You are on page 1of 24

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271101290

Parameter Sensitivity Analysis of Cylindrical LiFePO4 Battery Performance


Using Multi-Physics Modeling

Article  in  Journal of The Electrochemical Society · March 2014


DOI: 10.1149/2.048405jes

CITATIONS READS

68 5,361

7 authors, including:

Liqiang Zhang Chao Lyu


Ocean University of China Harbin Institute of Technology
22 PUBLICATIONS   459 CITATIONS    53 PUBLICATIONS   803 CITATIONS   

SEE PROFILE SEE PROFILE

Gareth Hinds Ma Kehua


National Physical Laboratory shinergy
148 PUBLICATIONS   3,234 CITATIONS    5 PUBLICATIONS   106 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

HiPoBat (EPSRC) View project

Battery modeling and control View project

All content following this page was uploaded by Gareth Hinds on 22 July 2015.

The user has requested enhancement of the downloaded file.


Parameter sensitivity analysis of cylindrical LiFePO4 battery performance
using multi-physics modelling
Liqiang Zhang1, Chao Lyu1*, Gareth Hinds2, Lixin Wang1, Weilin Luo1, Jun Zheng1, Kehua Ma1
1
School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin 150001, China
2
National Physical Laboratory, Teddington, Middlesex, TW11 0LW, United Kingdom

Highlights
- Coupled thermal-electrochemical model of cylindrical lithium ion battery performance
- Parameter sensitivity matrix established using clustering theory
- Stepwise experiments for parameter identification designed using sensitivity analysis results

Abstract
A multi-physics model for a cylindrical Li-ion battery has been developed by coupling a model of the thermal
distribution in the radial direction to an electrochemical P2D model. The model can predict both terminal voltage and surface
temperature, which has the advantage that it can be readily validated by measurement. A sensitivity analysis of up to 30
parameters was carried out using model simulation. A parameter sensitivity matrix was established to describe the parameter
sensitivity under different operating conditions and clustering theory used to group the parameters according to their average
sensitivity. Finally, a stepwise experiment was designed to validate the results of the sensitivity analysis. It was shown that
the stepwise approach to parameter identification results in significantly higher accuracy.

Keywords:
Multi-physics model
Parameter sensitivity
Parameter identification
Li-ion battery

1. Introduction
In recent years, Li-ion batteries have been widely used in electric vehicles due to their relatively high energy and power
density. In such demanding applications, Prognostics and Health Management (PHM) of Li-ion batteries is a key issue and
has attracted increasing interest from researchers. The aim of our research is to build a PHM database for Li-ion batteries by
determining which parameters will change and how they change under different aging patterns. This database could be used
for characterisation of State-of-Health (SOH) and estimation of remnant lifetime. The core work is to develop a numerical
model that can predict battery behavior and identify the critical model parameters accurately and nondestructively using
simple and rational experiments.
The Newman P2D model is a type of first principle model commonly used in the analysis of Li-ion battery performance
[1-3]. It addresses the complex interaction of physicochemical processes such as diffusion, ion migration, ohmic phenomena
and electrochemical reaction using a group of partial differential and algebraic equations. The parameters of the P2D model
have corresponding physical meaning and can be treated as an indicator of battery SOH, e.g. Zhang et al. used the
stoichiometric number of the electrode material to indicate the stages of capacity fading [4-5]. Schmidt et al. investigated the
relationship between the parameters εs, κe and the cycle number during aging and used them to evaluate the SOH of batteries
[6]. Ramadesigan et al. found that the parameters ks,a and Ds,a degraded with cycle number according to a power-law [7].
These papers have established relationships between a limited number of parameters and battery SOH. The aim of this work
was to investigate a much larger set of model parameters under different stages of aging. Therefore, a rapid, accurate and
non-invasive identification method is necessary.
However, the large number of parameters in the P2D model makes parameter identification computationally challenging.
For example, it takes about three weeks to identify 88 parameters on a cluster of five quad-core PCs using a dynamic test [8].
The identifiability of parameters and identification accuracy are also important issues. Forman et al. used the Fisher
information matrix to study parameter identifiability [8-9]. They concluded that some of the parameters in the P2D model are
unidentifiable from experimental data because they do not affect the model output in a unique and significant way. The
authors also summarized the important factors that determine parameter identifiability. Firstly, the identifiability is a function
of which parameters are being identified and which are assumed to be already known. Secondly, the identifiability also
depends on the values of the parameters after they have been fitted to the data. Finally, the experimental procedure can
greatly influence parameter identifiability.
Identifiability can be indicated by parameter sensitivity, i.e. the influence on battery model output when a parameter
changes value. When the value of a highly sensitive parameter changes a little, the model output, i.e. the terminal voltage and
surface temperature, will change noticeably. Thus it is easier to identify a highly sensitive parameter from experiments. On
the other hand, low sensitivity parameters have very little effect on the model output and are therefore hard to identify
correctly from experiments. Insensitive parameters cannot be identified and should be excluded from identification.
Schmidt et al. [10] also used the Fisher information matrix to analyze parameter sensitivity and excluded some
unidentifiable parameters to increase the identification quality of the remaining parameters. They also proposed an
identification scheme with five different experiments, in which the parameters were grouped and identified stepwise. Using
this method the variance and confidence interval can be obtained, but the identifiability matrix needs to be calculated, which
is very complex with a multi-physics model. Min et al. [11] and Wang et al. [12] both used model simulation to examine
parameter sensitivity for a PEM fuel cell model. They roughly classified the parameters into several groups without a
quantitative method. Srinivasulu et al. provided a simple method based on simulated polarisation curves to quantify the
parameter sensitivity for a PEM fuel cell model [13]. A similar method has not yet been applied to Li-ion batteries. Moreover,
there are few established standards for identification experiments; constant current tests [7,14] and dynamic tests [8,10] have
typically been used by researchers. These all tend to employ the terminal voltage as the primary model output, while the
surface temperature of the battery has been seldom used.
In this article, we analyze the parameter sensitivity of our multi-physics model and propose a feasible method to design
experiments for identifying the parameters more accurately using terminal voltage Uapp and surface temperature Tsh as two
identification objects. The remainder of the paper is organized as follows: Section 2 extends the P2D model to a
multi-physics model by adding the thermal behavior and temperature distribution in the radial direction of a cylindrical
Li-ion battery. In this way the surface temperature, which is a readily measurable quantity, can be simulated more accurately.
Section 3 analyses the parameter sensitivity and proposes a parameter sensitivity matrix to describe the sensitivity of model
output when parameters take different values on different operating conditions. In Section 4, a stepwise experiment is
designed based on the parameter sensitivity analysis to facilitate more accurate parameter identification. Validation studies of
the rationality and effectiveness of the stepwise experiment are also presented. Section 5 summarizes the main conclusions of
the paper.

2. Multi-physics model
The multi-physics model developed in this work consists of the P2D electrochemical model coupled to a thermal model of
heat flow and temperature distribution in a cylindrical Li-ion battery as described below.

2.1. Electrochemical model


The schematic of the P2D model for a Li-ion battery is shown in Fig. 1. It consists of two current collectors, a negative
electrode (anode), a separator and a positive electrode (cathode). The electrodes and separator have a porous structure. Two
inner boundaries (anode/separator interface 2 and separator/cathode interface 3) and two external boundaries (Cu/anode
interface 1 and cathode/Al interface 4) are also shown. The electrochemical reactions occurring in the electrodes during
charge and discharge processes can be expressed as follows:

anode Li x C6 discharge 6C  xLi +  xe 


 charge
 (1)
cathode Li y MO2 +xLi +  xe 
discharge

 charge
Li y  x MO2

where MO2 stands for metallic oxide, such as CoO2 or FePO4, and x and y are the stoichiometric numbers of anode and
cathode respectively, i.e. the ratio of Li+ concentration Cs and maximum Li+ concentration Cs,MAX.
During the discharge process, active lithium diffuses from within the bulk of the anode material to the surface, where
lithium ions (Li+) are generated and released into the electrolyte by the electrochemical reaction which occurs on the
solid/electrolyte interface, releasing an equal number of electrons (e-). Lithium ions move through the electrolyte towards the
cathode through the separator, then insert into the cathode active material via another electrochemical reaction. The electrons
cannot pass through the separator, so they migrate to the current collectors and the external circuit, thereby providing the
electrical work. These physical and chemical processes are described by several partial differential equations and algebraic
equations summarised in Appendix A.
On the other hand, thermal phenomena such as energy conservation, heat generation and conduction also occur during
the charge/discharge process and can be taken into account in the P2D model. In addition, some model parameters (e.g.
diffusion coefficient, conductivity and reaction rate) depend on the battery temperature (see Appendix B for details). A
coupled electrochemical-thermal model is presented by feeding back the relationship between temperature-dependent
parameters and thermal effects to the P2D model.
Fig. 1. Schematic of P2D model (discharge process).

2.2. Thermal distribution of cylindrical Li-ion battery


A typical cylindrical Li-ion battery consists of a spiral cell as shown in Fig. 2. Thin layers of anode, cathode, separator
and current collector are stacked as shown in Section B-B, rolled up on a central mandrel (Section A-A) and then inserted
into a can. The gaps are filled with liquid electrolyte.

Fig. 2. Cross-section of a cylindrical Li-ion battery [15].

Therefore, thermal distribution occurs in three different directions: (a) the thickness direction of the electrode plate, (b)
the radial direction of the cylindrical battery, and (c) the axial direction of the cylindrical battery. The Biot number, which is
used to determine whether or not the temperature gradient is negligible, is defined as:

Bi  hL /  (2)

where h is the heat-transfer coefficient, L is a characteristic length and λ is the effective thermal conductivity. If the Biot
number is greater than 0.1, the temperature gradient in that direction cannot be neglected [16].
Generally, the thickness of the electrode plate in a small battery is several hundred microns, while the heat transfer
coefficient and the thermal conductivity of the electrode material are of the same order of magnitude, so the Biot number is
usually well below 0.1 and therefore the temperature distribution in the through-thickness direction can be neglected. The
temperature distribution in the axial direction can be also be neglected because the axial thermal conductivity λa is far higher
than the radial thermal conductivity λr. The ratio λa/λr is more than 100 [17] and the characteristic length is equal to the
battery height which is just a few centimeters, yielding a Biot number less than 0.1. However, for the spiral roll configuration,
the heat-transfer coefficient h is typically one or two orders of magnitude greater than the radial thermal conductivity λr,
which is limited by large contact thermal resistances between the two different layers of battery materials. For example, the
radius of a 26650 type Li-ion battery is 0.013 m, the heat-transfer coefficient of the lateral surface is ~10-50 W m-2 K-1 and
the radial thermal conductivity is 0.2 W m-1 K-1, so the Biot number is greater than 0.1. In this case, temperature gradients in
the radial direction must be taken into account or the simulated temperature will be very different from the experimental
value measured at the surface of the battery.
By neglecting the temperature gradient in the thickness direction of the electrode plate and in the axial direction of a
cylindrical battery, the heat conduction term in Equation (B.1) can be reduced to one-dimensional form by Fourier’s Law:

 2T 1 T
(T )   (  ) (3)
R 2 R R

The temperature gradient at the center of the battery is zero and it is equal to the heat exchange rate at the surface:

 T
 R 0
 R 0
 (4)
  T q
 R R  Rcell

where Rcell is the radius of the cylindrical battery and the heat exchange rate q̇ is given by q  qc  qr , where q̇c and q̇r are the
heat exchange rates due to conduction and radiation respectively. The temperature gradient is continuous at the interfaces of
spiral roll, gap and can.
A thermal impedance model based on that presented by Fleckenstein et al. [17] is used to simulate the thermal
distribution in the radial direction of a cylindrical Li-ion battery, as shown in Fig. 3.

Fig. 3. 1-D thermal impedance model in radial direction of cylindrical battery.

In Fig. 3, the spiral roll is divided into NR grids in the radial direction using the finite volume method. Each finite
volume element is a cylindrical shell and the innermost one is a cylinder. The volume Vi and centroid radius ri of each
element are obtained from the geometric parameters and Ti is the temperature of the ith element. The thermal resistance
between two adjacent elements is given by [17]:

1 r
Ri  ln i 1 (5)
 2 H ri

where H is the effective height of the spiral roll. The thermal capacity is determined by:

Ci  Vi  C p (6)

Here the density ρ and heat transfer coefficient λ are volume-averaged values for all materials in the spiral roll.
The thermal impedance model can be described by the state equation in Equation (7). It is noticeable that the innermost
element is filled only with separator and electrolyte, so there is no heat source in it, while the gap and the can are treated as
independent elements.
 1 1 
 R C 
 NR NR
RNRC NR
  0 
 1 RNR  RNR 1 1   
R C   VNR 1Q 
 TNR   NR NR 1   TNR   CNR 1 
CNR 1 RNR RNR 1 RNR 1CNR 1
    T   
TNR 1     NR 1   
   1 Ri 1  Ri 1    ViQ  (7)

   Ri 1Ci Ci Ri 1 Ri Ri Ci  T   C 
 i   i  
T i
      
   R2  Rroll  gap  T   
  1   V1Q 
 T1   1 1

T   R2C1 C1 R2 Rroll  gap Rroll  gap C1   Tgap   C1 
 gap      
 Tcan   1 Rroll  gap  Rgap can 1   Tcan   0 
  
 Rgap can Cgap 

Rroll  gap Cgap Cgap Rroll  gap Rgap can
   Ash q 
 1 1   Ccan 

 Rgap can Ccan Rgap canCcan 

Here Cgap = mgapCp,gap is the heat capacity of the gap, Ccan = mcanCp,can is the heat capacity of the can, Rroll→gap and Rgap→can
represent the thermal resistances between spiral roll, gap and can, calculated in a similar manner to Equation (5), Ash q is the
heat exchange between battery shell and the external environment and Ash is the lateral surface area of the cylindrical battery.
Ti t 1  Ti t
The discrete form of the temperature state quantity Ti is Ti  ; when t=0, Ti=Tgap=Tcan=Tam. The temperature
t
distribution Ti(t) during the charge/discharge process can be calculated by iteration with time step Δt. The average value of T1
to TNR is used as T in the coupled electrochemical-thermal model.

2.3. Parameter set and simulation


A set of physical and chemical parameters for a cylindrical 2.3 Ah LiFePO4 battery is shown in Table 1. In the literature,
several expressions have been used to describe the temperature dependence of De, κe and Ds. Considering that these
expressions will probably be no longer valid following electrolyte and electrode degradation, we have opted to use a
reference value and the Arrhenius law instead. These parameters are treated as unknowns to be analyzed and identified
further.
Table 2 shows the thermal and geometrical parameters of the battery. The thermal properties of the spiral roll are
volume-averaged values for all the materials, except the thermal conductivity which was experimentally measured [18].
We improved DUALFOIL, a freely available FORTRAN code maintained by Newman’s research group, by adding
subroutines for local heat generation, heat exchange, temperature distribution and parameter updates. This program can
simulate both internal behavior (e.g. concentration distribution of electrolyte and reaction current distribution, etc.) and
external performance (e.g. terminal voltage and battery temperature) when inputting a set of parameters and an operating
condition. It takes about 8 seconds to simulate the 1C discharge process on a PC with INTEL Core i3-530 CPU (2.93GHz)
and 4G RAM. The simulation time is relevant to the operating condition time and will become slower when convergence is
an issue, because in this case the step size will become smaller automatically according to the variable step algorithm.
Fig. 4 shows a comparison of simulated discharge curves for 0.5C, 1.0C and 2.0C at 30 C. Simulated curves are
represented by solid lines and relevant experimental data by symbols. Reasonable agreement is obtained between the
simulated Uapp and Tsh curves and the experimental data, with minor differences attributed to the parameters of the
experimental battery not being exactly the same as those in Tables 1 and 2. This might be caused by a different production
batch or performance degeneration.
Fig. 5 shows the simulated temperature distribution during 2C discharge. The horizontal axis in Fig. 5(a) indicates the
radial direction of a cylindrical Lithium-ion battery, where R=0 represents the innermost element of the spiral roll and R=Rcell
represents the can. There is no heat source in either the gap or the can and the thermal conductivity of the can is very high
compared to that of the spiral roll, so a smaller temperature gradient appears there. The shell temperature simulated by our
thermal impedance model and the average temperature calculated using an isothermal model, in which the temperatures of all
elements in battery are considered the same, are compared with the experimental shell temperature in Fig. 5(b). Clearly, the
thermal impedance model gives a more accurate prediction of shell temperature, making it possible to identify model
parameters by fitting the measured shell temperature to simulation results.
Table 1 Physical and chemical parameters for a 2.3 Ah LiFePO4 battery.

Parameters Anode Separator Cathode Ref.


Design specifications
Acell (m2) 0.1755 - 0. 1694 b

b
L (um) 34 30 70
δ (um) 25 - 25
b
Rs (um) 3.5 - 0.0365
εs 0.55 - 0.43 a

εe 0.33 0.54 0.332 a

Li-ion concentration
b
CsMAX (mol m-3) 31370 - 22806
b
Cs0 CsMAX 0.8 - 0.03

Diffusion coefficient and conductivity


Ds (m2 s-1) 3.9e-14 - 1.18e-18 a

ĒDs (kJ mol-1) 35 - 35 a

σs (S m-1) 100 - 0.5 a

Rfilm (ohm m2) 0.003 - 0 e

Kinetic
ks (m2.5 mol-0.5 s-1) 3e-11 - 1.4e-12 a

Ēks (kJ mol-1) 20 - 30 a

α 0.5 - 0.5 b

γ 1.5 1.5 1.5 a

Electrolyte properties
Ce (mol m-3) 1200 a

De (m2 s-1) 2.0e-10 d

ĒDe (kJ mol-1) 11 d

κe (S m-1) 0.2 c

Ēκe (kJ mol-1) 26.6 d

Other parameters and constants


a
t+ 0.363
Rext (Ω m2) 0.002 e

R 8.3143
F (C mol-1) 96487

a
Ref. [18]
b
Ref. [19]
c
Estimated value from Ref. [3]
d
Ref. [20]
e
Estimated value

Table 2 Thermal properties and geometrical parameters for a 2.3 Ah LiFePO4 battery.

λ ρ Cp
Thermal properties
(W m-1 K-1) (kg m-3) (J kg-1 K-1)
Spiral roll 0.2 a 2189 c 1014 a
Electrolyte 0.6 a 1130 a 2055 a
Can 14 b 7917 b 460 b

Cylindrical battery geometry parameters


Diameter (m) 26e-3 b
Height (m) 65e-3 b
Height of roll (m) 57.2e-3 b
Diameter of roll (m) 24.8e-3 c
Mass of roll (kg) 60e-3 c
Can thickness (m) 0.3e-3 b
Mass of can (kg) 13e-3 c
Heat exchange parameters
Heat transfer coefficient (W m-2 K-1) 10 c
Surface emissivity (W m-2 K-4) 0.8 b

a
Ref. [18]
b
Ref. [15]
c
Estimated value
Fig. 4. Comparison of simulated and experimental data at 30 C.

Fig. 5. (a) Radial temperature distribution and (b) comparison of simulated shell temperature, simulated average temperature and
experimental shell temperature for cylindrical Li-ion battery during 2C discharge process.

3. Parameter sensitivity analysis


3.1. Parameters to be examined
Not all parameters in Table 1 and 2 need to be analyzed and identified. Some parameters are unaffected by ageing and
some can be obtained directly by dismantling the battery, such as the geometrical parameters or material characteristics. A set
of 30 parameters, which should be identified by a non-invasive method, is proposed and divided into 2 groups.
Group 1 includes 17 parameters. They are the radii of active particles, Rs, a and Rs, c, the diffusion coefficients in the solid
phase, Ds, a and Ds, c and the electronic conductivities in the solid phase, s, a and s, c. Subscripts a and c refer to anode and
cathode respectively. Changes in these parameters indicate degradation of material caused by active particle dissolution and
degeneration [7,21]. The initial stoichiometric numbers x0 and y0 and the volume fractions of active material s,a and s,c, are
representative of capacity fade caused by loss of cyclable lithium ions and active material [4]. The film resistances, Rfilm,a and
Rfilm,c, reflect film formation, which consumes active lithium ions [5]. In addition, the concentration, Ce, diffusion coefficient,
De, ionic conductivity, e, and the density, e, of electrolyte will change when reduction reactions and decomposition occur
[22]. The external resistance Rext is also in this group. These parameters characterize the basic properties of the battery, and
are most likely to change when the battery degenerates, so we treat them as internal characteristics in the PHM database.
Group 2 includes 13 parameters. They are the electrolyte volume fractions e,a, e,c and e,s, the reaction rate ks,a and ks,c,
the activation energy in the Arrhenius law (including Ēks,a, Ēks,c, ĒDs,a, ĒDs,c, ĒDe, and Ēκe), the average thermal conductivity,
λ, and the heat transfer coefficient, h. These parameters may not change as a result of degradation but are difficult to obtain
directly.

3.2. Methodology
In individual parameter sensitivity analysis, five values for each parameter were chosen from a specific range to analyze
their influence on the discharge terminal voltage curve and the shell temperature curve under different conditions with the
other parameters held constant. The ranges examined for each parameter are shown in Table 3, taking into account
benchmark values from the literature and the anticipated variation during the aging process. This method has a limitation in
that the results are only valid for a given set of benchmark values. If some parameter values are far away from the benchmark
value or taken from another battery type or system, the result may be significantly different. In this case, parameter sensitivity
should be re-examined in the new range with revised benchmark values.
Simulation of complete discharge of the battery from the fully charged state at five different ambient temperatures and
five different discharge rates is shown in Fig. 6. The ambient temperatures (Tam) are -5 °C, 10 °C, 25 °C, 40 °C and 55 °C and
the discharge current rates (Crate) are 0.2C, 0.5C, 1C, 2C and 4C. Thus 25 combined operating conditions are investigated in
this paper. Fig. 6 shows the curves for parameters Rs,a and Ēks,a as examples.

Fig. 6. (a) Terminal voltage curves and (b) shell temperature curves for various Rs,a at 1C and 25 C.
(c) Terminal voltage curves and (d) shell temperature curves for various Ēks,a at 2C and -5C.

Different influences of the varied parameters are observed on the simulated Uapp and Tsh curves. Both the Uapp and Tsh
curves show high sensitivity to Rs,a across the entire depth-of-discharge (dod) range, as shown in Fig. 6(a) and 6(b), with
higher values leading to lower Uapp, higher Tsh and a lower End-of-Discharge (EOD) voltage. In contrast, varying Ēks,a has a
more limited influence on the Uapp curve in the high dod region and no influence on EOD. It is very difficult to characterize
parameter sensitivity without a quantitative method, especially with 25 combined operating conditions, and the required
method is described below.

3.3. Parameter sensitivity matrix


The dispersion of Uapp curves indicates the parameter sensitivity of Uapp. It was calculated as follows. Firstly, the
standard deviation of five Uapp values at each dod point was determined as:

SU (dod , j, i )  std(U app (dod , k , j, i )) (8)


k

where Uapp(dod,k,j,i) represents the terminal voltage at the dodth point on the discharge curve for the kth value of the parameter,
at the jth ambient temperature and ith current, while std() is the standard deviation function.
Secondly, all dod points were divided into five dod regions represented by the variable d, as shown in Fig. 6. The mean
values of SU(dod,j,i) were calculated in each dod region using Equation 9, where EOD1 is equal to the dod at which the
curves first reach the EOD voltage, as indicated in Fig. 6(a).
 SMU (1, j, i )  mean [ SU (dod , j, i )]
 0 dod  0.2
 SMU (2, j, i )  mean [ SU (dod , j, i )]
 0.2  dod  0.4

 SMU (3, j , i )  mean [ SU (dod , j, i )] (9)
0.4 dod  0.6

 SMU (4, j, i )  0.6mean dod  0.8
[ SU (dod , j, i )]

 SMU (5, j, i )  0.8mean
dod  EOD1
[ SU (dod , j, i )]

If EOD1 appears in the region where d=4, then:

SMU (4, j, i )  SMU (5, j, i )  mean [ SU (dod , j, i )] (10)


0.6 dod  EOD1

Finally, a 3-D matrix SMU(d,j,i) which is termed the “parameter sensitivity matrix” of Uapp was obtained using
Equations (8) to (10) under 25 permutated conditions, so the dimensionality is (5,5,5). Each element in this matrix represents
the parameter sensitivity for a certain dod region, ambient temperature and current, expressed in units of volts. The SMU
matrices of 30 different parameters were calculated by this method.
The column “SMUavg” in Table 3 shows the mean value of all elements, while the column “SMUmax” shows the
maximum value in the SMU matrix for each parameter. Parameters whose mean sensitivity was greater than 0.01 were
grouped as “Highly Sensitive”, while those with mean sensitivity between 0.005 and 0.01 were grouped as “Sensitive”. “Low
Sensitive” parameters were defined as those with mean sensitivity lower than 0.005, but maximum sensitivity greater than
0.005. Parameters with maximum sensitivity lower than 0.005 were grouped as “Insensitive”. This rough grouping is
indicated in the “Sens. Uapp” column of Table 3.
The Uapp parameter sensitivity matrix is shown in Fig.7 using slice plots, taking Rs,a and Rfilm,a as examples. It can be
seen that the parameter sensitivity is continuously related to operating conditions and dod.

Fig. 7. Uapp parameter sensitivity matrix for (a) Rs,a and (b) Rfilm,a.

In the same way, the Tsh parameter sensitivity matrix can be obtained and expressed as SMT(d,j,i), with units of degrees
Celsius. The mean and maximum sensitivity values of the matrices are shown in the “SMTavg” and “SMTmax” columns in
Table 3. The parameters were also roughly grouped with threshold values of 0.05 and 0.2 as indicated in the “Sens. Tsh”
column in Table 3. The SMT matrices for parameters Rs,a and Rfilm,a are shown in Fig. 8.
The SMU matrix and SMT matrix for the same parameter show little correlation. Some parameters (e.g. heat transfer
coefficient h) have a high sensitivity in SMT but low sensitivity in SMU. Since temperature tends to increase with depth of
discharge, Tsh sensitivity is positively related to dod under certain operating conditions. Tsh sensitivity is also positively
related to current because larger currents generate more heat and the thermal response is more obvious. The maximum SMT
values of most parameters appear in the high dod and Crate regions.
The total sensitivity should take into account both the SMU and SMT by summing the number of asterisks in the
“Sens.tot” column in Table 3. Nine parameters (Rs,a, Ds,a, x0, εs,a, Rext, εe,a, ks,a, ks,c and Ēκe ) have 5 or 6 asterisks in total; they are
regrouped as “Highly Sensitive”. Eight parameters (Rs,c, σs,c, Rfilm,a, Ce, De, κe, Ēks,a and h) have 4 asterisks in total and they are
regrouped as “Sensitive”. Ten parameters (Ds,c, y0, εs,c, εe,s, εe,c, Ēks,c, ĒDs,a, ĒDs,c, ĒDe and λ) have 2 or 3 asterisks and they are
regrouped as “Low sensitive”. Three parameters (σs,a, Rfilm,c and ρ) have no asterisks; they are regrouped as “Insensitive” and
will be excluded from further analysis, because they have negligible influence on either terminal voltage or surface
temperature and the identification results will be not credible and cannot be used for studying the PHM characteristics.

Fig. 8. Tsh sensitivity value matrix for (a) Rs,a and (b) Rfilm,a.

3.4. Clustering analysis


Theoretically, a parameter can be identified most easily and accurately under the operating condition that provides the
highest sensitivity. This is termed the “Best Conditions for Identification (BCI)” and indicated by the index of the parameter
sensitivity matrix. Due to the continuous relationship between parameter sensitivity and operating conditions, conditions
similar to BCI could also provide a relatively high sensitivity, while conditions that are opposite to BCI will result in a
relatively low sensitivity. For example, the BCI for parameter Rs,a is “low Tam, very high Crate and very high dod”; under these
conditions the Uapp and Tsh sensitivity are both relatively high. The similar conditions “room temperature, medium Crate, and
medium dod” result in a medium sensitivity, while “high Tam, low Crate, and low dod”, which is opposite to BCI provides
relatively low sensitivity. This conclusion can be easily observed from inspection of Fig. 7(a) and Fig. 8(a).
The best identification strategy is to identify each parameter under its own BCI or under conditions similar to its BCI,
but this is at best extremely complicated and at worst impossible because of the diverse range of BCIs for all parameters. In
order to solve this problem, parameters that have the same or similar BCI should be grouped into clusters, allowing them to
be identified together under relatively few sets of operating conditions that render them sensitive and thus readily identified.
The clustering procedure is as follows:
Firstly, the maximum sensitive point which indicates the BCI was determined for both the SMU and SMT and each
parameter. For example, the maximum sensitive point of parameter Rfilm,a is (1,5,5) for SMU and (5,5,5) for SMT, i.e. the BCI
of SMU is “very low dod, very high Tam and very high Crate” and the BCI of SMT is “very high dod, very high Tam and very
high Crate”.

Fig. 9. The BCI characteristic for parameter Rfilm,a.

Secondly, the values of the SMU and SMT matrices were normalized, ensuring that all elements were in the range [0,1],
and values in three dimensionalities coordinating to the maximum sensitive point were determined to form a vector with 15
elements. This vector is defined as the “BCI characteristic” and characterizes how similar an operating condition is to the
BCI of a given parameter. Fig. 9 shows the BCI characteristic of parameter Rfilm,a. The blue line is the BCI characteristic of
SMU, the green line is that of SMT and the red line is the average BCI characteristic, which is obtained from averaging the
former two with weighting. The weights are the sensitivity of Uapp and Tsh, shown as the number of asterisks in Table 3.
Thirdly, the Fuzzy C-Means (FCM) method was used for clustering. The average BCI characteristic was used as
clustering feature. The “Clustering results” column in Table 3 shows the degree of association of each parameter with each of
the three clusters. The cluster centers are shown in Fig. 10. 13 parameters were allocated to Cluster A, 5 parameters to Cluster
B and 6 parameters to Cluster C, since their membership is greater than 0.5. Three parameters (y0, De, Ēks,a) could not be
obviously clustered.

Table 3 Parameter sensitivity analysis and cluster results.


Sensitivity of Uapp Sensitivity of Tsh Clustering results
Parameters Range Sens.tot
SMUavg SMUmax Sens.Uapp SMTavg SMTmax Sens.Tsh A B C

Rs,a (μm) 1~9 0.0294 0.2429 *** 0.4419 2.6006 *** *** 0.1511 0.7428 0.1061
Rs,c (μm) 0.02~0.06 0.0053 0.0571 ** 0.0904 0.893 ** ** 0.8009 0.1244 0.0747
2 -1
Ds,a (m s ) 0.5~10 e-14 0.0216 0.304 *** 0.1632 1.2066 ** *** 0.3207 0.5754 0.1039
Ds,c (m2 s-1) 0.5~10 e-18 0.0033 0.1672 * 0.0461 0.935 * * 0.5806 0.2943 0.1251
σs,a (S m ) -1
10~100 1.39E-05 5.01E-05 / 4.53E-04 0.0022 / / / / /
σs,c (S m ) -1
0.1~5 0.0052 0.0137 ** 0.0959 0.4818 ** ** 0.1008 0.1243 0.7748
x0 (-) 0.75~0.85 0.0225 0.1343 *** 0.0775 0.5575 ** *** 0.1209 0.7624 0.1167
y0 (-) 0.02~0.1 0.0025 0.0148 * 0.0534 0.2118 ** * 0.2174 0.4365 0.3461
εs,a (-) 0.5~0.6 0.0256 0.1877 *** 0.098 0.8265 ** *** 0.0847 0.847 0.0682
εs,c (-) 0.4~0.5 0.0019 0.0144 * 0.0474 0.4228 * * 0.6259 0.2869 0.0873
Rfilm,a (Ω m ) 2
0~0.01 0.0052 0.0122 ** 0.1288 0.6594 ** ** 0.072 0.079 0.849
Rfilm,c (Ω m2) 0~0.01 2.75E-05 1.48E-04 / 7.97E-04 0.0053 / / / / /
-3
Ce (mol m ) 800~1200 0.0055 0.0073 ** 0.0693 0.3853 ** ** 0.082 0.0814 0.8366
2 -1
De (m s ) 0.1~10 e-10 0.0061 0.0175 ** 0.0818 0.6509 ** ** 0.3401 0.2871 0.3728
κe (S m-1) 0.4~2 0.0051 0.0253 ** 0.1521 1.413 ** ** 0.5956 0.1658 0.2386
ρe (kg m ) -3
1000~1200 7.93E-05 0.0017 / 0.0116 0.0982 * / / / /
Rext (Ω m ) 2
0~0.002 0.0143 0.0413 *** 0.4367 2.243 *** *** 0.0725 0.0794 0.8481
εe,a (-) 0.3~0.4 0.0314 0.1937 *** 0.0921 0.6515 ** *** 0.1319 0.7543 0.1138
εe,s (-) 0.5~0.6 8.38E-04 0.0056 * 0.0267 0.1737 * * 0.526 0.1987 0.2753
εe,c (-) 0.3~0.4 0.0023 0.0946 * 0.0251 0.217 * * 0.6276 0.2758 0.0966
2.5 -0.5 -1
ks,a (m mol s ) 0.1~10 e-11 0.0528 0.079 *** 1.2472 4.1843 *** *** 0.0873 0.1234 0.7893
ks,c (m2.5 mol-0.5 s-1) 0.1~10 e-12 0.0165 0.0572 *** 0.4771 2.0416 *** *** 0.1258 0.1642 0.7101
Ēks,a (KJ mol ) -1
10~50 0.0057 0.0243 ** 0.1462 0.7194 ** ** 0.4439 0.3061 0.25
Ēks,c (KJ mol ) -1
10~50 0.0016 0.0184 * 0.0495 0.3761 * * 0.5534 0.2564 0.1903
ĒDs,a (KJ mol ) -1
10~50 0.004 0.1167 * 0.0288 0.6996 * * 0.5015 0.4029 0.0956
ĒDs,c (KJ mol-1) 10~50 9.29E-04 0.0196 * 0.0189 0.3788 * * 0.761 0.1605 0.0785
ĒDe (KJ mol ) -1
10~50 0.0019 0.0156 * 0.0531 0.6025 ** * 0.7366 0.148 0.1154
Ēκe (KJ mol ) -1
10~50 0.0105 0.0841 *** 0.301 2.5004 *** *** 0.6522 0.1926 0.1552
λ (W m K ) -1 -1
0.1~1 3.87E-04 0.0067 * 0.1464 1.6538 ** * 0.8032 0.1226 0.0742
h (W m-2 K-1) 5~50 0.0013 0.0226 * 0.4151 4.0856 *** ** 0.7477 0.172 0.0803

***: Highly sensitive


**: Sensitive
*: Low sensitive
/: Insensitive

The conditions indicated by the highest points of the cluster centers are defined as the “Best Practicable Conditions
(BPC)”. Under these conditions the parameters in the same cluster could be at or close to their own BCIs. Thus a simple set
of conditions could be used to identify a cluster of parameters, yielding a group of highly identifiable parameters. In short,
BCI is the condition that makes one parameter relatively most sensitive, and BPC is the condition that makes a cluster of
parameters relatively most sensitive synchronously. The “BPC characteristic” shows how similar an operating condition is to
a cluster’s BPC.

Fig. 10. Cluster centers, regarded as BPC characteristics.

In Fig. 10, these three BPC characteristics are very similar in the trend, the conditions of higher dod, lower Tam and
higher Crate are more similar to BPC, while in contrast the conditions of lower dod, higher Tam and lower Crate are far away
from BPC.
Finally, these three BPC characteristics were combined by averaging. A combined BPC characteristic was obtained as
shown in Fig. 11. Theoretically, the combined BPC “very high dod, very low Tam, and very high Crate” would make most of
the parameters relatively sensitive, making the experiments much simpler in this case. However, the sensitivity of some
parameters may be lower than others due to differences between the combined BPC and their own BCI. Hence, this combined
BPC characteristic must be used with caution in the design of identification experiments, and the rationality of such
experiments should always be subsequently validated.

Fig. 11. Combined BPC characteristics.

4. Stepwise experiment for parameter identification


Identifying all parameters synchronously using a simple experiment has some disadvantages. The parameter space is
very large and the optimal solution is difficult to obtain. In addition, it is computationally expensive to identify all parameters
using a parameter optimization algorithm and it is difficult to identify the low sensitive parameters correctly, because the
model output is dominated by highly sensitive parameters. Therefore, a stepwise identification is proposed based on the
results of parameter sensitivity analysis to overcome those disadvantages.

4.1. Experimental design


The stepwise experiment should be designed according to the following criteria:
(1) “Insensitive” parameters should be excluded from identification.
(2) “Highly sensitive” parameters should be identified first and the results used to identify the “Sensitive” parameters in
the second step. Similarly, the results of previous steps are used to identify the “Low sensitive” parameters.
(3) Operating conditions of the test in each step should refer to “BPC”.
(4) Tests in the stepwise experiment should be as simple as possible. Conditions that are harmful to the batteries should
be avoided.
The major challenge in stepwise identification is that when we identify parameters in the first step, the values of some
parameters (to be identified in later steps) are still unknown. Estimated values are input into the model because the model
needs a whole parameter set to complete the simulation. The choice of estimated value can have a significant influence on
model output and thus the identification results may be incorrect. To overcome this problem, the stepwise experiment should
be designed using sensitivity grouping results. The key concept is to ensure that the sensitivity of the parameters to be
identified is relatively high while that of the other parameters is low. Hence, in Step 1 the “highly sensitive” parameters
should be identified under conditions not similar to BPC, in order to ensure low sensitivity of the other parameters while
maintaining the identifiability of the “highly sensitive” parameters. Similarly, the conditions used for identification of the
“low sensitive” parameters should be similar to BPC, to provide a high identifiability.

Fig. 12. Flow chart for stepwise identification.

Therefore, the following three-step identification was designed based on the BPC characteristic (Fig. 11), taking into
account the ease and simplicity of experiments. The details of the experimental scheme are shown in Fig. 12.
Step 1: The operating conditions are dod1 = 0.6, Tam,1 = 20 C and Crate,1 = 2C. Under these conditions, most parameters
have relatively low sensitivity, but highly sensitive parameters should remain sensitive. Benchmark values are used for the
“sensitive” and “low sensitive” parameters, listed in Table 1 as estimated values. Using this strategy, even if the estimated
values are far from the true values, the model output will not be significantly affected and highly sensitive parameters will
still be identified correctly.
Step 2: The operating conditions are dod2 = 0.8, Tam,2 = 10 C and Crate,2 = 3C. Under these conditions, parameters are
more sensitive than in Step 1, and the “sensitive” parameters can meet the sensitivity demand of identification, while the
sensitivity of “low sensitive” parameters is still sufficiently low. In this step, the identified results of “highly sensitive”
parameters and benchmark values of “low sensitive” parameters are input into model.
Step 3: The operating conditions are dod3 = 1.0, Tam,3 = 0 C and Crate,3 = 4C. Under these conditions, all parameters are
relatively highly sensitive and the “low sensitive” parameters are now easier to identify. Previously identified values of other
parameters are used as model inputs.
It should be noted that the stepwise experiments described above are not universally applicable and should only be used
for this particular type of LiFePO4 battery, whose parameters are listed in Table 1 and Table 2. For other types of battery, it is
necessary to obtain benchmark values for parameters by consulting the manufacturer and/or relevant literature.

4.2. Rationality validation


Using the same methodology presented in Sections 3.2 and 3.3, the parameter sensitivities were analyzed under the
three-step conditions, with the results summarized in Table 4. The “Identify?” column indicates whether the parameter meets
the sensitivity demand of identification in a certain step, while a blank space means the parameter has been identified in a
previous step and need not be analyzed. The validation results were also used to revise the parameter groupings according to
their sensitivity.
Table 4 Parameter sensitivities in three-step experiment.
Step 1 Step 2 Step 3
Parameter sens.avg SMU sens.U SMT sens.T sens Identify? SMU sens.U SMT sens.T sens Identify? SMU sens.U SMT sens.T sens Identify?
Rs,a *** 0.0465 *** 0.9851 *** *** Y 0.0643 1.2954 0.0805 1.1608
Rs,c ** 0.0035 * 0.1046 ** * 0.0061 ** 0.2529 *** *** Y 0.0296 0.4645
Ds,a *** 0.0279 *** 0.2766 *** *** Y 0.0420 0.2815 0.0587 0.4696
Ds,c * 0.0018 * 0.0298 * * 0.0031 * 0.0657 ** * 0.0199 *** 0.1884 ** *** Y
σs,a / 0.0000 / 0.0007 / / 0.0000 / 0.0012 / / 0.0000 / 0.0016 / /
σs,c ** 0.0036 * 0.1474 ** * 0.0049 * 0.2519 *** ** Y 0.0060 0.3277
x0 *** 0.0105 *** 0.0714 ** *** Y 0.0098 0.1213 0.0106 0.1949
y0 * 0.0018 * 0.0562 ** * 0.0019 * 0.0809 ** * 0.0025 * 0.1093 ** ** Y
εs,a *** 0.0130 *** 0.0869 ** *** Y 0.0139 0.1430 0.0153 0.2577
εs,c * 0.0017 * 0.0644 ** * 0.0028 * 0.1241 ** * 0.0042 * 0.1988 ** ** Y
Rfilm,a ** 0.0049 * 0.2082 *** ** 0.0068 ** 0.3462 *** *** Y 0.0081 0.4500
Rfilm,c / 0.0001 / 0.0017 / / 0.0001 / 0.0022 / 0.0001 / 0.0031 / /
Ce ** 0.0034 * 0.1362 ** * 0.0046 * 0.2250 *** ** Y 0.0055 0.2971
De ** 0.0042 * 0.1535 ** * 0.0064 ** 0.2804 *** *** Y 0.0092 0.4299
κe ** 0.0258 *** 1.0204 *** *** Y 0.0425 2.0981 0.0594 3.6948
ρe / 0.0001 / 0.0131 / / 0.0002 / 0.0316 * / 0.0006 / 0.0528 ** *
Rext *** 0.0175 *** 0.7056 *** *** Y 0.0233 1.1732 0.0276 1.5063
εe,a *** 0.0121 *** 0.0729 ** *** Y 0.0126 0.1638 0.0137 0.3053
εe,s * 0.0010 * 0.0422 * * 0.0017 * 0.0863 ** * 0.0023 * 0.1355 ** * Y
εe,c * 0.0014 * 0.0412 * * 0.0038 * 0.0695 ** * 0.0147 *** 0.1632 ** *** Y
ks,a *** 0.0610 *** 2.4345 *** *** Y 0.0592 2.9274 0.0524 2.8313
ks,c *** 0.0214 *** 0.8766 *** *** Y 0.0264 1.4056 0.0283 1.7182
Ēks,a ** 0.0010 * 0.0336 * * 0.0039 * 0.2304 *** *** Y 0.0065 0.4203
Ēks,c * 0.0003 / 0.0087 / / 0.0014 * 0.0809 ** * 0.0034 * 0.2016 *** *** Y
ĒDs,a * 0.0009 / 0.0112 * / 0.0042 * 0.0398 * * 0.0099 ** 0.1184 ** ** Y
ĒDs,c * 0.0004 / 0.0043 / / 0.0010 * 0.0224 * * 0.0040 * 0.3821 *** ** Y
ĒDe * 0.0002 / 0.0081 / / 0.0014 * 0.0748 ** * 0.0040 * 0.2416 *** *** Y
Ēκe *** 0.0015 * 0.0515 ** * 0.0080 ** 0.4615 *** *** Y 0.0199 1.2843
λ * 0.0010 * 0.1988 ** * 0.0028 * 0.5323 *** ** Y 0.0051 1.0226
h ** 0.0029 * 1.1231 *** ** 0.0084 ** 2.1811 *** *** Y 0.0167 3.3176

***: Highly sensitive


**: Sensitive
*: Low sensitive
/: Insensitive

The validation results are consistent with the average sensitivity grouping. “Highly sensitive” parameters exhibit high
sensitivity in Step 1, while other parameter sensitivities are relatively low (even lower than their average sensitivity). There
are two special cases in Step 1. One is the parameter κe, which had been grouped as “sensitive” and was supposed to be
identified in Step 2, but in fact showed high sensitivity in Step 1. If its estimated value is used the identification results may
be adversely influenced, so the parameter κe must now be identified in Step 1. The other special case is the parameter Ēκe,
which showed low sensitivity in Step 1, and will therefore be identified in a later step. In Step 2, “sensitive” parameters
remain sensitive and may even become highly sensitive. The “low sensitive” parameter λ becomes “sensitive” and will thus
be identified in this step. In Step 3, most “low sensitive” parameters become “sensitive” or even “highly sensitive”, so they
can be identified more accurately, but there are still three parameters in the “low sensitive” grouping, implying that their
identification accuracy may be lower than the others.
There are 9 parameters in each step in the revised stepwise experiment, indicated by “Y” in the “Identify?” column in
Table 4. The parameters are relatively sensitive when they are identified, which means that their identifiability is high. The
parameters with estimated values before identification show relatively low sensitivity and have very little influence on model
output. Therefore we can conclude that this stepwise experiment is rationally designed.

4.3. Effectiveness of stepwise experiment for parameter identification


Synthetic experimental data is used to verify the effectiveness of the stepwise experiment. This synthetic experimental
data is not derived from real experiments but from simulation, thus it has many advantages:
(1) The “true values” of parameters can be treated as known, because we set them ourselves and use them to acquire
synthetic experimental data via simulation. The identification results are compared with the “true values”, so the identified
error can then be calculated.
(2) The use of synthetic experimental data circumvents the issues of cost, delay and measurement error associated with
experimental test programmes.
In order to assess the effectiveness of the stepwise experiment, the “true values” of 27 parameters were set, taking into
account the typical range and benchmark values. The set parameter values are shown in the “Value” column in Table 5. These
parameters were used for model simulation under the operating conditions obtained in Section 4.1 and the results are
regarded as synthetic experimental data. Finally, a weight-based multi-objective Genetic Algorithm (GA) was used for
parameter identification, in which the terminal voltage Uapp and surface temperature Tsh were treated as two objects. The
object function is the sum of squares error (SSE) of the simulated curve and synthetic experimental data. A lower SSE means
the simulated curve fits the synthetic experimental data better and the identified results are closer to the “true values”.

Table 5 Results of stepwise identification.


Parameter value result error %
Rs,a (m) 3.60E-06 3.4106E-06 5.26%
Rs,c (m) 4.00E-08 3.7472E-08 6.32%
Ds,a (m2 s-1) 4.00E-14 3.6844E-14 7.89%
Ds,c (m2 s-1) 3.00E-18 2.8929E-18 3.57%
σs,c (S m-1) 0.6 6.4448E-01 7.41%
x0 (-) 0.81 8.1351E-01 0.43%
y0 (-) 0.05 4.9810E-02 0.38%
εs,a (-) 0.6 6.1872E-01 3.12%
εs,c (-) 0.44 4.5131E-01 2.57%
Rfilm,a (Ω m2) 0.004 3.3216E-03 16.96%
Ce (mol m-3) 1100 1.0722E+03 2.52%
De (m2 s-1) 4.50E-10 4.7211E-10 4.91%
κe (S m-1) 0.8 8.0850E-01 1.06%
Rext (Ω m2) 0.002 2.0581E-03 2.91%
εe,a (-) 0.35 3.5135E-01 0.39%
εe,s (-) 0.5 5.3399E-01 6.80%
εe,c (-) 0.35 3.4689E-01 0.89%
ks,a (m2.5 mol-0.5 s-1) 5.00E-11 4.5901E-11 8.20%
ks,c (m2.5 mol-0.5 s-1) 3.00E-12 2.4436E-12 18.55%
Ēks,a (kJ mol-1) 25 2.5019E+01 0.07%
Ēks,c (kJ mol-1) 28 3.4004E+01 21.44%
ĒDs,a (kJ mol-1) 32 3.4283E+01 7.13%
ĒDs,c (kJ mol-1) 30 2.9182E+01 2.73%
ĒDe (kJ mol-1) 14 1.3890E+01 0.78%
Ēκe (kJ mol-1) 23.5 2.8429E+01 20.98%
λ (W m-1 K-1) 0.7 7.0631E-01 0.90%
h (W m-2 K-1) 25 2.5108E+01 0.43%
average error % 5.73%

The Genetic Algorithm Toolbox [23] was used to program the parameter identification in MATLAB. Since there is only
one test in each step, a single population GA was used. For cases in which there is more than one test in each step due to
complex BPC, the Cooperative Co-evolutionary Genetic Algorithm (CCGA) should be used, and the population partition
should be set according to the parameter clustering results. The model simulation program mentioned in Section 2.3 was
called 3×Ni×MAXGEN times during the whole identification process, which is computationally expensive and takes more
than one week to complete. In order to speed up the identification, a computing cluster with 20 cores was built using an HP
ProLiant DL160 G6 server (2×Intel L5520 @ 2.26GHz CPU, 16G RAM) and five PCs (Intel i3-530 @ 2.93GHz CPU, 4G
RAM) based on the MATLAB Distributed Computing Toolbox. This resulted in a 14-fold increase in the speed of the GA
identification process compared to our previous work [24]. Table 6 shows the preferences of GA for the three steps, where Nd
is the number of variables, Ni is the size of population, MAXGEN is the maximum generation, GGAP is the generation gap,
INSR is the insert rate and Pm is the mutation probability, For details of GA-based parameter identification, the reader is
referred to [25].

Table 6 Preferences of Genetic Algorithm in 3-step parameter identification.


STEP Nd Ni MAXGEN GGAP INSR Pm
1 9 600 50 0.8 0.8 1/9
2 18 600 50 0.8 0.8 1/18
3 27 600 50 0.8 0.8 1/27

Some changes were made to Step 2 and Step 3 during this process. The values of identified parameters were not used
directly, but were identified again over a smaller range [lower’, upper’] defined as follows:
lower   lower  (value  lower )  0.5
 (10)
upper   upper  (upper  value)  0.5

where value is the identified result in previous step and lower and upper represent the searching range in the previous step.
This was deemed necessary as although the operating conditions in Step 1 can decrease parameter sensitivity and reduce the
influence of estimated values on model output, it cannot guarantee the accurate identification of “highly sensitive” parameters.
The values of “highly sensitive” parameters are very close to the true values, which affects the model output slightly, thus
“sensitive” parameters still remain highly identifiable. In the same manner, “highly sensitive” parameters and “sensitive”
parameters were identified over a smaller range in Step 3.
The verification process was carried out as a single blind test in that the set “true values” were unknown to the
researcher carrying out the simulation. In this way the use of synthetic experimental data in the identification process is very
similar to testing a real battery. The identified results using synthetic experimental data are shown in Table 5, and the
“error%” column lists the relative error between the “true value” and the identified result.
As a comparison, a Dynamic Stress Test (DST) and a constant current (CC) test at room temperature were also used for
parameter identification. Fig. 13 shows the comparison of the relative error of each parameter in different tests, and the Avg.
stands for the average error of all parameters in a certain test. In each test, the parameters are identified both stepwise and
synchronously (respectively expressed by “-step” and “-all” in the legend).
In general, identification errors using the stepwise BPC test are lower than others. The highest errors were observed for
Ēks,c (21.44%) and Ēκe (20.98%) because their BCI is far away from the combined BPC, making these two parameters “low
sensitive” when identified. This is demonstrated clearly in the “Clustering results” column in Table 3, in which they
ambiguously belong to Cluster A due to small memberships.
The identification accuracy using the stepwise DST test is also high, albeit lower than that using BPC test. Though there
are many current levels included in the DST, there is only one ambient temperature condition and as a result some “low
sensitive” parameters related to temperature (e.g. activation energy) are unsatisfactorily identified. The CC test contains only
one current and one ambient temperature. It is far away from the BPC, and therefore yields high errors for “low sensitive”
parameters.

Fig. 13. Comparison of the relative errors in parameter identification using different tests.

The “low sensitive” and “sensitive” parameters such as Ds,c, Rfilm,a, De and activation energies are unsatisfactorily
identified using the DST or CC test when identified together, because they have very little influence on the terminal voltage
and surface temperature compared to “highly sensitive” parameters. As a result, the accuracy in identifying all 27 parameters
synchronously is much lower than that using a stepwise experiment. The stepwise experiment designed from BPC described
in this article clearly demonstrates better performance than others in identification accuracy.
It is interesting that although the parameters y0 and εs,c are “low sensitive” they can still be identified accurately in all
tests, because their values are directly related to the battery capacity and initial terminal voltage. This special case shows that
there remains much work to do on understanding parameter sensitivity analysis, particularly with regard to experiment
design.

5. Summary and conclusions


This article has presented a parameter sensitivity analysis of a cylindrical Li-ion battery based on a multi-physics model,
using a carefully designed stepwise experiment. The main conclusions are as follows:
(1) A multi-physics model of a cylindrical Li-ion battery has been developed by coupling a thermal distribution model to
an electrochemical P2D model, allowing simulation of the surface temperature, which can be easily validated through
measurement. 30 parameters are identified in this model, including many which can be used to characterize battery health.
(2) The sensitivity of terminal voltage Uapp and surface temperature Tsh to these parameters were examined using model
simulation under 25 different conditions. A parameter sensitivity matrix was proposed to describe the variation of Uapp and Tsh
curves with 5 different parameter values under a range of conditions and the 30 parameters were then roughly grouped
according to their average and maximum sensitivity. 9 “highly sensitive”, 8 “sensitive” and 10 “low sensitive” parameters
were identified. Three “insensitive” parameters were excluded from further analysis as they cannot be used to study PHM
characteristics.
(3) The sensitivity of a parameter is strongly dependent on operating conditions. A BCI characteristic was proposed to
indicate the operating conditions under which a parameter has the highest sensitivity. The best identification strategy is to
identify each parameter under its own BCI or under conditions similar to its BCI. The parameters were grouped into 3
clusters based on their BCI characteristic using the Fuzzy C-mean Method, with the clustering center defined as BPC, which
is defined as the conditions in which a cluster of parameters are synchronously most sensitive.
(4) Three BPC characteristics were combined because they are similar in the trend and the combined BPC characteristic
could simplify experiment design. A three-step experiment was designed for a 2.3 Ah LiFePO4 battery. In each step, 9
parameters were identified under a simple set of operating conditions. The rationality was validated using model simulation
under each set of conditions. The identification accuracy of stepwise BPC, DST and CC conditions was compared using
synthetic experimental data. The stepwise BPC showed the highest accuracy with an average relative error of 5.73%, while
the accuracy of identifying all 27 parameters synchronously was also higher.

Acknowledgment
This research was financially supported by the National Natural Science Foundation of China (No. 51107021), and the
Fundamental Research Funds for the Central Universities (Grant No. HIT. NSRIF. 2014021).

Appendix A. Electrochemical model

A.1. Faraday’s laws


The electrochemical reaction current density is used to describe the flux of Li + on the intercalation particle of the
electrode:

is  nFjLi (A.1)

where n=1 is the charge number of the lithium ion (all variables are described in the Nomenclature section).
The ionic current density in the solution phase modulation is determined by:

 i2  asis  as FjLi (A.2)

3 s
where as is specific surface area of active material particles defined as as  and the electron current density in the solid
Rs
phase is:

 i1  asis  as FjLi (A.3)

The applied current is equal to the product of electron current density at boundary 1 or 4 and the electrode area:

iapp  Acell i1 1  Acell i1 4 (A.4)

The total current density inside the battery is given by the sum of electron current density and ionic current density:

i  i1  i2 (A.5)

A.2. Charge balance and Ohm’s law


In active materials the solid phase potential is calculated from Ohm’s law:

i1   seff s (A.6)


where the effective conductivity of solid phase  seff   s  s1.5 , s is 0 at boundary 1 and the charge flux is zero at boundaries 2
and 3:

s 1  0

 s i (A.7)
 x    eff
 4 s ,a

s s
 0 (A.8)
x 2  x 3

In the electrolyte, the solution phase potential is determined by Ohm’s law and ionic transport of charges:

2 eeff RT  ln f 
i2   eeff e  (1  )(1  t0 ) ln Ce (A.9)
F  ln Ce

Liquid-junction potential is introduced with the expression:

2 RT  ln f 
K junc  (1  )(1  t0 ) (A.10)
F  ln Ce

and  eeff   e  e1.5 is the effective conductivity of the solution phase. e is taken to be continuous at boundaries 2 and 3:

e 2  e 2
 (A.11)
e 3  e 3

while it is set to zero at boundaries 1 and 4:

e 
 e 0 (A.12)
x 1 x 4

A.3. Mass balance and transport process


The mass balance of Li+ in a spherical particle of active material is described by Fick’s law:

Cs 1
 2  ( Ds r 2Cs ) (A.13)
t r

where r is the distance from the center of the particle. The lithium concentration at the surface is coupled to Ce/s and the
reaction current density, while the Li+ flux is set to zero at the center of the particle, because there is no species source:

 Cs jLi
 r 
 r  Rs Ds
 (A.14)
 Cs 0
 r r 0

The Li+ concentration in solution, which varies due to diffusion and reaction at the electrode/electrolyte interface, is
calculated from the following equation:

Ce a
e   ( Deeff Ce )  s (1  t0 )  is (A.15)
t F

Diffusion takes place in three regions (anode, separator and cathode). Current collectors present an impermeable wall to
the electrolyte, so the Li+ flux is zero at external boundaries 1 and 4 and shows a continuity condition at inner boundaries 2
and 3.

Ce Ce
 x  x  0
 1 4

 eff Ce Ce


  De ,a   Deeff, s (A.16)
 x 2  x 2
 eff C Ce
  De , s e
  Deeff,c
 x 3 x 3

At the initial time, the electrolyte concentration is equal to a constant value Ce,0.

A.4. Electrochemical kinetics


The electrochemical reaction occurring at the surface of the active material is described by the Butler-Volmer equation:

a F c F
is  nFjLi  i0 [exp(  )  exp(  )] (A.17)
RT RT

where αa and αc are the transfer coefficients of anode and cathode respectively and  is the overpotential, which is defined
as:

  s  e  Eocv  is R film (A.18)

Rfilm is the resistance of the solid electrolyte interface (SEI) film and Eocv is the open circuit potential, which is a function of
the solid phase Li+ concentration at the particle surface. The Eocv of the MCMB/LiFePO4 system is given by [19]:

x  0.1958
Eocv ,a  0.6379  0.5416exp(305.5309 x )  0.044 tanh( )
0.1088
(A.19)
x  1.0571 x  0.0117 x  0.5692
0.1978 tanh( )  0.6875 tanh( )  0.0175 tanh( )
0.0854 0.0529 0.0875
Eocv ,c  3.4323  0.8428exp(80.2493(1  y )1.3198 )
(A.20)
3.2474  106 exp(20.2645(1  y )3.8003 )  3.2482  106 exp(20.2646(1  y )3.7995 )

Ce / s ,n Ce / s , p
where x  is the stoichiometric number for the anode and y  is the stoichiometric number for the cathode.
Cs , n
MAX
CsMAX
,p

The exchange current density i0 acts as a bridge connecting concentrations in both solid and liquid phases, and indicates
the ease of electrode polarization. It may be calculated as follows:

i0  ksCea (Cs ,MAX  Ce / s )a Ce/cs (A.21)

where k s is the electrochemical reaction rate.


All of the above equations should be solved by a numerical method, after which the terminal voltage of the battery is
calculated by:

Uapp  s 4  s 1  iRext (A.22)

where Rext is the contact resistance between electrode and current collector.

Appendix B. Thermal model

B.1. Energy balance of the battery


The energy balance equation contains three terms, namely the heat capacity, the heat conduction and the heat generation,
respectively.

T
Cp  (T )  Q (B.1)
t

where  is the volume-averaged density of the battery, C p is the volume-averaged heat capacity and  is the effective
thermal conductivity.

B.2. Heat generation


The heat generation rate is defined by Q  Qrea  Qent  Qohm , comprising electrochemical reaction heat, entropic heat
and ohmic heat. These are calculated using following equations.
Electrochemical reaction heat:

Qrea  asis (s  e  Eocv ) (B.2)

Entropic heat:

dEocv
Qent  asisT (B.3)
dT

dEocv
where is the entropy coefficient of the electrode, related to the stoichiometric number. The entropy coefficient of the
dT
MCMB/LiFePO4 system can be obtained from [18].
Ohmic heat can be divided into three parts:

Qohm  Qohm,s  Qohm,e  Qext (B.4)

The first term represents solid phase ohmic heat, generated by the electron current in the active material, the second term is
the sum of solution phase ohmic heat and ionic migration heat and the third term is the ohmic heat associated with contact
resistance:

Qohm, s   seff s s




Qohm,e   e e e   e  ln Ce e
eff D , eff
(B.5)

Qext  iapp Rext L
2

2 RT  eeff 0
where L is the total thickness of the battery and  eD ,eff  (t  1) .
F

B.3. Heat exchange


Heat exchange between the battery and ambient is governed by convection and radiation. Heat convection is expressed
by Newton’s cooling law:

qc  h(Tsh  Ta ) (B.6)

while heat radiation is expressed by the Stefan-Boltzman equation:

qr   (Tsh4  Ta4 ) (B.7)

where h is the heat transfer coefficient,  is the emissivity, σ is the Stefan-Boltzman constant, Tsh is the shell temperature of
the battery and Ta is ambient temperature.

B.4. Temperature dependent parameters


Some parameters in the P2D model, such as the solid phase diffusion coefficients, Ds,a and Ds,c, solution phase
conductivity, κe, solution phase diffusion coefficient, De, and electrochemical reaction rates, ks,a and ks,c, are coupled with the
battery temperature. These parameters can be updated using the Arrhenius law:

Ei 1 1
X i  X i ,ref exp[ (  )] (B.8)
R Tref T

where Xi represents the parameters mentioned above, Xi,ref is the parameter value at the reference temperature, Tref, and Ei
denotes the corresponding activation energy of parameter Xi.
In addition, the open circuit potential depends on battery temperature according to the Nernst equation:

dEocv
Eocv  Eocv
ref
 (T  Tref ) (B.9)
dT

ref
where Eocv is the open circuit potential at the reference temperature.

Nomenclature
as: specific surface area (m-1)
Acell: area of electrode (m2)
Ce: electrolyte concentration (mol m-3)
Cs: concentration of Li+ in the intercalation particle (mol m-3)
Ce/s: concentration of Li+ at the surface of intercalation particle (mol m-3)
Cp: specific heat capacity (J kg-1 K-1)
De: electrolyte diffusion coefficient (m2 s-1)
Ds: Li+ diffusion coefficient in active material (m2 s-1)
Ē: activation energy (kJ mol-1)
Eocv: open circuit potential (V)
F: Faraday's constant (=96487 C mol-1)
h: heat transfer coefficient (W m-2 K-1)
i: total current density in electrodes or separator (A m-2)
iapp: applied current (A)
i0: exchange current density (A m-2)
i1: solid phase current density (A m-2)
i2: solution phase current density (A m-2)
is: electrochemical reaction current density (A m-2)
jLi: flux of Li+ on the intercalation particle of electrode (mol m-2 s-1)
ks: electrochemical reaction rate (m2.5 mol-0.5 s-1)
L: thickness of electrode or separator (m)
q : heat exchange rate (W m-2)
qc : heat convection rate (W m-2)
qr : heat radiation rate (W m-2)
Q: heat generation rate (W m-3)
R: ideal gas constant (=8.3143J mol-1 K-1)
Rext: contact resistance (Ω m2)
Rfilm: SEI film resistance (Ω m2)
Rs: intercalation particle radius of electrode (m)
t0 : Li+ transference number in the electrolyte
T: battery temperature (K)
Ta: ambient temperature (K)
Uapp: applied potential (V)
x: stoichiometric number of anode
y: stoichiometric number of cathode

Greek symbols
α: transfer coefficient of the electrochemical reactions
δ: thickness of current collector (m)
ε: emissivity
ε e: volume fraction of electrolyte in electrode or separator
εs: volume fraction of active material in electrode
η: overpotential (V)
κ e: ionic conductivity of electrolyte (S m-1)
λ: thermal conductivity (W m-1 K-1)
ρ: density (kg m-3)
σ: Stefan-Boltzman constant (=5.6704×10-8 W/m2/K4)
σs: solid phase conductivity (S m-1)
ϕs: solid phase potential (V)
ϕe : solution phase potential (V)

Subscript
a: anode
c: cathode
s: separator
0: initial value
MAX: maximum value
ref: reference value

References
[1] M. Doyle, T.F. Fuller, J. Newman, J. Electrochem. Soc. 140 (1993) 1526-1533. (doi:10.1149/1.2221597)
[2] M. Doyle, T.F. Fuller, J. Newman, J. Electrochem. Soc. 141 (1993) 1-10. (doi:10.1149/1.2054684 )
[3] M. Doyle, J. Newman, J. Electrochem. Soc. 143 (1996) 1890-1903. (doi:10.1149/1.1836921)
[4] Q. Zhang, R.E. White, Journal of Power Sources 179 (2008) 785-792. (doi:10.1016/j.jpowsour.2007.12.022)
[5] Q. Zhang, R.E. White, Journal of Power Sources 179 (2008) 793-798. (doi:10.1016/j.jpowsour.2008.01.028)
[6] A.P. Schmidt, M. Bitzer, A.W. Imre, L. Guzzella, Journal of Power Sources 195 (2010) 7634-7638.
(doi:10.1016/j.jpowsour.2010.06.011)
[7] V. Ramadesigan, D. Chen, N.A. Burns, V. Boovaragavan, R.D. Braatz, V.R. Subramanian, J. Electrochem. Soc. 158
(2011) A1048-A1054. (doi: 10.1149/1.3609926)
[8] J.C. Forman, S.J. Moura, J.L. Stein, H.K. Fathy, Journal of Power Sources 210 (2012) 263-275.
(doi:10.1016/j.jpowsour.2012.03.009)
[9] J.C. Forman, S.J. Moura, J.L. Stein, H.K. Fathy, American Control Conference (2011) 362-369.
[10] A.P. Schmidt, M. Bitzer, A.W. Imre, L. Guzzella, Journal of Power Sources 195 (2010) 5071-5080.
(doi:10.1016/j.jpowsour.2010.02.029)
[11] C.H. Min, Y.L. He, X.L. Liu, B.H. Yin, W. Jiang, W.Q. Tao, Journal of Power Sources 160 (2006) 347-385.
(doi:10.1016/j.jpowsour.2006.01.080)
[12] X.D. Wang, J.L. Xu, D.J. Lee, International Journal of Hydrogen Energy 37 (2012) 15766-15777.
(doi:10.1016/j.ijhydene.2012.04.029)
[13] G.N Srinivasulu, T. Subrahmanyam, V.D. Rao, International Journal of Hydrogen Energy 36 (2011) 14838-14844.
(doi:10.1016/j.ijhydene.2011.03.040)
[14] V. Boovaragavan, S. Harinipriya, V.R. Subramanian, Journal of Power Sources 183 (2008) 361-365
(doi:10.1016/j.jpowsour.2008.04.077)
[15] X. Zhang, Electrochimica Acta 56 (2011) 1246–1255. (doi:10.1016/j.electacta.2010.10.054)
[16] K.E. Thomas, J. Newman, R.M. Darling, Advances in Lithium-Ion batteries Chapter 12, Kluwer Academic Publishers,
New York, 2002.
[17] M. Fleckenstein, S. Fischer, O. Bohlen, B. Baker, Journal of Power Sources 223 (2013) 259-267.
(doi:10.1016/j.jpowsour.2012.07.144)
[18] Y. Ye, Y. Shi, A.A.O. Tay, Journal of Power Sources 217 (2012) 509-518. (doi:10.1016/j.jpowsour.2012.06.055)
[19] M. Safari, C. Delacourt, J. Electrochem. Soc. 158 (2011) A562-A571. (doi: 10.1149/1.3567007)
[20] E. Prada, D. Di Domenico, Y. Creff, J. Bernard, V. Sauvant-Moynot, F. Huet, J. Electrochem. Soc. 159 (2012)
A1508-A1519. (doi: 10.1149/2.064209jes)
[21] L. Cai, Y. Dai, M. Nicholson, R.E. White, K. Jagannathan, G. Bhatia, Journal of Power Sources 221 (2013) 191-200.
(doi:10.1016/j.jpowsour.2012.08.046)
[22] K.L. Huang, Z.Z. Lu, S.Q. Liu, Battery Bimonthly 31 (2001) 142-145.
[23] A. J. Chipperfield, P. J. Fleming, C. M. Fonseca, Proc. Adaptive Computing in Engineering Design and Control 21
(1994) 128-133.
[24] L. Zhang, C. Lyu, L. Wang, J. Zheng, W. Luo, K. Ma, Advances in Mechanical Engineering (2013) Article ID: 754653.
(doi: 10.1155/2013/754653)
[25] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison Wesley Publishing
Company, January 1989.

View publication stats

You might also like