9378 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO.

12, DECEMBER 2018

IPMSM Model Predictive Control in

Flux-Weakening Operation Using
an Improved Algorithm
Jinglin Liu , Member, IEEE, Chao Gong , Zexiu Han, and Haozheng Yu

Abstract—Three-phase interior permanent magnet syn- (EVs) that place intensive demands upon the good performance
chronous motor (IPMSM) drive systems have gotten widely (e.g., high dynamics, constrained current, and strong speed
used recently in various applications due to their speed regulation capacity) of their traction drive systems [4]. One
regulation characteristics and high power density. Com-
pared with classical field-oriented control (FOC) and flux- commonly accepted approach to satisfy these requirements is
weakening control, model predictive control (MPC) tech- inventing novel or improving classical control algorithms.
nique is more efficient and effective in achieving excellent In general, the conventional flux-weakening strategy based
performance without complicated controller tuning, but with on controlling d-axis stator current is required in most EV drive
explicit constraints. Yet, the optimum application of the MPC
systems for the high-speed operation [5], [6]. That method uses
algorithm in a machine control system is still in the ex-
ploratory stage. In order to lower the current and torque several proportional integrate (PI) controllers in the cascaded
ripples, this paper retains a modulator, but replaces all the control loops to calculate dq-axis current and voltage. Although
proportional integrate control loops, which are contained in it is relatively simple to bring into effect, three inevitable short-
the conventional FOC systems with a single MPC controller. comings exist [7]–[10]. First, because the outputs’ rise speed
Moreover, we adopt a brand-new linearization approach to of PI controllers cannot be constrained, the phase current might
tackle the strong coupled nonlinear IPMSM mathematical
model, obtaining an improved linear plant model, which is soar unexpectedly during the starting process, resulting in po-
suitable for the motor with constant load torque. When it tential risks. The second problem is that those (six, at least)
comes to flux-weakening control, the required d-axis cur- parameters of PI controllers have not gotten specific physical in-
rent is calculated, after which it is used as the input of the terpretation. Therefore, complicated parameter tuning schemes
proposed MPC controller, abandoning the previous natural for PI controllers are usually needed, not only enlarging time
field weakening method. Finally, for the purpose of lowering
the dynamic speed and current overshoot, a further con- cost in practical applications, but also burdening the overall
straint in the change rate of manipulated variables is dis- control performance optimization for various objectives. Third,
cussed. The improved control algorithm has been verified it is experimentally discovered that the phase currents fluctu-
in both simulation and experiment. ate obviously in the vicinity of the flux-weakening basic speed
Index Terms—Constraint, flux-weakening control, im- turning point, particularly when an IPMSM is under light load
proved linear plant model (ILPM), model predictive control condition, which has not been significantly improved by means
(MPC), modulator. of the conventional flux-weakening algorithm.
Model predictive control (MPC), proposed in 1970s, has been
I. INTRODUCTION
developed on a model basis as an alternative algorithm to the
NTERIOR permanent magnet synchronous motors PI control, which does not utilize the model [11], making it
I (IPMSMs) have higher efficiency due to the reluctance
torque and wide speed regulation range in flux-weakening
probable to overcome the shortcomings of PI controllers. Due
to the merits of online optimization control, quick response, and
operation on account of the salient-pole structure [1]–[3]. At manageable constraints, MPC causes wide academic concern
present, IPMSMs are heavily adopted in electrical vehicles within the electrical drive field [12]–[16]. Two typical applica-
tions are demonstrated as follows: first, model predictive direct
Manuscript received August 4, 2017; revised January 7, 2018 and torque control is an extension of direct torque control (DTC),
February 27, 2018; accepted March 12, 2018. Date of publication March showing dramatic promise in the control of medium-voltage
22, 2018; date of current version July 30, 2018. This work was supported drives [17]. Second, model predictive direct current control re-
by Shaanxi Science Technology Co-ordination and Innovation Project
of China under Grant 2016KTCQ01-49. (Corresponding author: Chao gards the phase currents as the controlled variables, optimizing
Gong.) the capacitor voltage variations and circulating currents [18].
The authors are with the Shaanxi Key Laboratory of Small and Special Moreover, [19] incorporates MPC into flux-weakening control
Electrical Machine and Drive Technology, School of Automation, North-
western Polytechnical University, Xi’an 710129, China (e-mail: jinglinl@ schemes and it is a direct speed control method that demands
nwpu.edu.cn; [email protected]; [email protected]; yhz29@ further investigation for both theoretical and practical progress.
qq.com). There are two implementation methods of MPC in flux-
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. weakening operation. One extends on conventional field-
Digital Object Identifier 10.1109/TIE.2018.2818640 oriented control (FOC) by replacing the control loops with
0278-0046 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: IPMSM MODEL PREDICTIVE CONTROL IN FLUX-WEAKENING OPERATION USING AN IMPROVED ALGORITHM 9379

an MPC-based controller, but still retaining the modulator loss are assumed negligible
[20], and it is employed in [19]. The other eliminates mod-
did Rs Lq ud
ulators altogether, with the outputs of MPC-based controller = − id + pωm iq +
directly selecting the optimal inverter switch states [21]–[23]. dt Ld Ld Ld
Comparatively speaking, the later strategy can bring compu- diq Ld Rs uq Ψf
= − pωm id − iq + − pωm
tational complexity reduction benefiting from a lookup table dt Lq Lq Lq Lq
that is obtained off-line, but the former provides more pre-
Te = 1.5p (Ψf iq + (Ld − Lq ) id iq ) (1)
cise switch positions, resulting in lower current and torque
ripples. dωm 1
= (1.5p (Ψf iq + (Ld − Lq ) id iq ) − Bωm − Tl ) (2)
MPC can be best implemented for the systems that accept dt J
a representation by a linear model with constraints, because in where id and iq are stator dq-axis current and ud , uq are dq-
that case, most of the optimization process can be moved offline axis control voltage. Ld and Lq are dq-axis inductance and the
[24], leading to the fact that the standard MPC design methods stator winding resistance is Rs . Te , Tl , and ωm are electromag-
usually require a linear plant model (LPM). But for the lack netic torque, load torque, and rotor mechanical angular speed,
of appropriate linearization approaches, LPMs for the strong respectively. Additionally, p represents the number of pole pairs
coupled nonlinear IPMSMs have not notably progressed yet, and ψf is permanent magnet flux linkage, while J and B are
limiting the control performance to some extent. the moment of inertia and the viscous coefficient, respectively.
The constraint handling capability of MPC is that it can deal In practice, (1) and (2) must be discretized in a time step of
with the input, output, and state constraints numerically over T (sampling time) to directly calculate the future states. When
the finite horizon. While designing an MPC controller, two con- forward Euler discretization is adopted, the prediction model is
straint subcategories are supposed to be considered. The best- obtained as follows:
known is to artificially set the maximum and minimum values
of variables, protecting devices from overload. Limitation on Ld − T Rs T Lq p
id (k + 1) = id (k) + ωm (k) iq (k)
the change rate of manipulated variables is directly related to Ld Ld
the steady state and dynamic performance likewise [25], but it T
is often overlooked in the previous studies. + ud (k) (3)
Ld
In this paper, a modulator is retained and all PI con-
T Ld p Lq − T Rs
trollers are replaced with a single linear multiple-input multiple- iq (k + 1) = − ωm (k) id (k) + iq (k)
outputpredictive controller [19], whereas the reference signals Lq Lq
conclude not only rotor angular speed but also d-axis cur- T T Ψf p
rent. Control of d-axis current contributes to achieving the + uq (k) − ωm (k) (4)
Lq Lq
goal of adjusting accurately air-gap field magnitude to dif-
ferent speed settings. The other crucial aims of this paper 1.5T pΨf 1.5T p (Ld −Lq )
ωm (k+1) = iq (k)+ id (k) iq (k) .
are to propose a linearization approach to obtain an improved J J
LPM(ILPM), which is expressed in state-space equations, and BT + J T Tl
+ ωm (k) − (5)
to introduce new constraint pattern about manipulated variable J J
change rate. The control effect of the proposed algorithm is
ultimately detailed according to simulation and experimental B. Classic IPMSM MPC Algorithm in Flux-Weakening
results. Operation
The structure of the rest of the paper is as follows. Sec- The classic IPMSM MPC algorithm is explained in terms of
tion II describes the nonlinear state-space model, the classical three different main techniques: linearization, control topology,
MPC-based flux-weakening algorithm and the improved MPC and constraints.
controller. In Section III, the control performance of the As said in the previous chapter, standard MPC design requires
two abovementioned schemes is analyzed according to the an LPM that can be expressed in the form
simulation results. The results of the successful experiments are
detailed in Section IV. Finally, Section V concludes this paper. x (k + 1) = Ax (k) + Bu (k)
y (k) = Cx (k) + Du (k) (6)

II. IPMSM MPC ALGORITHM where x is the vector of n state variables, u represents the ma-
nipulated variables, and y is a vector of the plant outputs. A, B,
A. IPMSM State-Space Model C, and D are the plant coefficient matrices. However, the non-
State-space models have an attractive advantage that they are linear terms ωm (k)iq (k), ωm (k)id (k) and id (k)iq (k) become
very easy to exploit the common behaviors of multivariable the biggest obstacle to convert (3)–(5) into (6). There are two
systems, making itself highly suitable for studying novel MPC classical approaches to remove the nonlinearities.
algorithms. The electrical dynamics and the mechanical dynam- The first way is achieved by choosing proper state variables.
ics of IPMSMs can be described by the following differential It needs to neglect the difference between Ld and Lq , in other
equations, where the iron saturation, eddy current and hysteresis words, assume Ld = Lq , so that the mechanical equation (5)

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
9380 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 12, DECEMBER 2018

is linear because id (k)iq (k) is equal to zero. Meanwhile, those undoubtedly simplify the system structure and take full advan-
nonlinear terms in (3) and (4) can be considered as measured dis- tage of the cost functions of MPC controllers. But for lack of
turbances because ωm (k), id (k), and iq (k) could be measured penalty on id , the optimization procedure might generate unsta-
at each sampling period [26]. Now, an LPM can be obtained ble control voltages, issuing in output torque ripple and speed
when the state variables of IPMSM are the currents id , iq and fluctuation in the constant power zone of IPMSM.
the speed ωm , along with ωm iq , ωm id Finally, controlling an IPMSM always needs to refuse both
 T overcurrent and overvoltage, so that the famous current and volt-
x (k) = id (k) , iq (k) , ωm (k) , ωm id (k) , ωm iq (k) . (7) age limit equations are welcomed in flux-weakening operation
Aside from selecting ωm iq and ωm id as the state variables, 
the second traditional linearization way is to assume a constant u2d + u2q ≤ Um ax (12)
angular speed for the nonlinear terms in the whole prediction 
horizon [19], that is ωm (k) = Ωm i , and Ld equals to Lq as well. i2d + i2q ≤ Im ax (13)
The electrical equations (3) and (4) can be described by
Ld − T Rs T Lq pΩm i T where Umax and Imax are the maximum allowable voltage and
id (k + 1) = id (k)+ iq (k)+ ud (k) current, respectively. Unfortunately, (11) and (12) cannot be
Ld Ld Ld
modeled as a function of constraints straightly because they are
(8)
nonlinear. But thanks to SVPWM technique, the voltage and
T Ld pΩm i Lq − T Rs current constraints can be expressed in hexagon shape
iq (k + 1) = − id (k) + iq (k)
Lq Lq
⎡ ⎤T

T T Ψf p −1 −1 0 0 1 1
ud (k)
+ uq (k) − ωm (k) . (9) ⎣ −1 1 2 −2 −1 1 ⎦ ≤ Um ax × R (14)
Lq Lq √ √ √ √ √ √ uq (k)
3 3 3 3 3 3
The state variables are only id , iq , and ωm
⎡ ⎤T

−1 −1 0 0 1 1
x (k) = [id (k) , iq (k) , ωm (k)]T . (10) ⎣ −1 1 2 −2 −1 1 ⎦
id (k)
≤ Im ax × R (15)
Although the IPMSM model is now linear, there still exist √ √ √ √ √ √ iq (k)
3 3 3 3 3 3
difficulties in obtaining an LPM on account of Tl . The previ-
ous study has introduced a torque observer and treated it as where R = [ 1 1 1 1 1 1] T . Note that there is no manipulated
an unmeasured disturbance calculated by the equivalent q-axis variable change-rate constraints here, which will degrade the
current iq 1 [27], [28] control performance, particularly, the speed overshoot.
Tl = 1.5pΨf iq 1 (k) . (11)
Certainly, iq 1 must be added to the state variable list. For ex- C. Proposed MPC-Based Flux-Weakening Algorithm
ample, if the second linearization method is employed, (10) will In allusion to the problems shown in Section II-B, the under-
turn into x(k) = [id (k), iq (k), ωm (k), iq 1 (k)]T . To be honest, mentioned measures are taken to establish the novel MPC-based
that method is not one hundred percent effective as a conse- flux-weakening algorithm to improve the classic one.
quence of two points, one of which is that the torque of IPMSM, This paper uses the first linearizing method in Section II-B,
unlike surface PMSM, is not only relevant to q-axis current, but where one of the most important priorities is ensuring that the
d-axis current. The approximate observer might be far from load torque will be included in an LPM. Assume that a constant
adapting to the flux-weakening cases. Second, the value of iq 1 torque is imposed on the rotor shaft, and then, the value of Tl
is unknown. Only when another permanent magnet synchronous is fixed. Regard Tl as a measured stable disturbance and add a
generator whose currents are measured works as the load ma- constant term 1 into the state variable vector (7), after that the
chine could Tl be estimated precisely, but the common method is linear drive model can be described as (16)–(20). So far, the
to let iq 1 equals to iq and this is only an approximation. Aiming ILPM of IPMSM can be online modified after offline identifi-
to a constant load torque, an improved algorithm is proposed in cation whenever the load torque varies during flux-weakening
Section II-C. operation.
As far as the classic control topology demonstrated in [19] Looking at the model at length, an extra feature of (16) is that
is concerned, it can be found that an inverter controlled by id (k)iq (k) is retained so as to be more appropriate for IPMSM
space vector pulse width modulation (SVPWM) algorithm is with no need for the assumption Ld = Lq . As is illustrated in
employed to reduce the current and torque ripples, but rotating (17), the dq-axis currents are seen as the output variables. In this
speed is the single reference signal during the natural MPC- paper, iq is only a measured output that needs to be constrained
based field-weakening operation. Comparatively speaking, con- and penalized during optimization process, but id as well as ωm
ventional PI-based flux-weakening strategy is realized by con- functions as the reference of MPC controller in marked contrast
trolling the d-axis current at the same time. In that course, the with the natural field-weakening algorithm. In detail, when a
negative reference current idref is maintained by utilizing closed- higher speed setpoint than basic speed ωbas is given, a negative
loop cascade method [29]. Fewer references in classic algorithm id requires to be controlled at a vested level, so that two tracking

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: IPMSM MODEL PREDICTIVE CONTROL IN FLUX-WEAKENING OPERATION USING AN IMPROVED ALGORITHM 9381

paths are wanted in the new algorithm rules:

 ⎧
⎪ i = 0, if ωm set ≤ ωbas
x (k) = id (k) , iq (k) , ωm (k) , id iq (k) , ωm id (k) , ⎨ dset
T ⎪ Ψ V . (23)
ωm iq (k) , 1 (16) ⎩ idset = − f + √ DC−bus , if ωm set > ωbas
Ld 
3Ce ωm set Ld
y (k) = [id (k) , iq (k) , ωm (k)]T (17)
Finally, the purpose of restraining the change rate of manip-
u (k) = [ud (k) , uq (k)]T (18) ulated variables is to reduce the acceleration of shaft, lowering
the dynamic speed overshoot, and meanwhile, that is conducive
[eq. (19) shown at the bottom of this page.] to preventing the motor and the IGBT from voltage and current
⎡ ⎤T surge to prolong the service life of the system. In addition to
T
0 0 0 0 0 0 those commonly used constraints in (14) and (15), the new flux-
⎢ Ld ⎥
B=⎢ ⎣
⎥ ,
⎦
weakening algorithm introduces one more constraint which is
T like
0 0 0 0 0 0
Lq
⎡ ⎤ |Δuq (k)| ≤ Δuq m ax
1000000
⎢ ⎥ |Δud (k)| ≤ Δud m ax (24)
⎢0 1 0 0 0 0 0⎥
C=⎢ ⎢ ⎥ , D = 0. (20)
⎥ where Δuq (k) and Δud (k) are manipulated voltage change rate
⎣0 0 1 0 0 0 0⎦
F4×7 = 0 and Δuq m ax and Δud m ax are the maximal q- and d-axis control
voltage variation during each time step, respectively.
However, another issue that how to preset the reference value The issue how to determine Δuq m ax and Δud m ax arises.
idset shows up. Considering that the IPMSM flux-weakening When the machine works under no load in the constant torque
process is characterized by full voltage operation and the line- region where id and iq are nearly zero, ud in (1) roughly equals
to-line back electromotive force approximately equals the dc- to zero as well. Then, uq can be approximately regarded as the
bus voltage when ignoring the turn-ON voltage of insulated gate sole control voltage during start-up course, and it peaks (no
bipolar transistors (IGBTs), the rotating speed ωm and the air- larger than VDC−bus ) when the IPMSM reaches the basic speed.
gap flux linkage ψf have the relationship expressed by Given that the required settling time of the system is ts and uq
grows monotonously and uniformly to the maximum voltage
VDC−bus , Δuq m ax is going to be approximated by
VDC−bus
Ψf = √ (21)
3Ce ωm T · VDC−bus
Δuq m ax = . (25)
ts
where Ce is the voltage constant relevant to the motor and
VDC−bus represents the dc-bus voltage. Meanwhile, the flux link- It is more complicated to decide Δudm ax theoretically.
age in the motor is also described as Nonetheless, now that Δuq m ax in (25) can guarantee lower
speed overshoot and avoid large voltage or current surge in the
Ψf = Ψf + Ld id . (22) constant torque region, it is believed that it is the same with
Δud m ax in the constant power region, that is
Because the dc-bus voltage is usually unchangeable once the Δudm ax = Δuq m ax . (26)
IPMSM system is established, the reference current idset is now
merely related to the motor speed setpoint value ωmset . More A block diagram of the proposed MPC algorithm realized on
specifically, idset is supposed to be controlled by the following an IPMSM control system is shown in Fig. 1.

⎡ L − TR T Lq p ⎤
d s
0 0 0 0 0
⎢ Ld Ld ⎥
⎢ ⎥
⎢ Lq − T Rs T Ψf p T Ld p ⎥
⎢ 0 − 0 − 0 0 ⎥
⎢ Lq Lq Lq ⎥
⎢ ⎥
⎢ ⎥
⎢ 1.5T pΨf BT + J 1.5T p(Ld − Lq ) T Tl ⎥
⎢ 0 0 0 − ⎥
A=⎢ J J J J ⎥ (19)
⎢ ⎥
⎢ 0 ⎥
⎢ 0 0 0 1 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 1 0 0 ⎥
⎢ ⎥
⎢ ⎥
⎣ 0 0 0 0 0 1 0 ⎦
0 0 0 0 0 0 1

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
9382 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 12, DECEMBER 2018

Fig. 2. Speed response. (a) Proposed MPC-based flux-weakening al-

gorithm. (b) Classic flux-weakening MPC algorithm.

Fig. 1. Block diagram of the proposed MPC algorithm.

TABLE I
MOTOR AND CONTROL PARAMETERS

Fig. 3. Response of currents under two kinds of algorithms. (a) Pro-

posed MPC-based flux-weakening algorithm. (b) Classic flux-weakening
MPC algorithm.

III. SIMULATION STUDY

A. Comparative Study on Control Performance
The improved MPC-based flux-weakening algorithm in com-
parison with a classic one is simulated to track forward com-
mand speeds in MATLAB/Simulink 2013b. The motor and con-
trol parameters of IPMSM are listed in Table I. In this section,
both no-load and under-constant-load conditions are analyzed.
First, the machine speeds up from 0 rad/s to 90 rad/s, where
id is supposed to remains zero and an external load of 4 N·m Fig. 4. IPMSM torque characteristics under two kinds of algorithms.
is applied at t = 2.0 s when the parameter Tl in the ILPM is (a) Proposed MPC-based flux-weakening algorithm. (b) Classic flux-
weakening MPC algorithm.
modified to 4 simultaneously. Second, set the reference speed
as 120 rad/s at t = 4.0 s, after that flux-weakening operation
launches. Finally, the load is removed immediately when t = proved by the response of currents iq and the torque characteris-
6.0 s. The simulation results are shown in Figs. 2–4. tics. Another advantage of the new MPC-based flux-weakening
1) Analysis on Steady-State Performance: The speed re- algorithm is that the current id in Fig. 3(a) is controlled to sta-
sponse results reflect marked differences between the pro- bilize at zero and −36.6 A when the speed setting point is 90
posed MPC-based flux-weakening algorithm and classic flux- and 120 rad/s, respectively. Comparatively, the id in Fig. 3(b) is
weakening MPC algorithm although both schemes can track not as stable as that in Fig. 3(a) during steady-state operation,
the commands well. Fig. 2(a) illustrates the motor speed re- acting on the air-gap magnetic field and further clearly reacting
mains stable after it arrives at the setpoint no matter whether the on the control performance.
IPMSM is loaded or not during the whole test range, but as is 2) Analysis on Dynamic Performance: Dynamic perfor-
seen in Fig. 2(b), the speed fluctuates slightly when the motor mance evaluation of an IPMSM control system needs to consider
is under load status. This phenomenon shows that the ILPM three crucial aspects, namely, settling time, speed overshoot, and
can contribute to achieving higher accuracy, which can also be current surge.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: IPMSM MODEL PREDICTIVE CONTROL IN FLUX-WEAKENING OPERATION USING AN IMPROVED ALGORITHM 9383

Fig. 6. Control performance of the proposed MPC controller consider-

ing dc-voltage fluctuations.

Fig. 7. Experimental equipment.

Fig. 5. Speed tacking properties of the proposed MPC controller.

C. Impact of DC-voltage Fluctuations on Stability and
Accuracy of Proposed Controller
In order to discuss the influence of dc-voltage fluctuations on
the control characteristics, a sinusoidal disturbance (amplitude
In terms of start-up speed, the rise time of the proposed MPC 2 V, frequency 1 HZ) is overlaid onto the voltage source. As in
algorithm is relatively longer (0.45 s) than that of the classic shown in Fig. 6, the machine speed can keep stable at 120 rad/s
algorithm (0.25 s). Nevertheless, thanks to the reformative con- if dc-link voltage experiences an upward change, while a rela-
straints in the proposed MPC-based flux-weakening algorithm, tively higher fluctuation of 2 rad/s (1.7%) is witnessed when the
Fig. 2(a) shows lower speed overshoot than Fig. 2(b). When fo- voltage goes down. Honestly, the proposed MPC-based algo-
cusing on the course of load step response, it is observed that the rithm does not use dc-link voltage to predict the system states,
current iq and the electromagnetic torque Te change faster, but a so that the system will not get out of control as long as the
noticeable damped oscillation occurs with the classic controller. voltage does not change sharply, and the speed can recover to
Meanwhile, the simulation results demonstrate relatively larger the set point when the voltage fluctuations disappear.
d-axis current surges whenever the load or the speed command
changes in Fig. 3(b), while fairly stable current id is visible in IV. EXPERIMENTAL RESULTS
Fig. 3(a). That phenomenon exactly proves that the proposed
algorithm is more powerful in controlling id . Apart from simulation, experiments are conducted on a three-
phase IPMSM whose parameters are also consistent with Table I.
The experimental equipment is shown in Fig. 7. A common dc
B. Analysis on Speed Tracking Characteristics for
power supply is available at 50 V. The 30 kW dynamometer is
Proposed Controller mechanically coupled to the IPMSM, providing constant load.
The speed tacking properties of the proposed MPC algorithm Both of the classic and the improved MPC algorithms are imple-
are specially discussed in this section taking both acceleration mented on DSP TMS32 28335 controller board. Motor phase
and deceleration in flux-weakening region. In Fig. 5, the IPMSM currents are measured using YOKOGAWA 96001 Clamp-on
works under no load between 0 and 2.5 s while a torque of 4 N·m Probes that are connected to WT3000 used for displaying wave-
is loaded over the left period. form, while dq-axis currents are calculated and recorded by the
Overall, the machine can remain stable at different speeds digital controller. The model DSP6001 is the bridge of PC and
(e.g., 120 and 110 rad/s) without steady-state errors. What’s dynamometer, directly controlling the load value, meanwhile,
more, the proposed controller is able to track both forward and sending the information of instantaneous torque and speed back
reversed command speed fast, and the overshoot is hardly seen. to the PC.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
9384 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 12, DECEMBER 2018

Fig. 8. Experimental results of steady states with no load. (a) Measured

speed and dq-axis currents of proposed MPC-based flux-weakening Fig. 9. Experimental results of steady states under 4 N·m. (a)
algorithm. (b) Measured speed and dq-axis currents of classic flux- Measured speed and dq-axis currents of proposed MPC-based flux-
weakening MPC algorithm. weakening algorithm. (b) Measured speed and dq-axis currents of clas-
sic flux-weakening MPC algorithm.

First, steady-state performances of the improved algorithm

and the conventional algorithm under no-load as well as load comparison, Fig. 10(a) shows the rise time of the proposed MPC
conditions are depicted in Figs. 8 and 9. It can be found that algorithm is longer (0.45 s), but there exists no speed overshoot
all of the speed waveforms nearly level off at 120 rad/s with no during speed-up process. During start-up, the maximum value
error, and id is about −37 A during flux-weakening operation. of iq and Te are more stable and smaller than the classic MPC
As a whole, higher fluctuations can be seen with load status, algorithm, 8.5 A and 1.2 N·m, respectively. One may notice that
but Fig. 9(a) shows obviously lower speed and current volatility the d-axis current of classic flux-weakening MPC algorithm
than Fig. 9(b), which is consistent with the simulation results. gets negative when a higher speed set point is given in constant
Fig. 10 proves that although both of the MPC algorithms torque zone while it is almost invariable for the proposed MPC-
are fundamentally for constant power zone, they are also well based flux-weakening algorithm, which is in agreement with the
applied to constant torque zone (e.g., 90 rad/s). simulated results in Fig. 3(a) and (b).
Second, dynamic performances of two control schemes in Then, the transient speed and current responses under abrupt
speed-up mode of operation are evaluated. In Fig. 10(b), the change of load are investigated and presented in Fig. 10. As
rotating speed quickly arrives at 90 rad/s with an overshoot expected, when IPMSM is loaded with 4 N·m in constant torque
of 3% within 0.3 s. iq and Te shift sharply and the maximum zone, the speed drops visibly before it returns 90 rad/s, and
value of them are 10 A and 1.6 N·m at startup. Overshoot (4%) conversely, ωm might jump gently immediately when the load
can also be seen as the IPMSM is accelerated to 120 rad/s. In is moved in flux-weakening process, followed by getting back

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: IPMSM MODEL PREDICTIVE CONTROL IN FLUX-WEAKENING OPERATION USING AN IMPROVED ALGORITHM 9385

Fig. 11. Speed tacking properties of the proposed MPC-based

algorithm.

Fig. 12. Measured control performance of the proposed MPC controller

considering dc-voltage fluctuations.

to 120 rad/s. What is indistinct in simulation, but clear here

includes that id suddenly becomes negative at both loading and
unloading moments. However, the proposed algorithm presents
smaller amplitude of variation (no more than 5 A).
Overall, the novel flux-weakening MPC algorithm presents
slower current and speed tracking performances, but fixed set-
tling time can be set, ensuring minimum overshoot. Another
competitive advantage is that dq-axis currents keep more stable,
especially at load condition.
The speed tacking properties of the proposed MPC algorithm
are depicted in Fig. 11. Whether the machine operates under
no-load or load conditions, the improved MPC algorithm is
capable of tracking any command speed in the flux-weakening
region efficiently, which is in accord with the simulation analysis
in Fig. 5.
Fig. 10. Measured speed, dq-axis currents and torque waveforms re-
Finally, when dc-voltage has fluctuations, as is shown in
sponse to start-up and load step-change before and after flux-weakening Fig. 12, the control system can remain stable in terms of machine
operation. (a) Proposed MPC-based algorithm. (b) Classic MPC speed. After the voltage fluctuations vanish, the speed gets back
algorithm.
to the set point quickly. What needs to be pointed out is that
the experimental test witnesses higher speed fluctuation than

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
9386 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 12, DECEMBER 2018

simulation, with 0.8 rad/s when the voltage rises and 3.5 rad/s [6] X. Liu, H. Chen, J. Zhao, and A. Belahcen, “Research on the performances
(2.9%) when the voltage decreases, respectively. and parameters of interior PMSM used for electric vehicles,” IEEE Trans.
Ind. Electron., vol. 63, no. 6, pp. 3533–3545, Jun. 2016.
[7] J. H. Park, D. J. Kim, and K. B. Lee, “Predictive control algorithm in-
cluding conduction-mode detection for PFC converter,” IEEE Trans. Ind.
V. CONCLUSION Electron., vol. 63, no. 9, pp. 5900–5911, Sep. 2016.
Many research works on applying MPC controllers in PMSM [8] B. S. Riar, T. Geyer, and U. K. Madawala, “Model predictive direct cur-
rent control of modular multilevel converters: Modeling, analysis, and
FOC systems have emerged, whereas there is relatively less experimental evaluation,” IEEE Trans. Power Electron., vol. 30, no. 1,
work about MPC-based flux-weakening control strategy used pp. 431–439, Jan. 2015.
in IPMSMs. The difficulties in real application include the [9] C. S. Lim, E. Levi, M. Jones, N. A. Rahim, and W. P. Hew, “A comparative
study of synchronous current control schemes based on FCS-MPC and
following respects: first, the IPMSM mathematical model for PI-PWM for a two-motor three-phase drive,” IEEE Trans. Ind. Electron.,
prediction is nonlinear; second, the optimal d-axis current con- vol. 61, no. 8, pp. 3867–3878, Aug. 2014.
trol method is uncertain; third, constraints need to be chosen [10] M. Cheng, F. Yu, K. T. Chau, and W. Hua, “Dynamic performance evalu-
ation of a nine-phase flux-switching permanent-magnet motor drive with
properly in practice. Although the exemplary MPC-based flux- model predictive control,” IEEE Trans. Ind. Electron., vol. 63, no. 7,
weakening algorithm presented a linear approach of PMSM, pp. 4539–4549, Jul. 2016.
the term of load torque is still an approximation. Besides, the [11] T. Geyer, “Model predictive direct current control: Formulation of the
stator current bounds and the concept of the switching horizon,” IEEE
natural field weakening method could generate unstable control Ind. Appl. Mag., vol. 18, no. 2, pp. 47–59, Mar./Apr. 2012.
voltages, lowering down the control performance. [12] J. Su, R. Gao, and I. Husain, “Model predictive control based field-
This paper proposed an improved MPC controller with a weakening strategy for traction EV used induction motor,” in Proc.
2016 IEEE Energy Convers. Congr. Expo., Milwaukee, WI, USA, 2016,
modulator for flux-weakening operation, which tries to take pp. 1–7.
those difficulties into account comprehensively. Meanwhile, a [13] A. Formentini, A. Trentin, M. Marchesoni, P. Zanchetta, and P. Wheeler,
comparative study with the prior conventional MPC algorithm “Speed finite control set model predictive control of a PMSM fed by matrix
converter,” IEEE Trans. Ind. Electron., vol. 62, no. 11, pp. 6786–6796,
and the proposed MPC algorithm was carried out and both of Nov. 2015.
the simulation and experimental results showed that the steady- [14] J. Lopez-Sanz, C. Ocampo-Martinez, and J. Alvarez-Florez, “Thermal
state and dynamic performance of the improved MPC algorithm management in plug-in hybrid electric vehicles: A real-time nonlinear
model predictive control implementation,” IEEE Trans. Veh. Technol.,
is better than the classic in some respects. vol. 66, no. 9, pp. 7751–7760, Sep. 2017.
Although the new MPC-based algorithm provides a more [15] W. Tang and Y. J. (Angela) Zhang, “A model predictive control approach
precise IPMSM predicting model for constant load conditions for low-complexity electric vehicle charging scheduling: Optimality and
scalability,” IEEE Trans. Power Syst., vol. 32, no. 2, pp. 1050–1063,
and manages to achieve better performance characteristics by Mar. 2017.
means of an improved control topology, there still exist some [16] E. Siampis, E. Velenis, and S. Longo, “Model predictive torque vectoring
limitations that should be addressed. First, it is highly required control for electric vehicles near the limits of handling,” in Proc. 2015
Eur. Control Conf. 2015, pp. 2553–2558.
to come up with an optimization approach for determining the [17] T. Geyer and S. Mastellone, “Model predictive direct torque control of a
maximal voltage change rate, which is obtained by approxi- five-level ANPC converter drive system,” IEEE Trans. Ind. Appl., vol. 48,
mation method currently. Second, the method to calculate the no. 5, pp. 1565–1575, Sep./Oct. 2012.
[18] M. Preindl and E. Schaltz, “Sensorless model predictive direct current
d-axis current reference by use of the speed reference is not control using novel second-order PLL observer for PMSM drive sys-
always efficient because it is generally parameter-dependent. tems,” IEEE Trans. Ind. Electron., vol. 58, no. 9, pp. 4087–4095, Sep.
So, an MPC controller that incorporates a brand-new idset de- 2011.
[19] Z. Mynar, L. Vesely, and P. Vaclavek, “PMSM model predictive con-
termination method without adopting extra PI controller should trol with field-weakening implementation,” IEEE Trans. Ind. Electron.,
be investigated. Third, the predicting model only works effec- vol. 63, no. 8, pp. 5156–5166, Aug. 2016.
tively when the load torque is constant and known in advance, [20] J. Scoltock, T. Geyer, and U. K. Madawala, “A comparison of model
predictive control schemes for MV induction motor drives,” IEEE Trans.
so that the proposed method requires prior knowledge of the Ind. Informat., vol. 9, no. 2, pp. 909–919, May 2013.
load torque when controlling an IPMSM. Consequently, it is by [21] C. K. Lin, J. T. Yu, Y. S. Lai, H. C. Yu, Y. H. Lin, and F. M. Chen,
far not widely applicable to all cases. “Simplified model-free predictive current control for interior permanent
magnet synchronous motors,” Electron. Lett., vol. 52, no. 1, pp. 49–50,
Jan. 2016.
[22] Y. Zhang and H. Yang, “Model predictive torque control of induction motor
REFERENCES drives with optimal duty cycle control,” IEEE Trans. Power Electron.,
[1] L. Rovere, A. Formentini, A. Gaeta, P. Zanchetta, and M. Marchesoni, vol. 29, no. 12, pp. 6593–6603, Dec. 2014.
“Sensorless finite-control set model predictive control for IPMSM drives,” [23] Y. Zhang and H. Yang, “Two-vector-based model predictive torque control
IEEE Trans. Ind. Electron., vol. 63, no. 9, pp. 5921–5931, Sep. 2016. without weighting factors for induction motor drives,” IEEE Trans. Power
[2] Y. Sun, M. Preindl, S. Sirouspour, and A. Emadi, “Unified wide-speed Electron., vol. 31, no. 2, pp. 1381–1390, Feb. 2016.
sensorless scheme using nonlinear optimization for IPMSM drives,” IEEE [24] A. Linder and R. Kennel, “Model predictive control for electrical drives,”
Trans. Power Electron., vol. 32, no. 8, pp. 6308–6322, Aug. 2017. Proc. 2005 IEEE 36th Power Electron. Specialists Conf., 2005, pp. 1793–
[3] T. Tarczewski and L. M. Grzesiak, “Constrained state feedback speed 1799.
control of PMSM based on model predictive approach,” IEEE Trans. Ind. [25] L. Wang, Model Predictive Control System Design and Implementation
Electron., vol. 63, no. 6, pp. 3867–3875, Jun. 2016. Using MATLAB (Advances in Industrial Control). London, U.K.: Springer,
[4] J. M. Mun, G. J. Park, S. Seo, D. W. Kim, Y. J. Kim, and S. Y. Jung, ch. 2, pp. 47–48, Mar. 2009.
“Design characteristics of IPMSM with wide constant power speed range [26] S. Bolognani, L. Peretti, and M. Zigliotto, “Design and implementation
for EV traction,” IEEE Trans. Magn., vol. 53, no. 6, pp. 1–4, Jun. 2017. of model predictive control for electrical motor drives,” IEEE Trans. Ind.
[5] Y. Zhang, W. Cao, S. McLoone, and J. Morrow, “Design and flux- Electron., vol. 56, no. 6, pp. 1925–1936, Jun. 2009.
weakening control of an interior permanent magnet synchronous motor for [27] J. Guzinski, H. Abu-Rub, M. Diguet, Z. Krzeminski, and A. Lewicki,
electric vehicles,” IEEE Trans. Appl. Supercond., vol. 26, no. 7, pp. 1–6, “Speed and load torque observer application in high-speed train electric
Oct. 2016. drive,” IEEE Trans. Ind. Electron., vol. 57, no. 2, pp. 565–574, Feb. 2010.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: IPMSM MODEL PREDICTIVE CONTROL IN FLUX-WEAKENING OPERATION USING AN IMPROVED ALGORITHM 9387

[28] M. Preindl and S. Bolognani, “Model predictive direct torque control with Zexiu Han was born in Shaanxi Province, China,
finite control set for PMSM drive systems, Part 1: Maximum torque per on May 4, 1990. He received the B.Eng. and
ampere operation,” IEEE Trans. Ind. Informat., vol. 9, no. 4, pp. 1912– M.Eng. degrees in electrical engineering from
1921, Nov. 2013. Northwestern Polytechnical University, Xi’an,
[29] Y. Zhang, W. Cao, S. McLoone, and J. Morrow, “Design and flux- China, in 2013 and 2015, respectively.
weakening control of an interior permanent magnet synchronous motor for His research interests include electrical ma-
electric vehicles,” IEEE Trans. Appl. Supercond., vol. 26, no. 7, pp. 1–6, chines design and drives and motion control.
Oct. 2016.

Jinglin Liu (M’01) received the B.Eng. degree

in electrical engineering from Tsinghua Univer-
sity, Beijing, China, in 1986, and the M.Eng.
and Ph.D. degrees in electrical engineering from
Northwestern Polytechnical University, Xi’an,
China, in 1990 and 2002, respectively.
Since 1994, he has been a Faculty Mem-
ber with Northwestern Polytechnical University,
Xi’an, China, where he is currently a Professor
in electrical engineering. His research interests
include electrical machines design and drives,
power electronics, fault diagnosis, and motion
control.
Haozheng Yu was born in Henan Province,
China, on July 3, 1992. He received the B.Eng.
Chao Gong was born in Shandong Province, degree in electrical engineering from Zhengzhou
China, on February 22, 1991. He received University, Zhengzhou, China, in 2015.
the B.Eng. and M.Eng. degrees in electrical His research interests include electrical ma-
engineering from Northwestern Polytechnical chines design and drives and motion control.
University, Xi’an, China, in 2014 and 2016,
respectively.
His research interests include electrical ma-
chines design and drives, power electronics, and
motion control.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2021 at 20:26:00 UTC from IEEE Xplore. Restrictions apply.