IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO.

3, SEPTEMBER 2010

821

Development and Validation of a Battery Model

Useful for Discharging and Charging Power
Control and Lifetime Estimation
Vivek Agarwal, Student Member, IEEE, Kasemsak Uthaichana, Member, IEEE, Raymond A. DeCarlo, Fellow, IEEE,
and Lefteri H. Tsoukalas, Member, IEEE

AbstractAccurate information on battery state-of-charge, expected battery lifetime, and expected battery cycle life is essential for many practical applications. In this paper, we develop a
nonchemically based partially linearized (in battery power) input
output battery model, initially developed for lead-acid batteries in
a hybrid electric vehicle. We show that with properly tuned parameter values, the model can be extended to different battery types,
such as lithium-ion, nickel-metal hydride, and alkaline. The validation results of the model against measured data in terms of power
and efficiency at different temperatures are then presented. The
model is incorporated with the recovery effect for accurate lifetime estimation. The obtained lifetime estimation results using the
proposed model are similar to the ones predicted by the Rakhmatov and Virudhula battery model on a given set of typical loads at
room temperature. A possible incorporation of the cycling effect,
which determines the battery cycle life, in terms of the maximum
available energy approximated at charge/discharge nominal power
level is also suggested. The usage of the proposed model is computationally inexpensive, hence implementable in many applications,
such as low-power system design, real-time energy management in
distributed sensor network, etc.
Index TermsBattery cycle life validation, battery model, lifetime estimation, recovery effect, validation.

I. INTRODUCTION
ATTERIES are widely used as a finite source of energy for
a variety of applications ranging from low-power design of
portable devices to high-power hybrid electric vehicle (HEV).
Battery performance is affected by various factors, such as operating temperature, humidity, discharging/charging cycles, etc.
Battery models are essential for any battery-powered system
design that aims at extending the batterys expected life and
in battery power management. This creates a need for battery
models that capture essential application-dependent character-

Manuscript received August 15, 2008; revised June 26, 2009; accepted
January 6, 2010. Date of current version August 20, 2010. The work of
K. Uthaichana was supported by grant Thailand Research Fund, Commission
on Higher Education. Paper no. TEC-00320-2008.
V. Agarwal and L. H. Tsoukalas are with the School of Nuclear Engineering, Purdue University, West Lafayette, IN 47907 USA (e-mail: agarwal1@
purdue.edu; tsoukala@purdue.edu).
K. Uthaichana is with the Department of Electrical Engineering, Chiang Mai
University, Chiang Mai 50200, Thailand (e-mail: kasemsak@chiangmai.ac.th).
R. A. DeCarlo is with the School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907 USA (e-mail: decarlo@
ecn.purdue.edu).
Digital Object Identifier 10.1109/TEC.2010.2043106

istics of real batteries to estimate/predict battery behavior under

various operation conditions.
The conversion of chemical energy within the battery into
electric energy is a complex electrochemical process. The battery behavior, depending upon the application can be approximated by an empirical model or by highly accurate and complex
electrochemical model. Rao et al. [1] classified battery models
as 1) physical models; 2) empirical models; 3) abstract models;
and 4) mixed models.
Physical models are also known as electrochemical models [1]. They are highly accurate and involve detailed consideration of electrochemical processes, thermal dynamic process, and
the physical construction for both charging/discharging characteristics [2], [3]. Since, the number of parameters to be carefully
selected in electrochemical models are large, i.e., computationally expensive, the usage of physical models in practical applications is limited.
Empirical battery models are easy to configure and are represented by a simple mathematical expressions [4], [5] with less
number of parameters. The parameters are usually obtained as a
solution of a least squares problem using chargingdischarging
data, operating conditions, and physical properties. Empirical
models often fail to provide accurate representation/estimation
under varying load conditions. On the other hand, abstract models are simplified equivalent representation of electrochemical process within a battery, and their choice is usually application specific. The equivalent representation, for example,
can be in terms of electric circuits [6] or stochastic process
[7].
Mixed models are based on high-level abstraction, avoiding
excessive details of physical laws (e.g., electrochemical process)
governing battery characteristics, which leads to the derivation
of a simplified analytical expression. The number of parameters
to be selected is usually low in mixed models. Rakhmatov and
Virudhula [8] and Rong and Pedram [9] battery models are
examples of mixed models.
In [10] and [11], a nonlinear empirical lead-acid battery model
was proposed and validated for supervisory level power flow
control of an HEV. The battery model in [10] has captured the
characteristics of discharging/charging efficiencies to estimate
the battery state-of-charge (SOC). Initially, the proposed model
assumes relatively constant operating temperature, and does not
consider the capacity fading effect. Further, the model was linearized in the (control) input to make the model amenable to a
large body of control literature.

0885-8969/$26.00 2010 IEEE

822

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

In this research, we show that with appropriately tuned parameters, the battery model in [10] can be extended to accurately describe behaviors of additional battery types [alkaline,
lithium-ion (Li-ion), nickel-metal hydride (Ni-MH)]. This extension is validated against the actual discharging and charging
data. Based on the validation results, we set forth a functional
relationship that exists between the model coefficients and temperatures for lead-acid and Ni-MH battery types alone. The
actual data at different temperatures for Li-ion and alkaline
battery types were not available to us. Although, Uthaichana
et al. [10] did not consider the recovery effect, by incorporating
the recovery effect, the model can be used for battery lifetime
estimation under various discharging load conditions.
A possible incorporation of the cycling effect, which determines the battery cycle life when it undergoes repeated discharging and charging is also suggested. Specifically, the maximum
available energy in a battery is characterized as a function of
number of cycles and nominal charge/discharge power level at
a specific depth of discharge (DOD). The quality of the approximation is validated against the actual data for a Li-ion
battery. Hence, the versatility of the model to different battery types widens its utility for practical applications. For example, in distributed sensor network application, estimation of
battery longevity, estimates of scheduled maintenance, and battery replacement are some of the key issues in such application [12][15].
The paper is organized as follows. The discussion on Rakhmatov and Virudhula battery model [8] and the development of
nonlinear and partially linearized battery model with recovery
is presented in Section II. The relationship between both battery
models is also discussed. The validation of the proposed model
for different battery types against the actual data measured at
different temperatures for both discharging and charging is illustrated in Section III. Section IV presents the validation of the
proposed model with the incorporated recovery model for Li-ion
battery lifetime estimation under different constant and varying
power load conditions. The approximation and the validation of
the maximum rated energy for a Li-ion battery to incorporate the
effect of cycling on battery cycle life is presented in Section V.
Finally, conclusions are drawn in Section VI.
II. BATTERY MODELING
Given various forms of battery models, we briefly summarize
the Rakhmatov and Virudhula battery model [8] and develop a
partially linearized inputoutput battery model of Uthaichana
et al. [10]. The generalization of Uthaichana et al. model [10]
to different battery types and its validation against the actual
data measured at different temperatures for both charging and
discharging are some of the main contributions of the paper. In
addition, a differential equation describing the recovery effect
is presented. Under the constant maximum energy assumption,
Uthaichana et al. model [10] with the incorporated recovery
model can be applied to solve the battery lifetime estimation
problem for nonrecharging load profiles and is presented in
Section IV. A possible relaxation on the constant maximum
energy assumption to capture the cycling effect and estimate

the battery cycle life under repeated discharging and charging

cycles are mentioned at the end of the paper.
A. Rakhmatov and Virudhula Battery Model
Rakhmatov and Virudhula [8] developed an analytical expression to estimate battery lifetime for various time-varying
loads by taking into account the changes in the concentration of
the electroactive species inside the battery. The model is based
on a 1-D diffusion process of the concentration of the species,
where the concentration of the species at time t, and at distance
x from the electrode is denoted by C(x, t). The battery lifetime L is defined as the time at which the concentration at the
electrode surface C(0, t) drops from the initial concentration of
the species C , to a specific threshold Ccuto . The cutoff concentration Ccuto depending on the battery type and size, is the
concentration level below which no further power can be drawn.
Rakhmatov and Virudhula obtained an analytical expression for
the concentration behavior by defining two partial differential
equations based on Ficks law as follows:
J(x, t) = D

C(x, t)
x

C(x, t)
2 C(x, t)
=D
x
2 x

(1)
(2)

where 1) J(x, t) denotes the flux of the species at time t

[0, L] and at distance x [0, w] from the electrode, in which
w is the length of the battery and 2) D denotes the diffusion
coefficient. The boundary conditions at the electrode surface
x = 0 based on the Faradays law, and at the other electrode
x = w based on constant concentration, are expressed in terms
of the concentration gradient as follows:
C(x, t)


i(t)
=D
(3)


x=0
nF A
x
C(x, t)


0=D
(4)


x=w
x
where F is Faradays constant, n is the number of electrons involved in the electrochemical reaction at the electrode surface,
A is the surface area of the electrode, and i(t) is the discharging
current and represents the load on the battery. Throughout this
paper, the electrode refers to the one that receives the electroactive species, and the other electrode refers to the one that
generates the species.
It can be shown that an analytical solution for (1) and (2), at
the electrode surface (x = 0) with the boundary conditions (3)
and (4) is as follows:

t
1

i(t)d
C(0, t) = C
nF Aw 0

 t

2 D ( t ) m 2

2
w
+ lim 2
i(t)e
d . (5)
0

m =1

The details on the derivation of (5) are presented in [16]. An

observation can be made from (5) that the concentration of
the species at the electrode surface decreases with time due to

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

the usage (i(t) > 0). Once the concentration C(0, t) drops to
Ccuto , we obtain the measure of battery lifetime L.
In certain applications, it is important to obtain an estimate
of the amount of energy consumed, and consequently, the remaining battery stored energy; this then allows estimates of the
battery lifetime given nominal loading profiles. To make the
model more amendable to various applications, (5) has been
expressed in terms of battery capacity as follows:


i(t)d + 2

=
0



m =1

i(t)e

m 2 (L )

d.

(6)

The first term in (6) is the total charge drawn by the load
and battery losses. The second term of (6) is the unavailable charge left in the battery due the nonuniform distribution of the electroactive species during discharging. Here, the
two battery parameters and are introduced. The parameter
= nF AwC (1 Ccuto /C ) denotes the battery capacity,
i.e., the estimated
amount of charge used during the battery lifetime and = D/w denotes the discharge time constant. The
values of both parameters depend on the battery type. Note that
the nonuniform distribution of the electroactive species in the
second term of (6) results in lower concentration at the electrode
surface C(0, t) than at the other electrode C(w, t). If a rest period is introduced, the nonuniformity decreases over time and
more charge are available at the electrode surface, which is a part
of the recovery effect. The equation capturing this phenomenon
is described in [17].
In general, the load i(t) is unknown and nonconstant. Even
if known, the nonlinear and random behavior would make the
usage of (6) numerically difficult. Nevertheless, if i(t) has an
average value over small intervals of time, the time-varying
discharge rate can be approximated by piecewise constant loads,
i.e., i(t) is approximated in the time interval [0, T ] by N equal
size staircase basis functions as follows:
i(t)

N
1




Ik U (t tk ) U (t tk +1 )

(7)

k =0

where U (t) is the unit step function.

Given a set of M constant experimental current loads Ii , i =
1, 2, . . . , M and corresponding set of M lifetimes Li , the battery
parameters 
and are selected as a least squares solution
2
is an estimate

, )
minimizing M
i=1 |Ii Ii | , where Ii = Ii (
of the experimental current load based on the most recent values
of
and within the iteration process.
Given the estimated battery parameters and N -step staircase
approximation of the load [see (7)], the battery lifetime computation involves two steps. The first step is to find the subinterval
(Tk = tk +1 tk , k = 1, 2, . . . , N 1), such that the concentration at the electrode surface C(0, tk +1 ) is below Ccuto at
tk +1 , i.e., find the subinterval Tk , such that L [tk , tk +1 ].
The second step is to determine the smallest t within the subinterval Tk , such that C(0, tk +1 ) Ccuto . The smallest t can
be found using the modified secant method, as in [18].

823

B. Development of Nonlinear and Partially Linearized

Battery Model
As mentioned earlier, the motivation to solve the power management problem of an HEV led to the development of partially linearized (control-oriented) battery model developed as
follows:
Here, normalized battery energy, denoted by W bat (t), is defined as the ratio of the instantaneous stored charge to the maxm ax
, where
imum stored charge, i.e., W bat (t) = Wbat (t)/Wbat
m ax
is
Wbat (t) is the instantaneous stored battery energy and Wbat
the maximum rated storage energy of the battery. As developed
in [19], Wbat (t) approximates the battery SOC under the reasonable assumption of a relatively constant open-circuit battery
voltage during operation.
To achieve a dynamic model for measuring the SOC of the battery, we differentiate the normalized stored energy W bat (t) =
m ax
with respect to time to obtain a relationship to
Wbat (t)/Wbat
the discharging and charging power of the battery
bat (t)
W
1

W
bat =
m ax = W m ax bat (W bat , Pbat , v)Pbat (t)
Wbat
bat
(8)
where 1) v is the battery mode of operation, where v = 0 means
the battery is discharging and v = 1 means the battery is charging; 2) bat (W bat , Pbat , v) is a generic efficiency for discharging and charging; and 3) Pbat (t) is the discharging (+)/charging
() battery power flow (input) as expressed in (9). The negative sign in (8) indicates decreasing Wbat (t) during discharging,
while (8) becomes positive during charging, thereby indicating
increasing Wbat (t)

> 0, discharging
Pbat (t) =
(9)
< 0, charging.
In general, during discharging, when 0 Pbat (t) < , where
is a small threshold and usually negligible compared to the discharging loads, a battery undergoes charge recovery, which is described in Section II-D. During discharge, the generic efficiency
0
0
(W bat , Pbat , v) > 1, where bat
bat (W bat , Pbat , v) = 1/bat
is the actual discharging efficiency. A desired power output
causes the battery to discharge more rapidly than what is required to account for losses. On the other hand, during charging
1
1
(W bat , Pbat , v) < 1, where bat
is
bat (W bat , Pbat , v) = bat
the actual charging efficiency, indicates that more power is
needed to overcome charging losses.
Given this relationship, the approximations of the generic efficiency bat (W bat , Pbat , v), v {0, 1} as a function of W bat
(state) and Pbat (input) for both cases is done by interpolating
a nonlinear function against the downscaled battery efficiency
map (see Appendix B) for v {0, 1}. Intuitively observing that
generic efficiency curves of a battery appear to have a logarithmic characteristic, the following approximation was empirically
determined:
bat (W bat , Pbat , v) = ln(W bat + d1,v ) + d2,v Pbat (t) + d3,v .
(10)
Given appropriate coefficients, the objective of this paper is to
show the quality of this approximation for four distinct battery

824

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

types. The coefficients dk ,v , k = 1, 2, 3 are chosen to fit the

generic battery efficiency map. The coefficients during battery
discharging and charging are computed as a solution to a nonlinear least-squares regression similar to that described earlier for
obtaining parameters of the Rakhmatov and Virudhula model.
Specifically, in each mode, the coefficients dk ,v for k = 1, 2, 3
are the numerical solution to the following minimization:
min bat (W bat , Pbat , v)

d k R

[ln(W bat + d1,v ) + d2,v Pbat (t) + d3,v ]2 . (11)

This nonlinear minimization problem is solved via a subroutine
nlinfit in MATLABs statistical toolbox in this study.
As per the aforementioned development, the differential equa
tion for W
bat (t) [see (8)] is nonlinear in the control input
(Pbat (t)). To make the model amenable to a large body of control
literature, (8) is partially linearized about the mode-dependent
v
. The resulting linnominal battery operating power Pbat,nom
earized differential equation is as follows:

2
d2,v

v
W bat (t) = m ax Pbat,nom
Wbat


Pbat (t)
v
ln(W bat (t) + d1,v ) + 2d2,v Pbat,nom
+ d3,v
m ax .
Wbat
(12)
The details on the methodology to obtain the partially linearized model [see (12)] are presented in Appendix A.
C. Relationship Between Rakhmatov and Virudhula and Partially Linearized Battery Model
In many practical applications, the actual knowledge about
the battery SOC is essential. In Rakhmatov and Virudhula [8]
battery model the concentration of the electroactive species at
the electrode surface is given by (5). Since the specific gravity
of the electroactive species is greater than that of water, the
higher the concentration of the electrospecies is the higher the
specific gravity. When a battery is fully charged (SOC = 1),
then the concentration in the electrolyte is at its maximum C
and so is the specific gravity. On the other hand, if the concentration of the species is at Ccuto (SOC = C
= 0), then the
specific gravity is at minimum operating value. In fact, there
is a linear relationship between the battery SOC and the specific gravity as shown in [20] under an assumption of uniform
concentration of the electroactive species at equilibrium. Based
on aforementioned discussion, there exists a relationship between the battery SOC and concentration in the Rakhmatov and
Virudhula battery model at equilibrium. Uthaichana et al. battery model [10] approximates the battery SOC as normalized
battery energy (W bat ) under the assumption of relatively constant open-circuit voltage. This relationship between the concentration in the Rakhmatov and Virudhula battery model [8]
and W bat in Uthaichana et al. battery model [10] is expressed
as follows:
C(t) = W bat (t)(C Ccuto ) + Ccuto

(13)

where C(t) is the concentration of the species. Note that at

equilibrium, the concentration of the species is uniform, i.e.,
C(0, t) = C(w, t) = C(t).
D. Battery Recovery Model
The recovery effect is the increase in W bat (the battery SOC)
due to the adjustment of the concentration gradient toward zero
under no load condition 0 Pbat (t) < , where is a small
threshold and usually negligible compared to the discharging
power loads. The recovery of the W bat , depending upon the
battery type, reaches its steady-state value after a few time constants. However, under the same condition, the derivative of
W bat (t) given by the partially linearized model (12) approxim ax
v
(Pbat,nom
)2 for v = 0. This is not
mately reduces to d2,v /Wbat
equivalent to the charge recovery effect. Therefore, to achieve a
better representation of the battery behavior, we incorporate the
recovery effect by augmenting an additional equation [see (14)]
to the partially linearized battery model
r (t) = W r (t) + (1 + )W (t ),
W
bat i
bat
bat

0 Pbat (t) <

(14)
r
where 1) = 1/ , is the recovery time constant; 2) W bat (t)
denotes the battery SOC during recovery; 3) the superscript r
denotes recovery; and 4) denotes the percentage of recovery,
whose value depends upon battery type and is usually obtained
from experimental data [21]. In (14), W bat (ti ) denotes the battery SOC at time ti , where ti is the time instance when recovery
r
starts, hence the initial condition of W bat (t) is W bat (ti ). The
value of is computed using the logarithmic approach (see
Appendix C).
III. PARTIALLY LINEARIZED BATTERY MODEL VALIDATION
FOR CHARGING AND DISCHARGING
In this section, we validate the partially linearized battery
model developed in Section II-B for different battery types without incorporating the recovery model. Since, the selected magnitude of the input power (Pbat ) during validation is greater than
the threshold , the recovery effect is negligible. The procedure
to validate the partially linearized model for each battery type
can be briefly summarized as follows: first, we obtain the battery
efficiency data. Second, we estimate the efficiency coefficients
dk ,v for charging and discharging, as described in Section II-B.
Finally, we evaluate the quality of the partially linearized model
without recovery effect against the actual data by computing the
two-norm normalized error. Specifically, we compute the error

v
between W
bat given by (12) at a given Pbat,nom and the actual
data for charging and discharging under various loads, SOC,
and temperatures.
The validation results during charging and discharging at
0 C, 25 C, and 50 C are shown for 17.2 Ah, 12 V lead-acid
battery type. The validation results during discharging alone are
shown for 6.5 Ah, 7.2 V Ni-MH battery type at 5 C, 20 C,
30 C, and 40 C, as the internal resistance data during charging
was not available to us. Based on the validation, we set forth
a functional relationship between the model coefficients, dk ,v ,
k = 1, 2, 3 and temperature (T ) during charging and discharging

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

TABLE I
MODEL COEFFICIENTS AND THE APPROXIMATION ERROR FOR 17.2 Ah, 12 V
LEAD-ACID BATTERY FOR THE NONLINEAR EFFICIENCY APPROXIMATION FOR
BOTH DISCHARGING AND CHARGING AT DIFFERENT TEMPERATURE

for both lead-acid and Ni-MH battery types. On the other hand,
the validation results during both charging and discharging for
1.41 Wh Li-ion and during discharging alone for 2850 mAh
alkaline batteries1 at room temperature are shown.
A. Efficiency Data for Validation
To obtain the efficiency data for lead-acid, Ni-MH, Li-ion,
and alkaline batteries, we follow the procedure described in [22].
In [22], the battery is represented by a simple resistive Thevenin
equivalent circuit. The internal resistance of the battery is a
function of load rate, SOC, and temperature. The internal resistance value is obtained as an average value under various
loads for each SOC at different temperature values for lead-acid
and Ni-MH batteries and for each SOC at room temperature
for Li-ion and alkaline batteries. The battery efficiency is then
obtained as the ratio of the terminal voltage to the open-circuit
voltage for discharging and vice versa for charging, where the
voltage difference is due to the losses through the internal resistance. The data on the internal resistance for each SOC at
different temperature values for 17.2 Ah, 12 V lead-acid, and
6.5 Ah, 7.2 V Ni-MH batteries is obtained from [23] and [24],
respectively. While, the internal resistance data for each SOC at
room temperature for 1.41 Wh Li-ion, and 2850 mAh alkaline
batteries is obtained from [25].
B. Validation Results for Lead-Acid at Different Temperature
Initially, we present the discharging and charging validation
results of the partially linearized battery model without considering the recovery model for 17.2 Ah, 12 V lead-acid battery
type at 0 C, 25 C, and 50 C.
In the case of a 17.2 Ah, 12 V lead-acid battery, we assume a
nominal load power of 27.5 W during discharging and 22 W
during charging for all the temperature values. The coefficients
dk ,v , k = 1, 2, 3 used to validate the linearized approximation
during discharging and charging at different temperatures are
presented in Table I along with the two-norm normalized error.
The approximation of the linearized generic battery efficiency
is accurate around the nominal input power of 27.5 W during
discharging and around the nominal input power of 22 W
during charging. Fig. 1, whose y-axis is Pbat bat , illustrates
1 Alkaline batteries in the market are mostly nonrechargeable as the existing
rechargeable alkaline batteries are not commonly used.

825

the accuracy of the efficiencies from the linearized approximation against the actual data for both discharging and charging,
respectively. The approximation error increases for large excursion from the nominal input power level, as observed in Fig. 1, as
expected. The partially linearized battery model, initially developed for an HEV application, showed a similar accuracy pattern
over the range of battery power in the case of a 30 of 12 Ah
12 V lead-acid battery [10], [11], [26].
From Table I, upon interpolation of the data, we see that
the variation in the model coefficients is essentially linear with
temperature and is given by a functional relationship in Table II
for both discharging and charging.
C. Validation Results for Ni-MH at Different Temperature
The validation results of the partially linearized battery model
for 6.5 Ah, 7.2 V Ni-MH battery type at 5 C, 20 C, 30 C, and
40 C for discharging alone are presented. A nominal load power
of 15 W is assumed during discharging for all the temperature
values. The approximation of the linearized battery model is
accurate around the nominal input power level of 15 W during
discharging. Fig. 2 shows a similar accuracy pattern over the
range of battery power. The coefficients used to validate the
linear approximation during discharging for 6.5 Ah, 7.2 V NiMH battery at 5 C, 20 C, 30 C, and 40 C are listed in Table III
along with the two-norm approximation error. From Table III,
we observe that the variation in the model coefficients during
discharging is essentially linear with temperature and is given
by a functional relationship in Table IV.
D. Validation Results for Li-ion and Alkaline Batteries
at Room Temperature
Initially, we present the discharging and charging validation
results of the partially linearized battery model for 1.41 Wh Liion battery at room temperature alone, as the internal resistance
data at different temperature values was not available to us.
Later, we present the validation result during discharging for
2850 mAh alkaline battery at room temperature.
In the case of a 1.41 Wh Li-ion battery, the same validation procedure is applied. The two-norm normalized error for
discharging is 3.15% and for charging is 1.38%. The approximation of the linearized battery model is accurate around the
nominal input power level of 253 mW during discharging and
250 mW during charging, as shown in Fig. 3. The approximation error increases during both discharging and charging as the
load power level deviates from respective nominal power level,
but not significantly over the range, as shown in Fig. 3.
In the case of 2850 mAh alkaline battery, the same validation
procedure is applied during discharging alone, since alkaline
batteries are mostly nonrechargeable. The existing rechargeable
alkaline batteries are not commonly used. Fig. 4 shows the accuracies of the linearized battery model for 2850 mAh alkaline
battery, which is linearized about the nominal input power of
125 mW. The two-norm normalized error is 4.01%. From Fig. 4,
a similar accuracy pattern over the range of battery power is
observed.

826

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

Fig. 1. Linearized approximation of 17.2 Ah 12 V lead-acid battery. (Left column) During discharging around the nominal power of 27.5 W at (a) 0 C,
(b) 25 C, and (c) 50 C. (Right column) During charging around the nominal power of 22 W at (d) 0 C, (e) 25 C, and (f) 50 C.

The coefficients used to validate the linear approximation during discharging and charging for 1.41 Wh Li-ion and 2850 mAh
alkaline batteries during discharging alone are listed in Table V.
The linearized approximation of the battery captures the fact
that the battery efficiency decreases as SOC decreases and bat-

tery input power increases. This is consistent with the behavior

of the battery efficiency maps. In the validation of all the battery
types during discharging and charging, at different temperatures
(including the room temperature), we observed that the partially
linearized model is accurate about its nominal input power level

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

827

Fig. 2. Linearized approximation of 6.5 Ah, 7.2 V Ni-MH battery during discharging around the nominal power of 15 W at (a) 5 C, (b) 20 C, (c) 30 C, and
(d) 40 C.
TABLE II
FUNCTIONAL RELATIONSHIP OF THE MODEL COEFFICIENTS WITH
TEMPERATURE DURING DISCHARGING AND CHARGING FOR 17.2 Ah, 12 V
LEAD-ACID BATTERY

TABLE III
MODEL COEFFICIENTS AND THE APPROXIMATION ERROR FOR 6.5 Ah, 7.2 V
Ni-MH BATTERY FOR THE NONLINEAR EFFICIENCY APPROXIMATION DURING
DISCHARGING AT DIFFERENT TEMPERATURES

TABLE IV
FUNCTIONAL RELATIONSHIP OF THE MODEL COEFFICIENTS WITH
TEMPERATURE DURING DISCHARGING FOR 6.5 Ah, 7.2 V Ni-MH Battery

and the approximation error increases slightly as the load power

deviates from the respective nominal input power level. The accuracy of the partially linearized can be improved by selecting
different power level based on the operating load power.2
IV. LI-ION BATTERY LIFETIME ESTIMATION USING PARTIALLY
LINEARIZED MODEL WITH RECOVERY
In this section, we present the description of typical power
load usage during battery operation and numerical estimation
2 A maximum of 5.24% error was observed during the validation, which is
quite reasonable. However, one can create a set of range of load powers and
select an associated nominal power for each range to lower the deviation between
the load and nominal power, and hence to reduce the approximation error.

828

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

Fig. 3. Linearized approximation of 1.41 Wh Li-ion battery during (a) battery discharging around the nominal power of 253 mW and (b) battery charging around
the nominal power of 250 mW.
TABLE V
MODEL COEFFICIENT FOR THE NONLINEAR EFFICIENCY APPROXIMATION FOR
BOTH DISCHARGING AND CHARGING FOR 1.41 Wh AND FOR DISCHARGING
ALONE FOR 2850 mAh ALKALINE BATTERIES AT ROOM TEMPERATURE

TABLE VI
DESCRIPTION OF ACTUAL [16] AND DOWNSCALED POWER LOADS
OF SUBSET P T

Fig. 4. Linearized approximation during discharging for 2850 mAh alkaline

battery around the nominal power of 125 mW.

of 1.41 Wh Li-ion battery lifetime using the partially linearized

model during discharging (v = 0) [see (12)] with the incorporated recovery model [see (14)]. The actual lifetime measurements on a real 1.41 Wh Li-ion battery for various typical load
usages are not available. The DUALFOIL simulator [3], which
numerically solves set of partial differential equations governing
the chemical reaction in the rechargeable Li-ion batteries, can
be used to obtain lifetime estimates of 1.41 Wh Li-ion battery
for a given set of typical power loads. However, we do not use
DUALFOIL simulator because it is a cumbersome processes
to appropriately choose over 50 parameters in order to obtain
accurate simulation results. Therefore, to estimate the lifetime
of 1.41 Wh Li-ion battery, we downscale the power loads used
during the actual lifetime measurement of same type 2.2 Wh
Li-ion battery in [16] and [27]. Under the downscaling assumption, (see Appendix B for details), the lifetime of 2.2 Wh Li-ion
battery subject to typical power loads summarized in Tables VI
and VII and the lifetime of 1.41 Wh Li-ion battery subject to
downscaled typical power loads also summarized in Tables VI
and VII must be the same. We compare the aforementioned ac-

tual battery lifetime with the estimate obtained using the battery
model in Section IIB and D [see (12) and (14)] and with the
estimate obtained using the Rakhmatov and Virudhula battery
model (6).

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

829

TABLE VII
DESCRIPTION OF ACTUAL [16] AND DOWNSCALED VARYING POWER LOADS OF SUBSET P c

In general, the battery open-circuit voltage Vo c is a function

of battery SOC and operating temperature. In this paper, all
the computation assumes room temperature. Initially, W bat (t =
0) = 1 corresponds to fully charged battery with the initial opencircuit voltage Vo c = Voc , while W bat (t = L) = 0 corresponds
to Vo c dropping to the cutoff voltage (Vcuto ). The value of Voc
and Vcuto depends upon the battery type. So under a typical
load, battery Vo c drops to Vcuto over time. Hence, the battery
lifetime L can also be defined as the time taken by battery Vo c
to drop to Vcuto below which no further power (current) can
be drawn from the battery. This definition is consistent with the
battery lifetime definition in Section II-A, i.e., the time duration
during which the concentration of the electroactive species drops
from its initial concentration C to a specific threshold Ccuto .
A. Description of Typical Power Load Usage
The typical power loads (both constant and varying) used in
this paper are the average power consumed during various operating modes of the Itsy pocket computer, which is powered by
a 2.2 Wh Li-ion battery [27]. The power loads used to estimate
battery lifetime in this paper were taken from [27] in which the
power consumption of Itsy pocket computer were reported. Let
P represent a set consisting of various actual average power
loads (both constant and varying). The set P is divided into two
subsets, PT and PC , where the subset PT consists of 15 constant
average power loads and the subset PC consists of seven varying
average power loads. The average power loads in both subsets
are downscaled and are used to estimate the lifetime of 1.41 Wh
Li-ion battery as mentioned earlier. See Appendix B, for details
on the computation of downscaling factor and assumptions. The
actual and downscaled power loads are summarized in Tables
VI and VII were computed using average current and average
voltage of 3.75 V [27]. So the lifetimes estimated and measured
are average lifetimes. Given the upper and lower bounds on
the load values, the models discussed in Section II can be used
to estimate upper and lower bounds on the actual lifetimes.
From here onward, we will refer average power load used to
estimate battery lifetime as power load for simplicity.
In Table VI, the constant power loads (T1T10) represent
the power consumed during different operating modes of Itsy
computer. In addition, peripheral devices, such as Microdrive
hard disk [28] and a WaveLAN wireless card [29] were attached
to Itsy, and power consumptions (T11T15) were recorded for
different operating modes of Itsy and peripheral devices [16],
[27]. The following letter abbreviations are used to describe the

operation modes of Itsy and peripheral devices in Table VI [16],

[27].
1) Itsy: IIdle, MMPEG, DDictation, TTalk1, WWAV1,
and SSleepDC.
2) Microdrive: SStandby and AAccess.
3) WaveLAN: DDoze, RReceive, and TTransmit.
The power loads in Table VII represent the measured power
consumption when Itsy operation is switched between different modes. In Table VII, the second column represents the
actual power measurements, the third column represents the
downscaled power measurements, and the fourth column represent the time duration in which each power level was measured.
Cases C1 and C2 have the same load power levels. However, the
order in which the battery is subjected to power load is reversed.
In case C1, load power is a decreasing staircase function, except
for the last period in which the load power jumps up and remains
constant. In case C2, load power is an increasing staircase function and that remains constant at the same power value as case
C1 during the final period. Case C3 is similar to C2, except it
has a rest period of 50 min. Case C4 is a periodic repetition of
C2 for ten cycles (represented by superscript 10 in Table VII),
but the time for each power level is downscaled by a factor of
ten (also represented by superscript 10 in Table VII); hence the
total load power consumption in cases C2 and C4 is identical.
Case C5 is similar to C4 except that the first two power levels in
C5 are lower than the ones in C4. The time interval column in
Table VII for cases C4 and C5 has a subscript 200, which means
that after ten load cycles, the total time duration is 200 min.
Cases C6 and C7 have different rest periods. Cases C3, C6, and
C7 are used to evaluate the recovery effect of the battery, since
they contain rest period.
The typical power loads summarized in Tables VI and VII
represent a broad range of loading condition used to exercise
the battery.
B. Numerical Estimation of Battery Lifetime
Given the partially linearized battery model during discharging (v = 0) with the incorporated recovery model [see (12) and
(14)], the discharging coefficients dk ,v , k = 1, 2, 3 for downscaled operating nominal power Pbat,nom of 181.2 mW presented in Table VIII is computed, as per the procedure described
in Section II-B, and various downscaled power loads in Tables
VI and VII, a numerical lifetime estimation of 1.41 Wh Li-ion
battery is computed in MATLAB. Note that all the parameters and the downscaled power load profiles used in the lifetime

830

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

TABLE VIII
MODEL COEFFICIENTS FOR THE NONLINEAR EFFICIENCY APPROXIMATION FOR
DISCHARGING (v = 0) AT THE NOMINAL POWER LEVEL OF 181.2 mW FOR
1.41 Wh LI-ION BATTERY AT ROOM TEMPERATURE

TABLE XI
MEASURE OF BATTERY RECOVERY

TABLE IX
MEASURED AND ESTIMATED LIFETIMES FOR EACH CONSTANT POWER LOAD

summarized in Tables IX and X along with measured lifetimes

in the first column. The Rakhmatov and Virudhula battery model
parameters: = 33706 and = 0.75 reported in [16] is used to
estimate battery lifetime. The error between the measured and
the estimated lifetime is computed using (15) and is also listed
in Tables IX and X
Error(%) =

measured lifetime estimated lifetime

100.
measured lifetime
(15)

C. Discussion

TABLE X
MEASURED AND ESTIMATED LIFETIMES FOR EACH VARYING POWER LOAD

estimation process is for 1.41 Wh Li-ion battery and is measured

at room temperature. The downscaled nominal operating input
power Pbat,nom is 181.2 mW, which is obtained by downscaling
the average battery power flow in the Itsy pocket computer.
The recovery time constant () computed using the downscaled
load data is 0.016 s1 (see Appendix C). The percentage of
recovery is selected as 3.27% because the partially linearized
battery model during discharging without the recovery model
underestimated the battery lifetime by 3.27% on average at nominal power.
The results of lifetime estimation using the partially linearized
battery model during discharging (v = 0) with the incorporated
recovery model for downscaled power load profiles PT and PC
are summarized in Tables IX and X, respectively. Based on the
downscaling assumption in Appendix C, the lifetime of 1.41 Wh
Li-ion battery computed using the downscaled power loads must
be equal to the lifetime of 2.2 Wh Li-ion battery computed (measured) using the actual power loads. Henceforth, the estimated
lifetimes of 2.2 Wh Li-ion battery using the Rakhmatov and
Virudhula battery model [8] [see (6)] for PT and PC are also

In this section, we analyze and discuss the performance of

the partially linearized model with the incorporated recovery
model under varying load conditions. A broad range of selected
load profiles used to exercise the battery model in this paper
is a real-time power consumption of Itsy computer under different operating modes [16], [27]. The typical loads presented
in Tables VI and VII captures both static and dynamic nature
of the power usage observed in many real-time applications.
In addition, to lifetime estimation using the partially linearized
model with the incorporated recovery model, we also compute
the percentage of recovery observed in cases C3, C6, and C7.
We define a few terms as in [8]. Let Loriginal represent the estimated lifetime for the power profile when no load relaxation (no
rest period) is exercised. When the rest period is present, let
denote the rest period in a load profile and let Lunaected denote
the expected lifetime as if no recovery took place during the rest
period of duration , i.e., Lunaected = Loriginal + . In general,
the lifetime L should be greater than Lunaected due to recovery
effect. The quantity ((L Lunaected )/Loriginal ) 100% represents the lifetime extension due to recovery effect. Table XI
summarizes the recovery percentage (i.e., lifetime extension)
obtained by using the partially linearized battery model along
with the incorporated recovery model with respect to Loriginal .
Cases C3, C6, and C7 having a rest period of 50, 35.2, and
24.9 min, respectively, show an increase in estimated battery
lifetime by 0.94%, 4.3%, and 6.7%, respectively, (with respect
to Loriginal ) due to recovery effect. These observations ascertain the fact that the partially linearized battery model with the
incorporated recovery model is able to capture the recovery effect. To establish the accuracy of the recovery percentage, we
compare the percentage increase in the lifetime computed using
the partially linearized model with the incorporated recovery
model and the percentage increase in the lifetime when the DUALFOIL simulator is used on the same cases C3, C6, and C7.
The percentage increase in the lifetime with the DUALFOIL
reported in [8] is 1%, 1.7%, and 3.7% for C3, C6, and C7, respectively. However, the percentage of recovery observed in C3,

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

C6, and C7 can be slightly higher than the actual recovery of real
battery. This is due to the fact that the partially linearized model
underestimates the power losses when the load power is higher
than Pbat,nom , as shown in Fig. 3. In general, the battery recovery has more profound effect when the power discharge rate is
high. Since the power discharge rate in C7 is higher than one in
C6 and C7 exhibits higher percentage recovery, as indicated in
Table XI.
Intuitively, a battery at a higher SOC must be able to handle
a large power load better than a battery at a lower SOC. This
assists in identifying the power load pattern that would result in
better lifetime. The cases C1 and C2 have the same power load
values except the order in which the battery is subjected to load
power is reversed (as described in Section IV-A). As expected,
the lifetime of C1 is slightly higher than the lifetime of C2. Case
C4 is a periodic repetition of an increasing staircase as in C2
for ten cycles, but the time for each power level is downscaled
by a factor of ten, while the total load power consumption in
both cases is identical. The battery in case C2 handles high-load
power at low SOC, while in case C4, the battery is subject to
periodic variation of both low- and high-load power at low SOC.
As a result, the lifetime estimated in case C4 is slightly better
than the lifetime estimated in case C2.
In the case of selected constant loads (see Table VI), the
lifetime estimation widely ranges from 2 h to 10 days. Table IX
shows the comparison of the accuracy of the lifetime estimations
between the partially linearized model and the chemical-based
Rakhmatov and Virudhula battery model [8]. The estimation
error of the partially linearized battery model is less than 10%
on average at nominal load power for the typical constant loads
listed in Table VI. We observe that the estimation error increases
as the load power level increases or decreases from the nominal
input power. This observation is consistent with the assumption
made during the linearization of nonlinear battery efficiency
[see (8)] and the validation results presented in Section III. For
example, the estimation error for 701.4 mW load power, which
is 520.2 mW more than nominal input power, is 9.8%, while
the error for 7.2 mW load power, which is 174 mW less than
nominal input power, is 7.2%. In the case of Rakhmatov and
Virudhula battery model (6), the estimation error increases as
the load on the battery decreases.
Based on the numerical results presented earlier, we observe
that the inputoutput partially linearized model with incorporated recovery model provides reasonably accurate lifetime estimation on average at nominal power in comparison to the
lifetime estimation result obtained from the Rakhmatov and
Virudhula model [8].
V. BATTERY CYCLE LIFE VALIDATION
In general, cycling a battery at different DODs and
charge/discharge power rates affects battery cycle life as battery
m ax
is a function of
ages. Data in the literature suggests that Wbat
number of cycles, DOD, and nominal charge/discharge power
level. In this section, we specifically model how the maximum
rated energy stored in a battery is affected by repeated charging
and discharging at different power levels given a specific level

831

of DOD. Full-discharge cycling test (equivalent to 100% DOD)

is one of many cycling tests. IEEE Standard recommends fulldischarge cycling test as it corresponds to a worst case scenario,
but there are lots of charge/discharge cycling test at different
DODs [30]. Incorporation of the DOD into the battery model is
beyond the scope of this paper.
In this paper, battery cycle life is defined as the number of
cycles (discharging than charging) the battery can withstand
before the maximum capacity of the battery reduces to 80% of
its initial ampere-hour rating at a specific level of DOD under
the assumed condition of constant temperature. Battery cycle
life represents a metric different from the chemical lifetime
definition in Section IV. For a fixed DOD, if the battery is
cycled at a rated charge/discharge power level, the battery last
for N cycle. Further, smaller rated power levels in each cycle
prolongs the battery cycle life (>N ), while the effect of higher
power shorten the battery cycle life (<N ), respectively.
In practical applications, rechargeable batteries are exposed
to a wide range of power loads. At higher power load for a
fixed DOD and constant temperature, the battery cycle life reduces significantly when compared to lower power loads under
same conditions. Drouilhet [31] argued that uneven power load
results in loss of conductivity of electroactive species in the battery. This leads to uneven energy distribution and higher stress.
Thus, unevenness in discharge power load increases the stress,
which likely leads to shorter battery cycle life, in a manner analogous to mechanical fatigue. Coleman et al. [32] used the metric
state of health to determine aging in value regulated lead-acid
(VRLA) and Li-ion battery types. In [31], an empirical model
was developed to predict battery cycle life for hybrid power
applications and validated on data provided by the manufacturer. Most of the experimental analysis on battery cycle life is
performed at a fixed charge/discharge rate, usually the rate for
which the batterys rated capacity is given [25], [31].
In order to account for the effect of different charge/discharge
power levels, we approximate the maximum rated energy term
m ax
in the partially linearized model [see (12)] as a function
Wbat
of number of cycles and nominal charge/discharge power load
Pbat,nom at a fixed DOD as follows:
m ax
m ax
Wbat
Wbat
(N, DOD, Pbat,nom ).

(16)

As an illustration, the approximation of the maximum rated

m ax
energy term Wbat
is obtained by interpolating the capacity data
at different charge/discharge power level obtained from [33] for
DOD = DOD = 50% at a constant temperature of 25 C. The
approximation is empirically selected as follows:
m ax
Wbat
= aN 2 + bPbat,nom + c.

(17)

Then, we validate the maximum rated energy approximation

[see (17)] for different charge/discharge power levels. The
coefficients are computed at nominal charge/discharge power
level for consistency because the initial model is partially
linearized about its nominal power during both charging and
discharging. The validation results for 965 mAh 4.1 V Li-ion
battery at nominal charge/discharge power level of 3780 mW
using the computed coefficients, a = 0.005, b = 0.014, and
c = 3.6 103 is shown in Fig. 5. The two-norm normalized

832

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

Fig. 5. Maximum rated energy term W bmaat x approximation at the nominal

charge/discharge power level of 3780 mW for 965 mAh 4.1 V Li-ion battery.

error for charge/discharge power level closer to nominal power

level is 2.5%. As the charge/discharge power level deviates
from the nominal power level, the two-norm normalized error
increases, as expected.
Thus, from Fig. 5, we observe that the validation result for
m ax
is accurate about its nominal
the maximum rated energy Wbat
charge/discharge power level and the error increases slightly
as the power level deviates from Pbat,nom . This validation pattern is consistent and similar to the pattern observed during
the validation of the linearized model for Li-ion battery type in
Section III-D. This is due to the fact that the battery efficiency
decreases with repeated discharging and charging, as the internal resistance of the battery increases, which is consistent with
the findings of Shim and Striebel [34] on a Li-ion battery.
Thus, in this section, we showed that by approximating the
m ax
as a function of number of
maximum rated energy term Wbat
cycle, nominal charge/discharge power level at a fixed DOD,
we can incorporate the cycling effect.
VI. CONCLUSION
In this paper, we have shown that the dynamic measurement of the battery SOC was approximated by the partially
linearized (control-oriented) battery model for both charging
and discharging. A complete mathematical development of the
partially linearized battery model was presented in Section II-B
and in Appendix A. The model was augmented with an additional differential equation to capture the recovery effect. Then,
the generalization of the scalable partially linearized model to
different battery types at different temperatures and its validation against the actual data for both charging and discharging
was demonstrated. Based on the validation results, a functional
relationship between the model coefficients and temperature
was set forth for lead-acid and Ni-MH battery types. The model
was found to be reasonably accurate around its nominal input
power level for various battery types at different temperatures.
The battery model with the incorporated recovery model was
exhaustively exercised under a variety of power loads for Li-ion
battery lifetime estimation. Actual lifetime measurements and

the estimated lifetimes using the Rakhmatov and Virudhula battery model were given for comparison. By approximating the
maximum rated energy term of the linearized model as a function of number of cycle and nominal charge/discharge power
level at a specific DOD level, the effect of repeated charging
and discharging on the overall battery cycle life was presented.
Although the developed model is linear in the input, it captures
all the essential nonlinear characteristics of the battery under
varying load conditions.
The partial linearized inputoutput battery model has been
used successfully for control-oriented power flow management
in an HEV application [11], [26]. Specifically, the model is very
useful for control when an inverter controls the currentvoltage
levels (and thus power levels) between the battery and any device, such as an electric drive, sensor, or actuator. In distributed
sensor network application, estimation of battery longevity, estimates of scheduled maintenance, and replacement are some
of the key issues. The usage of the model is computationally
inexpensive and can provide reasonably accurate information
on the SOC of the batteries. This information will assist in intelligent network activity scheduling, dynamic rerouting, and
battery energy aware protocol designs in scenarios, where battery maintenance and replacement is impractical.
The accuracy of the partially linearized battery model can be
improved by selecting different nominal input powers based on
the operating load power. For example, one can create a set of
range of load powers and select an associated nominal input
power for each range to lower the deviation between load power
and nominal input power, and hence to reduce the approximation
error.
APPENDIX A
Consider a general nonlinear state equation x = f (x, u),
where x is the state, and u is the (control) input. A Taylor
series expansion around the nominal operating input u is given
by
f (x, u)
|u u + H.O.T.
(18)
u
where H.O.T. means higher order terms that are assumed small
and negligible. Therefore, the linearization of (8) about a nomv
, v {0, 1}, becomes
inal battery power Pbat,nom
x = f (x, u)|u +

W
bat
v
Pbat,nom
m ax
Wbat


v
+ d3,v ]Pbat}

{[ln(W bat + d1,v ) + d2,v Pbat,nom


+

v
Pbat

v
= [ln(W bat + d1,v ) + d2,v Pbat,nom
+ d3,v ]

Pb a t, n o m

Pbat
m ax
Wbat

(19)

v
where Pbat = Pbat (t) Pbat,nom
. Now, we compute the parv
tial derivative in (19) and evaluate at Pbat,nom


{[ln(W bat + d1,v ) + d2,v Pbat + d3,v ]Pbat }



v
Pbat
Pb a t, n o m

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

= {[ln(W bat + d1,v ) + d2,v Pbat + d3,v ]

d2,v Pbat }|P bva t , n o m
v
= [ln(W bat + d1,v ) + 2d2,v Pbat,nom
+ d3,v ].

(20)

Substituting the expression in (20) into (19) results in the

following equation:

833

lifetime of 2.2 Wh battery is same as the lifetime of 1.41 Wh

battery, whenever the Li-ion battery 2 (1.41 Wh) is subject to
power load Pbat2 = Pbat1 , where Pbat1 denote the average
power delivered by 2.2 Wh Li-ion battery. Hence, the average
typical power load used in this study for 1.41 Wh Li-ion battery
is obtained.
APPENDIX C

Pbat,nom

v
W
bat = [ln(W bat + d1,v ) + d2,v Pbat,nom + d3,v ]
m ax
Wbat
v
[ln(W bat + d1,v ) + 2d2,v Pbat,nom
+ d3,v ]

Pbat
m ax .
Wbat
(21)

For our control purposes, we want W

bat equation to depend
on the actual battery input Pbat (t) rather than Pbat . For this,
v
into (21)
we substitute Pbat = Pbat Pbat,nom
v

Pbat,nom

v
W
bat = [ln(W bat + d1,v ) + d2,v Pbat,nom + d3,v ]
m ax
Wbat
v
[ln(W bat + d1,v ) + 2d2,v Pbat,nom
+ d3,v ]

v
Pbat (t) Pbat,nom
m ax
Wbat

(22)

and arrive at the final partially linearized equation for the battery
SOC
d2,v

v
2
W
bat (t) =
m ax (Pbat,nom ) [ln(W bat (t) + d1,v )
Wbat
v
+ 2d2,v Pbat,nom
+ d3,v ]

Pbat (t)
m ax .
Wbat

(23)

The coefficients dk ,v , k = 1, 2, 3 and v {0, 1} for different

battery types are presented in Tables I, III, and V.

In Section II-D, we discussed that batteries exhibit recovery

effect when, during which the SOC increases before, reaching a
steady-state value. The rate of increase in SOC during recovery
is approximated using a recovery time constant. In the recovery
model presented in Section II-D, the parameter [see (14)]
represents the recovery time constant. The value of depends
upon battery type and is often computed using experimental
data. The recovery time constant () of 1.41 Wh battery used in
this paper is computed by using the downscaled experimental
recovery data of same type 2.2 Wh Li-ion battery from [21]. The
actual data is monotonically increasing and reaches a steadystate value over time. In another words, the data consists of the
transient response of batterys SOC and its steady-state value.
A simple logarithmic approach is used to compute . This
common procedure is to compute the time constant of a RC
circuit from the transient response. Consider a general state
equation of an RC circuit with zero input. The state trajectory
(due to an initial condition) is exponentially decreasing and is
expressed as follows:
x(t) = x(0)et/

where x(t) is the state at time t, x(0) is the initial value of the
state, and = RC is the time constant with R and C being the
resistance and the capacitance, respectively.
Taking natural logarithmic on both side of (25), we get
ln(x(t)) = ln(x(0)) t/.

APPENDIX B
The discharging and charging efficiency of lead-acid (25 of
18 Ah 12.5 V) battery pack is obtained from [35]. In order to
obtain discharging and charging efficiency maps of lead-acid (30
of 12 Ah 12 V) battery pack used in this paper, a downscaling
is performed.
m ax
be the maximum rated energy of lead-acid battery
Let Wbat1
m ax
be the maximum rated energy of the leadpack 1 [35] and Wbat2
acid battery pack 2, the one used in this paper. We define scaling
m ax
m ax
/Wbat1
. The assumption used in downscaling
factor = Wbat2
the battery efficiency map of [35] to the size used in this paper is
as follows: let Pbat1 denote the power delivered by battery pack
1 and Pbat2 that of battery pack 2, whenever = Pbat2 /Pbat1 ,
then for v {0, 1}
bat1 (SOC, Pbat1 , v) = bat2 (SOC, Pbat2 , v).

(24)

A similar scaling procedure and assumption is adopted in downscaling the load powers of [16], [27] to the energy level of
m ax
Li-ion battery used in this paper: let Wbat1
represent the maxm ax
be
imum rated energy of Li-ion battery 1 (2.2 Wh) and Wbat2
the maximum rated energy of Li-ion battery 2 (1.41 Wh) used
m ax
m ax
/Wbat1
be a scaling factor. The
in this paper. Let = Wbat2

(25)

(26)

Equation (26) is a straight line when plotted against time t, with

the slope of 1/ and the intercept of ln(x(0)). Only the slope
of the line determines the time constant.
To compute the time constant of the complete response of (14)
due to both constant input and the initial condition, we modify
the data into an exponentially decreasing response. Therefore,
we first determine the steady-state value of the response, and
then, subtract the data from the steady-state value. This leads
to an exponentially decreasing response. Then, take the natural
logarithm and perform least-square fit to the response data. The
fit will be a straight line when plotted against time t and the
slope of the line is negative inverse of RC time constant.
The procedure described earlier to compute time constant
of a simple RC circuit with nonzero input is used to compute
. The downscaled data of 2.2 Wh Li-ion battery is subtracted
from the steady-state value to obtain an exponentially decreasing
tr
function. Let W bat (t) be the difference between the steadystate value and the complete response of the SOC, with an
tr
tr
initial condition of W bat (0). Similar to (25), W bat (t) decays
exponentially to zero as per
tr

tr

W bat = W bat (0)et/ .

(27)

834

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 25, NO. 3, SEPTEMBER 2010

Again, take natural logarithmic on both side of (27) and result

in
tr

tr

ln(W bat (t)) = ln(W bat (0)) t/.

(28)

tr
W bat (t)

Thus, the function

is linear with time t and with the
slope of 1/ . Hence, the estimate of is 1/ .
ACKNOWLEDGMENT
The authors would like to thank reviewers for their valuable
comments and suggestions. They would also like to thank Rick
Mayer (Purdue University) for his thoughtful discussions on
battery modeling.
REFERENCES
[1] R. Rao, S. Virudhula, and D. Rakhmatov, Battery models for energy
aware system desing, IEEE Comput., vol. 36, no. 12, pp. 7787, Dec.
2003.
[2] M. Doyle, T. F. Fuller, and J. S. Newman, Modeling of galvanostatic
charge and discharge of the lithium/polymer/insertion cell, J. Electrochem. Soc., vol. 140, no. 6, pp. 15261533, 1993.
[3] T. F. Fuller, M. Doyle, and J. S. Neuman, Simulation and optimization of
the dual lithium ion insertion cell, J. Electrochem. Soc., vol. 141, no. 1,
pp. 110, 1994.
[4] M. Pedram and Q. Wu, Design considerations for battery-powered electronics, in Proc. 36th Design Autom. Conf., 1999, pp. 861866.
[5] K. C. Syracuse and W. D. K. Clark, A statistical approach to domain
performance modeling for oxyhalide primary lithium batteries, in Proc.
12th Annu. Battery Conf. Appl. Adv., 1997, pp. 163170.
[6] S. Gold, A PSPICE macromodel of lithium ion batteries, in Proc. 12th
Annu. Battery Conf. Appl. Adv., 1997, pp. 215222.
[7] C. F. Chiasserini and R. R. Rao, Energy efficient battery management,
IEEE J. Sel. Areas Commun., vol. 19, no. 7, pp. 12351245, Jul. 2001.
[8] D. N. Rakhmatov and S. Vrudhula, An analytical high-level battery model
for use in energy management of portable electronic systems, in Proc.
IEEE/ACM Int. Conf. Comput. Aided Design, 2001, pp. 488493.
[9] P. Rong and M. Pedram, An analytical model for predicting the remaining
battery capacity of lithium-ion batteries, in Proc. Design Autom. Test Eur.
Conf. Expo., 2003, pp. 11481149.
[10] K. Uthaichana, R. DeCarlo, S. Pekarek, S. Bengea, and M. Zefran, Development of a control-oriented power flow model for a parallel hybrid
electric vehicle, Purdue Univ., West Lafayette, IN, Tech. Rep. TR-ECE06-13, Oct. 2006.
[11] K. Uthaichana, Modeling and control of a parallel hybrid electric vehicle,
Ph.D. dissertation, Purdue Univ., West Lafayette, Dec. 2006.
[12] A. Tiwari, P. Ballal, and F. L. Lewis, Energy efficient wireless sensor
network design and implementation for condition-based maintenance,
ACM Trans. Sens. Netw., vol. 3, no. 1, pp. 123, 2007.
[13] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson,
Wireless sensor networks for habitat monitoring, in Proc. 1st ACM Int.
Workshop Wireless Sens. Netw. Appl., 2002, pp. 8897.
[14] J. Khan and R. Vemuri, Energy management in battery-powered sensor
networks with reconfigurable computing nodes, in Proc. Int. Conf. Field
Programmable Logic Appl., 2005, pp. 543546.
[15] S. Park, A. Savvides, and M. B. Srivastava, SensorSim: A simulation
framework for sensor networks, in Proc. MSWIM, 2000, pp. 104111.
[16] D. Rakhmatov, S. Vrudhula, and D. A. Wallach, A model for battery
lifetime analysis for organizing applications on a pocket computer, IEEE
Trans. VLSI Syst., vol. 11, no. 6, pp. 10191030, Dec. 2003.
[17] M. R. Jongerden and B. R. Haverkort, Which battery model to use? in
Proc. 25th U.K. Perform. Eng. Workshop, 2008, pp. 7688.
[18] S. C. Chapra, Ed., Applied Numerical Methods with MATLAB for Engineers and Scientists. New York: McGraw-Hill, 2005.
[19] M. Hopka, A. Brahma, S. Dilmi, G. Ercole, C. Hubert, S. H. K. Huseman,
G. Paganelli, M. Tateno, Y. Guezennec, G. Rizzoni, and G. Washington,
The 2000 Ohio state university future truck, SAE, OH, Tech. Rep. SP167, 2008.
[20] J. Aylor, A. Thieme, and B. Johnson, A battery state-of-charge indicator for electric wheelchair, IEEE Trans. Ind. Electron., vol. 39, no. 5,
pp. 398409, Oct. 1992.

[21] D. Rakhmatov and S. Vrudhule, Energy management for battery-powered

embedded systems, ACM Trans. Embedded Comput. Syst., vol. 2, no. 3,
pp. 277324, 2003.
[22] H. L. N. Wiegman and A. J. A. Vandenput, Battery state control techniques for charge sustaining applications, in Proc. Int. Congr. Expo.,
1998, pp. 6575.
[23] M. Coleman, C. B. Zhu, C. K. Lee, and W. G. Hurley, A combined
SOC estimation method under varied ambient temperature for a leadacid battery, in Proc. IEEE Appl. Power Electron. Conf. Expo., 2005,
pp. 991997.
[24] L. Serro, Z. Chebab, Y. Guezennec, and G. Rizzoni, An aging model of
Ni-MH batteries for hybrid electric vehicles, in Proc. IEEE Conf. Veh.
Power Propulsion, 2005, pp. 7885.
[25] D. Linden and T. Reddy, Eds., Handbook of Batteries, 3rd ed. New
York: McGraw-Hill, 2001.
[26] K. Uthaichana, S. Benga, R. DeCarlo, S. Pekarek, and M. Zefran, Hybrid
model predictive control tracking of a sawtooth driving profile for an
HEV, in Proc. Amer. Contr. Conf., 2008, pp. 967974.
[27] M. Viredaz and D. A. Wallach, Power evaluation of a handheld computer:
A case study, Compaq WRL Res. Rep., Palo Alto, CA, Tech. Rep. WRL2001-1, May 2001.
[28] Microdrives: 2001. Product specification, 2001.
[29] Orinoco WaveLAN, 2001. Product specification.
[30] T. Guena and P. Leblanc, How the depth of discharge affects the cycle
life of lithium-metal-polymer batteries, in Proc. INTELEC 28th Annu.
Int. Telecommun. Energy Conf., 2006, pp. 18.
[31] S. Drouilhet, A battery life prediction method for hybrid power applicaitons, in Proc. 35th AIAA Aerosp. Sci. Meeting Exhibit., 1997.
[32] M. Coleman, W. G. Hurley and C. K. Lee, An improved battery characterization method using a two-pulsed load test, IEEE Trans. Energy
Convers., vol. 23, no. 2, pp. 708713, Jun. 2008.
[33] T. R. Crompton, Battery Reference Book, 3rd ed. Boston, MA: Newnes,
2000.
[34] T. R. Shim and K. A. Striebel, Cycling performance of low-cost lithiumion batteries with natural graphite and LiFePO4, J. Power Sources,
vol. 119, pp. 955958, Jun. 2003.
[35] C. C. Lin, H. Peng, J. W. Grizzle, and J. Kang, Power management
strategy for a parallel hybrid electric truck, IEEE Trans. Control Syst.
Technol., vol. 11, no. 6, pp. 839849, Nov. 2003.

Vivek Agarwal (S98) received the B.E. degree

from the University of Madras, Chennai, India, and
the M.S. degree from the University of Tennessee,
Knoxville, TN, in 2001 and 2005, respectively, both
in electrical engineering. He is currently working toward the Ph.D. degree from the School of Nuclear
Engineering, Purdue University, West Lafayette, IN.
His doctoral research focuses on power management
in battery powered interconnected wireless networks.
In 2005, he was a Research Associate at the
Hewlett-Packard Research Laboratories, Palo Alto,
CA . His current research interests include power management in wireless sensor networks, battery modeling, signal and image processing, nuclear material
detection, regulation techniques, and machine learning.

Kasemsak Uthaichana (M07) received the B.S. degree in electrical, computer, and systems engineering
from Rensselaer Polytechnic Institute, Troy, NY, in
2000, and the M.S. and Ph.D. degrees in electrical
and computer engineering from Purdue University,
West Lafayette, IN, in 2002 and 2006, respectively.
His doctoral research focused on modeling and power
flow control for parallel hybrid electric vehicles.
From 2007 to 2008, he was with Caterpillar Inc.,
Peoria, IL, where he was engaged in integrated power
systems control software for wheel loader machines.
He is currently in the Department of Electrical Engineering, Chiang Mai University, Chiang Mai, Thailand. His research interests include power management in
hybrid vehicles, operational control of fuel cells, battery management system,
and friction estimation and compensation.
Dr. Uthaichana is an Active Member of the IEEE Control Systems Society
and the IEEE Power and Energy Society.

AGARWAL et al.: DEVELOPMENT AND VALIDATION OF A BATTERY MODEL

Raymond A. DeCarlo (F89) He received the B.S.

and M.S. degrees in electrical engineering from the
University of Notre Dame, Notre Dame, IN, in 1972
and 1974, respectively, and the Ph.D. degree from
Texas Tech University, Lubbock, TX.
In 1977, he joined Purdue University, West
Lafayette, IN as an Assistant Professor of electrical
engineering, and became an Associate Professor in
1982 and Full Professor in 2005. During the summers
of 1985 and 1986, he was at the General Motors Research Laboratories. He has authored or coauthored
three books, more than 50 journal, more than 100 conference, and eight book
chapter/reprint articles. His current research interests include interdisciplinary
ranging from variable structure control, hybrid optimal control and stability,
hybrid electric vehicle modeling and control, biological modeling and control,
as well as the control and modeling of the software test process.
Dr. DeCarlo was a Secretary Administrator of the IEEE Control Systems
Society, and from 1986 to 1992 and from 1999 to 2003, a member of the Board
of Governors. He was an Associate Editor of Technical Notes and Correspondence, and Survey and Tutorial Papers, both for the IEEE TRANSACTIONS ON
AUTOMATIC CONTROL. He was a Program Chairman for the 1990 IEEE Conference on Decision and Control (CDC), Honolulu, and a General Chairman
of the 1993 IEEE CDC, San Antonio. During 2001 and 2002, he was the Vice
President for Financial Activities for the IEEE Control System Society (CSS).
He was the recipient of CSSs Distinguished Member Award in 1990, the IEEE
Third Millennium Medal in 2000, the EATON Award in Electrical and Computer Engineering (ECE) in 2002, the Motorola Excellence in Teaching Award
in 2006, and the award for Best Theoretical Paper in Automatica in 2008. He
was a Chair of the Purdue University Senate and a Committee on Institutional
Cooperation (CIC) Faculty Fellow.

835

Lefteri H. Tsoukalas (M89) received the Ph.D.

degree from the University of Illinois, UrbanaChampaign, in 1989.
He was a Professor and Former Head of the
School of Nuclear Engineering, Purdue University,
West Lafayette, IN, where he also holds a courtesy
professorial appointment in construction engineering
and management. His current research interests include in developing smart energy methodologies. He
has authored or coauthored more than 200 research
publications in the area including a book titled Fuzzy
and Neural Approaches in Engineering, (New York: John Wiley and Sons,
1997). He was with the faculty of the University of Tennessee, Knoxville, Aristotle University, and the University of Thessaly. He was also a Researcher and
an Advisor, and a Consultant at the Japan Atomic Energy Research Institute, the
Ministry of Education, ON, Canada; the International Atomic Energy Agency;
the Agency for Science, Technology and Research of the Government of
Singapore; and the U.S. Department of Energy.
Dr. Tsoukalas was the recipient of numerous awards and recognitions including Best Teacher Award at Purdue, Fellow of the American Nuclear Society,
and the 2009 recipient of the Humboldt Prize.