Cell-Level State of Charge Estimation for Battery

Packs Under Minimal Sensing

Dong Zhanga , Luis D. Coutoc , Ross Drummondd , Shashank Sripadb ,

Venkatasubramanian Viswanathanb,∗
arXiv:2109.08332v1 [eess.SY] 17 Sep 2021

a School of Aerospace and Mechanical Engineering, University of Oklahoma, 660 Parrington

Oval, Norman, OK, USA
b Department of Mechanical Engineering, Carnegie Mellon University, 5000 Forbes Avenue,

Pittsburgh, PA, USA

c Department of Control Engineering and System Analysis, Université Libre de Bruxelles,

B-1050 Brussels, Belgium

d Department of Engineering Science, University of Oxford, 17 Parks Road, OX1 3PJ,

Oxford, United Kingdom

Abstract

This manuscript presents an algorithm for individual Lithium-ion (Li-ion) bat-

tery cell state of charge (SOC) estimation in a large-scale battery pack under
minimal sensing, where only pack-level voltage and current are measured. For
battery packs consisting of up to thousands of cells in electric vehicle or station-
ary energy storage applications, it is desirable to estimate individual cell SOCs
without cell local measurements in order to reduce sensing costs. Mathemati-
cally, pure series connected cells yield dynamics given by ordinary differential
equations under classical full voltage sensing. In contrast, parallel–series con-
nected battery packs are evidently more challenging because the dynamics are
governed by a nonlinear differential–algebraic equations (DAE) system. The
majority of the conventional studies on SOC estimation for battery packs bene-
fit from idealizing the pack as a lumped single cell which ultimately lose track of
cell-level conditions and are blind to potential risks of cell-level over-charge and
over-discharge. This work explicitly models a battery pack with high fidelity
cell-by-cell resolution based on the interconnection of single cell models, and ex-
amines the observability of cell-level state with only pack-level measurements.

∗ Correspondingauthor
Email address: [email protected] (Venkatasubramanian Viswanathan)

Preprint submitted to Elsevier September 20, 2021

A DAE-based state observer with linear output error injection is formulated,
where the individual cell SOC and current can be reconstructed from minimal
number of pack sensing. The mathematically guaranteed asymptotic conver-
gence of differential and algebraic state estimates is established by considering
local Lipschitz continuity property of system nonlinearities. Simulation results
for Graphite/NMC cells illustrate convergence for cell SOCs, currents, and volt-
ages.
Keywords: Lithium-ion Battery Packs, Differential-Algebraic System,
Minimal Sensing, State Estimation.

1. Introduction

Lithium-ion (Li-ion) batteries have emerged as one of the most prominent

energy storage devices for large-scale energy applications, e.g., hybrid electric
vehicles (HEV), pure electric vehicles (EV), and smart grids, due to their high
energy and power density, low self-discharge, long lifetime, and rapidly falling
prices [1, 2]. A battery pack system generally consists of hundreds or thousands
of single cells connected via parallel and series connections in order to fulfill the
requirements of high-energy and high-power applications [3]. For example, the
75 kWh battery pack in Tesla Model 3 contains in total 4,416 Li-ion cylindrical
2170 cells organized in four modules (two 25 series and two 23 series) [4]. Each
series component has 46 battery cells wired in parallel. See an illustrative ex-
ample in Figure 1 for an EV with a battery pack as well as an on-board battery
management system (BMS). Meanwhile, it is also well-known that Li-ion cells in
battery packs are sensitive to over-(dis)charge, high currents, and degrade over
their lifetime [5]. Accurate estimation of the battery internal variables, such as
state of charge (SOC) and local currents, enables a battery management system
to ensure that individual in-pack cells do not violate safety constraints while
prolonging battery service life via online control and cell-level health diagnosis.
Battery pack system modeling can be divided into three categories. The
first approach treats the entire pack as one lumped single cell [6]. However,

2
Figure 1: An illustrative schematic of a battery pack inside an electric vehicle with an on-board
battery management system.

the internal states of individual cells within the pack are likely to be different.
Results from [7, 8] show that there exists considerable cell-to-cell variability
even for cells manufactured in the same batch. Therefore, within a pack, some
cells are more prone to violate safety-critical constraints than others, and such
issues cannot be resolved from the lumped single cell approach. The second
modeling approach also relies on a single cell model, but it focuses on specific
in-pack cells – the weakest and the strongest ones, as representatives of the pack
dynamics [3, 9]. Although convenient from the computational viewpoint, these
extreme cells need to be identified and they may change with time, and this is
generally not considered. The last approach is based on the interconnection of
single cell models [10, 11, 12]. This approach benefits from high fidelity cell-by-
cell resolution, but it might suffer from high real-time computational burden.
To counteract this computational challenge, most of these approaches resort to
equivalent circuit models, which tend to have a low complexity when compared
to more sophisticated electrochemical models.
Typical battery packs in EVs and storage applications, as mentioned previ-
ously, usually contain hundreds or thousands of cells. Conventionally, to care-
fully monitor every single cell with divergence of characteristics, voltage sensing
is placed on each cell or every group of cells in parallel [5]. In practice, this task
inevitably demands a significant amount of sensing hardware and labor, and is

3
considered to be uneconomical. Moreover, as pointed out in [13], the failure
rate of the battery system increases with more hardware components. Further,
large-scale battery packs generate massive amount of data if a large number
of voltage sensors are deployed, challenging the BMS’s storage and computa-
tional capability limits [14]. All these issues motivate a substantial reduction
on number of sensing hardware.
In battery systems-and-control community, most of the existing studies on
SOC estimation for battery packs resort to idealizing the entire battery pack as
a lumped cell and define a pack-level SOC [3, 15, 16, 17, 18]. However, such
methods completely lose track of cell-level conditions and are blind to poten-
tial risks of cell-level over-charge and over-discharge. To date, the cell-level
SOC estimation problem in a battery pack under reduced sensing has not been
sufficiently explored. Series cell SOC estimation using only the total voltage
measurement has been studied previously in [5, 19, 20], whereas the estimation
for parallel (and parallel-series) configuration has been overlooked for multiple
reasons. First, the cells in parallel are widely considered to behave as one single
cell. However, an implicit assumption behind this reasoning is that the applied
current is evenly split amongst the cells in parallel. This is rarely true in practice
due to cell heterogeneity, such as non-uniform parameter values and tempera-
tures [21]. Secondly, the estimation problem for battery cells in series is arguably
easier to solve than the parallel counterpart, because in the series case, the input
current to each battery cell is the same and it can be practically measured. In
the parallel and parallel-series cases, each cell’s local current is unknown and
determined by nonlinear algebraic constraints. Under reduced sensing scenario,
the parallel configuration turns out to be a differential-algebraic equation (DAE)
system that requires non-trivial state estimation theories [22].
A DAE system, a.k.a. a descriptor or singular system, involving both dif-
ferential and algebraic equations, is a powerful modeling framework that gener-
alizes ordinary differential (normal) systems [23]. The state observer design for
linear descriptor systems is a rich research topic [24, 25, 26]. In contrast, state
observers for nonlinear DAE systems is less prolific. Some relevant contribu-

4
tions encompass a local asymptotic state observer [27], looking at the system as
set of differential equations on a restricted manifold [28], and an index-1 DAE
observer [29]. Other works consider the case of Lipschitz nonlinearities [30],
which have served as a basis for Lyapunov-based observer design using the lin-
earized system [31], and linear matrix inequality (LMI) approaches producing
state observers in singular [32] and non-singular [33] forms. Another Lipschitz
system was considered in [34], where the temporal separation between slow and
fast dynamics was exploited to design a robust state observer. Nonlinear DAE
systems have also been estimated through moving horizon approaches [35] and
Kalman filters [36, 37, 38].
In light of the aforementioned literature and research gaps, it is imperative
to tackle the challenges in SOC estimation for individual cells in a battery pack
under reduced sensing. In fact, we will explore the setting where only the overall
pack voltage and current can be measured in real time denoted here as min-
imal sensing, which will be showcased in Figure 2. The proposed estimation
approach contrasts with the majority of existing studies that lump the entire
battery pack as one virtual cell, and others that propose the measurements of
few specific cells inside a pack. The proposed state estimator is realized by
adopting an interconnection of single cell models with high fidelity cell-by-cell
resolution. To the best of the authors’ knowledge, this is the first attempt in the
literature to estimate the cell-level state in a battery pack with parallel-series
configurations using only pack-level sensing, which is particularly challenging
due to multiple technical reasons. First, the pack models with cell-level reso-
lution are generally governed by complex differential-algebraic equations that
require more sophisticated analysis and observer design theories than those for
normal ordinary differential equation (ODE) systems. Second, under extremely
limited measurable signals, namely only pack-level voltage and current, system
observability is significantly deteriorated. This potentially renders locally un-
observable conditions. Ultimately, this paper departs from previously existing
works by

5
(1) proposing a novel modeling framework for Li-ion battery packs with parallel-
series connected heterogeneous cells as a nonlinear DAE (descriptor) system;
(2) rigorously analyzing the nonlinear local (smooth) observability conditions
of such a nonlinear DAE system;
(3) designing a provably convergent Lyapunov-based state observer for the bat-
tery pack nonlinear DAE model to estimate cell-level SOCs and currents,
utilizing only pack-level voltage and current measurements (minimal sens-
ing).

In particular, the system analysis and observer design proposed in this work
is directly performed on the original nonlinear DAE model of the battery pack
without any model reductions, thus retaining the physical significance of the
equations and of the phenomena that they represent. This notably further
separates our work from the majority of the existing efforts.
The reminder of this paper is organized as follows. Section 2 motivates
the importance of monitoring cell-level SOC with the presence of heterogeneity
among cells in a pack. Section 3 introduces the modeling framework for a battery
pack with minimal sensing. Section 4 provides the local observability analysis for
the nonlinear battery pack DAE system. Section 5 proposes the state observer
design and its asymptotic convergence analysis. Finally, the effectiveness of
the proposed approach is illustrated in Section 6 via numerical simulations.
Conclusions are drawn in Section 7.
Notation. Throughout the work, the symbol Ip×q denotes the identity ma-
trix with dimension p × q, and 0p indicates a zero column vector with dimension
p. First-order time derivative of a variable x is represented by ẋ, and high
derivatives are denoted as x(j) where j is the number of differentiation. R and
C indicate the set of real numbers and complex numbers, respectively. Further-
more, a matrix [A1 A2 ] is 1-full if rank([A1 A2 ]) = n + rank(A2 ), where A1
offers the first n columns.

6
 Ns
A Np Np Np

Figure 2: Battery pack configuration under consideration. It consists of Ns parallel modules

in series, whereas each parallel module has Np cells connected in parallel. The i-th battery
cell in the j-th parallel module is denoted by cell (i, j). In this work, each cell (i, j) is modeled
by an equivalent circuit model sketched in Figure 5.

2. Problem Formulation & Motivation

The battery pack configuration considered in this study is shown in Figure

2. It is composed of Ns battery modules connected in series, and each module
contains Np battery cells connected in parallel. Such configuration represents a
wide range of EV battery packs, e.g., Tesla Model S and Model 3. We assume
that all cells are heterogeneous. Namely, they may be characterized by differ-
ent model parameters (capacity, internal resistance etc.), different SOC levels,
and different temperature distribution [39]. The above-mentioned heterogene-
ity among cells can be caused by manufacturing, temperature variability, and
battery degradation. Moreover, in order to reduce hardware sensing efforts, we
further assume that only the pack-level voltage and the current of the string
are available for measurement (see Figure 2). This is referred to as the minimal
sensing scenario, in which the voltage and current of each single cell in the pack
are not explicitly under surveillance. It is noteworthy that reduced sensing sig-
nificantly lowers the cost of battery design and assembly by possibly minimizing
the required number of distributed sensors, but at the cost of diminishing the
ability to monitor cell-level real-time conditions. This calls for a significant

7
 (a)

(b)

(c)

Figure 3: Simulation results of two cells in parallel using coupled electrical-thermal dynamics
with temperature and SOC dependent electrical parameters. In (b)-(c), cells are initialized at
different SOCs. The total current distributes unevenly due to both parameter and initializa-
tion heterogeneities.

incentive to accurately estimate all cell-level SOCs under reduced sensing.

According to Kirchhoff’s laws, the measured pack-level total voltage is equal
to the summation of all module voltages, and the currents entering every paral-
lel module are the same. However, since cells are heterogeneous, it is expected
that the distributions of the total current to local branches in a module are
imbalanced. We demonstrate this via an open-loop numerical simulation for
two cells connected in parallel in Figure 3. This numerical study utilizes two
Graphite/NMC type cells with 2.8 Ah nominal capacity. In this embodiment,
the cells are parameterized distinctly but have identical SOC-OCV relationship.
A transient electric vehicle-like duty charge/discharge cycle generated from the
standardized Urban Dynamometer Driving Schedule (UDDS) is applied. More-
over, we initialize the cells at different SOCs. Specifically, the total applied
current (summation of local cell currents) is plotted in Figure 3(a). It can be

8
 (a)

(b)

(c)

Figure 4: Simulation results of two cells in series using coupled electrical-thermal dynamics
with temperature and SOC dependent electrical parameters. In (b)-(c), the initial cell SOCs
are distinct. This discrepancy will persist because all cells accept the same input current.

observed that even though the applied total current is small initially (around
zero, see Figure 3(a)), Cell 1 draws a large negative current (around −10 A)
and Cell 2 positions itself at a large positive current (around +10 A). This oc-
curs because Cell 1 SOC is initialized higher (see Figure 3(b)). In such cases,
single cells can violtae safety constraints (e.g. maximum current, see Figure
3(c)) while the pair of parallel cells do not, whereas the latter is the one that is
commonly supervised. In the long run, the values for SOC follow a similar trend
while seemingly approaching, but they never converge and cells accept different
current rates, which promote different aging patterns and further increase cell-
to-cell variations. This behavior can be ascribed to the parallel connection that
forces a natural voltage balance, but not SOC balance, between the cells.
Furthermore, we encapsulate the case of the series arrangement of two het-
erogeneous cells, shown in Figure 4. This can be considered as a milder scenario
compared to the parallel case in terms of local current behaviors, since every cell

9
accepts identical and possibly measurable current. However, SOC discrepancy
is worse and more persistent in time (see Figure 4(b)) than the parallel case,
since the SOC values for the two cells will never synchronize – a bias will always
exist unless an external active action is taken, such as cell balancing.
Under reduced/minimal sensing, the battery management systems do not
monitor the local current and local voltage of each cell in a pack, which might
translate in some cells operating outside their safe operating region whereas
this cannot be seen from the pack-level information. Therefore, it will be of
significant value to monitor local SOCs, currents, and voltages caused by cell
heterogeneity in order to ensure safe battery pack operation and mitigate cell
degradation.

3. Battery Pack Model Formulation

This section first reviews an equivalent circuit model (ECM) for a single
battery cell, which is then electrically interconnected with other cell models to
form a pack model. Although an ECM is employed in this work, the observer
design and analysis can be readily generalized to other types of battery cell
models, e.g., the reduced-order electrochemical models [40, 41].

3.1. Single Cell Model

Consider the ECM for a single battery cell [42], shown in Figure 5, repre-
sented by the following continuous-time state-space representation,

ṡi,j (t) = Ai,j si,j (t) + B i,j Ii,j (t), (1)

yi,j (t) = Vi,j = P (si,j , Ii,j ), (2)

where si,j = [zi,j Ui,j ]> ∈ R2 is the state vector for the i-th battery cell in the
j-th parallel module, depicted in Figure 5, in which zi,j is cell SOC and Ui,j
represents the relaxation voltage for the parallel R-C pair. In (1)-(2), Ii,j (t)
denotes the current passing through the i-th battery cell in the j-th parallel

10
 Ri,j

ri,j Ii,j
+ Ui,j −

g(zi,j ) − Vi,j
+
Ci,j

Figure 5: The schematic of an equivalent circuit model, which is leveraged to model each cell
in the pack given by Figure 2.

module and Vi,j ∈ R is the cell terminal voltage. The state matrix Ai,j ∈ R2×2
and input matrix B i,j ∈ R2×1 are given by
   
1
0 0  Qi,j 
Ai,j =  , B i,j =  , (3)

0 − Ri,j1Ci,j 1
Ci,j

where Qi,j represents battery capacity, and ri,j , Ri,j , Ci,j are resistances and
capacitance shown in Figure 5. The output equation (2) for the (i, j)-th cell
provides the voltage response characterized by the function

P (si,j , Ii,j ) = g(zi,j ) + Ui,j + ri,j Ii,j . (4)

In (4), function g(·) denotes the open circuit voltage (OCV) that is a nonlinear
function with respect to zi,j . The cell voltage Vi,j : [0, 1] × R × R → R is the
summation of the OCV, the relaxation voltage Ui,j , and voltage contributed by
the ohmic resistance ri,j .

Remark 1. Although the single cells in the battery pack under consideration
are modeled by ECM for its structural simplicity, the analysis and algorithm de-
signs are applicable to other modeling framework, e.g., classes of electrochemical
models.

11
3.2. Kirchhoff ’s Laws

For a module of Np cells in parallel (i.e., one of the parallel modules in Fig-
ure 2), Kirchhoff’s voltage law indicates that a parallel connection constraints
terminal voltages to the same value for all cells, and Kirchhoff’s current law
indicates that the total current is equivalent to the summation of all branch
currents. Mathematically, the following nonlinear algebraic constraints, accord-
ing to Kirchhoff’s voltage law, need to be enforced,

P (sk,j , Ik,j ) = P (s`,j , I`,j ), ∀k, ` ∈ {1, · · · , Np }, k 6= `, j ∈ {1, · · · , Ns }, (5)

which can be equivalently expanded as

g(zk,j ) + Uk,j + rk,j Ik,j = g(z`,j ) + U`,j + r`,j I`,j ,

∀k, ` ∈ {1, · · · , Np }, k 6= `, j ∈ {1, · · · , Ns }. (6)

Similarly, Kirchhoff’s current law poses the following linear algebraic constraint
with respect to cell local currents,

Np
X
Ii,j (t) = I(t), ∀j ∈ {1, · · · , Ns }, (7)
i=1

where I(t) is the measurable pack-level total current. It is worth highlighting

that, for every parallel module j, equation (6) imposes (Np − 1) nonlinear alge-
braic constraints with respect to local SOCs and local currents, whereas equation
(7) imposes 1 additional algebraic constraint with respect to local currents.

3.3. Parallel Module Model

When only the total current is measured, the local currents of cells are un-
known since they are unevenly distributed across cells due to cell heterogeneity.
Hence, a system of differential-algebraic equations must be solved such that the
algebraic equations (6) and (7) are fulfilled at all times t > 0. Such methodology
is realized by augmenting the local currents (algebraic states) to the differential

12
states (local SOCs) to form an aggregated state vector for a nonlinear DAE
system [23], which takes the form

Ej ẋj (t) = Aj xj (t) + φ(xj (t), I(t)), (8)

where xj = [s>
j u> >
j ] ∈R
3Np
, ∀j ∈ {1, · · · , Ns }, with

h i>
sj = s>
1,j s>
2,j ··· s>
Np ,j
∈ R2Np , (9)
h i>
uj = I1,j I2,j ··· INp ,j ∈ RNp . (10)

Equation (8) encodes both the dynamical equations (1) and algebraic constraints
(6)-(7) for the j-th parallel module, and the matrix Ej is a singular matrix of
the form  
I2Np ×2Np 02Np ×Np
Ej =   ∈ R3Np ×3Np . (11)
0Np ×2Np 0Np ×Np

Matrix Aj accounts for the linear part of the system equations with
 
A11,j A12,j
Aj =   ∈ R3Np ×3Np , (12)
A21,j A22,j

where

A11,j = diag(A1,j , A2,j , · · · , ANp ,j ),

A12,j = diag(B 1,j , B 2,j , · · · , B Np ,j ),

 
0 1 S 0 · · · 0
 
0 1 0 S · · · 0
 
. . . . . ..  Np ×2Np
h i
A21,j =  . . . . .. ∈ R , where S = 0 −1 ,
. . . . .
 
0 1 0 0 · · · S
 
 
0 0 0 0 ··· 0

13
  
r1,j −r2,j 0 ··· 0 
 
r1,j
 0 −r3,j ··· 0  
 . .. .. .. ..  Np ×Np
.
A22,j =
 . . . . .  ∈R . (13)
 
r
 1,j 0 0 ··· −rNp ,j 

 
1 1 1 ··· 1

Notice that matrix A22,j is full rank [43], i.e. the linear part of the DAE model
(8) is regular and impulsive free [23]. Function φ(xj , I) in (8) constitutes the
nonlinear portion in the system equations from the voltage algebraic constraints
(6),  
 02Np ×1

 
 g(z1,j ) − g(z2,j ) 
    

φs  g(z 1,j ) − g(z )
3,j 

φ(xj , I) =   =   ∈ R3Np , (14)

.
..
φu 



 
g(z ) − g(z
Np ,j )

 1,j
 
−I(t)

where φs represents the nonlinearities in the dynamical equations and corre-

sponds to row 1 through row 2Np of φ. Evidently we have φs = 02Np ×1 due
to the linear nature of system (1). φu ∈ RNp encodes the nonlinearities and in-
put appearing in the algebraic constraints, and it corresponds to row (2Np + 1)
through row 3Np of φ.

3.4. Pack Model

For a string of parallel modules wired in series (Figure 2), there are in to-
tal N = Np × Ns cells. The measured pack-level voltage is equivalent to the
summation of all parallel module voltages, and every parallel module shares the
same current, according to Kirchhoff’s laws. Let us denote n = 2N and

h i>
x = x>
1 x>
2 ··· x>
Ns
∈ Rn , (15)

14
where respective xj (the state vector for parallel module j) was previously intro-
duced in (8)-(10). x is essentially the state vector that aggregates all differential
(local SOCs) and algebraic (local currents) states of all cells in the pack. Based
on the parallel module model in (8), a string of Ns parallel modules can be
expressed in the compact form

E ẋ(t) = Ax(t) + Φ(x(t), I(t)), (16)

y(t) = H(x(t)) = Cx(t) + h(x(t)), (17)

in which y(t) = V (t) is the measured pack voltage, and

E = diag(E1 , E2 , · · · , ENs ),

A = diag(A1 , A2 , · · · , ANs ),
h i>
Φ(x, I) = φ(x1 , I)> φ(x2 , I)> · · · φ(xNs , I)> . (18)

where Ej , Aj , and φ(xj , I) have been introduced in (12) , (11), and (14), respec-
tively. In (17), the pack voltage expression has been partitioned into a linear
part Cx(t) and a nonlinear part h(x(t)), which will be detailed later. Note
that in the subsequent sections, we will slightly abuse the notations for the ease
of analysis and presentation. Specifically, in (16), the state vector x is com-
posed of xj , j ∈ {1, 2, · · · , Ns }, concatenated according to to module numbers.
Nonetheless, from now on, we alter the sequence in x such that the differential
state vector s = [s>
1 s>
2 ··· s> >
Ns ] comes before the algebraic state vector

u = [u>
1 u>
2 ··· u> >
Ns ] , i.e.,

h i>
x = s> u> . (19)

The system matrices E, A, and Φ are adjusted accordingly. In that event, we

also let the partition of the function Φ to be Φ(x, I) = [Φs (x) Φu (x, I)]> , where
vector Φs accounts for the nonlinearities in the differential equations whereas
Φu lumps the nonlinearities in the algebraic constraints. Note, Φs ≡ 0N since

15
battery dynamics described by the ECM (5) is linear. However, we maintain
the appearances of Φs in the forthcoming analysis to make the framework gen-
eralizable to other nonlinear battery dynamics.
Special care needs to be taken for the output function H(x(t)) in (17).
H(x(t)) represents the summation of all parallel module voltages, while each
parallel module voltage can be mathematically expressed in Np different ways.
Namely, each parallel module voltage is equivalent to any cell voltage in that
module. Thus, in order to increase the information contained in the pack sys-
tem output to maximize system observability, H(x(t)) must include all (Np )Ns
possible combinations, although the numerical values of them would all equal
to the measured pack voltage signal:

 >  >
V (t)  V1,1 + V1,2 + · · · + V1,Ns −1 + V1,Ns 
   
V (t)  V1,1 + V1,2 + · · · + V1,Ns −1 + V2,Ns 
H(x(t)) =  .  =   , (20)
   
 ..  ..
 

 . 

   
V (t) VNp ,1 + VNp ,2 + · · · + VNp ,Ns −1 + VNp ,Ns

where each Vi,j is given by (2).

Remark 2. In fact, not all (Np )Ns combinations are unique. That is, only a
subset of these combinations are linearly independent of others. As an example,
consider a simple case with Np = 2 and Ns = 2. In this case, 4 possible voltage
representations can be generated, namely
   
y1 (V1,1 + V1,2 )
   
   
y2  (V1,1 + V2,2 )
H(x(t)) =   = 
   .
 (21)
y3  (V2,1 + V1,2 )
   
y4 (V2,1 + V2,2 )

Nevertheless, it is easily observed that y4 = y2 + y3 − y1 , making y4 a redundant

output entry. Ultimately, the formulation of H(x(t)) in (20) should be carefully
calibrated based upon the structures of the pack to remove redundant entries,

16
which effectively reduces the mathematical complexity in the observability anal-
ysis in the upcoming Section 4.

Remark 3. In Kirchhoff’s laws (6)-(7), although these algebraic constraints

are nonlinear in the differential states zi,j (individual cell SOCs), they are in-
deed linear in the algebraic states Ii,j (individual cell currents). Consequently,
in a parallel module, one can solve for the algebraic states as a closed form of
differential states and substitute this expression into the system dynamics to
obtain a reduced-order ODE system, which is ultimately independent of the al-
gebraic states. This strategy was carried out in [43, 44, 45]. In spite of the fact
that this model reduction from DAEs to ODEs is useful for simulation studies
to understand imbalanced current distributions in parallel and parallel-series
configurations, the reduced-order ODEs are mathematically not the exact rep-
resentation of the original DAEs [46], primarily due to the following reasons:
(i) numerical methods that are commonly used for solving systems of ODEs do
not trivially apply to DAEs; (ii) the solution to the reduced-order ODEs does
not always satisfy the algebraic constraints; and (iii) when a state estimator
is designed for the reduced-order ODEs, the algebraic state will only be esti-
mated in an open-loop fashion, whereas the differential states will most likely be
estimated in a feedback (closed-loop) manner. Ultimately, due to these restric-
tiveness and limitations, we never perform model reduction in order to retain
the physical significance of the differential-algebraic nature and of the phenom-
ena that they represent. All analysis and observer designs are based upon the
high-fidelity DAE system (16)-(17). This, together with the minimal sensing
setup, substantially separates our method from the existing battery pack state
estimation works in the literature, which tend to lump the entire pack as a single
virtual cell whose dynamics is described by a reduced-order ODE that signifi-
cantly loses the tractability of individual cells or conduct the measurement of
few specific cells inside a pack.

Remark 4. In practice, the parallel modules in Figure 2 are wired with the
same number of cells, but the proposed pack modeling framework and the sub-

17
sequent state observability analysis in Section 4 can be readily generalized to
the case when the parallel modules have different number of cells under extreme
circumstances.

The battery pack model introduced above will be leveraged in the analysis
and designs in the subsequent sections.

4. Observability Analysis

Informally, observability analysis refers to the study of the conditions under

which it is possible to uniquely determine the states of a dynamical system
from measurements of its input and output [47]. In other words, a system,
e.g., (16)-(17), is said to be completely observable on an interval (t0 , t1 ) if the
initial state can be uniquely determined from knowledge of the output y(t) and
input I(t) over (t0 , t1 ) [48]. Specifically for our case, if the battery pack system
is unobservable, this means that it is impossible to infer the SOC of at least
some cells from the pack-level voltage and current data. In this section, we
mathematically analyze the conditions for the observability of individual cell
SOC in a battery pack under the minimal sensing scenario (Figure 2), in which
only pack-level voltage and current are measured. For mathematical tractability,
the observability conditions are derived from the pack configuration shown in
Figure 2 with Ns = 2 and Np = 2 (N = 4 cells in total), which is conveniently
referred to as 2P2S. We study the observability using two techniques, namely
(i) observability via linearization of the pack model (16)-(17), and (ii) nonlinear
DAE smooth observability inherited from DAE solvability.
Previously in the literature, observability analysis has been carried out for
cells connected in series [5] and in parallel [22] under reduced sensing scenarios.
The former assumes total voltage measurement for a string of heterogeneous
cells, and the latter faces unknown imbalanced current distributions in a par-
allel module. Both studies require non-flatness of high-order gradient of cell’s
OCV function. The observability matrices for the series case [5] are derived from
a conventional ODE setting, differing from that of a parallel (and series-parallel

18
considered in this work) arrangement. Namely, in the series arrangement, each
cell’s parameters/states appear in a column of observability matrix. See, for in-
stance, equation (10) in [5]. This is not the case for a battery pack with parallel
topology, where parameters/states of the cells are scattered all over the differ-
ent entries in the observability conditions (See Section IV in [22]). Therefore,
parameters/states of one cell influence the observability of the neighboring cells
in a pack containing parallel modules. Importantly, in light of Remark 3, the
observability analysis in this section will be conducted directly on the pack DAE
system, rather than a reduced-order ODE system, to examine if both differential
and algebraic states are observable.

Remark 5. The dynamics for the cell relaxation voltage Ui,j in (1) exponen-
tially decays to an equilibrium position. This state is generally observable (at
least detectable) from current-voltage measurements. In the subsequent observ-
ability analysis and observer designs, the effect of relaxation voltage is ignored
to simplify computations.

4.1. Observability via Model Linearization

To study the observability of any general form of a nonlinear system, one of
the most convenient strategies is to linearize the system around an equilibrium
point and derive the local observability conditions for the linearized system. If
the linearized system is observable at a given equilibrium point, then the non-
linear system is locally observable at that position. This method is practically
easy to implement but the results are only sufficient. That is, the observability
conditions arising from linearizing the nonlinear system can be conservative,
and nothing can be concluded for the original nonlinear system if the linearized
system is not observable. Thus, under the unobservable condition from the lin-
earized system, less conservative observability notions need to be explored. In
this section, we first study if local observability exists from the linearization of
the nonlinear pack DAE system (16)-(17).
The DAE system (16)-(17) with Ns = 2 and Np = 2 is employed to de-
scribe the dynamics for 2P2S with x = [x1 x2 x3 x4 x5 x6 x7 x8 ]> =

19
[z1,1 z2,1 z1,2 z2,2 I1,1 I2,1 I1,2 I2,2 ]> ∈ R8 , and
   
1
0 0 0 0 Q1,1 0 0 0 

0

 
0 1  
 0 0 0 0 Q2,1 0 0    0 
   
0 1  
 0 0 0 0 0 Q1,2 0    0 
   
1   
0 0 0 0 0 0 0 Q2,2 
 0 
A =  , Φ(x, I) =  ,

−I
0  
 0 0 0 1 1 0 0    
   
−I
0  
 0 0 0 0 0 1 1    
   
−r2,1 g(x1 ) − g(x2 )
0  
 0 0 0 r1,1 0 0  
   
0 0 0 0 0 0 r1,2 −r2,2 g(x3 ) − g(x4 )
 
1 0 0 0 0 0 0 0
 
 
0 1 0 0 0 0 0 0
 
 
0 0 1 0 0 0 0 0  



 g(x1 ) + r1,1 x5 + g(x3 ) + r1,2 x7
0 0 0 1 0 0 0 0  
E =  , H(x) =  .

  g(x 1 ) + r1,1 x 5 + g(x 4 ) + r2,2 x 8
0 0 0 0 0 0 0 0  



 g(x2 ) + r2,1 x6 + g(x3 ) + r1,2 x7
0 0 0 0 0 0 0 0
 
 
0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0
(22)

To study the observability of the pack model (16)-(17) with (22), we linearize
the system around an equilibrium point x = x and check the rank of the ob-
servability matrix of the linearized system. The linearized model takes the form

E ẋ(t) = T x(t) + BI(t), (23)

y(t) = Cx(t), (24)

where the state matrix T ∈ R8×8 and output matrix C ∈ R3×8 are given by

dΦ dH
T = A+ (x) , C= (x) , (25)
dx x=x dx x=x

20
with matrix A and H given in (22). Matrices T and C take the form
 
1
 0 0 0 0 Q1,1 0 0 0 
 
 0 1
 0 0 0 0 Q2,1 0 0  
 
 0 1
 0 0 0 0 0 Q1,2 0 

 
 0 1 
0 0 0 0 0 0 Q2,2 
T = ,

 0 0 0 0 1 1 0 0 
 
 
 0 0 0 0 0 0 1 1 
 
 
g 0 (x ) −g 0 (x ) 0 0 r1,1 −r2,1 0 0 
 1 2 
 
0 0
0 0 g (x3 ) −g (x4 ) 0 0 r1,2 −r2,2
 
g 0 (x1 ) 0 g 0 (x3 ) 0 r1,1 0 r1,2 0
 
C = g 0 (x1 ) g 0 (x4 ) r1,1 r2,2  . (26)
 
0 0 0 0
 
0 0
0 g (x2 ) g (x3 ) 0 0 r2,1 r1,2 0

Let us now introduce the definition of Complete Observability (C-Observability)

for the linearized DAE system (23)-(24), which represents a linearized battery
pack model for the topology in Figure 2.

Theorem 1 (Complete Observability [23]). The regular linear descriptor

system (23)-(24) is Complete Observable if and only if the following two condi-
tions hold:

(C1) rank([E > , C > ]> ) = 2N ;

(C2) rank([(λE − T )> , C > ]> ) = 2N, ∀λ ∈ C.

It follows from [23] that there exists a standard decomposition, such that
the descriptor linear system (23)-(24) is transformed into a slow subsystem (an
ODE system) and a fast subsystem (a DAE subsystem). The system (23)-(24)
is C-Observable if and only if both its slow and fast subsystems are observable.
See Section 4 of [23] for more precise definitions. Specifically, Condition (C1)
concerns observability of the fast (algebraic) subsystem while (C2) involves the
slow (differential) subsystem observability. Focusing first on condition (C1), it
can be verified that the expression in (C1) is rank deficient. Intuitively, the

21
singular matrix E offers rank N = 4 with its non-zero columns. Nonetheless,
matrix C does not have enough rows (or measurements) to fulfill the missing
rank. This means that the pack voltage measurement is potentially insufficient
to uniquely determine the branch current of every cell. Looking at condition
(C2), which needs to be verified for all λ ∈ C. Fortunately, for those λ’s that
are not one of the generalized eigenvalues of the pair (E, T ), condition (C2)
automatically validates. Thus, the verification of condition (C2) requires the
numerical computation of the generalized eigenvalues of the pair (E, T ), by
solving the characteristic equation det(λE − T ) = 0, and λ = 0 is guaranteed
to be one of the solutions to this equation. When λ = 0, condition (C2) boils
down to rank([−T > , C > ]> ) = 2N , which is not true. Consequently, neither
conditions in Theorem 1 can be verified. Hence, we cannot draw any conclu-
sions towards the individual cell SOC and current observability under minimal
sensing by checking the observability conditions of the linearized DAE system.
This incentivizes a more sophisticated (less conservative) strategy to derive the
observability conditions of the original nonlinear DAE system (16)-(17).

4.2. DAE Solvability

Before introducing a less conservative notion of observability, let us first

present the concept of solvability of a DAE system. Consider a general nonlinear
implicit DAE of the form

F (t, x(t), ẋ(t)) = u(t), (27)

y(t) = H(x, t), (28)

with x ∈ Rn , F (·, ·, ·) ∈ Rn , and Fẋ = ∂F/∂ ẋ identically singular [49]. Note

that the pack model (16)-(17) can be easily reformulated into this structure.
We also assume that H (or equivalently OCV) is sufficiently smooth so that the
high-order derivatives with respect to states are continuous. There have been
considerable amount of research on computing a solution for DAE (27)-(28),
which is a system of equations in the (2n + 1)-dimensional variable (t, x, ẋ). Let

22
us now first provide the formal definition of solvability of DAE (27) [50, 51].

Definition 1 (Solvability [50]). DAE (27) is solvable in an open set Ω ⊆

R2n+1 if the graphs (t, x, ẋ) of the solutions form a smooth manifold in Ω called
the solution manifold and solutions are uniquely determined by their value x0
at any t0 such that (t0 , x0 , ẋ0 ) ∈ Ω.

In general, the solution x of DAE (27) is dependent on derivatives of F . If

(27) is differentiated γ times with respect to t, we get (γ + 1)n equations [50]:
 
 F (t, x, ẋ) 
Ft (t, x, ẋ) + Fx (t, x, ẋ)ẋ + Fẋ (t, x, ẋ)x(2) 
 
Fγ (t, x, ẋ, w) =   = u, (29)
 
..

 . 

γ
 d 
γ
F (t, x, ẋ)
dt

where w denotes the high-order derivatives of x, i.e., w = [x(2) x(3) · · · x(γ+1) ],

and u is a column vector with u = [u u̇ · · · u(γ) ]> .
While the definition for DAE solvability in Definition 1 is obscure, the as-
sumptions below are verifiable sufficient conditions for DAE solvability:

(S1) Fγ is sufficiently differentiable in its arguments.

(S2) G ≡ Fγ = 0 is consistent as an algebraic equation.
(S3) [Gẋ Gw ] is 1-full and has constant rank.
(S4) [Gx Gẋ Gw ] has full row rank independent of (t, x, ẋ, w).

The minimum γ for which conditions (S1)-(S4) hold is called the uniform differ-
entiation index [52]. In addition, [Gẋ Gw ] is 1-full with respect to ẋ if the first
N columns are linearly independent, and linearly independent of the remaining
columns.
The definitions and assumptions in this subsection will be utilized to estab-
lish the smooth observability in the forthcoming sections.

23
4.3. Smooth Observability

Observability analysis using the linearized system failed as has been demon-
strated in Section 4.1. This motivates us to study a stronger notion of observabil-
ity. To elucidate if less conservative observability conditions exist, we analyze
the local observability of the nonlinear battery pack DAE system (16)-(17) by
introducing the concept of smooth observability.

Definition 2 (Smooth Observability [53]). The nonlinear DAE system (27)-

(28) is smoothly observable (of order (δ, γ)) on the interval K if there exists
smooth functions ∆l (t) and Γl (t) on K such that

δ
X γ
X
x(t) = ∆l (t)y (l) (t) + Γl (t)u(l) (t). (30)
l=0 l=0

Note that smooth is defined to mean infinitely differentiable, and the infinite
differentiability is only used to make sure the observability conditions are neces-
sary and sufficient [53]. In order to verify smooth observability, (30) is used to
determine the least δ and γ such that the solution of the DAE model x(t) can be
represented by a weighted sum of the time derivatives of input u(t) and output
y(t). The essence of smooth observability is that, if C in (24) is not full column
rank, then additional information to determine x is obtained by differentiating
the measurements.

Remark 6. It should be emphasized that there exist various forms of observ-

ability other than the previously noted complete and smooth observability in the
literature [48]. For instance, total observability refers to complete observability
on every sub-interval of (t0 , t1 ) and uniform observability stands for smooth ob-
servability of order (n − 1, n − 1). From Definition 2, it is evident that smooth
observability, which is selected in this work for analyzing battery pack system
observability, is a stronger type of observability than total and complete observ-
ability. For deeper discussions on smooth observability, interested readers may
refer to Section I of [53].

24
 Differentiating the output expression (28) δ times with respect to t yields
 
 H(x) 
 
 Hx (x)ẋ 
 
H ≡ Hδ (t, x, ẋ, w) =  ..  = y, (31)
.
 
 
 δ 
d 
δ
H(x)
dt

where y = [y ẏ ··· y (δ) ]> . Then we can express the combination of equa-
tion (29) with equation (31) as
 
u
O(t, x, ẋ, w) =   . (32)
y

Hence, the Jacobian matrix of O(t, x, ẋ, w) with respect to (x, ẋ, w) is then given
by  
Gx Gẋ Gw
JO =  . (33)
Hx Hẋ Hw

Now we are positioned to formally introduce the sufficient condition for smooth
observability for the nonlinear DAE system (27)-(28).

Theorem 2 (Smooth Observability Verifiable Conditions [54]). Suppose

system (27) satisfies condtions (S1)-(S4) in a neighborhood U of a consistent
point (t0 , x0 , ẋ0 , w0 ). Additionally, suppose the Jacobian matrix JO given in
(33) validates
 
 Gẋ Gw 
, for (t, x, ẋ, w) ∈ U
(O1) rank(JO ) = n + rank 
Hẋ Hw
(O2) JO has constant rank on U

Then system (27)-(28) is smoothly observable on U .

Remark 7. The smooth observability conditions in Theorem 2 is local in a

neighborhood of U , and the Jacobian matrix JO depends on state x, OCV
function g, and individual cell model parameters.

25
 Utilizing Theorem 2, we analyze the local smooth observability of the non-
linear battery pack DAE system (27)-(28) for 2P2S, in which
   
ẋ1 − x5 /Q1,1 0
   
ẋ2 − x6 /Q2,1
   
  0
   
ẋ3 − x7 /Q1,2
   
  0
   
ẋ4 − x8 /Q2,2
   
  0
F (t, x(t), ẋ(t)) = 
 ,
  ,
u(t) =   (34)
 x5 + x6  I 
   
   
 x7 + x8  I 
   
g(x1 ) − g(x2 ) + r1,1 x5 − r2,1 x6 
   
0
   
g(x3 ) − g(x4 ) + r1,2 x7 − r2,2 x8 0

and H(x) once again takes the same form as that in (22). Jacobians in (S3),
(S4), and (O1) can be computed by automatic differentiation [54, 55]. It is
also confirmed numerically that the uniform differentiation index is γ = 1. The
objective is to find the least indices γ and δ that would render the battery
pack system under minimal sensing smoothly observable. Numerically, utilizing
the output function (22), the least γ and δ for the DAE system (27)-(28) to
be smoothly observable are γ = 3 and δ = 3. It is noted that up to `-th
order gradients of g(x) must be checked, where ` = max{(γ + 1), (δ + 1)}. For
condition (O1) to be satisfied, g (i) (x) with i = {1, 2, · · · , `} should not be zero
simultaneously. As can be seen from the 2P2S case, up to 4-th order gradients
of OCV are involved. It has been studied extensively in the literature, e.g.,
[5, 40], that higher order gradients of OCV approach zero, in particular in
the middle SOC range (around 15%-90%). Hence, when only the pack total
voltage is measured, the observability of individual cell SOC is expected to
be weak. Nonetheless, at high and low SOC ranges, the OCV function g is
generally highly nonlinear with respect to SOC, rendering significant high-order
gradients g (i) (x), i = {1, 2, · · · , `}. As has been previously highlighted in [5],
the high and low SOC ends are the regions where the risks of over-charge and
over-discharge are critical, and high gradients in these regions should facilitate

26
notable individual cell SOC observability from only pack voltage measurements.
In addition, since the cells in the pack are heterogeneous, it is anticipated that
cells would not have identical SOC values at any given time instant t, which
ultimately enhances condition (O1) in Theorem 2. However, it is also worth
noting that in the extreme cases where all cells are identically parameterized,
observability for individual cell SOCs will be completely lost as the cells are not
distinguishable from one another. This trivial case just boils down to a single
cell estimation problem.
The classical approach to study observability of a nonlinear battery syste, is
to linearize the model, as done in Section 4.1 and e.g., [22, 56]. By doing so, the
cell observability condition is only determined by the first-order derivative of
OCV function. However, we can conclude from the analysis in Section 4.1 that
this is conservative and less informative. Same conclusion was also drawn in
[57] through the local observability analysis of a single cell. The more detailed
analysis based on smooth observability of the nonlinear battery pack DAE sys-
tem showed that OCV gradients must be different than zero to guarantee local
observability, but it does not need to be the first-order gradient. It relies on
the fact that the observability condition (O1) explicitly depends on high-order
gradients of OCV function, and it can be analytically obtained through e.g.,
symbolic software.
It is finally emphasized that in order for cell SOCs to be locally observable
from only pack-level voltage and current measurements (i.e., minimal sensing),
high-order gradients of OCV function must be checked. Although fewer mea-
surement signals weaken cell-level observability, which is expected, our analysis
provides considerable incentive to significantly reduce the number of sensors in
a battery pack while maintaining enough cell-level observability.

5. Design of State Observers

For the purpose of observer design, we investigated different notions of ob-

servability in Section 4 and discussed the conditions under which the battery

27
pack system with minimal sensing is smoothly observable. These conditions
essentially ensure that there exists a state observer for estimating the SOC and
current of individual cells modeled by the battery pack system (16)-(17). In
this section, we propose a Luenberger type state observer to asymptotically es-
timate cell-level SOCs and currents when only pack-level voltage and current
are available for measurement.
The following observer with linear output error injection is proposed for the
battery pack plant model (16)-(17):

˙
 
E x̂(t) = Ax̂(t) + Φ(x̂(t), I(t)) + K y(t) − C x̂(t) − h(x̂(t)) , (35)

ŷ(t) = C x̂(t) + h(x̂(t)), (36)

where x̂(t) indicates the estimation of x(t), and K ∈ Rn is the observer gain
vector to be designed such that the differential states (individual cell SOCs) and
the algebraic states (cell local currents) converge to the truth states asymptot-
ically, i.e., x̂(t) → x(t) as t → ∞. Moreover, K = [Ks> Ku> ]> with Ks ∈ RN
and Ku ∈ RN (recall that n = 2N and N = Np × Ns ). The above observer
structure adopts the linear output error injection method [30, 31], although the
battery pack plant model is nonlinear. Theorem 3 establishes the convergence
properties of the proposed observer.

Theorem 3. Consider the battery pack plant model dynamics (16)-(17), and
suppose the matrix [A22 C]> has rank N . Let
 
G11 G12
G = (A − KH) =  , (37)
G21 G22

and define the matrix

e = (G11 − G12 G−1 G21 ).
G (38)
22

Suppose the function

 
L(x, I) = Φs (x) − G12 G−1 −1
22 Φu (x, I) + G12 G22 Ku − Ks h(x), (39)

28
is Lipschitz continuous with respect to x, in which Φ = [Φs Φu ]> was defined
in the vicinity of (19). That is,

kL(x1 , I) − L(x2 , I)k ≤ γkx1 − x2 k (40)

for any feasible x1 , x2 ∈ X, where γ ∈ R is the Lipschitz constant. If the

observer gain K is chosen to ensure that G
e is stable, and

 
min σmin G̃ − jωIN ×N > γ, (41)
ω∈R+

where σmin (·) denotes the minimum singular value. Then the zero equilibrium
of the dynamics of estimation error e(t) = x(t) − x̂(t) given by

 
E ė = Ge + Φ(x, I) − Φ(x̂, I) − K h(x) − h(x̂) (42)

is asymptotically convergent to zero.

Proof. Let the state estimation error e = [e>

s e> >
u ] , with es = s − ŝ being

the estimation error for the differential states and eu = u − û the estimation
error for the algebraic states. Then (42) can be decomposed into
       
I 0 ės G11 G12 es Φs (x) − Φs (x̂)
 N ×N   =   +  
0 0 ėu G21 G22 eu Φu (x, I) − Φu (x̂, I)
 
Ks  
−   Cx − C x̂ + h(x) − h(x̂) . (43)
Ku

We highlight that G22 can be non-singular (i.e., invertible) if the linear part of
(16) is impulse observable [31], i.e., the matrix [A22 C]> has rank N . Then

29
the estimation error system (43) is equivalently described by

 
−1
ės = G11 − G12 G22 G21 es
h i  
+ Φs (x) − G12 G−1 Φ
22 u (x, I) + G G−1
12 22 Ku − Ks h(x)
h i  
− Φs (x̂) − G12 G−1
22 Φ u (x̂, I) − G12 G−1
22 Ku − Ks h(x̂)

e s + L(x) − L(x̂),
= Ge (44)

along with the algebraic equation

eu = −G−1 −1 −1
   
22 G21 es − G22 Φu (x) − Φu (x̂) + G22 Ku h(x) − h(x̂) . (45)

Consider the following Lyapunov function for the error dynamics (44), cor-
responding to the differential states es ,

1 >
W (t) = e P es , P = P >  0. (46)
2 s

The time derivative of the Lyapunov function W (t) along the trajectory of es
is computed by

1 > 1
Ẇ (t) = ės P es + e> P ės
2 2 s
1 > e> e s + e>
= e (G P + P G)e s P [L(x) − L(x̂)]
2 s
1 > e>
≤ e (G P + P G)ee s + kP es k · kL(x) − L(x̂)k
2 s
1 > e>
≤ e (G P + P G)ee s + γkP es k · kes k
2 s
1 > h e> i
e + γ 2 P P + IN ×N es ,
≤ es G P + P G (47)
2

where the inequality

2γkP es k · kes k ≤ γ 2 e> >

s P P es + es es (48)

has been utilized in the last inequality of (47). According to Theorem 2 in [30],

30
if G̃ is stable and minω∈R+ σmin (G̃ − jωIN ×N ) > γ, then there exists ε > 0 and
P = P >  0 such that

G̃> P + P G̃ + γ 2 P P + IN ×N + εIN ×N = 0. (49)

Therefore, in view of (47) and (49), Ẇ (t) < 0, and the estimation error es is
asymptotically stable. Under this scenario, when t → ∞, [Φu (x) − Φu (x̂)] → 0,
and [h(x) − h(x̂)] → 0. Hence, according to (45), the estimation error eu for
the algebraic states also converges to zero asymptotically. This completes the
proof.

Remark 8. Note that the first N columns in matrix A and matrix C are zero
columns. See, for example, matrix A in (22) for the 2P2S configuration. In
consequence, G11 and G21 are zero matrices, which yields a marginally stable
G̃ regardless of the choice of observer gain matrix K. This violates the observer
design conditions in Theorem 3. However, we observe that the matrix A can
essentially be freely assigned by adding stabilizing terms in the first N columns,
and the added terms are cancelled through the nonlinear term Φ, which could
potentially alter the Lipschitz constant γ.

Remark 9. The Lipschitz constant γ could be obtained by computing the in-

finity norm of function L(x) with respect to the state x, i.e., γ = k∂L/∂xk∞ .
In addition, condition (41) can be practically verified if

Re(−λ) > κ(R)γ, (50)

where λ is the eigenvalues of G̃ and κ(R) is the condition number of matrix R,

in which G̃ = RΛR−1 . Interested readers may refer to Theorem 5 in [30] for
more details and proof.

In view of Remark 3, the observer (35)-(36) is proposed directly on the non-

linear battery pack DAE model without any model reductions. It has been es-
tablished in Theorem 3 that both differential and algebraic states are estimated

31
Figure 6: Open circuit voltage for a Graphite/NMC cells. The fitted OCV functional form is
g(zi,j ) = p1 e(α1 zi,j ) + p2 e(α2 zi,j ) + p3 zi,j
2 adopted from [58].

in a feedback fashion with asymptotic convergence. Ultimately, the individual

cell SOCs and currents can be effectively estimated with guaranteed mathemat-
ical convergence using only pack-level voltage and current measurements.

6. Simulations

This section presents a simulation study that demonstrates the performance

of the proposed state observer for individual cell SOC estimation in a battery
pack under minimal sensing scenario. For the ease of presentation and without
loss of generality, the numerical implementation is conducted for the 2P2S case.
All cells in the pack are of Graphite/NMC type, whose open circuit voltage is
shown in Figure 6. We consider the situation in which the cells may differ in
their initial SOCs and model parameters, but subject to the same SOC-OCV re-
lationship. The assumption that SOC-OCV relationship is the same is based on
the fact that this is a thermodynamic property and only a function of the elec-
trode materials. Thus, processing variation at the material level is likely to not
impact the SOC-OCV variation, compared to other quantities such as capacity,
resistance, etc. The numerical values for the model parameters are enumerated
in Table 1. The considered setup guarantees local smooth observability based

32
 Table 1: Model Parameters in Simulations
Module 1 Module 2
Cell 1 Cell2 Cell 1 Cell 2 Units
r 0.1 0.22 0.3 0.13 [Ω]
Q 1500 1800 1200 2000 [A·sec]
z0 0.2 0.25 0.15 0.22 [–]
ẑ0 0.3 0.375 0.225 0.33 [–]

Figure 7: Plant model simulation for 2P2S with model parameters listed in Table 1. (a)
applied total voltage; (b) pack level voltage measurement.

on the analysis in Section 4.3. For all simulations, the state estimates are in-
tentionally initialized at incorrect values. The true SOC initial conditions are
20%, 25%, 15%, and 22%. The observer initial conditions are perturbed by 50%
of the true values, and they are given by 30%, 37.5%, 22.5%, and 33% (Table
1). In the presented simulations, we utilize the battery pack plant model simu-
lated data to validate the proposed observer. Ultimately, the DAE plant model
and the corresponding observer system are solved using the publicly available
numerical solvers in MATLAB® . A crucial step in the numerical integration is
to compute consistent initial conditions.

33
Figure 8: Individual cell SOC estimation performance. SOC trajectories will not synchronize
when cells are heterogeneous. All SOC estimates are initialized with 50% errors, then quickly
converge to the truth SOCs asymptotically using pack-level voltage measurement only. (a)
parallel module 1 cell SOC estimates; (b) parallel module 2 cell SOC estimates.

In this simulation study, the total applied current at the pack level is appro-
priately scaled from an Urban Dynamometer Driving Schedule (UDDS) drive
cycle to emulate a practical electric vehicle driving pattern, and it is sketched
in Figure 7(a). The simulated (and measured) pack voltage responses are re-
ported in Figure 7(b). The objective is to reconstruct individual cell SOC signals
asymptotically from the pack-level current and voltage measurements only, via
observer gain selection according to Theorem 3. The observer in (35)-(36) is
used to estimate the individual cell SOCs and the cell local currents.
Figure 8 and Figure 9 demonstrate the estimation performance. These re-
sults are numerically generated using the observer gain

 >
0.65 1.09 0.80 0.83 0.36 0.51 0.04 1.85
 
K = 1.31 1.87 0.33 1.84 1.59 1.15 0.88 0.52 ,
 
 
1.50 0.46 0.13 1.53 1.34 1.43 1.28 0.84

34
Figure 9: Individual cell current estimation performance. Cell current estimates are initialized
far from the truth, for the purpose of consistent initial conditions. These algebraic state
estimates converge instantly. (a) module 1 cell 1 current estimates; (b) module 1 cell 2
current estimates; (c) module 2 cell 1 current estimates; (d) module 2 cell 2 current estimates.

which satisfies the observer conditions in Theorem 3. Figure 8 illustrates the

convergence behavior of individual cell SOCs. The solid curves are the truth
SOCs that are simulated from the plant model, and the dashed curves represent
the SOC estimates. Notice again that when cells are heterogeneous, their SOC
trajectories will not synchronize, even within a parallel module that self balances
voltages. Despite significantly incorrect initial conditions (50% initial errors for
all cases), the SOC estimates quickly converge to the truth values. After a rapid
initial transient, the SOC estimates have root mean squared (RMS) errors of

35
0.13%, 0.37%, 0.25%, and 0.092% for module 1 cell 1, module 1 cell 2, module
2 cell 1, and module 2 cell 2, respectively. Furthermore, Figure 9 portrays the
estimates for the algebraic states (cell local currents). Note that the current
estimates are initialized considerably far from the real initial spots of the truth
currents (see the dashed magenta curves in Figure 9). This is because the initial
conditions of algebraic state estimates are calculated based on differential states’
initial conditions to form consistent initial conditions for the observer DAE sys-
tem (35)-(36). Drastic SOC initial estimation errors induce enormous errors in
cell local current estimates. Despite large initial errors, the algebraic state esti-
mates are able to recover the truth signals almost instantly. The performances
of the observer confirms the asymptotic zero error convergence conclusions from
Theorem 3.

7. Conclusion

In this study, the cell-level SOC estimation problem in a battery pack is

investigated. In contrast to conventional approaches where a battery pack is
represented by a lumped cell in which cell-level information is ignored, the
framework proposed in this paper rigorously exploits the high fidelity cell-by-cell
resolution using the interconnection of single cell models. We further challenge
the problem set-up with minimal sensing scenario, where the pack-level voltage
and current are the only measurable signals. It is shown that non-zero high-
order gradients of OCV function is required for the smooth observability of
cell-level SOCs and currents, whereas the observability analysis from linearizing
the nonlinear battery pack model does not provide conclusive results.
A nonlinear DAE system has been proposed to model the battery pack with
parallel-series arrangements, and a state observer for such a DAE system has
been developed. This modeling framework fits naturally with battery appli-
cations, given the interconnections arising from Kirchhoff’s laws. The design
procedure used to build the state observer from this model avoids linearization
or canonical transformations, and it only relies on the assumption of Lipschitz

36
nonlinearities. The resulting state observer benefits from considering the un-
known cell-level currents as algebraic states to be simultaneously estimated with
the differential states in a feedback fashion. The effectiveness of the proposed
estimation approach was demonstrated in simulation.
Consequently it is noted that although fewer measurement signals weaken
cell-level observability, the analysis in this paper provides considerable incentives
to significantly reduce the number of sensors in a battery pack. Future work
will explore the effects of temperature [59] on cell-level state estimation with
reduced sensing.

Acknowledgments

Luis D. Couto would like to thank the Wiener-Anspach Foundation for its
financial support.

References

[1] N. A. Chaturvedi, R. Klein, J. Christensen, J. Ahmed, A. Kojic, Algo-

rithms for advanced battery-management systems, IEEE Control systems
magazine 30 (3) (2010) 49–68. doi:10.1109/MCS.2010.936293.

[2] M. S. Ziegler, J. E. Trancik, Re-examining rates of lithium-ion battery

technology improvement and cost decline, Energy Environ. Sci. (2021) –
Publisher: The Royal Society of Chemistry. doi:10.1039/D0EE02681F.

[3] L. Zhong, C. Zhang, Y. He, Z. Chen, A method for the estimation of the
battery pack state of charge based on in-pack cells uniformity analysis,
Applied Energy 113 (2014) 558–564. doi:10.1016/j.apenergy.2013.08.
008.

[4] L. Ulrich, GM bets big on batteries: A new $2.3 billion plant cranks out
ultium cells to power a future line of electric vehicles, IEEE Spectrum
57 (12) (2020) 26–31. doi:10.1109/MSPEC.2020.9271805.

37
 [5] X. Lin, A. G. Stefanopoulou, Y. Li, R. D. Anderson, State of charge imbal-
ance estimation for battery strings under reduced voltage sensing, IEEE
Transactions on Control Systems Technology 23 (3) (2015) 1052–1062.
doi:10.1109/TCST.2014.2360919.

[6] S. Castano, L. Gauchia, E. Voncila, J. Sanz, Dynamical modeling procedure

of a Li-ion battery pack suitable for real-time applications, Energy Con-
version and Management 92 (2015) 396–405. doi:10.1016/j.enconman.
2014.12.076.

[7] T. Baumhöfer, M. Brühl, S. Rothgang, D. U. Sauer, Production caused

variation in capacity aging trend and correlation to initial cell performance,
Journal of Power Sources 247 (2014) 332–338. doi:10.1016/j.jpowsour.
2013.08.108.

[8] M. Baumann, L. Wildfeuer, S. Rohr, M. Lienkamp, Parameter varia-

tions within Li-ion battery packs–theoretical investigations and experimen-
tal quantification, Journal of Energy Storage 18 (2018) 295–307. doi:
10.1016/j.est.2018.04.031.

[9] Y. Hua, A. Cordoba-Arenas, N. Warner, G. Rizzoni, A multi time-scale

state-of-charge and state-of-health estimation framework using nonlinear
predictive filter for lithium-ion battery pack with passive balance control,
Journal of Power Sources 280 (2015) 293–312. doi:10.1016/j.jpowsour.
2015.01.112.

[10] Y. Zheng, M. Ouyang, L. Lu, J. Li, X. Han, L. Xu, H. Ma, T. A. Dollmeyer,

V. Freyermuth, Cell state-of-charge inconsistency estimation for LiFePO4
battery pack in hybrid electric vehicles using mean-difference model, Ap-
plied Energy 111 (2013) 571–580. doi:10.1016/j.apenergy.2013.05.
048.

[11] X. Zhang, Y. Wang, D. Yang, Z. Chen, An on-line estimation of battery

pack parameters and state-of-charge using dual filters based on pack model,
Energy 115 (2016) 219–229. doi:10.1016/j.energy.2016.08.109Get.

38
[12] T. Zhao, J. Jiang, C. Zhang, K. Bai, N. Li, Robust online state of
charge estimation of lithium-ion battery pack based on error sensitivity
analysis, Mathematical Problems in Engineering 2015 (2015) 11. doi:
10.1155/2015/573184.

[13] H. Kim, K. G. Shin, Efficient sensing matters a lot for large-scale batteries,
in: 2011 IEEE/ACM Second International Conference on Cyber-Physical
Systems, IEEE, 2011, pp. 197–205. doi:10.1109/ICCPS.2011.21.

[14] L. Zhou, L. He, Y. Zheng, X. Lai, M. Ouyang, L. Lu, Massive battery pack
data compression and reconstruction using a frequency division model in
battery management systems, Journal of Energy Storage 28 (2020) 101252.
doi:10.1016/j.est.2020.101252.

[15] X. Zhang, Y. Wang, D. Yang, Z. Chen, An on-line estimation of battery

pack parameters and state-of-charge using dual filters based on pack model,
Energy 115 (2016) 219–229. doi:10.1016/j.energy.2016.08.109.

[16] S. Sepasi, R. Ghorbani, B. Y. Liaw, Improved extended kalman filter for

state of charge estimation of battery pack, Journal of Power Sources 255
(2014) 368–376. doi:10.1016/j.jpowsour.2013.12.093.

[17] J. Wang, B. Cao, Q. Chen, F. Wang, Combined state of charge estimator for
electric vehicle battery pack, Control Engineering Practice 15 (12) (2007)
1569–1576. doi:10.1016/j.conengprac.2007.03.004.

[18] X. Hu, F. Sun, Y. Zou, Estimation of state of charge of a lithium-ion battery

pack for electric vehicles using an adaptive Luenberger observer, Energies
3 (9) (2010) 1586–1603. doi:10.3390/en3091586.

[19] L. D. Couto, S. Schons, D. Coutinho, M. Kinnaert, Observer design for

the series interconnection of Li-ion battery cells subject to reduced voltage
information, in: Dynamic Systems and Control Conference, Vol. 84270,
American Society of Mechanical Engineers, 2020, p. V001T19A003. doi:
10.1115/DSCC2020-3146.

39
[20] L. D. Couto, D. Zhang, A. Aitio, S. Moura, D. Howey, Estimation of pa-
rameter probability distributions for lithium-ion battery string models us-
ing bayesian methods, in: Dynamic Systems and Control Conference, Vol.
84270, American Society of Mechanical Engineers, 2020, p. V001T20A003.

[21] T. Bruen, J. Marco, Modelling and experimental evaluation of parallel con-

nected lithium ion cells for an electric vehicle battery system, Journal of
Power Sources 310 (2016) 91–101. doi:10.1016/j.jpowsour.2016.01.
001.

[22] D. Zhang, L. D. Couto, S. Benjamin, W. Zeng, D. F. Coutinho, S. J. Moura,

State of charge estimation of parallel connected battery cells via descriptor
system theory, in: 2020 American Control Conference (ACC), 2020, pp.
2207–2212. doi:10.23919/ACC45564.2020.9147284.

[23] G.-R. Duan, Analysis and design of descriptor linear systems,

Vol. 23, Springer Science & Business Media, 2010. doi:10.1007/
978-1-4419-6397-0.

[24] R. Nikoukhah, A. S. Willsky, B. C. Levy, Kalman filtering and Riccati

equations for descriptor systems, IEEE Transactions on Automatic Control
37 (9) (1992) 1325–1342. doi:10.1109/CDC.1990.203308.

[25] L. Chisci, G. Zappa, Square-root kalman filtering of descriptor sys-

tems, Systems & Control Letters 19 (4) (1992) 325–334. doi:10.1016/
0167-6911(92)90071-Y.

[26] M. Darouach, M. Boutayeb, Design of observers for descriptor systems,

IEEE Transactions on Automatic Control 40 (7) (1995) 1323–1327. doi:
10.1109/9.400467.

[27] M. Boutayeb, M. Darouach, Observers design for nonlinear descriptor sys-

tems, in: 34th IEEE Conference on Decision and Control, Vol. 3, 1995, pp.
2369–2374 vol.3. doi:10.1109/CDC.1995.480692.

40
[28] G. Zimmer, J. Meier, On observing nonlinear descriptor systems, Systems
& Control Letters 32 (1) (1997) 43–48. doi:10.1016/S0167-6911(97)
00054-6.

[29] J. Åslund, E. Frisk, An observer for non-linear differential-algebraic sys-

tems, Automatica 42 (6) (2006) 959–965. doi:10.1016/j.automatica.
2006.01.026.

[30] R. Rajamani, Observers for lipschitz nonlinear systems, IEEE transactions

on Automatic Control 43 (3) (1998) 397–401. doi:10.1109/9.661604.

[31] S. Kaprielian, J. Turi, An observer for a nonlinear descriptor system, in:

31st IEEE Conference on Decision and Control, 1992, pp. 975–976. doi:
10.1109/CDC.1992.371580.

[32] L. Guoping, D. W. C. Ho, Full-order and reduced-order observers for

Lipschitz descriptor systems: the unified LMI approach, IEEE Transac-
tions on Circuits and Systems II: Express Briefs 53 (7) (2006) 563–567.
doi:10.1109/TCSII.2006.875332.

[33] M. Darouach, L. Boutat-Baddas, Observers for a class of nonlinear singular

systems, IEEE Transactions on Automatic Control 53 (11) (2008) 2627–
2633. doi:10.1109/TAC.2008.2007868.

[34] D. N. Shields, Observer design and detection for nonlinear descriptor

systems, International Journal of Control 67 (2) (1997) 153–168. doi:
10.1080/002071797224234.

[35] J. S. Albuquerque, L. T. Biegler, Decomposition algorithms for on-line

estimation with nonlinear DAE models, Computers & chemical engineering
21 (3) (1997) 283–299. doi:10.1016/0098-1354(94)00107-Y.

[36] V. M. Becerra, P. D. Roberts, G. W. Griffiths, Applying the extended

Kalman filter to systems described by nonlinear differential-algebraic equa-
tions, Control Engineering Practice 9 (3) (2001) 267–281. doi:doi.org/
10.1016/S0967-0661(00)00110-6.

41
[37] R. Mandela, R. Rengaswamy, S. Narasimhan, L. N. Sridhar, Recursive state
estimation techniques for nonlinear differential algebraic systems, Chemical
Engineering Science 65 (16) (2010) 4548–4556. doi:10.1016/j.ces.2010.
04.020.

[38] Y. Puranik, V. A. Bavdekar, S. C. Patwardhan, S. L. Shah, An en-

semble kalman filter for systems governed by differential algebraic equa-
tions (DAEs), IFAC Proceedings Volumes 45 (15) (2012) 531–536. doi:
10.3182/20120710-4-SG-2026.00167.

[39] D. Zhang, L. D. Couto, P. Gill, S. Benjamin, W. Zeng, S. J. Moura, Interval

observer for SOC estimation in parallel-connected lithium-ion batteries, in:
2020 American Control Conference (ACC), IEEE, 2020, pp. 1149–1154.

[40] S. J. Moura, F. B. Argomedo, R. Klein, A. Mirtabatabaei, M. Krstic, Bat-

tery state estimation for a single particle model with electrolyte dynamics,
IEEE Transactions on Control Systems Technology 25 (2) (2016) 453–468.
doi:10.1109/TCST.2016.2571663.

[41] D. Zhang, S. Dey, L. D. Couto, S. J. Moura, Battery adaptive observer for a

single-particle model with intercalation-induced stress, IEEE transactions
on control systems technology 28 (4) (2019) 1363–1377.

[42] D. Zhang, S. Dey, H. E. Perez, S. J. Moura, Remaining useful life estimation

of lithium-ion batteries based on thermal dynamics, in: 2017 American
Control Conference (ACC), IEEE, 2017, pp. 4042–4047. doi:10.23919/
ACC.2017.7963575.

[43] R. Drummond, L. D. Couto, D. Zhang, Resolving kirchhoff’s laws for

state-estimator design of Li-ion battery packs connected in parallel (2020).
arXiv:2010.16264.

[44] X. Fan, W. Zhang, Z. Wang, F. An, H. Li, J. Jiang, Simplified bat-

tery pack modeling considering inconsistency and evolution of current dis-

42
 tribution, IEEE Transactions on Intelligent Transportation Systemsdoi:
10.1109/TITS.2020.3010567.

[45] M. H. Hofmann, K. Czyrka, M. J. Brand, M. Steinhardt, A. Noel, F. B.

Spingler, A. Jossen, Dynamics of current distribution within battery cells
connected in parallel, Journal of Energy Storage 20 (2018) 120–133. doi:
10.1016/j.est.2018.08.013.

[46] L. Petzold, Differential/algebraic equations are not ODE’s, SIAM Journal

on Scientific and Statistical Computing 3 (3) (1982) 367–384. doi:10.
1137/0903023.

[47] R. Vidal, A. Chiuso, S. Soatto, S. Sastry, Observability of linear hybrid

systems, in: Hybrid Systems: Computation and Control, Springer Berlin
Heidelberg, Berlin, Heidelberg, 2003, pp. 526–539.

[48] L. M. Silverman, H. Meadows, Controllability and observability in time-

variable linear systems, SIAM Journal on Control 5 (1) (1967) 64–73. doi:
10.1137/0305005.

[49] K. E. Brenan, S. L. Campbell, L. R. Petzold, Numerical solution of initial-

value problems in differential-algebraic equations, SIAM, 1995. doi:10.
1137/1.9781611971224.

[50] S. L. Campbell, E. Griepentrog, Solvability of general differential algebraic

equations, SIAM Journal on Scientific Computing 16 (2) (1995) 257–270.
doi:10.1137/0916017.

[51] S. L. Campbell, Least squares completions for nonlinear differential al-

gebraic equations, Numerische Mathematik 65 (1) (1993) 77–94. doi:
10.1007/BF01385741.

[52] S. L. Campbell, C. W. Gear, The index of general nonlinear

DAEs, Numerische Mathematik 72 (2) (1995) 173–196. doi:10.1007/
s002110050165.

43
[53] S. L. Campbell, W. J. Terrell, Observability of linear time-varying descrip-
tor systems, SIAM Journal on Matrix Analysis and Applications 12 (3)
(1991) 484–496. doi:10.1137/0612035.

[54] W. J. Terrell, Observability of nonlinear differential algebraic systems,

Circuits, Systems and Signal Processing 16 (2) (1997) 271–285. doi:
10.1007/BF01183279.

[55] S. L. Campbell, E. Moore, Y. Zhong, Utilization of automatic differentia-

tion in control algorithms, IEEE transactions on automatic control 39 (5)
(1994) 1047–1052. doi:10.1109/9.284891.

[56] M. Rausch, S. Streif, C. Pankiewitz, R. Findeisen, Nonlinear observability

and identifiability of single cells in battery packs, in: 2013 IEEE Inter-
national Conference on Control Applications, 2013, pp. 401–406. doi:
10.1109/CCA.2013.6662782.

[57] S. Zhao, S. R. Duncan, D. A. Howey, Observability analysis and state

estimation of lithium-ion batteries in the presence of sensor biases, IEEE
Transactions on Control Systems Technology 25 (1) (2017) 326–333. doi:
10.1109/TCST.2016.2542115.

[58] I. Baccouche, S. Jemmali, B. Manai, N. Omar, N. E. B. Amara, Improved

OCV model of a Li-Ion NMC battery for online SOC estimation using the
extended kalman filter, Energies 10 (6). doi:10.3390/en10060764.

[59] D. Zhang, L. D. Couto, P. S. Gill, S. Benjamin, W. Zeng, S. J. Moura,

Thermal-enhanced adaptive interval estimation in battery packs with het-
erogeneous cells, IEEE Transactions on Control Systems Technology.

44