This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
A lithium-ion battery-in-the-loop approach to
test and validate multi-scale dual H infinity
filters for state of charge and capacity estimation
Cheng Chen, Rui Xiong, Senior member, IEEE, Weixiang Shen, Member, IEEE
 effectively in EVs [3], [4]. However, the estimation of battery
Abstract—An accurate battery capacity and state estimation
method is one of the most significant and difficult techniques to
ensure efficient and safe operation of the batteries for electric
vehicles (EVs). Since capacity and state-of-charge (SoC) are
strongly correlated, the SoC is hardly to be accurately estimated
without knowing accurate battery capacity. Thus, a multi-scale
dual H infinity filters has been proposed to estimate battery SoC
and capacity in real-time with dual time scales in response to
slow-varying battery parameters and fast-varying battery state.
The proposed method is first evaluated and verified using offline
experimental data and then compared with the single/multi-scale
dual Kalman filters. The results show that the proposed
multi-scale dual H infinity filters has better robustness and higher
estimation accuracy than the single/multi-scale dual Kalman
filters. To further validate the feasibility of the proposed method
for EV applications, a lithium-ion battery-in-the-loop approach is
applied to verify the stability and accuracy of the SoC estimation
and it is found that the SoC estimated from the proposed method
can converge to the reference value gradually and be stabilized
within 2%.
Index Terms—battery-in-the-loop; capacity; dual H infinity
filters; Lithium-ion battery; multi-scale; state-of-charge.
I. INTRODUCTION
W ITH the superiority of high specific energy and specific
power, lithium-ion batteries are widely used in the
development of electric vehicles (EVs) [1], [2]. For safe and
efficient operation of EVs, it is essential for battery
management systems to accurately estimate the SoC and
capacity so that batteries can be balanced and managed
Manuscript received September 14, 2016; revised December 02, 2016;
accepted February 9, 2017. This work was jointly supported by the National
Natural Science Foundation of China under Grant No. 51507012, Australian
Research Council Linkage Grant under Grant No. LP110200302, Sino-polish
Collaborative research in e-mobility public transportation under Grant No.
2015DFG81930 and the Joint Funds of the National Natural Science
Foundation of China under Grant No. U1564206. (Corresponding authors: Rui
Xiong, e-mail: rxiong@bit.edu.cn).
Cheng Chen and Rui Xiong are all with National Engineering Laboratory
for Electric Vehicles, School of Mechanical Engineering, Beijing Institute of
Technology, Beijing, 100081, China, and also all with Collaborative
Innovation Center of Electric Vehicles in Beijing, Beijing Institute of
Technology, Beijing, 100081, China. (e-mail: chencheng@bit.edu.cn (C.
Chen)).
Weixiang Shen is with Faculty of Science, Engineering and Technology,
Swinburne University of Technology, Hawthorn, Victoria 3122, Australia
(e-mail: wshen@swin.edu.au).
0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
SoC and capacity is challenging in uncertain and complex EV
environments.
There are various SoC estimation methods. They can
generally be divided into four categories: look-up table
methods [5], ampere-hour methods [6], model-based methods
[7]–[9] and data-driven methods including neural networks
[10], fuzzy controllers [11] and support vector regressions [12].
Among them, the model-based methods provide the robust
performance for the SoC estimation due to their sustained error
correction mechanism through the close-loop feedback. In [7],
the Thevenin model is combined with the Lyapunov based
adaptive observer to obtain the accurate SoC. In [8], the dual
polarization (DP) model is used to estimate the SoC based on
the particle swarm optimization (PSO). In [9], the
resistance-capacitance (RC) model with the nonlinear
characteristics of SoC versus open circuit voltage for Li-PB is
used to estimate the SoC with the sliding-mode observer.
Obviously, the state estimation algorithm and the battery model
are the keys of the SoC estimation. For the former, the Kalman
filter (KF) and its improved forms are the typically
representative algorithms. The extended KF (EKF) shows a
good estimation of the SoC with a fast convergence at
erroneous initial values of SoCs in [10]. The unscented KF
(UKF) is proved to be higher estimation precision and less
computational burden than the EKF due to its no requirement of
linearizing nonlinear model in [13]. The adaptive UKF (AUKF)
whose process and measurement noise covariance can be
adjusted adaptively in the estimation process is proposed to
obtain more accurate results than the UKF in [14]. As a general
rule, the KFs can minimize the covariance of the estimation
error based on the assumptions of accurate battery models and
known statistical characteristics of noises, so they do not make
any guarantee of estimation error when the assumptions are not
met in the actual process. The H infinity filter (HIF) based on
the minimax error criterion, which is proposed and improved in
[15], [16], can still guarantee estimation accuracy in the worst
case and it is increasingly popular in the field of state estimation
[17]–[20]. The experimental results of the SoC estimation in
[21] show that the HIF can provide the accurate SoC estimation
at inexact or unknown statistical properties of the errors.
Furthermore, the adaptive algorithms for estimating noise
covariance matrices are introduced in [22]. In [23], the mixed
Kalman/H-infinity filtering approach is proved to have better
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

performance than either of the algorithms individually. For the system.

latter, battery model parameters significantly affect estimation
A. H infinity filters and single-scale dual H infinity filters
performances, offline identification of battery parameters is not
appropriate when the working environment of batteries is Based on the principle of minimizing the maximum
changing. For the practical applications in EVs, online estimation error, the HIF ensures the good estimation
identification of battery parameters is proposed to update performance even in the worst case, and it defines cost function
battery parameters regularly in [24], [25], such as a moving as follows:
N 1 2
window least square method in [25], which leads to the
exploration of the joint estimation of battery states and

k 0
xk  xˆk
J1 
Sk
(1)
 
N 1
parameters such as a recursive least square (RLS) and an AEKF
x0  xˆ0   wk  vk
2 2 2

in [26]. Since the RLS and the AEKF are independent modules P01 Qk1 Rk1
k 0
in this method, it is hard to ensure its convergence. In [27], the where xk is the system state and xˆk is its guess value; x0 is the
combination of state estimation and parameter identification is
initial system state and x̂0 is its guess value; wk is the process
used to estimate the SoC and state of energy. The experimental
results show good accuracy and robustness. noise vector for system states; vk is the measurement noise
In addition, the model-based methods usually obtain the SoC vector; P0 , Qk , Rk and Sk are the symmetric positive definite
state equation through an ampere-hour counting method, where matrixes that should be chosen by designers based on the
battery capacity is assumed to be a known parameter but specific applications. According to the mathematical deduction
actually changing slowly with aging. To overcome this problem, in [15], [16], the following recursive filtering algorithm is
many methods have been investigated to estimate battery obtained.
capacity in real time [28]–[31]. The open circuit voltage (OCV)  K  P  I   S P  C T R 1C P 1 C T R 1
is utilized to obtain the accurate SoC which is then used to  k k k k k k k k k k

calculate battery capacity online [28], [29]. The accuracy of the  xk 1  Ak xk  Ak K k  yk  Ck xk 
ˆ ˆ ˆ (2)

 Pk 1  Ak Pk  I   Sk Pk  Ck Rk Ck Pk  Ak +Qk
calculated capacity depends on the current which is difficult to T 1 1 T

be measured accurately in EV environments. In [30], the

degradation models are established to estimate battery capacity
based on a series of experimental data, which is very
time-consuming. In [31], battery capacity is considered as a
system parameter and then a dual EKF is proposed to estimate
the SoC and capacity online. However, fast-varying states (e.g.
SoC) and slow-varying parameters (e.g. capacity) require dual
estimation methods in different time scales [32], [33], which
will not only increase the stability of the estimation methods
but also reduce their computation complexity.
Based on this rationale, this paper uses the HIF to construct a
multi-scale dual HIF for the SoC and capacity estimation. To
further verify the reliability and accuracy of the proposed
multi-scale dual HIF for the SoC and capacity estimation, the
battery-in-loop test based on xPC target [34] is setup to closely
simulate EV operation environments for a real time validation
[35], [36].
The rest of this paper is organized as follows. The
description of a discrete-time nonlinear system and the
multi-scale dual HIF are presented in Section II. Section III
describes the implementation of the multi-scale dual HIF in
battery systems for the SoC and capacity estimation. The
experimental and simulation results as well as the evaluation of
the multi-scale dual HIF algorithm are reported in Section IV.
Section V further verifies the multi-scale dual HIF algorithm by
the battery-in-loop test, followed by the conclusions drawn in
Section VI.
II. MULTI-SCALE DUAL H INFINITY FILTERS
where  is the performance boundary designed by users; Ak
and Ck are the coefficient matrixes; Pk is obtained by
recursive operation from P0 ; K k is the gain matrix at time tk .
The HIF estimates the states while providing the uncertainty
vector of the states which indicates the error bounds of the
estimated states. Three confidence levels of 65.3%, 95.4% and
99.7% are very commonly adopted in practical applications.
The confidence level of 65.3% will cause the loss of
high-confidence results while the confidence level of 99.7%
will bring about the overmuch low-confidence results. Thus,
this paper selects the confidence level of 95.4%, which means
we are 95.4% confident that the true state xk will be bounded
by xˆk  2 diag  k  , where  k is the error covariance of state
which is expressed by
k  ( I  Kk Ck )k 1 ( I  Kk Ck )T  Kk Rk KkT
(3)
To achieve dual estimation of states and parameters, the
procedure includes states update prepared for parameter
identification and parameters update that is premise of state
estimation, which is summarized in Table I.
TABLE I
ALGORITHM OF SINGLE-SCALE DUAL H INFINITY FILTERS
Nonlinear system
 xk  f  xk 1 , k 1 , uk 1   wk 1

 k   k 1   k 1 (4)
 y  g x , , u  v
 k  k k k k
Definition

In this section, the H infinity filter (HIF) and the single-scale

dual HIF are first introduced as a preparation. Then, the
multi-scale dual HIF is proposed to estimate system states and
parameters based on the multi-scale discrete-time nonlinear

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

Ak 1 

df x,ˆk , uk 1  , Ckx 

dg x,ˆk , uk  characteristic. The following equation describes a multi-scale
discrete-time non-linear system.
dx dx
xk  xˆk1 xk  xˆk  xk  f  xk 1 ,l 1 , uk 1   wk 1
(5) 
dg  xˆ , , uk 

l  l  L   l  L (14)
Ck 
k
 y  g x , , u  v
d  k  k l k k
k ˆk

Initialization
xˆ0  E  x0  , P0x  E  x0  xˆ0  x0  xˆ0  
 
ˆ0  E 0  , P0  E  0  ˆ0 0  ˆ0 

where,  k is the system parameter and ˆk is its guess value;  0 is
the initial system parameter and ˆ0 is its guess value;
For k 1,2, ,  , calculation:
Step 1: Time-update equations for parameters
ˆk  ˆk1 , Pk ,  Pk,1  Qk1
Step 2: Time-update equations for states
where k indicates the time scale for system state estimation
accomplished at the sampling time tk , and here k  0, 1, 2, ;
T
(6)
xk is the system state vector at the time tk ; l indicates the time
 
T
(7) scale for system parameter estimation accomplished at the

sampling time tl , and here l  L, 2L, 3L, ( L is the limit
value for scale conversion);  l is the system parameter vector
updated at the time tl , which will be used to estimate system
states between the time tl +1 and tl +L and l  0 means the initial
value of the parameter, as shown in Fig. 1; uk is the system
(8)
input vector at time tk ;  l is the process noise vector for system

 
xˆk  f xˆk1 ,ˆk , uk 1 , Pkx ,  Ak 1Pkx,1 AkT1  Qkx1 (9) parameters; yk is the system measurement vector at time tk .
Step 3: Measurement-update equations for states The sampling time, k 1, 2, ,L L+1, L+2, ,L+L 2L+1, 2L+2, , 2L+L

e  yk  g xˆ ,ˆ , uk

x
k  
k

k  The time scale for system
 x
 
1 1, 2, ,L L+1, L+2, ,L+L 2L+1, 2L+2, , 2L+L
 K k  Pk I   x Sk Pk   Ck   Rk  Ck Pk
x , x x , x T x 1 x x , state estimation, k


   Ckx   Rkx 
T 1 The time scale for system
(10) parameter estimation, l
0 L 2L

 xˆk+  xˆk  K kx ekx
 Fig. 1. Time scales for state and parameter estimation.

 
1

 k k x k k  k   k 
 P x ,   P x ,  I   S x P x ,   C x T R x 1 C x P x , 
k k
Therefore, the single-scale dual HIF in Table I is improved
as the multi-scale dual HIF to estimate system states and
Step 4: Measurement-update equations for parameters parameters with different time scales as summarized in Table

ek  yk  g xˆk ,ˆk , uk

 II.

 
TABLE II
  1

 K k  Pk I   Sk Pk   Ck   Rk  Ck Pk
 ,   ,  T  1   ,  ALGORITHM OF MULTI-SCALE DUAL H INFINITY FILTERS
 Nonlinear system
   Ck   Rk 
1
 xk  f  xk 1 ,l 1 , uk 1   wk 1
T
(11)
 
ˆk  ˆk  K k ek l  l  L   l  L
  y  g x , , u  v
 k l k k
   k
1

k k k  k   k 
 P ,   P ,  I   S  P ,   C  T R 1 C  P , 
 k  k k
Definition

The system parameters C in Eq. (5) are the function of the

k
Ak 1 

df x,ˆk , uk 1  , Ckx 

dg x,ˆk , uk 
system states and can be calculated by dx dx
xk  xˆk1 xk  xˆk
dg  xk , , uk  g ( xk , uk , ) g ( xk , uk , ) dxk (15)
Ck    (12) dg  xˆ , , uk 

d  xk d Cl 
k

d
where l ˆl

dxk f ( xk 1 , uk 1 , ) f ( xk 1 , uk 1 , ) dxk 1 Initialization: install L and l ( here l =0 ),

=  (13)
d  xk 1 d xˆ0+  E  x0  , P0x ,  E  x0  xˆ0  x0  xˆ0  
T
(16)
 
The derivative calculations are recursive in nature and

ˆ0+  E 0  , P0 ,  E  0  ˆ0 0  ˆ0    
T
evolve over time. The term dx0 d is initialized to zero unless (17)

the priori information gives a better estimate of its value. This
For k 1,2, ,  , calculation:
algorithm is initialized by providing the system states and
Step 1: Time-update equations for state
parameters, the matrices and performance boundary to the best
guesses based on the prior information.  
xˆk  f xˆk1 ,ˆl+ , uk 1 , Pkx ,  Ak 1Pkx,1 AkT1  Qkx1 (18)
Step 2: Measurement-update equations for state
B. Multi-scale dual H infinity filters
The dual HIF provide the dual estimation of states and
parameters. In general, the states show the fast-varying
characteristic and the parameters show the slow-varying

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4



ekx  yk  g xˆk ,ˆl+ , uk  can be converted into a difference equation by standard
techniques, where the discrete-time battery model is modified
 x
 
1

 K k  Pk I   x Sk Pk   Ck   Rk  Ck Pk
x , x x , x T x 1 x x , to a multi-scale model as
   t   t 
   Ckx   Rkx 
1
U p,k 1 = exp  U p,k  [1  exp  ]iL,k R p ,l
T
(19)
   C p ,l R p ,l   C p ,l R p ,l  (23)
 xˆk+  xˆk  K kx ekx 
 U
 t ,k  U  U  i R

 
oc , k p,k L , k 0 ,l
1

 k k x k k  k   k 
 P x ,   P x ,  I   S x P x ,   C x T R x 1 C x P x , 
k k where t is the sampling period, and its value approximates to
be 1s in this paper. According to the ampere-hour counting
Step 3: Judge whether k divided by L with no remainder. If so,
method, the SoC state equation can be described as a
make l =l  L and continue to the next step 4; otherwise return to step
1 and ready for the calculation of next sampling time. discrete-time form
Step 4: Time-update equations for parameters   iL,k  iL,k t
ˆl  ˆl L , Pl ,  Pl,L  Ql L (20) zk 1 =zk  (24)
Ca ,l
Step 5: Measurement-update equations for parameters

 
where zk is the SoC at the time tk , iL,k is the load current at
el  yk  g xˆk ,ˆl , uk
 time tk ,   iL,k  is the coulomb efficiency of the battery, and
 
 
1
I   Sl Pl ,    Cl   Rl  Cl Pl , 
 , T 1
 K l  Pl Ca ,l is the battery capacity, which is a function of temperature

   Cl   Rl 
T 1
(21) and aging level [37], [38]. Together with the model parameters
 in Eq. (23), Ca ,l is considered as a system parameter and will be
ˆl  ˆl  Kl el
 updated at the time tl  L , where the subscript l  L means the
 
1

l  l l  l  l l l
 P ,   P ,  I   S  P ,   C  T R 1 C  P , 
 l sampling period for the battery capacity.
The relatively stable SoC-OCV curve is usually used to
III. APPLICATION TO BATTERY SYSTEMS calibrate battery SoC. In fact, this curve is affected by
temperature and aging to a certain extent which mainly reflects
A. Battery model in the change of battery capacity [37], [38]. To tackle the
It is important to select a suitable model for the application of capacity change, the capacity-SoC-OCV curve surface can be
the above methods. There are many battery models, such as used in the measurement-update process of SoC and capacity,
Rint model, RC model, Thevenin model and PNGV (the and it is obtained by the following equations.
Partnership for a New Generation of Vehicles) model. Among Firstly, the open circuit voltage can be described by an
them, Thevenin model is widely used because of its balance electrochemical equation:
between simplicity and accuracy [7], [26], [31]. So, Thevenin Uoc ,k  Ca ,l , zk   0  1 zk  2 zk2  3 zk3 + 4 / zk
model is selected in this study and its equivalent circuit is (25)
 5 ln  zk   6 ln 1  zk 
shown in Fig. 2.
where the coefficients of Eq. (25) are defined as the quadratic
function of capacity as:
+
0 1 2 3 4 5 6 
T
   Ca2,l Ca ,l 1
T
(26)
+
_
where 0 ,1 ,2 ,3 ,4 ,5 ,6 are the coefficients in the
_
electrochemical equation,  is a constant matrix, and it is
obtained by the experimental data.
Fig. 2. Thevenin model of a battery. Up to now, we have the system state vector xk and
According to circuit theory, this battery model can be
parameter vector  l :
described by
T
U t  U oc  U p  iL R0 xk  U p,k zk  (27)

 1 1 (22) T
U p   C R U p  C iL l   R0 ,l Rp ,l C p ,l Ca ,l  (28)
 p p p
Combining Eqs. (23) and (24) yields the state and
where U t is the terminal voltage; U oc is the open circuit
measurement equations as follows:
voltage; U p is the polarization voltage and U p is the time
derivative of U p ; iL is the current; R0 is the ohmic internal
resistance; C p and R p are the polarization capacitance and
resistance, respectively.
Considering slow-varying behaviors of a battery, we assume
that the battery is a time-invariant system and the load current is
constant at each sampling interval. As a result, Equation (22)

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

   -t   explained as follows.

 exp   0 Measurement The multi-scale H infinity filters

 xk 1  f  xk , l , uk     C p ,l R p ,l   xk start Initial value at tk

 
ˆl+ Cˆ a,l
   xˆk+1 , Pkx,1 ,ˆl , Pl ,

  0 1 Loading current
HIFx uk 1 U oc  xˆ ,ˆ

k 1 l

 ⑤ Measurement-update
(parameter)

60

xˆk ,ˆl Eqs. (20)  el , Kl ,ˆl+ , Pl ,

40

 
20

① Time-update (state)
 -t 

Current(A)

0

(29) -20

-40

Eqs. (17)  xˆk , Pkx ,

[1  exp  ]iL ,k R p ,l 
-60

u k , yk

-80
0 50 100 150 200 250 300
Time(min)

 + C R
 p ,l p ,l   cells
u k , yk
④ Time-update(parameter)

  L,k  L,k a ,l 
② Measurement-update

  i i  t C (state)
U oc  xˆk ,ˆl 
xˆk1 ,ˆl Eqs. (19)  ˆl , Pl ,

 Eqs. (18)  ekx , K kx , xˆk+ , Pkx ,

HIF
 y  g  x , , u   U  U  R i xˆk+  zˆk+  ③
Whether k divisible by L
 k
k=k+1 l=l+1
k l k oc , k p,k 0 ,l L , k Y
iL,k , U t ,k
N

where yk  Ut ,k which is the battery terminal voltage and

Fig. 3. Implementation flowchart of multi-scale dual H infinity filters.
uk  iL,k which is the battery current. While the batteries are discharging, the current and terminal
voltage of the batteries are measured in real-time by their
B. Implementation of multi-scale dual H infinity filters in corresponding sensors through the low-pass filter modules, and
battery systems
these measured data transmits to the multi-scale dual HIF to
According to the multi-scale dual HIF shown in Table II, we execute the dual estimation which can be shown in the
estimate the battery state vector with the known battery following five steps:
parameters, current and terminal voltage and then use the new (i) Based on the initial value settings at the time tk , HIFx
state vector to update the parameter vector. Because the battery
executes the time-update and calculates the priori parameter
states and parameters have different changing rates, we can
estimation xˆk and the matrix Pkx , by Eq. (18). This step
update those values in different time scales.
From Eqs. (15), (25) and (29), we obtain all the matrices and prepares for the next state estimation.
observation equations as: (ii) After the time-update of HIFx , the estimated terminal
  t   voltage at the time tk is obtained. The error between the
exp   0
Ak 1   C p ,l R p ,l   (30) estimated and measured terminal voltages is used to execute the
  measurement-update of HIFx by the correction of gain matrix,
 0 1
T where the posteriori state estimation xˆk+  zˆk+  and the matrix
 U oc  Ca ,l , zk  
C   1
x
k  (31) Pkx ,+ are computed by Eq. (19). In this step, the optimal state
 z z  zˆk 

T estimation at the time tk is used to calculate the priori state at
 U oc  Ca , z   x
Cl   iL,k

0 0   Ckx k (32) the next time.
 Ca    ˆl , x  xˆk (iii) Judge whether k divided by L with no remainder. If so,
 Ca Cˆ a ,l ,z  zˆk 
 

where make l =l  L and continue to the next step (iv); otherwise

Uoc return to step (i) and ready for the calculation of next sampling
 1  22 z  33 z 2  4 / z 2  5 / z  6 / 1  z  (33) time.
z
U oc (iv) On the basis of step (ii), HIF executes the time-update
 1 z z 2 z 3 1 / z ln  z  ln 1  z 
Ca  (34) and calculates the priori parameter estimation ˆl and the
   2Ca 1 0 matrix Pl , by Eq. (20). This step prepares for the next
dxk f ( xk 1 , uk 1 , ) dx parameter estimation.
=  Ak 1 k 1 (35)
d  d (v) After the time-update of parameter estimation, HIF is
f ( xk 1 , uk 1 , ) 0 b1 b2 0  ready to execute the measurement-update with the posteriori
  (36)
 0 0 0 b4  state estimation and priori parameter estimation at the time tl .
 U p,k -1t
b1 = 2 exp 
 t  iL ,k -1t
 
 t 
exp 
Then, we obtain the posteriori state estimation ˆl Cˆ a,l and the  
 R p ,l C p ,l   C R   ,
 C p ,l R p ,l  C p ,l R p ,l  p ,l p ,l  matrix Pl calculated by Eq. (21), which will return to step (i)

   t   and ready for the calculation of next sampling time.
  iL ,k exp  C R   1 There are totally 12 design parameters including x̂0 , P0x , Qkx ,
   p ,l p ,l  
 (37)
 U p,k -1t  t  iL ,k -1t  t  Rkx ,  x , Skx in the state estimation module and ˆ0 , P0 , Qk , Rk ,
b2 = 2 exp 
   2
exp 
    , Sk in the parameter estimation module. x̂0 is the initial
 C p ,l R p ,l  C p ,l R p ,l  C p ,l  C p ,l R p ,l 

b    iL ,k  iL ,k t
state vector including the initial SoC and the initial polarization
 4 voltage. In this study, the initial SoC is set to 0.8 for the fully
 Ca2,l
charged battery and the initial polarization voltage is set to 0.
Fig. 3 shows the implementation flowchart of the multi-scale P0x is a 2×2 symmetric positive definite matrix. The initial
dual HIF in Table II. The detailed implementation procedure is
values of P0x  i, i  are set correspondingly to the error of the

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

initial state x̂0  i  . If there is the bigger error of x̂0  i  , P0x  i, i  means the Li(NiCoMn)O2 battery. One NCM battery (NCM1)
should be set to the bigger value too. Hence, P0x  2,2  is set to is used in the verification process while the other NCM battery
(NCM2) is used in the battery-in-loop verification.
0.1 and P0x 1,1 is set very closely to zero. Similarly, Qkx is a TABLE III
THE SPECIFICATIONS OF THE CELLS AND ITS RELATED
2×2 symmetric positive definite matrix. The initial values of PREPARATION TESTS
Qkx  i, i  are set correspondingly to the level of system noise Battery types LMO NCM1
nominal capacity (Ah) 35 2
wk  i  . If there is the higher level of noise wk  i  , Qkx  i, i  rated voltage (V) 3.7 3.6
should be set to the higher value too. Rkx is a 1×1 symmetric cut-off voltage (V) 4.2/3.0 4.1/3.0
static capacity static capacity tests,
positive definite matrix. The initial value of Rkx is set test, hybrid open circuit voltage
pulse test, open tests, and dynamic
correspondingly to the level of the measurement noise vk .  x preparation tests circuit voltage stress tests at 10, 25,
is a performance boundary related to the model accuracy, test，and driving 40°C, dynamic stress
cycles test at test at room
smaller  x means a more accurate state-space model. If the 25°C temperature
state-space model is believed to be accurate,  x should be set The capacity test is to obtain the battery capacity prepared
very closely to zero. With the high accuracy mode based on for the hybrid pulse test; the hybrid pulse test is used to
online parameter identification,  x is selected to be less than 1. determine the battery parameters offline; both the driving
Skx is a 2×2 symmetric positive definite matrix. The initial cycles and dynamic stress test (DST) are used for verification
and comparison of the above-mentioned estimation algorithms.
values of Skx  i, i  are determined on the basis of the degree of
In the following Section B, we use the data of LMO battery
interest in the states. If one state xk  i  is more interested than to compare three algorithms that are the HIF, the EKF and the
the other states, Skx  i, i  corresponding to xk  i  should be set to dual HIF for the SoC estimation with the known battery
capacity, which aims to prove the superiority of the H infinity
the bigger value. The state of the SoC is the focus of this study,
filter and the dual H infinity filters. In Section C, we use the
thus Skx  2,2  is set to 10 and Skx 1,1 is set to 1. In the same data of NCM battery to demonstrate that the multi-scale dual
way, the parameters can be tuned in the parameter estimate HIF can provide more accurate estimation of SoC and capacity
module. than the single-scale dual HIF and EKF as well as the
multi-scale dual EKF.
IV. VERIFICATION AND DISCUSSIONS 1) Data of LMO battery
We can obtain the cell capacity by means of the capacity test,
A. Experimental setup
the OCV function Uoc  z  in Eq. (25) with the coefficients
The battery test bench is set up to acquire the experimental
data and is shown in Fig. 4, which consists of a battery testing shown as Table IV and the other parameters with the data of the
system (Arbin BT2000), three thermal chambers, a computer hybrid pulse test and the open circuit voltage test. They are all
and the batteries or battery pack. The battery testing system used to estimate the cell state with the offline EKF and HIF.
TABLE IV
regulates battery charging and discharging with the established COFFICIENTS OF OCV FUNCTIONS–UOC(z)
strategy and collects currents and voltages, which are then sent 0 1 2 3 4 5 6
to the computer. The computer is used for the human–machine 4.827 -2.197 2.529 -1.128 0.02153 0.5104 -0.01409
interface and experimental data storage. The thermal chambers
The battery in EVs will be fully charged but not be fully
are used to control the environment during the test. discharged because of the endurance mileage and the battery
Arbin BT2000
life, the SoC in the range of 100% to 20% is selected in this
study. The driving cycles is used to simulate the typical urban
Thermal Chamber driving cycle of EVs. Based on the Arbin data-logger, we apply
Thermal Chamber
the current profile based on the driving cycles to the battery as
shown in Fig. 5(a) and obtain the corresponding trajectory of
SoC as shown in Fig. 5(b) which is used as a reference (or true)
SoC, where(a)itloadiscurrent
assumed(Zoom)
that the coulomb efficiency is 100%.
(b) SoC as a function of time
50 1
(a) (b)
Current (A)

0
SoC

0.5
Computer
-50
0
60 70 80 90 100 50 100 150 200 250 300
Fig. 4. The physical map of battery test bench. Time (min) Time (min)
The accuracies of the voltage and current acquisition of Fig. 5. Test results of driving cycles (a) current profile (zoom) and (b)
Arbin BT2000 are about 0.05%. The batteries under test in this corresponding response SoC.
study include two kinds of lithium-ion batteries and their 2) Data of NCM1 battery
specifications and related preparation tests are shown in Table Similarly, we can obtain the cell capacities at 10°C, 25°C,
III, where the LMO means the LiMn2O4 battery and the NCM 40°C and the room temperature by means of the capacity tests

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7

as shown in Table V. With the data of the open circuit voltage true values in 70s while the EKF takes 140s.
tests at 10°C, 25°C, 40°C and the room temperature, we obtain 2) Evaluation of H infinity filter and dual H infinity filters
the OCV functions Uoc Ca , z  with different cell capacities. From the observation of the above results, it is found that the
Figure 6 shows the SoC trajectories corresponding to the DSTs fitting errors of battery terminal voltage are slightly high due to
at 10°C, 25°C, 40°C and room temperature. the inaccurate model with the parameters determined offline.
TABLE V This inaccurate model also affects the measurement-update
CELL CAPACITY AT DIFFERENT TEMPERATURES procedure for the SoC estimation. If we pay more attention to
Temperature 10°C 25°C 45°C room the SoC estimation accuracy, it may result in large errors of
Capacity/Ah 2.0568 2.0806 2.1165 2.0960 terminal voltage or other states which will in turn cause the
1
T10 large auto-covariance and error bound, and thus increase the
0.8 T25 uncertainty of the SoC estimation. The dual HIF (DHIF) can
SOC reference

T40
Troom
provide online estimation of the model parameters which can
0.6
reduce the fitting errors of terminal voltage and improve the
0.4 SoC estimation. The terminal voltage and SoC estimation
0.2 results of the HIF and DHIF with the erroneous initial SoC
50 100 150
(80%) are plotted in Fig. 8.
Time(min) 0.06
measured (b) HIF
4.2

Voltage error(V)
Fig. 6. SoC reference trajectory at 10°C (T10), 25°C (T25), 40°C (T40) and HIF 0.04 DHIF

Voltage(V)
4 DHIF
room temperature (Troom). 0.02
3.8
The experimental data at three different temperatures of 4.1 0
3.6 4
10°C, 25°C, 40°C are used to verify the estimation accuracy of 3.9
0 50 (a) -0.02
3.4
the multi-scale dual HIF algorithm. The experimental data at 0 100 200 300 0 100 200 300
Time(min) Time(min)
room temperature is used to validate its prediction accuracy and 1 3
reference HIF error
stability. HIF 2 -0.18 DHIF error

SoC error(%)
DHIF -0.2 HIF bound
SoC

1
0.5 1 -0.22
B. Evaluation of H infinity filter, EKF and dual H infinity filter 0.9
90 95 100 DHIF bound
0
Based on the experimental data of the LMO battery, the HIF, 0.8
0 1 2
(c)
0 -1 (d)
EKF and dual HIF are used to estimate SoC when the initial 0 100 200 300 0 100 200 300
Time(min) Time(min)
value of SoC is erroneous and capacity is known. Their
Fig. 8. Comparison results of HIF and DHIF: (a) estimated and measured
estimation results are compared in terms of the accuracies of voltage and (b) corresponding error, (c) estimated and reference SoC and (d)
the SoC and terminal voltage, the error bound of the SoC and corresponding error.
their convergence rates. It shows that the DHIF has the absolute value of voltage
1) Evaluation of H infinity filter and EKF estimation error less than 0.02 V in most of time except the
With the offline parameters, the HIF and the EKF are used to initial period and the absolute value of the SoC estimation error
estimate SoC when the initial value is erroneous. The terminal less than 0.5%. It is far superior to the HIF in terms of the
voltage and SoC estimation results of the HIF and the EKF with estimation accuracies of the SoC and terminal voltage, the error
the erroneous initial SoC (80%) are plotted in Fig. 7. bound of the SoC and the convergence rate.
measured
0.06
EKF
The comparison of the numerical results for the EKF, HIF
4.2
Voltage error(V)

EKF 0.04 HIF and DHIF is summarized in Table VI.

Voltage(V)

4 HIF
0.02 TABLE VI
3.8 COMPARISON OF EKF, HIF AND DHIF WITH ERRONEOUS INITIAL
4.1 0
3.6 4 SOC SETTING OF 80%
3.9 (a) -0.02 (b)
3.4 0 50
RMS error RMS Bounds convergence
0 100 200 300 0 100 200 300
Time(min) Time(min) of voltage error of error of time of SoC
1
reference
3
EKF error
(mV) SoC (%) SoC (%) (s)
0
EKF 2 HIF error EKF 12.99 1.04 0.20 140
SoC error(%)

HIF -0.5 EKF bound HIF 12.23 0.96 0.20 70

SoC

1 1
0.5 -1
HIF bound DHIF 5.59 0.35 0.02 30
0.9 90 100 110
0
0.8
0 1 2 3 (c) -1 (d)
C. Single-scale dual HIF and EKF and Multi-scale dual HIF
0
0 100 200 300 0 100 200 300 and EKF for SOC estimation and capacity
Time(min) Time(min)
Fig. 7. Comparison results of EKF and HIF: (a) estimated and measured voltage Based on the experiment data of the NCM1 battery, the
and (b) corresponding error, (c) estimated and reference SoC and (d) multi-scale dual HIF (L=60), the single-scale dual HIF (L=1),
corresponding error.
the single-scale dual EKF (L=1) and the multi-scale dual EKF
It shows that the absolute value of voltage estimation error is
(L=60) are utilized to estimate the SoC and capacity when the
less than 0.06 V and the absolute value of SoC estimation error
initial value of SoC and capacity are erroneous. Their results
is less than 1% when the SoC is estimated with the HIF. When
will be evaluated in terms of the estimation accuracies of the
the SoC is estimated with the EKF, it has very similar
SoC, capacity and terminal voltage as well as the convergence
estimation accuracy to the HIF. However, the HIF has higher
rate of the SoC and capacity.
convergence rate than the EKF, namely the HIF can reach the

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8

1) Comparison between dual H infinity filters and dual EKF 4.5

(a) measured MDEKF

Voltage error(V)
0.04
With the DST data at room temperature, the estimation MDEKF MDHIF

Voltage(V)
4 MDHIF 0.02 (b)
results of the single-scale dual HIF (SDHIF) and the
0
3.5
single-scale dual EKF (SDEKF) are plotted in Fig. 9 and the 4
-0.02
estimation results of the multi-scale dual HIF (MDHIF) and the 3
3.8
0 10
-0.04
0 50 100 150 200 0 50 100 150 200
multi-scale dual EKF (MDEKF) are plotted in Fig. 10, where Time(min) Time(min)
1 20
the erroneous initial SoC is set to 80% and the erroneous initial (c) reference 2 MDEKF
MDEKF 15 MDHIF
capacity is set to 1.5Ah while in fact the true SoC is 100% and

SoC error(%)
MDHIF 0
(d)

SoC
1 10
the true capacity is about 2Ah. 0.5
-2
0 100 200
0.9
4.5 5
(a) measured SDEKF 0.8
Voltage error(V) 0
0 5
SDEKF 0.04 SDHIF 0
Voltage(V)

4 SDHIF 0 50 100 150 200 0 50 100 150 200

0.02 (b)
Time(min) Time(min)
0 2.2 30
3.5 4 MDEKF
2

Capacity error(%)
-0.02 MDHIF

Capacity(Ah)
3.8 2
0 10 20 1
3 -0.04 (e) (f)
0 50 100 150 200 0 50 100 150 200 0
Time(min) Time(min) 1.8 2
10 -1
reference
1 20 0 100 200
(c) reference SDEKF MDEKF
2 1.6 1.5
SDEKF 15 SDHIF 0 1 2 MDHIF 0
SoC error(%)

SDHIF 0 (d) 0 50 100 150 200 0 50 100 150 200

SoC

0.5 1 10 Time(min) Time(min)

-2
0.9 5 0 100 200 Fig. 10. Results of MDHIF and MDEKF: (a) estimated and measured voltage
0.8
0 5 0
and (b) corresponding error, (c) estimated and reference SoC and (d)
0
0 50 100 150 200 0 50 100 150 200
corresponding error, (e) estimated and reference capacity and (f) corresponding
Time(min) Time(min) error.
2.2 30
SDEKF Similarly, Fig. 10(a) shows the measured and estimated
Capacity error(%)

2
SDHIF
Capacity(Ah)

2 20
0 terminal voltages from both MDHIF and MDEKF and Fig.
(e) (f)
1.8 2 10 -2 10(b) shows their voltage estimation errors, they both are
reference 0 100 200

1.6 1.5
0 1 2 3
0 50 100
Time(min)
Fig. 9. Results of SDHIF and SDEKF: (a) estimated and measured voltage and
(b) corresponding error, (c) estimated and reference SoC and (d) corresponding
error, (e) estimated and reference capacity and (f) corresponding error.
Fig. 9(a) shows the measured voltages and the estimated
terminal voltages from both SDHIF and SDEKF and Fig. 9(b)
shows that their estimation errors both are within 0.04V. But,
the SDHIF’s precision is higher than SDEKF between 0min
and 30min, which indicates the SDHIF is more stable than
SDEKF with the erroneous initial values. Figs. 9(c) and (d)
indicate the SoC estimation errors of SDHIF can not only
converge to the reference value in 5 min but also maintain
within 1% after convergence, and the SDEKF has obviously
lower convergence rate than SDHIF. Fig. 9(e) shows the
reference and estimated capacities from both SDHIF and
SDEKF and Fig. 9(f) indicates their capacity estimation errors.
They both can converge to the reference values within 2 min.
Fig. 9(f) indicates that the capacity estimation errors obtained
from the SDHIF can keep in 2% and remain stable around 0%
while the capacity estimation errors obtained from the SDEKF
go beyond 2% at about 100min and have an increasing trend to
negative half axis direction. The above analyses indicate that
the SDHIF has better robustness and higher estimation
accuracy than the SDEKF.
SDEKF
0
slightly less than 0.03V. The estimated terminal voltage from
SDHIF
150 200 0 50 100 150 200
the MDHIF is more stable than MDEKF with the erroneous
Time(min) initial value. Fig. 10(c) shows the reference and estimated SoCs
from both MDHIF and MDEKF and Fig. 10(d) indicates their
estimation errors. They both can not only converge to the
reference values in 5 min but also keep in 1% after convergence.
Fig. 10(c) also indicates the MDHIF has higher convergence
rate than the MDEKF. Fig. 10(e) indicates the reference and
estimated capacities from both MDHIF and MDEKF and Fig.
10(f) indicates their estimation errors, they both can converge
to the reference value in 1 min, after convergence the MDHIF
can keep the errors in 1% while the MDEKF can keep the errors
in 2%. The above analyses indicate the MDHIF has better
robustness and higher estimation accuracy than the MDEKF.
2) Evaluation of multi-scale and single-scale dual H infinity
filters
With the DST data at room temperature, the estimation
results of the SDHIF and MDHIF with the erroneous initial
SoC (80%) and the erroneous initial capacity (1.5Ah) are
plotted in Fig. 11. Figs. 11(a) and (b) show the measured and
estimated terminal voltages from the SDHIF and the MDHIF
and their voltage estimation errors, respectively. It can be seen
from Fig. 11(b) that the MDHIF has slightly smaller estimation
errors than the SDHIF. Furthermore, the MDHIF has faster
convergence rate than the SDHIF as shown in Fig. 11(c). Fig.
11(d) indicates the SoC estimation errors from the MDHIF are
less than those from the SDHIF all the time. Fig. 11(e) shows
the estimated capacity from the MDHIF converges to the
reference value faster than those from the SDHIF, where the
former only takes one update and the latter takes many updates
and more time. Fig. 11(f) indicates the capacity estimation
errors from the MDHIF are less than those from the SDHIF all
the time. From the above comparisons, it shows that the

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9

multi-scale algorithm has better robustness and higher A. Battery-in-loop test bench
estimation accuracy than single-scale one.
The battery parameters change slowly, but a single-scale
algorithm frequently updates the parameters which reduce the
stability of the parameters. This will weaken the state
correction based on the terminal voltage and affect the state
estimation accuracy, which then in turn affects the estimation
accuracy of the parameters. As a result, the single-scale
algorithm greatly reduces the estimation accuracy and stability
which is consistent with the experimental results as shown in
the above analysis. Thus, the multi-scale algorithm is superior
to the single-scale algorithm. The comparison of the numeric
results obtained from the single-scale and multi-scale
algorithms is showed in Table VII.
4.5
(a) measured SDHIF
Voltage error(V)

SDHIF 0.04 MDHIF

Voltage(V)

4 MDHIF Fig. 12. The physical map of battery-in-loop test bench.

0.02 (b)
3.5
0 The battery-in-loop test bench is shown in Fig. 12. It consists
4

3.8
-0.02 of host-PC, Target-PC, electrical load equipment, I/O interface
0 10
3
0 50 100 150 200
-0.04
0 50 100 150 200 and software part. The host-PC is used to design the models and
Time(min) Time(min)
1 20
generate the executable codes in Real Time Workshop in
reference SDHIF
(c)
SDHIF 15
2
MDHIF
MATLAB/Simulink. These codes will be downloaded to the
SoC error(%)

MDHIF 0 (d) Target-PC through TCP/IP; The Target-PC is used to runs the
SoC

0.5 1 10
0.9 5 -2
0 100 200
executable codes, which controls and operates the electrical
0.8
0 5 0 load equipment by CAN bus. The electrical load equipment
0
0 50 100 150 200 0 50 100 150 200 regulates battery charging and discharging with the established
Time(min) Time(min)
2.2 30 strategy and collects currents and voltages, which are then sent
SDHIF
Capacity error(%)

2
MDHIF to the Target-PC through CAN bus. In the Target-PC, these
Capacity(Ah)

2 20
2 (e) 0 (f) data of currents and voltages is further imported to the model
1.8
reference 10 -2
0 100 200 downloaded from host-PC for further calculation [27].
1.5 SDHIF
1.6 0 1 2
MDHIF 0 With the continuous operations in Target-PC, this bench can
0 50 100
Time(min)
150 200 0 50 100
Time(min)
150 200 achieve the online estimation of the SoC and capacity in the real
Fig. 11. Results of SDHIF and MDHIF: (a) estimated and measured voltage and working environment of the battery in EVs to the maximum
(b) corresponding error, (c) estimated and reference SoC and (d) corresponding extent. It should be noted that the accuracies of the voltage and
error, (e) estimated and reference capacity and (f) corresponding error.
TABLE VII
current acquisition of electrical load equipment are about 0.5%,
COMPARISON OF SINGLE-SCALE AND MULTI-SCALE ALGORITHMS which is aligned with the accuracy of the current and voltage
RMS RMS RMS acquisition in real EVs [27].
error of error of error of Simulation Considering that the minimum measurement voltage of
voltage SoC capacity time (s)
(mV) (%) (%)
electrical load equipment is 5V which is higher than the voltage
Single-scale of a single battery, we select two similar batteries in series for
11.08 1.59 1.24 2.26
dual EKF testing and the specifications of the selected two batteries is
Single-scale shown in Table VIII.
11.47 1.07 0.87 2.77
dual HIF
TABLE VIII
Multi-scale
10.37 0.90 0.79 1.58 SPECIFICATION OF SELECTGED TWO BATTERIES IN
dual EKF
BATTERY-IN-LOOP TEST
Multi-scale
9.38 0.64 0.47 1.99 Battery species NCM2
dual HIF
nominal capacity (Ah) 25
rated voltage (V) 3.6
V. BATTERY-IN-LOOP VERIFICATION cut-off voltage (V) 4.15/2.5
In order to further verify the estimation accuracy of the SoC B. Analysis and evaluation of multi-scale dual estimation of
and capacity, the battery-in-loop test bench based on xPC target SoC and capacity
is designed to closely simulate EV operation conditions for a In the case of the battery fully charged, the Target-PC loads
real time validation of the multi-scale dual estimation the DST testing current profile to the batteries using the
algorithms. electrical load equipment. The multi-scale dual estimation
algorithm based on the MDHIF downloaded from host-PC with
the erroneous initial value of the SoC and capacity is used to
estimate the SoC and capacity in real time. The online
estimation results of terminal voltage, SoC and capacity are

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10

plotted in Fig. 13.
It shows the absolute value of voltage estimation error is
slightly less than 0.06V, and the SoC and capacity estimation
errors can converge to the reference value gradually and
stabilize in 2% and 4%, respectively. The results demonstrate
that the proposed MDHIF meets the actual needs and verifies
its suitability of EV applications.
8.2
(a)
ageing mechanisms and estimations for automotive applications,” Journal
of Power Sources, vol. 241, pp. 680−689, 2013.
[3] Laxman Maharjan, Shigenori Inoue, Hirofumi Akagi, Jun Asakura,
“State-of-Charge (SOC)-Balancing Control of a Battery Energy Storage
System Based on a Cascade PWM Converter,” IEEE Transactions on
Power Electronics, vol. 24, no. 6, pp. 1628− 1636, 2015.
[4] Hongjie Wang, M. Muneeb Ur Rehman, Michael Evzelman, and Regan
Zane, “SIMULINK based Hardware-in-the-Loop Rapid Prototyping of an
Electric Vehicle Battery Balancing Controller,” Control and Modeling for
Power Electronics (COMPEL), pp. 1−6, 2015.
measured (b)
[5] Markus Einhorn, Fiorentino Valerio Conte, Christian Kral, Juergen Fleig,

Voltage error(V)
estimated 0.05
8
“A method for online capacity estimation of lithium ion battery cells
Voltage(V)

7.8
0 using the state of charge and the transferred charge,” IEEE Transactions
7.6
7.4
7.8

7.6
on Industry Applications, vol. 48, no. 2, pp. 734−741, 2012.
-0.05 [6] Kong Soon Ng, Chin-Sien Moo, Yi-Ping Chen, Yao-Ching Hsieh,
7.4
7.2 60 70

0 20 40 60 80 100 0 20 40 60 80 100
“Enhanced coulomb counting method for estimating state-of-charge and
Time(min) Time(min) state-of-health of lithium-ion batteries,” Applied Energy, vol. 86, pp.
1
measured
20 1506−1511, 2009.
(c) (d) 2
0.9 estimated 15 [7] Hicham Chaoui, Navid Golbon, Imad Hmouz, Ridha Souissi, Sofiène
SOC error(%)

1
Tahar, “Lyapunov-Based Adaptive State of Charge and State of Health
SOC

0.8 1 10
0
0 60
Estimation for Lithium-Ion Batteries,” IEEE Transactions on Industrial
0.7 0.9
5 Electronics, vol. 62, no. 3, pp. 1610−1618, 2015.
0.6 0.8
0 5 10
0 [8] Taesic Kim, Wei Qiao, Liyan Qu, “Real-time state of charge and
0 20 40 60 80 100 0 20 40 60 80 100 electrical impedance estimation for lithium-ion batteries based on a
30
Time(min) Time(min)
hybrid battery model,” Applied Power Electronics Conference and
15
(e) reference (f) 6 Exposition (APEC), pp. 563−568, 2013.
Capacity error(%)

4
estimated Il-Song Kim, “Nonlinear State of Charge Estimator for Hybrid Electric
Capacity(Ah)

28 10 2 [9]
0
26 5
-2 Vehicle Battery,” IEEE Transactions on Power Electronics, vol. 24, no. 4,
-4
0 50 100 pp. 2027−2034, 2008.
24 0 [10] Mohammad Charkhgard, Mohammad Farrokhi, “State-of-charge
22 -5 estimation for lithium-Ion batteries using neural networks and EKF,”
0 20 40 60 80 100 0 20 40 60 80 100
Time(min) Time(min)
IEEE Transactions on Industrial Electronics, vol. 57, no. 12, pp.
Fig. 13. Results of multi-scale dual estimation of SoC and capacity in 4178−4187, 2010.
battery-in-loop test, (a) comparison of estimated and measured voltage, (b) [11] Pritpal Singh, Ramana Vinjamuri, Xiquan Wang, David Reisner, “Design
model error, (c) comparison of estimated and reference SoC, (d) SoC and implementation of a fuzzy logic-based state-of-charge meter for
estimation error, (e) comparison of estimated and reference capacity, (f) Li-ion batteries used in portable defibrillators,” Journal of Power Sources,
capacity estimation error. vol.162, no. 2, pp. 829−836, 2006.
[12] Juan Carlos Alvarez Anton, Paulino Jose Garcia Nieto, Cecilio Blanco
Viejo, Jose Antonio Vilan Vilan, “Support vector machines used to
VI. CONCLUSIONS estimate the battery state of charge,” IEEE Transactions on Power
Electronics, vol.28, no. 12, pp. 5919−5926, 2013.
This paper proposes a multi-scale dual H infinity filters for [13] Htet Aung, Kay Soon Low, Shu Ting Goh, “State-of-Charge Estimation
the SoC and capacity estimation. Firstly, the dual HIF is of Lithium-Ion Battery Using Square Root Spherical Unscented Kalman
compared with the HIF and the EKF for the SoC estimation at Filter (Sqrt-UKFST) in Nanosatellite,” IEEE Transactions on Power
Electronics, vol. 30, no. 9, pp. 4774−4783, 2015.
the erroneous initial SoC and known battery capacity based on
[14] Jinhao Meng, Guangzhao Luo, Fei Gao, “Lithium Polymer Battery
the experimental data of the LMO battery. The results indicate State-of-Charge Estimation Based on Adaptive Unscented Kalman Filter
the dual HIF has the best performance and its SoC estimation and Support Vector Machine,” IEEE Transactions on Power Electronics,
can converge to the true value within 30 seconds while keeping vol. 31, no. 3, pp. 2226−2238, 2016.
[15] Ravi Banavar, “A game theoretic approach to linear dynamics estimation,”
the high accuracy within 0.5%. Secondly, the multi-scale dual Doctoral Dissertation, University of Texas at Austin, May 1992.
HIF is compared with the single-scale dual HIF and EKF as [16] X. Shen, “Discrete H∞ filter design with application to speech
well as the multi-scale dual EKF for the SoC and capacity enhancement,” IEEE International Conference on Acoustics, Speech and
Signal Processing, vol.2, pp. 1504−1507, 1995.
estimation at the erroneous initial SoC and capacity based on [17] H.K. Sahoo, P.K. Dash, N.P. Rath, R. Kant, “Estimation of power quality
the experimental data of the NCM battery. The results show using Complex H∞ Filter,” International Conference on Power Systems
that the multi-scale dual H infinity filter has the best robustness (ICPS), no. 3, pp. 1−6, 2009.
[18] K. Ramakrishnan and G. Ray, “Robust H∞ Filter Design for a Class of
and estimation accuracy. Its SoC and capacity estimation can Uncertain Systems with Time-varying State-Delays,” International
converge to the true values in 5 min. and 1 min., respectively, Conference on Electrical & Computer Engineering (ICECE), pp.
while maintaining the high accuracy within 1%. Finally, the 775−778, 2012.
[19] Ivan Vitanov, Nabil Aouf, “Fault detection and isolation in an inertial
stability and accuracy of multi-scale dual H infinity filters for navigation system using a bank of unscented H∞ filters,” UKACC
battery SoC estimation is verified by the battery-in-the-loop International Conference on Control, pp. 250−255, 2014.
approach. [20] Edo Abraham, Eric C. Kerrigan, “Lower-Order H∞ Filter Design for
Bilinear Systems With Bounded Inputs,” IEEE Transactions on Signal
Processing, vol. 63, no. 4, pp. 895− 906, 2015.
REFERENCES [21] Fei Zhang, Guangjun Liu, Lijin Fang, Hongguang Wang, “Estimation of
[1] Guizhou Ren, Guoqing Ma, Ning Cong, “Review of electrical energy Battery State of Charge With H∞ Observer: Applied to a Robot for
storage system for vehicular applications,” Renewable and Sustainable Inspecting Power Transmission Lines,” IEEE Transactions on Industrial
Energy Reviews, vol. 41, pp. 225− 236, 2015. Electronics, vol. 59, no. 2, pp. 1086−1095, 2012.
[2] Anthony Barré, Benjamin Deguilhem, Sébastien Grolleau, Mathias [22] Mohammad Charkhgard, Mohammad Haddad Zarif, “Design of adaptive
Gérard, Frédéric Suard, Delphine Riu, “A review on lithium-ion battery H∞ filter for implementing on state-of-charge estimation based on battery
state-of-charge-varying modelling,” IET Power Electronics, vol. 8, no. 10,
pp. 1825−1833, 2015.

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2017.2670081, IEEE
Transactions on Power Electronics
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 11

[23] Christoph Unterrieder, Robert Priewasser, Stefano Marsili, Mario
Huemer, “Battery State Estimation Using Mixed Kalman/Hinfinity,
Adaptive Luenberger and Sliding Mode Observer,” Vehicle Power and
Propulsion Conference (VPPC), pp. 71−76, 2013.
[24] Hicham Chaoui, Asmae Elmejdoubi, Hamid Gualous, “Online Parameter
Identification of Lithium-Ion Batteries with Surface Temperature
Variations,” IEEE Transactions on Vehicular Technology, vol. PP, no. 99,
pp. 1−10, 2016.
[25] Habiballah Rahimi-Eichi, Federico Baronti, Mo-Yuen Chow, “Online
Adaptive Parameter Identification and State-of-Charge Coestimation for
Lithium-Polymer Battery Cells,” IEEE Transactions on Industrial
Electronics, vol. 61, no. 4, pp. 2053−2061, 2014.
[26] Fengchun Sun, Rui Xiong, “A novel dual-scale cell state-of-charge
estimation approach for series-connected battery pack used in electric
vehicles,” Journal of Power Sources, vol. 74, pp. 582–594, 2015.
[27] YongZhi Zhang, Rui Xiong, HongWen He and Weixiang Shen, “A
lithium-ion battery pack state of charge and state of energy estimation
algorithms using a hardware-in-the-loop validation”, IEEE Transactions
on Power Electronics, vol. PP, no. 99, pp. 1−10, 2016.
[28] J.T.B.A. Kessels, B. Rosca, H.J. Bergveld, P.P.J. van den Bosch, “On-line
battery identification for electric driving range prediction,” Vehicle
Power and Propulsion Conference (VPPC), pp. 1–6, 2011.
[29] Xidong Tang, Xiaofeng Mao, Jian Lin, and Brian Koch, “Capacity
Estimation for Li-ion Batteries,” American Control Conference (ACC),
vol. 145, no. 2, pp. 947–952, 2011.
[30] Sébastien Grolleau, Arnaud Delaille, Hamid Gualous, Philippe Gyan,
Renaud Revel, Julien Bernard, Eduardo Redondo-Iglesias, Jérémy Peter,
“Calendar aging of commercial graphite/LiFePO4 cell-Predicting
capacity fade under time dependent storage conditions,” Journal of Power
Sources, vol. 255, pp. 450−458, 2014.
[31] Jonghoon Kim, Seongjun Lee, B. H. Cho, “Complementary Cooperation
Algorithm Based on DEKF Combined With Pattern Recognition for
SOC/Capacity Estimation and SOH Prediction,” IEEE Transactions on
Power Electronics, vol. 27, no. 1, pp. 436−451, 2012.
[32] Haifeng Dai, Xuezhe Wei, Zechang Sun, Jiayuan Wang, Weijun Gu,
“Online cell SoC estimation of Li-ion battery packs using a dual
time-scale Kalman filtering for EV applications,” Applied Energy, vol. 95,
pp. 227−237, 2012.
[33] Yiran Hu, Yue-Yun Wang, “Two time-scaled battery model identification
with application to battery state estimation,” IEEE Transactions on
Control Systems Technology, vol.23, no.3, pp. 1180−1188, 2015.
[34] K. H. Low, Heng Wang, Michael Yu Wang, “On the development of a
real time control system by using xPC target: solution to robotic system
control,” IEEE International Conference on Automation Science and
Engineering, pp. 345–350, 2005.
[35] Wai Chung Lee, David Drury, “Development of a Hardware-in-the-Loop
Simulation System for Testing Cell Balancing Circuits,” IEEE
Transactions on Power Electronics, vol.28, no.12, pp. 5949−5959, 2013.
[36] Yongsheng He, Wei Liu, Brain J. Koch, “Battery algorithm verification
and development using hardware-in-the-loop testing,” Journal of Power
Sources, vol. 195, no. 9, pp. 2969−2974, 2010.
[37] Yonghuang Ye, Yixiang Shi, Andrew A.O. Tay, “Electro-thermal cycle
life model for lithium iron phosphate battery,” Journal of Power Sources,
vol. 217, pp. 509−518, 2012.
[38] Rui Xiong, Fengchun Sun, Zheng Chen, Hongwen He, “A data-driven
multi-scale extended Kalman filtering based parameter and state
estimation approach of lithium-ion polymer battery in electric vehicles,”
Applied Energy, vol. 113, pp. 463−476, 2014.
Rui Xiong (S’12–M’14–SM’16) received
the M.Sc. degree in vehicle engineering
and the Ph.D. degree in mechanical
engineering from Beijing Institute of
Technology, Beijing, China, in 2010 and
2014, respectively. He conducted
scientific research as a joint Ph.D. student
in the DOE GATE Center for Electric
Drive Transportation at the University of
Michigan, Dearborn, MI, USA, between 2012 and 2014.
Since 2014, he has been an Associate Professor in the
Department of Vehicle Engineering, School of Mechanical
Engineering, Beijing Institute of Technology, Beijing, China.
During 2015, he was a Visiting Associate Professor in the
Faculty of Science, Engineering and Technology, Swinburne
University of Technology, Australia. He has conducted
extensive research and authored more than 70 peer-reviewed
articles. His research interests mainly include electrical/hybrid
vehicles, energy storage, and battery management system.
Dr. Xiong received the Excellent Doctoral Dissertation from
Beijing Institute of Technology in 2014 and the first prize of the
Chinese Automobile Industry Science and Technology
Progress Award in October 2015. He received the Best Paper
Awards from the journal Energies in 2015. He is an Associate
Editor of the Energy, Ecology and Environment. He serves on
the Editorial Board of the Energies.
W.X. Shen (S’00–M’02) received Ph.D.
degree from University of Hong Kong,
Hong Kong, China, in 2002.
From 2002 to 2003, he was a Lecturer in
Ngee Ann Polytechnic, Singapore. From
2003 to 2008, he was a Lecturer and then a
Senior Lecturer in School of Engineering,
Monash University Malaysia. He then
worked as a Research Fellow for one year in
School of Electrical and Electronics Engineering, Nanyang
Technological University, Singapore. Currently, he is an
Associate Professor in Electrical Engineering in Faculty of
Science, Engineering and Technology, Swinburne University
of Technology, Australia. His research interests focus on
electric vehicles, renewable energy and power systems.

Cheng Chen received the M.S. Degree in

vehicle engineering in 2014 from Beijing
Institute of Technology, Beijing, China. He
is now studying for the Ph.D. degree in
vehicle engineering at Beijing Institute of
Technology, Beijing, China.
His research mainly focuses on the state
estimation of battery and the application of
the core algorithms in battery management
system.

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.