Received December 24, 2018, accepted January 7, 2019, date of publication January 14, 2019, date of current version

February 14, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2892746

Electrical Vehicle Path Tracking Based Model

Predictive Control With a Laguerre Function
and Exponential Weight
BING ZHANG 1, CHANGFU ZONG1 , GUOYING CHEN1 , AND BANGCHENG ZHANG2
1 Department of State Key Laboratory of Automotive Simulation and Control, Jilin University, Changchun 130022, China
2 Automotive Engineering Research Institute, Changchun University of Technology, Changchun 130012, China

Corresponding authors: Changfu Zong ([email protected]) and Bangcheng Zhang ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61751304, Grant 51575224, and
Grant 51505178, and in part by the China Postdoctoral Science Foundation under Grant 2014M561289.

ABSTRACT Model predictive control (MPC) is advantageous for designing an electrical vehicle
path-tracking controller, but the high computational complexity, mathematical problem, and parameteriza-
tion challenge adversely affect the control performance. Hence, based on a fully actuated-by-wire electrical
vehicle (FAW-EV), a novel path-tracking controller based on improved MPC with a Laguerre function and
exponential weight (LEMPC) is designed. The massive optimization control parameters of MPC with a long
control horizon are reduced by introducing a fitting orthogonal sequence consisting of Laguerre functions,
thereby substantially reducing the computational complexity without sacrificing the tracking accuracy.
An exponential weight with decreasing characteristic is introduced to MPC to solve the mathematical
problem, thereby improving the robustness of the path tracking controller. In addition, the parameterization
access for online adjusting path tracking control performance can be provided by the proposed method. The
path tracking motion realization for FAW-EV is subsequently illustrated. Finally, several simulations are
implemented to verify the advantages of the proposed method.

INDEX TERMS Path track, electrical vehicle, model predictive control (MPC), Laguerre function,
exponential weight.

I. INTRODUCTION
In the development of sensing, wire-control, and electricity-
storage capacities, the autonomous electrical vehicle has
attracted considerable attention because of its convenience,
intelligence and environmental protection [1]– [3], and an
increasing number of researchers are focusing on this
advanced device. The driving control method is the core of
the autonomous electric vehicle; its control task can often be
functionally divided into path planning and path tracking [4],
FIGURE 1. Relationship between autonomous driving control tasks.
with the relationship shown in Fig. 1. Path tracking control
is very significant since it is the cornerstone for realizing
autonomous driving of electrical vehicles [5]– [8]. predictive control (MPC) [17], [18], linear quadratic regula-
The primary target of a path tracking control system is tor (LQR) optimal control [19], and output constraint con-
to control the vehicle to accurately follow a reference path trol [20]. Each of these technologies can converge the vehicle
given by path planning. This process is not easy due to the trajectory to the expectations, especially, MPC is more suit-
requirement of simultaneously ensuring tracking accuracy able since the constrained mechanism can take the vehicle
and vehicle dynamic stability [9]– [11]. For this function, dynamic limit into consideration, and the control input for
many control methods have been applied, such as slide mode path tracking can be achieved concerning overall situation
control (SMC) [12]– [14], robust control [15], [16], model during prediction horizon [21], [22].

2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
17082 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 2. Complete control structure for autonomous driving of the FAW-EV.

The essential idea of MPC is to achieve control input by influence the path tracking control performance are very
solving the optimization problem of minimizing the error difficult to adjust finely online.
between the future system output and reference output, The previous studies show that the MPC controller with
although this is only a basic control framework, and many Laguerre fitting can reduce the computational complexity
researchers revise it to meet different path tracking control of MPC optimization problem [26], and the unexpected
requirements. For tracking the planned collision-avoiding sensitivity of MPC can be improved by introducing expo-
path, multiple constraints are introduced to the MPC frame- nential weight [27], therefore, motivated by the problems
work in [4] to track a reference path that minimizes the possi- exited in MPC path tracking controller, a novel path tracking
bility of collision. In [23], an electrical vehicle path tracking controller based on improved MPC with a Laguerre func-
controller based on the MPC framework is proposed in which tion and exponential weight (LEMPC) is designed for the
the steering angle and wheel slip ratio are the control inputs. fully actuated-by-wire electrical vehicle (FAW-EV), in which
Reference [24] combines the MPC and SMC to robustly track the four wheels can be independently steered, driven and
a reference path; this study mainly addresses the micro- braked [28]. The main contributions of this paper are listed
vehicle, and the tire stability limit is considered to harmonize as follows:
the path tracking accuracy and vehicle robustness. For path (1) In the prediction of future system outputs within
tracking control of wheeled robots, reference [25] proposes MPC path tracking controller, a Laguerre function orthogonal
a backstepping kinematic controller based MPC with dual sequence is employed to fit the long control input trajectory
heuristic programming. by a linearized combination; therefore, the computational
However, the basic framework of above MPC path tracking complexity can be reduced without sacrificing the tracking
controllers encounters its limits in practical applications of performance by transforming the optimization control param-
electrical vehicle path tracking. First, the long control horizon eters from massive control input increments to a few fitting
is always introduced to ensure the favorable tracking con- coefficients.
trol performance. Consequently, the optimization problem (2) The mathematical problem inherent in MPC is ana-
of MPC possesses massive optimization control parameters lyzed and solved by introducing an exponential weight to
defined as future control inputs or future control input incre- the cost function of the MPC path tracking controller; there-
ments. Therefore, a high computational complexity arises fore, the unexpected sensitivity of the MPC path tracking
and can be made worse under multiple dimensions of con- controller with a long prediction horizon can be ame-
trol inputs and constraints. Second, there is a mathematical liorated, and the path tracking control robustness can be
problem in MPC in that the instability of the predictive plant improved.
is cumulated during predicting future system outputs, which (3) We propose an LEMPC path tracking control method
leads to high sensitivity even to slight disturbances. Such in which the attenuating property diversity of a Laguerre
unexpected sensitivity will be deteriorated by setting a long function orthogonal sequence with different parameters pro-
prediction horizon, however, which is always necessary for vides parameterization for finely online-adjusting path track-
the dynamic-unstable driving conditions of electrical vehicle ing control performance to varied path tracking requirements;
path tracking. Finally, the control requirement for path track- the prediction horizon can be adjusted online to variations in
ing control performance is varied as different driving condi- the vehicle dynamic stability since the mathematical prob-
tions; unfortunately, the MPC parameters that significantly lem is improved. Moreover, the parameterized path tracking

VOLUME 7, 2019 17083

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

controller basing LEMPC can be adjusted to compensate the where X and Y denote the vehicle longitudinal position and
linearized error. lateral position, respectively, Clf and Clr denote the front and
For convenient and flexible control, the autonomous driv- rear equivalent longitudinal cornering stiffness, respectively,
ing control system of an FAW-EV is based on a modularized Ccf and Ccr denote the front and rear equivalent lateral
structure, as shown as Fig. 2. This paper emphasizes the cornering stiffness, respectively, sf and sr denote the front
LEMPC path tracking controller design, which is illustrated and rear longitudinal slip rates, respectively, ψ denotes the
as follows: Section II gives both the predictive plant for yaw angle, and Vx and Vy denote the longitudinal and lateral
LEMPC path tracking control and the model capturing the vehicle velocities, respectively. δfd denotes the desired front
FAW-EV dynamic characteristics. Section III illustrates the steering angle.
improved path tracking control method based on LEMPC, MPC method is computationally complicated, and the non-
and the realization of path tracking motion with the FAW-EV linearity of the predictive plant aggravates the computational
is briefly discussed. Section IV validates the availabilities of burden. Previous studies show that formulating the vehicle
the proposed method by several simulations, and the main dynamic linearly combining with a quadratic cost function
ideas for adjusting the parameters are given in this section. can reduce the computational effort and improve the effi-
Finally, a succinct conclusion is given in section V. ciency [32], [33]. Therefore, the plant (1) is linearized around
current vehicle states x(t) and previous control input u(t − 1)
II. PREDICTIVE PLANT AND DYNAMIC MODEL with assumption that the control input is unchangeable
A. PREDICTIVE PLANT FOR LEMPC PATH TRACKING within prediction horizon [33], and the achieved time-varied
For LEMPC, a predictive plant is employed to predict future form as
system outputs [29], [30]. In this paper, the single-track 
vehicle model is introduced to build the predictive plant. ẋ = At x(t) + Bt u(t)
(2)
As shown in Fig. 3, the model ignores the vertical, pitch y = Ct x(t)
and roll motions, and the sideslip angles of the left wheels T
with the vehicle state vector x = Vx , Vy , ψ, ψ̇, Y , X ,

are assumed to be equal to those of right [31]. For general-
ity, the model is considered under the condition of a small choosing the control input u as δfd , which is used to describe
front steering angle. The desired tracking motion is described the desired tracking motion, and the brake and propulsion
by front steering angle, therefore the rear steering angle is are not concerned in path tracking controller. At and Bt are
assumed to be 0, with the given initial value of Vx set to system matrices. In the path tracking system, we divide the
desired longitudinal velocity. Combining with the stiffness path tracking control issue into yaw angle tracking and lateral
tire model, and according to Newton law, the predictive plant position tracking; the system output is y = [ψ, Y ]T , and the
is described as system output matrix Ct is therefore
Vy +lf ψ̇

mV̇y = −mVx ψ̇ + 2[Ccf (δfd − Vx )
" #

 0 0 1 0 0 0
l ψ̇−V Ct =

+Ccr r Vx y ]



 0 0 0 0 1 0
mV̇x = mVy ψ̇ + 2[Clf sf + Ccf (δfd



V +l ψ̇ (1)
 − y Vxf )δfd + Clr sr ] Predictive plant is used to predict future vehicle states

V +l ψ̇ l ψ̇−V as well as future system outputs, for reducing the lin-
Iz ψ̈ = 2 [lf Ccf (δfd − y Vxf ) − lr Ccr r Vx y ]




 earized error, a simple compensation is given in formula-
Ẏ = Vx sin ψ + Vy cos ψ



 tion of MPC optimal problem described in part A of next
Ẋ = Vx cos ψ − Vy sin ψ

section.
Significantly, the predictive plant is used only to describe
the future tracking error between the vehicle motion and the
desired motion according reference path. Since the wheel
steering angels of FAW-EV are controlled independently,
the control input δfd of (2) is used only to describe the desired
path tracking motion and subsequently transformed to tire
forces of four wheels.

B. DYNAMIC MODEL FOR FAW-EV MOTION REALIZATION

For realizing path tracking motion with the FAW-EV, the rela-
tionships among the vehicle integral motions, total efforts,
and tire forces are captured by the double-track model with
3 degrees of freedom (DOFs), which takes only the planar
motion into consideration. As shown in Fig. 4(a), the dynamic
FIGURE 3. Predictive plant based single-track vehicle model. characteristics for the longitudinal, lateral and yaw directions

17084 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

For realizing the desired lateral tire force, the empirical

relation curve between the tire sideslip angle αij and the
lateral tire force Fy_ij is fitted by the arctangent function [34]:
s
Fx_ij 2 µij −1 kij
   
Fy_ij = −Cα_ij 1 − tan αij (5)
µij Fz_ij kij µij

where kij = Cα_ij π pFz_ij and the value of curve-fitting



constant p is set as 2.9 to improve the fitting accuracy under

a relatively large sideslip angle. µij denotes the tire-road
friction factor of each wheel, and a large µij often gives a high
dynamic stability margin. Fz_ij denotes the tire vertical load,
which can be achieved by referencing [35]. Angle σij between
the velocity of the wheel center and the longitudinal axle of
the vehicle body is defined as σij = αij + δij and computed by

σlf ,rf = tan−1 Vy + lf ψ̇
 Vx ∓ tf ψ̇  2


 


(6)
σlr,rr = tan−1 Vy − lr ψ̇ Vx ∓ tr ψ̇ 2
where the sign ‘‘∓’’ refers to the left/right wheel.

III. PARTH TRACKING CONTROL BASED LEMPC

The basic idea of the LEMPC path tracking controller is that a
Laguerre function orthogonal sequence is employed to fit the
trajectory of future control input increments within the MPC
path tracking optimization problem. In addition, an exponen-
tial weight is introduced to the path tracking cost function
to solve the mathematical problem. The control principle is
shown in Fig. 5. This improved method is based on the MPC
path tracking control framework, whose illustration is shown
at the beginning of this section.

FIGURE 4. Schematic of the FAW-EV dynamic model: (a) Integral vehicle A. BASIC MPC FRAMEWORK FOR PATH TRACKING
dynamic. (b) Single wheel dynamic.
An integrator is introduced to converge the steady-state error
are described as to zero, and the control input of the MPC path tracking
 r r controller is therefore changed to the incremental form 1u.
a = V̇ − ψ̇V = 1
F = 1 PP
Xw_ij



 x x y m xd m With a discrete time step of T , plant (2) can be discretized as
 i=l j=f


 r P r follows:
ay = V̇y + ψ̇Vx = m1 Fyd = m1
P
Yw_ij (3) ξ ( k + 1| t) = Ãk,t ξ ( k| t) + B̃k, t 1u( k| t)

 i=l j=f (7)
y( k| t) = C̃k,t ξ ( k| t)

t
Mzd = Iz ψ̈ = 2f Xw_rf −
Xw_lf + t2r Xw_rr −




Xw_lr




+lf Yw_lf + Yw_rf − lr Yw_lr + Yw_rr

where augmented state matrix
where the total efforts of Fxd , Fyd , and Mzd are the vehi- T
ξ ( k| t) = x( k| t) u( k − 1| t )

cle longitudinal force, lateral force, and vehicle yaw torque,
respectively. Iz denotes the rotary inertia of the vehicle around and
the vehicle vertical axle. Xw_ij and Yw_ij denote the longitudi-  
nal and lateral tire forces, respectively, acting on the point I + T At T Bt
Ãk, t =
between the tire and ground in the vehicle-body coordinate Om×n Im
 
system, and they can be described as T Bt
B̃k, t =
( Im
Xw_ij = Fx_ij cos δij − Fy_ij sin δij
∀i ∈ {l, r} , ∀j ∈ {f , r}
 
0010000
Yw_ij = Fx_ij sin δij + Fy_ij cos δij C̃k, t =
0000100
(4)
With (7), the predicted future system outputs within the
where Fx_ij and Fy_ij denote the longitudinal and lateral tire prediction horizon can be achieved by
forces, respectively, acting on the center of the tire and δij
denotes the wheel steering angle, as shown in Fig. 4(b). Ysys (k) = 9k ξ ( k| t) + 2k 1U ( k| t) (8)

VOLUME 7, 2019 17085

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 5. Schematic diagram of the LEMPC path tracking controller.

c −1
NX
where
+ k1u(k + i)k2R + ρ ε 2
i=0
 

y( k + 1| t)
 C̃k,t Ãk,t
 C̃k,t Ã2  s.t. 1Umin ≤ 1U (k) ≤ 1Umax (9b)
 y( k + 2| t)  k,t 
(9c)

   ···  Umin ≤ U ≤ Umax
· · ·
 y( k + Nc | t)  9k(Np ×1) = 
 
Ysys (k)(Np ×1) =
 
 Nc  Ymin ≤ Ysys ≤ Ymax (9d)
   C̃k,t Ãk,t 
···    ··· 
  
N where yref is the reference output and consists of the reference
y( k + Np t) p
C̃k,t Ãk,t yaw angle ψref and reference lateral position Yref . Q and R are

2 k(Np ×Nc ) weights for the system outputs and control input increments,
respectively. With the first term in the cost function (9a),
 
C̃k,t B̃k,t 0 0 0
 C̃ Ã B̃ C̃k,t B̃k,t 0 0  the predictive errors between the predicted system outputs
 k,t k,t k,t 
 ..  and future reference outputs are penalized through a weighted

 ··· ··· . ···

 norm. This approach ensures that the vehicle tracks the refer-
 Nc −1 Nc −2 
=  C̃k,t Ãk,t B̃k,t C̃k,t Ãk,t B̃k,t · · ·
 C̃k,t B̃k,t  ence path as accurately as possible. The second term of (9a)
 C̃k,t ÃNc B̃k,t C̃k,t ÃNc −1 B̃k,t · · ·

C̃k,t Ãk,t B̃k,t  is a weighted norm on the input increments, which aims to
k,t k,t
.. .. ..
 
 ..  minimize the control effort. The final term of (9a) is a penalty

 . . . . 
 on the slack variable added to avoid the infeasibility under the
Np −1 Np −2 Np −Nc −1
C̃k,t Ãk,t B̃k,t C̃k,t Ãk,t B̃k,t · · · C̃k,t Ãk,t B̃k,t strict constraints. Large weight of slack variable gives more

1u( k| t)
 relaxation to the optimization control problem, however, with
 1u( k + 1| t)  sacrificing the path tracking performance; therefore, the slack
1U (k)(Nc ×1) =   variable weight can be chosen as a large value only under
 ··· 
1u( k + Nc − 1| t) condition that the optimization control problem is insolu-
ble [36]. Equations (9b)-(9d) describe the constraints for the
control input increments according to the tolerances of the
where Np and Nc represent the lengths of the prediction
path tracking control system outputs and FAW-EV actuators.
horizon and control horizon, respectively.
In this paper, we omit the constant term of the cost func-
The control objective for path tracking can be expressed
tion; the optimization control problem (9) for path tracking
as the followed optimization control problem with a receding
is equivalently arranged into quadratic programming (QP)
horizon.
form as
Np
yref (k + i) − y(k + i) 2 min J = V T Hk V + Gk V (10a)
X
min J = Q
(9a)
i=1 s.t. 1Ũmin ≤ 1U ≤ 1Ũmax (10b)

17086 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

where (10b) is the integration of (9b)-(9d) and the optimiza- where a and NL denote the scale factor and term number,
tion control parameters defined as V = [1U , ε] are obtained respectively, of the discrete-time Laguerre function sequence,
by solving the QP problem as in (10). The coefficient matrices which possesses orthogonality as
are defined as 
1 Rπ
2k Q 2t 0
 T   0m (ejω )0m (ejω )dω = 1 , m = n
2π −π

Gk = 2EkT Q 2k 0
 
Hk = (13)
0 ρ  1 π 0m (ejω )dω0n (ejω ) = 0 , m 6 = n
 R
where Ek is the matrix consisting of predicted errors between 2π −π
the predicted system outputs and future reference outputs Since the inverse z-transform of the Laguerre function
within the prediction horizon. For further ensuring the track- does not become more concise, the discrete-time Laguerre
ing accuracy, during predicting future system outputs, the lin- function sequence is simply expressed in vector form as
earized error for achieving predictive plant (2) are considered
L(i) = [l1 (i) l2 (i) · · · lNL (i)]T (14)
in Ek , which can be described as
where li (k) represents the inverse z-transform of 0k (z, a),
Ek = YREF − Ysys − Eline
whose difference equation can be described as
where YREF is the matrix consisting of future reference out-
z−1 − a
puts within prediction horizon, Eline is the linearized error 0k (z) = 0k−1 (z) k = 1, 2, · · · NL (15)
matrix of system outputs within prediction horizon and can 1 − az−1
be achieved by Thus, the inverse z-transform discrete-time Laguerre func-
tion sequence satisfies the difference equation as
Eline = MDk
L(k + 1) = Al L(k) (16)
with the Dk is the matrix about the linearized error dk0 at the
next time. According to [37], dimension of matrix M depends with
on the length of prediction horizon, and its elements can be
Al(NL ×NL )
described as  
(
i−j
a 0 0 0 ···
C̃k,t Ãk,t , j ≤ i  β a 0 0 ···
M (i, j) =
j>i β
 
0, =
 −aβ 0 a ···
.. .. ..  .. ..

The dimension of 1U within V is Nc , and only the first . . . . .


element can be used as a control input by integral action as (−1) NL −2 NL −2
a β (−1) NL −3 NL −3
a β ··· β a
in (11).
where β = (1 − a2 ). Since a system H (i) can be fit as
u(k) = u(k − 1) + 1u(k |t ) (11)
H (i) = c1 l1 (i) + c2 l2 (i) + · · · + cNL lNL (i) (17)
B. LAGUERRE FUNCTION INTRODUCTION the control input increment at an arbitrary future time step
The principle of reducing the MPC computational complexity within the control horizon can be represented as
is to approximately represent the trajectory of future control
NL
input increments 1U (k) by linearly combining a sequence X
1u(k + i) = L(i) η =T
cj (k)lj (i), i = 0, 2, · · · Nc (18)
of orthogonal functions with few fitting coefficients. Conse-
j=1
quently, the whole control horizon can be covered without the
need for massive optimization control parameters. where η = [c1 c2 · · · cNL ]T and cj is the fitting coefficient.
Because of its programming simplicity and favorable Therefore, (8) can be approximately transformed as
approximation ability for variances of the control plant,
a Laguerre function orthogonal sequence is chosen as the Ysys (k) = 9k ξ ( k| t) + 2k KL η (19)
appropriate fitting agent for FAW-EV path tracking con- where the augmented matrix
troller [38]. The sequence is discretely described using the
L(0)T
 
following z-transformed expression.
√  L(1)T 
1 − a2 KL(Nc ×NL ) = 

..

(20)
01 (z) = .

1√− az−1
 
1 − a2 z−1 − a L(Nc − 1)T
02 (z) =
1 − az−1 1 − az−1 With the orthogonality described with (13), the QP
.. problem (10) can be simply rewritten as
.√
1 − a2 z−1 − a NL −1 min JL = VLT HL VL + GL VL (21a)
0NL (z) = ( ) (12)
1 − az−1 1 − az−1 s.t. 1Ũmin ≤ KL η ≤ 1Ũmax (21b)

VOLUME 7, 2019 17087

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

where the vector of optimization control parameter is changed ill-conditioning, resulting in high sensitivity to even slight
into VL = [η, ε], in which the dimension of η is NL . disturbances [39], [40].
The coefficient matrices of (21) are arranged as The condition number, a quantization index for system
ill-conditioning, can be computed by
KL 2Tk Q 2k KLT 0
 
GL = 2EkT Q 2k KLT 0
 
HL =
ρ γc (HL ) = λmax (HL ) λmin (HL )
  
0 (28)
(22) where λmax (HL ) and λmin (HL ) are maximal and minimal
Paper [26] gives the stability proof of the MPC with characteristic values, respectively. γc can be considered a
Laguerre function, and the control input increment at the magnified degree of the unexpected sensitivity caused by
current time can be achieved by the ill-conditioning. As Np elongates, γc (HL−1 ) dramatically
increases to a value that is far greater than ‘‘1’’ and tends to
1u(k |t ) = L(0)T η (23) infinity. This large value represents a high level of numeri-
cal sensitivity to disturbances and consequently reduces the
Remark 1: The scale factor a is closely related to the term
robustness of the path tracking control system.
number that is primarily required to fit 1U (k). While a = 0,
To address this issue, the exponential weight α −2j with
traditional MPC path tracking performance can be uniformly
decreasing characteristic is introduced to the cost function of
realized by setting NL equal to Nc [38]. For reducing the
the MPC path tracking controller. The introduction of α −2j
dimension of VL , we set 0 < a < 1; therefore, the similar
aims to restore a stability margin by placing less weighting
path tracking control performance can be achieved with the
on future tracking errors and control input increments, and
optimization control parameter dimension of NL + 1, which
the cost functions (9a) can be changed to
is far less than Nc + 1 within traditional MPC of (10). This
Np
advantage reduces the computational complexity without sac- X 2
α −2i yref (k + i |k ) − y(k + i |k ) Q

rificing the path tracking control performance. J =
i=1
Nc−1
C. SOLVING MATHEMATICAL PROBLEM X 2
+ α −2i k1u( k + i| k)kR + ρ ε 2 (29)
WITH EXPONENTIAL WEIGHT
i=0
Actually, the mathematical problem is inherent in MPC tech-
nology. The first derivative of the cost function JL described Both λmax (HL ) and λmin (HL ) can be converged into a
in (21a) is achieved by bounded region, significantly reducing the condition number
and therefore ameliorating the ill-conditioning of the path
∂JL tracking system with MPC, and paper [27] gives the feasi-
= 2HL VL + GL (24)
∂VL bility proof of the exponential weight introduction.
Without considering constraints, the equation Actually, the exponential weighted cost function acts
through the transformed control input increments and state
∂JL variables instead of cost function design; consequently,
=0 (25)
∂VL 1U and ξ change to
is necessary for minimizing JL ; thus, the optimal solution is 1Û = [1û(k) 1û(k + 1) 1û(k + 2) · · · 1û(k + Nc − 1)]T
.
VL = −HL−1 GL 2 (26) and
ξ̂ = [ξ̂ (k + 1)T ξ̂ (k + 2)T · · · ξ̂ (k + Np )T ]T
with the assumption of existing HL−1 , which is called the
Hessian matrix of the MPC algorithm. where
When predicting future system outputs within (21a), we set 1û(k + i) = α −i 1u(k + i)T = α −i L(i)T η
the convolution sum matrix φ = 2k KL , and the difference
ξ̂ (k + i) = α −i ξ (k + i)T
relationship between the convolution matrix elements is
This transformation is achieved by changing the predictive
Ãk,t φ(m − 1) + φ(1)(Am−1 )T , 0 ≤ m < N c

φ(m) = l (27) plant in (7) to
Ãk,t φ(m − 1) , Nc ≤ m ≤ Np
ξ̂ (k + 1 |t ) = Âk,t ξ̂ (k) + B̂k,t 1û(k) (30)
Because of the integral action in the MPC framework,
with

norms of the matrix powers Ãmk,t and kφk do not decay
. .
Âk,t = Ãk,t α, B̂k,t = B̃k,t α (31)
to zero as Np increases; thus, the Hessian matrix HL−1pos-
sesses a large magnitude. For vehicle path tracking control In summary, the LEMPC path tracking control can be
based on MPC, a long Np is always chosen to ensure vehicle formulated as an optimization control problem as
dynamic stability during severe driving conditions, but the
large magnitude of HL−1 accumulates strongly to the insta- min ĴL = VLT ĤL VL + ĜL VL (32a)
bility of the predictive plant and thereby leads to system s.t. 1Ũmin ≤ 3KL η ≤ 1Ũmax (32b)

17088 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

where ĤL and ĜL are achieved by introducing (31) into (7) Second, the system ill-conditioning is largely improved
and (22), respectively. The transformation matrix in (32b) is by introducing the exponential weight α −2j ; the system
defined as ill-condition gap between different Np values can also be
 
I 0 ··· 0 reduced considerably, so the Np can be adjusted online to
  different driving conditions without large fluctuation.
 0 α1I · · · 0 
  The regulation principles are illustrated through several
3= (33)
 
simulations in part C of Section IV
 .. .. .. .. 
. . . . 

 
E. MOTION REALIZATION FOR FAW-EV
0 0 ··· α N c −1 I
As the foundation of actualizing path tracking, a motion real-
Remark 2: To ensure closed-loop stability, α should be ization method with FAW-EV was presented and studied in
chosen to be slightly larger than ‘‘1’’. our previous paper [41]. To aid in understanding, the motion
The current control input of the LEMPC path tracking realization process is illustrated briefly here.
controller can be achieved by (11), with the current control The path tracking motion with FAW-EV is realized by a
input increment computed by (23). hierarchical structure that is functionally divided into three
tasks: desired motion computation, tire force distribution, and
D. PARAMETERIZATION ILLUSTRATION
actuator implementation.
Actually, the path tracking controller should adjust its per- The desired motion computation transforms the desired
formance to the variable driving condition. For the MPC front steering angle and desired vehicle velocities to total
path tracking controller with a basic framework, the value vehicle efforts, which includes Fxd , Fyd , and Mzd . For sta-
of Nc significantly affects the control performance, such as bility, the desired lateral vehicle velocity Vyd is set as 0. With
the tracking accuracy. The influence of Np also cannot be the desired longitudinal velocity Vxd preset and the desired
ignored. For example, an MPC path tracking controller with front steering angle δfd achieved by the LEMPC path tracking
long Np controls the vehicle with greater stability; however, controller, the desired motions are defined as
tracking accuracy is sacrificed. In contrast, a controller with
short Np often encounters challenges in ensuring dynamic 
ψ̇d = Vx lf + lr · δfd 1 + Kg Vx
2




stability during severe driving conditions.

Unfortunately for the path tracking controller based on the Vxd = Vxd (34)

traditional MPC, it is challenging to finely adjust the control Vyd = 0

performance because Nc and Np must be integers. Moreover,
in addition to the severe system ill-conditioning, the system where Kg is the understeering degree. For decoupling the
ill-condition gap between different Np is very large. These nonlinear relationship between motions on those three direc-
mathematical problems lead to violent fluctuation in the con- tions, SMC technology is employed to achieve the desired
trol input during Np adjustment, therefore largely reducing total efforts as (35), as shown at the bottom of this page.
the path tracking control performances of both the tracking In addition,
 p3 and q3 denote positive odd numbers, and
accuracy and vehicle dynamic stability. 1 < p3 q3 < 2, s1 , s2 , and s3 are the sliding mode sur-
Within the LEMPC path tracking controller, the parame- faces defined as minimizing differences away from desired
terization ability is mainly summarized as follows: motions on those three directions, and the function sat is
First, the Laguerre function possesses an attenuating prop- defined as
erty whose extent depends on the value of the scale factor a.
s φ , if |s | < φk
   
Such a Laguerre function with larger a needs more sample sk
sat = k k 
k (36)
points to decay to 0 [26]. The difference of the attenuating φk sgn sk φk , if |sk | ≥ φk
property determined by a becomes the key to finely regu-
late the control performance of the LEMPC path tracking For the tire force distribution, the desired total efforts
controller. achieved by (35) are divided into eight tire forces by solving

Fxd = m −Vy ψ̇ + V̇xd − η1n sat s1 φ1 , {Fxd | |Fxd | ≤ µmg}




(35a)
Fyd = m Vx ψ̇ − η2n sat s2 φ2 , Fyd | Fyd  ≤ µmg


  
(35b)

!

δfd lf + lr 2Kg Vx2 δfd lf + lr


Mzd = Iz ax + Vy ψ̇ −
1 + Kg Vx2

2
1 + Kg Vx2

2−p3 /q3 −α3n s3 − β3n sq33n /p3n
!!
Vx δ̇fd lf + lr α3 q3


+ + ψ̇ − ψ̇d −1 (35c)
1 + Kg Vx2 β3 p 3 α3 ψ̇ − ψ̇d

VOLUME 7, 2019 17089

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

the following QP problem: IV. SIMULATION RESULTS AND ILLUSTRATION

The simulations are carried out using the software of
r X2 + Y 2
r X
X w_ij w_ij Simulink combined with Carsim, in which a high-fidelity
min JCA = (37a)
i=l j=f
µ2ij Fz_ij
2 model of the FAW-EV is built. The major simulation param-
eters are listed in Table 1. The FAW-EV is controlled to
s.t. Aeq uCA = beq (37b)
track a double-change lane (DLC), with the desired outputs
Alim uCA ≤ blim (37c) defined by
where the vector of optimization control parameters is dy1 dy2
Yref (X ) = (1 + tanh(z1 )) − (1 + tanh(z2 ))
defined as 2 2
1 1.2
 T ψref (X ) = tan−1 (dy1 ( )2 ( )
uCA = Xw_lf Xw_rf Xw_lr Xw_rr Yw_lf Yw_rf Yw_lr Yw_rr cosh(z1 ) dx1
1 1.2
which consists of eight tire forces as shown in Fig. 4. (37a) is − dy2 ( )2 ( )) (43)
cosh(z1 ) dx2
the cost function defined as the sum of the tire workloads to
maximize the stable margin. According to the relationships
TABLE 1. Simulation parameters.
between the tire forces and total efforts described by (3),
the equivalent constraints are introduced as (37b). To ensure
tire stability, the linearized tire stability limit is introduced
as (37c), corresponding to

−µij Fz_ij < Xw_ij < µij Fz_ij ,

−µij Fz_ij < Yw_ij < µij Fz_ij ,
√ √
− 2µij Fz_ij < Xw_ij + Yw_ij < 2µij Fz_ij ,
√ √
− 2µij Fz_ij < Yw_ij − Xw_ij < 2µij Fz_ij (38)

Actuator implementation translates the tire forces to wheel where

driving torques and steering angles, which can be directly 
z1 = 2.4(X − 27.19) 25 − 1.2
realized by the actuators. First, δij ≈ σij is assumed according 
to the infinitesimal sideslip angle; with (4), it can be obtained z2 = 2.4(X − 56.46) 21.95 − 1.2
that and dx1 , dx2 , dy1 , and dy2 are set as 25, 21.59, 4.05, and

Fy_ij ≈ −Xw_ij sin σij + Yw_ij cos σij 5.7, respectively. The reason for choosing the DLC as the
(39) reference lane is that the multiple-curvature characteristic of
w_ij cos σij + Yw_ij sin σij
∗ =X
Fx_ij
it can feature different driving conditions by combing with
∗ denotes the estimated longitudinal tire force. With
where Fx_ij different vehicle velocities and road conditions.
the inversion of the tire model as (5), the sideslip angle of the The initial values for the QP problems of (32) and (37) are
tire can be computed by set as zero in this paper, and the Matlab function of ‘‘quad-
  prog’’ is used to solving the QP problems. The simulation
verification is mainly divided into three parts: first, control
µij

 kij Fy_ij
 performance and advantage of MPC path tracking controller
αij = −

tan 
µ · s  (40) with the incorporated Laguerre function are interpreted and
kij  ij
 ∗ ◦
 2
Fx_ij cos 22.5  verified in part A, and then part B proves the availability
Cα_ij 1 − µij Fz_ij for improving the path tracking control robustness by solv-
ing the mathematical problem with introduction of exponen-
The wheel steering angle is computed by tial weight. Finally, the parameterization ability and main
adjusted ideas for path tracking control with LEMPC are
δij = σij − αij (41)
interpreted by combining with several simulation results in
with σij achieved by (6). The wheel driving torque is part C.
computed by
A. INTRODUCTION OF THE LAGUERRE FUNCTION
Tw_ij = Jw_ij ω̇ij + Fx_ij Rw_ij + Tb_ij (42) For ensuring the close-loop stability and to consider practical
feasibility, we set a long Np as 36. With the constant friction
where ω̇ij , Rw_ij , Tb_ij and Jw_ij denote the wheel rotation rate, factor µij as 0.75, the simulations with different longitudinal
effective rolling radius, friction braking torque and wheel velocities of 15 m/s and 30 m/s are implemented.
rolling inertia, respectively. The longitudinal tire force Fx_ij Set the term number NL as 4 and scale factor a as 0.9; the Nc
is computed according to equation (4). of MPC introduced Laguerre function (LMPC) is set as 36.

17090 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 6. Path tracking errors for Vxd = 15 m/s, µij = 0.75. (a) Lateral position tracking error (b) Yaw angle tracking error.

FIGURE 7. Path tracking errors for Vxd = 30 m/s, µij = 0.75. (a) Lateral position tracking error. (b) Yaw angle tracking error.

FIGURE 8. Path tracking control inputs. (a) Vxd = 15 m/s, µij = 0.75. (b) Vxd = 30 m/s, µij = 0.75.

For outstanding performance of the proposed path tracking Fig. 6 and part of the Fig. 7 show that vehicles controlled by
controller, the traditional MPC controllers with control hori- the two MPC controllers with long Nc (LMPC and TMPC-36)
zon length of 4 (TMPC-4) and 36 (TMPC-36) are introduced possess higher path tracking accuracy than that controlled
as comparisons. The results are shown in Fig. 6 to Fig. 8. For by the traditional MPC controller with short Nc (TMPC-4).
in-depth researching of the tracking performance of different Notably, Fig. 7 shows a lower tracking accuracy of LMPC and
controllers, an index that quantifies the tracking performance TMPC-36 during simulation duration from 2nd to 4th second.
is given as Fig. 8(b) indicates the reason that the radical control motions
v
u tsim /Tq used to achieve high tracking accuracy reach the constraint
u P
2 boundary during severe driving conditions. Such control sat-
u yref (j) − ysys (j)
u
t j=1 uration leads to unexpected control performances [40], which
Qtrack_i =  , ∀i ∈ {ψ, Y } (44) influences the tracking index in Table 3 but can be mitigated
tsim Tq − 1
by adjustment of the Laguerre function parameters. Addition-
where tsim is the simulation duration time and Tq is the con- ally, the large tracking error mostly influenced by the long
troller sampling step. The tracking performance indexes of prediction horizon can be shrank by adjusting Np , which is
simulations for the three controllers are given in Table 2 and illustrated in the last part of this section.
Table 3, in which the dimension of the optimization control Discarding the control saturation, the higher tracking
parameter matrix V or VL is denoted as n. accuracy can be achieved by controllers with a long

VOLUME 7, 2019 17091

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

TABLE 2. Indexes for condition of Vxd = 15 m/s, µij = 0.75.

TABLE 3. Indexes for the condition of Vxd = 30 m/s, µij = 0.75.

FIGURE 9. Condition number of different prediction horizons.

disturbances with different amplitudes are typically imposed

to yaw angle signal at different simulation times (3 ψ at the
3rd second and 50 ψ at the 10th second). The results are
control horizon. Notably, Table 2 and Table 3 indicate that shown as Fig. 10 to Fig. 12.
relative to TMPC-36, the LMPC path tracking controller Fig. 12 comparably shows that for LMPC with a longer
can realize similar path tracking control performance with prediction horizon, the control input influence caused by
only 5 (computed by NL + 1) optimization control param- disturbances is much more severe than that with shorter Np .
eters, which is far less than that of TMPC-36 (with an This finding illustrates the incremental progression of sys-
optimization control parameter quantity of 37 computed by tem ill-conditioning as Np elongates, which conforms to the
Nc + 1). This result indicates that the proposed MPC con- results shown in Fig. 9. The large scale of vibration of the
troller with an incorporated Laguerre function can substan- control input indicates high sensitivity and weak robustness
tially reduce the computational complexity. of LMPC to the disturbance.
Moreover, it is observed that the LMPC can control the Relative to LMPC, it can be found in Fig. 10 and
vehicle to track the reference path with more accuracy than Fig. 11 that with two prediction horizon lengths, the LEMPC
that controlled by TMPC-36. The reason is because as NL path tracking controller with exponential weight possesses
increases, the control trajectory of LEMPC can be converged stronger robustness, which is embodied mainly in the slighter
to an underlying trajectory instead of the control trajectory of shake caused by the disturbance and in a faster return to the
the traditional MPC; this trajectory is interpreted in the last stable state.
part of this section. Additionally, with the two prediction horizon lengths,
the introduction of exponential weight can improve the
B. MATHEMATICAL PROBLEM IMPROVEMENT robustness without sacrificing tracking accuracy.
WITH EXPONENTIAL WEIGHT
For the MPC path tracking controller, the robustness deteri- C. PARAMETERIZATION ABILITY FOR ONLINE ADAPTION
oration by mathematical problems is closely related to Np , An important advantage of the proposed method is the param-
so this part sets the driving condition of the middle desired eterization of the LEMPC path tracking controller, which
longitudinal velocity (Vxd = 17 m/s) and friction factor provides access for fine online adjustment of the control
(µij = 0.5), and the control performance of MPC controller performance to driving conditions with different dynamic
with Laguerre function and exponential weight (LEMPC) and path tracking requirements. The parameterized variables
is compared with that of the controller without exponential are the Laguerre function parameters (NL , a) and prediction
weight (LMPC). horizon Np .
Since the mathematical problem becomes severe along First, we set different longitudinal velocities and friction
with Np elongating, first, for quantifying the ability to factors; with the reference path aforementioned, several sim-
solve the mathematical problem, we vary Np to increase ulations are implemented with different NL and a. Part A of
from 20 to 36, and the curves of condition number are shown this section indicates that the constraints dynamically affect
in Fig. 9. the control motion; for studying the influence of Laguerre
It is observed that the condition number is significantly function parameters without confusion, the constraints are
reduced by the proposed method (with 95.11% reduction at ignored, and the results are shown as Fig. 13 to Fig. 15, with
Np = 36), which means that the ill-conditioning has been the curve label form of ‘‘NL − a’’.
largely ameliorated with LEMPC. Fig. 13 shows that with the same value of NL , a larger a is
Then, for further proving the path tracking robustness associated with a higher tracking accuracy. Fig. 14 indicates,
improvement according to the relationship between Np and however, that for the FAW-EV path tracking during severe
the tracking system ill-conditioning, two sets of simulations conditions, the LEMPC path tracking controller with larger
with Np set to constants of 25 and 30 are implemented, and the a leads to a radical control input of the desired front steering

17092 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 10. Path tracking control performance for Np = 25, Vxd = 17 m/s, µij = 0.5, with introduced disturbances. (a) Lateral position tracking.
(b) Yaw angle tracking.

FIGURE 11. Path tracking control performance for Np = 30, Vxd = 17 m/s, µij = 0.5, with introduced disturbances. (a) Lateral position tracking.
(b) Yaw angle tracking.

FIGURE 12. Path tracking control inputs with introduced disturbances. (a) Control inputs for Np = 25. (b) Control inputs for Np = 30.

angle, which produces a very dangerous large yaw overshoot. set a low term number (NL ≤ 4); therefore the scale factor a
Therefore, online adjustment of a to different driving condi- can be regarded as a fine-tuning knob, which can be adjusted
tions is recommended for maximizing the control advantage online to the current dynamic stability and tracking accuracy
of the LEMPC path tracking controller. requirements.
Additionally, Fig. 15 shows that with the constant scale Then, it can be found in Fig. 9 that both the condition
factor a, as NL increases, the control performance converges number and the condition number gaps between different Np
to an underlying trajectory that is uniquely decided by the are significantly reduced by LEMPC. Considering constant
DLQR as soon as Q and R are identified; this is the reason Laguerre parameters (a = 0.9,NL = 2), switching Np at the
for the higher tracking accuracy with LMPC than that with 3rd second; the control inputs and control input increments of
TMPC-36 as discussed in part A of this section. Fig. 13 and LEMPC and traditional MPC are shown as Fig. 16. It shows
Fig. 15 indicate that with large NL , adjustment of a cannot that the fluctuation of control input during switching Np can
efficiently affect the path tracking control performance. be largely reduced by employing LEMPC.
As discussed above, for realizing online adapting control For illustrating the impact of Np on LEMPC path tracking
performance to the varied driving conditions, a regulated control performance, set a severe driving condition (Vxd =
principle for Laguerre parameters is simply given as follows: 20 m/s,µij = 0.4), the compared results of controllers with

VOLUME 7, 2019 17093

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 13. Path tracking control performances for different NL and a under conditions of Vxd = 15 m/s, µij = 0.5. (a) Lateral position tracking.
(b) Yaw angle tracking. (c) Lateral position tracking. (d) Yaw angle tracking.

FIGURE 14. Path tracking control performances for different a under severe conditions of Vxd = 22 m/s, µij = 0.3. (a) Control inputs of path
tracking controllers. (b) Yaw angle tracking.

FIGURE 15. Path tracking control performances for different NL under conditions of Vxd = 15 m/s, µij = 0.5. (a) Lateral position tracking.
(b) Yaw angle tracking.

different prediction horizon settings are shown in Fig.17, of controller with varied Np . It can be found that vehicle
in which the LEMPC controllers both with fixed Np of 20 and manoeuvred by controller with fixed short Np experiences
36 are set as the comparisons to highlight the advantage serious dynamic unstability during severe driving condition;

17094 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

FIGURE 16. Control input of the path tracking controller during switching Np . (a) Control input increments of the path tracking controllers.
(b) Control inputs of the path tracking controllers.

FIGURE 17. Path tracking control performances for different Np settings. (a) Lateral position tracking. (b) Yaw angle tracking.

controller with fixed long Np can ensure the dynamic stability path tracking controller. The proposed method significantly
of the vehicle, but during dynamic stable driving conditions, reduces the computational complexity without sacrificing
the tracking accuracy makes concession for the conservative path tracking control performance (the dimension of the opti-
dynamic stability; for controller with variable Np , the rela- mization control parameter matrix is reduced from 37 to 5),
tively high tracking accuracy is ensured by online choosing and the robustness of FAW-EV path tracking control sys-
a short Np , which is elongated online to a long length to tem is enhanced by solving the mathematical problem (with
enlarge stable margin only during dynamic unstable con- 95.11% reduction of ill-conditioning at Np = 36). Moreover,
ditions, therefore the vehicle dynamic stability can also be the proposed method provides parameterization, which is key
guaranteed. to adjust the path tracking control performance online for
In the last simulation, the Np is online adjusted at given diverse driving conditions.
simulation times, actually, which should be adjusted accord- The main idea for adjusting parameters of LEMPC path
ing to current or predicted future vehicle states and control tracking controller in this paper is only a preliminary work.
requirements. The advanced adjustment method should be For fully taking advantage of the adjustability of the LEMPC
studied in future work, and the main idea can be summarized path tracking controller, the relation between the vehicle per-
as follows: First, build the stability boundary in the stability formances (such as tracking accuracy and dynamic stability)
phase plane, and judge whether the current or future vehicle and parameters of LEMPC (NL , a, and Np ) should be finely
dynamic stability exceeds dynamic stable region. Then, if the quantized, and a more refined online-regulated method for
vehicle be manoeuvred within the dynamic stable region, adaptive path tracking control should be developed in future
the prediction horizon can be set as a minimum value; oth- work.
erwise, the prediction horizon should be online adjusted to a
relative long length to ensure the vehicle dynamic stability. REFERENCES
[1] H. N. de Melo, J. P. F. Trovão, P. G. Pereirinha, H. M. Jorge, and
C. H. Antunes, ‘‘A controllable bidirectional battery charger for elec-
V. CONCLUSION AND FUTURE WORK tric vehicles with vehicle-to-grid capability,’’ IEEE Trans. Veh. Technol.,
vol. 67, no. 1, pp. 114–123, Jan. 2017.
An improved path tracking controller based on LEMPC [2] B. Li, H. Du, and W. Li, ‘‘A potential field approach-based trajectory
for a FAW-EV is proposed. A parsimonious description for control for autonomous electric vehicles with in-wheel motors,’’ IEEE
the future control input increment trajectory is achieved by Trans. Intell. Transp., vol. 18, no. 8, pp. 2044–2055, Aug. 2017.
[3] J. Q. James and A. Y. S. Lam, ‘‘Autonomous vehicle logistic system: Joint
Laguerre function fitting, and an exponential weight is intro- routing and charging strategy,’’ IEEE Trans. Intell. Transp., vol. 19, no. 7,
duced to solve the mathematical problem inherent in the MPC pp. 2175–2187, Jul. 2018, doi: 10.1109/TITS.2017.2766682.

VOLUME 7, 2019 17095

 B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

[4] J. Ji, A. Khajepour, W. W. Melek, and Y. Huang, ‘‘Path planning and [25] C. Lian, X. Xu, H. Chen, and H. He, ‘‘Near-optimal tracking control of
tracking for vehicle collision avoidance based on model predictive con- mobile robots via receding-horizon dual heuristic programming,’’ IEEE
trol with multiconstraints,’’ IEEE Trans. Veh. Technol., vol. 66, no. 2, Trans. Cybern., vol. 46, no. 11, pp. 2484–2496, Nov. 2016.
pp. 952–964, Feb. 2017. [26] L. Wang, ‘‘Discrete-time MPC using laguerre function,’’ in Model Pre-
[5] H. Guo, J. Liu, D. Cao, H. Chen, R. Yu, and C. Lv, ‘‘Dual- dictive Control System Design and Implementation Using MATLAB,
envelop-oriented moving horizon path tracking control for fully auto- M. J. Grimble, Ed. London, U.K.: Springer, 2009, pp. 85–148.
mated vehicles,’’ Mechatronics, vol. 50, pp. 422–433, Apr. 2018, [27] L. Wang, ‘‘Use of exponential data weighting in model predictive control
doi: 10.1016/j.mechatronics.2017.02.001. design,’’ in Proc. 40th IEEE Conf. Decis. Control, Orlando, FL, USA,
[6] C.-L. Hwang, C.-C. Yang, and J. Y. Hung, ‘‘Path tracking of an autonomous Dec. 2001, pp. 4857–4862.
ground vehicle with different payloads by hierarchical improved fuzzy [28] J. Guo, Y. Luo, and K. Li, ‘‘An adaptive hierarchical trajectory following
dynamic sliding-mode control,’’ IEEE Trans. Fuzzy Syst., vol. 26, no. 2, control approach of autonomous four-wheel independent drive electric
pp. 899–914, Apr. 2018, doi: 10.1109/TFUZZ.2017.2698370. vehicles,’’ IEEE Trans. Intell. Transp. Syst., vol. 19, no. 8, pp. 2482–2492,
[7] H. Jiang and Y. Q. Liang, ‘‘Online path planning of autonomous UAVs for Aug. 2018.
bearing-only standoff multi-target following in threat environment,’’ IEEE [29] K. Worthmann, M. Reble, L. Grüne, and F. Allgöwer, ‘‘The role of
Access, vol. 6, pp. 22531–22544, 2018. sampling for stability and performance in unconstrained nonlinear model
[8] M. M. G. Plessen and A. Bemporad, ‘‘Reference trajectory planning under predictive control,’’ SIAM J. Control Optim., vol. 52, no. 1, pp. 581–605,
constraints and path tracking using linear time-varying model predictive Feb. 2014.
control for agricultural machines,’’ Biosyst. Eng., vol. 153, pp. 28–41, [30] S.-K. Oh, B. J. Park, and J. M. Lee, ‘‘Point-to-point iterative learning model
Jan. 2017. predictive control,’’ Automatica, vol. 89, pp. 135–143, Mar. 2018.
[9] S. Liu, C. Xu, and L. Zhang, ‘‘Hierarchical robust path following [31] J. Nilsson, J. Fredriksson, and A. C. E. Ödblom, ‘‘Reliable vehicle pose
control of fully submerged hydrofoil vessels,’’ IEEE Access, vol. 5, estimation using vision and a single-track model,’’ IEEE Trans. Intell.
pp. 21472–21487, 2017. Transp. Syst., vol. 15, no. 6, pp. 2630–2643, Dec. 2014.
[10] X. Ji, Y. Liu, X. He, K. Yang, X. Na, C. Lv, and Y. Liu, ‘‘Interactive control [32] B. Gutjahr, L. Gröll, and M. Werling, ‘‘Lateral vehicle trajectory opti-
paradigm-based robust lateral stability controller design for autonomous mization using constrained linear time-varying MPC,’’ IEEE Trans. Intell.
automobile path tracking with uncertain disturbance: A dynamic game Transp. Syst., vol. 18, no. 6, pp. 1586–1595, Jun. 2017.
approach,’’ IEEE Trans. Veh. Technol., vol. 67, no. 8, pp. 6906–6920, [33] P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, ‘‘Predictive
Aug. 2018. active steering control for autonomous vehicle systems,’’ IEEE Trans.
[11] N. Wang, Z. Sun, J. Yin, S. F. Su, and S. Sharma, ‘‘Finite-time Control Syst. Technol., vol. 15, no. 3, pp. 566–580, May 2007.
observer based guidance and control of underactuated surface vehicles [34] S. Sakai, H. Sado, and Y. Hori, ‘‘Dynamic driving/braking force distri-
with unknown sideslip angles and disturbances,’’ IEEE Access, vol. 6, bution in electric vehicles with independently driven four wheels,’’ Elect.
pp. 14059–14090, 2018. Eng. Jpn., vol. 138, no. 1, pp. 79–89, Jan. 2002.
[12] J. Taghia, X. Wang, S. Lam, and J. Katupitiya, ‘‘A sliding mode controller [35] B. Ren, H. Chen, H. Zhao, and L. Yuan, ‘‘MPC-based yaw stability
with a nonlinear disturbance observer for a farm vehicle operating in the control in in-wheel-motored EV via active front steering and motor torque
presence of wheel slip,’’ Auton. Robots, vol. 41, no. 1, pp. 71–88, Jan. 2017. distribution,’’ Mechatronics, vol. 38, pp. 103–114, Sep. 2016.
[13] A. A. Janbakhsh, M. B. Khaknejad, and R. Kazemi, ‘‘Simultaneous [36] M. N. Zeilinger, M. Morari, and C. N. Jones, ‘‘Soft constrained model
vehicle-handling and path-tracking improvement using adaptive dynamic predictive control with robust stability guarantees,’’ IEEE Trans. Autom.
surface control via a steer-by-wire system,’’ Proc. Inst. Mech. Eng., D, J. Control, vol. 59, no. 5, pp. 1190–1202, May 2014.
Automobile Eng., vol. 227, no. 3, pp. 345–360, Aug. 2012. [37] P. Falcone, F. Borrelli, J. Asgari, E. H. Tseng, and D. Hrovat, ‘‘Linear
[14] S. A. Tchenderli-Braham, F. Hamerlain, and N. Saadia, ‘‘Experimentations time-varying model predictive control and its application to active steering
on the adaptive sliding mode control for a trajectory tracking applied on a systems: Stability analysis and experimental validation,’’ Int. J. Robust
bi-steerable car,’’ Int. J. Vehicle Des., vol. 69, nos. 1–4, pp. 285–303, 2015. Nonlinear Control, vol. 21, no. 8, pp. 943–944, May 2011.
[15] P. Hang and X. Chen, ‘‘Integrated chassis control algorithm design for path [38] L. Wang, ‘‘Discrete model predictive controller design using Laguerre
tracking based on four-wheel steering and direct yaw-moment control,’’ functions,’’ J. Process Control, vol. 14, no. 2, pp. 131–142, Mar. 2004.
Proc. Inst. Mech. Eng., I, J. Syst. Control Eng., pp. 1–17, Oct. 2018. [39] T. T. V. Yap, A. H. Tan, and W. N. Tan, ‘‘Identification of higher-
doi: 10.1177/0959651818806075. dimensional ill-conditioned systems using extensions of virtual trans-
[16] A. Gray, Y. Gao, J. K. Hedrick, and F. Borrelli, ‘‘Robust predictive con- fer function between inputs,’’ J. Process Control, vol. 56, pp. 58–68,
trol for semi-autonomous vehicles with an uncertain driver model,’’ in Aug. 2017.
Proc. IEEE Intell. Vehicle Symp., Gold Coast, QLD, Australia, Jun. 2013, [40] N. Emmart, Y. Chen, and C. C. Weems, ‘‘Computing the smallest eigen-
pp. 208–213. value of large ill-conditioned Hankel matrices,’’ Commun. Comput. Phys.,
[17] Y. Cai, Q. Zhan, and X. Xi, ‘‘Path tracking control of a spherical mobile vol. 18, no. 1, pp. 104–124, Jul. 2015.
robot,’’ Mech. Mach. Theory, vol. 51, pp. 58–73, May 2012. [41] P. Song, M. Tomizuka, and C. Zong, ‘‘A novel integrated chassis controller
[18] J. Backman, T. Oksanen, and A. Visala, ‘‘Navigation system for agri- for full drive-by-wire vehicles,’’ Vehicle Syst. Dyn., Int. J. Vehicle Mech.
cultural machines: Nonlinear model predictive path tracking,’’ Comput. Mobility, vol. 53, no. 2, pp. 215–236, 2015.
Electron. Agricult., vol. 82, pp. 32–43, May 2012.
[19] P. Hang, X. Chen, F. Luo, and S. Fang, ‘‘Robust control of a four-wheel-
independent-steering electric vehicle for path tracking,’’ SAE Int. J. Veh.
Dyn., Stab. NVH, vol. 1, no. 2, pp. 307–316, Jul. 2017.
[20] C. Hu, R. Wang, F. Yan, and N. Chen, ‘‘Output constraint control on
path following of four-wheel independently actuated autonomous ground
vehicles,’’ IEEE Trans. Veh. Technol., vol. 65, no. 6, pp. 4033–4043,
Jun. 2016.
[21] C. J. Ostafew, A. P. Schoellig, and T. D. Barfoot, ‘‘Robust constrained BING ZHANG received the B.Eng. degree
learning-based NMPC enabling reliable mobile robot path tracking,’’ Int. from the Humanities and Information College,
J. Robot. Res., vol. 35, no. 13, pp. 1547–1563, 2016.
Changchun University of Technology, Changchun,
[22] Y. Yoon, J. Shin, H. J. Kim, Y. Park, and S. Sastry, ‘‘Model-predictive active
China, in 2012, and the M.Eng. degree from the
steering and obstacle avoidance for autonomous ground vehicles,’’ Control
Eng. Pract., vol. 17, no. 7, pp. 741–750, Jul. 2009.
Changchun University of Technology, in 2015.
[23] G. Yin, J. Li, X. Jin, C. Bian, and N. Chen, ‘‘Integration of motion plan- She is currently pursuing the Ph.D. degree with
ning and model-predictive-control-based control system for autonomous the Department of State Key Laboratory of Auto-
electric vehicles,’’ Transport, vol. 30, no. 3, pp. 353–360, Oct. 2015. motive Simulation and Control, Jilin University,
[24] T. Oda, K. Nonaka, and K. Sekiguchi, ‘‘Robust path tracking control using Changchun.
model predictive control and sliding mode control—Application to the Her research interests include control-by-wire
JSAE-SICE benchmark problem,’’ in Proc. 54th Annu. Conf. Soc. Instrum. vehicle control, autonomous vehicle control, and vehicle dynamic stability
Control Eng. Japan (SICE), Jul. 2015, pp. 1332–1336. control.

17096 VOLUME 7, 2019

B. Zhang et al.: EV Path Tracking-Based MPC With a LEMPC

CHANGFU ZONG received the B.Eng. degree BANGCHENG ZHANG received the B.Eng. and
from the Liaoning University of Technology, M.Eng. degrees from the Changchun University
Jinzhou, China, in 1986, and the M.Eng. and of Technology, Changchun, China, in 1995 and
Ph.D. degrees from the Jilin University of Tech- 2004, respectively, and the Ph.D. degree from Jilin
nology, Changchun, China, in 1994 and 1998, University, Changchun, in 2011.
respectively. He is currently a Professor with the Changchun
He is currently a Professor with the Department University of Technology. He has been an Aca-
of State Key Laboratory of Automotive Simulation demic Visitor with Tsinghua University, Beijing,
and Control, Jilin University, Changchun. He has China, in 2007. He has published over 20 arti-
been an Academic Visitor with Cambridge Univer- cles. His research interests include mechatronics
sity, U.K, in 2005, and a Senior Academic Visitor with the University of measurement technique and fault diagnosis.
California at Berkeley, USA, in 2013. He has published over 200 articles.
His research interests include vehicle control stability, new energy vehicles,
and autonomous vehicle control.

GUOYING CHEN received the B.Eng. degree

from China Agricultural University, Beijing,
China, in 2005, the M.Eng. degree from Xinhua
University, Chengdu, China, in 2009, and the
Ph.D. degree from the Jilin University of Technol-
ogy, Changchun, China, in 2012.
He is currently a Professor with the Department
of State Key Laboratory of Automotive Simu-
lation and Control, Jilin University, Changchun.
His research interests include vehicle dynamics
simulation and control.

VOLUME 7, 2019 17097