Hindawi

Mathematical Problems in Engineering

Volume 2020, Article ID 1204260, 21 pages
https://doi.org/10.1155/2020/1204260

Research Article
Optimal Torque Split Strategy of Dual-Motor Electric Vehicle
Using Adaptive Nonlinear Particle Swarm Optimization

Qingxing Zheng ,1,2,3,4 Shaopeng Tian ,1,2,3,4 and Qian Zhang 1,2,3,4

1
School of Automotive Engineering, Wuhan University of Technology, Wuhan 430070, China
2
Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology,
Wuhan 430070, China
3
Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan 430070, China
4
Hubei Research Center for New Energy & Intelligent Connected Vehicle, Wuhan University of Technology,
Wuhan 430070, China

Correspondence should be addressed to Shaopeng Tian; [email protected]

Received 16 January 2020; Revised 30 March 2020; Accepted 7 April 2020; Published 21 May 2020

Guest Editor: Chi-Hua Chen

Copyright © 2020 Qingxing Zheng et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
In order to exploit the potential of energy saving of dual-motor powertrain over single-motor powertrain, this paper proposes a
time-efficient optimal torque split strategy for a front-and-rear-axle dual-motor electric powertrain. Firstly, a physical model of
electric vehicle powertrain is established in Matlab/Simulink platform and further validated by real-vehicle experiments.
Subsequently, a three-layer energy management strategy composed of demanded torque calculation layer, mode decision layer,
and torque split layer is devised to enhance the total operating efficiency of two motors. Specifically, the optimal torque split
strategy using adaptive nonlinear particle swarm optimization (ANLPSO) is embedded in the torque split layer. Finally, two
conventional strategies (even distributed strategy and rule-based strategy) for dual-motor powertrain are considered for
comparison to verify the efficacy of the proposed strategy. Tremendous results demonstrate that the dual-motor powertrain with
this proposed optimal torque split strategy develops energy saving by 11.88% and 12.18% against single-motor powertrain in the
NEDC and WLTP. Compared to two conventional torque split strategies, it is able to reduce the total motor loss by 12.17% and
8.1% in NEDC and 11.91% and 8.07% in WLTP, respectively, which indicates the prominent optimization performance and a
great potential in realistic applications.

1. Introduction green transportation and sustainable future mobility [6].

However, BEVs are still facing driving range concern,
To relieve increasing burden caused by environmental mainly owing to the long charging time and unsatisfactory
pollution and energy shortage, electric vehicles (EVs) energy density of power battery when compared to gasoline
emerged as a promising approach all over the world [1]. and diesel [7].
While the energy efficiency of internal combustion engine
(ICE) vehicles is merely from 15% to 20%, typical EVs utilize
50–75% of the energy effectively to propel the vehicle [2, 3]. 1.1. Literature Review. To improve the driving range of the
There are three main categories of EVs, including hybrid conventional single-motor-driven BEVs, from the per-
electric vehicles (HEVs), fuel cell electric vehicles (FCHEVs), spective of powertrain topology and control, various pow-
and battery electric vehicles (BEVs) [4], among which BEVs ertrain structures and relevant control strategies have been
have the best “tank-to-wheel” efficiency [5]. Furthermore, proposed and investigated. These advanced configurations
the salient upsides of no CO2 emissions and other pollutants offer additional degrees of freedom to achieve more efficient
lead BEVs to be one of the most promising alternatives for torque distribution and independent control of each driving
2 Mathematical Problems in Engineering

shaft/wheel. The literature review mainly focuses on two driven by two motors, respectively (shown in Figure 1(a)). It
following aspects: powertrain topology and energy man- exhibits great enhancement of drivability and potential of
agement strategy. energy saving [21]. Due to the low manufacturing cost,
simple architecture, energy efficiency improvement poten-
tial, and great dynamics performance, the front-and-rear-
1.1.1. Powertrain Topology. According to the number of axle dual-motor powertrain is investigated in this paper.
motors used in the vehicle and their configuration, existing
powertrain topology of BEVs can be classified into two types
including the centralized single-motor-driven powertrain 1.1.2. Energy Management Strategy. In dual-motor power-
and the distributed multimotor-driven powertrain. train, energy management strategy plays an important role
The centralized single-motor one-speed-driven power- in power split among two motors and energy saving. Hu
train is the most common structure in modern BEVs. As the et al. proposed the multimode control strategy and power
motor substitutes the ICE to propel the vehicle through a split strategy for a dual-motor powertrain with a planetary
one-speed reducer, it can be easily installed on the con- gear. The results show great improvement in economic
ventional chassis without excessive modification. Compared performance and driving range [18]. Urbina Coronado et al.
to one-motor multispeed powertrain, this kind of power- designed a two-parameter rule-based control strategy for a
train is known as the simplest structure, whereas it would dual-motor powertrain. The results indicate that a combi-
increase the working burden of both motor and controller. nation of high-speed transition threshold and high torque
Moreover, it has relatively low energy efficiency [8]. To transition threshold could achieve the best results in energy
address such drawbacks, some researchers proposed the saving and drivability [22]. In previous study [23], theo-
single-motor with multispeed gearbox powertrain. The re- retical solutions of power loss optimization were conducted
sults indicated that multispeed gearbox could downsize the for a front-rear-induction-motor EVs. It is confirmed that
motor and improve the overall energy efficiency by about single-motor mode takes priority over dual-motor mode
5–12% through the improvement of motor operating points under required low torque condition, while even torque
[9, 10]. However, this powertrain is rarely used in practice distribution is the optimal control in dual-motor mode
because the additional multispeed gearbox might cause the under required middle or high torque condition. Yuan and
increase in cost and control difficulty of shift smoothness Wang investigated the optimal torque distribution strategy
(especially with the increase in gears) [11]. for an EV with two identical PMSMs to improve motor
Recently, some researchers have focused on distributed efficiency in the region of low torque and high speed. This
multimotor-driven powertrain. Existing multimotor-driven study reveals that the motor efficiency could be significantly
powertrain is mainly classified into three categories [12]: improved in low torque region when one motor is operating
dual-motor powertrain, triple-motor powertrain (shown in and the other motor is fully turned off through a clutch. And
Figure 1(c)) [13], and four-motor powertrain (shown in in the region of middle or high torque, the required torque
Figure 1(d)) [14–16]. Specially, dual-motor powertrain has should be equally distributed to minimize power loss when
been actively investigated as a promising alternative to only the longitudinal dynamics is considered [24]. Wang
single-motor powertrain. Hu et al. studied a novel dual- et al. suggested a mode decision system based on fuzzy logic
motor powertrain in which the speed of two motors is for a dual-motor powertrain. The frequent mode switching is
coupled through planetary gear and the torque is coupled reduced by the defined offset and compensation for mode
through shaft-fixed gear. The simulation shows that this switching line [25]. Zhao et al. developed a novel rule-based
powertrain could improve energy efficiency effectively [17]. strategy extracted from DP optimization results. The sim-
However, the speed and torque coupling devices increase ulation results show that the improved strategy could reach
structure complexity, manufacturing cost, and control dif- 95% energy-saving effect of that under DP [26].
ficulty. Gao et al. researched a practical dual-motor coaxial
powertrain for battery electric buses with dynamic pro-
gramming (DP) method, which could eliminate power in- 1.2. Motivation and Innovation. This study was carried out
terruption and realize a great reduction of energy in order to prove the potential of energy efficiency im-
consumption [18]. Actually, it is one kind of planet gear- provement of a front-and-rear-axle dual-motor powertrain
based coupling system, and additional coupling devices, over single-motor powertrain. In aforementioned literature,
such as clutches and synchronizers, are still needed, which tremendous research studies were conducted on control
burden the mechanism design and control. In previous study strategy for front-and-rear-axle-driven powertrain. It is
[19], there is another type of powertrain with two wheel-hub concluded that the switching threshold between single-
motors directly driving the left/right wheel through a one- motor mode and dual-motor mode is related to motor speed.
speed reducer, shown in Figure 1(b). But this powertrain is In practice, the motor speed varies with unexpected road
rarely used in practice. Ruan and Song devised a novel dual- conditions, which requires the switching threshold and
motor two-speed powertrain as one of the parallel direct torque split strategy to be optimized in real time for better
drive layouts for BEVs. This structure could achieve a great control. However, a vast majority of existing torque split
efficiency improvement and smooth shifting process without strategies are designed based on theoretical analysis or
the increase in mechanical complexity [20]. Kang et al. empirical rules, using a fixed torque switching threshold for
proposed an electric powertrain with the front and rear shaft control simplicity but with the sacrifice of optimality.
Mathematical Problems in Engineering 3

E E E E E E
Differential
M M M M M M

EM

Battery Battery Battery Battery

EM EM

Differential Differential E E
M M

(a) (b) (c) (d)

Figure 1: BEV powertrains.

To address this research gap, this paper developed a studied in this paper because the vehicle vertical vibration
time-efficient optimal three-layer energy management and vehicle yaw have little influences on energy efficiency.
strategy based on adaptive nonlinear particle swarm opti- Generally, the overall resistance force over wheels consists of
mization (ANLPSO). The main contributions incorporate air resistance, rolling resistance, and climbing resistance.
three aspects. Firstly, the optimal torque split strategy was However, in this study, the climbing resistance is neglected
established to search the best torque split ratio of two motors since the driving cycles are horizonal. Hence, the final re-
instantaneously and thus achieve the optimal energy effi- sistance force over wheels is calculated by equations shown
ciency. Secondly, the algorithm architecture and main below:
influencing factors of PSO were improved to construct the Cd Av(k)2 J a(k)
proposed ANLPSO, which could perform a tradeoff between Fr � + mgCr + t 2 , (1)
optimality and real-time applicability. Eventually, a com- 21.15 r
prehensive comparison was conducted between two com-
Jt � Jv + Jw + JEM1 + JEM2 􏼁i2t , (2)
mon torque split strategies and the optimal one.
The organization of this paper is as follows: the pow-
ertrain architecture is described and modeled in Section 2. dv(k) T + TEM2 􏼁it ηt
m � EM1 − Fr , (3)
The three-layer energy management strategy and the real- dt r
time optimal torque distribution strategy are designed in where Fr is the overall resistance force over wheels; Cd is the
Section 3. In Section 4, the physical model is validated by air drag coefficient; A is the frontal area of vehicle; v is the
real-vehicle experiment, followed by the comparison and velocity of vehicle; k is the kth time step; m is the mass of
analysis of the dual-motor powertrain simulation results vehicle; g is the gravitational acceleration; Cr is the rolling
between even distribution strategy, the rule-based strategy in resistance coefficient; a is the longitudinal acceleration of
literature, and the proposed real-time optimal control vehicle; Jt is the total rotary inertia at the wheels; Jv is the
strategy. Finally, conclusions are summarized in Section 5. rotary inertia of vehicle chassis; Jw is the total rotary inertia
of four wheels; JEM1 and JEM2 are the rotary inertia of EM1
2. Powertrain Architecture and Model and EM2, respectively; TEM1 and TEM2 are the output torque
of EM1 and EM2, respectively; it is the transmission ratio; ηt
2.1. Electric Powertrain Configuration. Figure 1(a) illustrates
is the transmission efficiency; and r is the rolling radius of
the configuration of the front-and-rear-axle dual-motor
wheel.
electric powertrain. In this architecture, two identical
PMSMs are coupled to the front axle or the rear axle through
a one-speed gear set and a differential mechanism, re-
2.3. Driver Model. The driver model is designed based on
spectively. The target dual-motor BEV in this paper is a
feedforward-feedback PI controller. Based on the bias be-
commercial vehicle prototype (shown in Figure 2), with all
tween the desired velocity (given by driving cycle) and the
relevant parameters detailed in Tables 1–3.
actual velocity, acceleration or braking command will be
given by the driver model. To improve the control precision
2.2. Vehicle Longitudinal Dynamics Model. This research and reduce the fluctuation of driver commands, a feedfor-
focuses on the energy efficiency of the dual-motor power- ward control module based on desired velocity, a feedback
train. Thus, only the longitudinal dynamics vehicle model is control module based on driver commands, and an
4 Mathematical Problems in Engineering

Figure 2: Target dual-motor BEV.

Table 1: Vehicle parameters of target vehicle. Table 2: Parameters of driving motor.

Vehicle characteristics Name Value
Vehicle mass (half load) (m) 2000 kg Type PMSM (ETG029)
Dynamic rolling radius (r) 0.273 m Rated/peak power 15/44.6 kW
Frontal area (A) 5 m2 Rated/peak torque 31/117 Nm
Transmission efficiency (ηt ) 0.9504 Maximum rotation speed 7500 rpm
Air drag coefficient (Cd ) 0.4 Nominal voltage 115 V
Rolling resistance coefficient (Cr ) 0.01 Peak efficiency 96%
Rotary mass coefficient (Jt ) 1.12
Gravitational acceleration (g) 9.8 m/s2
Final drive ratio (front/rear axle) 6.732 Table 3: Parameters of power battery.
Top speed 85 km/h
Name Value
Climbing capacity 20% at 25 km/h
Acceleration time (0–50 km/h) 7s Type LiFePO4
Capacity (C) 375 Ah
Nominal voltage 115.2 V
additional high-pass filter were added. The driver command Operating temperature range −20–55°C
can be calculated by the following equation: Mass 320 kg

Accpdl (k) � KFF Vd (k) + Kp Vd (k) − Va (k)􏼁 proportional gain, integral gain, and feedback module, re-
k spectively; Vd (k) is the desired velocity; Va (k) is the actual
+ 􏽚 Ki Vd (k) − Va (k)􏼁 + KFB Dc 􏼁dk, velocity; Dc is the driver command given by driver model;
0
and Brkprs is the demanded brake pressure.
Vd (k) − Va (k) ≥ 0,
(4)
2.4. Driving Motor Model. The driving motors in this
Brkprs (k) � Gbrk ∗ 􏼐KFF Vd (k) + Kp Vd (k) − Va (k)􏼁 powertrain consist of two identical PMSMs, which can
k
work as the driving motor to propel the vehicle or as a
+ 􏽚 Ki Vd (k) − Va (k)􏼁 + KFB Dc 􏼁dk􏼁, generator to recover braking energy. The data of motor
0 external characteristics and efficiency characteristics are
Vd (k) − Va (k) < 0, based on experimental data, which is shown in Figure 3.
(5) The output torque of motor is simplified as a function of
motor speed and motor torque command, as shown in
where Accpdl is the acceleration pedal command; equation (6). The efficiency characteristics are related to
Gbrk , KFF , Kp , Ki , and KFB are constant gain of the motor speed and motor output torque, as shown in
demanded brake pressure, feedforward module, equation (7):
Mathematical Problems in Engineering 5

100

85 88 90 91 92
80
93
50 75 93.5
70
Torque (N∗m)

70
0 93
90
88
85
91
75
80
93
70 93
–50 75 93
92
80 85 90 91
88

–100

0 1000 2000 3000 4000 5000 6000 7000

Speed (rpm)

Full load chracteristics (Nm)

Efficiency (%)
(a) (b)

Figure 3: (a) External characteristics and efficiency characteristics; (b) motor test bench.

l (k) × TchgmaxEMi nEMi (k)􏼁, lEMi (k) < 0; i � 1, 2, 2

⎪
⎧
⎨ EMi
TEMi (k) � ⎪ Pb (k) � 􏽘 PEMi (k) + I(k)2 R(SOC(k)), (9)
⎩ i�1
lEMi (k) × TdismaxEMi nEMi (k)􏼁, lEMi (k) ≥ 0; i � 1, 2,
(6) Pb (k)
I(k) � , (10)
UOC (SOC(k))
ηEMi (k) � f nEMi (k), TEMi (k)􏼁, i � 1, 2, (7)
k
nEMi (k)TEMi (k) 􏽒 Ib (k)dk (11)
⎪
⎧
⎪ SOC(k) � SOCini − 0 ,
⎪
⎪ , TEMi (k) ≥ 0; i � 1, 2, 3600C
⎪
⎨ 9550ηEMi (k)
PEMi (k) � ⎪ where Pb is the total battery power, I is the battery current,
⎪
⎪
⎪
⎪ n (k)TEMi (k)ηEMi (k) R(SOC(k)) is the internal resistance at SOC(k),
⎩ EMi , TEMi (k) < 0; i � 1, 2,
9550 UOC (SOC(k)) is the open circuit voltage at SOC(k), SOCini
(8) is the initial value of SOC, and C is the battery capacity.

where nEM is the motor speed, TEM is the output torque of 2.6. Driving Cycle. The driving cycle is a standard test pro-
motor, lEM is the motor torque command, TchgmaxEM (nEM ) is cedure in which the desired velocity is fixed with respect to
the maximum charge torque of motor at nEM , time. The driver is supposed to propel the vehicle to reach the
TdismaxEM (nEM ) is the maximum discharge torque of motor desired velocity within a certain error tolerance range to ensure
at nEM , ηEM is the efficiency of motor, and PEM is the electric accuracy of the test or simulation. Numerous driving cycles are
power of motor. proposed to evaluate energy consumption, fuel consumption,
and emissions for sorts of vehicles in different countries, such
as New European Driving Cycle (NEDC), Urban Dyna-
2.5. Battery Model. The electrochemistry reaction inside mometer Driving Schedule (UDDS), Japan Cycle 08 (JC08),
the battery is quite complicated, which is not the scope of World Light Vehicle Test Procedure (WLTP). In this research,
this study. Hence, the open equivalent circuit model [27] NEDC and WLTP are selected as the standard test cycle for the
is adopted, which reflects the relationship between open target vehicle. As mentioned above, the target vehicle is a
circuit voltage, inner resistance, and the state of charge commercial sample vehicle with a maximum velocity of 85 km/
(SOC). In this case, when the battery output power is h. Therefore, the portions with the velocity exceeding 85 km/h
given, the current of battery and SOC can be calculated as in NEDC and WLTP are limited to 85 km/h. Figure 4 shows
follows: the driving cycles with velocity limitation.
6 Mathematical Problems in Engineering

90 90

80 80

70 70

60 60
Velocity (km/h)

Velocity (km/h)
50 50

40 40

30 30

20 20

10 10

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200 1400 1600 1800
Time (s) Time (s)
Desired velocity Desired velocity
(a) (b)

Figure 4: (a) NEDC with velocity limitation; (b) WLTP with velocity limitation.

3. Optimal Energy Management Strategy Regenerative brake mode: if the battery SOC is below
charge upper limit (SOC_h) and the demanded brake
To optimize energy efficiency and dynamics performance, torque is nonzero and not greater than the available
energy management strategy plays a paramount role in maximum torque of two motors and the current ve-
power split for dual-motor powertrain. In this research, the locity is greater than a fixed value, the vehicle will
optimal energy management strategy consists of three layers: operate in the regenerative brake mode. In order to
demanded torque calculation (upper layer), vehicle oper- adapt to different braking requirements, the regener-
ating mode decision (decision layer), and real-time optimal ative brake mode is further divided into three sub-
torque split (lower layer), with the schematic diagram shown operating modes including energy recovery priority
in Figure 5. mode to maintain high regenerative braking efficiency,
braking performance priority mode to obtain maxi-
mum braking force, and safety priority mode for the
3.1. Demanded Torque Calculation. In the upper layer, the sake of safety in wet/snow road surface [28]. In this
demanded torque can be calculated based on the driver’s study, the energy recovery priority mode is the main
command, state of two motors, and sensor signals of hy- working mode for regenerative braking.
draulic brake. In detail, the demanded driving torque is
Hydraulic brake mode: considering vehicle safety,
determined by the acceleration pedal signal and the available
whenever the demanded brake torque exceeds the
maximum output torque of two motors, while the demanded
threshold defined by the ECE safety regulation or the
brake pressure is decided by driver’s brake command and
current velocity is smaller than a specified value, the
constant gain of the demanded brake pressure (as mentioned
hydraulic brake mode is switched on.
in Section 2.3).

3.2. Vehicle Operating Mode Decision. The decision layer is 3.3. Real-Time Optimal Torque Split Strategy. The lower layer
designed to define the vehicle operating mode based on is torque split strategy, composed of the optimal driving
driving torque command, brake pressure, SOC of battery, torque split strategy and braking torque split strategy.
and current velocity. The simple and robust “if-else” control However, in this paper, the proposed real-time optimal
strategy is adopted to yield the most appropriate operating driving torque split strategy plays the major role and would
mode for this powertrain through state machine. The mode be detailed further. At the same time, to explore its potential,
switching logic is illustrated in Figure 6. the even distribution strategy and the control strategy
proposed by some researchers would be included for
Electric mode: if the demanded driving torque is comparison in terms of energy efficiency.
greater than zero and the battery SOC is greater than
discharge lower limit (SOC_l), the vehicle is in electric
mode. And the magnitude of the demanded driving 3.3.1. Baseline Driving Torque Split Strategy. For dual-motor
torque determines whether the vehicle is driven by a powertrain, the simplest and most common torque split
single motor or two motors, which is further deter- strategy is even distribution strategy, in which the demanded
mined by the torque split strategy. driving torque is always distributed equally to two motors to
Mathematical Problems in Engineering 7

Optimal energy S Current

management strategy O velocity
C

Upper layer Decision layer Lower layer

Vehicle operating Torque split strategy

mode decision
Motor maximum TEM1
Motor speed Real-time optimal driving
available torque torque split strategy based
calculation Electric mode TEM2
on ANLPSO
Energy recovery
priority
Regenerative
Acceleration pedal Demanded Braking TEM1,reg
brake mode
torque performance priority
Brake pedal TEM2,reg
calculation Safety priority
Hydraulic Hydraulic brake
brake mode Pure hydraulic pressure
braking

Figure 5: Schematic of optimal energy management strategy.

Key signal
Start_switch > 0 No Parking mode

Yes

Demanded
brake pressure No SOC > SOC_l Yes Torque_required > 0 Yes Electric mode
>0

Yes No Charge mode

Demanded
Regenerative
brake pressure No SOC < SOC_h Yes Current_velocity > 5 Yes
brake mode
> 0.7

Strategy

Yes No No Energy Braking

recovery performance Safety priority
priority priority

Hydraulic brake
mode

Figure 6: Mode switching logic.

propel the vehicle. Thus, even distribution strategy is defined motors. To define a proper torque threshold T switch, the
as the baseline torque split strategy. optimization procedure is proceeded using the genetic
In the previous research, another critical strategy is the algorithm (GA), in which T switch is selected as the
rule-based torque split strategy, shown in Figure 7. If the optimization variable and the energy consumption of
demanded driving torque is smaller than a preset torque typical cycles simulation is chosen as the optimization
value T switch, the demanded driving torque will be objective. The optimal T switch (102.6 Nm) is obtained
totally distributed to a single motor to propel the vehicle. after 100 iterations. As a contrast, it is also researched in
Else, the driving torque is distributed equally to two this paper.
8 Mathematical Problems in Engineering

Start

No T_required > T_switch

Yes
EM1 and EM2
EM2 only
Even distributed strategy

End

Figure 7: Rule-based torque split strategy.

3.3.2. Real-Time Driving Torque Split Strategy Based on Objective Function. For the given demanded driving
ANLPSO torque and motor speed at time k, the main purpose
of the optimal driving torque split strategy is to
(1) Optimization problem description
maximize total efficiency of two motors. Hence, the
Optimization Variables. The power split ratio is total efficiency is used as objective function, which
considered as the optimization variable, which is can be formulated as follows:
defined as the required power of front motor di-
vided by the total demanded power: Poutput Td
η� �
PEM2 (k) TEM2 (k)nEM2 (k) Pin αTd /ηEM2 􏼁 +(1 − α)Td /ηEM1
α(k) � �
Pd (k) TEM1 (k)nEM1 (k) + TEM2 (k)nEM2 (k)
1
� ,
TEM2 (k) T (k) α/f nEM2 , αTd 􏼁 +(1 − α)/f nEM1 , (1 − α)Td 􏼁
� � EM2 ,
TEM1 (k) + TEM2 (k) Td (k) (15)
(12)
where η is the total efficiency of two motors, Poutput is
where α is the power split ratio and Pd and Td are the total output power of two motors, and Pin is the
demanded power and torque of driver, respectively. input power of the powertrain, which is supplied by
According to Equation (12), since the powertrain the battery.
architecture of the front axle is the same as that of (2) Fundamental principle of adaptive nonlinear particle
rear axle, the motor speed of EM1 is the same as that swarm optimization
of EM2. Thus, the power split ratio is the same as the
torque split ratio. The optimized output torque of Based on the powertrain model in Section 2 and the
EM1 and EM2 can be given as optimization problem description, it could be found
that the optimization of the torque split ratio is a
TEM2 (k) � α(k)Td (k); highly nonlinear optimization problem. The optimal
(13)
TEM1 (k) � (1 − α(k))Td (k). torque split ratio can be described as the optimi-
zation of α within the simplified range, given the
time-varying inputs (Td (k), nEM1 (k), and nEM2 (k)),
According to equations (12) and (13), it is obvious
subject to a set of time-varying constraints
that the range of optimization variable is 0-1. Be-
(TdismaxEM1 (k), TdismaxEM2 (k)).
cause of the symmetrical front and rear powertrain
architecture, the range of optimization variables can PSO is a population-based heuristic algorithm in-
be simplified to 0–0.5. troduced by Kennedy and Eberhart in 1995 [29],
which mimics the social behavior of animal groups
Constraints. The output torque of EM1 and EM2 is
supposed not to exceed the available maximum dis- such as a flock of birds or fish shoals guiding particles
charge torque of motor. It is also nonnegative as the to find global optimal solution. The optimization
energy efficiency is lower when one motor drives the procedure can be abstracted as particles flying in the
other one to generate electricity. In detail, it can be designed N-dimensional space. During each flying
expressed as process, each particle would update its velocity and
position according to previous individual best po-
0 ≤ TEM1 (k) ≤ TdismaxEM1 nEM1 (k)􏼁, sition and the best position of the swarm. Then, the
(14)
0 ≤ TEM2 (k) ≤ TdismaxEM2 nEM2 (k)􏼁. particle will be compared with the best solution. It
Mathematical Problems in Engineering 9

has experienced to update individual best position. each particle is randomly generated within the range of
And the best position of the swarm is further selected corresponding segment, shown in Figure 8(b). Thus,
among all individual best positions. The velocity and the particle is distributed within the designated range as
position of each particle can be calculated using the uniformly as possible, as shown in the following
following equation: equation:
vN N N N
i (t + 1) � ξvi (t) + c1 r1 􏼐pBesti (t) − pi (t)􏼑
(i − 1) pN N N N
max − pmin 􏼁 r3 pmax − pmin 􏼁
pN N
i � pmin + + ,
sizepop sizepop
+ c2 r2 􏼐gBestN (t) − pN
i (t)􏼑, (16)
(17)
pN
i (t + 1) � pN
i (t) + vN
i (t + 1),
where pN N
min and pmax are lower bound and upper bound
th
where vN
i (t) is the current velocity of the i particle, of the designated range, respectively, and r3 is a random
i � 1, . . ., sizepop; sizepop is the population size of the number between 0 and 1.
swarm; t means the tth iteration, t � 1, . . ., maxgen; N Second, it is concluded that the bigger ξ value results in
represents the dimensions of solution space; ξ is the the occurrence of exploration behavior, while the
inertia weight, which is used to inherit from previous smaller ξ value results in the occurrence of exploitation
velocity feature to next iteration; c1 and c2 are the behavior [31]. In the ANLPSO, the inertia weight de-
learning factors of individual learning and social creases nonlinearly with the increase in iteration. In the
learning, respectively; r1 and r2 are the random number early optimization process, the inertia weight is rela-
between 0 and 1; pBestN i (t) is the best position of the i
th
tively big to ensure the convergence speed. And the
th N
particle in t iteration; pi (t) is the current position of inertia weight decreases with the increasing iteration to
the ith particle in tth iteration; and gBestN (t) is the best improve the quality of solution. When compared to
position of the whole swarm after t iterations. linear decreasing inertia weight, the inertia weight in
It is proved that PSO is a simple, efficient, and robust ANLPSO is a little smaller, resulting in better quality of
stochastic optimization method to solve a nonlinear solution. The inertia weight can be determined as
optimization problem using the concept of social in- follows:
teraction. However, the general PSO suffers from some 0.5
t−1
drawbacks, which may result in premature convergence ξ � ξ max + 􏼠 􏼡 ξ min − ξ max 􏼁, (18)
and slow convergence speed [30]. maxgen − 1
To improve the performance of general PSO, this paper
adopts an adaptive nonlinear particle swarm optimi- where ξ min and ξ max are the designated minimum value
zation algorithm (ANLPSO). First, in the general PSO, and maximum value of the inertia weight.
the initial position and velocity of each particle are Third, the learning factors c1 and c2 are dynamically
randomly generated within the designated range (range changing with the fitness value of current particle,
of the optimization variables). Assuming most of the shown in equation (19). Particles with above average
initial particles centralized in a small region (subset of fitness values can obtain the smaller c1 value and bigger
the designated range), the PSO would be highly pos- c2 value to strengthen their social-learning abilities and
sible to get into the local optimum, shown in improve solutions’ qualities. On the contrary, particles
Figure 8(a). As such, the initialization process is with below average fitness values can obtain the bigger
redesigned to guarantee the population diversity in the c1 value and smaller c2 value to strengthen their global
early optimization process. The designated range is search abilities and improve convergence speed:
divided into sizepop segments. The initial position of

⎪
⎧
⎪ 􏼐c1, max − c1,min 􏼑 Fiti − Fitmin 􏼁
⎪
⎪ c � c1, max − , c2 � c − c1 , Fiti < Fitavg ,
⎪
⎨ 1 Fitavg − Fitmin
⎪ (19)
⎪
⎪
⎪
⎪
⎩c �c Fiti ≥ Fitavg ,
1 1, min , c2 � c − c1 ,

where c1,min and c1,max are minimum and maximum value of the ith particle; and Fitmin and Fitavg are the
individual-learning factor value, respectively; c is the minimum and average fitness value of the whole
designated sum value of c1 and c2 ; Fiti is the fitness swarm, respectively.
10 Mathematical Problems in Engineering

Sizepopth
segment
Local optimum

……

……
1st segment 2nd segment 3rd segment
Particles ……
centralization Best solution
Best solution

1st iteration 2nd iteration 3rd iteration …… Maxgen – 1th Maxgenth 1st iteration 2nd iteration 3rd iteration …… Maxgen – 1th Maxgenth
(Initialization) iteration iteration (Initialization) iteration iteration

(a) (b)

Figure 8: Differences in initialization process between (a) PSO and (b) ANLPSO.

(3) Procedures of adaptive nonlinear particle swarm (4) Fitness value calculation: as mentioned, the total
optimization efficiency is used as objective function of the opti-
To optimize the energy efficiency in this dual-motor mization problem. In ANLPSO, the fitness value of
vehicle, the ANLPSO algorithm was adopted to the particle is calculated based on this objective
control the torque split between two motors. The function. And the particle with the maximum fitness
optimization procedure of the torque split ratio value would be selected as the best solution:
based on ANLPSO (shown in Figure 9) is described 1
Fit � η � .
as follows [32]: α/f nEM2 , αTd 􏼁 +(1 − α)/f nEM1 ,(1 − α)Td 􏼁
(1) Adaptive initialization: the initial position of (21)
particles is generated according to equation (17)
to guarantee the population diversity, while the (5) Update pBesti and gBest: after the calculation of
initial velocity for the particles is randomly fitness value, each particle should choose the particle
generated. The optimization problem in this with the maximum fitness value, pBesti , according to
study is a one-dimensional (N � 1) problem. its own experience. Besides, the particle with global
Hence, the position and velocity of the particles best fitness value of the whole swarm should be
are described as αi and vi , respectively, which are selected as gBest among all pBesti . In the first iter-
generated as follows: ation, each particle is directly selected as its own
pBesti .
(i − 1) αmax − αmin 􏼁 r3 αmax − αmin 􏼁
αi � αmin + + , (6) Stopping rule: Steps (2)–(5) are repeated until the
sizepop sizepop
optimization converges or the maximum number of
vi � R􏼂vmin , vmax 􏼃, iteration is reached. The best position with the
maximum fitness value would be outputted as the
(20) optimal torque split ratio of two motors.

where αmin and αmax are lower and upper boundaries

of torque split ratio, respectively, and R[vmin , vmax ] 4. Results and Discussion
means the random process to generate velocity In this study, the simulation was conducted for four
within the range of minimum velocity and maxi- cases, including single-motor powertrain, dual-motor
mum velocity. powertrain with even distributed strategy, dual-motor
(2) Update inertia weight and learning factors: the powertrain with rule-based strategy, and dual-motor
nonlinear-decreasing inertia weight and dynamic powertrain with ANLPSO strategy. The results would be
learning factors can be determined according to analyzed with comparison in this section. It is noted that
equations (18) and (19), respectively. the torque split strategy was not applied into the control
(3) Update velocity and position: the velocity and po- strategy for single-motor powertrain. Thus, only the
sition of each particle are updated according to results of energy consumption would be compared with
equation (16). other three cases.
Mathematical Problems in Engineering 11

Adaptive initialization

The initial position of each particle is randomly

generated within the range of according segment

Update inertia weight and

learning factors

Update nonlinear-decreasing inertia

weight and dynamic learning factors

Update velocity and position

Fitness value calculation No

Update pBesti and gBest

Optimization converges or
maximum iteration is reached

Yes

Output the best solution

Output the optimal torque split ratio

Figure 9: Optimization procedure of ANLPSO.

4.1. Simulation Program and Settings. The simulation was real vehicle is shown in Figure 11(a). In the development
mainly based on Matlab/Simulink platform. The sampling stage, the allocation of driving torque for this sample vehicle
time was set as a fixed sampling time of 0.1 s, and the is based on the simplest even distributed strategy and the
Bogacki–Shampine solver was selected. Both the driver braking energy is not regenerated. Due to the restricted
model, physical vehicle model, and vehicle control unit experiment condition, the real-vehicle experiment was
(VCU) model were conducted, as shown in Figure 10. The conducted on chassis dynamometer using the NEDC (as
ANLPSO strategy was programmed as a S-function block in shown in Figure 11(b)). Both the vehicle velocity and battery
the torque split layer of VCU. All key parameters of SOC results were recorded from VCU through CAN bus
ANLPSO are listed in Table 4. acquisition tool and used to validate the corresponding
simulation model. It is noted that all parameters and the
control strategy of the simulation model are consistent with
4.2. Simulation Model Validation. In order to verify the the experiment.
validity and accuracy of the physical simulation model, the As shown in Figure 12(a), both the simulation model and
real vehicle experiment is conducted. As mentioned in experimental vehicle could track the driving cycle well. The
Section 2, the target vehicle is a commercial dual-motor vast majority of velocity tracking error of experiment results
electric vehicle prototype. The powertrain architecture of is within ±5 km/h, while the velocity tracking error of
12 Mathematical Problems in Engineering

ANLPSO S-Function

Desired velocity Torque_split_ratio

Acc_pedal Hyd_brk_prs T_brk T_brk Current velocity
Acceleration Hydraulic brake pressure
Desired velocity pedal command Hydraulic brake model Tm,out Motor speed
Demanded_brk_prs
Demanded Vehicle
TEM1 Tm, out
brake pressure longitudinal
Current_velocity TEM1
Driving cycle Current velocity TEM2 Pm, req dynamic model

Driver model Motor_speed Motor speed Pbrk, reg

TEM2
Motor model
SOC
VCU Pm, req
SOC
Pbrk, reg
Battery model

Figure 10: Physical vehicle model with VCU based on Matlab/Simulink platform.

Table 4: Parameters value of ANLPSO.

Rear motor
Parameters Value
Sizepop 5
Maxgen 20
N 1
αmin 0
αmax 0.5
ξ min 0.4 Front motor
ξ max 0.9
Rear motor controller
vmin −0.25
vmax 0.25 Battery
c1,min 1.2
c1,max 1.5
(a)
c 3

simulation is within ±0.5 km/h. As for the changing ten-

dency of battery SOC, simulation and experimental results
are consistent. Because of the sampling accuracy (0.4%) of
battery management system, the SOC results of experiment
decrease in a ladder shape. The SOC starts with an initial
value of 90% and ends with 84.87% and 84.8% in simulation
and experiment, respectively, implying a high validity and
accuracy of the simulation model, shown in Figure 12(b).

4.3. Simulation Results. With the validated simulation (b)

model, comparative simulation results are presented using
three different strategies (note that regenerative braking Figure 11: Real-vehicle experiment: (a) powertrain architecture;
strategy is added to controller in this section). The results in (b) vehicle experiment on dynamometer.
NEDC at the initial SOC 90% are demonstrated in Figures 13
and 14. As illustrated in the first row of Figure 13, the actual
velocity (V c) tracks the desired velocity (V d) very well for obvious that the torque split ratio is 0.5 during acceleration
all three cases. The velocity track error is within ± 0.4 km/h and uniform velocity conditions (T_r > 0), which means the
since more precise driver model based on the feedforward- desired driving torque is evenly distributed between EM1
feedback PI control is adopted, indicating the reliability of and EM2. However, the torque split ratio is 0 in the third row
the simulation results. The second row presents the torque for rule-based strategy, indicating that the EM1 provides
split ratio (α) results for even distributed strategy. It is driving torque solely in the whole cycle run. In the fourth
Mathematical Problems in Engineering 13

90 90.0

80 89.5
89.0
70
88.5
60
Velocity (km/h)

88.0
50

SOC (%)
87.5
40 87.0
86.5
30
86.0
20
85.5
10 85.0
0 84.5
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Time (s) Time (s)
Simulation Simulation
Experiment Experiment
(a) (b)

Figure 12: Comparison of simulation and experiment results: (a) driving cycle tracking; (b) battery SOC.

100
75
Velocity
(km/h)

50
25
0
0 200 400 600 800 1000 1200
NEDC
V_d
V_c

50
Alfa (%)

25

0
0 200 400 600 800 1000 1200
Even distributed
Alfa
50
Alfa (%)

25

0
0 200 400 600 800 1000 1200
Rule-based
Alfa
50
Alfa (%)

25

0
0 200 400 600 800 1000 1200
ANLPSO
Alfa
Figure 13: Velocity and driving torque split ratio of three strategies in NEDC.
14 Mathematical Problems in Engineering

100

Torque (N∗m)
50

–50
0 200 400 600 800 1000 1200
Even distributed
T_r
T_m1
T_m2
100
Torque (N∗m)

50

–50
0 200 400 600 800 1000 1200
Rule-based
T_r
T_m1
T_m2
100
Torque (N∗m)

50

–50
0 200 400 600 800 1000 1200
ANLPSO
T_r
T_m1
T_m2
90
89
SOC (%)

88
87
86
85
0 200 400 600 800 1000 1200
Time (s)
Even distributed
Rule-based
ANLPSO
Figure 14: Torque curves and the battery SOC of three strategies in NEDC.

row, the real-time torque split ratio results based on change curve of EM1 is the same as that of demanded
ANLPSO is revealed. Compared with the other two strat- torque. The torque results are in concert with the torque
egies, the optimal torque split ratio changes with torque split ratio results shown in Figure 13. With respect to
requirements during acceleration condition, which implies ANLPSO, the torque of two motors is split based on α. The
that two motors participate in propulsion more flexibly to EM1 predominates in the low torque area (when T_r is
achieve a higher total efficiency. small, such as uniform velocity condition), while two
In order to elucidate the difference of those three motors are both activated to propel the vehicle in high
strategies specifically, the torque results of three cases are torque area (when T_r is relatively big, such as acceler-
illustrated in Figure 14. For the even distributed strategy ation condition). Within the high torque area, the torque
in the first row, the demanded torque is equally distributed split ratio is 0.5 mostly. However, it is noticeable that the
to EM1 and EM2. The torque change trajectory of two required torque is not equally distributed to two motors in
motors is exactly the same. As for rule-based strategy, the some of high torque area, which indicates that even
vehicle is propelled only by the EM1, and the torque distributed strategy is not always the optimal torque split
Mathematical Problems in Engineering 15

Velocity (km/h)
100
75
50
25
0
0 200 400 600 800 1000 1200 1400 1600 1800
WLTP
V_d
V_c
50
Alfa (%)

25

0
0 200 400 600 800 1000 1200 1400 1600 1800
Even distributed
Alfa
50
Alfa (%)

25

0
0 200 400 600 800 1000 1200 1400 1600 1800
Rule-based
Alfa
50
Alfa (%)

25

0
0 200 400 600 800 1000 1200 1400 1600 1800
ANLPSO
Alfa
Figure 15: Velocity and driving torque split ratio of three strategies in WLTP.

strategy in high required torque area (in literature review, acceleration and deceleration condition is more compared
for dual-motor powertrain with two identical motors, the with NEDC. Consequently, EM2 works more frequently to
optimal torque split strategy is even distributed strategy, assist EM1 to propel the vehicle in rule-based strategy, and the
while the high driving torque is required). The fourth row optimal torque split ratio based on ANLPSO also changes
shows SOC change trajectory of three cases. Since the more frequently to achieve the best energy efficiency. Fig-
powertrain is the same, the difference between three cases ure 16 demonstrates the comparison of torque results of three
is small. As can be seen, there are three green dotted strategies. Similar with the results shown in Figure 14, it is
circles, which present three zoom-in pictures of SOC more efficient that single motor supplies the power solely in
change curve in low velocity area, medium velocity area, low torque area, while two motors propel the vehicle together
and high velocity area, respectively. The battery SOC of with an appropriate torque split ratio in high torque area. In
ANLPSO is always higher than the rest two cases, indi- the fourth row of Figure 16, the results of battery SOC is
cating that the torque split strategy with ANLPSO con- presented. It is obvious that more electricity is consumed in
sumes least energy. The rule-based strategy is least efficient WLTP relative to in NEDC. In the first 1000s, frequent rapid
because the required torque is relatively high in most time acceleration indicates a higher required torque. Hence, the
of first 800s as it is inefficient that single motor drives the battery SOC of rule-based strategy drops more than that of
vehicle in high torque area. After 800s, most time of cycle even distributed strategy. Since the best split ratio is not
run is in the area of uniform velocity, which means the always 0.5 in high torque area, energy efficiency of ANLPSO is
required torque is low. It turns out that single motor is still higher than that of even distributed strategy.
more efficient to supply the power. Furthermore, it can be After 1100 s, rule-based strategy consumes little energy
seen that the difference of battery SOC between even than even distributed strategy because of the smoother
distributed strategy and rule-based strategy becomes acceleration process and uniform velocity period. And the
smaller and the SOC with rule-based strategy is higher change tendency of the SOC is similar to the results in the
than that with even distributed strategy eventually. fourth row of Figure 14. In general, ANLPSO exhibits
Figures 15 and 16present the simulation results in WLTP. highest energy efficiency while the rule-based strategy is
As illustrated in Figure 15, the adaptive driver model still relatively efficient than even distributed strategy in both
tracks the desired velocity precisely. It is apparent that NEDC and WLTP.
16 Mathematical Problems in Engineering

150

Torque (N∗m)
100
50
0
–50
0 200 400 600 800 1000 1200 1400 1600 1800
Even distributed
T_r
T_m1
T_m2
150
Torque (N∗m)

100
50
0
–50
0 200 400 600 800 1000 1200 1400 1600 1800
Even distributed
T_r
T_m1
T_m2
150
Torque (N∗m)

100
50
0
–50
0 200 400 600 800 1000 1200 1400 1600 1800
Rule-based
T_r
T_m1
T_m2
90
88
SOC (%)

86
84
82
80
0 200 400 600 800 1000 1200 1400 1600 1800
Time (s)
Even distributed
Rule-based
ANLPSO
Figure 16: Torque curves and the battery SOC of three strategies in WLTP.

4.4. Optimization Performance and Real-Time Applicability particles are randomly generated within a small range
Analysis of ANLPSO. As detailed in Section 3, the torque (0.35–0.36). As a result, the optimization process is trapped
split ratio is searched based on ANLPSO, which is designed in local optimum at the 7th iteration, and thus, the optimal
to ameliorate shortcomings of general PSO. To further solution is not found eventually. By contrast, because of the
compare the optimization performance and real-time ap- improvement of the initialization process, the nonlinear
plicability between general PSO and ANLPSO, the evolution decreasing inertia weight, and dynamic learning factors, the
process was investigated. Figure 17 shows the optimization converge speed and optimization performance of ANLPSO
process to search the best torque split ratio at T_r � 85 Nm are greatly enhanced. As shown in the second row and the
and Motor_speed � 5400 rpm (to compare two methods third row, the evolution process is more stable and swifter.
fairly, this condition is randomly selected from cycle run). The optimal fitness value (fitness value � 93.77345) is found
Both the best position of each particle Pbest, the best position only after 9 iterations, and the best position of each particle
of all particles Gbest, and the best fitness value in each it- Pbest both converges to the optimal position (α � 0.449) after
eration are demonstrated. As shown in the first row, four 16 iterations. Therefore, ANLPSO is verified to possess better
Mathematical Problems in Engineering 17

0.5
0.4
0.3

Alfa
0.2
0.1
0.0
0 5 10 15 20
Iteration
Pbest1 Pbest4
Pbest2 Pbest5
Pbest3 Gbest
(a)
0.5
0.4
0.3
Alfa

0.2
0.1
0.0
0 5 10 15 20
Iteration
Pbest1 Pbest4
Pbest2 Pbest5
Pbest3 Gbest

(b)
93.774
93.773
93.772
Fitness value

93.771
93.770
93.769
93.768
93.767
0 5 10 15 20
Iteration
PSO
ANLPSO
(c)

Figure 17: Comparison of convergence performance between PSO and ANLPSO.

convergence performance and global optimization ability iterations and 5 populations) to ensure real-time applica-
than general PSO. bility. By contrast, it would be better to determine a large
Furthermore, the computational time of the optimal maximum iteration size and a large population size (30
torque split strategies is 377 s and 665 s for NEDC (1180 s) iterations and 10 populations) to take priority of the opti-
and WLTP (1800 s), respectively. Then, the equivalent mality without sacrifice in real-time implementability.
computational time per second is approximately 0.031 s and
0.036 s, which shows a great real-time applicability and
potential to be implemented in VCU of real vehicle. Before 4.5. Comparison of Efficiency, Motor Loss, and Energy
practical applications, to decide the population size and Improvement. To further demonstrate energy-saving po-
maximum number of iteration, there should be a tradeoff tential of dual-motor powertrain over single-motor pow-
between the optimality and real-time applicability. For ex- ertrain and the difference of three control strategies, the total
ample, as for VCU with limited computational capability, it efficiency of powertrain, the power loss of two motors, and
is probably advisable to use a relatively small maximum energy consumption were investigated. Figure 18 illustrates
iteration size and a relatively small population size (10 the results of total efficiency of two motors in NEDC and
18 Mathematical Problems in Engineering

100 100
90 90
80 80
70 70
Total efficiency (%)

Total efficiency (%)

60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200 1400 1600 1800
Time (s) Time (s)
Even distributed Even distributed
Rule-based Rule-based
ANLPSO ANLPSO
(a) (b)

Figure 18: Total efficiency of two motors of three strategies in NEDC and WLTP.

5000
3000

4000
Motor power loss (W)

Motor power loss (W)

2000 3000

2000
1000
1000

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200 1400 1600 1800
Time (s) Time (s)
Even distributed Even distributed
Rule-based Rule-based
ANLPSO ANLPSO
(a) (b)

Figure 19: Motor power loss of two motors of three strategies in NEDC and WLTP.

WLTP. The total efficiency of ANLPSO is above 90% in most is the most in high torque area and the motor power loss of
of cycle run, and even the lowest efficiency is not lower than even distributed strategy is the most in low torque area, as
65%, which is the highest among three strategies. Conse- shown in Figure 19. It is worth noting that the total effi-
quently, the power loss of two motors is the lowest with ciency of even distributed strategy is basically the same as
ANLPSO, as shown in Figure 19. ANLPSO in high torque area, while the total efficiency of
As for rule-based strategy, the total efficiency is the rule-based strategy is the same as ANLPSO in low torque
lowest in the first 800 s in NEDC and in the first 1000 s in area.
WLTP. However, it becomes more efficient than even The energy consumption is obtained by integrating the
distributed strategy when the cycle run enters in low torque battery power in NEDC and WLTP, as listed in Table 5.
area (after 800 s in NEDC and after 1000 s in WLTP). And it Assuming that the braking energy of each axle is totally
is obvious that the motor power loss of rule-based strategy regenerated by the motor mounted on according axle, it is
Mathematical Problems in Engineering 19

Table 5: Energy consumption comparison.

Energy saving
Electricity
Total motor loss compared to single-
consumption
(kWh) motor powertrain
(kWh/100 km)
(%)
NEDC WLTP NEDC WLTP NEDC WLTP
Single-motor powertrain — — 20.29 23.16 — —
Dual-motor powertrain with even distributed strategy 0.2324 0.5330 18.15 20.65 10.55 10.83
Dual-motor powertrain with rule-based strategy 0.2221 0.5107 18.05 20.54 11.04 11.31
Dual-motor powertrain with ANLPSO 0.2041 0.4695 17.88 20.34 11.88 12.18

proved that the energy saving of dual-motor powertrain with optimization performance than conventional PSO,
three different strategies in NEDC has been improved by which is provably real-time available in a real con-
around 10.55%, 11.04%, and 11.88%, respectively, while the troller to achieve the optimal torque allocation.
energy saving improvement of dual-motor powertrain with (4) Based on the detailed result analysis of three strat-
three different strategies in WLTP is 10.83%, 11.31%, and egies, it is concluded that, for dual-motor with two
12.18%, respectively, compared with single-motor power- identical motors, the single motor is more efficient to
train. In addition, when compared with even distributed provide the power in low torque area. However,
strategy and rule-based strategy for dual-motor powertrain, within high torque area, both two motors should be
the proposed torque split strategy shows an improvement in activated to propel the vehicle with an appropriate
total motor loss of 12.17% and 8.1% in NEDC, respectively, torque split ratio. In most cases, the optimal torque
and 11.91% and 8.07%, respectively, in WLTP. split ratio is 0.5, which means the torque should be
equally distributed to two motors. However, 0.5 is
5. Conclusion not always the best split ratio according to searching
results of ANLPSO. Hence, a proper methodology to
In this study, a real-time optimal energy management figure out the real-time optimal torque split ratio is
strategy for a dual-motor electric powertrain was proposed. essential to manage the torque distribution of dual
Firstly, the physical vehicle model was implemented in motor powertrain, which is also the main contri-
Matlab/Simulink platform. Then, the optimal torque split bution of this study.
strategy based on ANLPSO was integrated with VCU by
Our future work could focus on the applicability of the
using S-function. Moreover, three cases (single-motor
optimal torque split strategy in the real vehicle test. Fur-
powertrain, dual-motor powertrain with even distributed
thermore, the research would be expanded into the inte-
strategy, and dual-motor powertrain with rule-based strat-
grated optimization framework for component sizing and
egy) were also modeled for comparison. Several important
optimal energy management strategy for dual-motor
findings are summarized as follows:
powertrain.
(1) In NEDC and WLTP driving cycles, the dual-motor
powertrain with two identical motors could achieve
energy saving of (10.55%, 10.83%), (11.04%, 11.34%),
Data Availability
and (11.88%, 12.18%) using even distributed strategy, The following data, models, or results that support the
rule-based strategy, and the optimal torque split findings of this study are available from the corresponding
strategy, compared to single-motor powertrain, re- author upon reasonable request: (1) vehicle parameters; (2)
spectively. This confirms that the dual-motor pow- motor efficiency data; (3) vehicle physical model; and (4)
ertrain has a great energy efficiency improvement results data.
over single-motor powertrain.
(2) With respect to dual-motor powertrain, compared to Conflicts of Interest
even distributed strategy, the proposed optimal
torque split strategy could reduce the total motor loss The authors declare that there are no conflicts of interest.
by 12.17% and 11.91% in NEDC and WLTP, re-
spectively. And the improvement over rule-based
strategy is 8.1% and 8.07% accordingly.
Acknowledgments
(3) As for the ANLPSO, the real-time best torque split The authors would like to express their gratitude to Tsinghua
ratio could be searched swiftly with three inputs University under project 2017YFB0103902, Wuhan Uni-
(motor speed, the required torque, and the available versity of Technology under the 111 project (no. B17034),
maximum torque of each motor) and fitness func- and Innovative Research Team Development Program of
tion of total efficiency of two motors. It is also shown Ministry of Education of China (no. IRT_17R83) for sup-
that it has better convergence speed and global porting this study.
20 Mathematical Problems in Engineering

Supplementary Materials [12] F. C. Correa, J. J. Eckert, L. C. A. Silva, F. M. Santiciolli,

E. S. Costa, and F. G. Dedini, “Study of different electric
In the “experiment results” folder, the excel file named vehicle propulsion system configurations,” in Proceedings of
“experiment results” is the results from real-vehicle exper- the 2015 IEEE Vehicle Power and Propulsion Conference
iment validation, while the excel file named “simulation (VPPC), pp. 1–6, Montreal, Quebec, Canada, October 2015.
model” is the results of the validated simulation model. [13] J. J. Eckert, L. C. A. Silva, E. S. Costa, F. M. Santiciolli,
These results are mainly for simulation model validation. In F. G. Dedini, and F. C. Corrêa, “Electric vehicle drivetrain
the “simulation results” folder, there are two folders named optimisation,” IET Electrical Systems in Transportation, vol. 7,
“NEDC” and “WLTP,” respectively, which present the no. 1, pp. 32–40, 2017.
[14] H. Qian, G. Xu, J. Yan, T. L. Lam, Y. Xu, and K. Xu, “Energy
simulation results in NEDC and WLTP, respectively. The
management for four-wheel independent driving vehicle,” in
excel file named “EVEN” is the results of even distributed Proceedings of the 2010 IEEE/RSJ International Conference on
strategy. The excel file named “Rule-based” is the results of Intelligent Robots and Systems, pp. 5532–5537, IEEE, Taipei,
the rule-based strategy. The excel file named “ANLPSO” is Taiwan, 2010.
the results of the optimal torque split strategy using ANL- [15] V. Ivanov, D. Savitski, and B. Shyrokau, “A survey of traction
PSO. These results are mainly for comparison between control and antilock braking systems of full electric vehicles
different torque split strategies. (Supplementary Materials) with individually controlled electric motors,” IEEE Transac-
tions on Vehicular Technology, vol. 64, no. 9, pp. 3878–3896,
2015.
References [16] Y. Chen, X. Li, C. Wiet, and J. Wang, “Energy management
[1] L. C. A. Silva, J. J. Eckert, F. M. Santiciolli, E. S. Costa, and driving strategy for in-wheel motor electric ground ve-
F. G. Dedini, and F. C. Correa, “A study of battery power for a hicles with terrain profile preview,” IEEE Transactions on
different electric vehicle propulsion system,” in Proceedings of Industrial Informatics, vol. 10, no. 3, pp. 1938–1947, 2014.
the 2015 International Conference on Electrical Systems for [17] M. Hu, J. Zeng, S. Xu, C. Fu, and D. Qin, “Efficiency study of a
Aircraft, Railway, Ship Propulsion and Road Vehicles (ESARS), dual-motor coupling EV powertrain,” IEEE Transactions on
pp. 1–6, IEEE, Aachen, Germany, March 2015. Vehicular Technology, vol. 64, no. 6, pp. 2252–2260, 2015.
[2] T. Wilberforce, Z. El-Hassan, F. N. Khatib et al., “Develop- [18] Y. Gao, W. Wang, and Y. Li, “Optimization of control strategy
ments of electric cars and fuel cell hydrogen electric cars,” for dual-motor coupling propulsion system based on dynamic
International Journal of Hydrogen Energy, vol. 42, no. 40, programming method,” in Proceedings of the 2019 3rd Con-
pp. 25695–25734, 2017. ference on Vehicle Control and Intelligence (CVCI), pp. 1–6,
[3] K. T. Chau and C. C. Chan, “Emerging energy-efficient Hefei, China, September 2019.
technologies for hybrid electric vehicles,” Proceedings of the [19] B. Wang, M. Xu, and L. Yang, “Study on the economic and
IEEE, vol. 95, no. 4, pp. 821–835, 2007. environmental benefits of different EV powertrain topolo-
[4] G. Wu, X. Zhang, and Z. Dong, “Powertrain architectures of gies,” Energy Conversion and Management, vol. 86, pp. 916–
electrified vehicles: review, classification and comparison,” 926, 2014.
Journal of the Franklin Institute, vol. 352, no. 2, pp. 425–448, [20] J. Ruan and Q. Song, “A novel dual-motor two-speed direct
2015. drive battery electric vehicle drivetrain,” IEEE Access, vol. 7,
[5] H. Eckard and P. Marx, “Electric cars: technical characteristics pp. 54330–54342, 2019.
and environmental impacts,” Environmental Sciences Europe, [21] J. Kang, J. Yoo, and K. Yi, “Driving control algorithm for
vol. 24, no. 1, pp. 1–15, 2012. maneuverability, lateral stability, and rollover prevention of
[6] O. van Vliet, A. S. Brouwer, T. Kuramochi, M. van den Broek, and 4WD electric vehicles with independently driven front and
A. Faaij, “Energy use, cost and CO2 emissions of electric cars,” rear wheels,” IEEE Transactions on Vehicular Technology,
Journal of Power Sources, vol. 196, no. 4, pp. 2298–2310, 2011. vol. 60, no. 7, pp. 2987–3001, 2011.
[7] S. Chopra and P. Bauer, “Driving range extension of EV with [22] P. D. U. Coronado and H. Ahuett-Garza, “Control strategy for
on-road contactless power transfer: a case study,” IEEE power distribution in dual motor propulsion system for
Transactions on Industrial Electronics, vol. 60, no. 1, electric vehicles,” Mathematical Problems in Engineering,
pp. 329–338, 2013. vol. 2015, Article ID 814307, 10 pages, 2015.
[8] M. A. Miller, A. G. Holmes, B. M. Conlon, and P. J. Savagian, [23] B. Sun, S. Gao, C. Ma, and J. Li, “System power loss opti-
“The GM “voltec” 4ET50 multi-mode electric transaxle,” SAE mization of electric vehicle driven by front and rear induction
International Journal of Engines, vol. 4, no. 1, pp. 1102–1114, motors,” International Journal of Automotive Technology,
2011. vol. 19, no. 1, pp. 121–134, 2018.
[9] F. Di Nicola, A. Sorniotti, T. Holdstock, F. Viotto, and [24] X. Yuan and J. Wang, “Torque distribution strategy for a
S. Bertolotto, “Optimization of a multiple-speed transmission front- and rear-wheel-driven electric vehicle,” IEEE Trans-
for downsizing the motor of a fully electric vehicle,” SAE actions on Vehicular Technology, vol. 61, no. 8, pp. 3365–3374,
International Journal of Alternative Powertrains, vol. 1, no. 1, 2012.
pp. 134–143, 2012. [25] D. Wang and B. Wang, “Research on driving force optimal
[10] Q. Ren, D. A. Crolla, and A. Morris, “Effect of transmission distribution and fuzzy decision control system for a dual-
design on electric vehicle (EV) performance,” in Proceedings motor electric vehicle,” in Proceedings of the 2015 34th Chinese
of the 2009 IEEE Vehicle Power and Propulsion Conference, Control Conference (CCC), pp. 8146–8153, Chinese Associa-
pp. 1260–1265, IEEE, Dearborn, MI, USA, September 2009. tion of Automation, Hangzhou, China, July 2015.
[11] J. Ruan, P. Walker, and N. Zhang, “A comparative study [26] M. Zhao, J. Shi, and C. Lin, “Energy management strategy
energy consumption and costs of battery electric vehicle design for dual-motor coaxial coupling propulsion electric
transmissions,” Applied Energy, vol. 165, pp. 119–134, 2016. city-buses,” Energy Procedia, vol. 152, pp. 568–573, 2018.
Mathematical Problems in Engineering 21

[27] P. Shen, Z. Zhao, X. Zhan, and J. Li, “Particle swarm opti-

mization of driving torque demand decision based on fuel
economy for plug-in hybrid electric vehicle,” Energy, vol. 123,
pp. 89–107, 2017.
[28] J. Ruan, P. D. Walker, P. A. Watterson, and N. Zhang, “The
dynamic performance and economic benefit of a blended
braking system in a multi-speed battery electric vehicle,”
Applied Energy, vol. 183, pp. 1240–1258, 2016.
[29] S. Amaran, N. V. Sahinidis, B. Sharda, and S. J. Bury,
“Simulation optimization: a review of algorithms and appli-
cations,” Annals of Operations Research, vol. 240, no. 1,
pp. 351–380, 2016.
[30] H. S. Ramadan, A. F. Bendary, and S. Nagy, “Particle swarm
optimization algorithm for capacitor allocation problem in
distribution systems with wind turbine generators,” Inter-
national Journal of Electrical Power & Energy Systems, vol. 84,
pp. 143–152, 2017.
[31] M. N. Ab Wahab, S. Nefti-Meziani, and A. Atyabi, “A
comprehensive review of swarm optimization algorithms,”
PLoS One, vol. 10, no. 5, 2015.
[32] S.-Y. Chen, Y.-H. Hung, C.-H. Wu, and S.-T. Huang, “Op-
timal energy management of a hybrid electric powertrain
system using improved particle swarm optimization,” Applied
Energy, vol. 160, pp. 132–145, 2015.