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

net/publication/283129079

State-of-charge estimation for lithium-ion batteries via a coupled thermal-

electrochemical model

Conference Paper in Proceedings of the American Control Conference · July 2015

DOI: 10.1109/ACC.2015.7172260

CITATIONS READS

21 559

6 authors, including:

Shu-Xia Tang Yebin Wang

University of California, Berkeley Mitsubishi Electric Research Laboratories
38 PUBLICATIONS 897 CITATIONS 132 PUBLICATIONS 2,534 CITATIONS

SEE PROFILE SEE PROFILE

Zafer Sahinoglu Toshihiro Wada

Mitsubishi Electric Research Laboratories Mitsubishi Electric Corporation
112 PUBLICATIONS 6,915 CITATIONS 25 PUBLICATIONS 821 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Shu-Xia Tang on 20 June 2018.

The user has requested enhancement of the downloaded file.

 Automatica 83 (2017) 206–219

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

State-of-Charge estimation from a thermal–electrochemical model of

lithium-ion batteries✩
Shu-Xia Tang a , Leobardo Camacho-Solorio a , Yebin Wang b , Miroslav Krstic a
a
Department of Mechanical & Aerospace Engineering, University of California, San Diego, La Jolla, CA 92093-0411, USA
b
Mitsubishi Electric Research Laboratories, 201 Broadway, Cambridge, MA 02139, USA

article info a b s t r a c t
Article history: A thermal–electrochemical model of lithium-ion batteries is presented and a Luenberger observer is
Received 4 August 2016 derived for State-of-Charge (SoC) estimation by recovering the lithium concentration in the electrodes.
Received in revised form 10 February 2017 This first-principles based model is a coupled system of partial and ordinary differential equations,
Accepted 19 May 2017
which is a reduced version of the Doyle–Fuller–Newman model. More precisely, the subsystem of Partial
Differential Equations (PDEs) is the Single Particle Model (SPM) while the Ordinary Differential Equation
(ODE) is a model for the average temperature in the battery. The observer is designed following the PDE
Keywords:
Infinite dimensional systems backstepping method. Since some coefficients in the coupled ODE–PDE system are time-varying, this
Lithium-ion batteries results in the time dependency of some coefficients in the kernel function system of the backstepping
Battery management systems transformation and it is non-trivial to show well-posedness of the latter system. Adding thermal dynamics
PDE backstepping to the SPM serves a two-fold purpose: improving the accuracy of SoC estimation and keeping track
of the average temperature which is a critical variable for safety management in lithium-ion batteries.
Effectiveness of the estimation scheme is validated via numerical simulations.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction required that the SoC remains within appropriate bounds all the
time during the battery operation for safety reasons. Hence, a
1.1. Motivation reliable and accurate estimation of the SoC is required for proper
battery management.
Due to its high power and energy storage density, its lack of
memory effect and low self discharge, lithium-ion technology is 1.2. Lithium-ion battery models
a common choice among the rechargeable battery family (Chu &
Majumdar, 2012). Besides its wide employment in portable elec- Accuracy of the SoC estimation depends highly on the quality
tronics, lithium-ion batteries are now being adopted in electrified of the selected model. Thus, one is encouraged to compare the dif-
transportation (Stewart, Christensen, Chaturvedi, & Kojic, 2015) ferent models available for describing the battery dynamics. Mod-
such as electric vehicles and hybrid electric vehicles. Lithium-ion els for lithium-ion batteries can be categorized into two classes.
technology is being considered for grid energy storage as well. The first class consists of empirical models, in which the most
The key indicator for the amount of electrical energy available in frequently used ones are Equivalent Circuit Models (ECMs) (Chi-
batteries is the SoC which, simply put, is the ratio of instantaneous asson & Vairamohan, 2005; Plett, 2004). ECMs use electric circuit
remaining battery charge to its maximum capacity (Chaturvedi, elements such as voltage sources, resistances and RC networks to
Klein, Christensen, Ahmed, & Kojic, 2010). Thus, in order to predict approximate the dynamics of the battery. Currently, most battery
the available power and energy in the battery during operation, management systems employ ECMs for various tasks: power and
online estimation of the SoC serves as an important factor for energy estimation, cell balancing, thermal management, state-of-
regulating both charging and discharging. Besides, it is generally health estimation and charge/discharge control. The second class
of models are based on first principles (Ramadesigan, Northrop,
✩ A part of this work was carried out during S.-X. Tang’s visit to Mitsubishi Electric De, Santhanagopalan, Braatz, & Subramanian, 2012). These electro-
Research Laboratories, USA. The material in this paper was partially presented at chemical models account for the main underlying physics in the
the 2015 American Control Conference, July 1–3, 2015, Chicago, IL, USA. This paper battery, more precisely, they offer an explicit description of the
was recommended for publication in revised form by Associate Editor Yoshihiko
Miyasato under the direction of Editor Thomas Parisini.
battery dynamics in terms of the main electrochemical parame-
E-mail addresses: [email protected] (S.-X. Tang), [email protected] ters and variables. The need for accurate SoC estimation as well
(L. Camacho-Solorio), [email protected] (Y. Wang), [email protected] (M. Krstic). as visibility of important electrochemical states and parameters,

http://dx.doi.org/10.1016/j.automatica.2017.06.030
0005-1098/© 2017 Elsevier Ltd. All rights reserved.
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 207

specially in high power and high energy applications, motivates the design estimation and control algorithms for lithium-ion batteries
study of estimation based on electrochemical models. from electrochemical models without relying on the discretization
The widely studied electrochemical Doyle–Fuller–Newman of the PDEs in these models. The main technical challenges in
(DFN) model has been shown to accurately describe the main phe- our design consist of proving the well-posedness of the kernel
nomena in lithium-ion batteries (Chaturvedi et al., 2010; Doyle, function system for the backstepping transformation. The fact that
Fuller, & Newman, 1993). However, the complexity of the model is some coefficients in the thermal–electrochemical model system
too high for online SoC estimation (Klein, Chaturvedi, Christensen, are time-varying results in a kernel function system with time-
Ahmed, Findeisen, & Kojic, 2013). Among the various approxima- varying coefficients, for which the well-posedness is non-trivial
tions to the DFN model, the SPM (Guo, Sikha, & White, 2011; Haran, to derive. This paper is a continuation of a previous result for SoC
Popov, & White, 1998) is commonly used to derive online SoC estimation from a thermal–electrochemical model of lithium-ion
estimation algorithms (Moura, Chaturvedi, & Krstic, 2014). In the batteries in Tang et al. (2015).
SPM, diffusion of lithium ions in each electrode is simplified as dif-
fusion in a single spherical particle and electrolyte concentration is 1.5. Organization
assumed to be constant. Still, the SPM has several limitations, for
example, being accurate only at low currents (Chaturvedi et al., The rest of this paper is organized as follows. In Section 2,
2010). Another limitation is that the SPM is restricted to the cases a temperature-compensated SPM model is presented; and the
with small variation in internal temperature, which comes from corresponding SoC estimation problem is formulated in Section 3.
the fact that SPM ignores the dependence of the battery parameters In Section 4, a linear observer is developed for estimation of the
on temperature. In fact, lithium-ion batteries meet issues such lithium concentration in the electrodes through boundary state
as an increase in internal resistance and decrease of capacity, as measurements via the backstepping method. The observer error
functions of battery internal average temperature (Guo et al., system is proved to be exponentially stable with an arbitrarily
2011; Thomas, Newman, & Darling, 2002). designated decay rate, for which the well-posedness is derived by
making use of the abstract evolution equation theory. It is worth
1.3. Estimation algorithms noting that solving the kernel function system for the backstepping
transformation is not trivial because of its dependence on the
Extensive efforts have been devoted to developing SoC estima- temperature (Izadi & Dubljevic, 2015; Smyshlyaev & Krstic, 2005).
tion algorithms, for example, Extended Kalman Filters (EKFs) for Under some more relaxed assumptions and simplifications than
ECMs (Plett, 2004) and for the SPM (Santhanagopalan & White, those in Tang et al. (2015), the existence and regularity of the so-
2006). Estimation algorithms have also been derived for reduced lution to the system are proved in this section. The SoC estimation
electrochemical models with temperature dynamics, e.g., a linear accuracy is verified by the numerical simulation results presented
observer derived to satisfy the conservation of lithium ions (Klein in Section 5. Finally, some concluding remarks and possible future
et al., 2013) and a linear observer using pole placement (Tanim, research topics are given in Section 6.
Rahn, & Wang, 2015). These estimation algorithms, together with
others based on the unscented Kalman filter or particle filters, rely 2. SPM-T model
on some discretization of the diffusion phenomena.
Discretization generally implies a trade-off between high ac- In this section, the working mechanism of lithium-ion batteries
curacy of the approximation, i.e., a large number of states, and a is briefly introduced through an overview of the DFN model. Then,
small number of tuning gains in the observer, i.e. a small number the single particle model with temperature dynamics (Guo et al.,
of states. The backstepping approach can be employed to design 2011; Tang et al., 2015), named SPM-T model, is presented for
boundary observers for PDEs in which the discretization is not the purpose of SoC estimation, which can be viewed either as a
required. The readers can refer to Smyshlyaev & Krstic (2003) for simplification of the DFN model or as temperature-compensated
a preliminary example of boundary observer design for diffusion SPM.
PDEs via backstepping. This method has been used for the stabiliza-
tion of various unstable PDE systems, see the tutorial book (Krstic 2.1. Working principles of lithium-ion batteries
& Smyshlyaev, 2008), in which backstepping boundary controllers
and observers are designed for some unstable parabolic, hyperbolic A lithium-ion battery cell consists of three main regions: nega-
PDEs and other types of PDEs. It has also been applied for stabilizing tive electrode, separator and positive electrode; all of them char-
some coupled PDE–ODE systems (Tang & Xie, 2011a,2011b). acterized by a porous structure. Each electrode includes active
materials, conductive fillers, a current collector and a binder. The
1.4. Contribution porous structure of the electrodes provides a large surface area and
small distances between lithium ions and active material surfaces
The main contribution of this paper is the derivation of a for reactions to occur. The separator is placed between the negative
linear observer for SoC estimation from a simplified thermal– and positive electrodes to forbid the flow of electrons between two
electrochemical model of lithium-ion batteries, i.e., a coupled electrodes while allowing the movement of lithium ions dissolved
ODE–PDE model composed by the SPM and a model for the in the electrolyte. The active materials, intercalated in the lattices
averaged internal temperature (Guo et al., 2011; Tang, Wang, of the corresponding electrode, are insertion compounds, i.e. these
Sahinoglu, Wada, Hara, & Krstic, 2015). Adding thermal dynamics are host structures in which lithium can be reversibly inserted or
to SPM serves a two-fold purpose: improving the accuracy of SoC extracted. Electrolyte fills all remaining parts of the battery.
estimation and keeping track of the average temperature which is The DFN model is derived based on the porous structure all
a critical variable for safety management in lithium-ion batteries. through the lithium-ion battery (Chaturvedi et al., 2010; Thomas
The observer is designed following the PDE backstepping et al., 2002). In the DFN model, each electrode is viewed as super-
method. It is worth noting that backstepping observers have not position of active materials, inert filler and electrolyte; justified by
been introduced to the problem of battery SoC estimation until the porous configuration. As depicted in Fig. 1, all intercalation par-
very recently (Moura et al., 2014), and by this means the discretiza- ticles are assumed to be spheres with a uniform, averaged radius,
tion of the diffusion PDEs in the model is avoided. We consider the and the battery is formulated as a pseudo two-dimensional model.
result presented in this paper as an additional step in the efforts to The first dimension represents the path along the spatial direction
208 S.-X. Tang et al. / Automatica 83 (2017) 206–219

Table 1
Nomenclature.
Variables
cs Lithium concentration in the solid phase
css Lithium concentration at the surface of the particle
ce Constant lithium concentration in the electrolyte
c̄s Volume averaged lithium concentration in the solid phase
j Molar flux of lithium at the surface of the particle
φs Electric potential in the solid phase
φe Electric potential in the electrolyte
η Reaction overpotential
U Open-circuit potential
i0 Exchange current density
ie Electrolyte current density normalized by cross-sectional area
T Internal average temperature
Tamb Ambient temperature
I External current density normalized by cross-sectional area
V Terminal voltage
q̄s Volume averaged flux
Fig. 1. DFN schematic.
Parameters
L Length
Ds Diffusion coefficient of lithium in the solid phase
x from the anode, through the separator, to the cathode; and the De Diffusion coefficient of lithium in the electrolyte
second dimension is a radial direction rs used to represent the csmax Maximum lithium concentration in the solid phase
intercalation and diffusion of lithium ions in the active materials Rs Radius of the particle
αa Anodic transfer coefficient
(see Table 1).
αc Cathodic transfer coefficient
Lithium ions move from the negative electrode to the positive reff Effective reaction rate in the solid phase
electrode during discharging and in the opposite direction during Rf Film resistance of the solid-electrolyte interphase
charging. Lithium concentration in the solid phase, i.e. concentra- Rc Contact resistance between the electrode and current collector
tion of lithium ions in the active materials, follows the Fick’s law of εs Volume fraction of the active material
εe Volume fraction of the electrolyte
diffusion: as Interfacial surface area
∂ cs± 1 ∂ 2 ∂ cs
±
[ ]
F Faraday’s constant
(t , x, rs ) = 2 D± (T (t))r (t , x , rs ) , R Universal gas constant
∂t rs ∂ rs s s
∂r NLi,s Total number of lithium ions in the solid phase
( ± ±) s σ
t > 0, x ∈ 0 , L , rs ∈ 0, Rs ,
( ±) Electronic conductivity in the solid phase
(1)
κ Ionic conductivity in the electrolyte
∂ cs±
tc0 Transference number of the ions w.r.t. the solvent velocity
(t , x, 0) = 0, t > 0, x ∈ 0± , L± ,
( )
(2) fc/a Mean molar activity coefficient in the electrolyte
∂ rs ρ avg Average density
∂ cs± 1 cP Heat capacity
(t , x, R± j± (t , x), t > 0, x ∈ 0± , L± ,
( )
s ) = − ± (3)
∂ rs Ds (T (t)) hcell Heat transfer coefficient
E Activation energy coefficient
cs± (0, x, rs ) = cs0±
(x, rs ), x ∈ 0± , L± , rs ∈ 0, R± s ,
[ ] [ ]
(4)
Super- and subscripts
where the temporal variable is t, the spatial variables are x and rs . + Positive electrode
The states of the PDE model (1)–(4) are cs± (t , x, rs ) ∈ R; the solid − Negative electrode
sep Separator
phase lithium concentration. The molar fluxes j± (t , x) are related to
s Solid phase
the reaction overpotential η± (t , x) by the Butler–Volmer equation e Electrolyte

0 (t , x)
i± αa F αc F ±
η± (t ,x) η (t ,x)
[ ]
− RT
j± (t , x) = e RT (t) −e (t) .
F
The reaction overpotentials η± (t , x) are computed from where I(t) is the external current density normalized by cross-
sectional area. Charge conservation in the electrodes provides a
η (t , x) = φs (t , x) − φe (t , x) − U (css (t , x), T (t))
± ± ± ± ±
e (x, t) and molar
relation between electrolyte current densities i±
− FR± f (T (t))j (t , x).
±
fluxes j± (t , x):
Lithium concentration in the liquid phase ce (t , x), i.e. concentration ∂ i+
of lithium ions in the electrolyte, satisfies the diffusion equation
e
(t , x) = −a+s Fj (t , x),
+
∂x
∂ ce ∂
[
∂ ce 1 − tc0
] ∂ i−
(t , x) = De (t , x) + ie (t , x) . (5)
e
(t , x) = a−
s Fj (t , x),
−
∂t ∂x ∂x εe F ∂x
with boundary conditions i− e (t , 0 ) e (t , L )
−
Equations for solid electric potential φs (t , x) and electrolyte electric = i+ +
= 0 and
potential φe (t , x) are i−
e (t , L −
) = i+
e (t , 0+
) = I(t). In the separator, ie (t , x) = I(t). Output
voltage is the difference between the two solid electric potentials
∂φs ie (t , x) − I(t)
(t , x) = , (6) computed as
∂x σ
∂φe ie (t , x) V (t) = φs (t , L+ ) − φs (t , 0− ).
(t , x) = −
∂x κ
The readers should refer to Chaturvedi et al. (2010) for a complete
∂ ln ce
( )
2RT (t) d ln fc/a
+ (1 − tc0 ) 1 + (t , x) (t , x), description of the DFN model and for the boundary conditions
F d ln ce ∂x for Eqs. (5)–(7). Note that we are using the convention: positive
(7) current for discharging and negative current for charging.
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 209

± ±
2.2. The SPM-T model Parameters D± s (T (t)), reff (T (t)) and Rf (T (t)) are functions with an
Arrhenius-like dependence (Klein et al., 2013) on the battery cell
The DFN model accurately describes many aspects of the internal average temperature T (t), i.e.,
lithium-ion cells working mechanism; however, the complexity of T (t)−T (0)
E ± T (t)T (0)
the model is too high for online SoC estimation. For this reason, D± ±
s (T (t)) = Ds (T (0))e
Ds
, (14)
we present a simplified model which is the single particle model T (t)−T (0)
E ± T (t)T (0)
with temperature dynamics, i.e., the SPM-T model. The SPM-T
±
reff ±
(T (t)) = reff (T (0))e r
eff , (15)
model (Guo et al., 2011; Tang et al., 2015) is derived by making T (t)−T (0)
E ± T (t)T (0)
the following assumptions and simplifications: R± ±
f (T (t)) = Rf (T (0))e
R
f , (16)
where ED± , Er ± , ER± , are activation energy coefficients. Internal
• concentration of lithium ions in the electrolyte ce (t , x) is s eff f

uniform in both time and space, average temperature satisfies the following ODE (Thomas et al.,
2002, Section 12.3.7)
• molar fluxes j± (t , x) are uniform in the x-direction,
• concentration of lithium ions in the active materials cs (t , x) dT
ρ avg cP (t) = hcell (Tamb (t) − T (t)) − I(t)V (t)
is uniform in the x-direction. dt {

Moreover, each electrode is modeled as a single spherical par- + I(t) U + (c̄s+ (t), T (t)) − U − (c̄s− (t), T (t))
ticle in this simplification; representative of all particles in the
∂ U (c̄s (t), T (t)) ∂ U − (c̄s− (t), T (t))
[ + + ]}
electrode. Compared with the SPM-T model presented in Guo et
− T (t) −
al. (2011), here we choose not to take into account the electrolyte ∂T ∂T
resistance Rcell , i.e., we set Rcell = 0 (Klein et al., 2013). + Rc I(t)2 , t > 0, (17)
In the coupled SPM-T model, the SPM subsystem is
T (0) = Tamb (0), (18)
∂ cs± 1 ∂ 2 ∂ cs
±
[ ]
(t , rs ) = 2 D± (T (t))r (t , r s ) , ±
where c̄s (t) are the average concentrations defined as
∂t rs ∂ rs s s
∂ rs
R±
t > 0, rs ∈ (0, R± s ),
∫
(8) 3 s
c̄s± (t) = rs2 cs± (t , rs )drs .
∂ cs± (R±
s )
3
0
(t , 0) = 0, t > 0, (9)
∂ rs The system states are the solid phase lithium ion concentrations
∂ cs± 1 cs± (t , rs ) ∈ R in the PDE (8)–(11) and the internal average temper-
(t , R±s ) = − ± j± (t), t > 0, (10) ature T (t) in the ODE (17)–(18).
∂ rs Ds (T (t))
cs± (0, rs ) = cs±,0 (rs ), rs ∈ [0, R±s ]. (11) 3. Problem formulation
The states of the system (8)–(11) are cs (t , rs ) ∈ R, with the tem-
±

poral variable t and the spatial variables rs . The relation between 3.1. Estimation objective
molar fluxes j± (t) and current I(t) becomes linear:
Our objective is to estimate the battery SoC, defined as the nor-
I(t) I(t) malized averaged lithium concentration in the negative electrode,
j+ (t) = − , j− (t) = .
a+
s FL
+ a−
s FL
− i.e.,
R−
By assuming the same value to the anodic and cathodic transfer cs− (t , rs ) c̄s− (t)
∫
3 s

coefficients, i.e., α ≜ αa = αc , a simple relation between reaction SoC(t) = rs2 drs = , (19)
(R− )3
s 0 cs−,max cs−,max
overpotentials and molar fluxes can be found as
( ) from measurements of the input current I(t) and the output volt-
RT (t) F age V (t). For this purpose, boundary observers can be constructed
η± (t) = sinh−1 j± (t) ,
αF 2i±
0 (t) to estimate the concentrations of lithium ions in the electrodes
cs− (t , rs ) by using the boundary values, i.e., the corresponding sur-
where −
face concentrations css (t). Note however that in (13), boundary
]αc [ )]αa −
i± ± ±
ce,0 cs±,max − css
±
, values css (t) are not directly available from measurement of V (t).
[ (
0 (t) = reff (T (t)) css (t) (t) (12)
Instead, they appear in V (t) with a nonlinear fashion, i.e., the
±
with css (t) ≜ cs± (t , R±
s ). The parameter ce,0 in Eq. (12) denotes the nonlinearities within the exchange current densities i− 0 (t) and
electrolyte concentration at equilibrium. Solid electric potentials within the subtraction between the nonlinear Open-Circuit Poten-
are computed from tial (OCP) functions U − (t , css −
(t)). Therefore, in order to overcome
the lack of boundary value measurements required by the bound-
φs± (t) = η± (t) + U ± (css± (t), T (t)) + FR±
f (T (t))j (t).
±
ary observers, an inversion of the output function V (t) with respect
Output voltage is now the difference between solid electric poten- to the boundary values is needed.
tials in the positive electrode and negative electrode: To ease inversion of the output voltage, the lithium concentra-
tion dynamics in one of the electrodes will be simplified. Inversion
V (t) = φs+ (t) − φs− (t) will then be done with respect to the surface concentration of the
[ ( )
RT (t) 1 I(t) electrode with unsimplified lithium concentration dynamics. The
=− sinh−1 leading terms in the output voltage are the OCP functions and we
αF 2i+
0 (t)
a+ L +
( )] are assuming that the OCP functions are invertible with respect
−1 1 I(t) to the surface concentration in the corresponding unsimplified
+ sinh
2i−0 (t)
a− L − electrode. In this paper, we will simplify the lithium concentration
( +
Rf (T (t)) −
Rf (T (t))
) dynamics in the negative electrode and invert the output function
− + +
+ − −
I(t) with respect to the surface concentration in the positive electrode,
a L a L
and a boundary observer will then be derived for estimation of
+ U + (css+ (t), T (t)) − U − (css− (t), T (t)). (13) lithium ion concentration in the positive electrode.
210 S.-X. Tang et al. / Automatica 83 (2017) 206–219

Remark 1. The decision of simplifying the negative electrode dy- Thus, we can rewrite Eq. (22) as
namics instead of the positive one is made based on the sensitivity
dŤ
of OCPs to surface concentrations. For some common lithium ion ρ avg cP (t) = χ (t)Ť (t) + ω(t), (23)
active materials, ∂ U /∂ css+ is larger that ∂ U /∂ css− in magnitude, thus
+ +
dt
+
making it easier to recover css from the voltage measurement V (t). where
[ ( )
One can easily prove that the total amount of lithium ions in R 1 I(t)
solid phase NLi,s is conserved (Klein et al., 2013), i.e., χ (t) = − hcell + I(t) sinh −1
αF + + +
2iˇ0 (t) a L
dNLi,s ( )]
= 0, (20) 1 I(t)
dt + sinh−1 − − −
where 2iˇ0 (t) a L
[ ]
NLi,s = ε + L+ c̄s+ (t) + ε − L− c̄s− (t). ∂ Ǔ2+ ∂ Ǔ2−
− I(t) (t) − (t) ,
Since we assume NLi,s is a known quantity, i.e. a parameter in the
∂T ∂T
( )
model, then we can also compute the battery SoC in the negative Ř+ Ř−
f (t) f (t)
electrode from the averaged lithium concentration in the positive ω(t) = hcell Tamb (t) + + − Rc I(t)2
electrode, i.e. a+ L + a− L −
[ ] [ ]
NLi,s − ε + L+ c̄s+ (t) − I(t) Ǔ1+ (t) − Ǔ1− (t) + I(t) Ǔ2+ (t) − Ǔ2− (t) ,
SoC(t) = . (21)
ε − L− cs−,max
then, it holds that
1
∫t
avg c χ (τ )dτ
3.2. Output function inversion Ť (t) = Ť (0)e ρ P
0

∫ t ∫ t −τ
1 1
χ (σ )dσ
The goal of the output function inversion is to write V (t) as a + eρ
avg c
P
0
ω(τ )dτ . (24)
+
function of only of css (t) and I(t). ρ avg c
P 0

Substituting (24) into (13) yields the following simplified output

±
3.2.1. Write V (t) as a function of css (t), c̄s± (t) and I(t) function:
The first step is to simplify the internal average temperature [ ( )
RŤ (t) 1 I(t)
dynamics to derive an expression for T (t) only in terms of time V (t) = − sinh−1
and current, i.e., a time-varying function Ť (t) ≜ Ť (t , I(t), Tamb (t)) αF 2i+
0 (t)
a+ L +
independent of the concentrations cs± (t), css ±
(t). We start by substi-
( )]
1 I(t)
tuting the output voltage equation (13) into the original average + sinh−1
2i−
0 (t)
a− L −
temperature equation (17): ( )
dT R+
f (Ť (t)) R−
f (Ť (t))
ρ avg cP (t) = hcell (Tamb (t) − T (t)) − + I(t)
dt a+ L + a− L−
[ ( ) ( )]
RT (t) 1 I(t) 1 I(t) + U + (css+ (t), Ť (t)) − U − (css− (t), Ť (t)),
+ I(t) sinh−1 + sinh −1
αF 2i+
0 (t)
a+ L + 2i−
0 (t)
a− L − ≜h1 (t , css
±
(t), c̄s± (t), I(t)). (25)
( + − )
Rf (T (t)) R (T (t))
+ + +
+ f − − − Rc I(t)2
a L a L 3.2.2. Write V (t) as a function of css +
(t) and I(t)
− I(t) U + (css+ (t), T (t)) − U − (css− (t), T (t))
[ ]
In order to further simplify the output function, we are to
+
{ establish relations between css (t) and the other concentrations
+ I(t) U + (c̄s+ (t), T (t)) − U − (c̄s− (t), T (t)) −
css (t), c̄s+ (t), c̄s− (t). Consider the following approximate polyno-
mial solution profiles of the electrode diffusion dynamics (Subra-
∂ U (c̄s (t), T (t)) ∂ U − (c̄s− (t), T (t)) manian, Diwakar, & Tapriyal, 2005)1 :
[ + + ]}
− T (t) − . (22)
∂T ∂T 8R± R±
c̄s± (t) = css
±
(t) − s
q̄±
s (t) +
s
j± (t), (26)
± ± 35 ±
We assume that the functions U (·, T (t)) and i0 (·) are independent 35Ds (Ť (t))
of concentrations but possibly time-varying, and we replace their
where the volume averaged fluxes q̄±
s (t) satisfy
dependence on temperature T (t) with dependence on the ambient
temperature Tamb (t). For this purpose, denote d 30D±
s (Ť (t)) 45
q̄±
s (t) = − q̄±
s (t) − j± (t).
Ǔ1 (t) ≜ U (css (t), Tamb (t)),
± ± ± dt (R± )2
s 2(R±
s )
2

Ǔ2± (t) ≜ U ± (c̄s± (t), Tamb (t)), Moreover, from the conservation (20) of lithium ions in solid phase
±
NLi,s , we can write the relation
0 (Tamb (t)),
iˇ0 (t) ≜ i±
c̄s− (t) = α c̄s+ (t) + β, (27)
where subscript 1 in Ǔ is used to denote the approximation of
U ± (·, T (t)) when they are evaluated at the surface concentration 1 Note that (26) is obtained by assuming the following polynomial solution
and subscript 2 when they are evaluated at the averaged con-
profile
centration. Similarly, for the temperature dependent parameters
39 35
R± ±
f (T (t)) and reff (T (t)), we also replace dependence on T (t) with cs± (t , r) = ±
css (t) − 3q̄± ±
s (t)Rs − c̄s± (t)
4 4
Tamb (t), i.e., ) (rs± )2
+ −35css± (t) + 10q̄± ± ±
(
± s (t)Rs + 35c̄s (t)
ˇ (t) ≜ reff (Tamb (t)),
reff ± ± 2
(Rs )
rs4
( )
± 105 105
Rˇf (t) ≜ Rf (Tamb (t)).
± .
± ± ± ±
+ css (t) − 7q̄s (t)Rs − c̄s (t)
4 4 (R±
s )
4
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 211

ε + L+ NLi,s
where α = − and β =
ε − L− ε − L−
. It then immediately follows from interval. Without loss of generality, consider t ∈ [0, tmax ] where
(26) and (27) that tmax is an appropriate finite time for the regularities to hold.
(
8R+ R+
) A Luenberger-type observer for the normalized and trans-
c̄s− (t) = α +
css (t) − s
q̄+
s (t) +
s
j+ (t) + β, (28) formed PDE system (31)–(34) can be designed:
35 35D+
s (Ť (t))

and ∂ ĉ D+ (Ť (t)) ∂ 2 ĉ

(t , r) = s + (t , r) + p1 (t , r)(c(t , 1) − ĉ(t , 1)),
8R− R− ∂t (Rs )2 ∂ r 2
− s s
css (t) = c̄s− (t) + q̄−
s (t) − j− (t) t > 0, r ∈ (0, 1), (35)
35 35D− s (Ť (t))
8R+ R+ ĉ(t , 0) = 0, t > 0, (36)
( )
= α css+ (t) − s q̄+ (t) + s
j+
(t) +β ∂ ĉ
35 s 35D+ s (Ť (t)) (t , 1) − ĉ(t , 1) = I1 (t) + p10 (t)(c(t , 1) − ĉ(t , 1)), t > 0, (37)
8R− R− ∂r
+ s −
q̄s (t) − s
j− (t). (29) ĉ(0, r) = ĉ0 (r), r ∈ [0, 1], (38)
35 −
35Ds (Ť (t))
Therefore, from (25), (26), (28) and (29), we obtain a further sim- which is a copy of the plant together with output error injection
plified version of the output function: terms. Here, ĉ0 (r) denotes the initial condition of the observer, and
the boundary state error injection gains p1 (t , r) and p10 (t) are to
V (t) = h2 (t , css
+
(t), I(t)). (30) be determined to guarantee the stability of the estimation error
system
3.2.3. Inversion of the function h2 ∂ c̃ D+ (Ť (t)) ∂ 2 c̃
As long as the function (30) is a one-to-one correspondence (t , r) = s + (t , r) − p1 (t , r)c̃(t , 1),
+ ∂t (Rs )2 ∂ r 2
w.r.t. css (t), uniformly in I(t), one could invert it to derive the
boundary concentration in the positive electrode as t > 0, r ∈ (0, 1), (39)
c̃(t , 0) = 0, (40)
+
css (t) = h0 (t , V (t), I(t)).
∂ c̃
(t , 1) − c̃(t , 1) = −p10 (t)c̃(t , 1), (41)
∂r
3.3. Normalization and state transformation c̃(0, r) = c0 (r) − ĉ0 (r) ≜ c̃0 (r), (42)

We perform normalization and state transformation to simplify with c̃(t , r) ≜ c(t , r) − ĉ(t , r). In order to find the output injection
the system and thus also the structure of to-be-designed observer. gains, the PDE backstepping method (Krstic & Smyshlyaev, 2008)
Let r = rs /R+s for normalization and proceed the state transfor- is employed. We would like to find an invertible transformation
mation c(t , r) = rs cs+ (t , rs ), then the PDE subsystem (8)–(11) is
1
transformed into2
∫
c̃(t , r) = w̃ (t , r) − p(t , r , ι)w̃ (t , ι)dι (43)
∂c D+ (Ť (t)) ∂ 2 c r
(t , r) = s + (t , r), t > 0, r ∈ (0, 1), (31)
∂t (Rs )2 ∂ r 2 so that w̃ satisfies the following exponentially stable target sys-
c(t , 0) = 0, t > 0, (32) tem:
∂c R+ I(t) ∂ w̃ D+ (Ť (t)) ∂ 2 w̃
(t , 1) − c(t , 1) = s
≜ I1 (t), t > 0, (33) (t , r) = s + (t , r) + λw̃ (t , r), (44)
∂r + + +
Ds (Ť (t)) a FL ∂t (Rs )2 ∂ r 2
c(0, r) = c0 (r) = R+ s rcs (0, Rs r), r ∈ [0, 1].
+ +
(34) w̃(t , 0) = 0, (45)
Our objective now is to design an observer for this normalized and ∂ w̃ 1
(t , 1) = − w̃ (t , 1), (46)
transformed PDE system. ∂r 2
w̃(0, r) = w̃0 (r), (47)
4. Backstepping state observer
where w̃0 (r) denotes the initial condition to be determined for
With the inversion of the output function in Section 3.2, the the target system, and λ < mint ≥0 {D− + 2
s (Ť (t))}/(4(Rs ) ) is a free
boundary concentration in the positive electrode is then available parameter to be chosen, which determines the convergence rate of
for observer design. Again, we assume that the internal averaged the observer state in (39)–(42) to the system state in (31)–(34). The
temperature is a time-varying and concentration-independent following lemma states the exponential stability of the w̃ -system
function which can be computed from the simplified Eq. (24). Thus, (44)–(47).
the function D+ s (Ť (t)) will be treated as known. Moreover, assume
that I(t), U ± (·, Ť (t)) and V (t) are piecewise (real) analytic. In what Lemma 2. Let t ∈ [0, tmax ]. If
follows, we only consider the proof piecewisely so that both I(t)
1
and V (t) are analytic in each separate time interval. Then, from λ< min{D+
s (Ť (t))}, (48)
(24) and with the assumption that ∂ U /∂ T are also analytic in each
± 4(R−
s )
2 t ≥0

corresponding time interval, we can prove by induction that the

then for any initial data w̃0 (·) ∈ L2 (0, 1), the w̃ -system (44)–(47)
nth order derivative of Ť (t) is differentiable for any nonnegative
admits a (mild) solution w̃ (t , ·) ∈ L2 (0, 1) and is exponentially stable
integer n. Further, we can derive that Ť (t) is analytic in each time
at w̃ ≡ 0. Moreover, if the boundary compatibility condition is
satisfied, the solution is classical.
2 The normalization transformation t̄ = D+ (Ť (t))/(R+ )2 t employed in Moura et
s s
al. (2014) is not used in this paper. The reason is that T (t) is not known a priori in this
case, needing to be measured or derived at each time step. Thus, the corresponding Proof. Consider the state space H = L2 (0, 1). For every t ∈
inverse transformation cannot be trivially obtained. [0, tmax ], define a linear operator A(t) : Dom(A(t)) ⊂ H → H
212 S.-X. Tang et al. / Automatica 83 (2017) 206–219

as follows: where (44), (46) and integration by parts have been used in the
calculation. Differentiating (43) with respect to r gives
D+
s (Ť (t))
A(t)ϕ = ϕ ′′ + λϕ, ∀ϕ ∈ Dom(A(t)), c̃r (t , r) = w̃r (t , r) + p(t , r , r)w̃ (t , r)
(R+
s )
2
∫ 1
{ }
1 − pr (t , r , ι)w̃ (t , ι)dι, (53)
Dom(A(t)) = ϕ ∈ H 2 (0, 1); ϕ (0) = 0, ϕ ′ (1) = − ϕ (1) . r
2
c̃rr (t , r) = w̃rr (t , r) + p(t , r , r)w̃r (t , r)
( )
Then, the system (44)–(47) can be written into the following ab- d
+ p(t , r , r) + pr (t , r , r) w̃ (t , r)
stract equation: dr
∫ 1
d − prr (t , r , ι)w̃ (t , ι)dι. (54)
w̃(t , ·) = A(t)w̃(t , ·), 0 ≤ t ≤ tmax , (49) r
dt
w̃(0, ·) = w̃0 (·). (50) From (39)–(42), (43), (45)–(47) and (52)–(54), we derive that the
kernel function p(t , r , ι) needs to satisfy the following PDE system:
Note that Dom(A(t)) is dense in H and independent of t, and it D+
s (Ť (t))
pt (t , r , ι) = [(prr (t , r , ι) − pιι (t , r , ι)] − λp(t , r , ι), (55)
can be proved that A(t) is for each t the infinitesimal genera- (R+
s )
2

tor of an exponential stable semigroup. Indeed, all the assump- p(t , 0, ι) = 0, (56)
tions (P1)–(P3) in Pazy (1983, Section 5.6) are satisfied. Thus, (R+ 2
s )
p(t , r , r) = λr , (57)
from Pazy (1983, Section 5.6, Theorem 6.1), there exists a unique 2D+ s (Ť (t))
evolution system corresponding to (49)–(50) and (44)–(47) as p(0, r , ι) = p0 (r , ι), (58)
well. Furthermore, by considering the Lyapunov function V (t) =
for which the domain is T = {(t , r , ι); 0 ≤ t ≤ tmax , 0 ≤ ι ≤ r ≤
1
2
∥w̃(t , ·)∥2L2 (0,1) , we get 1}. Here, p0 (r , ι) denotes the initial condition for the kernel system
[ ] and satisfies
s (Ť (t)) ∂ w̃
1 2
D+ 1
∫ ∫
V̇ (t) = w̃(t , r) (t , r) + λw̃ (t , r) dr p0 (r , ι)w̃ (t , ι)dι = c0 (r) − ĉ0 (r) − w̃0 (r). (59)
0 (R+ )2
s ∂ r2 r

D+ Moreover, the observer gains need to be chosen as

[ ]
s (Ť (t)) 1
= − w̃ 2 (t , 1) − ∥w̃r (t , ·)∥2L2 (0,1)
(R+ )2 2 D+
( )
s s (Ť (t)) 1
p1 (t , r) = − pι (t , r , 1) + p(t , r , 1) , (60)
+ λ∥w̃(t , ·)∥2L2 (0,1) (R+
s )
2 2
2
(R+
( )
D+ 3 s )
s (Ť (t)) p10 (t) = − λ. (61)
≤−2 − λ V (t), (51) 2 2D+
4(R+
s )
2 s (Ť (t))

In more detail, first, plugging (43), (52) and (54) into (39) gives (55),
where (44) is used in the first line, (45), (46) and integration by (60) and the boundary condition
parts are applied in the second line, and the Poincaré Inequal- d (R+ 2
s )
p(t , r , r) = λ. (62)
ity (Krstic & Smyshlyaev, 2008, Lemma 2.1) dr 2D+
s (Ť (t))

Second, plugging (40) and (45) into (43) gives (56). Third, (57) is
∥w̃(t , ·)∥2L2 (0,1) ≤ 4∥w̃r (t , ·)∥2L2 (0,1) derived from (62) and (56). Then, (61) is derived from (41), (43),
(46), (53) and (57). Finally, (59) is derived by plugging (42) and (47)
is employed in the last line. As a result, from (48), exponential into (43).
stability of the w̃ -system (44)–(47) is proved. □
4.1. Well-posedness of the kernel function p(r , ι, t)

For notation simplicity we will denote the L2 (0, 1)-norm by ∥ · ∥ Lemma 3. The initial data p0 (·, ·) is an analytic function in T =
in the sequel. {(r , ι); 0 ≤ ι ≤ r ≤ 1}, and the system (55)–(58) admits an analytic
solution p(t , ·, ·) in T .
Differentiating the transformation (43) with respect to t gives
Proof. We first transform the system (55)–(58) into an equivalent
s (Ť (t)) ∂ w̃
2
D+
[
c̃t (t , r) = (t , r) + p(t , r , r)w̃r (t , r) integral equation. Let ξ = r +ι, η = r −ι and q(t , ξ , η) = p(t , r , ι),
(R+ )2 ∂r2 then we have from (55)–(58) that q satisfies the following PDE:
(s ) ]
1
+ pι (t , r , 1) + p(t , r , 1) w̃ (t , 1) D+
s (Ť (t))
2 qt (t , ξ , η) = 4 qξ η (t , ξ , η) − λq(t , ξ , η), (63)
(R+
s )
2
D+
[ ]
s (Ť (t)) q(t , ξ , −ξ ) = 0, (64)
+ λ− pι (t , r , r) w̃ (t , r)
(R+ s )
2
(Rs ) + 2
∫ 1[ q(t , ξ , 0) = λξ , (65)
− pt (t , r , ι) + λp(t , r , ι) 4D+
s (Ť (t))
r
with the initial condition
D+
]
s (Ť (t)) ξ +η ξ −η
pιι (t , r , ι) w̃ (t , ι)dι,
( )
+ (52)
(R+ )2
s
q(0, ξ , η) = p 0, , .
2 2
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 213

Eq. (63) can be rewritten as Assume that (72) holds for an integer n ≥ 0, then, for any integer
m ≥ 0, we derive from (70) that
2
(R+
s )
qξ η (t , ξ , η) = (qt (t , ξ , η) + λq(t , ξ , η)) . ⏐∂ q (t , ξ , η)⏐
⏐ m n+1 ⏐
(66) t
+
4Ds (Ť (t)) ⏐ [ ∫ ξ∫ η ]⏐
= ⏐⏐∂tm Cf (t) qt (t , α, β ) + λqn (t , α, β ) dβ dα ⏐⏐
⏐ [ n ] ⏐
Integrating (66) with respect to η from 0 to η and using boundary
−η 0
condition (65), we have
⏐ m
⏐∑ [( )
m
(R+ 2
(R+ 2 = C ⏐⏐ ∂tm−i f (t)
s ) s ) i
qξ (t , ξ , η) = λ+ i=0
+
4Ds (Ť (t)) 4D+
s (Ť (t)) ∫ ξ ∫ η ]⏐
η
∂t q (t , α, β ) + λ∂ti qn (t , α, β ) dβ dα ⏐⏐.
∫ [ i+1 n ] ⏐
×
× (qt (t , ξ , β ) + λq(t , ξ , β )) dβ. (67) −η 0
0
Through further calculation, we obtain the following estimates
Integrating (67) with respect to ξ from −η to ξ gives the following
⏐∂ q (t , ξ , η)⏐
⏐ m n+1 ⏐
integro-differential equation (IDE): t
m
{( )
2 2
∑ m
(R+
s ) (R+
s ) ≤C Cfm+n+2 (m − i)!
q(t , ξ ,η) = λ (ξ + η ) + i
4D+
s (Ť (t)) 4D+
s (Ť (t))
i=0
λ
[ ]
∫ ξ ∫ η
× (qt (t , α, β ) + λq(t , α, β )) dβ dα, (68)
× λC n+1
Cf + (Cf + λ)n
i + 2n + 1
−η 0 }
(i + 2n + 1)! ξ n+1 ηn+1 (ξ + η)
where (64) is used. ×
2n n! (n + 1)!(n + 2)!
Second, we apply the method of successive approximation. Let
2 ≤ λC n+2 Cfm+n+2 (Cf + λ)n+1
(R+
s ) E + /Ť (t)
C = , f (t) = e D s , m [( )
(i + 2n + 1)! ξ n+1 ηn+1 (ξ + η)
]
E + /Ť (0)
∑ m
4D+
s (Ť (0))e
D s ×
i
(m − i)!
2n n! (n + 1)!(n + 2)!
i=0
then from (14), we look for a solution of (68) in the form of
= λC n+2 Cfm+n+2 (Cf + λ)n+1
∞
(m + 2(n + 1))! ξ n+1 ηn+1 (ξ + η)
∑
q(t , ξ , η) = qn (t , ξ , η), × ,
n=0
2n+1 (n + 1)! (n + 1)!(n + 2)!
where the following equalities have been used:
where
ξ η
α n β n (α + β ) ξ n+1 ηn+1 (ξ + η)
∫ ∫
dβ dα = ,
q0 (t , ξ , η) = λC (ξ + η)f (t), (69) −η 0 n!(n + 1)! (n + 1)!(n + 2)!
m ( )
and
∑ m (m + j + 1)!
(m − i)!(i + j)! = .
i j+1
qn+1 (t , ξ , η) i=0

∫ ξ ∫ η By induction, (72) holds for any integer n ≥ 0.

qnt (t , α, β ) + λqn (t , α, β ) dβ dα. Finally, the existence of q(t , ξ , η) and p(t , r , ι) can be proved.
[ ]
= Cf (t) (70)
−η 0 Fixing m = 0 in (72) gives

⏐q (t , ξ , η)⏐ ≤ λC n+1 C n+1 (Cf + λ)n (2n)! ξ η (ξ + η) .

⏐ n ⏐ n n
Recall that Ť (t) is analytic, and since it is physically impossible f
2n n! n!(n + 1)!
for the temperature to reach zero Kelvin, i.e., Ť (t) ̸ = 0, then it ∑∞ n
is reasonable to assume that 1 is an analytic function in t ∈ Then we can show that the series n=0 q (t , ξ , η) converges abso-
Ť (t)
[0, tmax ], and thus there exists a constant Cf such that for every lutely and uniformly. Indeed, the following bound holds:
∞
nonnegative integer k, the following bound holds: ∑
|q(t , ξ , η)| ≤ ⏐q (t , ξ , η)⏐
⏐ n ⏐
⏐ k ⏐
⏐f (t)⏐ := ⏐ d f (t)⏐ ≤ C k+1 k!. n=0
⏐ (k) ⏐ ⏐ ⏐
f (71) ∞
(2n)! ξ n ηn (ξ + η)
⏐ dt k ⏐ ∑
≤ λC n+1 Cfn+1 (Cf + λ)n
Since the composition of analytic functions is analytic, then 2n n! n!(n + 1)!
n=0
q0 (t , ξ , η) is an analytic function in t ∈ [0, tmax ] and it can be ∞
∑
derived from (69) and (71) that = λCCf (ξ + η) φ1 (ξ , η; n),
n=0
|∂ti q0 (t , ξ , η)| ≤ λCCfi+1 i!(ξ + η), ∀i ∈ N, where
(2n)! 1
with respect to (ξ , η), uniformly for t ∈ [0, tmax ]. φ1 (ξ , η; n) = [CCf (Cf + λ)ξ η]n .
In what follows we are to prove by induction that for any integer 2n n! n! · (n + 1)!
n ≥ 0 the following estimate holds: Since
φ1 (ξ , η; n + 1)
⏐∂ q (t , ξ , η)⏐ ≤ λC n+1 C m+n+1 (Cf + λ)n
⏐ m n ⏐ lim
t f n→∞ φ1 (ξ , η; n)
(m + 2n)! ξ n ηn (ξ + η) (2n + 1)
× . (72) = lim [CCf (Cf + λ)ξ η] = 0 < 1,
2n n! n!(n + 1)! n→∞ (n + 1)(n + 2)
214 S.-X. Tang et al. / Automatica 83 (2017) 206–219
∑∞
then from the ratio criterion, the series n=0 φ1 (ξ , η; n) is conver- Now, our main result can be presented. Consider an appropriate
gent. Consequently, the existence of q(t , ξ , η) and p(t , r , ι) is es- time interval [0, tmax ] for the assumed regularities in (A4) to hold.
tablished, which are analytic in T . Moreover, the following bound With the well-posedness of the kernel function in the transforma-
holds for p(t , r , ι) tion (43) together with the invertibility of the transformation, the
∞ following main theorem holds.
∑
|p(t , r , ι)| ≤ 2λCCf r φ2 (r , ι; n),
Theorem 4. Let t ∈ [0, tmax ]. Under the simplifications and assump-
n=0
tions (A1)–(A4), if
where
1
λ< min{D+
s (Ť (t))},
φ2 (r , ι; n) = φ1 (ξ , η; n). □ 4(R+
s )
2 t ≥0

then for any initial value ĉ(0, ·) ∈ L2 (0, 1), the observer error c̃-
4.2. Invertibility of the transformation (43) system (39)–(42) is exponentially stable at c̃ ≡ 0 in the L2 norm,
which means the designed observer (35)–(38) is exponentially con-
vergent to the system (31)–(34).
Indeed, the continuity of the kernel p(t , r , ι) in (43) guarantees
the existence of an inverse transformation. We write the inverse
Proof. It follows directly from (51) that
transformation as ( + )
1 D (Ť (t))
− s + 2 −λ t
∫
w̃(t , r) = c̃(t , r) + ρ (t , r , ι)c̃(t , ι)dι, (73) ∥w̃(t , ·)∥ ≤ ∥w̃(0, ·)∥e 4(Rs )
. (75)
r
From the state transformations (43) and (73), the equivalence of
then we could derive from (43) and (73) that the kernel ρ (t , r , ι)
the states c̃(t , r) and w̃ (t , r) is established, i.e., there exist positive
needs to satisfy
constants M1 , M2 such that
∫ ι
ρ (t , r , ι) = p(t , r , ι) + p(t , r , σ )ρ (t , σ , ι)dσ . (74) M1 ∥w̃ (t , ·)∥ ≤ ∥c̃(t , ·)∥ ≤ M2 ∥w̃ (t , ·)∥. (76)
r

In order to solve Eq. (74), a similar (successive approximation) Then, the proof is completed with (75) and (76). □
procedure as in Section 4.1 can be followed, see also, Krstic and
Smyshlyaev (2008, Section 4.4). A similar well-posedness result for Remark 5. Theorem 4 is rigidly proved under the assumption that
its inverse can also be obtained and this derivation is omitted here. the averaged internal temperature is independent of the lithium
Note also that the initial condition w̃0 (r) for the target w̃ -system ion concentrations in the electrodes; computed from the linear
(44)–(47) is determined by ĉ0 (r) and ρ0 (r , ι) = ρ (0, r , ι). Indeed, ODE (23). Here, we would like to clarify that these assumptions
from (42) and (73), w̃0 (r) can be calculated as are posed solely for the theoretical derivations. Indeed, in the
next section we are to present some simulation results showing
∫ 1
that the original unsimplified equation for the averaged internal
w̃0 (r) = c0 (r) − ĉ0 (r) + ρ0 (r , ι)[c0 (ι) − ĉ0 (ι)]dι.
r
temperature (17) can be used in the implementation of the esti-
mation algorithm, which depends on lithium ion concentrations in
the electrodes, and still achieve convergence of the SoC estimate.
4.3. Exponential convergence of the observer Since only estimates of lithium ion concentration are available
to compute the internal averaged temperature, we are actually
Some assumptions and simplifications have been made to ease computing an open-loop estimate calculated from (17) and use the
the analysis. For completeness and clarity we summarize these notation T̂ (t), i.e.,
assumptions and simplifications before stating our main result.
dT̂ (t)
ρ avg cP (t)
+
(A1) To derive an output inversion function, i.e., to recover css (t) dt( )
from the voltage measurement, we have assumed that T (t) is a = hcell Tamb (t) − T̂ (t)
time-varying function independent of concentrations. We have [ ( )
used the notation Ť (t) and compute its value from Eq. (23). For RT̂ (t) −1 1 I(t)
+ I(t) sinh
this assumption to hold, some underlying simplifications and αF ˆ
2i+ a+ L +
0 (t)
assumptions have been made: ( )]
−1 1 I(t)
± ± + sinh
(i) Parameters Rf (T (t)) and reff (T (t)) are approximated by
ˆ
2i− a− L −
ر ± ± ± 0 (t)
f (t) ≜ Rf (Tamb (t)) and řeff (t) ≜ reff (Tamb (t)). ( )
±
(ii) Functions U (·, T (t)) are assumed to be independent of R+
f (T̂ (t)) R−
f (T̂ (t))
concentrations, and their dependence on T (t) has been
+ + − Rc I(t)2
a+ L + a− L −
replaced with dependence on Tamb (t). We have used the [ ]
notation Ǔ1± (t) ≜ U ± (css ±
(t), Tamb (t)) and Ǔ2± (t) ≜ − I(t) U + (ĉss+ (t), T̂ (t)) − U − (ĉss− (t), T̂ (t))
U (c̄s (t), Tamb (t)).
± ± {
+ −
(A2) To derive an output inversion function, diffusion of lithium + I(t) U + (c̄ˆ s (t), T̂ (t)) − U − (c̄ˆ s (t), T̂ (t))
in the negative electrode has been simplified. This is done by as- [ + ]}
suming a polynomial solution profile for the diffusion dynamics ∂ U + (c̄ˆ s (t), T̂ (t)) ∂ U − ˆ −
− T̂ (t) − (c̄ (t), T̂ (t)) , (77)
in the negative electrode. ∂T ∂T s
(A3) For observer design, we have used Ť (t) to replace T (t).
(A4) For observer design, functions I(t), U ± (·, T (t)), V (t) and where i±ˆ
0 (t) are computed from (12) with concentration values
∂ U ±/∂ T (·, T (t)) are assumed to be piecewise analytic. replaced by their estimates.
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 215

In the original state variables and unnormalized coordinates,

the observer for lithium-ion concentration in the positive electrode
reads
∂ ĉs+ s (T̂ (t)) ∂
D+ 2 ∂ ĉs
+
[ ]
(t , rs ) = r (t , rs )
∂t rs2 ∂ rs s ∂ rs
+ p̄1 (t , rs )(css+ (t) − ĉss+ (t)), (78)
∂ ĉs+
(t , 0) =0, (79)
∂t
∂ ĉs
+
I(t)
(t , Rs ) =
∂r +
Ds (T̂ (t))a+ FL+
+ p̄10 (t)(css+ (t) − ĉss+ (t)), (80)
with
Fig. 2. Current profile.
p1 (t , rs
Rs
) p10 (t)
p̄1 (t , rs ) = , p̄10 (t) = . (81)
rs Rs
The SoC estimation can then be derived from (19) and (21), with 5.1.1. Simulation with square current profile
cs+ (t , rs ) replaced by their estimated values ĉs+ (t , rs ). Additionally, Figs. 2–5 correspond to the first set of simulation tests, which
estimates ĉs− (t , rs ) and ĉss
−
(t , rs ) on the negative electrode can be use a square current profile shown in Fig. 2. The current consists
computed from (28)–(29), with the estimates ĉss +
(t , rs ) and ĉss
+
(t , rs ) of repeated cycles of: 36 min of 1 C-rate constant discharging
on the positive electrode obtained from (78)–(81) and the open- followed by 54 min of resting, i.e., zero input, then 36 min of
1 C-rate constant charging ending with 54 min of zero input.
loop estimate T̂ (t) from (77).
Only the first 250 min of the simulation results are shown in the
figures. True and estimated SoC are shown in Fig. 3 using (a) SPM-T
5. Simulation results
measurements and (b) DFN measurements. The initial errors in SoC
estimation are due to intentionally chosen, incorrect initialization
The ambient temperature is assumed to be constant; Tamb =
of lithium concentrations. Convergence of output voltage coincides
298 [K] = 24.85 [◦ C]. Simulations are performed with parame-
with convergence of SoC, and this is shown in Fig. 4. The estimate
ters of a LiCoO2 –LiC6 cell. Parameters and OCP functions U ± are
of internal average temperature is shown in Fig. 5 using voltage
borrowed from Mao, Tiedemann, and Newman (2014) and the measurements from the (a) SPM-T model and (b) DFN model;
references within. Note that the OCP functions depend on the compared against the true average temperature of the respective
internal average temperature and here a linear approximation is models. Note that, since the internal average temperature is mon-
employed: itored in an open-loop fashion, one needs to correctly initialize its
U ± (css
±
(t), T ) = U ± (css
±
(t), Tamb ) value. This condition is satisfied at thermal equilibrium, i.e., the
internal average temperature of the battery coincides with the
∂ U ± (css± (t), Tamb )
+ (T − Tamb ). ambient temperature.
∂T
The magnitude of input current is described in terms of the cell 5.1.2. Simulation with UDDS current profile
C-rate (per unit area), which is computed from Figs. 6–9 correspond to simulation tests using a current input
derived from a set of UDDS data and scaled to a current density
min εs+ L+ cs+,max , εs− L− cs−,max
{ }
profile within the range of ±4 C-rate of the battery. This current
C-rate = F .
3600[s] profile, shown in Fig. 6, is representative of current demands in
automotive applications. SoC estimation is shown in Fig. 7 with the
initial errors coming from incorrect initialization. As seen in Fig. 8,
5.1. Simulation tests
convergence of the output voltage coincides with convergence of
the SoC as well. Finally, Fig. 9 compares the open-loop estimates
Simulation tests are performed to evaluate the effectiveness
of internal average temperature with the true internal average
of the proposed observer with two different current profiles: a
temperature from the (a) SPM-T model and (b) DFN model.
square profile (constant charge, discharge and rest) and a current
Fig. 10 shows (a) the difference in output voltage values be-
profile obtained from the Urban Dynamometer Driving Sched-
tween SPM-T model and SPM and (b) the difference in temperature
ule (UDDS). For each current profile two cases of measurements
values from SPM-T model and DFN model for constant discharge
are considered: voltage measurements generated from the SPM-T currents. One can see that the difference in output voltage values
model and voltage measurements generated from the DFN model from SPM-T model and SPM accentuates at high currents rate
serving as true data. To generate voltage measurements, lithium (a) while temperature values from SPM-T model and DFN model
concentration in the negative electrode is initialized at 80% of the remain relatively close for currents as high as 4 C-rate (b).
maximum value and lithium concentration in the positive elec-
trode is initialized at 50% of the maximum value. For the observer, 5.2. Numerical implementations
lithium concentration in the negative electrode is initialized at 50%
of the maximum value and the one in the positive electrode is Numerical implementations of the SPM-T and the DFN models
initialized at 67% of the maximum value. The tuning parameter λ follow the equations presented in Sections 2.1 and 2.2, respec-
in the observer is set as −1 for all tests.3 tively. A finite volume method is used for the spatial discretization

3 Ideally, the convergence rate of the designed observer can be made arbi- (calculated) boundary measurement, there exists a design trade-off between high
trarily high by choosing a small enough λ, i.e., a large enough |λ|. However, convergence rate of the observer and effective attenuation of the approximation
since accurate/direct measurement of boundary concentration is not available and error/measurement noise. In particular, choosing large values for |λ| makes the
approximations required in output inversion unavoidably introduce error in the system more sensitive to measurement noise and results in larger estimation errors.
216 S.-X. Tang et al. / Automatica 83 (2017) 206–219

Fig. 3. True and estimated SoC. (a). Observer with SPM-T measurements. (b). Fig. 5. True and estimated internal average temperature. (a). Observer with SPM-T
Observer with DFN measurements. measurements. (b). Observer with DFN measurements.

Fig. 6. Current profile.

concentration in the negative electrode is approximated by the

polynomial profile presented in Subramanian et al. (2005), as
described briefly in Section 3.2.
For the numerical implementation of the kernel function
p(t , r , ι) and the computation of the observer gain, a trapezoidal
approximation of the IDE (68) is used to obtain an ODE, which
is then discretized in time with the Euler-backward method. As
s (T (t))/(Rs ) t
+ 2
mentioned in Section 4, time normalization t ′ = D+
by the temperature-dependent function is not preferable; here
the normalization is performed by a constant instead, i.e., t ′ =
s (Tamb )/(Rs ) t.
+ 2
Fig. 4. True and estimated voltage. (a). Observer with SPM-T measurements. (b). D+
Observer with DFN measurements.
6. Conclusions and future work

of PDEs in the models, and then the Euler-backward method is used This paper discusses the problem of SoC estimation for the
for the temporal discretization of the resulting system of ODEs. The lithium-ion batteries based on a thermal–electrochemical model.
observer is implemented using the same discretization procedure. In this regard, an infinite-dimensional Luenberger observer is pro-
Note that in the numerical implementation of the observer, lithium posed. For the transformation between the observer error system
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 217

Fig. 7. True and estimated SoC. (a). Observer with SPM-T measurements. (b). Fig. 9. True and estimated interval average temperature. (a). Observer with SPM-T
Observer with DFN measurements. measurements. (b). Observer with DFN measurements.

Fig. 10. SPM-T model validation. (a) Output voltage error between SPM and DFN,
Fig. 8. True and estimated voltage. (a). Observer with SPM-T measurements. (b). and between SPM-T and DFN. (b) Internal averaged temperature computed from
Observer with DFN measurements. the DFN model and from the SPM-T model.

and the exponentially stable target system, well-posedness of the We consider this result as an additional step in the effort to design
time-varying PDE backstepping kernel functions are rigorously estimation (and control) algorithms for lithium-ion batteries from
proved. Then, exponential stability of the observer error system is electrochemical models, without relying on the discretization of
established, which proves effectiveness of the designed observer. the PDEs in these models.
218 S.-X. Tang et al. / Automatica 83 (2017) 206–219

The observer requires only one design/tuning parameter as Pazy, A. (1983). Semigroups of linear operators and applications to partial differential
compared with the possibly large number of tuning parameters re- equations. New York: Springer-Verlag.
Plett, Gregory L. (2004). Extended Kalman filtering for battery management systems
quired in estimation methods based on finite dimensional battery
of LiPB-based HEV battery packs: Part 2. Modeling and identification. Journal of
models, e.g., EKF. Compared with estimation based on the infinite Power Sources, 134(2), 262–276.
dimensional SPM alone, it takes into account the temperature de- Ramadesigan, Venkatasailanathan, Boovaragavan, Vijayasekaran, Arabandi,
pendence of model parameters and catches the battery responses Mounika, Chen, Kejia, Tsukamoto, Hisashi, Braatz, Richard, & Subramanian,
better than SPM, especially at high C-rates. Simultaneously, the Venkat (2009). Parameter estimation and capacity fade analysis of lithium-
ion batteries using first-principles-based efficient reformulated models. ECS
internal average temperature can be monitored in an open-loop Transactions, 19(16), 11–19.
fashion. Ramadesigan, Venkatasailanathan, Northrop, Paul W. C., De, Sumitava
Some simplifications are made in this paper, and their re- Santhanagopalan, Shriram, Braatz, Richard D., & Subramanian, Venkat R. (2012).
laxation could be considered as a future research direction. An- Modeling and simulation of lithium-ion batteries from a systems engineering
perspective. Journal of the Electrochemical Society, 159(3), R31–R45.
other possible extension is to retain the concentration dynamics
Santhanagopalan, S., & White, R. E. (2006). Online estimation of the state of charge
in the negative electrode and design one observer for each elec- of a lithium ion cell. Journal of Power Sources, 161(2), 1346–1355.
trode (Moura, Bribiesca Argomedo, Federico, Klein, Mirtabatabaei, Smyshlyaev, A., & Krstic, M. (2003). Explicit state and output feedback boundary
& Krstic, 2016). One could also consider multiple active materials controllers for partial differential equations. Journal of Automatic Control, 13(2),
1–9.
in the electrodes (Camacho-Solorio, Klein, Mirtabatabaei, Krstic, &
Smyshlyaev, A., & Krstic, M. (2005). On control design for PDEs with space-
Moura, 2016) or add models for degradations (e.g., capacity fade) to dependent diffusivity or time-dependent reactivity. Automatica, 41(9),
the battery model (Ramadesigan, Boovaragavan, Arabandi, Chen, 1601–1608.
Tsukamoto, Braatz, & Subramanian, 2009). Observer design for the Stewart, S., Christensen, J., Chaturvedi, N., & Kojic, A. (2015). Challenges in batteries
battery internal, core and surface (Lin, Perez, Mohan, Siegel, Ste- for electric vehicles. In Frontiers of engineering: Reports on leading-edge engineer-
ing from the 2014 symposium.
fanopoulou, Ding, & Castanier, 2014), or even the distributed (Kim, Subramanian, V. R., Diwakar, V. D., & Tapriyal, D. (2005). Efficient macro-micro scale
Mohan, Siegel, Stefanopoulou, & Ding, 2013) temperature is a coupled modeling of batteries. Journal of the Electrochemical Society, 152(10),
subject worth investigating as well. A2002–A2008.
Tang, S.-X., Wang, Y., Sahinoglu, Z., Wada, T., Hara, S., & Krstic, M. (2015).
State-of-charge estimation for lithium-ion batteries via a coupled thermal–
Acknowledgment
electrochemical model. American Control Conference, 5871–5877.
Tang, S.-X., & Xie, C. (2011a). Stabilization for a coupled PDE–ODE control system.
The authors would like to acknowledge Jean-Michel Coron for Journal of the Franklin Institute, 348(8), 2142–2155.
suggesting the application of the abstract evolution equation the- Tang, S.-X., & Xie, C. (2011b). State and output feedback boundary control for a
ory in the well-posedness derivation of the solution to the PDE coupled pDE-ODE system. Systems & Control Letters, 60, 540–545.
Tanim, Tanvir R., Rahn, Christopher D., & Wang, Chao-Yang (2015). State of charge
systems with time-varying coefficients. estimation of a lithium ion cell based on a temperature dependent and elec-
trolyte enhanced single particle model. Energy, 80, 731–739.
References Thomas, K., Newman, J., & Darling, R. (2002). Mathermatical modeling of lithium
batteries. In Advances in lithium-ion batteries (pp. 345–392). New York, NY USA:
Camacho-Solorio, L., Klein, R., Mirtabatabaei, A., Krstic, M., & Moura, S. (2016). Kluwer Academic/Plenum Publishers, (Ch. 12).
State estimation for an electrochemical model of multiple-material lithium-ion
batteries. ASME Dynamic Systems and Control Conference.
Chaturvedi, N., Klein, R., Christensen, J., Ahmed, J., & Kojic, A. (2010). Algorithms for
advanced battery-management systems. IEEE Control Systems Magazine, 30(3),
Shu-Xia Tang received her Ph.D. in Mechanical Engi-
49–68.
neering in 2016 from the Department of Mechanical &
Chiasson, J., & Vairamohan, B. (2005). Estimating the state of charge of a battery.
Aerospace Engineering, University of California, San Diego,
IEEE Transactions on Control Systems Technology, 13(3), 465–470.
USA. She is currently a postdoctoral research fellow and
Chu, S., & Majumdar, A. (2012). Opportunities and challenges for a sustainable
lecturer at the Department of Applied Mathematics, Uni-
energy future. Nature, 488(7411), 294–303.
versity of Waterloo, Canada. Her main research interests
Doyle, M., Fuller, T. F., & Newman, J. (1993). Modeling of galvanostatic charge and
discharge of the lithium/polymer/insertion cell. Journal of the Electrochemical are control and estimation in distributed parameter sys-
Society, 140(6), 1526–1533. tems. Recent research also includes optimal actuator and
Guo, M., Sikha, G., & White, R. E. (2011). Single-particle model for a lithium-ion cell: sensor design problems.
thermal behavior. Journal of the Electrochemical Society, 158(2), A122–A132.
Haran, B. S., Popov, B. N., & White, R. E. (1998). Determination of the hydrogen
diffusion coefficient in metal hydrides by impedance spectroscopy. Journal of Leobardo Camacho-Solorio received the B.S. degree in
Power Sources, 75(1), 56–63. Mechatronics Engineering from Tecnologico de Monter-
Izadi, M., & Dubljevic, S. (2015). Backstepping output-feedback control of moving rey, Queretaro in 2014. He is now working towards the
boundary parabolic pDEs. European Journal of Control, 21, 27–35. Ph.D. degree in Mechanical Engineering at the University
Kim, Youngki, Mohan, Shankar, Siegel, Jason B., Stefanopoulou, Anna G., & Ding, Yi. of California, San Diego. His research interests include
(2013). The estimation of radial temperature distribution in cylindrical battery estimation and control of systems described by partial dif-
cells under unknown cooling conditions. In 52nd IEEE conference on decision and ferential equations with applications to electrochemical
control (pp. 5680–5685). models of lithium-ion batteries.
Klein, R., Chaturvedi, N., Christensen, J., Ahmed, J., Findeisen, R., & Kojic, A. (2013).
Electrochemical model based observer design for a lithium-ion battery. IEEE
Transactions on Control Systems Technology, 21(2), 289–301.
Krstic, M., & Smyshlyaev, A. (2008). Boundary control of PDEs: A course on backstep-
ping designs. SIAM. Yebin Wang received the B.Eng. degree in Mechatronics
Lin, Xinfan, Perez, Hector E., Mohan, Shankar, Siegel, Jason B., Stefanopoulou, Anna Engineering from Zhejiang University, Hangzhou, China,
G., Ding, Yi, & Castanier, Matthew P. (2014). A lumped-parameter electro- in 1997, M.Eng. degree in Control Theory & Control Engi-
thermal model for cylindrical batteries. Journal of Power Sources, 257, 1–11. neering from Tsinghua University, Beijing, China, in 2001,
Mao, J., Tiedemann, W., & Newman, J. (2014). Simulation of li-ion cells by dualfoil and Ph.D. in Electrical Engineering from the University of
model under constant-resistance load. ECS Transactions, 58(48), 71–81. Alberta, Edmonton, Canada, in 2008. Dr. Wang has been
Moura, S., Chaturvedi, N., & Krstic, M. (2014). Adaptive partial differential equation with Mitsubishi Electric Research Laboratories in Cam-
observer for battery state-of-charge/state-of-health estimation via an electro- bridge, MA, USA, since 2009, and now is a Senior Principal
chemical model. Journal of Dynamic Systems, Measurement, and Control, 136(1), Research Scientist. From 2001 to 2003 he was a Software
011015. Engineer, Project Manager, and Manager of R&D Dept. in
Moura, S. J., Bribiesca Argomedo, F., Federico, Klein, R., Mirtabatabaei, A., & Krstic, industries, Beijing, China. His research interests include
M. (2016). Battery state estimation for a single particle model with electrolyte nonlinear control and estimation, optimal control, adaptive systems and their
dynamics. IEEE Transactions on Control System Technology, 25, 1086–1090. applications including mechatronic systems.
 S.-X. Tang et al. / Automatica 83 (2017) 206–219 219

Miroslav Krstic is Distinguished Professor of Mechanical Career, and ONR Young Investigator awards, the Axelby and Schuck paper prizes,
and Aerospace Engineering, holds the Alspach endowed the Chestnut textbook prize, and the rst UCSD Research Award given to an engineer.
chair, and is the founding director of the Cymer Center Krstic has also been awarded the Springer Visiting Professorship at UC Berkeley,
for Control Systems and Dynamics at UC San Diego. He the Distinguished Visiting Fellowship of the Royal Academy of Engineering, the
also serves as Associate Vice Chancellor for Research at Invitation Fellowship of the Japan Society for the Promotion of Science, and hon-
UCSD. As a graduate student, Krstic won the UC Santa orary professorships from four universities in China. He serves as Senior Editor in
Barbara best dissertation award and student best paper IEEE Transactions on Automatic Control and Automatica, as editor of two Springer
awards at CDC and ACC. Krstic is Fellow of IEEE, IFAC, book series, and has served as Vice President for Technical Activities of the IEEE
ASME, SIAM, and IET (UK), Associate Fellow of AIAA, and Control Systems Society and as chair of the IEEE CSS Fellow Committee. Krstic has
foreign member of the Academy of Engineering of Serbia. coauthored twelve books on adaptive, nonlinear, and stochastic control, extremum
He has received ASME Oldenburger Medal, ASME Nyquist seeking, control of PDE systems including turbulent ows, and control of delay
Lecture Prize, ASME Paynter Outstanding Investigator Award, the PECASE, NSF systems.

View publication stats