You are on page 1of 6

Proceedings of the American Control Conference

Arlington, VA June 25-27, 2001

Battery State-of-Charge Estimation


Shuo Pang, Jay Farrell, Jie Du, and Matthew Barth
Electrical Engineering, University of California, Riverside
Abstract
This paper discusses the problem of lead acid battery state-of-charge (SOC) estimation for (hybrid) electric vehicles. The problem is to accurately estimate the
remaining battery capacity for both driver notification
and automated energy management. The article contains a review of the problem, existing solution methods,
presentation of a new sohtion method, and experimental analysis of the performance of that method.

1 Introduction
Drivers of internal combustion engine driven cars are
accustomed to checking a gas gauge to estimate the
amount of driving for which their vehicle is currently capable. Measurement of the remaining gasoline available
is a fairly simple problem solved by measuring the level
of fluid in the tank. Electric or hybrid electric vehicles
use batteries for energy storage. The electric-only driving range is determined predominantly by the battery
SOC. It is therefore necessary to have methods capable of accurately estimating battery SOC for EV driver
notification and for HEVl battery SOC management to
succeed. In fact, [5] claims that Research leading to the
determination of an optimum battery pack management
strategy is probably the single most important technical issue in the successful commercialization of EVs.
Accurate SOC determination is a critical component of
battery pack management.
The HEV energy management application is the
problem motivating this article. In contrast to the EV
application which involves primarily battery discharge,
the HEV application may involve repeated charge discharge cycles. This distinction is critical. In EV applications, complete stationary recharge allows the SOC
estimation algorithm to be reset to an accurate initial
condition; therefore, estimation errors do not accumulate over multiple discharge/recharge cycles. A primary
goal of HEVs is to attain long vehicle range by allowing
the batteries to be recharged while in motion. Proper
operation of the energy manager depends on the accuAn HEV has an auxiliary power unit (e.g., an internal combustion engine) capable of either charging the batteries or providing traction power. The onboard energy management system attempts to optimize the use of the vehicles available energy sources
(e.g.lbattery or fuel) to
driver satisfaction while minimizing total fuel usage and emissions.

0-7803-6495-3/01/$10.00 0 2001 AACC

racy of the battery SOC estimate. If this estimate is


inaccurate, the battery pack may either be over-charged
or depleted to the point that some cells are driven into
reverse polarity. HEV SOC estimation is more difficult
than the EV application, since the HEV SOC estimation
problem can allow errors to accumulate over numerous
discharge/recharge cycles before a full recharge resets
the SOC estimate to an accurate initial condition.
SOC estimation for lead acid batteries is discussed
for stationary applications in [l,6 , 141 and for traction
applications in [2, 3, 41. There are five categories of
methods for estimating SOC: electrolyte specific gravity,
stabilized float current, coulometric measurement, open
circuit battery voltage, and loaded battery voltage.
The first two approaches are discussed in IEEE Std.
450 [l]and compared in [SI. The conclusion of [6] is
that under float charging conditions, the Aoat current is
a more accurate and rapid indicator of battery SOC than
specific gravity. The specific gravity method is based on
the fact that the electrolyte specific gravity, which is a
function of amount of acid in the electrolyte, is linearly
related to the remaining battery capacity (See Figure
3 in [3]). In stationary applications, specific gravity is
measured via a hydrometer. A long rest period is required after discharge/charge before specific gravity can
be accurately measured due to electrolyte diffusion dynamics. ManuaI measurement of specific gravity is not
feasible for mobile (traction) applications nor for sealed
lead acid batteries. The stabilized float current method
relies on the fact that at the conclusion of fully recharging a battery, when the battery is held at a constant
float voltage the charging current will stabilize (i.e., no
significant change for three hours). The float current
method is obviously not practical for real-time estimation of SOC in mobile applications.
Section 2.1 reviews lead acid battery chemistry as
it relates to the problem of in-vehicle SOC estimation.
Subsequent sections present and analyze various existing
and new methods for in-vehicle SOC estimation that
rely on coulometric, open circuit battery voltage, and
loaded battery voltage methods.
2 B a t t e r y Models

This section presents an overview of battery chemical


processes and electrical models. More detailed discussions can be found, for example, in [lo].

1644

2.1 Chemical Model ([5, 6, 10, 121)


The ionization and discharge processes for a lead acid
battery are summarized in [5, lo]. A fully charged lead
acid battery consists of a lead dioxide positive plate, a
lead metallic plate, and a sulfuric acid electrolyte solution. Ionization of the electrolytic solution and the positive plate results in sulfate ions, hydrogen ions, hydroxide ions, and tetravalent lead ions. When a current carrying path exist (either internally or externally) between
the positive and negative plates, electrons flow from the
negative plates to the positive plates through the current
path. An ideal fully discharged battery would consist of
a water solution and electrodes converted to lead sulfate. An actual reaction may not proceed to this ideal
completion if either electrode or the electrolyte becomes
deficient in (usable) active material.
For the electrode reactions to continue, the sulfate
ions must be able to diffuse into the active portions of
the porous electrodes. If the current drawn from the
battery is greater than the diffusion rate, the cell terminal voltage will drop. According to [SI,the cell voltage is
wheren0(JW2)
given by E = 1.87+.0291og
and 7~,(Pb+~)
are the number of divalent and tetravalent
lead ions available for reaction. When the current drawn
from the battery decreases, and sulfate diffusion supplies
additional sulfate ions for reaction, the cell voltage will
again increase. This diffusion process explains voltage
curves such as that shown in Fig. 4, where the terminal
voltage slowly increases after the terminal current goes
to zero.
Although batteries are energy storage devices,
batteries are rated by current capacity in AmpHrs, which after unit conversion is equivalent to
Coulombs. Theoretically, the current capacity is
Q, = min ( 2 * n,(Pb), 2 * n,(PbO2), no(H2S04)) where
n o ( X ) indicates the number of X ions in the fully
charged battery, which is nominally known by the system design. The actual integrated current that can be
removed from the battery will never exceed Qc. The
percentage of Qe that is achieved b. any application
is dependent on such factors as the load current and
operating temperature. This dependence on operating
conditions complicates the use of Qc for SOC determination, see Section 3.1.
Let E, denote the energy capacity of the battery and
E,.($)
denote the energy stored in the battery at time t.
Then,

(--)

Ec =

J_,

vO(T)id(T)dT 4- Er(t)-

The percentage remaining energy (PE,) is

Under the assumption that the battery terminal voltage

Figure 1: Resistive Thevenin equivalent circuit model


is constant, the tradition definition of state-of-charge
(SOC)

results, where the battery has been assumed to be fully


charged (SOC = 100%) at t = 0. Estimation of SOC
by means of eqn. (l),referred to as the Coulombic approach, is discussed Section 3.1.
2.2 Electrical Models [ll,121
Circuit models of varying complexity have been developed for the lead acid battery. A resistive Thevenin
model, see Fig. 1, is derived in [12] under the assumptions that (1) the electrodes are made of porous materials, (2) the electrolytic resistance is constant throughout
discharge, (3) the discharge occurs at constant current,
and (4) polarization is a linear function of the active
material current density. Based on empirical data, [12]
determines the following per cell parameters for a lead
acid battery:

v,= 2.0030 Volts,

Rd

= (-&

- 0.0150) R.

(2)
For the batteries used in the experiments in [12], Qc =
58.31 Amp-hrs. Based on the circuit in Figure 1, the
discharge loop equation is

& = vo - ibRd.

(3)

With Vo held constant at 2.003 V, the terminal voltage V, based on eqn. (3) is portrayed in Fig. 2 as a
function of integrated battery discharge current for various battery discharge currents. Note that, even with
this simple resistive Thevenin model, the amount of integrated current drawn from the battery before the cell
terminal voltage reaches any specified level is a function
of the rate of discharge.
Since [12] is concerned only with discharge, R, is not
defined therein. In a real battery, the open circuit steady
state voltage V, is dependent on SOC. This relationship
is approximately linear, which motivates open-circuit
SOC estimation methods when the open circuit voltage
is known. The resistance parameters Rd and R, model
all forms of energy loss (e.g., gassing), not necessarily
just electrical losses.

1645

verge exponentially toward V, with a time constant determined by RdC (when previously being discharged).
2.3 SOC versus Open Circuit Voltage
The steady-state open circuit voltage U,, is known to
be a function of temperature and approximately linearly
related to remaining SOC (see Fig. 3 in [3]). Based on
battery data, it is possible to estimate parameters a and
b such that

voc(t) = a S ( t ) b
-b
S(t) = Figure 2: Battery output voltage from Eqn. (2) versus
integrated battery current for battery discharge
currents of 0.1, 1, 5, 10, 25, and 50 amps.
The model of [12] is not dynamic, but derived based
on the assumption of a constant discharge current. The
non-dynamic nature of the model implies that the model
will not be able to account for such effects as the capacitive diffusion dynamics under the transient current conditions expected in (H)EV applications. Models suitable for (H)EV applications have been considered in
[4,9, 11, 131.
Two dynamic models are presented in [ll].The simpler of the two models is portrayed in Fig. 3. Based on
the circuit schematic of Fig. 3, the circuit equations for
charge and discharge are

(6)
(7)

where woc(t) is the open circuit battery pack voltage


and S(t) is the percentage remaining battery SOC. For
the battery pack being considered herein, a full battery
pack (S = 100%) corresponds to a voltage of 330 volts.
And, an empty battery pack (S = 0%) corresponds to a
voltage of 270 volts. Therefore,

a = .6- % SOC a n d b = 2 7 0 V .
Based on eqn. (7), there has been interest inusing
to calculate SOC. For example, during periods of
no load, the SOC can be calculated via eqn. . (7) and
the measured U,,. The difficulty in using eqn. (7) to
calculate the estimated SOC during vehicle operation is
in determining the battery open circuit voltage. The
new method of this article, which estimates U,, even
under loaded conditions is presented in Section 4.
U,,

3 SOC Estimation Methods

9.

where it, =
In the special case where Rb = 0,
V, = Vt and the battery current is determined by the
load. The nominal value of the polarization capacitance
C is 40 F [ll].The capacitance parameter also does not
necessarily represent purely electrical processes. For example, a portion of the effective capacitance is due to
the chemical diffusion within the battery. The resistance and capacitance parameters are dependent on the
temperature, SOC, and device design.
Based on the circuit model of Fig. 3, under no load
both of which conconditions (i.e., i b = o), & =

v,

This section reviews and analyzes the performance


of two existing SOC estimation methods. In a battery
stack, each individual cell can have a distinct SOC. Id+
ally, the SOC of each cell would be monitored and controlled individually; however, this is usually not practical or cost effective. The methods described in the following sections can be applied to either individual cells
or stacks of cells. For simplicity, we will not distinguish
between the two cases, but will generically use the term
batteries.
The introduction mentioned five categories of methods for estimating SOC. Electrolyte specific gravity and
stabilized float current methods were discussed in the
introduction and stated to be inappropriate for HEV
applications. The coulometric measurement and a new
loaded battery voltage estimation approach are discussed in the following.
3.1 Integrated Current

Figure 3: &active Battery Circuit Model ([ll, 131)

Theoretically, Qc is limited by the amount of active


material in the electrolyte and electrodes. Standard test

1646

procedures for experimentally determining Qc axe specified for example in [l].Knowledge of the specified current capacity, Qc, motivates the following simple SOC
estimation algorithm.
Define the integrated battery discharge current,
Qi(t) =

(9)

ib(T)dT.
0

The percentage remaining SOC is calculated as


SOC(t) = 100

Qc

nnm.mh

).

-QcQ i ( t )

where SOC = 100% means the battery pack is full.


Although this method is straightforward to implement, it has a few practical drawbacks. First, it requires that the initial SOC be known. Any error in
the initial value is reflected in all subsequent SOC values. Second, this algorithm assumes that the current
capacity is a known fixed quantity. More realistically,
the (usable) current capacity is time variant and dependent upon the current profile and temperature. Experience has shown that the rated current capacity is rarely
achieved in (H)EV applications [5].
Any problems experienced with this method in EV
applications will be compounded in HEV applications.
EV applications typically involve partial discharges followed by complete recharging. Recharging occurs in
a stationary mode, so that the reliable float current
method [l]of determining 100% SOC can be used. In
addition, a float current condition can be maintained for
sufficienttime to at least partially rebalance the battery
pack. In HEVs, several discharge/charge cycles may occur before an opportunity for stationary recharge and
balancing arises. During the on road discharge/charge
cycles, the energy management system will be scheduling APU usage (for recharging) based on the estimated
SOC. During discharge, the integrated current method
overestimates the remaining current capacity. During
charging, the integrated current method overestimates
the recovered current capacity, since some of the battery
charging current goes to processes such as electrolysis of
water instead of battery recharging. Therefore, over numerous discharge/charge cycles without resetting, an energy management system that relied on integrated current to determine SOC would have a growing disparity
between the calculated and actual battery SOC.
Fig. 4 presents an example of the use of the current
integration approach. The top graph depicts the calculated SOC based on the integrated current, assuming
that the initial SOC is 100%. The center graph shows
the ib(t). The lower graph shows V,(t). Note that although the battery pack performance is deteriorating
by about 17 minutes and performance is significantly
affected by 30 minutes, the calculated SOC is still high.
Even when the battery pack is dead at 60 minutes, the

Figure 4: Top - SOC estimate determined by current integration; Middle - Battery terminal current; Bottom - Battery terminal voltage.

calculated SOC is approximately 50%. These calculations used the manufacturers stated current capacity of
38 Amp-Hrs for Q c .
Note that the current integration approach is open
loop in the sense that the battery performance, as determined by V,, does not affect the estimated SOC. Also,
the approach is dependent on the initial value Q(0) and
Qc. Changing Q(0) shifts the entire SOC(t) curve vertically. The value of Qc is time varying and usage dependent. Therefore, it is evident that closed loop methods
that either estimate or adapt to uncertain initial conditions or parameters are required.

4 SOC & Parameter Estimation

This section presents an SOC estimation algorithm


that estimates the battery open circuit voltage, even
during loaded conditions, so that the open circuit voltage can be used to estimate SOC.
According to eqn. (4), the dynamic model for the
battery during discharge is

b is the measurement, I b is the


where v, is the state, v
input, and Rd, v,, and Rb are model parameters.
The objective of this section is to define an algorithm
to accurately estimate V, so that SOC can be estimated
via eqn. (7). To make the algorithm robust to different
batteries and aging effects, the battery parameters Rd,
Rb, and C will also be estimated. For the purpose of
estimator design, define the estimator state to be 2 1 =
v,, 22 = d c , 23 = $$, 5 4 = and 2 5 = Rb,where

1647

c,

+-

6,

9:m
..'..-I
m

the physical model parameters have been combined in


x2 and 5 3 to yield a simpler model representation. The
physical model parameters can be recovered from the
state estimates according to

ZBO

c=-

&=-

v"=G

1/p=x1

54

23

&=55.

For the estimator design, x i , i = 1 , . .., 5 are modeled


as first-order Markov colored noise processes [7]. Due
to these states combining several physical parameters,
there time evolution would be correlated. This correlation is neglected for simplicity of implementation. The
resulting model is written in state space as

+ Z 3 - I b X 4 + w1

$1

-2152

x2

a222+w2

x3

a353+w3

=
x5
=
& =
x4

a424 +w4

(13)
(14)
(15)
(16)

+ 205

(17)

Q555

where each wi and U are mutually uncorrelated white


noise processes.
The linearized state equation is

X(t) = F ( t ) x ( t )+ ~ ( t )
z ( t ) = H ( t ) x ( t ) v(t)

(19)
(20)

where F ( t ) =

I o

-xl(t)

-Ib(t)

o 1

a4
o

E{w(t)w(T)T} = Q(t)G(t- T )
E{v(ti)v(tj)T} = R(ti)Sij
E{w(t)v(ti)T} = 0.

"1
5

(21)
(22)
(23)

The solution to the state equation (19) is


= @(t,t o ) X ( t o )

1:

@(t,T)W(T)dT

(24)

where +(t,T) is the state transition matrix.


To obtain an equivalent discrete time model valid at
the measurement instants, consider t o = t k and t = t k + i
in (24), to obtain
X(tk+l)

~
I,min

= +(tk+l,tk)X(tk)

+J

lnln

Figure 5: Estimates using the Kalman filter approach. a)


Left - Estimated VO;b) Right - Estimated SOC.
which can also be written as

X(k

+ 1) = @ ( k ) x ( k )+ Wd(k)

(26)

where +(k) = @ ( t k + l , t k ) and Wd(k) is a white Gaussian


sequence that is statistically equivalent through its first
two moments to
L:b+'

*(tk+l,

T)W(T)dT.

(27)

Let Qd = E{wd(k)wz(k)}. By defining the sampling


period to be small relative t o the time constants of the
battery, it is valid t o assume that F(t) and Q(t) are approximately constant during the time interval [ t k , t k + l ] .
Therefore,

F(t) 2: F k ,

Q(t) E Q

for t E

[tk,tk+i].

Based on this assumption,

@ ( t ,=
~ )eFr(t--r)and *(k) = e F k T .

z ( t ) = &(t),and H(t) = [ l o 0 0 - I b ( t ) ] . A discrete time


estimator will be implemented with the following technical assumptions: E{w(t)} = 0 for all t , E{v(ti)} = 0
for all t i , and

X(t)

(18)

51 - I b x 5 + U

r -x2(t)

54

52

t&+l
Wk+l,T)W(T)dT

th

(25)

(28)

Based on similar assumptions, we approximate


Qd(k)

= *(k)&(k)aT(k)T.

(29)

Therefore, the discrete time state-variable model of the


battery used for the estimator design is

x(k

+ 1)

= @ ( k ) x ( k ) wd(k)

z(k) = H ( k ) x ( k ) v(k)

(30)
(31)

with the process noise covariance matrix defined by (29).


With this model, the standard EKF algorithm [7] is used
to implement the state estimator.
Results of applying this estimation based approach,
t o the voltage and current portrayed in Fig. 4, are shown
in Figs. (5 and 6). The sampling rate of & and l a is
6.7 H z . Fig. 5 show the estimated V, and the resulting
SOC. The calculated SOC accurately portrays the battery status after a short ( 5 minute) period during which
the initial battery parameter estimates are being corrected. Fig. 6 shows the estimated battery parameters
for runs from two Werent sets of initial battery parameters. The final conditions of the first run were used as
the initial conditions of the second run. The purpose

1648

discharge and charge process, knowledge of the functional dependence of these inefficiencies on the SOC
would allow the energy management system to identify
the most efficient range of SOC for battery operation.
If successful, this would result in more efficient overall
vehicle operation.

0.6

f 0.4
02

o:

30

20

U)

so

J
Bo

References

I-

[l] IEEE Recommended Practice for Maintenance, Testing, and Replacement of Large Lead Storage Batteries for
Generating Stations and Substations,IEEE Std. 450, 1987.
[2] J. Aldeu, H. Smimite, and C. Glaize, Improvement
of intelligent battery controller: state-of-chargeindicator and
associated functions, Power Sources 67, 1997, 157-161.

Figure 6: Estimated model parameters. a) Top - &; b)


Middle - I&,; c) Bottom - C.

of including two runs is to graphically show the convergence time for each parameter. The convergence time
i , the actual data (observis determined by Q , R, ~ and
ability). Since Rb directly affects the output when the
discharge current is not zero, Ra is quickly estimated.
The parameters C and Rd are less observable and hence
converge more slowly.

J. Aylor, A. Thieme, and B. Johnson, A Battery


[3]
State-of-Charge Indicator for Electric Wheelchairs, IEEE
Zkans. on IE, Vol. 39 (5), 1992, pp. 398-409.

C. Barbier and H. Meyer, A Battery State of Charge


[4]
Indicator for Electric Vehicles, Proc. of the Inst. of Mech.
Eng., 1994, pp. 29-34.
[5] B. Dickinson and D. Swan, EV Battery Pack Life:
Pack Degradation and Solutions, SAE paper 951949, pp.
145-154.
[6] K. Floyd, 2. Noworolski, and J. Noworolski, and W.
Sokolski, Assessment of Lead-acid Battery State of Charge
by Monitoring Float Charging Current, Proc. of 16th Int.
Tele. Energy Conf., 1994, pp. 602-608.

5 Conclusion

This article has reviewed the problem of battery SOC


estimation. The article has also presented a new algorithm and demonstrated its application on experimental
data. The experimental data of this article only involves
battery discharge. As additional data sets become available, future research will also address battery recharge.
The experimental results to date, as shown herein, indicate that the method is very promising. Advantages of
the new algorithm are (1) it accurately estimates SOC
without undesirable increases in SOC when the load
current changes, and (2) it provides estimates of the
battery internal parameters. These estimated parameters yield robustness across aging and different batteries. Additional advantages of the parameter estimation
and Vo pabased SOC algorithm are that the
rameters are available or battery monitoring and fault
detection purposes.
Note in Fig. 6 that all three battery parameters appear to be functions of the battery SOC. This is most
clearly shown in the plot of Rb, which rapidly converges
to the same value for the same SOC. This is also true
across data sets for the experiments that we have run.
This conclusion motivates future research that will attempt to learn these functions, and maintain the estimated functions accuracy as they change with battery
usage. Since Rd and Rb represent inefficiencies in the

&,a,

Gelb, Applied Optimal Estimation, MIT Press, Cam[7]


bridge, MA 1974.
[8]
F. Huet, A review of impedance measurements for
determinationof the state-of-charge or state-of-health of secondary batteries, Power Sources 70, 1998, 59-69.
[9]
M. Jayne and C. Morgan, Novel Voltage Models of
Lead Acid Battery Systems for Electric Vehicles, Polytechnic of Wales, U. K., October 1987.
[lo] D. Linden, Handbook of Batteries, McGraw Hill,
New York, 1995.
[ll] B. Powell and T. Pilutti, Series Hybrid Dynamic
Modeling and Control Law Synthesis, Ford Scientific Research TR SR-93-201, Dec. 1993.
[12] C. Shepherd, Design of Primary and Secondary
Cells, J. of the Electrochemical Soc. Vol. 112 (7), pp. 657664, 1965.

R. Sims, J. Carnes, M. Dzieciuch, and J. Fenton,


Computer Modeling of Automotive Lead Acid Batteries,
Ford Scientific Research TR SR-90-154, Sept. 1990.
[13]

[14] F. Vaccaro, J. Rhodes, and B. Le, The Effect of


Temperature on VRLA Reaction Rates and the Determination of Battery State of Charge: Part 2- Fundamental Considerations, Proc. of 19th Int. Tele. Energy Conf., 1997, pp.
230-237.

1649

You might also like