Академический Документы
Профессиональный Документы
Культура Документы
=
.
(3)
Acquisition of the battery resistance could let the
charging system determine the real battery voltage from the
terminal voltage in the simple battery model. In order to
overcome the voltage loss across the battery resistance, the
targeted apparent terminal voltage (V) should be set as the
actual battery voltage (V
BATO
), or said to be the OCV, plus
the voltage drop, i.e.,
BATO s CHRG
V R I V + =
(4)
Based on the identification of V
BATO
, the SOC could be
identified with the curve of SOC and OCV, hence the OCV
method.
According to Coulomb counting principles, it does not
contain any reference point and can suffer from long-term
drifting due to error introduced from system noise and
measurement noise. The value of system error can also be
introduced as the differentiating Peukert capacity and
Fig. 2. Schematic of the complex battery (top) and the complex battery
model (bottom).
Fig. 1. Terminal voltage during charging unit-step pulses showing the
desired model characteristics.
Peukert ideality coefficient as the batterys history
elongates. CC must be reinitialized with an accustomed
Peukert capacity and Peukert coefficient from a newly
acquired log-log chart from battery tests. Currently, there is
no correction for this within the CC simulation and it would
be difficult to find more recent values every few cycles.
Implementation of the KF requires a continuous time
observer to give a prediction to the filter itself for this
simulation, which is called a Kalman-Bucy filter. For
estimation of the true battery OCV, the observer will read
the current, terminal voltage, and the SOC from the CC-
OCV estimation in order to output a value. This is done
using an input of a stochastic state-space model, which
describes an estimation of the batterys parameters to
estimate the voltage across the batterys equivalent
capacitance and polarization capacitance. Solving the KCL
and KVL equations,
p
Cp
Cp
R
v
i i + =
(5)
Cp Cb t S
v v v iR =
,
(6)
respectively, gives system equations that result in the
necessary dynamic equations () = () + () + ()
and () = () + (). The dynamic system is
finalized in the form
) (
1
0
0
1
1
0
1 1
t s
i
v
C
C R
v
v
C R
C R C R
v
v
t
p
b S
Cp
Cb
p p
b S b S
Cp
Cb
+
(
(
(
(
(
+
(
(
(
(
(
=
(
&
&
(7)
) (
0
1
) ( t m
v
v
t y
Cp
Cb
+
(
=
,
(8)
where ) (t s and ) (t m represent system noise and
measurement noise varying in respect to time, respectively.
The observer estimates the rate of change with respect to
time and solves for the state variables,
Co
v and
Cp
v . Then,
the observer calculates the gain of the system in order to
achieve the desired output.
The Kalman filter is used to filter out noise,
measurements, and other inaccuracies resulting in values
closer to real values in dynamic linear systems. Upon
reception of the input values, the Kalman filter estimates the
uncertainty of the values and computes a weighted average
between the measured value and the predicted value. It
compiles all noise influences, even those unconsidered
within this paper, such as thermal differences. The
uncertainty is qualified by using the best linear depiction of
the input value, which allows for cancellation of noise by
narrowing the sensor and noise values to a minimum. The
weighted average is calculated placing heavier weights on
those values that are more likely, which is determined by the
covariance on the uncertainty. The filter then feeds back
and recursively uses prior prediction to determine the new
best guess at each time step. This functionality gives it the
title of an adaptive filter for digital signal processing.
Then, the KF uses a linear-quadratic regular in
combination with a Gaussian controller to form a linear-
quadratic-Gaussian controller, or LQG, to estimate the state-
space variables within our linear dynamic system. The cost
function is used to share the control input history and is
stated as
[ ]
)
`
+ + =
T
dt t u t R t u t x t Q t x T Fx T x E J
0
) ( ) ( ) ( ' ) ( ) ( ) ( ' ) ( ) ( '
,
(9)
where F, Q, and R are functions of time and are greater than
0 given t t t y < ' 0 ), ' ( and T is the final simulation time.
The LQG controller calculates
[ ]
[ ]
) ( ) (
) 0 ( ) 0 (
) ( ) ( ) ( ) ( ) ( ) (
t x L t u
x E x
t x t C t y K t Bu t x A t x
=
=
+ + =
&
,
(10)
where K is the Kalman gain, L is the feedback gain, and E is
the average value, or the expectation value.
The Kalman gain is described as
( ) ( )
) ( ' ) ( ) (
) 0 ( ' ) 0 ( ), ( ), ( , , ) (
1
t W C t P t K
x x E w W v V C A f t K
=
=
,
(11)
and to find P(t)
( ) ) 0 ( ' ) 0 ( ) 0 (
) ( ) ( ) ( ' ) ( ' ) ( ) ( ) (
1
x x E P
t V t CP t W C t P A t P t AP t P
=
+ + =
&
.
(12)
The feedback gain is described as
( )
) ( ) ( ' ) ( ) (
), ( ), ( , , ) (
1
t S t B t R t L
F t R t Q B A f t L
=
=
,
(13)
and to find S(t)
F T S
t Q t S t B t BR t S A t S t S A t S
=
+ + =
) (
) ( ) ( ) ( ' ) ( ) ( ) ( ) ( ' ) (
1
&
.
(14)
The dynamic matrix equations for P(t) and S(t) are defined
as Riccati differential equations. Individually, the first
matrix Riccati differential equation solves the linear-
quadratic estimation, whereas the second matrix Riccati
differential equation solves the linear-quadratic regulator.
Appending these calculations to the stochastic state-space
model allows for the LQG control algorithm to be solved,
resulting in a data reduced SOC output [8].
Referring to the previous section, CC, the batterys
Peukert capacity and Peukert ideality coefficient would now
be considered part of the system error. Thus, the Kalman
filter adjusts for the constantly changing Peukert values and
noise given from the sensors and system.
IV. RESULTS AND DISCUSSION
The pulse charging method made the internal Ohmic
impedance available to determine. It is concluded that Li-
ion battery internal Ohmic resistances are not constant
through all potential ranges, but are relatively within
tolerance range as can be seen in Fig. 3 and Fig. 4. A failed
battery is easily noticeable from Ohmic resistance testing
shown in Fig. 5. However, it is very likely that every
manufacturer holds different resistive properties depending
on battery architecture and materials, and these differences
must be taken into account.
The pulse charging method also determines the OCV vs.
SOC and compares it to the manufacturers specifications,
shown in Fig. 3. The test differentiation is caused by
deterioration of the Peukert capacity.
The results determine a difference between the
manufacturers and the two trial runs mostly likely due to
the Peukert capacity and ideality coefficient altering over
the batterys short history or differences in the
manufacturing process. The internal Ohmic resistance
remains below 15 until the SOC reaches the overcharge
regions, where the impedance increases drastically to
infinite. It is concluded that the Li-ion SOC performed
properly with a linear depiction with respect to time.
The CC simulation was implemented to accumulate the
SOC measured based on Peukerts ideality equations. The
automated simulation was developed in National
Instruments Labview 2009 SP using a National Instruments
myDAQ, a portable data acquisition unit.
The simulation, shown in accurately models long-term SOC
with introduced error, but is compensated for in the KF
algorithm.
The KF was implemented to correct the CC SOCs error
and result in a more true value with consideration to noise
in the system. This is done using the algorithm stated
previously. Several tests were done to observe the influence
of induced scenarios or errors on the OCV-CC-KF system.
The first test was done through Labview to implement a
random signal generation in the readings to simulate large
noise measurements in an automotive application. The test
was performed in both charging and discharging scenarios.
The result was a precise but less accurate system by a
magnitude of approximately 0.14% weighted mean.
The second test was done through automated pulse
charging and data acquisition to determine the difference of
SOC values. The KF SOC was measured during the charging
moments and the OCV was measured during the idle
moments given an 8 second rest prior to reading. This test
was performed in both charging and discharging scenarios.
The results of this test were in agreement with the research
done by Smith, Rahn, and Wang on pulse charging of
lithium-ion batteries [8]. According to the data found, the
KF SOC estimation had an error / tolerance of 1.76% in
comparison to the OCV method estimation, which was
calculated to be statistically insignificant.
The third test was done through Labview to implement a
constant DC offset on the KF inputs to determine the
correction abilities for large CC estimation error. The DC
offset implemented to each of the KF inputs results in a
percent error as shown in Table 1.
Fig. 6. Labview schematics for CC SOC estimation for cycling.
Fig. 5. Internal connection resistance testing at various currents of a bad
battery (experimental data using OCV).
Fig. 4. Internal connection resistance testing at various currents of a good
battery (experimental data using OCV).
Fig. 3. SOC CC-OCV estimation method verification of Peukerts ideality
principles.
Table 1 shows the dependency of the KF on specific inputs
and these values are well understood to be results of the
weighting average principles behind the functions.
The Labview 2009 code was proven to be useful during
this process. Fig. 7 and Fig. 8 show estimation of the SOC
from OCV, OCV-CC, and OCV-CC-KF accumulated
methods at a low SOC and a high SOC using the original
prototype algorithm. Fig. 9 is the newest prototype using
Labview for pulse charging and discharging. This
representation allows us to easily view the battery SOC
during online testing and verification with OCV SOC
detection.
V. CONCLUSION
This paper shows the implementation and results of
combining open circuit voltage, Coulomb counting, and
Kalman filter methods in order to more accurately estimate
the SOC of Li-ion battery cells by various factors into
consideration. The combination of the methods accurately
estimates the SOC with an error of 1.76% in comparison to
the OCV method estimation. This error may vary
depending on hardware used for the data acquisition. This
set up was done with a National Instruments myDAQ. This
error tolerance was calculated to be statistically insignificant
and therefore usable. The KF observes the terminal voltage
rate of change as well as discharging or charging currents to
calculate electrode surface dynamics, electro-chemical
reactions, and electrode particle transfer, along with other
side reactions. In addition to the rate of change, it also
receives the output of SOC from the CC method to edit
the needed values for the filtered SOC output. The tests in
this paper used the single polarization impedance model
shown at the beginning of this paper in Fig. 2, called the
complex battery model assembled within this study.
VI. RECOMMENDATIONS
It is recommended to enhance the KF processing
capabilities by incorporating thermal differentiation and
capacity differentiation. This is an essential part of our goal
to produce an overall battery management system for
PHEVs. This technology can be additionally supportive
within small electronic devices, which use an ESS.
REFERENCES
[1] Li, J. Murphy, E. Winnick, J. Kohl, P.A. J. Power Source 102 (2001)
302-309.
[2] H. L. Martin and R. E. Goodson, New concept in battery
performance simulation and monitoring, in EV EXPO 80, Cervantes
Convention Center, St. Louis, Missouri, May 1980.
[3] B. Thomas, W. B. Gu, J. Anstrom, C. Y. Wang, and D. A. Streit, In-
vehicle testing and computer modeling of electric vehicle batteries,
in Proc. 17th Int. Electric Vehicle Symp., Montreal, Oct. 2000,
[Online]. Available: http://mtrl1.me.psu.edu.
[4] V. S. Bagotzky and A. M. Skundin, Chemical Power Sources:
Academic Press, 1980.
[5] NREL, ADVISOR HEV Simulation Model, 1996.
[6] X. He and J.W. Hodgson, Modeling and Simulation for Hybrid
Electric Vehicles-Part I: Modeling, IEEE Trans. Intelligent
TransportationSystems, 3.4 December 2002.
[7] C. J. Hoff, Hybrid Vehicle Propulsion, Kettering University 2010.
[8] K. A. Smith, C. D. Rahn, and C. Y. Wang, Model-based
electrochemical estimation and constraint management for pulse
operation of lithium-ion batteries: IEEE Trans. Control Systems
Technoloy. 18.3, May 2010.
Fig. 9. Second prototype code using pulse charging and discharging.
Fig. 8. Pulse charging at stage 2 constant voltage for SOC accuracy
estimation. Note that charging current is negative.
Fig. 7. Pulse charging at stage 1 constant current for SOC accuracy
estimation. Note that charging current is negative.
TABLE 1
SOC ERROR FROM FORCED DC OFFSETS IN INPUTS OF CODE.