Accurate Thevenin’s Circuit-based Battery Model

Parameter Identificaiton
Cong-Sheng Huang (Student Member) Mo-Yuen Chow (IEEE Fellow)
Electrical and Computer Engineering Electrical and Computer Engineering
North Carolina State University North Carolina State University
Raleigh, NC, U.S.A. Raleigh, NC, U.S.A.
chuang15@ncsu.edu chow@ncsu.edu

Abstract—Batteries are becoming the main energy storage
devices in following decades. Highly accurate battery models are
needed to fulfill accurate controlling and monitoring purpose. For
real-time operations, the calculation complexity of the battery
model is an issue. Thevenin’s circuit model is a tradeoff between
fast calculation and accuracy. The model uses resistor and RC
pairs to capture the dynamic of terminal voltage. Even though
there are few components in the model, to accurately identify those
components value is a difficult question. In this paper, an
algorithm to separate effective data and ineffective data is
proposed by comparing the condition number. This algorithm can
help identify correct parameters for future model-based battery
monitoring and controlling.
Keywords—Thevenin’s circuit model; parameter identification;
linear regression; condition number
I. INTRODUCTION
Batteries are becoming the main energy storage devices in
the following decades. The battery technologies along with their
monitoring and management technologies are growing rapidly.
Novel chemical compounds have been developed to build higher
power and energy density batteries. From application
perspectives, it is critical to develop intelligent algorithms to
estimate the state of charge (SOC), state of health (SOH), and
state of function (SOF) of batteries with high accuracy,
robustness for real-time applications. Many battery models have
been developed for different purposes. To name a few:
electrochemical model [1], analytical model [2], electric circuit
model [3]–[5], etc. Electrochemical model is well-known for its
accuracy because it describes the chemical reaction inside the
battery during charging/discharging. However, the model
requires a full knowledge of the specific battery. This model is
difficult to be used for real-time applications due to its
computation complexity. Analytical model, which is also known
as Kinetic battery model or KiBam model, use the Kinetic
process to model the chemical process of large lead-acid
batteries. However, the analytical models adopt flat discharge
profiles, which suits only large lead-acid batteries, even though
This paper focuses on the electric circuit model because we
are mainly interested in real-time battery monitoring and
management. This model is more suitable to be used for
estimating SOC, SOH, and SOF in real-time operations with a
good tradeoff between computation complexity and accuracy.
Many algorithms have been developed for battery SOC, SOH,
SOF estimation. To name a few: Columb counting [6], observer
[7], Kalman filter[8], Fuzzy logic [9]. Most of these algorithms
are model-based. Model-based approaches always face a serious
problem: how to acquire the accurate parameters to build a
correct model.
This paper proposes the use of condition number to identify
effective on-line battery measurements in order to acquire
accurate parameters of the electric circuit model, which in turns
can significantly improve the accuracy of model-based SOC,
SOH, SOF estimations. In this paper, Section II describes the
background information of battery system modeling and system
identification, Section III describes the definition of battery
discharge cycle, the parameter profiles and the region
separation. Section IV discusses the identification result and
compare the relationship with condition number. Finally, the
conclusion is stated in Section V.
II. BATTERY MODELING AND PARAMETER IDENTIFICATION
A. Thevenin’s Circuit Model
A typical battery Thevenin’s circuit model includes one dc
resistance, R0, and one (R1C1) [3], two (R1C1, R2C2)[5], to
multiple pairs to describe the dynamic of the terminal
voltage, as shown in Fig. 1 [4]. More RC pairs capture more
terminal voltage dynamics, while increases the computation
complexity. The more pairs used will increase more states to
be used in the corresponding state-space model, and increase the
order of the corresponding transfer function.
VRC2 ...
VRC1 VRCn iL
...
R0 R1 R2
...
Rn
VQ =
iL QR
SOC
Voc = f(SoC)
C1 C2 Cn VT

they capture the recovery effect and rate capacity effect. Electric
circuit model, also called Thevenin’s circuit model, is
commonly adopted to describe battery characteristic such as
transient responses of the battery. Fig. 1. THEVENIN’S CIRCUIT MODEL.

This paper uses one pair model to illustrate the condition

Acknowledgement: This work is partially supported by the National number based measurement selection for battery parameter
Science Foundation (NSF) under Award Number IIP-1500208 identification. One pair model captures the terminal voltage

978-1-5090-0873-5/16/$31.00 ©2016 IEEE 274

dynamic in terms of seconds to minutes, and is suitable to be . (7)
used for SOC estimation on electric vehicle (EV), hybrid Choosing as state, the state-space model of the
plugged-in electric vehicle (HPEV) real-time operations. circuit can be represented in standard Linear Time Invariant
form , as:
A typical open-circuit voltage (VOC) and SOC relation,
Voc(SOC), of a battery is shown in Fig. 2. This nonlinearity is a 0 0 ⁄
challenge if we want to use linear system theory for battery , (8)
0 1⁄ 1⁄
modeling, parameter identifications, state estimation, and
control. In this paper, we approximate the nonlinear function 1 1 , 9
VOC(SOC) with a set of piece-wise linear functions over its SOC where 0, 0; 0, , , , 1,1 , .
operating range [10]. The SOC-VOC curve can be separated into
eight linear segments as Fig. 3 by the second derivative of Voc For each i-th linearized region, the continuous-time transfer
versus SOC map. function can be derived as [11], note we skip the subscript i for
simplicity purpose in the following, all the derivation are for each i-th
region:

10
.
SOC1 SOC2 SOC3 SOC4 SOC5 SOC6 SOC7 We will use linear regression analysis in discrete-time
Fig. 2. SOC - VOC MAP.
domain to perform parameter identification of the battery model.
The continuous-time transfer function is transformed into
discrete-time domain with bilinear transform (s → ), where
is the sampling time. In this paper, we set T = 1 second.

. 11
Once the coefficients (c0, c1, c2, a1, a2) of the discrete-time
transfer function is identified, the bilinear transfer function is
used to transfer the discrete-time transfer function back to
continuous-time domain. Hence, the value of the components in
the electric circuit model can be acquired from the continuous-
time transfer function.
B. Linear Regression Analysis
Linear regression analysis [12] is used to identify
Fig. 3. (A) FIRST AND (B)SECOND DERIVATIVE OF VOC VERSUS SOC MAP.
coefficients of the transfer function. The linear system in eqn.
The VOC(SOC) function is separated by a set of linearized (11) can be rewritten in an autoregressive-moving average,
function as: ARMA, model form [11]:
1 ⋯
(12)
, , , 1 ⋯ ,
, , , Or 1 ⋯
, (1) (13)
… 1 ⋯ ,
, , , We can also represent the output y(k) as the multiplication of
⋯ coefficient matrix ( ) and regression matrix ( ) as:
where , and , are the linear equation coefficients to be
k , (14)
identified. is related to load current IL and the battery
with:
capacity QR as [10]:
1 2 ⋯
∗ . (2) (15)
1 ⋯ .
Thus, the dynamics of the SoC and VOC can be described by
⋯ ⋯
the following two first order differential equations: (16)
.
, (3)
C. Condition Number
. (4) The linear regression matrix is widely used to find the
Based on the voltage Kirkoff Law, the voltage between the coefficients of relevant transfer function [13]. However, the
pair, VRC, and the terminal voltage, VT, can be represented as: ability to find accurate coefficient values can depend on many
factors such as whether the data set used contains sufficient
, (5)
information for the identification purposes.
, (6)

275
 In this paper, we propose to use the condition number as an We separate the charging/discharging cycle into four
indicator on what type of data should be used to perform the regions, D1, D2, D3, and D4 as shown in Fig. 5 by four time
parameter identification of a battery model. The condition stamps k1, k2, k3, and k4.
number [13], , presenting the accuracy of computing J-1 with
Fig. 5 shows more details for the charging/discharging cycle,
given J, is defined as the product of two matrices norms, ‖ ‖ and
‖ ‖ as: the battery is discharged with a constant current load. 1A for 378
seconds in period D1, which discharge the battery to 5% of its
‖ ‖∙‖ ‖. (17) total capacity. After discharging the battery, rest the battery for
When is small, J is well-conditioned, and J-1 can be at least 30 mins as suggested in [14], and is denoted D2. Then,
computed with high accuracy. Conversely, J with large is one discharge cycle is 50 minutes.
called ill-conditioned. These kind of matrices usually are close
to singular and give large computational error for inverse D1: Nonlinear Discharge Region
calculation. ∀ ∈ , 3 . (“-3” is related to the
transfer function order.)
III. REGIONS SEPERATION ALGORITHM VRC iL
To build an accurate model for model-based algorithm, how
R0 R1
to acquire correct parameters values is an important issue. An
important key is to efficiently distinguish effective data and iL QR VQ = SoC Voc = f(SoC)
C1 VT
ineffective data for the battery parameter identification. We
propose to separate the charging/discharging time responses into
four different time regions based on the nature of the
charging/discharging cycles, then use the condition number of Fig. 6. THEVENIN’S EQUIVALENT CIRCUIT MODEL IN D1.
the data set collected in each time region to identify effective
data sets for battery parameter identification. The data belongs to this region are defined as effective data
and can use linear approximation to acquire the parameters
For illustration purposes, we discharge a 2.1Ah battery from because the condition number is low in . The condition
fully charged to empty in 20 cycles. The battery discharges 5% number is listed in TABLE I. In this region, the battery is
of its SOC in every cycle. Fig. 4 shows SOC, load current IL, connected to a constant current load, completes a loop to
and terminal voltage VT of one cycle. discharge. consumes most of the energy, makes the Ohmic
drop at the begin of discharge. The pair provides the
k1 k2 k3 k4 k1
transient of the terminal voltage profile because the of the
pair is charging. All the electric components (resistor and RC
pair) contribute to the terminal voltage profile, therefore, all the
components can be identified in this region. Note that the
identification should include at least three past effective data to
capture the Ohmic drop on terminal voltage profile, the effect of
, due to the second order we chose for linear regression
approach.
D2: Linear Discharge Region
∀ ∈ , .
VRC iL
Fig. 4. (A)THE SOC PROFILE IN A CYCLE, (B) THE LOAD CURRENT PROFILE IN
A CYCLE, (C) THE LOAD VOLTAGE PROFILE IN A CYCLE. R0 R1
iL QR VQ = SoC Voc = f(SoC) VT
k1 k2 k3 k4 C1

Fig. 7. THEVENIN’S EQUIVALENT CIRCUIT MODEL IN D2.

In this region, the linear regression approach fails to identify

parameters and the condition number is very large. The
condition number in D2 can be found in TABLE I. In this region,
the of the pair is fully charged. The reactance of this
capacitance is infinity. In other words, the connection to the
of the pair is opened. Only and affect the terminal
voltage profile, and strictly follows the Ohm’s law. Therefore,
D4 D1 D2 D3 D4
the model of this region is a linear transfer function.
. (18)
Fig. 5. (A)THE SOC PROFILE IN DETAIL, (B) THE LOAD CURRENT PROFILE IN
DETAIL, (C) THE LOAD VOLTAGE PROFILE IN DETAIL. where . Note that in this case cannot be
decomposed into and .

276
D3: Nonlinear Relaxation Region
∀ ∈ , 3 .
The linear characteristic of terminal voltage profile ends
once the battery disconnects with load. All the parameters of this
region can be identified again and the condition number is small.
The condition number can be found in TABLE I. Once the
circuit is open, the battery start to relax. of the pair starts
to discharge, provides an excitation to terminal voltage [12].
different regions. The effect of nonlinear parameter vs. SOC in
parameter identification is beyond the scope of this paper. In this
paper, the parameters (R0, R1, C1, and R1C1 constant) are
assumed to remain constant, zero order hold (ZOH), in each
discharge cycle as shown in the red lines in Fig. 10 to Fig. 13.
The ZOH error respect to the real value is smaller than 6%,
which is small enough to be adopted in this paper. The nonlinear
parameter identification will be discussed in another paper.

R0 R1
QR VQ = SoC Voc = f(SoC) VT
C1

Fig. 8. THEVENIN’S EQUIVALENT CIRCUIT MODEL IN D3.

D4: Deep Relaxation Region Fig. 10. THE R0 VS. SOC CURVE.
∀ ∈ , .
VRC

R0 R1
QR VQ = SoC Voc = f(SoC) VT
C1

Fig. 9. THEVENIN’S EQUIVALENT CIRCUIT MODEL IN D4.

Fig. 11. THE R1 VS. SOC CURVE.
Once of the pair is fully discharged, the battery goes
to region . This region also called deep relaxation period. All The resistor of pair, , profile is shown in Fig. 11. The
the electric components cannot be identified because the lack profile is similar with the profile of because they both are
of excitation signal and the condition number is infinity in the resistance of the Thevinen’s circuit model.
TABLE I. There is no chemical reaction inside the battery until
next charge/discharge operation. Therefore, remains
constant without any change. Terminal voltage, which is same
as , also remains as constant.
IV. SIMULATION AND RESULTS
In order to validate the proposed condition number based
parameter identification under a fully controlled environment,
we developed a circuit model based battery simulator using
Simulink to represent the actual battery. The battery parameter
values are acquired from [10]. Fig. 12. THE C1 VS. SOC CURVE.

The resistor, , profile is shown in Fig. 10. R0 is

large when the battery is nearly fully discharged (low SOC)
because there not enough ions in the battery to draw larger
current from the battery. R0 decreases as the amount of ions
increases inside the battery when SOC increases, this is why R0
drops in the mid SOC region. When SOC continue to increase
to a point that the battery is full with ions and make themselves
(thus current) hard to flow, thus shows the slightly higher
resistance with large SOC.
The blue lines in Fig. 10 - Fig. 13 show the nonlinear Fig. 13. THE R1C1 VS. SOC CURVE.
relationships between R0, R1, C1, R1C1 and SOC, respectively. The capacitance of pair, , profile is shown in Fig. 12.
Together with these nonlinearity, the battery parameter Theoretically, the RC time constant of the battery remains
identification will become a complicated process. However, as constant or linear with different SOC, which is shown in Fig. 13.
mentioned in the introduction, the emphasis of this paper is in Thus, the profile will be lower when the battery is fully
the separation of the effective and ineffective data by using

277
charged and fully discharged, makes the product of the pair applied in due to the condition number is infinity and matrix
remains linear. inverse to find the parameter values is undefined.

TABLE I. Condition Number in Different Region and Different SOC.

Region
SOC

0.00 50958 34554716173194300000 49695 Inf.

0.05 52197 42048768976985700000 50944 Inf.
0.10 53197 6895897406297550000 51815 Inf.
0.15 53684 8055074687891220000 52252 Inf.
0.20 53750 3645740863232120000 52230 Inf.
0.25 53473 34389776576241700000 52013 Inf.
0.30 53114 14659978942349200000 51716 Inf.
0.35 52812 44026105942697600000 51424 Inf. Fig. 15. ABSOLUTE VALUE OF RELATIVE ERROR OF R0.
0.40 52552 30426597540581600000 51186 Inf.
0.45 52216 134266874142775000000 50895 Inf. The identification result of are both accurate in , .
0.50 51523 152839054796338000000 50300 Inf. The absolute value of relative error is almost zero except the
0.55 50373 855733757024051000000 49237 Inf. result at some SOC in . However, the error is still lower than
0.60 48747 129069785643046000000 47514 Inf.
0.5%. Alternatively, the absolute value of relative error in is
0.65 46659 24239365849927000000 45430 Inf.
0.70 44401 36488609763394800000 43089 Inf. very inaccurate, the value is about 15000%.
0.75 42056 52100298898659000000 40644 Inf.
0.80 39669 3921036994977210000 38294 Inf.
0.85 37265 37661357509331800000 36018 Inf.
0.90 34832 3366299089400070000 33488 Inf.
0.95 32118 1744889459158210000 29977 Inf.

Data

iL≠0 No

Yes
Fig. 16. ABSOLUTE VALUE OF RELATIVE ERROR OF R1.
dVt/dt ≥ dVt/dt <
No No
threshold threshold
The results of shows almost the same as the results of
Yes Yes and . The absolute relative error value is almost zero at
Effective Ineffective Effective Ineffective
different SOC in , . Even though there still have some
data (D1) data (D2) data (D3) data (D4) results not so accurate, the error is about 0.2%, which is very
small. However, the absolute relative error value is about 99%
All All
parameters parameters Cannot
in .
Only Rall can
(R0, R1, C1) (R0, R1, C1) identify any
be identified
can be can be parameter
identified identified

Fig. 14. FLOW CHART OF SEPARATING EFFECTIVE DATA AND INEFFECTIVE

DATA.

The condition number at different regions ( , , , )

and different SOC are listed in TABLE I. The condition number
in and is substantially smaller than the condition number
and . Also, the linear regression can accurately identify the
correct parameters in and but not in and due to the
set of data includes enough information or not. Therefore, the Fig. 17. ABSOLUTE VALUE OF RELATIVE ERROR OF C1.
set of data in and are defined as effective data, the set of The relationships between the identification performance,
data in and are defined as ineffective data. The flow chart absolute value of relative error, and condition number of R0, R1,
to separate effective data and ineffective data can be C1 are plotted in Fig. 18.
summarized as a flow chart as Fig. 14. Note that the threshold
of ⁄ is set as a very small value (10 ⁄ ). There are two groups in the relationship of R0, shows the
positive relationship between performance and condition
The identification result of is plotted in Fig. 15. The x- number. The group presents the relationship of and has
axis presents different SOC value. In and , the absolute low condition number and good performance. Another group
values of relative error of at different SOC are lower than presents the relationship of , alternatively, has high condition
0.2%. However, the absolute relative values error in are
about 49%. Note that the parameter identification cannot be

278
 number and poor performance. In other words, the identification REFERENCES
is accurate at low condition number condition.
[1] A. Seaman, T.-S. Dao, and J. McPhee, “A survey of mathematics-
based equivalent-circuit and electrochemical battery models for
hybrid and electric vehicle simulation,” J. Power Sources, vol. 256,
pp. 410–423, 2014.
D2
[2] J. Li, Q. Fang, H. Wu, Y. Liu, and P. Wen, “Investigation into
D1 and D3 diffusion induced plastic deformation behavior in hollow lithium
ion battery electrode revealed by analytical model and atomistic
simulation,” Electrochim. Acta, vol. 178, pp. 597–607, 2015.
[3] W. Waag, C. Fleischer, and D. U. Sauer, “Critical review of the
methods for monitoring of lithium-ion batteries in electric and
Fig. 18. PERFORMANCE V.S. CONDITION NUMBER OF R0. hybrid vehicles,” J. Power Sources, vol. 258, pp. 321–339, 2014.

Therefore, the accuracy of identification is highly related to [4] H. Chaoui, N. Golbon, I. Hmouz, and R. Souissi, “Lyapunov-Based
condition number. The identification is very accurate when Adaptive State of Charge and State of Health Estimation for
condition number is low, which is and . However, the Lithium-Ion Batteries,” IEEE Trans. Ind. Electron., vol. 62, no. 3,
identification fails in , the region with very high condition pp. 1610–1618, 2015.
number. [5] L. Pei, T. Wang, R. Lu, and C. Zhu, “Development of a voltage
Though and are both defined as ineffective data set relaxation model for rapid open-circuit voltage prediction in
because high condition number, some component values can lithium-ion batteries,” J. Power Sources, vol. 253, pp. 412–418,
still be identified in , where the battery follows Ohm’s. 2014.
Hence, can be identified correct in , but cannot [6] C. Y. Chun, J. Baek, G.-S. Seo, B. H. Cho, J. Kim, I. K. Chang, and
further separate into and . The identification result shows S. Lee, “Current sensor-less state-of-charge estimation algorithm for
that the absolute value of relative error in is about 0.06%. lithium-ion batteries utilizing filtered terminal voltage,” J. Power
However, the absolute value is about 4% in because the
Sources, vol. 273, no. 0, pp. 255–263, 2015.
reactance of of pair affects the linearity of Ohm’s law.
Note that the Ohm’s law fails in and bacause the load [7] M. a. Roscher, O. S. Bohlen, and D. U. Sauer, “Reliable state
current equals to zero. estimation of multicell Lithium-ion battery systems,” IEEE Trans.
Energy Convers., vol. 26, no. 3, pp. 737–743, 2011.
Inf. [8] S. Sepasi, R. Ghorbani, and B. Y. Liaw, “A novel on-board state-of-
charge estimation method for aged Li-ion batteries based on model
adaptive extended Kalman filter,” J. Power Sources, vol. 245, pp.
337–344, 2014.
[9] L. Kang, X. Zhao, and J. Ma, “A new neural network model for the
state-of-charge estimation in the battery degradation process,” Appl.
Energy, vol. 121, pp. 20–27, 2014.
[10] H. Rahimi-Eichi, F. Baronti, and M. Y. Chow, “Online adaptive
parameter identification and state-of-charge coestimation for
Fig. 19. ABSOLUTE VALUE OF RELATIVE ERROR OF RALL. lithium-polymer battery cells,” IEEE Trans. Ind. Electron., vol. 61,
V. CONCLUSION no. 4, pp. 2053–2061, 2014.
[11] R. Isermann, Fault-diagnosis systems: an introduction from fault
The condition number can be used as an index to
successfully evaluated and separate effective data and detection to fault tolerance. Springer Science & Business Media,
ineffective data, which helps to derive accurate parameter values 2006.
and help improve the accuracy of model-based battery related [12] K. J. Åström and B. Wittenmark, Adaptive control. Courier
estimation algorithm. The condition number is relatively small Corporation, 2013.
in and compare to . All the parameters, and and [13] P. L. Bonate and J.-L. Steimer, Pharmacokinetic-pharmacodynamic
, can be correctly identified through linear regression approach
modeling and simulation. Springer, 2011.
in and , which are defined as effective data. However,
only , which is the summation of all the resistance in the [14] M. A. Roscher and D. U. Sauer, “Dynamic electric behavior and
circuit but cannot be further separate into and , can be open-circuit-voltage modeling of LiFePO4-based lithium ion
identified in the linear region . In addition, all the parameters, secondary batteries,” J. Power Sources, vol. 196, no. 1, pp. 331–
and and , cannot be identified in . Therefore, the 336, 2011.
effective data can be found through small condition number.
Effective data can help identify accurate components in the
Thevenin’s equivalent circuit model.

279

Powered by TCPDF (www.tcpdf.org)