2784 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO.

4, DECEMBER 2023

Torque Distribution Optimization for a Dual-Motor

Electric Vehicle Using Adaptive Network-Based
Fuzzy Inference System
Chi T. P. Nguyen , Bảo-Huy Nguyễn , Member, IEEE, João Pedro F. Trovão , Senior Member, IEEE,
and Minh C. Ta , Senior Member, IEEE

Abstract—The development of electric vehicles (EVs) has been Index Terms—4WD dual-motor, ANFIS, Electric vehicle (EV),
considered one of the most efficient ways to reduce the carbon Energy efficiency optimization, Torque distribution strategy.
footprint of the transportation system. Among battery EV designs,
a dual-motor configuration is introduced as a promising solution
to improve dynamic performance and energy efficiency. In this NOMENCLATURE
study, a novel energy management strategy framework based on
an Adaptive Network-based Fuzzy Inference System (ANFIS) is Variables
proposed for torque distribution optimization between two differ- C Battery capacity [Ah]
ent motors. At first, Dynamic Programming (DP) is employed to d Partial derivative/
find global optimization of the torque distribution. After training
with the DP-obtained data set, the ANFIS model can execute E Energy [kWh]
a torque distribution online. By minimizing the battery energy F Force [N]
consumption, the motor torque-speed solution pair is found under I Current [A]
representative driving cycles (only three cycles). In addition, the J Cost function/
best ANFIS model has been selected using a clustering technique, k Ratio/
goodness-of-fit metrics, and sensitivity analysis. This distinguishes
the optimization problem in this study from previously published M Mass [kg]
literatures. The simulation results show that over an unknown SoC State of charge [%]
Urban Dynamometer Driving Schedule (UDDS) cycle, the torque T Torque [Nm]
prediction using this ANFIS model achieves 98.3% of the bench- U Voltage [V]
mark DP result. As a result, the overall efficiency of the proposed u Control variable/
strategy is increased to 73.3%, which is 3.4% higher than that of
the rule-based method. Furthermore, the signal hardware-in-the- V Velocity [km/h]
loop (S-HIL) experiments validate the real-time prediction of the x State variable /
ANFIS-based approach. w State disturbance/
ω Angular speed [rad/s]
P Power [w]
Manuscript received 4 January 2023; revised 13 April 2023 and 5 June 2023; R Resistance [Ω]
accepted 7 June 2023. Date of publication 13 June 2023; date of current version μ Mean/
28 November 2023. This work was supported in part by Canada Research
Chairs Program under Grant 950-230672, in part by the Natural Sciences and σ Standard deviation/
Engineering Research Council of Canada under Grant RGPIN-2017-05924, and
in part by the FCT-Portuguese Foundation for Science and Technology Project Subscripts
under Grant UIDB/00308/2020. Paper no. TEC-00014-2023. (Corresponding _act Actual
author: Bảo-Huy Nguyễn.)
Chi T. P. Nguyen is with the Department of Electrical Engineering and _bat Battery
Computer Engineering, University of Sherbrooke, Sherbrooke, QC J1K 2R1, _g Generator
Canada, and also with the Thai Nguyen University, Thai Nguyen City 24000, _gear Gearbox
Vietnam (e-mail: [email protected]).
Bảo-Huy Nguyễn is with the CTI Lab4EV, School of Electrical and Elec- _m Motor
tronic Engineering, Hanoi University of Science and Technology, Hanoi 10000, _max Maximum
Vietnam (e-mail: [email protected]). _min Minimum
João Pedro F. Trovão is with the Department of Electrical Engineer-
ing and Computer Engineering, University of Sherbrooke, Sherbrooke, QC _OCV Open-circuit voltage
J1K 2R1, Canada, and also with the Polytechnic Institute of Coimbra _pre Predicted
(IPC-ISEC) and INESC Coimbra, 3030-199 Coimbra, Portugal (e-mail: _ref Reference
[email protected]).
Minh C. Ta is with the Department of Electrical Engineering and Computer _req Required
Engineering, University of Sherbrooke, Sherbrooke, QC J1K 2R1, Canada _roll Rolling
(e-mail: [email protected]). _tot Total
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TEC.2023.3285225. _trac Traction
Digital Object Identifier 10.1109/TEC.2023.3285225 _tran Transmission

0885-8969 © 2023 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: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2785

_veh Vehicle human experience in different driving conditions, this method

_wheel Wheel is unable to produce optimal solutions.
_inv Inverter By contrast, the optimization-based methods have been de-
veloped for optimal control, which is based on minimizing an
objective function over time while respecting the constraints
I. INTRODUCTION [13]. The optimization-based EMSs can be classified into global
ECENTLY, the sale of electric vehicle (EV) is increasing optimization and real-time optimization. Dynamic Program-
R incredibly all over the world. This reflects the awareness
of the users involved in the greenhouse gas emission and their
ming (DP) and Pontryagin’s Minimum Principle (PMP) are
examples of the global optimization approaches. The use of
belief in EV systems. Therefore, EV techniques need to be the DP algorithm for the torque distribution control in the
rapidly developed for both dynamic performance and energy HEVs was presented in [14], [15]. PMP was also employed
economy [1]. In the early stage, due to the simplicity, good for the EMSs of multi-motor systems to minimize the energy
reliability, and low cost, single-motor single-gear powertrains consumption of the battery in the studies [16]. Although the DP
were mainly adopted in battery EVs (BEVs) [2]. However, and PMP methods can return globally optimal solutions, a priori
the fixed-speed transmission limited the vehicle dynamics. In knowledge of the entire driving cycle is required. Additionally,
addition, as presented in [3], operating points distributed in an since the optimization process is based on the discretization of
efficiency map of the motor cannot be adjusted since the driving state variables and input variables, the high computational cost is
conditions identify them. This results in low average efficiency challenging on many DP-based applications. Consequently, it is
of the system. To deal with this issue, multi-speed transmission not possible to implement them in real-life scenarios. However,
has been added to the powertrain. However, the multi-speed their solutions can be used as a benchmark to design online
transmission has some drawbacks in terms of extra system control rules or to evaluate the performance of other strategies
weight, mechanical loss, and manufacturing cost, which limit [17]. On the other side, real-time optimization divides a global
the system efficiency improvement. As an alternative, BEVs optimization problem into local optimization sub-problems. The
driven by multiple motors have been introduced for higher aver- reduction in computational effort makes it applicable in real-
age energy efficiency and better dynamic performance [4], [5]. time executions. The instantaneous optimization methods are
Among them, a dual-motor powertrain, with proper mechanical well-documented such as Equivalent Consumption Minimum
coupling and an optimal energy management strategy (EMS), Strategy (ECMS) and Model Predictive Control (MPC). By the
still can have high overall efficiency. According to Mutoh et al. use of the ECMS in [18], [19], the energy efficiency of the vehicle
[6], global efficiency is increased by making use of different can be improved, which is close to the global optimization over
high-efficiency regions of two independent motors. Using the unseen driving conditions. Since the equivalent factor of the
same two PMSMs can contribute to an efficiency improvement algorithm is sensitive to the driving cycles, it is not easily to
as presented in [7]. Besides, the dual-motor model can deliver design an adaptive equivalent factor. The MPC is a popular
superior drivability once the gear ratios connected to the motors real-time control strategy that can be seen as a compromise
are correctly designed as discussed in [8]. The results in [9] show solution between the DP and ECMS [20], [21]. In principle, its
that the dual-motor increased the overall efficiency compared to control decisions are obtained by calculating the cost function
the traditional single-motor powertrain. In fact, a large number in real time. Consequently, a heavy computational load can
of commercial EVs are being expected under the dual-motor challenge the hardware processor.
all-wheel-drive configurations including the Audi E-Tron Sport- So far, the main challenges of optimization-based methods
back, BMW iX3, Model XC40 of Volvo, and Model S from Tesla are between the complexity of online computations and the
[10]. hardware capability. With the development of artificial intel-
The dual-motor configuration brings a challenge to the EMS ligence technology, learning-based EMSs have achieved near-
aimed at minimizing energy consumption, while still ensuring global optimal solutions without any prior knowledge. Adnane
driving quality. Many strategies involved in torque distribution et al. [22] improved the energy performance of HEVs through
are developed for battery EVs and hybrid EVs (HEVs). Gen- a driving mode predictor based on supervised learning. In the
erally, they can be divided into two main groups: rule-based study of Lingxiong et al. [21], the deep Q-learning was found to
(RB), and optimization-based [11]. Deterministic control and solve MPC problem for a predictive EMS of a HEV. An indirect
fuzzy logic control (FLC) are the two main types of the RB reinforcement learning (RL) was introduced in two layers [23],
methods. By using the efficiency map of two motors, the torque [24]. In the offline state, an initial model was built based on
allocation ratio was determined through the look-up table. In training cycles. Then, a Q function was generated. In the online
addition, fuzzy rule-based control is more commonly adopted state, the RL controller chose the best actions by using the Q
in real-time EMSs. The benefit of the FLC is the efficient and function. Thus, the engine power was determined to minimize
simple implementation in multi-domain, and nonlinear systems. the fuel consumption of HEVs. Other deep RL algorithms, such
It can emulate the human experience to deal with uncertain con- as double deep Q-learning (DDQL) presented in [25], and deep
ditions in practice by using fuzzy rules. However, the torque split deterministic policy gradient (DDPG) presented in [26], were
ratio is realized by pre-defined membership functions and fuzzy also proposed to the power split problem of HEV and dual-motor
rules based on experts’ knowledge [12]. Due to the limitation of EV platforms. The benefits of learning-based methods include

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
2786 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO. 4, DECEMBER 2023

model-free and well-learning without the need for human in-

tervention. However, the large amount of computation required
can make RL difficult to be used in real-time. Also, it can be
difficult to design a reward function, which can result in biased
or ineffective policies.
Recently, the combination of the global optimization solution
and real-time performance in the EMS has been developed. In
other words, the results derived from the global optimization-
based methods are used to extract control rules considering the Fig. 1. Configuration of the dual-motor EV with the torque-split strategy.
real-time applications. In [27], two neural networks (NNs) were
used to generalize the future distribution of torque in a plug-in 2) In current literature, there is a lack of explanation of the
HEV, where the NNs were trained based on the DP results. Here, ANFIS architecture or details on its hyperparameters.
this learning system is seen as complicated because the authors 3) In terms of control variables, the torque distribution ratio
considered 11 variables for the NN1, while the NN2 included 7 is used in the reviewed literature. However, this ratio might
inputs. In the study of Silva et al. [28], the torque of the motor not produce the best efficiency. This is because different
was indicated by the fuzzy-logic power management control. As motors have different efficiency maps, and the ideal torque
pointed out, the fuzzy functions and the rules were optimized distribution ratio may not always correspond to the most
by the genetic algorithm (GA). Although the results demon- efficient operating points for each motor.
strated an improvement in driving range and battery life, like In this study, the framework is novel in the following four
optimization-based methods, GA is computationally expensive. aspects to fill in the aforementioned research gaps: 1) a rep-
For example, Yu et al. in [29], [30] proposed an optimal rule resentative dataset for all driving conditions is built. A small
extraction of EMS of a dual-motor configuration. The optimal dataset including urban, rural, and highway driving cycles is
working points obtained by DP were classified by using a self- used to represent general driving conditions; 2) using a clustering
organizing feature map (SOM) in [29]. As [30] presented a self- technique, goodness-of-fit metrics, and sensitivity analysis, a
learning NN, it can adaptively approximate the optimal control clear process is proposed for finding the best ANFIS model and
cost dependent on the DP actions. Nonetheless, a large amount hyperparameters; and 3) the high-efficiency operating points for
of historical data is required to provide a big training data. each motor are directly found. This is achieved by using the
Among these near-optimal solutions, there is a combina- motor torque-speed solution pair as the control variable instead
tion of NN and FLC introduced by Jang [31] in an adaptive of the torque split ratio.
network-based fuzzy inference system (ANFIS). In this way, The remainder of the article is organized as follows. The
the fuzzy rules are determined by training a NN based on configuration and the mathematical model of the dual-motor
the global solutions. Using a NN to extract control rules from configuration are fully described in Section II. The optimal
global optimization solutions improves precision. The ability of torque distribution optimization is acquired by DP method in
NNs to learn complex relationships between input and output Section III. The proposed ANFIS-based torque distribution is
variables allows ANFIS to produce more accurate rules. The designed and illustrated in detail in Section IV. Then, the simu-
second advantage is real-time performance, which is achieved lation and the S-HIL results are provided and discussed in Sec-
through the quick implementation of control rules. Furthermore, tion V and Section VI. Finally, Section VII concludes the article.
because NNs can adapt to changes in the system, the control
rules can be more flexible. Another advantage of ANFIS is that, II. THE STUDIED EV POWERTRAIN ARCHITECTURE
unlike machine learning algorithms, it does not require a large
amount of training data. This makes it an excellent choice for A. EV Configuration
applications where data is limited or expensive to obtain. For The architecture of the studied dual-motor configuration is
example, Tian et al. [19], [32] adopted the ANFIS for a real-time described in Fig. 1. The target vehicle is developed based on
EMS of a parallel hybrid electric bus. In [32], the training data the off-road e-Commander platform of the e-TESC laboratory
were collected from the optimal solution produced by an iterative at the University of Sherbrooke. The original topology consists
DP algorithm. The fuel consumption over the China typical city of an induction motor (IM) located in the front axle, driving two
bus driving cycle (CCBC) cycle for the iterative DP-ANFIS or four wheels. A dual-motor configuration was considered due
was smaller than that of the ECMS and rule-based method. to the poor dynamics and efficiency performance of using IM
Compared to the iterative DP, the fuel consumption increment alone. Among electric motors, permanent magnet synchronous
was only about 3.83%. motors (PMSMs) can produce high efficiency at high torque.
In comparison to other methods, the ANFIS method used in Meanwhile, the high-efficiency region of IM is spreading at
[19], [32] shows the potential for a real-time torque distribution high speed. Because IM and PMSM have different efficiency
strategy. However, based on the literature review, there are still distributions [33], [34], combining the two motors can benefit
the following research gaps: from higher energy efficiency [35]. This dual-motor topology
1) The existing ANFIS methods are focused on limited driv- provided superior driving performance with well-designed gear-
ing conditions. In both studies, only the CCBC - city bus box ratios, as demonstrated in our previous study [8]. The vehicle
routes in China have been considered. parameters are listed in Table I.

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2787

TABLE I
VEHICLE PARAMETERS

Fig. 2. Efficiency maps of two electric motors.

B. Vehicle Model
The traction force required at the wheels is determined by the are different. The maximal speed of IM reaches 8000 rpm,
acceleration force and the total resistance forces according to a while the maximal speed of PMSM is 5000 rpm. However, the
certain speed profile defined as: driving performance is successfully guaranteed with a two-speed
gearbox designed for PMSM.
1 2 By neglecting the impact of temperature change, the terminal
Freq = froll Mveh g cos θ + ρair Cx AVveh
2 battery voltage is defined by the open-circuit voltage and a
dVveh voltage drop across the equivalent resistance:
+ Mveh g sin θ + Mveh (1)
dt Ubat = UOCV (SoC) − Rbat Ibat (6)
where Mveh is the vehicle mass, Vveh the vehicle velocity, g the
gravity acceleration, froll the rolling resistance, θ the angle of the in which Rbat is the battery resistance, Ibat the battery current,
road, ρair the air density, Cx the aerodynamic drag coefficient, and the open-circuit voltage (UOCV ) is in a nonlinear relation-
and A the frontal area of this vehicle. ship with the state of charge (SoC) determined by:
 t
The total torque at the wheels is transmitted from the required 1
force given by: SoC = SoC0 − Ibat dt × 100% (7)
Cbat 3600 0
Ttot = Freq Rwheel (2) where SoC0 is the initial value of the battery SoC, and Cbat the
nominal capacity of the battery.
in which Rwheel is the wheel radius.
The power consumption of the battery is formulated as:
This torque is the sum of torques generated by the two motors
and then transferred to the vehicle’s wheels: Pbat = Ubat Ibat (8)
Ttot = TwheelIM + TwheelP M (3) Accordingly, Ibat is deduced from (6)–(8) by:

UOCV − UOCV 2 − 4Rbat Pbat
The torque and the speed at the wheels are described by the
Ibat = (9)
transmission system as follows: 2Rbat

Twheel = Tm kgear ηtran where the battery power can be presented in the relationship with
ω = ωm (4) the torque-speed and the efficiency of two motors given by:
wheel kgear
TIM ωIM TP M ωP M
where Tm is the motor torque (Nm), ωm the motor angular Pbat = + (10)
speed (rad/s), η tran the transmission efficiency, and kgear the ηinv ηIM ηinv ηP M
gear ratio. where ηinv = 0.95 is the efficiency of the inverter, TIM,P M
In the proposed configuration using two different motors, each denote the torque (Nm), ωIM,P M the angular speed (rad/s), and
motor can operate in two modes, motoring and braking. The ηIM,P M the efficiency of the IM and PMSM, respectively.
consumed or generated power of the motor is calculated by: For the sake of simplicity, we assume that the efficiency
 values of two inverters and the transmission are fixed because
km −1 (Tm ≥ 0)
Pm = T m ω m η m , with km = (5) the main goal is to optimize the efficiency of electric motors. The
1 (Tm < 0)
efficiency of two motors can be identified from their efficiency
where ηm is the motor efficiency. The motor efficiency can be maps which are provided by manufacturers. These maps give
identified through a static map of speed and torque as shown details on the motor efficiency at various torques and speeds. A
in Fig. 2. As we can see, the speed ranges of two motors more sophiticated dynamical lossess model of the motors and

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
2788 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO. 4, DECEMBER 2023

through the torque-speed relationship. Hence, Ebat is related to

the motor torque-speed pair. Because the speed of each motor
is determined by the vehicle’s speed through its own gear ratio,
each motor’s speed is fixed at each sample time. Meanwhile,
thanks to the torque distribution strategy, the torque of each
motor can be controlled independently. At the corresponding
speed, this torque control can move the operating points in the
(T − n) map. Ref. [36] analyzed a case when the vehicle speed
(Vveh ) and requested power (Preq ) were given. The gearboxes
of motor 1 and motor 2 defined the different speeds of each
motor. On a constant motor speed line of motor 1, the different
Fig. 3. Framework of the ANFIS-based torque distribution strategy. torque points illustrated the corresponding power and efficiency
values. This was similar to motor 2. As a result, while the total
amount of power remains constant, the efficiency value of the
inverters can be considered for future studies. The computational motor varies due to different speed-related torque distributions.
burden would, however, be significant. With the purpose of controlling the operating points of each
motor to reach the highest efficiency regions, the torque-speed
pair assigned to the rear motor is selected as the control variable:
III. TORQUE DISTRIBUTION OPTIMIZATION BY DYNAMIC
PROGRAMMING uk = {(TP Mk , nP Mk ) } (13)
A. Energy Management Strategy Framework where n is the rotational speed of motor in rpm.
Fig. 3 describes the framework of the torque distribution
strategy including offline and online stages. The offline stage, C. DP-based Torque Distribution Algorithm
shown in the dashed line boxes, carries out at first the global By discretizing variables, the DP algorithm deals with the
optimization based on the DP algorithm. The optimal motor cost function at every stage subjected to the constraints. Based
torque sequences produced by the DP over historical driving on Bellman’s equation, the optimal control policy is obtained by
cycles are collected as the database for ANFIS training. a backward calculation starting from the final step N :
The ANFIS model, after training, enables us to extract optimal ∗
parameters and control rules of the final FIS. This FIS is then JN (xN ) = min (L (xN , uN )) (14)
sent to the torque distribution strategy box in the real-time where N is the number of time steps (that is defined by the length
implementation, shown in the solid line boxes. The DP and of a driving cycle and step time).
ANFIS algorithms involved in this study are programmed and The k th stage cost function is:
trained in Matlab/SimulinkTM . The detailed description of the  
proposed strategy development is presented in the following. Jk∗ (xk ) = min L (xk , uk ) + Jk+1
∗
(xk+1 ) (15)
where L(xk , uk ) is the cost-to-go function from the state k th to
B. Problem Formalism the state (k + 1)th , and J ∗k+1 (xk+1 ) is the cost-to-go function
To optimize energy management, the current state of the from the state (k + 1)th to the end of the driving cycle.
system needs to be included in this problem. Among them, the In this study, vehicle speed and required force are considered
energy state variable referred to the battery SoC is selected to inputs, while the control variables are the output. The state
describe energy consumption. Two driving state variables, the variable (xk ) and the control variables (uk ) are defined in
vehicle speed, and the required force, are state disturbances to Section III-B. Two consecutive stages of the battery SoC as the
describe the vehicle dynamics. state variable are given by:
 ⎧  √ 2 
xk = [SoCk ] ⎪
⎪ UOCV − UOCV −4Rbat Pbat
(11) ⎨SoCk+1 = SoCk −
Wk = [Vvehk , Freqk ] 2Rbat Cbat
Δt (16)
⎪
⎪ × 3600 × 100%
The main purpose of EMS in this dual-motor powertrain is to ⎩
SoC0 = SoC (0)
reduce electric energy consumption. The battery’s energy con-
sumption (Ebat ) is calculated by integrating the battery power Finding the optimal torque-speed pairs to achieve the energy
as follows: consumption minimization over a specified driving schedule is
 t  t  the goal of the DP strategy for dual-motor topology. As a result,
TIM ωIM TP M ωP M
Ebat = Pbat dt = + dt (12) the minimum battery energy of each step can be considered as
0 0 ηinv ηIM ηinv ηP M the cost function at this stage, described by:
It shows that Ebat is proportional to the product of torque and  
Jk∗ (xk ) = min Ebat (xk , uk ) + Jk+1
∗
(xk+1 ) (17)
speed, and inversely proportional to the efficiency of the motors
and the inverters. The efficiency of the inverter is assumed to be Considering the battery energy in the cost function is a con-
fixed at 95%, whereas the efficiency of the motor is specified venient technique to find optimal energy solutions, because it

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2789

Rule 2: If x1 is A2 and x2 is B2 then y = p2 x1 + q2 x2 + r2

Layer 1 is called the fuzzification layer where every node (i)
is a square node. The output of this layer (Oi1 ) can be expressed
by:
 1
Oi = μAi (x1 )
(19)
Oi1 = μBi (x2 )
where x1 and x2 are the inputs of node i; Ai and Bi are the
linguistic labels; μAi and μBi are the membership functions
associated with the linguistic labels. If the μAi is a type of the
Gaussian membership function (MF).
μAi can be presented by:
2
−(x1 −ci )
2σ 2
μAi (x1 ) = e i (20)
where {σi , ci } parameter set is used to determine the shape of
the membership function, and they are referred to as premise
Fig. 4. ANFIS scheme.
parameters.
Layer 2 (rule layer): every node (i) in this layer is a circle node
and labeled Π. This layer uses the AND operator to multiply the
refers to the total amount of energy used from the battery during
incoming signals. The node output calculates the firing strength
a certain driving cycle. By optimizing this energy consumption,
(wi ) of a rule as follows:
the EMS aims to increase overall efficiency.
The physical constraints on state variables and control vari- Oi2 = wi = μAi (x1 ) × μBi (x2 ) (21)
ables are taken into account as follows:
⎧ Layer 3 (normalization layer): in this layer, the ith node, a
⎪SoCmin ≤ SoCk ≤ SoCmax
⎪
⎪ circle node labeled N, computes the ratio of the firing strength
⎪
⎨TP Mmin ≤ TP Mk ≤ TP Mmax of the ith rule to the sum of all rules’ firing strengths as given
TIMmin ≤ TIMk ≤ TIMmax (18) in:
⎪
⎪
⎪
⎪ ω P M ≤ ω P M ≤ ω P M wi
⎩ min k max
Oi3 = wi =
ωIMmin ≤ ωIMk ≤ ωIMmax w1 + w2
(22)

Due to the dynamics and all the constraints, it is possible where (wi ) denotes the normalized firing strength.
to complete all the calculations needed for DP in a reasonable Layer 4 (defuzzification layer): every node (i) in this layer
amount of time. However, if driving cycle length and grid density is a square node. The output of the nodes is determined by the
increase, the computation will quickly rise. To account for normalized firing strength (wi ) and the first-order polynomials
the computation’s precision and complexity, here we use three (yi ) as given in:
representative driving cycles with the time step Δt = 1s.
Oi4 = wi yi = wi (p1 x1 + q1 x2 + r1 ) (23)
IV. PROPOSED TORQUE DISTRIBUTION STRATEGY BASED ON where {pi , qi , ri } are the consequence parameters.
ANFIS Layer 5: each node is a circle node labeled ࢣ. The output of
A. The Principle of ANFIS this layer sums all incoming signals which are represented by:

For a comprehensive understanding, there are three blocks Oi5 = wi yi (24)

included in the ANFIS scheme in Fig. 4. The main component
The hybrid learning rule is applied for ANFIS to update the
is the five-layer ANFIS structure in block B. The corresponding
parameters in the adaptive nodes. In this manner, the gradient
FIS (the block A) is then delivered by the trained ANFIS, where
descent and the least-squares estimation methods are combined.
its parameters are well tuned by a NN. In block C, the hybrid
In each epoch, the hybrid learning process is carried out in both
learning process is illustrated.
the forward pass and the backward pass. In the forward pass,
In ANFIS architecture, a five-layer with two inputs (x1 and
the input data and functional signals go forward until layer 4.
x2 ) and one output (y) is considered. To represent different
The consequent parameters are identified by the least-squares
adaptive abilities, square and circle nodes are used in layers. The
method and the error measure is calculated, while the premise
square node referred to as the adaptive node includes parameters,
parameters are kept fixed. In the backward pass, the conse-
while the circle node referred to as the fixed node has no
quent parameters are fixed, and the error measure is propagated
parameters.
backward. The gradient descent is used to update the premise
Suppose that for a first order Sugeno FIS, two fuzzy if-then-
parameters. Accordingly, this hybrid learning algorithm not only
else rules are as follows:
reduces the convergence time but also prevents the local minima
Rule 1: If x1 is A1 and x2 is B1 then y = p1 x1 + q1 x2 + r1 issue of the gradient descent.

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
2790 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO. 4, DECEMBER 2023

Fig. 5. Relationship between V veh and T P M from the DP results. Fig. 7. Data and cluster centers with default SC’s parameters.

data, while the testing data is used to control the over-fitting of

the model. In this study, the first 3009 data pairs (80% of the
database) are used for training ANFIS, while the remaining 752
data pairs (20% of the database) are used for testing.
The optimal working points, as shown by the DP results, are
chaotic and complex (see red points in Fig. 7). This leads to
difficulties in initializing a FIS. For this reason, the subtractive
clustering (SC) method is employed. The purpose of SC is to
cluster a large data set into natural groups and then assign them
to cluster centers. Through the cluster centers, SC allows ANFIS
Fig. 6. ANFIS process. to determine the initial number of MFs and the corresponding
rules.
In the SC method, each data point is assumed as a potential
cluster center. The first cluster center is selected by the highest
B. Strategy Development Using ANFIS potential of the neighboring data points given by:
1) Analysis of the DP Results: Three driving cycles, includ- n
 
ing the New European Driving Cycle (NEDC), Worldwide har- xi − x2j
Pi = exp − (25)
monized Light-duty vehicles Test Cycles (WLTC) class 2, and j=1
(ra /2)2
ARTEMIS rural, are selected to solve the problem (17) by the DP
algorithm. Because these cycles represent urban, highway, and where ra is the neighborhood radius.
rural routes, DP-based laws for common rules to various driving After determining the first cluster center (xc1 ), the potential
patterns can be developed. The DP result of the relationship measure of each data point is revised by:
 
between torque of PMSM and speed, including the 3761 data ||xi − x2c1 ||
pairs in total, is shown in Fig. 5. Pi = Pi − Pc1 ∗ exp − (26)
(rb /2)2
From that, at very low speeds, TP M is preferable to be
zero. From about 10 to 30 km/h, the average value of TP M is where rb defines a neighborhood that has measurable reductions
optimal. When PM operates at high speeds, the torque values are in potential ( rb = 1.5ra ).
distributed near the limit. This torque allocation based on the PM According to (26), the data points near the first cluster center
speed ranges is used in this study for the RB method to compare xc1 will have reduced potential. The second one is selected with
the simulation results in the following section. However, the the highest remaining potential. This process is repeated until
approach faces difficulties in extracting intelligent rules for the all the data points are inside the radius of a cluster center.
various conditions. Therefore, the ANFIS-based technique is The number of clusters is automatically identified by four
developed in this study. parameters, such as cluster radius, squash factors, accept and
2) ANFIS Process: The ANFIS process consists of several reject ratios. The cluster radius (CR) has a strong impact on
steps. First, the required data is collected from DP sources. the number of cluster centers. If CR is small, the number of
Second, the initial FIS is created and trained using MatlabTM clusters is increased, meaning many cluster centers (many rules)
2021a. The training phase ends when the desired accuracy or the and vice versa. The squash factor (SF) is used to determine
maximum number of epochs is reached. Finally, the final model the neighborhood of a cluster center by multiple radius values.
is evaluated over training, testing data, and sensitivity analysis. The other criteria for cluster centers are accept and reject ratios.
The parameters of FIS are optimized after several trials. The Above the accept ratio (AR), another data point will be accepted
ANFIS process is described in Fig. 6. as a cluster center. Meanwhile, below the reject ratio (RR), a
To be specific, after implementing the DP algorithm over three data point is rejected as a cluster center. It is noted that the
known driving cycles, an input-output data set is collected. For different sets of parameters define different numbers of cluster
the use of the ANFIS training, the input-output variables are centers. For instance, there are three centers detected by the
normalized into the interval [0, 1]. Then, the normalized data default values of these parameters in MatlabTM (Fig. 7). In other
is divided into two parts: training data and testing data. The words, a FIS is generated with only three MFs and three rules.
training data is used to adjust the MF parameters to the best-fit It is not good enough to initialize a proper FIS. For this issue, a

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2791

selection of the SC’s parameters needs to be made around their TABLE II

ANFIS-SC SUB-MODEL PERFORMANCE METRICS (CR = 0.15-0.5, SF =
default values. The values of the CRs are examined between 1.2-1.4, AR = 0.5, RR = 0.15)
0.15 and 0.5, the SF from 1.2 to 1.4 [37], the AR from 0.4 to
0.5, and the RR from 0.01 to 0.15.
The clustered data is then imported into an ANFIS structure.
After that, the ANFIS is trained using the training data set. As
aforementioned, the hybrid learning algorithm is employed to
update the parameters of the FIS model. The training process is
done if the convergence criterion, or the maximum number of
epochs, is reached. TABLE III
ANFIS-SC SUB-MODEL PERFORMANCE METRICS (CR = 0.15, SF = 1.2,
To evaluate the training effectiveness of the ANFIS model, AR = 0.4–0.5, RR = 0.01-0.15)
the goodness-of-fit of the actual output and the reference output
is measured by various statistical metrics. Mean-square error
(MSE) and mean-square reference output (MSR) are given by:
N  2
i=1 TP M ref,i − TP M act,i
M SE = ,
N
N
i=1 TP M 2ref,i
M SR = (27)
N
where TP M ref and TP M act the reference and the actual values
of the PMSM torque, respectively.
The below MSPE and RMSPE describe the error of the model
in data prediction over the samples. Thereby, a lower value
implies a good model.
M SE Fig. 8. Data and cluster centers with selected SC17’s parameters.
M SP E = × 100%,
M SR

M SE
RM SP E = × 100% (28) ANFIS-SC17 model returns the best accuracy (MSPE, RMSPE,
M SR and R of 0.23%, 4.8%, and 0.989, respectively). As a result,
R, as calculated in the following equation, reflects the strength the SC17 is selected as the proper model for ANFIS training.
of the linear relationship between two variables. The R value that With this model, the number of cluster centers is indicated up
is closer to 1 implies a strong correlation between the reference to 29 (see Fig. 8), which corresponds to a FIS including 29
and the predicted values. MFs for each input and output. It should be noted that 29 MFs
N were chosen because the RMSPE reaches its minimum value
i=1 (TP Mref,i −TP Mref )(TP Mact,i −TP Mact )
R = 
2 N 2 without being overfit. This can ensure both training accuracy
i=1 (TP Mref,i −TP Mref ) × i=1 (TP Mact,i −TP Mact )
N

and strong prediction performance. More importantly, by using

(29)
SC technique, the selected ANFIS model produces only 29
where TP Mref and TP Mact are the average values of the refer- fuzzy rules. As a result, the model implementation is practical
ence and the actual torques of the PMSM, respectively. given the capability of the microcontroller thanks to reduced rule
calculation burden.
V. SIMULATION FOR COMPARATIVE EVALUATION Additionally, there are several hyperparameters for ANFIS
training. Each input variable in ANFIS with SC technique
A. Optimal ANFIS Model Selection
employs Gaussian MF “gaussmf” for each fuzzy cluster. Be-
Since the performance of the SC method is influenced by cause the model using a Sugeno-type fuzzy system, each output
four factors, two tests are carried out to select the proper model. variable is represented by the “linear” function. The number of
For the first test, AR and RR remain unchanged at their default epochs in ANFIS training is another hyperparameter. In general,
values of 0.5 and 0.15, respectively, whereas the CR and SF the greater the number of epochs, the better the fit and accuracy of
values are observed in the range of 0.15 to 0.5, and from 1.2 to the training data. Overfitting, however, can result from too many
1.4, respectively. As seen in Table II, it is indicated that the SC epochs. Therefore, a separate training and validation dataset
model using CR = 0.15 and SF = 1.2 provides greater accuracy should be used to determine the number of epochs. In this study,
(MSPE = 0.62%, RMSPE = 8.0%, R = 0.97). The impact of the minimal testing error value is 0.0356 at epoch 1000, while
two remaining factors, AR and RR, on ANFIS performance is the minimal training error value is 0.0029 also at epoch 1000.
investigated in the second test presented in Table III. In detail, It is pointed out that there is no over-fitting at the price of the
the AR value is increased from 0.4 to 0.5, and the RR is between training and testing errors, and thus the fine-tuned FIS at epoch
0.01 to 0.15. Meanwhile, the fixed values of CR 0.15 and SF 1000 is selected. After training, the final membership functions
1.2 which are selected by the previous test. It is found that the of two inputs are shown in Figs. 9 and 10.

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
2792 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO. 4, DECEMBER 2023

Fig. 9. Final membership functions for input speed.

Fig. 12. [Simulation] the vehicle velocity comparison over unknown UDDS.

Fig. 10. Final membership functions for input force.

Fig. 11. Elementary effects of speed and force on the output.

B. Sensitivity Analysis
A sensitivity analysis is then performed to consider the im-
portance of the input factors in the ANFIS model development.
To assess qualitative sensitivity, the elementary effect method is
used [38]. In this analysis, the practical input values are rescaled
within the range of 0 and 1 with an increment step Δ = 0.001.
The elementary effect of each input factor is determined by the Fig. 13. [Simulation] torque predictions of IM and PMSM based on ANFIS
corresponding partial derivative as follows: compared to DP and RB over unknown UDDS.

f (Vveh + Δ, Freq ) − f (Vveh , Freq )

d1 (Vveh ) = values are slightly lower (μ(x2 ) = 2.34, σ(x2 ) = 3.33). This
Δ
implies that the speed factor is more important and has more
non-linear impacts on the output. However, these sensitivity
f (Vveh , Freq + Δ) − f (Vveh , Freq ) measurements are not significantly different between the two
d2 (Freq ) = (30)
Δ factors. Thus, both proposed input factors effectively contribute
It can be seen in Fig. 11 that as both input factors change, to the development of the ANFIS model. This method can
the output function varies accordingly. Because the surface of provide an effective tool to select relevant variables for the
d1 (Vveh ) is rougher than the surface of d2 (Freq ), the effect of development of the ANFIS model.
the speed factor is more substantial than that of the force factor.
To quantify the sensitivity of each input, the sensitivity indices C. Comparative Simulation Results
are calculated by:
 To assess the effectiveness of the proposed torque distribution
r     r    2
1  j 
 1 strategy using ANFIS, the simulation results are compared with
μi = d i x i  , σ i =  di xji − μi the RB method as presented in Section IV.B.1, and the DP over
r j=1 r − 1 i=1
the unknown modified Urban Dynamometer Driving Schedule
(31) (UDDS). The results for the vehicle speed are shown in Fig. 12.
where μ is the mean, σ the standard deviation, x the input The measured speed tracks the reference accurately with negli-
variables, and r the number of sample points. gible tracking error.
The higher values of mean and standard deviation are found by The torques of two electric motors that are allocated by ANFIS
the speed factor (μ(x1 ) = 2.44, σ(x1 ) = 4.94). The force factor are presented in Fig. 13. In this figure, the blue-dashed line in
Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2793

Fig. 16. [Simulation] battery SoC comparison over unknown UDDS.

TABLE IV
PREDICTION VALIDATION OF ANFIS OVER UNKNOW CYCLES

Fig. 14. [Simulation] comparison of operating point distributions of PMSM

and IM over unknown UDDS.

Fig. 17. Signal-HIL experimental testbench.

Fig. 15. [Simulation] overall efficiency comparison over unknown UDDS.

to UDDS, the torque of PMSM is predicted more accurately

the torque of PMSM is indicated by the ANFIS generalization with only 11.9% RMSE using the Japan cycle, while the bigger
under the unknown cycle. difference is seen over the FTP and WLTC class 3 cycles.
Overall, a good agreement is achieved between the PMSM However, the strong agreement between reference and predicted
torque predicted by the ANFIS method and the reference of torque confirms the capacity of the proposed model to generate
the DP method. The mean absolute percentage error (MAPE) is data with different cycles.
used to evaluate the prediction accuracy of the PMSM torque as
follows: VI. SIGNAL-HIL SIMULATION FOR REAL-TIME ABILITY
N   EVALUATION
1  TP Mref,i − TP Mpre,i 
M AP E =   × 100% (32)
N i=1  TP Mref,i  The prediction performance of the proposed ANFIS is ver-
ified in real-time by the S-HIL simulation (see Fig. 17). The
As calculated by (32), the difference between the predicted main hardware setup includes a host PC, a TMS320F28335
and reference values is only 14.5%. It also can be seen that the digital signal controller (DSC) board, and the OP4510 real-time
remaining IM torque generated by the ANFIS closely follows simulator from OPAL-RT. In this simulation, the model of the
the DP optimization results. vehicle is built in the OP4510 simulator. In the meantime, the
As can be seen in Fig. 14, the operating points of IM and proposed control strategy is loaded onto the DSC board, and the
PMSM predicted by the proposed method tend to coincide with PC monitors the data and status of the vehicle model.
the corresponding points of the DP method. Meanwhile, the Fig. 18 depicts the S-HIL results with the unknown off-road
green points in both RB-generated efficiency maps are outside driving cycle (ORDC) built in [8]. The good velocity tracking
of the DP data. reveals that the vehicle dynamic performance is satisfied. Next,
As illustrated in Fig. 15, the overall efficiency using the the validation of the PMSM and IM torque predictions is deter-
proposed ANFIS is about 3.4% higher than that of the RB mined. When compared to the simulation results, the predicted
method. In another comparison, this efficiency value is only torques in the S-HIL testbench are well fitting. This demonstrates
1.3% less than the DP’s efficiency. Therefore, as seen in Fig. 16, the potential ability of the proposed ANFIS model to generate
the final SoC value is recorded at 97.58%, 97.54%, and 97.42% data under uncertain real-world cycles in real-time. In fact, we
corresponding to the DP, ANFIS, and RB methods, respectively. can observe acceptable forecast errors of IM and PMSM torques
Furthermore, the prediction ability of the torque of PMSM in Fig. 18.
based on the proposed ANFIS model is validated using three This is explained by unanticipated noises in data acquisition,
additional unknown cycles. The Table IV shows that, compared data transmission, and device communication. The proposed

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
2794 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 38, NO. 4, DECEMBER 2023

The simulation results under the unknown UDDS cycle have

demonstrated that the proposed ANFIS method can provide
accurate predictions for motor torque. The overall efficiency of
this method can reach 98.3% of the DP’s efficiency. Compared
to the RB method, the overall efficiency has been improved by
3.4%. In addition, the S-HIL experimental results have validated
the proper torque distribution by the proposed method in the real
controller layout using an automotive microcontroller. Thanks
to the good online performance, the proposed approach can
provide a promising solution to design the real-time EMSs for
multi-motor EV configurations using a reduced data set.

REFERENCES
[1] W. A. Lynch and Z. M. Salameh, “Realistic electric vehicle battery
evaluation,” IEEE Trans. Energy Convers., vol. 12, no. 4, pp. 407–412,
Dec. 1997.
[2] J. De Santiago et al., “Electrical motor drivelines in commercial all-electric
vehicles: A review,” IEEE Trans. Veh. Technol., vol. 61, no. 2, pp. 475–484,
Feb. 2012.
Fig. 18. [Experiment] prediction results over unknown ORDC. [3] V. Ivanov, D. Savitski, and B. Shyrokau, “A survey of traction control
and antilock braking systems of full electric vehicles with individually
controlled electric motors,” IEEE Trans. Veh. Technol., vol. 64, no. 9,
pp. 3878–3896, Sep. 2015.
[4] Y. Tian, N. Zhang, S. Zhou, and P. D. Walker, “Model and gear shifting
control of a novel two-speed transmission for battery electric vehicles,”
Mechanism Mach. Theory, vol. 152, 2020, Art. no. 103902.
[5] C. Zhang, S. Zhang, G. Han, and H. Liu, “Power management com-
parison for a dual-motor-propulsion system used in a battery elec-
tric bus,” IEEE Trans. Ind. Electron., vol. 64, no. 5, pp. 3873–3882,
May 2017.
[6] N. Mutoh and Y. Nakano, “Dynamics of front-and-rear-wheel-
independent-drive-type electric vehicles at the time of failure,” IEEE
Trans. Ind. Electron., vol. 59, no. 3, pp. 1488–1499, Mar. 2012.
[7] X. Yuan and J. Wang, “Torque distribution strategy for a front- and rear-
wheel-driven electric vehicle,” IEEE Trans. Veh. Technol., vol. 61, no. 8,
pp. 3365–3374, Oct. 2012.
[8] C. T. P. Nguyen, B.-H. Nguyễn, J. P. F. Trovão, and M. C. Ta, “Opti-
mal drivetrain design methodology for enhancing dynamic and energy
performances of dual-motor electric vehicles,” Energy Convers. Manage.,
vol. 252, 2022, Art. no. 115054.
[9] J. Wu, J. Liang, J. Ruan, N. Zhang, and P. D. Walker, “Efficiency com-
Fig. 19. [Experiment] operating point distribution of PMSM and IM over parison of electric vehicles powertrains with dual motor and single motor
unknown ORDC. input,” Mechanism Mach. Theory, vol. 128, pp. 569–585, 2018.
[10] “Electric car coming in 2020,” autotrader.ca. Accessed: Dec. 19, 2022.
[Online]. Available: https://www.autotrader.ca/newsfeatures/20200113/
here-is-every-single-electric-car coming-in-2020/
method demonstrates the efficiency improvement by dividing [11] Z. Wang, J. Zhou, and G. Rizzoni, “A review of architectures and control
each motor’s operating points into high-efficiency zones (see strategies of dual-motor coupling powertrain systems for battery electric
Fig. 19). The maximum and average values of the overall vehicles,” Renewable Sustain. Energy Rev., vol. 162, 2022, Art. no. 112455.
[12] E. Mohammadpour and M. N. Khajavi, “Modelling and control of power-
efficiency are determined at 83.8% and 77.2%, respectively. split hybrid electric vehicle using fuzzy logic method,” Int. J. Eng. Appl.
Therefore, the S-HIL experiment results prove that the proposed Sci., vol. 6, no. 1, pp. 1–13, 2014.
ANFIS-based torque distribution optimization is successfully [13] S. Barsali, C. Miulli, and A. Possenti, “A control strategy to minimize
fuel consumption of series hybrid electric vehicles,” IEEE Trans. Energy
integrated into the dual-motor configuration. Convers., vol. 19, no. 1, pp. 187–195, Mar. 2004.
[14] B. C. Chen, Y. Y. Wu, and H. C. Tsai, “Design and analysis of power
management strategy for range extended electric vehicle using dynamic
VII. CONCLUSION programming,” Appl. Energy, vol. 113, pp. 1764–1774, 2014.
[15] M. Kargar, T. Sardarmehni, and X. Song, “Optimal powertrain en-
In this study, a real-time torque allocation for the dual-motor ergy management for autonomous hybrid electric vehicles with flex-
powertrain has been developed based on ANFIS to improve ible driveline power demand using approximate dynamic program-
overall efficiency. Furthermore, a novel framework for EMS ming,” IEEE Trans. Veh. Technol., vol. 71, no. 12, pp. 12564–12575,
Dec. 2022.
has been presented. To be specific, a training data set was [16] S. Zhang, R. Xiong, and C. Zhang, “Pontryagin’s minimum principle-
first obtained by employing the DP method over a data set based power management of a dual-motor-driven electric bus,” Appl.
for only three cycles. After that, different ANFIS models were Energy, vol. 159, pp. 370–380, 2015.
[17] B. H. Nguyen, T. Vo-Duy, M. C. Ta, and J. P. F. Trovao, “Optimal energy
initialized and trained. The best ANFIS model for real-time management of hybrid storage systems using an alternative approach of
torque distribution implementation was chosen based on the Pontryagin’s minimum principle,” IEEE Trans. Transp. Electrific ., vol. 7,
SC technique, performance evaluations, and sensitivity analysis. no. 4, pp. 2224–2237, Dec. 2021.

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.
NGUYEN et al.: TORQUE DISTRIBUTION OPTIMIZATION FOR A DUAL-MOTOR EV USING ANFIS 2795

[18] J. Han, Y. Park, and D. Kum, “Optimal adaptation of equivalent factor of Chi T. P. Nguyen received the B.S. and M.S. de-
equivalent consumption minimization strategy for fuel cell hybrid electric grees in electrical engineering from the Thai Nguyen
vehicles under active state inequality constraints,” J. Power Sources, University of Technology, Thai Nguyen, Vietnam, in
vol. 267, pp. 491–502, 2014. 2014 and 2017, respectively. She is currently working
[19] X. Tian, R. He, X. Sun, Y. Cai, and Y. Xu, “An ANFIS-based ECMS toward the Ph.D. Program in electrical engineering
for energy optimization of parallel hybrid electric bus,” IEEE Trans. Veh. with the e-TESC Laboratory, University of Sher-
Technol ., vol. 69, no. 2, pp. 1473–1483, Feb. 2020. brooke, Sherbrooke, QC, Canada. Her research in-
[20] H. Borhan, A. Vahidi, A. M. Phillips, M. L. Kuang, I. V. Kolmanovsky, terests include electric machines, control engineer-
and S. Di Cairano, “MPC-based energy management of a power-split ing and energy management optimization for multi-
hybrid electric vehicle,” IEEE Trans. Control Syst. Technol., vol. 20, no. 3, motor multi-source EV applications. She was the
pp. 593–603, May 2012. recipient of the 2022 Premium Award for best paper
[21] L. Guo, H. Liu, L. Han, N. Yang, R. Liu, and C. Xiang, “Predictive energy in IET Electrical Systems in Transportation.
management strategy of dual-mode hybrid electric vehicles combining dy-
namic coordination control and simultaneous power distribution,” Energy, Bảo-Huy Nguyễn (Member, IEEE) received the
vol. 263, no. PA, 2023, Art. no. 125598. Ph.D. degree in electrical engineering from the Uni-
[22] M. Adnane, B. H. Nguyen, A. Khoumsi, and J. P. F. Trovao, “Driving mode versité de Lille, Lille, France, and the Université de
predictor-based real-time energy management for dual-source electric Sherbrooke, Sherbrooke, QC, Canada, in 2019. Since
vehicle,” IEEE Trans. Transp. Electrific ., vol. 7, no. 3, pp. 1173–1185, 2013, he has been with the Control Technique and
Sep. 2021. Innovation Laboratory for Electric Vehicles, School
[23] W. Zhang, J. Wang, Z. Xu, Y. Shen, and G. Gao, “A generalized energy of Electrical and Electronic Engineering, Hanoi Uni-
management framework for hybrid construction vehicles via model-based versity of Science and Technology, Hanoi, Vietnam,
reinforcement learning,” Energy, vol. 260, no. 5988, 2022, Art. no. 124849. where he is currently a Lecturer. From 2019 to 2021,
[24] N. Yang, L. Han, C. Xiang, H. Liu, T. Ma, and S. Ruan, “Real-time he was also a Postdoctoral Fellow with the Université
energy management for a hybrid electric vehicle based on heuristic de Sherbrooke. His research interests include con-
search,” IEEE Trans. Veh. Technol ., vol. 71, no. 12, pp. 12635–12647, trol engineering, power electronics, machine drives, and their applications for
Dec. 2022. electric vehicles and inverter-based resource power systems. Dr. Nguyễn was
[25] G. Du, Y. Zou, X. Zhang, L. Guo, and N. Guo, “Energy management for the Technical Chair of the IEEE Vehicular Technology Society Motor Vehicles
a hybrid electric vehicle based on prioritized deep reinforcement learning Challenge 2021. He is the co-author of the paper received the IET Premium
framework,” Energy, vol. 241, 2022, Art. no. 122523. Awards 2022.
[26] K. Zhang, J. Ruan, T. Li, H. Cui, and C. Wu, “The effects investigation of
data-driven fitting cycle and deep deterministic policy gradient algorithm
on energy management strategy of dual-motor electric bus,” Energy, João Pedro F. Trovão (Senior Member, IEEE) re-
vol. 269, 2023, Art. no. 126760. ceived the M.Sc. and Ph.D. degrees in electrical
[27] Z. Chen, C. C. Mi, J. Xu, X. Gong, and C. You, “Energy management for engineering from the University of Coimbra, Coim-
a power-split plug-in hybrid electric vehicle based on dynamic program- bra, Portugal, in 2004 and 2013, respectively. From
ming and neural networks,” IEEE Trans. Veh. Technol., vol. 63, no. 4, 2000 to 2014, he was a Teaching Assistant and an
pp. 1567–1580, May 2014. Assistant Professor with the Polytechnic Institute of
[28] S. F. da Silva, J. J. Eckert, F. C. Corrêa, F. L. Silva, L. C. A. Silva, and Coimbra—Coimbra Institute of Engineering, Portu-
F. G. Dedini, “Dual HESS electric vehicle powertrain design and fuzzy gal. Since 2014, he has been a Professor with the
control based on multi-objective optimization to increase driving range Department of Electrical Engineering and Computer
and battery life cycle,” Appl. Energy, vol. 324, 2022, Art. no. 119723. Engineering, University of Sherbrooke, Sherbrooke,
[29] X. Yu, C. Lin, M. Zhao, J. Yi, Y. Su, and H. Liu, “Optimal energy QC, Canada, where he currently holds the Canadian
management strategy of a novel hybrid dual-motor transmission system Research Chair position in efficient electric vehicles with the Hybridized Energy
for electric vehicles,” Appl. Energy, vol. 321, 2022, Art. no. 119395. Storage Systems. He is also the Founding Member and the Director of the
[30] X. Yu, C. Lin, Y. Tian, M. Zhao, H. Liu, and P. Xie, “Real-time and hier- electric-Transport, Energy Storage and Conversion Lab, University of Sher-
archical energy management-control framework for electric vehicles with brooke. He is the author or co-author of more than 190 journal articles and
dual-motor powertrain system,” Energy, vol. 272, 2023, Art. no. 127112. conference papers. His research interests include electric vehicles, hybridized
[31] J.-S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference sys- energy storage systems, energy management, and rotating electrical machines.
tem,” IEEE Trans. Syst., Man, Cybern ., vol. 23, no. 3, pp. 665–685, He was the General Chair of the 2018 IEEE Vehicle Power and Propulsion
May/Jun. 1993. Conference, Chicago, IL, USA. He was the Guest Editor of the Special Issue of
[32] X. Tian, R. He, and Y. Xu, “Design of an energy management strategy IET Electrical Systems in Transportation on energy storage and electric power
for a parallel hybrid electric bus based on an IDP-ANFIS scheme,” IEEE sub-systems for advanced vehicles and the Special Issue of IEEE TRANSACTIONS
Access, vol. 6, pp. 23806–23819, 2018. ON VEHICULAR TECHNOLOGY on electric powertrains for future vehicles and on
[33] Y. Hu, S. Zhu, C. Liu, and K. Wang, “Electromagnetic performance advanced vehicle power propulsion systems. He is also the Senior Editor of the
analysis of interior PM machines for electric vehicle applications,” IEEE automotive electronics topic of the IEEE VEHICULAR TECHNOLOGY MAGAZINE.
Trans. Energy Convers., vol. 33, no. 1, pp. 199–208, Mar. 2018.
[34] S. M. N. Ali et al., “Robust sensorless control against thermally degraded Minh C. Ta (Senior Member, IEEE) received the
speed performance in an im drive based electric vehicle,” IEEE Trans. Ph.D. degree in electrical engineering from Laval
Energy Convers., vol. 35, no. 2, pp. 896–907, Jun. 2020. University, Quebec City, QC, Canada, in 1998. From
[35] K. Li, A. Bouscayrol, S. Cui, and Y. Cheng, “A hybrid modular cascade 1998 to 2004, he was with Kyushu University,
machines system for electric vehicles using induction machine and perma- Fukuoka, Japan, The University of Tokyo, Tokyo,
nent magnet synchronous machine,” IEEE Trans. Veh. Technol ., vol. 70, Japan, and NSK Steering Systems Ltd. Co., Japan.
no. 1, pp. 273–281, Jan. 2021. In 2009, he became an Associate Professor with the
[36] K. Kwon, M. Seo, and S. Min, “Efficient multi-objective optimization of Hanoi University of Science and Technology, Hanoi,
gear ratios and motor torque distribution for electric vehicles with two- Vietnam, and was the Founding Director of the Center
motor and two-speed powertrain system,” Appl. Energy, vol. 259, 2020, for Technology Innovation during 2009–2018. He
Art. no. 114190. is currently with the Department of Electrical and
[37] O. Adeleke, S. A. Akinlabi, T. C. Jen, and I. Dunmade, “Prediction Computer Engineering, University of Sherbrooke, Sherbrooke, QC, Canada. His
of municipal solid waste generation: An investigation of the effect of main research interests include motor drives and advanced control techniques
clustering techniques and parameters on ANFIS model performance,” and their applications for electric vehicles and energy conversion systems. Dr.
Environ. Technol., vol. 43, pp. 1634–1647, 2022. Ta was the recipient of the Second Prize Paper Award of the IEEE IAS Industrial
[38] D. M. King and B. J. C. Perera, “Morris method of sensitivity analysis Drives Committee in 2001, Best Paper Award of the IEEE VPPC 2018, 2012
applied to assess the importance of input variables on urban water supply NSK Patent Awards, and 2017 Nagamori Awards. He was the General Chair of
yield - A case study,” J. Hydrol., vol. 477, pp. 17–32, 2013. the IEEE VPPC 2019, Hanoi.

Authorized licensed use limited to: Chung-ang Univ. Downloaded on April 26,2024 at 12:42:59 UTC from IEEE Xplore. Restrictions apply.