Proceedings of the 2008 IEEE/ASME

International Conference on Advanced Intelligent Mechatronics

July 2 - 5, 2008, Xi'an, China

Regenerative-Braking Sliding Mode Control of Electric

Vehicle Based on Neural Network Identification
Jianbo Cao, Binggang Cao, Peng Xu, and Zhifeng Bai
Research & Development Center of Electric Vehicle
Xi’an Jiaotong University
Xi’an, Shaanxi Province, China
caojianbo2008@163.com

Abstract – Aiming at the main problems of electric vehicle so as to improve the energy-regenerative efficiency and the
(EV): short driving range, short life of batteries, and variation of stability of EV.
model parameters, based on constructing the main circuit Due to the advantages of simple structure, easy to use,
diagram of the EV’s control system, the mathematical model of PID controller has been widely used in the field of EVs, yet it
regenerative-braking process was established, and a novel
is unsuitable for the target which is nonlinear and uncertain.
regenerative-braking controller was designed, which combined
neural network (NN) with traditional sliding mode controller As we know, EV’s working condition is complicated and
(SMC). The controller comprises a back propagation NN changeful, so the research on advanced control algorithm for
(BPNN), a radial basis function NN (RBFNN) and a SMC. The EV is necessary. J. C. Li [3] designed a sliding mode
BPNN is used to adaptively adjust the switching gain of the SMC controller (SMC) for EV by using rotational speed and
on-line so as to avoid the whippings. The RBFNN is used to acceleration as the state variables, which could improve the
perform system identification and parameter prediction. The robustness and real-time performance of the system. Yet it
experimental results show that the NNSMC could improve the could not resist disturbances while S-function is on the way of
stability and reliability of the system, and is superior to approaching sliding plane. J. Moreno, et al. [4] designed
traditional SMC at response speed, steady-state tracking error
energy-management system for a hybrid EV by using
and resisting disturbance in the regenerative-braking process.
Additionally, it can recover more energy, lengthen batteries’ life, ultracapacitors and neural networks, which has the merits of
and increase the driving range than SMC by about 6%. This is not needing the accurate model of controlled target. Yet it
very significant for saving energy. doesn’t have the strong robustness of SMC.
In this paper, neural network (NN) was combined with
Index Terms - Electric vehicle. Regenerative braking. Neural SMC so as to optimally control the EV. NN as an intelligent
network. Sliding mode control control algorithm is widely known for its strong capacities of
self-learning, self-adapting and self-organization, and it is
I. INTRODUCTION outstanding for the control of nonlinear systems [5]. The novel
It has become a general trend of the development of NNSMC designed in this paper comprises a back propagation
electric vehicles (EVs), because of the dual pressure from NN (BPNN), a radial basis function NN (RBFNN) and a
environmental pollution and energy crisis [1]. EVs have SMC. The BPNN is used to adaptively output the switching
improved their performance and made suitable for commercial gain adjustment of the SMC on-line and ensure the robustness
and domestic use during the last decades. At present, EVs while S-function is approaching the sliding plane. The
commonly comprise five parts: control system, motor, RBFNN as on-line identifier is used to establish nonlinear
batteries, bodywork and charger. Control system as EV’s brain prediction model and perform parameter prediction. The
is the key to improve motor efficiency, lengthen batteries’ life, prediction model established by RBFNN can approximate any
increase driving range and optimize control, so it to a great nonlinear function with arbitrary precision, and has the merits
extent determines the entire vehicle performance. As we of simple structure and fast convergence [6]. The intelligent
know, the problem that EVs still have not achieved ranges as composite controller with NN and SMC was applied to the
good as gas-powered conventional vehicles restricts EVs fast regenerative-braking process of EV. Additionally, SMC was
developing. It is very significant for the increase in range that also provided in this paper, and the experimental comparison
the technology of energy-regenerative baking is applied to between SMC and NNSMC was done and analyzed. We found
EVs [2]. For permanent magnet motor can work as a dynamo that NNSMC could obtain better control effect and recover
in braking process, this paper designed an intelligent control more energy.
system for EV, which can recover electrical energy to batteries II. CONTROL SYSTEM DESIGN FOR EV
from vehicle’s kinetic energy while braking, and has the
advantages: saving energy, increasing driving range, having In this paper, the EV employed a brushless DC motor
no mechanical braking noise. Nevertheless, the road state and (BLDCM), which has been widely applied in the field of EVs,
batteries’ voltage are variational in the regenerative-braking due to a series of advantages: simple structure, reliable
process, so the optimal control strategy should be researched performance, high efficiency, large starting torque, etc [7].
The control system of EV designed in this paper is shown

This project was supported by the innovation funding of Shaanxi Province, China

978-1-4244-2495-5/08/$25.00 © 2008 IEEE. 1219

in Fig. 1, which includes power converter module, control
chip module, three-phase inverter module, parameters
detection module, driving module, braking module, flash
program online module, panel display module, etc.
III. REGENERATIVE-BRAKING PROCESS AND ITS
MATHEMATICAL MODEL [8][9]
While braking, EV works in the regenerative state, and
BLDCM works as a dynamo. The main circuit topology of the
control system is shown in Fig. 2, includes four batteries, a
three-phase full bridge circuit and a BLDCM. In this paper,
the three-phase BLDCM works according to six states, and in
any state, two-phase work principle is the same. The BLDCM
could be seen as a DC motor under the condition of two Fig. 1 EV’s control system
phases conducting. In this, we take phases A and B as an
example to establish the regenerative-braking equivalent
circuit and mathematical model. While energy-regenerative
braking, T2 is controlled with PWM, other MOSFETs are off,
the equivalent circuit topology and the current flow direction
are shown in Fig. 3. At this time, batteries are charged by
boost converter. In the whole regenerative-braking process,
T2, T4, T6 are controlled with PWM in turn according to the
Hall sequence, other MOSFETs are off. As a result, the
vehicle’s kinetic energy is converted into electrical energy,
which can avoid the waste of energy, lengthen batteries’ life,
increase the driving range, and doesn’t have mechanical Fig. 2 Regenerative-braking circuit topology
braking noise.
The regenerative-braking mathematical model can be
established as follows (take Fig. 3 as an example):
While T2 is in conducting state and cut-off state, the state
equations are respectively described as:
T2_on ( 0 ≤ t ≤ d ⋅ T )
di
( La + Lb ) m = ea + eb − im ⋅ (ra + rb + r4 + rt ) (1) Fig. 3 Equivalent circuit topology as boost converter
dt
T2_off ( d ⋅ T ≤ t ≤ T ) ­ ª 2rm + rd + rt ⋅ d + (rd + rbat ) ⋅ (1 − d ) Ke º ª 1− d º
° «− 2 Lm Lm » « − 2 L vb »
di °° x = « »x+« m
»
( La + Lb ) m = ea + eb − vb − im ⋅ (ra + rb + r1 + r4 + rbat ) (2) « » «
dt ® K TL »
° « − t 0 » « − »
where d is the duty cycle of T2_PWM, T is the operation ¬ J ¼ ¬ J ¼
°
period, rbat is the equivalent series resistance of the batteries. °̄ y = [1 − d 0] x
Because the motor’s three phases are similar, in permitted (7)
scope, we have: La = Lb = Lm , ea = eb = em , ra = rb = rm , Suppose intermediate variable is u = d . After handling by
r1 = r4 = rd , then (1) and (2) can be simplified as: perturbation method and separation of steady-state variables
and instant variables, the linear small-signal mathematical
dim model of regenerative braking can be gotten as follows:
2 Lm = 2em − im ⋅ (2rm + rd + rt ) (3)
dt ­ ª 2rm + rd + rt ⋅ D + ( rd + rbat ) ⋅ (1 − D) K e º
dim ° «− Lm »
2 Lm = 2em − vb − im ⋅ (2rm + 2rd + rbat ) (4) ° x = « 2 Lm
»x
dt ° « Kt »
According to the deflecting couple equilibrium, have: ° « − 0 »
° ¬ J ¼
dω ®
J = − Kt ⋅ im − TL (5) r +
ª d bat t
r − r º ª b º
v
° 0»
° +« « 2L » ⋅ u
dt
«
2 Lm
» ⋅ X ⋅ u + « m»
em = K eω (6) °
° «
¬ 0 0 »
¼ ¬« 0 ¼»
Suppose state variables are x = [im ω ] , output is y = ib ,
T
° y = [1 − D 0] x + [ −1 0] ⋅ X ⋅ u (8)
¯
after average processing, then:

1220
where D and X are respectively the circuit steady values of B. SMC Design [11]
d and x . Consider the uncertainty of the control system, its state
equation could be expressed as:
IV. REGENERATIVE-BRAKING NNSMC DESIGN FOR EV
x = ( A + ΔA) x + ( B + ΔB )u + ( P + ΔP ) f (9)
A. NNSMC Introduction
NNSMC combines NN and SMC together, and has the where x ∈ R n , u ∈ R , f ∈ R , ΔA and ΔB are the variable
advantages of both. The control target changes from tracking uncertain parameters, P and ΔP are the disturbances.
Suppose the system satisfies the match condition:
error to sliding mode function, the input is not [e, e, 
( n -1)
e..., e ] ,
yet is ( s , s) , which is always two-dimensional. The system is ΔA = B ⋅ A , ΔB = B ⋅ B , P + ΔP = B ⋅ P
controlled by making S-function tend to zero, and the tracking Then:
error will also tend to zero. Additionally, NNSMC doesn’t x = ( A + B ⋅ A ) x + ( B + B ⋅ B )u + ( B ⋅ P ) f . (10)
require the precise mathematical model, and could soften
control signal to reduce or avoid the whippings that common The system can be simplified as:
SMC has, so it has better robustness [10]. x = Ax + B ⋅ [u (t ) + E (t )] (11)
The NNSMC designed in this paper uses the switching
gain adjustment to eliminate the disturbance according to the where E (t ) = A ⋅ x + B ⋅ u (t ) + P ⋅ f , includes the uncertainty
condition of sliding mode motion, thereby could avoid the and disturbance.
whippings and improve the stability and reliability of the From the regenerative-braking mathematical model like
system. The structure of NNSMC is shown in Fig. 4, which is (8), have:
made up of three parts: a neural network controller (NNC), a
x = [ x1 x2 ] = [im ω] ,
T T
neural network identifier (NNI) and a SMC. NNC employs a
BPNN, and NNI employs a RBFNN. The SMC is directly ªA A12 º
responsible for controlling the motor in the closed loop of A = « 11 »
current. The NNC adaptively adjusts the parameters of SMC ¬ A21 A22 ¼
on-line according to the system state so that the optimal ª 2rm + rd + rt ⋅ D + (rd + rbat ) ⋅ (1 − D) Ke º
control effect can be achieved. The NNI establishes nonlinear «− 2Lm Lm »
prediction model and performs parameter prediction on-line =« »,
for NNC. « K »
« − t 0 »
In order to further speed up the convergence rate of NN, ¬ J ¼
this paper employed variable-rate learning algorithm, which
ª rd + rbat − rt º ª vb º
could adaptively adjust the learning rate η (t ) and η ′(t ) on- ª B1 º « 0» « 2L »
B=« »=« 2 Lm
»⋅ X + « m » .
line according to the error’s value in convergence process. B
¬ 2¼ «
That is, if current error is bigger than the last one, then this ¬ 0 0 ¼» ¬« 0 ¼»
iteration is ineffective, and the learning rate should be reduced Suppose the uncertainty and disturbance can be expressed
to reiterate; otherwise, then this iteration is effective, and the as:
learning rate should be increased. The algorithm could further
improve the system’s dynamic response speed. K ( t ) = max( B1 ⋅ E ( t ) ) + ε (12)
The off-line training of NN is also important to reduce on- where ε >0.
line computing time and improve the response speed. The In order to satisfy global sliding mode condition, define
parameters choice determines the network’s performance. The
e = x1d − x1 , where x1d is the reference state vector of x1 ,
training data can be acquired from the experiments. After off-
line training, NN can be constructed and initialized. Then and define the S-function as follows:
control system carries out on-line training and on-line control, s = e + c ⋅ e − F (t ) (13)
which could weaken or eliminate the disturbance influence on
the control system. where c >0, and F (t ) must satisfy the following conditions:
(1) F (0) = e0 + c ⋅ e0 ;
s ŷ
(2) When t → ∞ , F (t ) → 0 ;
s
(3) F (t ) has one order derivative.
− λt
So define F (t ) = F (0)e , where λ >0.
The sliding mode control rule could be designed as:
−1 −1
u (t ) = B1 [c ⋅ e + x1d − A11 ⋅ x1 − A12 ⋅ x2
(14)
+ K (t ) ⋅ sgn( s) − c ⋅ F (t )]
−1
Fig. 4 NNSMC diagram

1221
 In order to verify the stability of the S-function, define negative sigmoid function as its activation function, yet the
Lyapunov function as: hidden layer chooses tangent sigmoid function as its activation
1 2 function.
V = s . (15) For the input layer, there is
2
Then: j (t ) = x ( j )
O (1) j = 1, 2 (18)

V = ss = s[e + c ⋅ e − F (t )] where O (1)

j (t ) is the output of this layer.

And for the hidden layer, there is

= s[e + c ⋅ x1d − c ⋅ x1 − F (t )]
(16) ­ (2) 2

= s[
e + c ⋅ x1d − c ⋅ A11 ⋅ x1 − c ⋅ A12 ⋅ x2 − c ⋅ B1 ⋅ u (t ) N
° i ( t ) = ¦ wij(2) O (1)
j (t )
® j = 1 (19)
− c ⋅ B1 ⋅ E (t ) − F (t )] °O (2) (t ) = f [ N (2) (t )] i = 1, 2,3, 4,5
¯ i i

When the u (t ) is substituted by (14), (16) can be where Oi(2) (t ) is the output of the hidden layer, wij(2) is the
simplified as:
weight between the input layer and the hidden layer, and f [⋅]
V = s[ − c ⋅ K (t ) ⋅ sgn( s ) − c ⋅ B1 ⋅ E (t )] is the hidden layer’s activation function as follows:
= − c[ K (t ) ⋅ s + B1 ⋅ E (t ) ⋅ s ] (17) e x − e− x
f [⋅] = tanh( x) = x . (20)
e + e− x
≤ −c ⋅ ε ⋅ s
The input and output of the output layer are
Because c >0, ε >0, we can get V ≤ 0 , and the S- ­ (3) 5

° Nl (t ) = ¦ wli Oi (t )
(3) (2)

function is stable. ° i =1
In the control rule (14), K (t ) is the source that causes ®O (3) (t ) = g[ N (3) (t )] l =1
° l l
whippings, and is used to counteract E ( t ) so that the sliding ° K (t ) = O (t )
(3)
(21)
¯ l
mode condition could be satisfied. If E ( t ) is time-varying,
where wli(3) is the weight between the hidden layer and the
K (t ) should be also time-varying, and can make the sliding
output layer, and g[⋅] is the output layer’s activation function
mode motion approach the sliding mode plane. Its value
should be great enough to eliminate the influence caused by as follows:
the uncertain and disturbance. 1 ex
g[⋅] = [1 + tanh( x)] = x . (22)
C. BPNN Design for Adjusting SMC 2 e + e− x
In order to improve the performance of the control The quadratic performance index of error is used to
system, BPNN is employed as NNC to adaptively adjust the modify the network weights, and defined as follows:
switching gain K (t ) of SMC on-line according to the system 1
J1 = [r (t ) − y (t )]2 . (23)
state. 2
In this paper, BPNN is a three-layer neural network. Its For realizing the optimal control of SMC, the
structure is shown in Fig. 5. The BPNN is fixed as 2-5-1, that modification of weights should be searched along the negative
is, two input nodes, five hidden nodes and one output node. grads direction of J1 . So the rule [12] is given as
The input nodes are defined as: x = [ s s]T . And the output ∂J1
node is the switching gain of SMC: K (t ) . Considering the Δwli(3) (t ) = −η (t ) + αΔwli(3) (t − 1) (24)
∂wli(3)
positive characteristic, the output layer could choose non-
where η (t ) is the learning rate, α is the momentum factor,
and
∂J1 ∂J1 ∂y (t ) ∂u (t ) ∂Ol(3) (t ) ∂N l(3) (t )
= ⋅ ⋅ ⋅ ⋅ (25)
∂wli(3) ∂y (t ) ∂u (t ) ∂Ol(3) (t ) ∂N l(3) (t ) ∂wli(3) (t )
x1
# Because ∂y (t ) / ∂u (t ) is unknown, in this paper, we
x2 adopted the output value ∂yˆ(t ) / ∂u (t ) of the prediction model
K (t )
# #
based on NNI to substitute ∂y (t ) / ∂u (t ) .
∂u (t ) / ∂Ol(3) (t ) can be gotten from (14):
xj
∂u (t ) ∂u (t )
= = B1−1 ⋅ sgn( s ) (26)
∂Ol(3) (t ) ∂K (t )
From (18) to (26), the output layer’s modified weight
Fig. 5 BPNN structure formulas can be deduced as follows:

1222
 ­ Δwli(3) (t ) = η (t )δ l(3) Oi(2) (t ) + αΔwli(3) (t − 1) where w j represents the weight between the jth hidden node
° and the output node.
® (3) ∂yˆ(t ) ∂u (t ) (27)
°δ l = [r (t ) − y (t )] ⋅ ∂u (t ) ⋅ ∂O (3) (t ) ⋅ g ′[ Nl (t )] l =1
(3)
The function of quadratic performance index is:
¯ l
1
where g ′[⋅] = 2 ⋅ g ( x)[1 − g ( x)] . J 2 = [ y (t ) − yˆ (t )]2 (31)
2
The same as above, the hidden layer’s modified weight According to the above performance index and the
formulas can be gotten as: gradient descent method, the iterative algorithms of w j , b j
°­ Δwij (t ) = η (t )δ i O j (t ) + αΔwij (t − 1)
(2) (2) (1) (2)
and c ji are shown as:
® (2) (28)
°̄δ i = f ′[ N i (t )]δ l wli (t ) i = 1, 2,3, 4,5
(2) (3) (3)
w j (t ) = w j (t − 1) + η ′(t )[ y (t ) − yˆ (t )]h j
(32)
where f ′[⋅] = 1 − f ( x) .
2 +α ′[ w j (t − 1) − w j (t − 2)]
2
D. RBFNN Design for on-Line Identification X −Cj
To get better control effect, this paper acquired b j (t ) = b j (t − 1) + η ′(t )[ y (t ) − yˆ (t )]w j h j
b3j (33)
∂yˆ(t ) / ∂u (t ) through NNI to substitute the ∂y (t ) / ∂u (t ) in
(25). +α ′[b j (t − 1) − b j (t − 2)]
In this paper, the RBFNN as NNI shown in Fig. 6 is also a xi − c ji
three-layer neural network, which is fixed as 3-5-1. RBFNN c ji (t ) = c ji (t − 1) + η ′(t )[ y (t ) − yˆ (t )]w j h j
b 2j (34)
can reflect the input locally via the function of hidden layer,
namely the radial basis function. The output of the whole +α ′[c ji (t − 1) − c ji (t − 2)]
network is the weight sum of hidden nodes’ outputs. The
where i = 1, 2,3 , j = 1, 2,3, 4,5 , η ′(t ) is the learning rate, and
mapping from the input to the output is nonlinear as a whole
[13]. α ′ is the momentum factor.
In this, The input vector is defined as: The gradient parameter ∂yˆ(t ) / ∂u (t ) can be gotten as
X = [u (t ) y (t ) y (t − 1)]T . Considering the convenience of bellow, which is transferred to the BPNN.
identification, the output layer uses linear activation function ∂y (t ) ∂yˆ (t ) 5 c j1 − u (t )
while the hidden layer’s activation function is the nonlinear ≈ = ¦ wj hj (35)
Gauss function, namely radial basis vector: ∂u (t ) ∂u (t ) j =1 b 2j
H = [h1 h2 h3 h4 h5 ] . The hidden layer’s output
T
V. EXPERIMENTS AND ANALYSES
formula is
2 The controller designed in this paper was tested at the EV
X −Cj made in Xi’an Jiaotong University Boyuan EV Company. The
h j = exp(− ) j = 1, 2,3, 4,5 (29)
2b 2j regenerative-braking experiments were respectively done with
the SMC and the NNSMC. In addition, for different control
where h j is the output of the jth hidden node, b j is the width
methods, the comparative experiments of the driving range
of the jth hidden node, b j > 0 , and C j is the center vector of were also done.
the jth hidden node, C j = [c j1 c j2 c j 3 ]T . A. Regenerative-Braking Control Experiments
The output of this neural network is At the initial speed of 25 km/h, the regenerative-braking
5 experiments were respectively done with the SMC and the
yˆ(t ) = ¦ w j h j (30) NNSMC. The experimental results are shown as Fig. 7, where
j =1 (a) is the charge current waveform with SMC, and (b) is the
charge current waveform with NNSMC. To facilitate analysis
and comparison, the output voltage of sensor was converted to
h1 the current value. We used approximate step response to test
x1
w1 the performance of the controllers. The reference value of the
h2 w2 charge current is: ibref =5 amperes. As we see, during the
x2
Σ ŷ regenerative-braking process, the actual value of the charge
# # wj current with the SMC is smaller than the reference one and has
big undulation. Yet the NNSMC has better performance at
xi response speed, steady-state tracking error, and resisting
hj
disturbance than the SMC, because it has good robustness and
self-adaptability. As a result, it can recover more energy, and
Fig. 6 RBFNN structure improve the energy-regenerative efficiency and the stability of
EV.

1223
B. Driving Range Experiments VI. CONCLUSIONS
The important aim of researching the technology of
In this paper, NNSMC was designed and applied to the
energy-regenerative baking is to increase the driving range of
regenerative-braking process of EV, and a lot of experiments
EV. The experimental results under different conditions are
were done. The following conclusions can be gotten from
compared and shown in Fig. 8. The experiments were
them:
performed at the same EV by braking one time per 500
Due to using on-line parameter adjustment and system
meters. The voltage of batteries was recorded one time per 3
identification, NNSMC possesses strong capacities of self-
km. The initial voltage of the batteries was 54 volts, and the
learning, self-adapting and self-organization. In the
final voltage was 42 volts. As shown in Fig. 8, the NNSMC
regenerative-braking process, comparing with traditional
with regeneration can increase more driving range than the
SMC, NNSMC has better performance at response speed,
SMC without regeneration by about 24%, and than the SMC
steady-state tracking error and resisting disturbance, and can
with regeneration by about 6%. The proposed method has the
improve the stability and reliability of EV. In addition, the
application value.
NNSMC with regeneration can recover more energy in the
braking process, consequently it could lengthen batteries’ life,
and increase more driving range than the SMC without
Batteries Current (A)

regeneration by about 24%, and than the SMC with

regeneration by about 6%. This is very significant for saving
energy.
REFERENCES
[1] B. G. Cao, C. W. Zhang, Z. F. Bai, and J. C. Li, “Technology progress
and trends of electric vehicles,” Journal of Xi’an Jiaotong University, vol.
38, no. 1, pp. 1-5, Jan. 2004.
[2] N. Mutoh, Y. Hayano, H. Yahagi, and K. Takita, “Electric braking control
methods for electric vehicles with independently driven front and rear
(a) SMC wheels,” IEEE Transactions on Industrial Electronics, vol. 54, no. 2, pp.
1168-1176, Apr. 2007.
[3] J. C. Li, Study of Drive and Regenerate Brake Control for an Electric
Vehicle, Xi’an: Mechanical Engineering School of Xi’an Jiaotong
University, 2003.
[4] J. Moreno, M. E. Ortúzar, and J. W. Dixon, “Energy-management system
for a hybrid electric vehicle, using ultracapacitors and neural networks,”
IEEE Transactions on Industrial Electronics, vol. 53, no. 2, pp. 614-623,
Apr. 2006.
[5] W. Wu, Neural Network Computing, Beijing: Higher Education Press,
2003.
[6] D. Langoni, M. H. Weatherspoon, S. Y. Foo, and H. A. Martinez, “A
speed and accuracy test of backpropagation and RBF neural networks for
small-signal models of active devices,” Engineering Applications of
Artificial Intelligence, vol. 19, no. 8, pp. 883-890, Dec. 2006.
(b) NNSMC [7] L. Y. Liao, Z. F. Wang, Y. D. Deng, and X. L. Hu, “Simulation of a
Fig. 7 Regenerative-braking control experiments permanent magnet brushless DC motor for electric vehicle using the bond
graph method,” Journal of WUT, vol. 25, no. 6, pp. 192-195, Dec. 2003.
[8] Z. F. Bai, Research and Development on Ultracapacitor-Battery
Composite Power System and Its H Robust Control of Electric Vehicle,
Xi’an: Mechanical Engineering School of Xi’an Jiaotong University,
2006.
[9] H. X. Wu, S. K. Cheng, and S. M. Cui, “A controller of brushless DC
motor for electric vehicle,” IEEE Transactions on Magnetics, vol. 41, no.
1, pp. 509-513, Jan. 2005.
[10] C. F. Zhang, Research on Intelligent Control Theory and Applications of
Sliding Mode Variable Structure, Changsha: Hunan University, 2001.
[11] J. K. Liu, MATLAB Simulation for Sliding Mode Control, Beijing:
Tsinghua University Press, 2005.
[12] C. L. Xia, M. Xue, and Z. R. Chen, “Adaptive PID control and on-line
identification for switched reluctance motors based on BP neural
network,” in Proc. The IEEE International Conference on Mechatronics
& Automation (ICMA), Niagara Falls. Canada, Jul. 2005, pp. 1918-1922.
[13] C. L. Xia and J. Xiu, “RBF ANN nonlinear prediction model based
adaptive PID control of switched reluctance motor,” Proceedings of the
CSEE, vol. 27, no. 3, pp. 57-62, Jan. 2007.

Fig. 8 EV’s driving-range experimental comparison

1224