor λf between torque and flux is set to 25. The
Zero Vector: The optimal real vector found in stage 1 deter-
parameters of the induction motor and control system used in
mines the candidate virtual vectors to be enumerated in Stage 2.
this article are listed in Table III.
The same cost function is evaluated for these candidate virtual
vectors in Stage 2 in order to find the final optimal vector which
is the closest to the desired voltage vector. If the desired vector A. Experimental Analysis on Suboptimization
(1) 3 is the
is located in sector III as shown in Fig. 3(b), v opt = V Suboptimization occurs if the cost function value produced
optimal real vector in Stage 1 as V 3 is the closest to the desired by a simplified DSVM-PTC method is higher than that of the
vector among all the real vectors. The candidate voltage vectors conventional DSVM-MPTC method. If the cost function values
for Stage 2 are: V 0 , v300 , v230 , v340 , v330 , v233 , and v334 . are the same, the suboptimality is zero, and the same steady-
Among these seven candidate voltage vectors, v330 has the state and dynamic performances as the conventional method
(2)
closest distance to vsref . Therefore v opt = v330 is the optimal are obtained. In summary, the suboptimality of reduced control
(2) set-MPFC with discrete switching states can be defined as (18).
virtual vector, which produces gopt in Stage 2. ⎧
3) Stage 3. Cost Function Value Evaluation From Stage 1 ⎨ 0, if |Δg| = 0
(2) (1)
and 2: Comparing gopt and gopt determines the optimal vector Suboptimality = ∞ (18)
(2) (1)
⎩ n, otherwise
from Stage 1 or 2. For example, if gopt < gopt , vopt = v330 , and n=1
(2) (1) 3 . The switching state corresponding to
if gopt ≥ gopt , vopt = V where |Δg| = |gLUT − gconventional |, gLUT is the cost function

v330 or V3 (depending on evaluation in Stage 3) is applied to value produced by a lookup table based-MPTC, gconventional
the inverter in the following cycle. The calculation time for the is the cost function value for all vector-based MPTC using the

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
OSMAN et al.: DSVM-MPTC WITH NO SUBOPTIMIZATION 8169

TABLE III the counter over a preset time indicates the total number of
PARAMETERS OF THE MACHINES AND CONTROLLER
suboptimizations caused by each of the simplification methods.
The suboptimization measurement is done for the entire range
of conditions for Method-2 and the proposed method as shown
in Fig. 5. The speed range is from 200 to the rated speed.
For each speed, the load varies from 0 to the rated torque,
7.4 Nm. As shown in Fig. 5, the proposed method maintains
zero suboptimality over a full range of speed and load. This
implies that the proposed method performs the same as the
conventional method because of no suboptimization occurring.
Method-2 leads to suboptimality due to the improperly reduced
number of vectors, as also indicated in Fig. 5. The selected sector
in Method-2 for voltage preselection is not always optimal.
The problem becomes severe at high-speed range when the flux
vector changes the sector more frequently. As shown in Fig. 5,
the percentage of suboptimization reaches as high as 39.9% at
1200 rpm with full load. Due to suboptimization of Method-2,
the steady-state and dynamic performance is degraded, as shown
in Figs. 6–8.

B. Steady-State Performance Comparison for

Medium and Low Speed
The steady-state responses at 1000 rpm speed with rated
load are presented in Fig. 6 for the proposed DSVM-MPTC,
torque and flux error based DSVM-MPTC (Method-2), and basic
MPTC (Method-1). The responses are presented in terms of
stator-current is , magnitude of stator flux |ψs |, torque Te , and
the difference in cost-function value Δg from its benchmark.
For the proposed method, Δg = (gproposed − gconven1 ), where
gproposed and gconven1 are the cost function values of proposed
and conventional algorithm I, respectively (15). Similarly, for
Method-2, Δg = (gmethod2 − gconven2 ), where gmethod2 is the
cost function value of Method-2 and gconven2 is the conventional
algorithm II (12). It should be noted that two different conven-
Fig. 5. Suboptimization analysis at different operating conditions. tional methods are adopted here to measure the suboptimality
of the proposed algorithm and Method-2.
From Fig. 6(a) and (b), it can be observed that the torque and
flux ripples are significantly reduced in the proposed method,
same cost function as the corresponding lookup table-based- compared to Method-2. The proposed algorithm always takes
MPTC. Conventional algorithm I using cost function (15), is optimal decisions as Δg is always at zero as shown in Fig. 6(a)
used as the benchmark for measuring suboptimization of the On the other hand, larger ripples in torque and flux are
proposed method. The proposed method is executed together observed due to suboptimal decisions made by Method-2. As
with the conventional algorithm I simultaneously in order to can be seen in Fig. 6(b), Δg is nonzero in most of the time during
detect the suboptimality. It should be noted that the conventional steady-state operation, which confirms that the performance
algorithm I is executed simultaneously only for monitoring of Method-2 is degraded due to suboptimization. Method-1
and measuring the degree of suboptimization. The switching has large torque and flux ripples due to the application of a
states applied to the inverter are generated only by the proposed real voltage vector. Compared to Method-1 and Method-2, the
method. proposed algorithm has improved steady-state performance.
Similarly, conventional algorithm II is used as the benchmark Steady-state responses for all three methods at 200 rpm with
for Method-2 and both methods with the same cost function half load are shown in Fig. 7. From Fig. 7(a), (b), and (c), it can be
and weighting factor are executed simultaneously. A 64-bit observed that the torque and flux ripples of the proposed method
digital counter is used to record the number of suboptimizations and Method-2 are significantly reduced compared to Method-1.
occurring in the proposed method and Method-2. It counts up The proposed algorithm always takes optimal decisions as Δg is
when a suboptimization is detected in each sampling cycle. This always zero as shown in Fig. 7(a). Although the overall perfor-
suboptimization measurement is conducted for a sufficiently mance of Method-2 at low speed and half load operation is better
long time for the operating conditions of Fig. 5. The value in than its operation at 1000 rpm with full load, notches in torque

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
8170 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO. 10, OCTOBER 2020

Fig. 6. Steady-state performances of all three methods at 1000 rpm

with 7.4 Nm (rated load). (a) Proposed method. (b) Method-2 (Torque
and flux error based DSVM-MPTC). (c) Method-1(basic real vectors Fig. 7. Steady-state performances at 200 rpm with 3.7 Nm (50% rated
based MPTC). load). (a) Proposed method. (b) Method-2. (c) Method-1.

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
OSMAN et al.: DSVM-MPTC WITH NO SUBOPTIMIZATION 8171

TABLE IV
EXPERIMENTAL COMPARISONS AT 80 µS SAMPLING TIME

and flux are still present due to nonzero Δg, as shown in Fig. 7(b).
The experimental results for all three methods are summarized
in Table IV. The total harmonic distortion (THD) of the stator
current ia for all three methods are included in Table IV. It is
observed that both proposed method and Method-2 have better
THDs than Method-1. However, the proposed method shows
about 50% lower THD compared to Method-2 at 1000 rpm with
full load operation.

C. Comparison in Torque-Transient Responses

A comparative study of transients in torque with the pro-
posed method, Method-2, and Method-1 is presented in Fig. 8.
The IM is operated with a step speed command from 200 to
1200 rpm with no load. Fig. 8 shows the transient responses
of the speed, current, torque, and suboptimization. The same
torque rise time, 0.72 ms, is obtained in each of three methods
tested. Furthermore, the proposed method achieves zero sub-
optimization during transient operation as shown in Fig. 8(a).
On the other hand, the suboptimality in Method-2 increases
rapidly as speed increases. The suboptimization results in higher
torque and current ripples during transient operation. Compared
to Method-1, the proposed method and Method-2 both attain
better torque ripples in both transient and steady-state operations
because of the DSVM used.

D. Performance Comparison With Basic FCS-PTC at

Three Times Lower Switching Frequency
Another comparison among the proposed algorithm,
Method-2, and basic MPTC (Method-1) is presented in Fig. 9.
The sampling frequency of Method-1 is set to three times
higher than the other two methods. Accordingly, the sam-
pling time for Method-2 and the proposed method are set to
120 μs while 40 μs for Method-1. The steady-state responses
at 1000 rpm speed with rated load are presented in Fig. 9 for
all three algorithms. The responses are presented in terms of
stator-current is , magnitude of stator flux |ψs |, torque Te , and
the suboptimization Δg. From Fig. 9(a) and (b), it can be seen
that the torque and flux ripples are significantly reduced in the
Fig. 8. Torque-transient performances of all three methods with a step
proposed method, compared to Method-2.
change in speed from 200 to 1200 rpm at no-load. (a) Proposed method. It should be noted that the performance of Method-2 and the
(b) Method-2. (c) Method-1. proposed method are not comparable to the one at 80 μs in

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
8172 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO. 10, OCTOBER 2020

Fig. 10. FFT analysis of ia at 1000 rpm with 7.4 Nm. (a) Proposed
method at 120 µs. (b) Method-2 at 120 µs. (c) Method-1 at 40 µs.

Fig. 6 because sampling time is prolonged to 120 μs in this com-

parison. Method-1 has improved torque and flux ripples due to
higher sampling frequency and the responses are comparable to
Method-2. However, steady-state performances of the proposed
algorithm is superior to that of Method-1 and Method-2.
Fig. 9. Steady-state performances investigation at higher sampling
The Fast Fourier Transform (FFT) analysis of the stator cur-
time at 1000 rpm with 7.4 Nm. (a) Proposed method at 120 µs. rent ia for all three methods are presented in Fig. 10. It should
(b) Method-2 at 120 µs. (c) Method-1(Basic real vectors based MPTC) be noted that the THD of the proposed method and Method-2 is
at 40 µs.
slightly increased compared to the performances at 80 μs due to
higher sampling time.

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
OSMAN et al.: DSVM-MPTC WITH NO SUBOPTIMIZATION 8173

[3] C. A. Rojas, J. Rodriguez, F. Villarroel, J. R. Espinoza, C. A. Silva, and M.

Trincado, “Predictive torque and flux control without weighting factors,”
IEEE Trans. Ind. Electron., vol. 60, no. 2, pp. 681–690, Feb. 2013.
[4] T. Geyer, “Algebraic weighting factor selection for predictive torque
and flux control,” in Proc. IEEE Energy Convers. Congr. Expo., 2017,
pp. 357–364.
[5] P. Cortes, J. Rodriguez, C. Silva, and A. Flores, “Delay compensation in
model predictive current control of a three-phase inverter,” IEEE Trans.
Ind. Electron., vol. 59, no. 2, pp. 1323–1325, Feb. 2012.
[6] Y. Zhang and H. Yang, “Model predictive torque control of induction motor
drives with optimal duty cycle control,” IEEE Trans. Power Electron.,
vol. 29, no. 12, pp. 6593–6603, Dec. 2014.
[7] Y. Zhang and H. Yang, “Two-vector-based model predictive torque control
without weighting factors for induction motor drives,” IEEE Trans. Power
Electron., vol. 31, no. 2, pp. 1381–1390, Feb. 2016.
[8] M. Ouhrouche, R. Errouissi, A. M. Trzynadlowski, K. A. Tehrani, and
A. Benzaioua, “A novel predictive direct torque controller for induction
motor drives,” IEEE Trans. Ind. Electron., vol. 63, no. 8, pp. 5221–5230,
Aug. 2016.
[9] M. Habibullah and D. D. Lu, “A speed-sensorless FS-PTC of induction
motors using extended Kalman filters,” IEEE Trans. Ind. Electron., vol. 62,
no. 11, pp. 6765–6778, Nov. 2015.
[10] M. Habibullah, D. D. Lu, D. Xiao, and M. F. Rahman, “A simplified finite-
Fig. 11. Switching Frequency analysis of Proposed DSVM-MPTC state predictive direct torque control for induction motor drive,” IEEE
Trans. Ind. Electron., vol. 63, no. 6, pp. 3964–3975, Jun. 2016.
[11] M. Habibullah, D. D. Lu, D. Xiao, I. Osman, and M. F. Rahman, “Selected
prediction vectors based FS-PTC for 3L-NPC inverter fed motor drives,”
E. Switching Frequency Analysis IEEE Trans. Industry Appl., vol. 53, no. 4, pp. 3588–3597, Jul./Aug. 2017.
[12] M. Habibullah, D. D. Lu, D. Xiao, and M. F. Rahman, “Finite-state
In Fig. 11 the average switching frequency of the proposed predictive torque control of induction motor supplied from a three-level
NPC voltage source inverter,” IEEE Trans. Power Electron., vol. 32, no. 1,
algorithm is presented. The operating speed ranges up to the pp. 479–489, Jan. 2017.
rated speed and the load varies from 0 to rated value for each [13] S. A. Davari, D. A. Khaburi, and R. Kennel, “An improved FCS-MPC
speed. The average switching frequency is calculated from the algorithm for an induction motor with an imposed optimized weighting
factor,” IEEE Trans. Power Electron., vol. 27, no. 3, pp. 1540–1551,
average switching transitions of the inverter [12]. The analysis Mar. 2012.
reveals that the average switching frequency for the proposed [14] S. Kwak and S. Mun, “Model predictive control methods to reduce
method maintains within 2.2–3.1 kHz. common-mode voltage for three-phase voltage source inverters,” IEEE
Trans. Power Electron., vol. 30, no. 9, pp. 5019–5035, Sep. 2015.
[15] Y. Zhang, B. Xia, and H. Yang, “Performance evaluation of an improved
VII. CONCLUSION model predictive control with field oriented control as a benchmark,” IET
Electric Power Appl., vol. 11, no. 5, pp. 677–687, May 2017.
This article thoroughly investigated the suboptimization is- [16] P. Karamanakos, P. Stolze, R. M. Kennel, S. Manias, and H. D. T. Mouton,
sues of the existing simplified DSVM-MPTC method for a “Variable switching point predictive torque control of induction machines,”
IEEE J. Emerg. Sel. Topics Power Electron., vol. 2, no. 2, pp. 285–295,
2-level inverter-driven IM drive. A new zero-suboptimization Jun. 2014.
DSVM-MPTC is proposed with significant and effective com- [17] I. Gonzalez-Prieto, M. J. Duran, J. J. Aciego, C. Martin, and F. Barrero,
putational load reduction. An analysis and method of evaluation “Model predictive control of six-phase induction motor drives using virtual
voltage vectors,” IEEE Trans. Ind. Electron., vol. 65, no. 1, pp. 27–37,
of zero suboptimality was discussed. The proposed algorithm Jan. 2018.
achieves the same performance of the conventional DSVM- [18] I. M. Hassine, M. W. Naouar, and N. Mrabet-Bellaaj, “Model predictive-
MPTC in wide operating conditions of an IM drive while using sliding mode control for three-phase grid-connected converters,” IEEE
Trans. Ind. Electron., vol. 64, no. 2, pp. 1341–1349, Feb. 2017.
a substantially reduced number of voltage vectors. The compu- [19] D. Casadei, F. Profumo, G. Serra, and A. Tani, “FOC and DTC: Two
tational burden in the prediction loop, as a result, was reduced viable schemes for induction motors torque control,” IEEE Trans. Power
by 65%. Experimental results demonstrated that the proposed Electron., vol. 17, no. 5, pp. 779–787, Sep. 2002.
[20] G. Foo and M. F. Rahman, “Sensorless direct torque and flux-controlled
algorithm achieves zero suboptimality and better steady-state IPM synchronous motor drive at very low speed without signal injection,”
and dynamic performances, compared to the simplified DSVM- IEEE Trans. Ind. Electron., vol. 57, no. 1, pp. 395–403, Jan. 2010.
MPTC method (Method-2) and the basic MPTC (Method-1) [21] D. Casadei, G. Serra, and K. Tani, “Implementation of a direct control
algorithm for induction motors based on discrete space vector modulation,”
over a wide range of operating conditions. The switching fre- IEEE Trans. Power Electron., vol. 15, no. 4, pp. 769–777, Jul. 2000.
quency analysis revealed that the proposed method maintains [22] X. Wei, D. Chen, and C. Zhao, “Minimization of torque ripple of direct-
the switching frequency of the drive in a narrow range. torque controlled induction machines by improved discrete space vector
modulation,” Electric Power Syst. Res., vol. 72, no. 2, pp. 103–112,
2004.
REFERENCES [23] S. Vazquez et al., “Model predictive control with constant switching
frequency using a discrete space vector modulation with virtual state
[1] P. Cortes, M. P. Kazmierkowski, R. M. Kennel, D. E. Quevedo, and J. vectors,” in Proc. IEEE Int. Conf. Ind. Technol., 2009, pp. 1–6.
Rodriguez, “Predictive control in power electronics and drives,” IEEE [24] Z. Zhou, C. Xia, Y. Yan, Z. Wang, and T. Shi, “Torque ripple minimization
Trans. Ind. Electron., vol. 55, no. 12, pp. 4312–4324, Dec. 2008. of predictive torque control for PMSM with extended control set,” IEEE
[2] J. Rodriguez, R. M. Kennel, J. R. Espinoza, M. Trincado, C. A. Silva, and Trans. Ind. Electron., vol. 64, no. 9, pp. 6930–6939, Sep. 2017.
C. A. Rojas, “High-performance control strategies for electrical drives: [25] Y. Wang et al., “Deadbeat model-predictive torque control with discrete
An experimental assessment,” IEEE Trans. Ind. Electron., vol. 59, no. 2, space-vector modulation for PMSM drives,” IEEE Trans. Ind. Electron.,
pp. 812–820, Feb. 2012. vol. 64, no. 5, pp. 3537–3547, May 2017.

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.
8174 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 67, NO. 10, OCTOBER 2020

[26] H. Moon, J. Lee, and K. Lee, “A robust deadbeat finite set model pre- Sayed Mohammad Showybul Islam Shakib
dictive current control based on discrete space vector modulation for a (S’16) received the B.Sc. degree in electrical
grid-connected voltage source inverter,” IEEE Trans. Energy Convers., and electronic engineering from the Rajshahi
vol. 33, no. 4, pp. 1719–1728, Dec. 2018. University of Engineering and Technology, Ra-
[27] T. Wang, C. Liu, G. Lei, Y. Guo, and J. Zhu, “Model predictive direct jshahi, Bangladesh, in 2012. He received the
torque control of permanent magnet synchronous motors with extended M.Eng.Sc. degree in the power electronics and
set of voltage space vectors,” IET Electric Power Appl., vol. 11, no. 8, renewable energy from the Department of Elec-
pp. 1376–1382, Sep. 2017. trical Engineering , University of Malaya, Kuala
[28] M. Amiri, J. Milimonfared, and D. A. Khaburi, “Predictive torque con- Lumpur, Malaysia, in 2016. He is currently work-
trol implementation for induction motors based on discrete space vector ing toward the Ph.D. degree in motor drives and
modulation,” IEEE Trans. Ind. Electron., vol. 65, no. 9, pp. 6881–6889, power electronic converters with the School of
Sep. 2018. Electrical Engineering, University of New South Wales, Sydney, Aus-
[29] R. Jose and C. Patricio, “Delay compensation,” in Predictive Control of tralia. His research interests include Motor Drives and Power Electronic
Power Converters and Electrical Drives, Hoboken, NJ, USA: Wiley, 2012, Converters.
pp. 177–189.
[30] C. A. Rojas, J. R. Rodriguez, S. Kouro, and F. Villarroel, “Multiobjective
fuzzy-decision-making predictive torque control for an induction motor
drive,” IEEE Trans. Power Electron., vol. 32, no. 8, pp. 6245–6260,
Aug. 2017.
[31] M. Habibullah, D. D. Lu, D. Xiao, J. E. Fletcher, and M. F. Rahman, “Low
complexity predictive torque control strategies for a three-level inverter
driven induction motor,” IET Electric Power Appl., vol. 11, no. 5, pp. 776–
783, May 2017.

Ilham Osman (M’15) received the B.Sc. (Hons.)

Md. Parvez Akter (S’16) received the B.Sc.
degree in electrical and electronics engineering
from the United International University, Dhaka, Eng. degree in electrical and electronics
engineering from the Chittagong University
Bangladesh, in 2011. She is currently working
of Engineering and Technology, Chittagong,
toward the Ph.D. degree in simplified model
Bangladesh, in 2011, and the M.Eng.Sc. degree
predictive control algorithms for induction mo-
in bidirectional power conversion and energy
tor drives with the Energy Systems Research
storage from the University of Malaya, Kuala
Group, School of Electrical Engineering, Univer-
Lumpur, Malaysia, in 2015. He is currently work-
sity of New South Wales, Sydney, Australia.
Her research interests are in Electric Drives, ing toward the Ph.D. degree in bidirectional
power conversion and energy storage at the
Model predictive control of Electrical Machines
Faculty of Engineering, University of New South
and power converters.
Wales, Sydney, NSW, Australia.
His research interests include power converters and electrical drives,
bidirectional power conversion, energy storage, smart grid, renewable
energy, and energy efficiency.
Dan Xiao (M’06) received the bachelor’s and
master’s degrees from the Shenyang University
of Technology, Shenyang, China, in 2001 and
2004, respectively, and the Ph.D. degree from
the University of New South Wales (UNSW Aus-
tralia), Sydney, Australia, in 2010, all in electrical
engineering.
He is currently working as a Technical Sup-
port with the Energy System Research Labo-
ratories, UNSW. His research interests include
sensor-less control and online parameter esti-
mation of ac machines, model predictive control for power converter and
drives, matrix converters, and solid-state transformers.
Muhammed F. Rahman (LF’14) received the
B.E. (Hons.) degree in electrical engineering
from the Bangladesh University of Engineering
and Technology, Dhaka, Bangladesh, in 1972,
Kazi Saiful Alam (M’15) received the B.Sc. and the M.Sc. and Ph.D. degrees in investi-
Eng. (Hons.) degree in electrical engineering gation of dynamics and high-speed resonance
from Bangladesh University of Engineering and instability of variable-reluctance type stepping
Technology, Dhaka, Bangladesh, in 2012. He motors from the University of Manchester In-
received the Ph.D. degree in control of power stitute of Science and Technology, Manchester,
converters and solid-state transformers from the U.K., in 1975 and 1978, respectively.
University of New South Wales, Sydney, Aus- He subsequently worked as a Systems De-
tralia, in 2018. sign Engineer with the General Electric Company at Rugby, UK, for two
Currently, he is working as a Postdoctoral years before joining the National University of Singapore, Singapore, in
Research Associate with the University of New 1980, as a Lecturer. He is currently a Professor in Energy Systems with
South Wales. His current research interests in- the School of Electrical Engineering, the University of New South Wales,
clude predictive control of power converters, control of isolated dc–dc Sydney, Australia. His research interests are in power electronics, elec-
converters and solid-state transformers. tric drives, electrical machines and motion control systems.

Authorized licensed use limited to: ESIEE PARIS. Downloaded on November 07,2024 at 09:26:49 UTC from IEEE Xplore. Restrictions apply.