Automatica 133 (2021) 109849

Contents lists available at ScienceDirect

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

Battery internal temperature estimation via a semilinear thermal PDE

model✩
∗
Dong Zhang a , , Satadru Dey b , Shu-Xia Tang c , Ross Drummond d , Scott J. Moura e
a
Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA, USA
b
Department of Mechanical Engineering, Pennsylvania State University, State College, PA, USA
c
Department of Mechanical Engineering, Texas Tech University, Lubbock, TX, USA
d
Department of Engineering Science, University of Oxford, Oxford, UK
e
Department of Civil and Environmental Engineering, University of California, Berkeley, CA, USA

article info a b s t r a c t

Article history: Accurate Lithium-ion (Li-ion) battery internal temperature information enables high-fidelity monitor-
Received 6 November 2020 ing and safe operation in battery management systems, thus prevents thermal faults that could cause
Received in revised form 5 June 2021 catastrophic failures. This paper proposes an online temperature estimation scheme for cylindrical
Accepted 20 June 2021
Li-ion batteries based on a one-dimensional semilinear parabolic partial differential equation (PDE)
Available online 7 August 2021
model subject to in-domain and output uncertainties, using temperature measurements at the battery
Keywords: surface only. The thermal state observer design exploits PDE backstepping method, with a mild
Lithium-ion batteries assumption on the Lipschitz continuity of the nonlinear heat generation rate. A sufficient condition on
Battery temperature estimation the Lipschitz constant to achieve exponential convergence is derived. Furthermore, when the thermal
Infinite dimensional systems system uncertainties are present, an analytic bound on the temperature estimation error is formulated
PDE backstepping in the sense of spatial L2 norm, in terms of Lipschitz constant, design parameters, and bounds on
Robust observer system uncertainties. Simulation studies on various practical current profiles are demonstrated to
illustrate the effectiveness of the proposed thermal estimation framework on a commercial cylindrical
Li-ion battery cell.
© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction Zhao, Chu, Sun, and Chen (2012). As well as safety issues, thermal
effects have also been shown to be key factors in the rate of
Due to the penetration of the electric vehicles (EV) and con- battery degradation (Broussely, Biensan, Bonhomme, Blanchard,
sumer electronics, lithium-ion (Li-ion) batteries are ubiquitous. Herreyre, Nechev, & Staniewicz, 2005; Wang, Liu, Hicks-Garner,
The reason for this widespread penetration is that Li-ion bat- Sherman, Soukiazian, Verbrugge, Tataria, Musser, & Finamore,
teries possess one of the best energy-to-weight ratios, exhibit 2011). Hence, in order to improve battery safety and longevity, it
no memory effect, and have low self-discharge when not in is crucial to develop thermal management strategies to alleviate
operation (Chaturvedi, Klein, Christensen, Ahmed, & Kojic, 2010). the effects of temperature and prevent the drastic failure of the
However, as batteries insert themselves more and more into our battery from happening. Moreover, recent work in Yang, Liu,
society, their operation is becoming increasingly safety critical, and Wang (2021) demonstrates improved battery power and fast
meaning that any battery failure can have an increasingly impor- charging ability for Lithium-Iron Phosphate (LFP) cells operated
tant impact upon systems of increasing size (Kim & Shin, 2011). at 60◦ C ambient condition. This high temperature is closer to a
One of the existing challenges that substantially impacts the bat- thermal runaway regime and the hot spots need to be monitored
tery safety and performance is its thermal instability. In particu- carefully. In light of the above concerns and recent advancements,
lar, many cases of thermal runaways leading to fire and explosion this work proposes an online model-based algorithm for battery
of Li-ion batteries have been previously reported in Wang, Ping, internal temperature estimation based on a high-fidelity semi-
linear thermal partial differential equation (PDE) model, using
✩ The material in this paper was not presented at any conference. This paper battery surface temperature measurements only.
was recommended for publication in revised form by Associate Editor Angelo Modeling of battery thermal performance has been exten-
Alessandri under the direction of Editor Thomas Parisini.
∗ Corresponding author. sively studied in the literature (Doughty, Butler, Jungst, & Roth,
2002). Comprehensive high-dimensional thermal models, e.g.,
E-mail addresses: [email protected] (D. Zhang), [email protected]
(S. Dey), [email protected] (S.-X. Tang), [email protected] Chen, Wan, and Wang (2005), Kim, Pesaran, and Spotnitz (2007)
(R. Drummond), [email protected] (S.J. Moura). and Song and Evans (2000), provide an accurate and thorough

https://doi.org/10.1016/j.automatica.2021.109849
0005-1098/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

understanding of the cell temperature behavior from an electro- in Nambu (1984) by means of an auxiliary functional observer.
chemical point of view. Nevertheless, since these high- Smyshlyaev et al. apply the backstepping method to controller
dimensional models are too complicated and demand a great design for a model with space-dependent diffusivity or time-
amount of computational power, their application to real-time es- dependent reactivity (Smyshlyaev & Krstic, 2005). The techniques
timation and control in a battery management system (BMS) will introduced in Smyshlyaev and Krstic (2010) provide a thorough
not be feasible for applications outside of industrial/stationary analysis of the stability of adaptive control for linear parabolic
storage (Dey, Biron, Tatipamula, Das, Mohon, Ayalew, & Pisu, PDEs with spatially varying coefficients. However, extending
2016). To balance physical relevance and model structural sim- these results from linear to nonlinear PDEs require a more in-
plicity, reduced-order PDE thermal models have been proposed tricate analysis, with only few attempts so far. For instance, an
(Al Hallaj, Maleki, Hong, & Selman, 1999; Gu & Wang, 2000; Mu- extended Luenberger observer is developed for a class of semi-
ratori, Canova, & Guezennec, 2012). A few other finite- linear parabolic PDEs in Meurer (2013). It verifies the exponential
dimensional approaches for battery thermal modeling stand out, stability of the linearized observer error dynamics, in which the
and one such model is a two-state thermal model that predicts design extends the well-known backstepping method (Krstic &
the surface and core temperature of a cylindrical battery cell (Park Smyshlyaev, 2008) to include the Volterra transformation with a
& Jaura, 2003; Zhang, Couto, Park, Gill, & Moura, 2020b). A time-dependent kernel function. A series of studies by Vazquez
lumped thermal model has also been proposed by Smith et al. et al. discuss the control design for a 1-D parabolic PDE with
in Smith and Wang (2006). Note that battery thermal behavior Volterra nonlinearities (Vazquez & Krstic, 2008a, 2008b). Bound-
exhibits certain nonlinearity that are originated from resistive ary controller have also been designed towards Burgers’ equa-
heat generation rate, reversible entropic heat, and enthalpy of tion (Krstic, 1999). Additionally, designs based upon the plant
reactions, etc. models with modeling and parametric uncertainties, have also
Catastrophic thermal failures stimulate numerous works on been explored. Cheng considers the stabilization of the heat
the development of thermal management strategies. Previous equation with parameter variation and boundary uncertainties by
studies on battery state estimation can be categorized into two designing a sliding mode controller (Cheng, Radisavljevic, & Su,
main groups, i.e., the equivalent circuit model-based estima- 2011). Parabolic PDEs that are subject to in-domain and bound-
tion (Hu & Yurkovich, 2012; Plett, 2004; Zhang, Dey, Perez, ary parameter uncertainties are examined in Ahmed-Ali, Giri,
& Moura, 2017, 2019b) and the electrochemical model-based Krstic, Burlion, and Lamnabhi-Lagarrigue (2016), Zhang, Tang,
estimation (Dey, Ayalew, & Pisu, 2015; Klein, Chaturvedi, Chris-
and Moura (2019c). However, none of the aforementioned works
tensen, Ahmed, Findeisen, & Kojic, 2013; Moura, Argomedo, Klein,
discuss observer design for semilinear parabolic PDE systems
Mirtabatabaei, & Krstic, 2017; Tang, Camacho-Solorio, Wang, &
simultaneously subjected to Lipschitz nonlinearity and model
Krstic, 2017; Zhang, Dey, Couto, & Moura, 2019a). The existing
uncertainties.
BMS generally uses a thermal sensor attached to the surface
In this context, we advance the aforementioned works by
of the cell to measure surface temperature. As such, in Lin,
developing a provably convergent PDE observer for Li-ion battery
Perez, Siegel, Stefanopoulou, Li, Anderson, Ding, and Castanier
thermal estimation, with the only measurement from the battery
(2013), an adaptive observer for battery core temperature es-
cell surface. The thermal dynamics for a cylindrical battery in the
timation was designed for a two-state thermal model using
radial direction can be described by a one-dimensional semilinear
surface measurements. As well as estimation, studies on thermal
parabolic PDE with a mixture of Neumann and Robin boundary
fault diagnosis, e.g., the detection and isolation of the faults
conditions. Provided that the nonlinear in-domain heat genera-
that influence the surface and core temperature in the two-state
tion rate is Lipschitz continuous, this work marks one of the first
thermal model (Dey et al., 2016) and a PDE observer based fault
to simultaneously consider PDE model nonlinearity, in-domain
detection (Dey, Perez, & Moura, 2019), have also demonstrated
benefits. A dual Kalman filter based temperature distribution modeling uncertainty, and boundary measurement uncertainty
estimation for cylindrical batteries under unknown cooling, based within a single observer design framework, which addresses sev-
on a reduced radially distributed one-dimensional thermal mod- eral key practical challenges often faced within EV applications
eling, was proposed in Kim, Mohan, Siegel, Stefanopoulou, and (i.e., in-situ modeling and measurement uncertainties). This prob-
Ding (2014). More work on cylindrical cell internal temperature lem is challenging because the conventional PDE backstepping
distribution estimation by combining measured electrochemi- control and estimation techniques (Krstic & Smyshlyaev, 2008)
cal impedance and surface temperature, relying on a combined are well suited for linear PDEs without modeling uncertainties.
thermal-impedance model is introduced in Richardson, Ireland, This paper meets these challenges by deriving an analytical ex-
and Howey (2014). However, these methodologies suffer from pression for the nonlinear estimation error in the sense of L2
one or more of the following drawbacks: (i) Early lumping ap- norm in terms of observer gains, Lipschitz constant, and bounds
proaches, where the thermal PDE models are discretized and on the uncertainties, and further proposing an optimization prob-
approximated by systems of ordinary differential equations (ODE) lem to achieve robust estimation by minimizing the size of the
a priori, are used to design the observers. This leads to the loss of derived bound. In this paper, we choose to adopt the PDE back-
the physical significance of the PDE models and of the phenomena stepping method due to its unique feature to eliminate in-domain
that they represent, and the state estimates from the discretized destabilizing terms using control and observations only on the
systems may not always converge to the true states; (ii) Only boundary. This makes it a suitable method for battery internal
a small portion of these works provide theoretically certified temperature distribution estimation from surface temperature
convergence properties for the proposed estimation scheme; measurements only. It is further noted that although the ob-
(iii) Last but not the least, most of these works do not consider the server design for a general nonlinear PDE remains an unsolved
inherent nonlinearity of battery thermal model. Hence, this paper problem, this paper offers a new and different perspective on
attempts to collectively address these challenges and research state observer for a class of semilinear PDE subject to Lipschitz
gaps by proposing a provably convergent observer directly on the continuous nonlinearity with respect to Meurer (2013). The pro-
high-fidelity nonlinear thermal PDE model. posed state estimation approach does not require an extended
Methods for control/estimation of linear parabolic PDEs with linearization with respect to the estimated state during observer
late lumping have been well studied. The stabilization of unsta- design and convergence analysis. The design is unique, since it
ble heat equations using boundary observation was addressed exploits fundamental thermal dynamic properties and adopts a
2
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

high-fidelity thermal model for the purpose of battery temper- battery surface (system boundary). This is a standard practice as
ature monitoring, which ultimately enhances battery safety and the cell internal temperature measurement is intractable and may
longevity in EV applications. lead to danger.
The remainder of the paper is organized as follows. A coordinate transformation and normalization is now per-
Section 2 formulates the one-dimensional semilinear thermal PDE formed to simplify the structure of the model and translate it into
model with nonlinear Lipschitz continuous heat generation rate. a form for which an observer can be derived. First, we transform
Section 3 presents the state estimation scheme based on the the system (1)–(4) from the cylindrical to Cartesian coordinates
preceding model without the in-domain and output uncertainties, with spatial variable x. The coordinates and parameters are then
and the corresponding stability analysis of the estimation error normalized by defining x = x/Rc , t = α t /R2c , k = k/R2c , h = h/Rc ,
dynamics. Section 4 proposes the strategies for observer gain se- T (x, t) = T (r , t) to give the following PDE system:
lection. Section 5 develops a robust state estimation for the plant
∂T ∂ 2T Q̇ (x, t)
model with in-domain and output uncertainties. The performance (x, t) = (x, t) + , (5)
of the observers is demonstrated via simulations in Section 6. The ∂t ∂ x2 k
conclusions and future works are discussed in Section 7. with boundary conditions
∂T
Notation. Throughout the manuscript, T (x, t) denotes the plant’s (0, t) = 0, (6)
state variable, which depends on nondimensionalized space x and ∂x
∂T
time t. The x and t subscripts represent partial derivatives with (1, t) = δ (T∞ − T (1, t)), (7)
respect to the notated variable: ut = ∂ u/∂ t, ux = ∂ u/∂ x, and ∂x
uxx = ∂ 2 u/∂ x2 . The dot symbol denotes derivative with respect where x ∈ [0, 1] and δ = h/k. The surface temperature measure-
to time t, e.g., u̇ = du/dt. The spatial L2 norm is defined as ment signal in the transformed coordinate becomes

ym (t) = T (1, t).

√
∫ 1 (8)
∥T (·, t)∥ = T 2 (x, t)dx.
0 The formula for computing heat generation rate proposed by
Bernardi, Pawlikowski, and Newman (1985) is employed fre-
2. Distributed parameter thermal model quently in its simplified form (Dey et al., 2019; Forgez, Do,
Friedrich, Morcrette, & Delacourt, 2010):
This section presents the development of a one-dimensional ∂ Voc
nonlinear PDE thermal model for batteries, oriented towards state Q̇ (x, t) = I 2 (t)Rs (T (x, t), z(t)) − I(t)T (x, t) , (9)
∂T
estimation design.
where I(t) is the applied current, z(t) represents battery state
2.1. One-dimensional thermal model of charge (SOC), and Voc denotes open circuit voltage. Symbol Rs
stands for volumetric battery internal resistance and is generally
For a sufficiently long cylindrical battery cell, the heat transfer dependent on battery temperature and SOC (Lin et al., 2013;
resistance of spirally wound in the axial direction is significantly Samad, Siegel, & Stefanopoulou, 2014). The term Rs (T (x, t), z(t))
smaller than that in the radial direction, which allows to model can be characterized experimentally, e.g., see Fig. 6 in Samad
the thermal dynamics using radial heat transfer only (Al Hal- et al. (2014) for an illustrative example. For the heat generation
laj et al., 1999; Evans & White, 1989). Furthermore, the Biot rate expression in (9), the first term is the heat generated from
number in the radial direction is sufficiently large such that the resistive dissipation, which is always positive. The second term is
temperature gradients are not negligible. As such, the following the reversible entropic heat Keyser, Pesaran, Li, Santhanagopalan,
one-dimensional thermal model for the radially distributed tem- Smith, Wood, Ahmed, Bloom, Dufek, Shirk, et al. (2017), Thomas
perature profile of a cylindrical battery cell is adopted from Al and Newman (2003). The entropic coefficient (∂ Voc /∂ T ) varies
Hallaj et al. (1999). By assuming that heat is generated uni- with SOC and can be determined by offline experimental studies,
formly throughout the cell, the cell can be treated as a thermally e.g., see Forgez et al. (2010).
homogeneous body:
Remark 1. It is worth noting that tabs in cylindrical cells could
1 ∂T ∂ 2T 1 ∂T Q̇ (r , t) lead to heterogeneity in current distribution and hence tempera-
(r , t) = (r , t) + (r , t) + , (1)
α ∂t ∂r2 r ∂r k ture under high current density. The thermal model (1)–(9) does
not specifically consider this effect, because the unsteady-state
∂T
(0, t) = 0, (2) one-dimensional (radial direction) thermal model with lumped
∂r properties are sufficient for modeling purposes under normal
∂T h conditions of battery use (Al Hallaj et al., 1999). Besides, the
(Rc , t) = (T∞ − T (Rc , t)), (3)
∂r k choice of the one-dimensional thermal model intelligently strikes
ym (t) = T (Rc , t), (4) a balance between modeling accuracy and tractability towards
state observer design. Furthermore, exemplified by the develop-
where T (r , t) is the battery temperature distribution with respect ment of Tesla’s forthcoming tabless 4860 cells, it is expected that
to radial position and time. r ∈ [0, Rc ] represents the radial the thermal gradient will be less severe than current cell designs.
coordinate, Rc is the battery radius, and t ∈ R+ is time. T∞ is am- As such, the restrictiveness of the assumption to neglect the heat
bient temperature, and Q̇ denotes the volumetric heat generation generation around the tabs may be less severe in future battery
rate. Moreover, parameters k and h are the thermal conductivity packs (Tranter, Timms, Shearing, & Brett, 2020).
and effective surface heat transfer coefficient of the battery cell,
respectively. Convective cooling through the battery surface is 2.2. Model reduction and analysis
modeled by a Robin boundary condition in (3). In addition, α =
k/(DCp ) is the thermal diffusivity, where D is the mass density In this work, the heat generation rate formulation (9) is simpli-
and Cp denotes the specific heat capacity. Furthermore, in (4), it fied by neglecting the entropic heat generation in the subsequent
is assumed that the only temperature measurements are from the discussions as the entropic coefficient (∂ Voc /∂ T ) is significantly
3
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

small for certain types of cell chemistry, e.g., LFP cell (Forgez The initial condition of the plant model is T (x, 0) = T0 (x). For the
et al., 2010). The simplified heat generation rate is adopted, which design of the state observer, the temperature at the surface of the
is used to strike a balance between the need for a model of battery is measured:
sufficient accuracy yet is simple enough to enable observer design
with convergence guarantees. Furthermore, the simplified heat y(t) = T (1, t) + µ(t), (18)
generation rate is considered to be sufficient for electric vehicle
where we impose output uncertainty µ(t) to account for distur-
and portable electronics applications (Du, Hu, Xie, Hu, Zhang, &
bances from ambient environment and thermal sensor inaccu-
Lin, 2020), and its dependence on temperature and SOC further
racies. Let U = L2 (Tmin , Tmax ) denote the state space of T (x, t).
enhances the prediction accuracy for cells over a wide range of
According to (10), Rs (T , z) is C 1 in T , for all z ∈ [0, 1] and T ∈ U .
operational conditions (for instance, various current rates and
different temperatures). Furthermore, since function f (T , z) satisfies (15), f (T , z) is also C 1
in T . The next theorem establishes Lipschitz continuity of function
Remark 2. As noted in Käbitz, Gerschler, Ecker, Yurdagel, Em- f (T , z).
mermacher, André, Mitsch, and Sauer (2013), Liaw, Roth, Jungst,
Nagasubramanian, Case, and Doughty (2003) and Lin et al. (2013) Theorem 1 (Lipschitz Continuity). The nonlinear function f : U ×
the behavior of Rs due to the variation in temperature should [0, 1] → U is Lipschitz continuous with respect to T in the range
nearly follow an Arrhenius type relationship: T ∈ [Tmin , Tmax ]. That is, for all T1 , T2 ∈ U , there exists a positive
[ ( )] constant Γ = Γ (z) such that
Ea 1 1
Rs (T , z) = M(z) · Rs,ref exp − , (10)
R T Tref ∥f (T1 , z ) − f (T2 , z )∥ ≤ Γ (z) · ∥T1 − T2 ∥ (19)

where Tref is a given reference temperature, Ea is activation en- holds for all z ∈ [0, 1] and t ∈ [0, ∞). Moreover, let γ =
ergy, R denotes universal gas constant, and Rs,ref = Rs (Tref ). maxz (Γ (z)) > 0 which is independent of z, then γ can be regarded
M(z) > 0 is usually a nonlinear function. Such a monotoni- as the Lipschitz constant for function f (T , z).
cally decreasing Arrhenius expression is retrieved from Lin et al.
(2013). Function M(z) characterizes the nonlinear dependence on Practically, a candidate for the Lipschitz constant may be ob-
SOC (Lin, Perez, Mohan, Siegel, Stefanopoulou, Ding, & Castanier, tained by computing the infinity norm of (df /dT ). According to
2014). Observe that the gradient of Rs with respect to T is (15) and Remark 2,
[ ( )]
dRs Ea Ea 1 1
  ⏐ ⏐
 df 
< 0. γ =   = 1 |I |2 ⏐ dRs (T , z)⏐
⏐ ⏐
= −M(z) · Rs,ref exp − (11)
dT RT 2 R T Tref  dT  k max ⏐
dT ⏐
∞
[ ( )] max
Li-ion battery manufacturers specify safety restrictions on allow- Ea 1 1
able operation temperature range, which is generally between = σ · exp − , (20)
R Tmin Tref
−30 ◦ C (243.15 K) and 60 ◦ C (333.15 K) for typical Li-ion battery
operation (Koniak & Czerepicki, 2017). Within this temperature where σ = |M(z)|max |I |2max Rs,ref Ea /(kRTmin
2
). Herein, |·|max denotes
range, the function dRs /dT in (11) is monotonically increasing the absolute maximum value.
with respect to T . Hence, suppose the maximum and minimum
temperature regulated by an application are Tmax and Tmin , we Theorem 2. Theorem 1 guarantees the existence and uniqueness of
have a classical solution to the PDE system (14)–(17).
dRs dRs dRs
(Tmin , z) ≤ (T , z) ≤ (Tmax , z) < 0, (12) Proof. It is observed from (9) and (13) that ν (or equivalently
dT dT dT
ε ) is globally Lipschitz continuous with respect to T in the range
for all z ∈ [0, 1].
T ∈ [Tmin , Tmax ]. Hence, the function F := f + ε is also glob-
Consequently, the modeling error from the aforementioned ally Lipschitz continuous with respect to T in the range T ∈
reductions on heat generation is compensated by an in-domain [Tmin , Tmax ]. According to Meurer (2013, Assumption 1) and Pazy
uncertainty ν (x, t), i.e., (2012, Theorem 6.1.5), the existence and uniqueness of a classical
solution can be ensured.
Q̇ (x, t) = I 2 Rs (T (x, t), z(t)) + ν (x, t), (13)

where ν (x, t) is upper and lower bounded, and these bounds can Assumption 1. The in-domain uncertainty ε (x, t) is finite and
be numerically retrieved by experiments. Hence, substituting the bounded by ε (x, t) ≤ ε , ∀(x, t) ∈ [0, 1] × [0, ∞), where ε ≥ 0.
reduced-order heat generation rate (13) into the dynamics (5) The disturbance µ(t) in the output equation is finite and bounded
yields the plant model under consideration in the subsequent by µ(t) ≤ µ, ∀t ∈ [0, ∞), where µ ≥ 0. We further assume that
studies: µ(t) is continuously differentiable with respect to time t.
∂T ∂ 2T
(x, t) = (x, t) + f (T (x, t), z(t)) + ε (x, t), (14) Our objective is to design a provably convergent observer to
∂t ∂ x2 estimate battery radially distributed temperature profile T (x, t)
where ε (x, t) = ν (x, t)/k and by utilizing only cell’s surface temperature measurements y(t) =
1 T (1, t). As such, in Section 3, we first present an observer design
f (T (x, t), z(t)) = I 2 Rs (T (x, t), z(t)), (15) and the corresponding observer convergence analysis based on
k
the uncertainty-free plant model, i.e., ε (x, t) = 0 and µ(t) = 0,
with Rs given by (10). The boundary conditions are
∀x ∈ [0, 1] and t ∈ [0, ∞). Furthermore, in Section 5, we
∂T develop a robust state estimation scheme based on the plant
(0, t) = 0, (16)
∂x model with uncertainties, i.e., ε (x, t) ̸ = 0 and µ(t) ̸ = 0, in the
∂T sense of minimizing the estimation error via optimal observer
(1, t) = δ (T∞ − T (1, t)). (17)
∂x gain scheduling.
4
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

3. State estimation for uncertainty-free plant model which maps the error system (30)–(32) to the target system

In this section, the following uncertainty-free thermal plant ωt (x, t) = ωxx (x, t) + ψ (x, t) − c ω(x, t), (36)
model, namely ε (x, t) = 0 and µ(t) = 0, for all x ∈ [0, 1] and ωx (0, t) = 0, (37)
t ∈ [0, ∞), is considered:
ωx (1, t) = −(c1 + δ )ω(1, t), (38)
Tt (x, t) = Txx (x, t) + f (T (x, t), z(t)), (21)
where constants c and c1 are parameters to be designed. More-
Tx (0, t) = 0, (22) over, ψ (x, t) is the transformed function from φ (x, t) by the same
Tx (1, t) = δ (T∞ − T (1, t)), (23) backstepping transformation structure in (35) Cheng et al. (2011),
as follows,
T (x, 0) = T0 (x), (24) ∫ 1

y(t) = T (1, t). (25) φ (x, t) = ψ (x, t) − K (x, y)ψ (y, t)dy. (39)
x
An observer for reconstructing the spatial and temporal evolution To explicitly determine the kernel function K (x, y) we differen-
of cell temperature is proposed in the subsequent sections. tiate both sides of the transformation in (35) with respect to x
and t and take the target system dynamics (36)–(38) into account.
3.1. State observer structure
An example of a similar procedure can be retrieved in Krstic and
Smyshlyaev (2008, Chapter 4). The computation reveals that the
A distributed parameter state observer system is designed by
kernel function K (x, y) must satisfy the following Klein–Gordon
using a copy of the plant model with output error injection,
PDE,
T̂t (x, t) = T̂xx (x, t) + f (T̂ (x, t), z(t))
Kxx (x, y) − Kyy (x, y) = −cK (x, y), (40)
+ p1 (x) y(t) − T̂ (1, t) ,
[ ]
(26)
Kx (0, y) = 0, (41)
T̂x (0, t) = 0, (27) c
K (x, x) = − x, (42)
T̂x (1, t) = δ (T∞ − T̂ (1, t)) + p10 y(t) − T̂ (1, t) ,
[ ]
(28) 2
in which the boundary condition (41) emerges from evaluating
T̂ (x, 0) = T̂0 (x) ̸ = T0 (x), (29)
(35) together with the boundary conditions (31)–(32). A unique
where T̂ (x, t) denotes the estimation of T (x, t), and T̂ (1, t) is and closed-form analytic solution exists for the PDE (40)–(42):
the boundary state estimate. Symbols p1 (x) and p10 are spatially (√ )
I1 c(y2 − x2 )
varying and constant observer gains to be designed to guarantee K (x, y) = −cy √ , (43)
the stability of the estimation error T̃ (x, t) := T (x, t) − T̂ (x, t). c(y2 − x2 )
where I1 (·) is the modified Bessel function of the first kind.
Remark 3. In this paper, we assume that the SOC of the battery
Solution (43) is derived by converting the PDE (40)–(42) into an
is either known or can be reliably estimated in real time by
integral equation and applying the method of successive approx-
algorithms developed in, e.g., Dey et al. (2015), Hu and Yurkovich
imations. Moreover, the observer gains are computed as
(2012), Klein et al. (2013), Moura et al. (2017), Plett (2004), Tang
et al. (2017) and Zhang et al. (2019a, 2017, 2019b). The estimated p1 (x) = −Ky (x, 1) − (c1 + δ )K (x, 1), (44)
SOC is then used to evaluate function f in the observer dynamics c
(26). p10 = c1 − K (1, 1) = c1 + . (45)
2
Subtracting (26)–(29) from (21)–(24) yields the estimation Therefore, the observer gains (44)–(45) can be determined offline
error dynamics: utilizing the kernel PDE solution (43).
T̃t (x, t) = T̃xx (x, t) + φ (x, t) − p1 (x)T̃ (1, t), (30)
Remark 4. The fact that ψ (x, t) is the transformed version of
T̃x (0, t) = 0, (31) φ (x, t) under the same backstepping transformation used for esti-
T̃x (1, t) = −(δ + p10 )T̃ (1, t), (32) mation error state transformation makes the kernel PDE
(40)–(42) to exhibit the same structure as that of the linear
T̃ (x, 0) = T (x, 0) − T̂ (x, 0), (33) case, e.g., Krstic and Smyshlyaev (2008, Chapter 4). The challenge
where hereafter lies in the stability verification of the target system
(36)–(38).
φ (x, t) := f (T (x, t), z(t)) − f (T̂ (x, t), z(t)). (34)
Additionally, it is necessary to verify the existence and unique-
Although the explicit dependence of φ (x, t) on T (x, t) and T̂ (x, t) ness of the inverse backstepping transformation, so that the sta-
is suppressed in the notation above, this explicit dependence bility of the target system implies the stability of the original error
remains, and is crucial, in the subsequent error system stability system. Consider the inverse backstepping transformation (Krstic
analysis. & Smyshlyaev, 2008),
∫ 1
3.2. Backstepping transformation
ω(x, t) = T̃ (x, t) + ℓ(x, y)T̃ (y, t)dy, (46)
x
To determine the appropriate observer gains, we adopt the
with the kernel function ℓ(x, y). Similarly, differentiating both
backstepping approach (Krstic & Smyshlyaev, 2008). We seek a
sides of the inverse backstepping transformation with respect to
linear backstepping transformation, i.e., a Volterra integral trans-
x and t yields the kernel PDE for ℓ(x, t):
formation, that transforms the state of the error system T̃ (x, t) to
a target state ω(x, t), by making use of the following expression ℓxx (x, y) − ℓyy (x, y) = c ℓ(x, y), (47)
with kernel function K (x, y):
ℓx (0, y) = 0, (48)
∫ 1
c
T̃ (x, t) = ω(x, t) − K (x, y)ω(y, t)dy, (35) ℓ(x, x) = − x, (49)
x 2
5
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

which has an analytic solution of the form Therefore, in view of (55) and (56), function ψ (x, t) verifies the
(√ ) inequality ∥ψ (·, t)∥ ≤ κ (c ; γ )∥ω(·, t)∥.
J1 c(y2 − x2 )
ℓ(x, y) = −cy √ , (50) Through Lemma 1, the stability of the error system (30)–(34)
c(y2 − x2 ) in terms of the L2 norm can then be established via the stability
where J1 (·) is the Bessel function of the first kind. of the target system.

3.3. Stability of the target system Theorem 3 (Convergence of State Observer). Consider the observer
error dynamics (30)–(34), and let the observer gains p1 (x) and p10
Based on the analysis in Section 3.2, the analytic solutions be as in (44) and (45). Given Lipschitz constant γ , if the design
for the backstepping and inverse backstepping kernel functions parameters c and c1 are chosen such that
K (x, y) and ℓ(x, y) exist and are unique. Therefore, the stability
1 1
properties of the target system (36)–(37) imply the stability of c > κ (c ; γ ) − , c1 ≥ − δ, (57)
the original error system (30)–(32). In this section, we perform 4 2
Lyapunov analysis to establish the stability of the target system then the origin of the error dynamics T̃ (x, t) = 0 is exponentially
in the sense of spatial L2 norm. Prior to that, we first present and stable in the sense of L2 norm, without the presence of the in-
prove the following lemma. domain uncertainty and the output uncertainty, i.e., ε (x, t) = 0 and
µ(t) = 0 for all x ∈ [0, 1] and t ∈ [0, ∞).
Lemma 1. The nonlinear function ψ (x, t) in the target system (36)
verifies the following inequality for all t ∈ [0, ∞), Proof. Consider the Lyapunov functional candidate for system
∥ψ (·, t)∥ ≤ κ (c ; γ ) ∥ω(·, t)∥ , (51) (36)–(39):
∫ 1
with κ (c ; γ ) := γ [1 + ρ (c)][1 + η(c)], where 1 1
W1 (t) = ω2 (x, t)dx = ∥ω(x, t)∥2 . (58)
√ 2 0 2
∫ 1∫ 1
ρ (c) := ℓ2 (x, y; c)dydx, (52) The time derivative of the Lyapunov function W1 (t) along the
0 x state trajectory can be computed as
√
∫ 1 ∫ 1
∫ 1

η(c) := K 2 (x , y; c)dydx, (53) Ẇ1 = ω(x, t)ωt (x, t)dx

0 x 0
∫ 1 ∫ 1 ∫ 1
and c is the design variable of the target system in (36). = ωωxx dx + ωψ dx − c ω2 dx. (59)
0 0 0
Proof. The inverse backstepping transformation from ψ (x, t) to Applying the integration by parts to the first term, and Cauchy–
φ (x, t) specified by the kernel function ℓ(x, y) in (50) is written Schwarz inequality to the second term at the right-hand side of
as (59) result in
∫ 1
ψ (x, t) = φ (x, t) + ℓ(x, y)φ (y, t)dy. (54) Ẇ1 ≤ −(c1 + δ )ω2 (1, t) − ∥ωx ∥2 + ∥ψ∥∥ω∥ − c ∥ω∥2 . (60)
x
Applying Poincaré inequality, −∥ωx ∥2 ≤ ω2 (1) − ∥ω∥2 , to the 1
2
1
4
Applying the triangle inequality produces
∫ 1  second term at the right-hand side of (60), and substituting the
third term with (51), yields
∥ψ (x, t)∥ ≤ ∥φ (x, t)∥ +  ℓ , φ ,
 
(x y) (y t)dy  ( ) [ ]

x
 1 1
√ Ẇ1 ≤ − c1 + δ − ω2 (1) − c + − κ (c ; γ ) ∥ω∥2 . (61)
∫ 1 (∫ 1 )2 2 4
= ∥φ (x, t)∥ + ℓ(x, y)φ (y, t)dy dx
0 x If the design parameters are chosen such that c1 ≥ 1
2
− δ and
√ (∫
∫ 1 1 ∫ 1 ) c > κ (c ; γ ) − 14 , (61) is simplified to
≤ ∥φ (x, t)∥ + ℓ2 (x, y)dy φ 2 (y, t)dy dx d
[
1
]
0 x x ∥ω∥ ≤ − c + − κ (c ; γ ) ∥ω∥, (62)
√
∫ 1 (∫ dt 4
1
)
≤ ∥φ (x, t)∥ + ℓ2 (x, y)dy ∥φ (y, t)∥2 dx which confirms the exponential stability of ω(x, t), as well as
0 x T̃ (x, t), in the sense of L2 norm.
⎛ √ ⎞
∫ 1∫ 1
= ⎝1 + ℓ2 (x, y)dydx⎠ ∥φ (x, t)∥ Remark 5. The state observer design imposes a simple linear out-
0 x put error injection for a nonlinear plant model, without having to
⎛ √ ⎞ perform linearization and compute time-varying kernel functions
∫ 1 ∫ 1 as in Meurer (2013). However, its limitation is analogous to that
≤ γ ⎝1 + ℓ2 (x, y)dydx⎠ ∥T̃ (x, t)∥, (55) of observer design for Lipschitz nonlinear ODE system (Rajamani,
0 x 1998), which will be discussed in the next section.
where the second inequality originates from the Cauchy–Schwarz
4. Numerical selection of design parameters
inequality, and the last inequality stems from the Lipschitz con-
tinuity of function f in Theorem 1. A similar computation can
The design criteria (57) for the stability of the origin of the
be performed based on the forward backstepping transformation
error system enforces c1 to be greater than a fixed constant
(35) to conclude
(1/2 −δ ), and an affine function of c to be greater than a nonlinear
function κ (c ; γ ). It should be highlighted that for a Lipschitz
⎛ √ ⎞
∫ 1 ∫ 1
∥T̃ (x, t)∥ ≤ ⎝1 + K 2 (y, t)dydx⎠ ∥ω(x, t)∥. (56) constant γ , κ (c ; γ ) increases near exponentially with respect to
0 x c since the backstepping kernel function K (x, y) is dependent
6
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

on the modified Bessel function I1 (·). The conditions in (57)

essentially require a linear function to dominate an exponential
function, no matter that the argument of function I1 (·) in (43) is
a square root of c. Satisfaction of the condition in (57) is then
entirely dependent upon the Lipschitz bound of the nonlinearity
γ . A numerical study is now presented to explore the space of γ
for which a feasible solution to the inequality of (57) exists.
Fig. 1 provides the visualization for the affine function (c +
1/4) and the nonlinear function κ (c ; γ ) with respect to c. As an
illustrative example, when γ < γ ∗ , κ (c ; γ ) (the black solid line)
intersects with (c + 1/4) (the blue solid line) at two distinct
points, namely c = c and c = c. Hence, under this scenario,
(c + 1/4) > κ (c ; γ ) if and only if c < c < c. With an increasing γ ,
the position of κ (c ; γ ) shifts towards the up-left direction, and the
span on the c-axis between c = c and c = c shrinks accordingly.
At the critical point when γ = γ ∗ , the straight line (c + 1/4) is
tangential to κ (c ; γ ) (the black dashed line) and they intersect
at a single point c = c ∗ . Ultimately, when γ > γ ∗ , κ (c ; γ ) will Fig. 1. An illustration of design parameter selection to guarantee observer
never intersect with (c + 1/4) for any values of c. Thus, when convergence. As γ increases, the function κ (c ; γ ) shifts towards the upper-left
direction and the span between c = c and c = c shrinks accordingly. When
γ < γ ∗ , there always exists a c such that (c + 1/4) > κ (c ; γ ) stays γ < γ ∗ ≈ 1.053, it is ensured that there always exists a design parameter c
valid. Using a bisection method, the critical value of the Lipschitz such that the conditions in (57) hold. (For interpretation of the references to
constant is γ ∗ ≈ 1.053. color in this figure legend, the reader is referred to the web version of this
article.)
5. State estimation for uncertain plant model

In the previous section, a state observer as well as its con- Remark 6. The backstepping transformation (69) for the uncer-
vergence properties for an uncertainty-free thermal plant model tain terms facilitates the construction of a target system of the
was proposed and analyzed. In this section, the plant model form (66)–(68), which is a key enabling step to integrate the in-
with in-domain and output uncertainties, emerged from the heat domain uncertainty into the target system. Moving forward, since
generation rate model reduction and temperature measurement r(x) is unknown, the challenge is to determine the functional form
noise, is under examination. We employ the same observer struc- of r(x) such that there exists a unique analytic solution to the
ture as in (26)–(29) for the uncertain plant model (14)–(17), and inverse backstepping transformation kernel function ℓ(x, y), and
investigate the boundedness of the estimation error in the sense that the target system (66)–(68) is stable.
of L2 norm.
Similarly, to explicitly determine K (x, y) and r(x), we differ-
Subtracting (26)–(29) from (14)–(17) produces the observer’s
entiate both sides of the backstepping transformation (35) with
error dynamics:
respect to x and t and take into account the target system dy-
T̃t (x, t) = T̃xx (x, t) + φ (x, t) − p1 (x)T̃ (1, t) namics (66)–(68). This direct computation leads to (70) which
is given in Box I, and note that (70) has to hold for all (x, t) ∈
− p1 (x)µ(t) + ε (x, t), (63)
[0, 1] × [0, ∞). Hence, the kernel function K (x, y) exhibits the
T̃x (0, t) = 0, (64) same structure as that in (40)–(42) from the uncertainty-free
T̃x (1, t) = − (δ + p10 )T̃ (1, t) − p10 µ(t), (65) case. Moreover, the observer gains also remain unchanged from
those in (44) and (45). Finally, in view of the third term in (70),
The analysis in this scenario is different from the traditional the function r(x) has to verify ∆(x) = 0 with
backstepping procedure where the boundary conditions and dy- ∫ 1
namics are uncertainty-free. It is noteworthy that the system of ∆(x) := r(x) − K (x, y)r(y)dy + p1 (x) + p10 K (x, 1). (71)
interest here, (63)–(65), imposes not only a nonlinearity but also x
modeling uncertainties. The closed-form expression of r(x) will be derived in the forth-
coming sections with the aid of the inverse backstepping trans-
5.1. Backstepping transformation formation.

The backstepping transformation (35) maps the error system 5.2. Inverse backstepping transformation
(63)–(65) to a target system
We adopt the same inverse backstepping transformation (46)
ωt (x, t) = ωxx (x, t) + ψ (x, t) − c ω(x, t)
which was used for the uncertainty-free case. Differentiating both
+ r(x)µ(t) + ζ (x, t), (66) sides of (46) with respect to x and t produces (72) given in Box I.
Again, (72) has to hold for all (x, t) ∈ [0, 1]×[0, ∞). Interestingly,
ωx (0, t) = 0, (67)
in this case, the inverse backstepping kernel ℓ(x, y) still satisfies
ωx (1, t) = − (c1 + δ )ω(1, t) − p10 µ(t), (68) the kernel PDE (47)–(49), whose analytic solution is given by (50).
Now, define the following:
where r(x) is an unknown function to be determined and ψ (x, t)
satisfies (39). The term ζ (x, t) is the transformed function from (⋆) := p1 (x) + ℓy (x, 1) + (δ + p10 )ℓ(x, 1)
ε (x, t) by the backstepping transformation with kernel function ∫ 1
K (x, y): + ℓ(x, y)p1 (y)dy, (73)
∫ 1 x
1
ε (x, t) = ζ (x, t) − K (x, y)ζ (y, t)dy.
∫
(69)
x
(⋆⋆) := r(x) + p1 (x) + p10 ℓ(x, 1) + ℓ(x, y)p1 (y)dy. (74)
x
7
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

∫ 1
ω(y, t) Kxx (x, y) − Kyy (x, y) + cK (x, y) dt + ω(1, t) p1 (x) + Ky (x, 1) + (c1 + δ) K (x, 1)
[ ] [ ]
x
1
∂
[ ∫ ] [ ]
+ µ(t) r(x) − K (x, y)r(y)dy + p1 (x) + p10 K (x, 1) − ω(x, t) 2 K (x, x) + c = 0. (70)
x ∂x

1 1
∫ [ ∫ ]
T̃ (y, t) −ℓxx (x, y) + ℓyy (x, y) + c ℓ(x, y) dt − T̃ (1, t) p1 (x) + ℓy (x, 1) + (δ + p10 ) ℓ(x, 1) + ℓ(x, y)p1 (y)dy
[ ]
x x
1
∂
[ ∫ ] [ ]
− µ(t) r(x) + p1 (x) + p10 ℓ(x, 1) + ℓ(x, y)p1 (y)dy − T̃ (x, t) 2 ℓ(x, x) + c = 0. (72)
x ∂x

Box I.

Hence, a sufficient condition for (72) to hold for all (x, t) ∈ [0, 1]× Substitute r(x) from (75), p1 (x) in (44), and p10 from (45), alto-
[0, ∞) is (⋆) = 0 and (⋆⋆) = 0. gether into (71), and we arrives at (84) which is given in Box II, in
which the second equality stems from substituting ℓy (x, 1) using
Proposition 4. Let (⋆) and (⋆⋆) be as in (73) and (74). Then (⋆) = 0 (83), and the last equality is obtained by (76).
and (⋆⋆) = 0 for all x ∈ [0, 1], if and only if
5.3. Stability of the target system
r(x) = ℓy (x, 1) + δℓ(x, 1). (75)
We prove Proposition 4 with the assistance of the following The disturbance µ(t) that appears in the target system via
Lemma. the boundary condition (68) becomes problematic when proving
the stability of the target system, due to cross terms involving
Lemma 2. For K (x, y) and ℓ(x, y), which are the kernel functions uncertainty µ and state ω if L2 norm is employed as the Lya-
for the backstepping and the inverse backstepping transformations, punov functional candidate. To tackle this issue, we utilize the
it holds from Krstic and Smyshlyaev (2008) and Meurer (2013) that following invertible transformation ω(x, t) ↦ → z(x, t) to transfer
∫ y the boundary uncertainties into the in-domain dynamics:
ℓ(x, y) = K (x, y) + K (x, ξ )ℓ(ξ , y)dξ (76) 1
∫x y ω(x, t) = z(x, t) − p10 µ(t)(x2 − 1). (85)
2
= K (x, y) + ℓ(x, ξ )K (ξ , y)dξ . (77) The second term at the right-hand side of (85) is leveraged to can-
x
cel the uncertainty term in the boundary condition of the target
We are now positioned to provide the proof for Proposition 4 system. It is also worth noting that this transformation preserves
utilizing Lemma 2. z(1, t) = ω(1, t). Differentiating the above transformation with
respect to x and t yields the dynamics for z(x, t):
Proof. Differentiate both sides of (77) with respect to y, and
evaluate at y = 1: zt (x, t) = zxx (x, t) + ψ (x, t) − cz(x, t)
∫ 1 + eµ (x, t) + ζ (x, t), (86)
ℓy (x, 1) = Ky (x, 1) + ℓ(x, 1)K (1, 1) + ℓ(x, ξ )Ky (ξ , 1)dξ . (78)
x where eµ (x, t) is a term associated with the uncertainty, given by
Next, we substitute Ky (x, 1) and K (1, 1) from (44)–(45) into (78): p10
eµ (x, t) = [r(x) − p10 ]µ(t) +(x2 − 1)[c µ(t) + µt (t)]. (87)
ℓy (x, 1) = − p1 (x) − (c1 + δ )K (x, 1) + ℓ(x, 1)(c1 − p10 ) 2
∫ 1 The boundary conditions for z(x, t) is given by
− ℓ(x, ξ ) [p1 (ξ ) + (c1 + δ )K (ξ , 1)] dξ . (79)
zx (0, t) = ωx (0, t) = 0, (88)
x

Re-arranging terms in (79) yields zx (1, t) = ωx (1, t) + p10 µ(t) = −(c1 + δ )z(1, t), (89)
1
from which we could observe that the uncertainty µ is no longer
[ ∫ ]
(⋆) = −(c1 + δ ) K (x, 1) − ℓ(x, 1) + ℓ(x, ξ )K (ξ , 1)dξ . (80) at the boundary conditions in the z-system. We provide the
x
estimation error stability in the sense of L2 norm through the
In addition, note from (77) that analysis of the transformed system (86)–(89).
∫ 1
ℓ(x, 1) = K (x, 1) + ℓ(x, ξ )K (ξ , 1)dξ . (81) Theorem 5 (Convergence of Robust Observer). Consider the ob-
x
server error dynamics (63)–(65), and let the observer gains p1 (x)
In view of (80) and (81), we have (⋆) = 0. Moreover, and p10 be as in (44) and (45). Given Lipschitz constant γ , if the
design parameters c and c1 are selected such that (57) is satisfied,
(⋆) − (⋆⋆) = ℓy (x, 1) + δℓ(x, 1) − r(x). (82)
then in the presence of the in-domain uncertainty and the output
Hence, (⋆⋆) = 0 if and only if r(x) verifies (75). Finally, we must uncertainty, i.e., ε (x, t) ̸ = 0 and µ(t) ̸ = 0, the estimation error
check if the proposed r(x) in (75) satisfies ∆(x) = 0. First, notice
 
remains bounded in the sense of L2 norm denoted by T̃ (·, t) ≤
from (76) that RB (c , c1 ) as t → ∞, where
1
β (c , c1 )
∫ [ ]
ℓy (x, 1) = Ky (x, 1) + ℓ(1, 1)K (x, 1) + K (x, ξ )ℓy (ξ , 1)dξ . (83) RB (c , c1 ) := [1 + η(c)] · + λ(c , c1 ) . (90)
x α (c)
8
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

∫ 1
( c)
∆(x) = ℓy (x, 1) + δℓ(x, 1) − Ky (x, 1) − (c1 + δ )K (x, 1) + c1 + K (x, 1) − K (x, ξ ) ℓy (ξ , 1) + δℓ(ξ , 1) dξ
[ ]
2 x
1
[ ∫ ]
= δ ℓ(x, 1) − K (x, 1) − K (x, ξ )ℓ(ξ , 1)dξ ≡ 0. (84)
x

Box II.

Specifically, Substituting (97)–(99) and (101) into the right-hand side of (96),
1 and applying the Poincaré inequality, −∥zx ∥2 ≤ 21 z 2 (1) − 14 ∥z ∥2 ,
α (c) := c − κ (c ; γ ) + , (91) to the second term on the right-hand side of (96), yields
4
β (c , c1 ) := κ (c ; γ )λ(c , c1 ) + ∥r(x) − p10 ∥µ
( )
1
Ẇ2 ≤ − c1 + δ − z 2 (1) − α (c)∥z ∥2 + β (c , c1 )∥z ∥. (102)
2
√
2
+ |p10 | U + (1 + ρ (c))ε, (92)
If c1 is chosen such that c1 ≥ 1/2 −δ and α (c) > 0, or equivalently
15
√ c > κ (c ; γ ) − 1/4, the comparison principle (Khalil, 1996) applied
2
λ(c , c1 ) := |p10 | µ, (93) to (102) gives
15
β (c , c1 ) β (c , c1 ) −α(c)t
[ ]
where U = maxt {c µ(t) + µ̇(t)}. ∥z(·, t)∥ ≤ + ∥z(·, 0)∥ − e . (103)
α (c) α (c)
Proof. Consider the Lyapunov functional candidate, Since the inverse transformation z(x, t) ↦ → ω(x, t) is unique,
1
∫ 1
1 1
W2 (t) = z 2 (x, t)dx = ∥z(x, t)∥2 . (94) z(x, t) = ω(x, t) + p10 µ(t)(x2 − 1), (104)
2 2 2
0

The time derivative of the Lyapunov function W2 (t) along the and the backstepping transformation (35) is invertible, an upper
trajectory of z(x, t) can be computed as bound of ∥T̃ (·, t)∥ is computed via the sequence of inequalities
leveraging (56) and (97):
∫ 1
Ẇ2 (t) = z(x, t)zt (x, t)dx ∥T̃ (·, t)∥ ≤ [1 + η(c)] · [∥z(·, t)∥ + λ(c , c1 )] . (105)
0
∫ 1 ∫ 1 ∫ 1 When t → ∞, the exponential terms at the right-hand side of
= zzxx dx + z ψ dx − c z 2 dx (105) decays to zero. Under this scenario,
0 0 0
β (c , c1 )
[ ]
∫ 1 ∫ 1
+ zeµ dx + z ζ dx. (95) ∥T̃ (·, t)∥ ≤ [1 + η(c)] · + λ(c , c1 ) . (106)
α (c)
0 0

Applying the integration by parts to the first term, and Cauchy– This completes the proof.
Schwarz inequality to the second, fourth, fifth, and sixth term at
the right-hand side of (95) results in Remark 7. The estimation error bound RB is characterized
by the observer gains, Lipschitz constant, and bounds on the
Ẇ2 (t) ≤ − (c1 + δ )z (1) − ∥zx ∥ − c ∥z ∥
2 2 2
uncertainties. For a given Lipschitz constant γ < γ ∗ , one can
+ (∥ψ∥ + ∥eµ ∥ + ∥ζ ∥)∥z ∥. (96) minimize the size of the error ball RB to achieve robust estimation
by formulating and solving the optimization problem
Now consider the transformation (85) and apply triangular in-
equality to conclude min RB (c , c1 )
c , c1
1 1
∥ω∥ ≤ ∥z ∥ + ∥p10 µ(t)(1 − x2 )∥ ≤ ∥z ∥ + λ(c , c1 ), (97) s.t. κ (c ; γ ) − c − <0
2 4 (107)
so that the upper bound of ∥ψ∥ given by (51) is explicitly ex- 1
− c1 + −δ ≤0
pressed in terms of state in z system: 2
c>0
∥ψ∥ ≤ κ (c ; γ )∥z ∥ + κ (c ; γ )λ(c , c1 ). (98)
Furthermore, from (87), the following upper bound is imposed on 6. Simulation studies
eµ ,
√ In this section, we present simulation studies to demonstrate
2 the performance of the proposed thermal observer on a com-
∥eµ ∥ ≤ ∥r(x) − p10 ∥µ + |p10 | U . (99)
15 mercial Lithium-Iron Phosphate (LFP) A123 26650 cylindrical cell
Finally, by recognizing the inverse backstepping transformation with rated capacity of 2.3 Ah. In this study, the one-dimensional
ζ (x, t) ↦→ ε (x, t), thermal model (14)–(18) is used as the plant model. The model
parameters have been previously identified in Dey et al. (2019)
∫ 1
and Liao, Zuo, Ma, Chen, An, Gao, and Yin (2012) using exper-
ζ (x, t) = ε (x, t) + ε (y, t)ℓ(x, y)dy, (100)
x imental data, and are enumerated in Table 1. The calculation
leveraging the formula in (20) using these parameters yields
similar strategy as that in Lemma 1 and (55) can be employed to
derive an upper bound on ∥ζ (·, t)∥:
γ = 0.0072 |I |2max . Hence, the analysis in Section 4 indicates that,
to guarantee the sufficient conditions of observer convergence,
∥ζ (·, t)∥ ≤ [1 + ρ (c)]ε. (101) the absolute maximum value of the applied current is |I |max
9
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

Table 1
Model Parameters for a commercial LFP A123
26650 cylindrical Cell.
Symbols Values Units
Tref 298.15 K
Tamb 298.15 K
Rs,ref 15 m
Ea 33.8 kJ/mol
k 0.61 W/(m K)
h 69.89 W/(m2 K)
R 8.314 J/(mol K)
D 2118 kg/m3
cp 711 J/(kg K)
Cbatt 2.26 Ah

Fig. 3. Observer results for a 4C constant current. Observer initial condition:

T̂ (x, 0) = T0 (x) + 10 K for all x ∈ [0, 1]. The estimation error converges to zero
exponentially in the sense of L2 spatial norm, thus confirming the conclusion
from Theorem 3.

difference method is employed to numerically discretize the plant

model. More precisely, the plant model is discretized spatially
with 100 grid points. The left and right boundary conditions are
handled by the second-order forward and backward approxima-
tions, respectively. The simulation is numerically implemented in
MATLAB⃝ R
. Suppose the battery cell has been in resting prior to
the simulation onset such that the initial condition of the plant
model is uniformly set to the ambient temperature, i.e., T0 (x) =
Tamb = 298.15 K for all x ∈ [0, 1]. As such, the measured
boundary signals and the spatially distributed temperature profile
are sketched in Fig. 2(b) and Fig. 2(c), respectively. Further, the
absolute maximum current in this specific case is |I |max = 4C.
With this, the expression in (20) provides the numerical value of
the Lipschitz constant for the nonlinearity function f with respect
to T in the plant model, i.e., γ = 0.59 < γ ∗ . This guarantees
the existence of a design parameter c such that the observer
convergence conditions presented in Theorem 3 are fulfilled. In
Fig. 2. The plant model response under a 4C constant current. Plant model initial fact, the feasible ranges for the design parameters in this case are
condition: T0 (x) = Tamb = 298.15 K for all x ∈ [0, 1]. 0.49 < c < 13.5 and c1 ≥ −0.53. With the choice of c = 8
and c1 = 2, the in-domain observer gain p1 (x) and the boundary
observer gain p10 can be computed with the assistance of back-
√ ∗
= γ /0.0072 ≈ 12A (equivalently 5.4C). This current rate is stepping kernels as well as (44) and (45). Ultimately, Fig. 3(a)
generally well beyond the operating current limits of a cylindrical presents the observer state, whereas Fig. 3(b) shows exponential
LFP cell — reaffirming the applicability of the proposed observer convergence of L2 norm of temperature estimation error to zero,
even under extreme practical conditions. Let the premises in despite incorrect initial conditions (T̂0 (x) = T0 (x) + 10 K for all
Assumption 1 hold, we evaluate the effectiveness of the observer x ∈ [0, 1]). The estimates T̂ converge to the true value T within
with and without the presence of in-domain and output uncer- 100 s. The numerical simulation results confirm our analysis in
tainties. For the purpose of demonstrating the estimation results, Theorem 3. It additionally demonstrates promising estimation
the state estimates are initialized with incorrect values to imitate performance under high-rate operating modes which are suitable
uncertainty in initial conditions. Furthermore, we quantify the for various battery fast charging applications.
performance of the observers in terms of L2 estimation error Next, we apply an electric vehicle charge/discharge cycle. This
convergence time, defined as the time spent to reach within ±0.2 input current profile shown in Fig. 4(a) is generated from concate-
K band of zero starting from a non-zero initial condition. nating two Federal Urban Driving Schedule (FUDS) drive cycles
referenced in Xing, He, Pecht, and Tsui (2014). The absolute
6.1. No in-domain and output uncertainties maximum C-rate is scaled to 5.3C — confirming the existence
of a design parameter c such that the observer convergence
We first encapsulate the performance of the proposed temper- conditions in Theorem 3 are verified. From a simple computation,
ature estimation approach under no uncertainties, i.e., ε (x, t) = 0 this current profile would roughly cause 50.1% change in battery
and µ(t) = 0, ∀(x, t) ∈ [0, 1] × [0, ∞). A 4C constant current, SOC. In particular, the boundary measurements and temperature
shown in Fig. 2(a), is applied to the plant model to generate bat- responses are plotted in Fig. 4(b) and Fig. 4(c). It can be ob-
tery thermal responses. The simulation time is t ∈ [0, 800 s], and served that this dynamic current profile produces less aggressive
the spatial domain is r ∈ [0, 0.013 m], i.e., x ∈ [0, 1]. The finite temperature variations across spatial and temporal domains. This
10
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

Fig. 4. The plant model response under a FUDS drive cycle. Plant model initial Fig. 6. The plant model response under a UDDS drive cycle. Plant model initial
condition: T0 (x) = Tamb = 298.15 K for all x ∈ [0, 1]. condition: T0 (x) = Tamb + 1 K = 299.15 K for all x ∈ [0, 1]. Gaussian noise
with 0.3 K variance and zero mean, i.e., µ(t) ∼ N (0, 1), is injected to the
boundary measurement signal y(t). Uniformly distributed random noise ε (x, t)
is introduced into the plant model (14).

6.2. With in-domain and output uncertainties

Finally, we demonstrate simulation results for the robust ob-

server acting on the plant model with uncertainties. Gaussian
noises with 0.3 K variance and zero mean, i.e., µ(t) ∼ N (0, 0.3),
is manually injected to the boundary measurement signal T (1, t)
(see Eq. (18)). Further, a uniformly distributed random noise
ε (x, t) is introduced to the plant model (14). The trajectory of the
applied current, which is extracted from the Urban Dynamometer
Driving Schedule (UDDS) driving cycle, is shown in Fig. 6(a). The
choice of UDDS is to imitate a practical electric vehicle driving
pattern. The noise-corrupted boundary measurement y(t) and
the spatially distributed temperature profile T (x, t) are plotted in
Fig. 6(b) and Fig. 6(c), respectively. Notice that the same observer
gains p1 (x) and p10 as those from the uncertainty-free case are
Fig. 5. Observer results for a FUDS drive cycle. Observer initial condition: utilized. Once again, the observer is initialized uniformly to an
T̂ (x, 0) = T0 (x) + 5 K for all x ∈ [0, 1]. The estimation error converges to zero incorrect value, and the L2 norm of the estimation error quickly
exponentially within 120 s in the sense of L2 norm. converges to a ball of radius RB around the equilibrium point
∥T̃ ∥ = 0, as can be seen in Fig. 7(b). Recall that Theorem 5
only guarantees boundedness of estimation error, in the sense
is because unlike the high-rate constant current input from the of L2 norm. These results are generated by setting c = 0.5 and
previous study, this dynamic cycle is characterized mostly by c1 = 1/2 − δ = −0.99, and the numerical value of RB (c , c1 )
non-sustained C-rates. To further emulate real-world scenarios computed by (90) is also provided in Fig. 7(b) to illustrate the
and evaluate the observer’s robustness with respect to parametric effectiveness of the derived estimation error bounds. Notably,
uncertainty in the surface heat transfer coefficient h, we inject the convergence time is within 540 s, whereas RB tightly bounds
10% error in parameter h used in the observer simulation. Con- the estimation error for all t ≥ 390 s. The enabling step for
sequently, regardless of an incorrect initialization, the observer the proposed robust observer approach is to solve the nonlinear
converges exponentially in the sense of L2 norm after the initial optimization problem in Remark 7 to minimize the magnitude of
transient (∼110 s), as visualized in Fig. 5(b). Moreover, the simu- RB . Herein, it is also highlighted that RB expands monotonically
lation reported in Fig. 5 demonstrates no apparent estimation bias with respect to parameter c, whereas the observer convergence
from the uncertainty in parameter h, highlighting the observer’s rate becomes faster as c increases. There is then a trade-off
robustness against the imperfection in h values. between observer convergence rate and the allowable magnitude
11
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

Al Hallaj, Said, Maleki, Hossein, Hong, Jong-Sung, & Selman, J. Robert (1999).
Thermal modeling and design considerations of lithium-ion batteries. Journal
of Power Sources, 83(1–2), 1–8.
Bernardi, D., Pawlikowski, E., & Newman, John (1985). A general energy balance
for battery systems. Journal of The Electrochemical Society, 132(1), 5–12.
Broussely, Michel, Biensan, Ph, Bonhomme, F., Blanchard, Ph, Herreyre, S.,
Nechev, K., et al. (2005). Main aging mechanisms in Li ion batteries. Journal
of Power Sources, 146(1–2), 90–96.
Chaturvedi, Nalin A., Klein, Reinhardt, Christensen, Jake, Ahmed, Jasim, &
Kojic, Aleksandar (2010). Algorithms for advanced battery-management
systems. IEEE Control Systems Magazine, 30(3), 49–68.
Chen, S. C., Wan, C. C., & Wang, Y. Y. (2005). Thermal analysis of lithium-ion
batteries. Journal of Power Sources, 140(1), 111–124.
Cheng, Meng-Bi, Radisavljevic, Verica, & Su, Wu-Chung (2011). Sliding mode
boundary control of a parabolic PDE system with parameter variations and
boundary uncertainties. Automatica, 47(2), 381–387.
Couto, Luis D., Zhang, Dong, Aitio, Antti, Moura, Scott, & Howey, David (2020). Es-
timation of parameter probability distributions for lithium-ion battery string
models using Bayesian methods. In Dynamic systems and control conference,
(Vol. 84270). American Society of Mechanical Engineers, V001T20A003.
Dey, Satadru, Ayalew, Beshah, & Pisu, Pierluigi (2015). Nonlinear robust ob-
servers for state-of-charge estimation of lithium-ion cells based on a reduced
Fig. 7. Observer results for a UDDS cycle, subject to in-domain and output electrochemical model. IEEE Transactions on Control Systems Technology, 23(5),
uncertainties. Observer initial condition: T̂ (x, 0) = T0 (x) + 5 K for all x ∈ [0, 1]. 1935–1942.
The estimation error converges to a ball of radius RB around the equilibrium Dey, Satadru, Biron, Zoleikha Abdollahi, Tatipamula, Sagar, Das, Nabarun, Mo-
point T̃ = 0 in the sense of L2 norm, thus justifying the conclusion from hon, Sara, Ayalew, Beshah, et al. (2016). Model-based real-time thermal fault
Theorem 5. diagnosis of lithium-ion batteries. Control Engineering Practice, 56, 37–48.
Dey, S., Perez, H. E., & Moura, S. J. (2019). Model-based battery thermal fault
diagnostics: Algorithms, analysis, and experiments. IEEE Transactions on
Control Systems Technology, [ISSN: 1063-6536] 27(2), 576–587. http://dx.doi.
of RB . The choice of parameter c for this simulation study was org/10.1109/TCST.2017.2776218.
carefully calibrated such that convergence rate and derived error Doughty, Daniel H., Butler, Paul C., Jungst, Rudolph G., & Roth, E. Peter (2002).
bound simultaneously remain reasonable. Lithium battery thermal models. Journal of Power Sources, 110(2), 357–363.
Du, Ronghua, Hu, Xiaosong, Xie, Shaobo, Hu, Lin, Zhang, Zhiyong, & Lin, Xianke
(2020). Battery aging-and temperature-aware predictive energy management
7. Conclusion and future works for hybrid electric vehicles. Journal of Power Sources, 473, Article 228568.
Evans, T. I., & White, Ralph E. (1989). A thermal analysis of a spirally wound
The knowledge of real-time battery internal temperature in- battery using a simple mathematical model. Journal of The Electrochemical
formation enables safe and reliable operations. In this regard, Society, 136(8), 2145.
an infinite-dimensional PDE boundary observer is proposed for a Forgez, Christophe, Do, Dinh Vinh, Friedrich, Guy, Morcrette, Mathieu, & Dela-
court, Charles (2010). Thermal modeling of a cylindrical LiFePO4 /graphite
one-dimensional battery thermal model subject to in-domain and lithium-ion battery. Journal of Power Sources, 195(9), 2961–2968.
output uncertainties. The estimation error system is converted Gu, W. B., & Wang, Chao-Yang Thermal and electrochemical coupled modeling
to a prescribed stable target system, from which the exponential of a lithium-ion cell. In Proceedings of the ECS, (Vol. 99) (pp. 748–762).
stability of the error dynamics is mathematically derived without Hu, Y., & Yurkovich, S. (2012). Battery cell state-of-charge estimation using linear
the presence of system uncertainties, thanks to the Lipschitz parameter varying system techniques. Journal of Power Sources, 198, 338–350.
Käbitz, Stefan, Gerschler, Jochen Bernhard, Ecker, Madeleine, Yurdagel, Yusuf,
continuous nonlinearity. In the presence of system uncertainties,
Emmermacher, Brita, André, Dave, et al. (2013). Cycle and calendar life study
we transfer the bounded uncertainties in the boundary conditions of a graphite | LiNi1/3 mn1/3 co1/3 o2 Li-ion high energy system. Part A: Full
of the target system into the in-domain dynamics, and the cor- cell characterization. Journal of Power Sources, 239, 572–583.
responding estimation error converges to an error ball around Keyser, Matthew, Pesaran, Ahmad, Li, Qibo, Santhanagopalan, Shriram,
the equilibrium point, which can be explicitly characterized by Smith, Kandler, Wood, Eric, et al. (2017). Enabling fast charging–battery
thermal considerations. Journal of Power Sources, 367, 228–236.
uncertainty bounds, observer parameters, and model parameters.
Khalil, Hassan K. (1996). Noninear systems, (Vol. 2). (5), New Jersey: Prentice-Hall,
This scheme is regarded as a leap forward in the effort to de- 5–1.
sign estimation algorithms for battery high-fidelity thermal mod- Kim, Youngki, Mohan, Shankar, Siegel, Jason B., Stefanopoulou, Anna G., &
els, without relying on prior spatial discretization of the PDEs. It Ding, Yi (2014). The estimation of temperature distribution in cylindrical
is also one of the first to simultaneously consider PDE model non- battery cells under unknown cooling conditions. IEEE Transactions on Control
Systems Technology, 22(6), 2277–2286.
linearity, modeling uncertainty, and measurement uncertainty
Kim, Gi-Heon, Pesaran, Ahmad, & Spotnitz, Robert (2007). A three-dimensional
within a single observer design framework. thermal abuse model for lithium-ion cells. Journal of Power Sources, 170(2),
Certain model reductions are made in this study, and their re- 476–489.
laxation incentivizes future research directions. This includes ex- Kim, Hahnsang, & Shin, Kang G. (2011). Efficient sensing matters a lot for
tending the proposed estimation algorithm to high-dimensional large-scale batteries. In 2011 IEEE/ACM second international conference on
cyber-physical systems (pp. 197–205). IEEE.
thermal PDE models, and exploring less conservative necessary
Klein, Reinhardt, Chaturvedi, Nalin A., Christensen, Jake, Ahmed, Jasim, Find-
and sufficient conditions to ensure observer’s exponential con- eisen, Rolf, & Kojic, Aleksandar (2013). Electrochemical model based observer
vergence. One could also consider coupling thermal effects with design for a lithium-ion battery. IEEE Transactions on Control Systems
battery electrochemical models (Zhang, Couto, & Moura, 2020a). Technology, 21(2), 289–301.
Besides, optimal thermal control and sensor placement could be Koniak, M., & Czerepicki, A. (2017). Selection of the battery pack parameters
explored in battery packs (Couto, Zhang, Aitio, Moura, & Howey, for an electric vehicle based on performance requirements. In IOP conference
series: Materials science and engineering, (Vol. 211). IOP Publishing, Article
2020). 012005.
Krstic, Miroslav (1999). On global stabilization of Burgers’ equation by boundary
References control. Systems & Control Letters, 37(3), 123–141.
Krstic, Miroslav, & Smyshlyaev, Andrey (2008). Boundary control of PDEs: A course
Ahmed-Ali, Tarek, Giri, Fouad, Krstic, Miroslav, Burlion, Laurent, & Lamnabhi- on backstepping designs, (Vol. 16). Siam.
Lagarrigue, Françoise (2016). Adaptive boundary observer for parabolic PDEs Liao, Lixia, Zuo, Pengjian, Ma, Yulin, Chen, XinQun, An, Yongxin, Gao, Yunzhi, et
subject to domain and boundary parameter uncertainties. Automatica, 72, al. (2012). Effects of temperature on charge/discharge behaviors of LiFePO4
115–122. cathode for Li-ion batteries. Electrochimica Acta, 60, 269–273.

12
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

Liaw, Bor Yann, Roth, E. Peter, Jungst, Rudolph G., Nagasubramanian, Ganesan, Zhang, Dong, Couto, Luis D., Park, Saehong, Gill, Preet, & Moura, Scott J. (2020b).
Case, Herbert L., & Doughty, Daniel H. (2003). Correlation of arrhenius Nonlinear state and parameter estimation for li-ion batteries with thermal
behaviors in power and capacity fades with cell impedance and heat coupling. IFAC-PapersOnLine, 53(2), 12479–12484.
generation in cylindrical lithium-ion cells. Journal of Power Sources, 119, Zhang, Dong, Dey, Satadru, Couto, Luis D., & Moura, Scott J. (2019a). Battery
874–886. adaptive observer for a single-particle model with intercalation-induced
Lin, Xinfan, Perez, Hector E., Mohan, Shankar, Siegel, Jason B., Ste- stress. IEEE Transactions on Control Systems Technology.
fanopoulou, Anna G., Ding, Yi, et al. (2014). A lumped-parameter electro-
Zhang, Dong, Dey, Satadru, Perez, Hector E., & Moura, Scott J. (2017). Remaining
thermal model for cylindrical batteries. Journal of Power Sources, 257,
useful life estimation of lithium-ion batteries based on thermal dynamics. In
1–11.
2017 American control conference (pp. 4042–4047). IEEE.
Lin, Xinfan, Perez, Hector E., Siegel, Jason B., Stefanopoulou, Anna G., Li, Yonghua,
Anderson, R. Dyche, et al. (2013). Online parameterization of lumped Zhang, Dong, Dey, Satadru, Perez, Hector E., & Moura, Scott J. (2019b). Real-time
thermal dynamics in cylindrical lithium ion batteries for core temperature capacity estimation of lithium-ion batteries utilizing thermal dynamics. IEEE
estimation and health monitoring. IEEE Transactions on Control Systems Transactions on Control Systems Technology.
Technology, 21(5), 1745–1755. Zhang, Dong, Tang, Shu-Xia, & Moura, Scott J. (2019c). State and disturbance
Meurer, Thomas (2013). On the extended luenberger-type observer for semilin- estimator for unstable reaction-advection-diffusion PDE with boundary dis-
ear distributed-parameter systems. IEEE Transactions on Automatic Control, turbance. In 2019 Proceedings of the conference on control and its applications
58(7), 1732–1743. (pp. 67–74). SIAM.
Moura, Scott J., Argomedo, Federico Bribiesca, Klein, Reinhardt, Mirta-
batabaei, Anahita, & Krstic, Miroslav (2017). Battery state estimation for a
single particle model with electrolyte dynamics. IEEE Transactions on Control
Systems Technology, 25(2), 453–468.
Muratori, Matteo, Canova, Marcello, & Guezennec, Yann (2012). A spatially- Dong Zhang received his Ph.D. and M.S. degrees from
reduced dynamic model for the thermal characterisation of Li-ion battery the University of California at Berkeley, Berkeley, CA,
cells. International Journal of Vehicle Design, 58(2), 134. USA, in 2020 and 2016, respectively, all in systems
Nambu, Takao (1984). On the stabilization of diffusion equations: boundary and control engineering. He also earned his B.S. degree
observation and feedback. Journal of Differential Equations, 52(2), 204–233. from the University of Michigan, Ann Arbor, MI, USA,
Park, Chanwoo, & Jaura, Arun K. (2003). Dynamic thermal model of Li-ion battery
in 2015, in Civil and Environmental Engineering. He
for predictive behavior in hybrid and fuel cell vehicles: Technical report, SAE
was also awarded a double B.S. degree in Electrical
Technical Paper.
and Computer Engineering from Shanghai Jiao Tong
Pazy, Amnon (2012). Semigroups of linear operators and applications to partial
University, Shanghai, China, in 2015. He is currently a
differential equations, (Vol. 44). Springer Science & Business Media.
Postdoctoral Research Associate with the Department
Plett, Gregory L. (2004). Extended Kalman filtering for battery management
of Mechanical Engineering at Carnegie Mellon Univer-
systems of LiPB-based HEV battery packs: Part 3. State and parameter
sity, Pittsburgh, PA, USA. His current research interests include dynamical system
estimation. Journal of Power Sources, 134(2), 277–292.
estimation and controls, optimization, machine learning, renewable energy sys-
Rajamani, Rajesh (1998). Observers for Lipschitz nonlinear systems. IEEE
tems, energy storage, and advanced Lithium-Ion battery management systems.
Transactions on Automatic Control, 43(3), 397–401.
Dr. Zhang is a recipient of the American Society of Mechanical Engineers
Richardson, Robert R., Ireland, Peter T., & Howey, David A. (2014). Bat-
(ASME) Energy System Best Paper Award and Finalist at 2020 American Control
tery internal temperature estimation by combined impedance and surface
Conference (ACC) and 2020 Dynamic Systems and Control Conference (DSCC).
temperature measurement. Journal of Power Sources, 265, 254–261.
Samad, Nassim A., Siegel, Jason B., & Stefanopoulou, Anna G. (2014). Param-
eterization and validation of a distributed coupled electro-thermal model
for prismatic cells. In Dynamic systems and control conference, (Vol. 46193). Satadru Dey received the master’s degree in control
American Society of Mechanical Engineers, V002T23A006. systems from IIT Kharagpur, India, in 2010, and the
Smith, Kandler, & Wang, Chao-Yang (2006). Power and thermal characterization Ph.D. degree in automotive engineering from Clemson
of a lithium-ion battery pack for hybrid-electric vehicles. Journal of Power University in 2015. From 2015 to 2017, he was a Post-
Sources, 160(1), 662–673. Doctoral Researcher with the University of California at
Smyshlyaev, Andrey, & Krstic, Miroslav (2005). On control design for PDEs with Berkeley, Berkeley. From August 2017 to August 2020,
space-dependent diffusivity or time-dependent reactivity. Automatica, 41(9), he was an Assistant Professor with the University of
1601–1608. Colorado Denver. He is currently an Assistant Profes-
Smyshlyaev, Andrey, & Krstic, Miroslav (2010). Adaptive control of parabolic PDEs. sor with the Department of Mechanical Engineering,
Princeton University Press. The Pennsylvania State University. His current research
Song, Li, & Evans, James W. (2000). Electrochemical-thermal model of lithium interests include controls, smart cities, energy, and
polymer batteries. Journal of The Electrochemical Society, 147(6), 2086–2095. transportation systems.
Tang, Shu-Xia, Camacho-Solorio, Leobardo, Wang, Yebin, & Krstic, Miroslav
(2017). State-of-charge estimation from a thermal–electrochemical model of
lithium-ion batteries. Automatica, 83, 206–219. Shu-Xia Tang received her Ph.D. in Mechanical Engi-
Thomas, Karen E., & Newman, John (2003). Thermal modeling of porous insertion neering in 2016 from the Department of Mechanical
electrodes. Journal of The Electrochemical Society, 150(2), A176. & Aerospace Engineering, University of California, San
Tranter, Thomas George, Timms, Robert, Shearing, Paul R., & Brett, Dan (2020). Diego, USA. She is currently an assistant professor at
Communication—-Prediction of thermal issues for larger format 4680 cylin- the Department of Mechanical Engineering, Texas Tech
drical cells and their mitigation with enhanced current collection. Journal of University, USA. She is an IEEE senior member and
The Electrochemical Society. is an IEEE CSS (Control Systems Society) Technical
Vazquez, Rafael, & Krstic, Miroslav (2008a). Control of 1-D parabolic PDEs with Committee member on Distributed Parameter Systems.
Volterra nonlinearities, Part I: design. Automatica, 44(11), 2778–2790. She serves as an associate editor of Journal of Control,
Vazquez, Rafael, & Krstic, Miroslav (2008b). Control of 1D parabolic PDEs with Automation and Electrical Systems and as an Editorial
Volterra nonlinearities, Part II: analysis. Automatica, 44(11), 2791–2803. Board member of IEEE CSS and ASME DSCC (Dynamic
Wang, John, Liu, Ping, Hicks-Garner, Jocelyn, Sherman, Elena, Soukiazian, Souren, Systems and Control Division). Her main research interests are stability analysis,
Verbrugge, Mark, et al. (2011). Cycle-life model for graphite-LiFePO4 cells. estimation and control design of distributed parameter systems.
Journal of Power Sources, 196(8), 3942–3948.
Wang, Qingsong, Ping, Ping, Zhao, Xuejuan, Chu, Guanquan, Sun, Jinhua, &
Chen, Chunhua (2012). Thermal runaway caused fire and explosion of lithium
ion battery. Journal of Power Sources, 208, 210–224. Ross Drummond received a D.Phil. in control engineer-
Xing, Yinjiao, He, Wei, Pecht, Michael, & Tsui, Kwok Leung (2014). State of charge ing from the University of Oxford, U.K (2013–2017)
estimation of lithium-ion batteries using the open-circuit voltage at various where he is currently a UKIC research fellow in the
ambient temperatures. Applied Energy, 113, 106–115. Control Group. His main research interests include
Yang, Xiao-Guang, Liu, Teng, & Wang, Chao-Yang (2021). Thermally modulated the stability analysis of nonlinear feedback systems
lithium iron phosphate batteries for mass-market electric vehicles. Nature which he combines with more applied research on the
Energy, 1–10. modeling and design of electrochemical energy storage
Zhang, Dong, Couto, Luis D., & Moura, Scott J. (2020a). Electrode-level state devices, including Li-ion batteries and supercapacitors.
estimation in lithium-ion batteries via Kalman decomposition. IEEE Control
Systems Letters, 5(5), 1657–1662.

13
D. Zhang, S. Dey, S.-X. Tang et al. Automatica 133 (2021) 109849

Scott J. Moura is an Associate Professor in Civil and Centre Automatique et Systémes, MINES ParisTech, Paris, France. His research
Environmental Engineering and Director of the Energy, interests include control, optimization, and machine learning for batteries,
Controls, & Applications Lab (eCAL) at the University of electrified vehicles, and distributed energy resources.
California, Berkeley. He is also a faculty member at the Dr. Moura is a recipient of the National Science Foundation (NSF) CAREER
Tsinghua-Berkeley Shenzhen Institute. He received the Award, Carol D. Soc Distinguished Graduate Student Mentor Award, the Hellman
B.S. degree from the University of California, Berkeley, Fellowship, the O. Hugo Shuck Best Paper Award, the ACC Best Student Paper
CA, USA, and the M.S. and Ph.D. degrees from the Award (as advisor), the ACC and ASME Dynamic Systems and Control Confer-
University of Michigan, Ann Arbor, in 2006, 2008, ence Best Student Paper Finalist (as student and advisor), the UC Presidential
and 2011, respectively, all in mechanical engineering. Postdoctoral Fellowship, the NSF Graduate Research Fellowship, the University
From 2011 to 2013, he was a Post-Doctoral Fellow at of Michigan Distinguished ProQuest Dissertation Honorable Mention, the Uni-
the Cymer Center for Control Systems and Dynamics, versity of Michigan Rackham Merit Fellowship, and the College of Engineering
University of California, San Diego. In 2013, he was a Visiting Researcher at the Distinguished Leadership Award.

14