2018 4th International Conference on Green Technology and Sustainable Development (GTSD)
Advanced Speed Control of PMSM Motor Using Neural FOC Method
Ho Pham Huy Anh, Cao Van Kien, Tran Thien Huan, Pham Quoc Khanh
Abstract - This paper investigates an artificial neural network
(ANN)-based field-oriented control (NN-FOC) for a surface-
mounted permanent magnet synchronous machine (SPMSM)
drive. The ANN was trained by using the modified Levenberg-
Marquardt MLM algorithm. The novel Maximum Torque per
Ampere (MTPA) and flux weakening (FW) approach are
undertaken for electric vehicle (EV) application. The proposed
control is validated in simulation tests and compared with a PI-
based field-oriented control (PI-FOC). The simulation results
show that the maximum operating speed of the proposed control
is improved by 48% compared to the PI-FOC based control.
Consequently the proposed neural NN-FOC controller can be a
potential replacement of the existing control schemes, such as
PID, fuzzy logic, or others, and provides adequate traction control
in advanced electric drives, particularly in electric vehicle EV
application.
Keywords: surface-mounted permanent magnet synchronous
machine (SPMSM), Neural-based field-oriented control (NN-
FOC), electric vehicle EV application, Maximum Torque per
Ampere (MTPA) control, flux weakening (FW) control, modified
Levenberg-Marquardt MLM algorithm
I. INTRODUCTION
In general, the current industries often use a PI-based
field-oriented control (PI-FOC) to operate PMSMs. The
advantage of the PI controller is its simplicity and ease of
understanding. However, recent studies show that such
control strategy is experiencing the following bottlenecks
related to difficulty to obtain the optimal PI coefficients that
satisfy both stability and performance and limited stability
and performance [1-2]. Up to now, as to overcome these
obstacles, various control schemes were proposed. Lee et al.
[1] proposed an adaptive H∞ controller to enhance the
stability in system under parameter perturbations and
external disturbances. Mishra et al. [3] and Kamel et al. [4]
utilized a fuzzy logic controller to drive the PMSM. Lu et al.
[5] employed a self-constructing recurrent fuzzy neural
network (SCRFNN) to control the PMSM online. The
proposed SCRFNN was incorporated with two control
schemes: self-constructing fuzzy neural network and
recurrent neural network. Xia et al. [6] presented a novel
direct torque and flux control for matrix converter based
PMSM drive. The study investigated all matrix converter
switching states to control input and output variable. Four
switching schemes were based on the tables, which have
optimal selections of switching. Compared with the
conventional PI based controller, the matrix converter based
PMSM drive proved better in performance. Khan et al [7]
proposed a novel neural-network-based self-tuned controller
for IPMSM drives. However, only minimal improvement
was observed. J. Moreno et al [8] applied a neural network
control in hybrid vehicle energy management system to
minimize the discharge of the batteries. The experiment was
validated in both simulation and hardware. Z. Chen et al. [9]
utilized two neural-based modules that are trained by
dynamic programming (DP) to optimally manage the energy
between engine and battery in power-split hybrid electric
vehicle.
These abovementioned results show that there is still no
research conducted on an artificial neural network-based
vector control for PMSM including both flux weakening
control and electric vehicle EV implementation. Thus, this
paper proposes the new neural-based vector control NN-
FOC for surface-mounted permanent magnet synchronous
machine (SPMSM) with flux weakening control. The
SPMSM motor drive with the proposed control will be
validated through simulation.
The rest of this paper is organized as follows. Section 2
gives the full descriptions of PMSM model while Section 3
introduces the conventional PI-based vector control and
proposed neural network vector control with full
descriptions on training algorithm. Section 4 introduces
novel maximum torque per ampere (MTPA) and flux
weakening approaches. Section 5 shows the simulation
result of the proposed NN-FOC control for SPMSM drive.
Finally, the conclusion of the paper is provided in Section 6.
II. MATHEMATICAL PMSM MODEL
The concept of the mathematical PMSM model begins with
the stator d- and q-axis flux linkage equations, which are
written as:
ªOsd º ª Lsd .isd  O fd º
« O » « L .i » (1)
¬ sq ¼ ¬ sq sq ¼
where Lsd,sq are the stator d- and q-axis inductances and
equal to Lmd,mq + Lls while λfd is the flux linkage constant
of the stator d-axis winding due to flux produced by the
magnets in the rotor, where it is assumed that the d-axis is
always aligned with the rotor magnetic axis, which is a-axis.
In terms of the aforementioned flux linkage equation, the
stator winding voltages can be written as follows.
ª dO º
R i  t   sd  Ze Osq »
Ho Pham Huy Anh is with Faculty of Electrical and Electronics ªvsd º « s sd dt
Engineering (FEEE), Ho Chi Minh City University of Technology, VNU- «v » « » (2)
HCM, Viet Nam (e-mail: hphanh@hcmut.edu.vn) ¬ sq ¼ « R i  t   d Osq  Z O »
Cao Van Kien is is with Faculty of Electrical and Electronics «¬ s sq dt
e sd »
¼
Engineering (FEEE), Ho Chi Minh City University of Technology, VNU-
HCM, Viet Nam (e-mail: kiencv@hcmut.edu.vn) in which ωe denotes the electrical angular velocity in
Tran Thien Huan is with Ho Chi Minh City University of Technology electrical rad/s. The electrical angular velocity can be
and Education (HCM-UTE), Viet Nam (e-mail: huantt@hcmute.edu.vn) calculated by multiplying the electromechanical angular
Pham Quoc Khanh is with Industrial University of HCM City, Viet
Nam (e-mail: phamquockhanh@iuh.edu.vn)

978-1-5386-5126-1/18/$31.00 ©2018 IEEE 696

 2018 4th International Conference on Green Technology and Sustainable Development (GTSD)

velocity, ωmech, with the pole pairs, p. If (1) and (2) are § Rs Ze Lq · § 1 ·
¨  ¸
combined, the equation becomes d § isd · Ld Ld ¸ § isd · ¨¨ Ld ¸§ vsd ·
¨ ¸  ¨¨ ¨ ¸ ¸¨
¸¸ (7)
ªvsd º ª Rs  Lsd P Ze Lsq º ªisd º ª 0 º dt ¨© isq ¸¹ L Rs ¸ ¨© isq ¸¹ ¨ 1 ¨
¸ © vsq  ZeO f ¹
 ¨Z d ¸ ¨ ¸
«v » « ZL ¨ e Lq Lq ¹¸
Rs  Lsq P »¼ «¬isq »¼ «¬O fd Ze »¼
(3)
© © Ld ¹ u
¬ sq ¼ ¬ e sd B
A

with P represents the derivative with respect to time. Under in which isd and isq are the stator d- and q-axis currents; A
a balanced sinusoidal steady state condition, the terms represents the system matrix; B is the input matrix; and u
associated with the time derivative are neglected. Thus, the signifies the input vector. As illustrated, (7) is in a
dq currents become DC. As a result, (3) in steady state can continuous state-space form. Since the controller will be
be rewritten as, implemented in a digital controller, the discretization of (7)
is required and obtained by utilizing either a zero-order or
ªvsd º ª Rs Ze Lsq º ªisd º ª 0 º first-order hold discrete equivalent mechanism as shown as,
«v » «Z L « »« » (4)
¬ sq ¼ ¬ e sd Rs »¼ ¬ isq ¼ ¬O fd Ze ¼
ªisd  kTs  Ts º ªisd kTs º ª vsd  kTs   0 º
« » A« » B« » (8)
The electro-magnetic torque for synchronous machines is «¬isq  kTs  Ts  »¼ i
¬ sq kTs ¼ «¬vsq  kTs   Ze O fd »¼
described as
in which Ts denotes the sampling period. For Ts is present
Tem
p
2

Osd isq  Osq isd  (5) on both sides, the equation further simplifies into

If substituted with (1), the equation (5) becomes,

isdq  k  1 
A.isdq  k   B vsdq  k   vrdq  (9)

where k is an integer time step; isdq = (isd,isq)’; vsdq = (vsd,vsq)’

Tem
p
2
  
O fd isq  Lsd  Lsq isd isq 
III. PROPOSED NEURAL NN-FOC CONTROL
The main objectives of NN vector control are to achieve
the decoupled d- and q-axis current control and to make the
control system more robust against parameter variations and
unknown disturbance such as an impulse of load torque. As
to develop a current neural NN-based FOC controller, the
equation (4) has to be rewritten into the standard state-space
form as follows,
(6) are the control actions; and vrdq =(0,ωeλfd)’ illustrates the
induced voltage of the rotor permanent magnet.
Figure 1 illustrates overall PMSM motor drives using
neural NN-based current controller and neural network
structure. As it shows, the proposed neural model is
structured into three different layers: an input layer, an
action hidden network layer, and an output layer. The main
role of the input layer is taking the error terms and the
integrals of the error terms, edq and sdq, and then, transform
those by dividing the appropriate gains and applying the
hyperbolic tangent function. The error and integrals of the
error terms are expressed as
k
edq  k  idq  k   i*dq  k  ,sdq  k  ³ e  k  dt
dq (10)
0

Figure 1.PMSM conventional PI- and neural network NN-based FOC control

697
 2018 4th International Conference on Green Technology and Sustainable Development (GTSD)

defineV  k   edq  k  
 
Then, these outputs of the input layer feed forward to
¦ U edq  k  m ¦ V  k 
N N 2

action network layer. This layer is comprised of two hidden C k 1

oC k 1
(16)
layers of six nodes. Like in the input layer, each node is
Then the gradient of the cost function with respect to the
computed using hyperbolic tangent functions. The last layer
weight vector is
is called output layer. The primary objective of the output
layer is translating the outputs of the action network layer
w¦ k V  k 
2
wV  k 
N
wC
¦ 2V  k  2J p  w  V (17)
N T
into the reference dq voltages. This was achieved by 1

applying the hyperbolic tangent. Then, because these outputs ww ww k 1

ww
are normalized, a gain has to be multiplied. This gain is in which the Jacobian matrix and the error function are
equal to the maximum allowable voltage, which is depended defined as
on amplitude of DC voltage and type of PWM. For example,
since SVPWM is used, the maximum allowable voltage is ª wV  1 wV  1 º
« »
ksvpwm Vdc * 3 / 2 / 3 (11) « wwww1 wwM » ª V  1 º
« », V « »
J p  w « » (18)
As a consequent, the final control action vsdq is defined as « »
« wV  N  wV  N  » «V  N  »
follows, ¬ ¼
« »
¬ ww1 wwM ¼
vsdq  k  
ksvpwm .A edq  k  ,sdq  k  ,w  (12)
where the weight update is expressed as follows,
where w is an weight vector while A(·) represents the action
1
'w  ª J p  w J p  w   P I º J p  w  V
T T
neural network. It is important to notice that recurrent RNN (19)
is used for training while forward FNN is used when it is ¬ ¼
implemented in FOC control. IV. PROPOSED MTPA ALGORITHM
The dynamic programming cost function for NN-based A. Methodology
FOC PMSM control is
Generally, there are two operating regions in PMSM

C idq  j   ¦ f
k j
J k  jU  edq  k  , j ! 0,0  J d 1 (13) drives: constant torque and constant power region. Figure 2
shows a typical torque versus speed graph with two
where γ is a discount factor and U(∙) the local cost or utility operating regions. The relation between power, torque
function. The objective of DP is to minimize the error of the versus speed is described as,
dq current. The utility function is defined as
Power VI T (20)
 
D
U edq  k  ª¬e  k   e  k  º¼
2
d
2
q where T is the motor torque and ω is the motor speed. Since
(14) the power is constant as shown in green curve in Fig. 2, this
^
ª¬id  k   idref  k º¼  ª¬iQ  k   iqref  k º¼
2
`
2 D
region is called constant power region. In the constant torque
region, many studies utilize a control called maximum
in which α is 1 for the motor drive application. torque per ampere (MTPA) control because it finds the most
optimal combination of dq currents to produce the demanded
Furthermore, the modified Levenberg-Marquardt MLM torque [10]. This technique satisfies not only to minimize the
algorithm is used to train the weights of proposed NN-FOC copper loss, but also to improve the overall driving distance,
controller. A MLM algorithm is a powerful tool that is used since less current is consumed.
to solve nonlinear least squares minimization [11-12]. Usual
nonlinear function is of the following special form
1 m 2
f  x ¦ rj  x  (15)
2 j 1
where x is comprised of n vectors while rj is the function of
x. Like other minimization method, the proposed MLM
algorithm also uses iterative technique to find the minimized
value. To accomplish this task, the MLM interpolates
between the steepest descent method and the Gauss-Newton
algorithm. The main objective of this method is to find the
most optimal minimized solution.
To apply the MLM in the PMSM motor application, the
first step is rearranging the cost function, C(∙) (13), into the
sum-of-squares as (15). The equation becomes
Figure 2. Torque versus speed curve of PMSM with other descriptions

698
 2018 4th International Conference on Green Technology and Sustainable Development (GTSD)
In the constant power region, a control called flux
weakening control is used. This control sacrifices the current
to increase a motor’s speed range by producing the flux
against the flux that is produced by the magnet. This control
is very important because usually, without this control in
HEV application, the maximum vehicle speed is only around
50 or 60 mph. The main objective is applied NN-FOC to
translate the reference torque into optimal d- and q-axis
currents for MTPA control and into adequate d- and q-axis
currents for FW control.
If the speed exceeds the base speed, the motor operation
is in the second section. This section operates between the
base speed and ω2, which is a rated speed. The rated speed is
a speed where voltage-limit ellipse, current-limit circle, and
constant torque curve meet. When the motor is in this
region, MTPA control cannot be used. Instead, the reference
torque moves along the voltage-limit ellipse within the
current-limit circle to prevent the back-EMF to exceed the
maximum voltage. Therefore, optimal control cannot be
used.
B. Proposed MTPA Algorithm
This paper proposes MTPA control since it uses both d-
and q-axis currents for control. Moreover it help to minimize
the copper loss, which is one of dominant losses in the
PMSM motor [10] and eventually it is quite simple relative
to other approaches.
The concept of MTPA control begins with the stator dq
voltage equations of PMSMs (4) and the electromagnetic
torque equation (5). The detailed explanations of these
equations were covered in previous section.
In case the motor is SPMSM, the electro-dynamic torque
equation is only dependent on the q-axis current because the
stator dq inductances are almost equal to each other; thus,
the reluctance torque, which is the second portion of (5), is
ignored. Thereby, the electro-magnetic torque of SPMSM
can be written as
Tem
p
2
test, parameter variation test, and flux weakening test. Note
that the initial load torque of 0.1 Nm is used.
First, the acceleration and deceleration test is conducted
as follows. At t = 0.56 s, the input speed is increased from 0
rad/sec to 100 rad/sec, and then stays at 100 rad/sec for 1
second. At t = 1.56s, the speed is decreased to 50 rad/sec and
retains its value for 1 second. At t = 2.56 s, the speed is
increased to 200 rad/sec and decreases again to 100 rad/sec
at t = 3.56 s. After the acceleration and deceleration test, the
load torque is varied as follows. At t = 4.36 s, the load
torque is changed from 0.1 to 0.3 Nm. At t = 5 s, the load
torque is dropped to 0.1 Nm. Eventually, the parameter
variation test is conducted as follows. At t = 5.5 s, the stator
resistance, synchronous inductance, and flux linkage
constant is changed as, the stator resistance is doubled while
the inductance and flux linkage constant is reduced 20%.
Lastly, the flux weakening control is tested for both
controllers. At t = 6.75 s, the reference speed is accelerated
at a rate of 150 rad/s until it reached just slightly below the
maximum speed, which is about 400 rad/s. At t = 9.75 s, the
reference speed was decreased to 200 rad/s to check whether
the control can return to normal control from the flux
weakening control.
The traction result of speed, dq currents, abc currents,
and normalized system voltage for both conventional PI and
proposed neural network NN controller are shown in
following figures (from Figure 3 to Figure 8), respectively.
As the figures show, both controllers perform similarly
until the flux weakening test. As Fig. 3 shows, when the
reference speed is about 250 rad/s, the conventional PI-FOC
control loses the stability. Even though the PI gains for
conventional control are changed in numerous times, none of
the PI gains could solve the problem. Nevertheless, unlike
the conventional control, proposed neural NN-FOC control
exhibits very stable speed traction in high speed region.
When the SPMSM motor reaches 400 rad/sec, as Fig. 4
shows, the d-axis current is producing the negative value to
prevent the back-EMF to exceed the maximum voltage as
O fd isq  (21) presented in Fig. 8.

A MTPA control for SPMSM machine is very efficient,

since the electromagnetic torque is only dependent on the q-
axis current, as shown in (21).
When the motor speed exceeds the base speed, the flux
weakening control is applied to prevent the back-EMF to
exceed the maximum allowable voltage. In this paper, real-
time flux weakening controls are focused rather than a priori
method because determining exact motor parameters in high
speed is difficult.
V. SIMULATION RESULTS
Aforementioned, total of four tests are conducted to
comparatively verify the performance of the proposed NN-
FOC controller and conventional PI-FOC controller. They
are acceleration and deceleration test, load torque variation Figure 3. Comparative speed traction results of SPMSM (simulation)

699
 2018 4th International Conference on Green Technology and Sustainable Development (GTSD)

Figure 4. Proposed NN-FOC control related to d- and q-axis current Figure 7. Conventional PI-FOC control related to abc current traction of
traction of SPMSM (simulation) SPMSM (simulation)

Figure 5. Conventional PI-FOC control related to d- and q-axis current Figure 8. Proposed NN-FOC control related to normalized amplitude of
traction of SPMSM (simulation) voltage of SPMSM (simulation)

VI. CONCLUSIONS
This paper proposes a novel neural-based field-oriented
control (NN-FOC) for a surface-mounted permanent magnet
synchronous machine (SPMSM) drive. The neural model is
trained by modified Levenberg-Marquardt MLM algorithm.
Furthermore the novel Maximum Torque per Ampere
(MTPA) and flux weakening (FW) approaches are
undertaken for PMSM-based electric vehicle (EV)
application. The proposed control is validated in simulation
tests and compared with a PI-based field-oriented control
(PI-FOC). The simulation results show that the proposed
neural NN-FOC controller demonstrates quite better
performance than the conventional PI-FOC control and then
provides an excellent traction method applied in advanced
electric drives.
Figure 6. Proposed NN-FOC control of abc current traction of SPMSM
(simulation) ACKNOWLEDGEMENTS
This paper is funded by Vietnam National Foundation for
Science and Technology Development (NAFOSTED) under
grant number MDT 107.01-2018.10 and by Vietnam
National University Ho Chi Minh City (VNU-HCM).

700
 2018 4th International Conference on Green Technology and Sustainable Development (GTSD)
REFERENCES
[1] T. Lee, C. Lin, and F. Lin, “An Adaptive H∞ Controller Design for
Permanent Magnet Synchronous Motor Drives,” Control Engineering
Practice 13, pp. 425-439, 2005.
[2] S. Li, T.A. Haskew, and L. Xu, “Conventional and Novel Control
[3] Designs for Direct Driven PMSG Wind Turbines,” Electric Power
[4] System Research, vol. 80, pp. 328-338, March 2010.
[5] A. Mishra, V. Mahajan, P. Agarwal, and S.P. Srivastava, “Fuzzy Logic
Based Speed and Current Control of Vector Controlled PMSM Drive,”
2012 2nd International Conf. on Power, Ctrl, and Embedded Sys, 2012.
[6] H. Kamel, H. Hasanien, and H. Ibrahim, “Speed Control of PMSM
using Fuzzy Logic Controller,” Electric Machines and Drives Conf.,
pp. 1587-1591, 2009.
[7] H. Lu, and M. Chang, “Online Speed Control of PMSM using Self-
constructing Recurrent Fuzzy Neural Network,” 2008 Int. Conf. on
Machine Learning and Cyber., vol.7, pp. 3857 – 3862, 2008.
[8] C. Xia, J. Zhao, Y. Yan, and T. Shi, “A Novel Direct Torque and Flux
Control Method of Matrix Converter-Fed PMSM Drives,”IEEE Trans.
on Power Electron., vol. 29(10), pp.5417-5430, 2014.
701
[9] M. Khan, M. Uddin, and M. Rahman, “A Novel Wavelet-Neural
Network-Based Robust Controller for IPM Motor Drives,” IEEE Trans.
Ind. Appl., vol. 49(5), pp. 2341-2351, Sep./Oct. 2013.
[10] J. Moreno, M. E. Ortuzar, and J. W. Dixon, “Energy-management
system for a hybrid electric vehicle, using ultra-capacitors and neural
networks,” IEEE Trans. on Ind.Electron., vol. 53(2), pp. 614-623, Apr.
2006.
[11] Z. Chen, C. C. Mi, J. Xu, X. Gong, and C. You, “Energy management
for a power-split plug-in hybrid electric vehicle based on dynamic
programming and neural networks,” IEEE Trans. on Veh. Tech., vol.
51(3), pp. 2321-2330, Oct. 2014.
[12] S. Jung, J. Hong, and K. Nam, “Current Minimizing Torque Control of
the IPMSM using Ferrari’s Method,” IEEE Trans. On Power Electron.,
vol. 28(12), Dec, 2013.
[13] M. A. Lourakis, “A Brief Descriptions of the Levenberg-Marquardt
Algorithm Implemented by levmar,” matrix, vol. 3, p. 2, 2005.
[14] X. Fu and S. Li, “Training Recurrent Neural Networks with the
Levenberg-Marquardt Algorithm for Optimal Control of a Grid-
Connected Converter,” IEEE Trans. on Neural Netw.and Learning
Syst., 2014.