ADAPTIVE PDE OBSERVER FOR BATTERY SOC/SOH ESTIMATION

Scott J. Moura Nalin A. Chaturvedi

Miroslav Krstic Research and Technology Center
Cymer Center for Control Systems and Dynamics Robert Bosch LLC
Mechanical and Aerospace Engineering Palo Alto, California 94304
University of California, San Diego nalin.chaturvedi@us.bosch.com
La Jolla, California 92093
smoura@ucsd.edu, krstic@ucsd.edu

ABSTRACT j Positive (+) or negative (-) electrode

This paper develops an adaptive PDE observer for battery
state-of-charge (SOC) and state-of-health (SOH) estimation. Real-
time state and parameter information enables operation near
physical limits without compromising durability, thereby unlock-
ing the full potential of battery energy storage. SOC/SOH esti-
mation is technically challenging because battery dynamics are
governed by electrochemical principles, mathematically modeled
by partial differential equations (PDEs). Simultaneous state and
parameter estimation is extremely challenging in PDE models.
Consequently, several new theoretical ideas are developed, inte-
grated together, and tested. These include a backstepping PDE
state estimator, a Padé-based parameter identifier, nonlinear pa-
rameter sensitivity analysis, and adaptive inversion of nonlinear
output functions. The end result is the first combined SOC/SOH
battery estimation algorithm that identifies physical system vari-
ables via an electrochemical model, from measurements of voltage
and current only.
NOMENCLATURE
A Cell cross sectional area [m2 ]
a j Specific interfacial surface area [m2 /m3 ]
c0e Li concentration in electrolyte phase [mol/m3 ]
j extreme importance due to their applications in electrified trans-
cs Li concentration in solid phase [mol/m3 ]
j portation and energy storage systems for renewable sources. The
css Li concentration at particle surface [mol/m3 ]
j
cs,max Max Li concentration in solid phase [mol/m3 ]
j
Ds Diffusion coefficent in solid phase [m2 /sec3 ]
F Faraday’s constant [C/mol]
I Input current [A]
j
i0 Exchange current density [V]
k j Reaction rate [A·mol1.5 /m5.5 ]
L j Electrode thickness [m]
nLi Total number of Li ions [mol]
q Boundary input coefficient parameter
R Universal gas constant [J/mol-K]
R f Lumped current collector resistance [Ω]
j
Rs Particle radius [m]
r Radial coordinate [m] or [m/m]
T Cell temperature [K]
t Time [sec] or [sec/sec]
U j Equilibrium potential [V]
V Output voltage [V]
α j Anodic/cathodic transfer coefficient
ε Diffusion parameter
j
εs Volume fraction of solid phase
1 INTRODUCTION
This paper investigates an adaptive PDE observer for com-
bined state-of-charge (SOC) and state-of-health (SOH) estimation
in batteries, using an electrochemical model.
Accurate battery SOC estimation algorithms are currently of
relevancy of this topic is further underscored by the 27.2 billion
USD federal government investment in energy efficiency and re-
newable energy research, including advanced batteries, under the
American Recovery and Reinvestment Act (ARRA) of 2009. As
such, battery management systems within these advanced trans-
portation and energy infrastructures must have accurate knowl-
edge of internal battery energy levels [1]. Such knowledge enables adaptive output function inversion. The final result is an adaptive
them to efficiently route energy while satisfying power demands observer for simultaneous SOC/SOH estimation which identi-
and device-level operating constraints [2]. fies physical battery system variables, from current and voltage
Monitoring battery SOC and SOH is particularly challenging measurements only.
for several technical reasons. First, directly measuring Li concen- The paper is organized as follows: Section 2 describes the
tration or physical examination of cell components is impractical electrochemical-based single particle model. Sections 3-6 de-
outside specialized laboratory environments [3, 4]. Second, the scribe the subsystems of the adaptive observer, including the state
dynamics are governed by partial differential algebraic equations estimator, PDE parameter identifier, output function parameter
derived from electrochemical principles [5]. The only measurable identifier, and adaptive output function inversion. Sections 7 and
quantities (voltage and current) are related to the states through 8 respectively provide simulation results and a summary of the
boundary values. Finally, the model’s parameters vary widely key contributions.
with electrode chemistry, packaging, and time. In this paper we
directly address these technical challenges. Namely, we design
an adaptive observer using a reduced-form PDE model based 2 ELECTROCHEMICAL CELL MODEL & ANALYSIS
upon electrochemical principles. As such, the algorithm estimates The single particle model (SPM) was first applied to lithium
physical variables directly related SOC and SOH, a first to the battery systems in [14] and is the model we utilize in this work.
authors’ knowledge. The key idea is that the solid phase of each electrode can be ide-
Over the past decade research on battery SOC/SOH estima- alized as a single spherical particle. This model results if one
tion has experienced considerable growth. One may divide this assumes the electrolyte Li concentration is constant in space and
research by the battery models each algorithm employs. time. This assumption works well for small currents or elec-
The first category considers estimators based upon equivalent trolytes with large electronic conductivities. However, it induces
circuit models (ECMs). These models use circuit elements to errors at large C-rates. Moreover, we assume constant temperature.
mimic the phenomenological behavior of batteries. For exam- Figure 1 provides a schematic of the SPM concept. Mathemati-
ple, the seminal work by Plett [6] applies an extended Kalman cally, the model consists of two diffusion PDEs governing each
filter to simultaneously identify the states and parameters of an electrode’s concentration dynamics, where input current enters
ECM. Verbrugge and his co-workers used ECMs with combined as a Neumann boundary condition. Output voltage is given by
coulumb-counting and voltage inversion techniques in [7] and a nonlinear function of the state values at the boundary and the
adaptive parameter identification algorithms in [8]. More recently, input current.
a linear parameter varying approach was designed in [9]. The Although this model captures less dynamic behavior than
key advantage of ECMs is their simplicity. However, they of- other electrochemical-based estimation models [14], its mathe-
ten require extensive parameterization for accurate predictions. matical structure is amenable to adaptive observer design.
This often produces models with non-physical parameters, whose
complexity becomes comparable to electrochemical models. 2.1 Single Particle Model
The second category considers electrochemical models, Diffusion in each electrode is governed by Fick’s law in
which account for the diffusion, intercalation, and electrical dy- spherical coordinates:
namics. Although these models can accurately predict internal
state variables, their mathematical structure is generally too com-  
plex for controller/observer design. Therefore, these approaches ∂c− 2 ∂c− ∂2 c−
s
(r,t) = D−
s
s
(r,t) + s
(r,t) (1)
combine model reduction and estimation techniques. Some of ∂t r ∂r ∂r2
the first studies within this category use a “single particle model”  
∂c+ 2 ∂c+ ∂2 c+
(SPM) of electrochemical battery dynamics in combination with
s
(r,t) = D+
s
s
(r,t) + s
(r,t) (2)
∂t r ∂r ∂r2
an extended Kalman filter [10, 11]. A separate research group em-
ployed residue grouping for model reduction and linear Kalman
filters for observers [12]. The authors of [13] apply simplifica- with Neumann boundary conditions:
tions to the electrolyte and solid phase concentration dynamics to
perform SOC estimation. To date, however, simultaneous SOC ∂c− ∂c− I(t)
and SOH estimation using electrochemical models remains an
s
(0,t) = 0, s
(R−
s ,t) = − − (3)
∂r ∂r Ds Fa− AL−
open question.
∂c+
s
+
∂cs + I(t)
In this paper we extend the aforementioned research by de- (0,t) = 0, (Rs ,t) = + + + (4)
signing an electrochemical model based adaptive observer for
∂r ∂r Ds Fa AL
simultaneous SOC/SOH estimation. Several novel theoretical
The Neumann boundary conditions at r = R+ −
ideas are developed, integrated, and tested. These include a PDE s and r = Rs signify
backstepping state estimator, Padé-based PDE parameter identi- that the flux entering the electrode is proportional to the input
fier, nonlinear identifiability analysis of the output equation, and current I(t). The Neumann boundary conditions at r = 0 are
 Anode Separator Cathode This model contains the property that the total number of
lithium ions is conserved [13]. Mathematically, dtd (nLi ) = 0
Solid
Electrolyte Z R+
ε+ L + A s
nLi = 4 s + 3 4πr2 c+
s (r,t)dr
Li 3 π(Rs ) 0
+
Z R−
ε− −
s L A s

V(t) + 4 − 3 4πr2 c−
s (r,t)dr (8)
3 π(Rs ) 0

---Single Particle Model---

I(t) This property will become important, as it relates the total con-
Li+
centration of lithium in the cathode and anode. We leverage this
cs-(r,t) cs+(r,t) fact to perform model reduction in the state estimation problem.
I(t) Li+
In the following sections we describe each subsystem of the
adaptive observer. A block diagram of the composed system is
r r
provided in Fig. 2.
Rs- Rs+

V(t) = h(cs-(Rs-,t), cs+(Rs+,t), I(t)) 3 STATE ESTIMATION

Figure 1. EACH ELECTRODE IS IDEALIZED AS A SINGLE POROUS 3.1 Observability & Model Reduction
SPHERICAL PARTICLE. DURING CHARGING/DISCHARGING, LI IN- For the purpose of observer design we reduce the SPM by
TERCALATES INTO/OUT OF THESE SPHERICAL PARTICLES. approximating the cathode diffusion dynamics (2) by its equilib-
rium. This step is mathematically motivated by the fact that the
required for well-posedness. Note that the states for the two SPM states are weakly observable from voltage measurements, as
PDEs are dynamically uncoupled, although they have proportional has been previously noted in the literature [11]. It turns out that
boundary inputs. approximating the cathode dynamics as instantaneous produces
The measured terminal voltage output is governed by a com- a reduced system whose states are locally strongly observable.
bination of electric overpotential, electrode thermodynamics, and Moreover, physical motivation exists for this reduction when dif-
Butler-Volmer kinetics. The end result is: fusion dynamics are significantly faster in the cathode than the
anode, a common characteristic of certain anode/cathode combi-
 
RT I(t) nations. We discuss these points in succession.
V (t) = sinh−1
αF 2a+ AL+ i+ +
0 (css (t)) Lack of observability can be shown using a number of tech-
  niques. For example, one may (i) approximate the PDEs by ODEs
RT −1 I(t)
− sinh using the finite difference method, producing a tri-diagonal matrix
αF 2a− AL− i− −
0 (css (t))
− −
A, (ii) linearize the output equation about the states, producing a
+U + (c+
ss (t)) −U (css (t)) − R f I(t) (5) matrix C, (iii) and compute the rank of the observability matrix
for the pair (A,C) [15].
j
where the exchange current density i0 and solid-electrolyte surface The reduced SPM has a PDE given by (1), boundary condi-
j tions given by (3), and output equation:
concentration css are, respectively:
q  
j j j j
i0 (cssj ) =k j
c0e css (t)(cs,max − css (t)) (6) RT I(t)
V (t) = sinh−1 −
cssj (t) = csj (Rsj ,t), j ∈ {+, −} (7) α+ F 2a+ AL+ i+
0 (αcss (t) + β)
 
RT I(t)
− − sinh−1 (9)
The functions U + (·) and U − (·) are the equilibrium potentials of α F 2a− AL− i− −
0 (css (t))
each electrode material, given the surface concentration. Mathe- +U + (αc− − −
ss (t) + β) −U (css (t)) − R f I(t)
matically, these are strictly monotonically decreasing functions
of their input. This fact implies that the inverse of its derivative is
Note that c+ −
always finite, a property which we require later. Further details ss (t) has been replaced by αcss (t) + β. This is the
on the electrochemical principles used to derive these equations critical detail of the reduced SPM. The equilibrium of the cathode
can be found in [1, 5]. states (i.e., c+ +
s (r,t) = css (t)) can be computed from the conserva-
 I(t) css+ V(t) Subsystem Legend
Diffusion
css- h(css+,css-, I)
PDE
I(t) Plant
φ ζ φ(V, I; θh)
Padé-based 1 I(t)
LSQ Adaptive
Regression Λ(s) čss-
Update Law Observer
z Model σ
θˆ pde
PDE
" ε̂ % _+ Param ID
θˆ pde → $ '
Diffusion ĉ -
$# q̂ '& ss Φ NL-LSQ θˆh Output Fcn
PDE Copy + Regressor Param ID
Update Law
Output Inj.
ε̂ , q̂ Adaptive
-(r,t) I(t) Output Fcn
I(t) ĉs
Inversion

Figure 2. BLOCK DIAGRAM OF ADAPTIVE OBSERVER. IT IS COMPOSED OF THE BACKSTEPPING STATE ESTIMATOR (BLUE), PDE PARAMETER
IDENTIFIER (GREEN), OUTPUT FUNCTION PARAMETER IDENTIFIER (RED), AND ADAPTIVE OUTPUT FUNCTION INVERSION (ORANGE).

tion of Li property in (8) to produce the relationship1 : LiFePO4 cells with doped nano-scale cathode materials [19]. For
other cells, the requisite diffusive time scale separation property
1   may not exist.
c+
ss (t) = nLi − ε− − −
s L Acss (t) (10)
ε+ +
s L A The SPM comprises linear dynamics and a nonlinear output
function. In general an output injection-based estimator would be
− − nonlinear for this class of systems. However, we design a linear
where α = − εεs+ LL+ and β = ε+nLLi+ A .
s s estimator in this paper by injecting the boundary state error. This
One can show this system is locally observable (i.e. in the idea requires us to calculate the boundary state from the measured
linear sense) by using the same finite difference and linearization voltage, demonstrated visually by the block diagram in Fig. 2.
approach described above. Ultimately, we guarantee observability In [20] we show the output function (9) is invertible with respect
for this reduced SPM by designing the observer gains such that the to the boundary state c−ss , uniformly in the input current I(t).
estimation error dynamics mimic an exponentially stable target
system. This is the core concept behind backstepping observer
design [16].
Physical motivation sometimes exists for approximating the 3.2 Normalization and State Transformation
cathode diffusion dynamics as instantaneous. Significant research
Next we perform normalization and state transformation to
efforts on manufacturing and material science techniques for cath-
simplify the mathematical structure of the observer. First scale
ode materials has enabled researchers to attain nano-scale particle
the radial r and time t coordinates as follows:
sizes and faster diffusion rates [17]. The result is characteristic
diffusion times (mathematically R2s /Ds ) which are often orders
of magnitude less in the cathode than the anode. Parallel studies
have been performed on the anode side (see e.g. [18]), however r D−s
r̄ = t¯ = t (11)
they are less prevalent. Hence, approximating cathode diffusion R−s (R−
s )
2

by its equilibrium is a reasonable approximation for certain cath-

ode/anode combinations. This insight was also observed through a
previous parameter identification study on commercially available Henceforth we will drop the bars over the space and time coordi-
1 To be technically correct, the cathode concentration should depend
nates to simplify notation. Next we perform a state transformation
on the anode concentration summed over the spherical volume: c+ to eliminate the first spatial derivative in the spherical diffusion
ss (t) =
1 3ε− − R −
s L A Rs 2 c− (r,t)dr . However, this results in a nonlinear
equation (1). Namely, let
ε+ L+ A
nLi − −3 0 4πr s
s 4πRs
output equation which depends on the in-domain states, as well as the boundary
state. This would create additional complexity to the backstepping approach we
employ in this paper. c(r,t) = rc−
s (r,t) (12)
This normalization and state transformation produces the follow- Also note the parameters ε in (16), (20)-(22) and q in (18). In
ing PDE with Dirichlet and Robin boundary conditions: the subsequent section we design an identifier for these parame-
ters. We form an adaptive observer by replacing these parameters
with their estimates, via the certainty equivalence principle [21].
∂c ∂2 c
(r,t) = ε 2 (r,t) (13)
∂t ∂r
c(0,t) = 0 (14)
∂c 4 PDE PARAMETER IDENTIFICATION
(1,t) − c(1,t) = −qρI(t) (15)
∂r Next we design an identification algorithm for the diffusion
and boundary input coefficients in (13) and (15), respectively.
and nonlinear output map given by (9) where c+ ss = αc(1,t) + β Identification of the diffusion coefficient ε from boundary mea-
(see (10)) and c− − − −
ss = c(1,t). The parameter ρ = Rs /(Ds Fa AL )
−
surements is a significant fundamental challenge [22], for the
groups together known parameters. The parameters ε and q are following reason. In finite-dimensional state-space systems we
nominally equal to one. Respectively, they represent uncertainty typically write the system in observable canonical form. This
in the diffusion and boundary input coefficients, which we identify structure enables one to uniquely identify state-space parameters
in Section 4. from input-output data. In our problem we require a parametric
model where the diffusion coefficient multiplies measured data
only. There is no clear way to do this for PDEs. This moti-
3.3 Backstepping PDE State Estimator
vates our new contribution: utilizing a reduced-order model (Padé
In previous work [20], we designed a backstepping PDE state
approximation) for the parameter identification.
estimator for the SPM. The state estimator structure consists of a
copy of the plant (13)-(15) plus boundary state error injection, as
follows:
4.1 Padé Approximates
∂ĉ ∂2 ĉ The PDE model (13)-(15) can be written in the frequency
(r,t) = ε 2 (r,t) + p1 (r)c̃(1,t) (16) domain as a transcendental transfer function:
∂t ∂r
ĉ(0,t) = 0 (17)
p 
∂ĉ
(1,t) − ĉ(1,t) = −qρI(t) + p10 c̃(1,t) (18) css (s) −qρ sinh s/ε
∂r G(s) = = p  p  p  (23)
I(s) s/ε cosh s/ε − sinh s/ε
where the boundary state error is given by:

c̃(1,t) = ϕ(V (t), I(t)) − ĉ(1,t) (19) We now apply Padé approximations of the transcendental
transfer function (23). Padé approximants represent a function by
a ratio of two power series. The defining characteristic of a Padé
and the function estimation gains are given by: approximate is that its Taylor series matches the Taylor series of
the function it is approximating. Another useful property of Padé
 
−λr 2λ approximates is that they naturally contain poles and zeros. The
p1 (r) = I1 (z) − I2 (z) (20) Padé expansion takes the following form:
2εz εz
r
λ 2
where z = (r − 1) (21)
ε  ∑Nk=0 bk sk
1 λ G(s) = lim (24)
p10 = 3− (22) N→∞ 1 + ∑N ak sk
k=1
2 ε

and I1 (z) and I2 (z) are, respectively, the first and second order
modified Bessel functions of the first kind. Please consult [20]
for the stability proof.
Note that the estimator is linear in the state because we use
the boundary state for error injection. The plant boundary state
is computed by inverting the nonlinear output mapping w.r.t. the
boundary state, given a current input (i.e. ϕ(V (t), I(t))). The
output function inversion is discussed in detail in Section 6.
Figure 3 provides bode plots of G(s) and several Padé approxi-
mates. Their analytical expressions are supplied in Table 1. The
Padé approximates capture low frequency dynamics well. Accu-
racy at high frequency increases as the Padé order increases. We
low-pass filter the input-output signals such that data is retained
where the Padé approximation is sufficiently accurate.
Our immediate goal is to design a parameter identification
scheme for the Padé approximation of the original PDE model.
 Since the parametric model contains time derivatives of measured
6 Transcendental TF
10 1st Order Pade signals, we employ filters [21] to avoid differentiation:
2nd Order Pade
Magnitude

4 3rd Order Pade

10
σ̇1 = σ2 (27)
2
σ̇2 = −λ0 σ1 − λ1 σ2 + css (28)
10
ζ˙ 1 = ζ2 (29)
0
10 −3 ζ˙ 2 = −λ0 ζ1 − λ1 ζ2 + I (30)
−2 −1 0 1 2 3
10 10 10 10 10 10 10
−40
where the polynomial Λ(s) = s2 + λ1 s + λ0 is chosen Hurwitz.
−50 One can analytically show that selecting the roots of Λ(s) results
Phase [deg]

−60
in a trade-off between convergence rate (via level of persistence
of excitation) and parameter bias (error induced by Padé approxi-
−70 mation). Consequently, the parametric model is:
−80
1 2
−90 −3 (−λ0 σ1 − λ1 σ2 + css ) = −3ρqε2 ζ1 − ρqεζ2 − εσ2 (31)
10 10
−2
10
−1
10
0
10
1 2
10 10
3
35 7
Frequency [rad/s]

Figure 3. BODE PLOTS OF THE TRANSCENDENTAL TRANSFER

Let us denote the vector of unknown parameters by:
FUNCTION (23) AND PADÉ APPROXIMATES IN TABLE 1.
 T
θ pde = qε2 qε ε (32)

4.2 Least Squares Identification Then the parametric model can be expressed in matrix form as
z pde = θTpde φ, where
We utilize the first order Padé approximant as the nominal
model. Namely,
1
z pde = (−λ0 σ1 − λ1 σ2 + css ) (33)
35
 T
Css (s) −3qρε − 27 qρs 2
≈ P1 (s) = 1 2
(25) φ = −3ρζ1 − ρζ2 − σ2 (34)
I(s) s + 35ε s 7

Given this linearly parameterized model, we chose a least-squares

Assuming zero initial conditions and applying the inverse Laplace update law of the form [21]:
transform produces the following linearly parameterized model:
T
z pde − θ̂ pde φ
θ̂˙ pde = Ppde φ (35)
1 2 m2pde
c̈ss (t) = −εċss (t) − 3ρqε2 I(t) − ρqεI(t)
˙ (26)
35 7 φφT T
Ṗpde = −Ppde Ppde , Ppde (0) = Ppde0 = Ppde0 > 0 (36)
m2pde
m2pde = 1 + γ pde φT φ, γ pde > 0 (37)
Table 1. PADÉ APPROXIMATES OF THE PDE MODEL (13)-(15)

Order, k Pk (s) 4.2.1 Managing Overparameterization with the

−qρ( 72 s+3ε) Pseudoinverse An important implementation issue with the
1 1 s+1
s( 35ε ) proposed Padé approximation approach is overparameterization.
−qρ( 165ε 1 s2 + 4 s+3ε
) That is, the physical parameters must be uniquely determined
 11 
2 1 3 s+1 from the parameter vector θ̂ pde
s 2 s2 + 55ε
3465ε
 
4 3 2 2 2
−qρ 2 s + 195ε s + 5 s+3ε  
3 
1
75075ε
3 2 s2 + 1 s+1
 c2
qε  
s 3s + 2 15ε   ε̂
675675ε 2275ε θ̂ pde = qεb −→ = θ̂εq (38)
q̂
ε̂
Coincidently, the particular nonlinear form (products and pow- where c+
ss (t; θ1 ) and the parameter vector θ are
ers) of the elements in vector θ̂ pde allows us to write a set of
linear equations using a logarithmic nonlinear transformation and
ε−
s L
−
− θ1
properties of the logarithm function: c+
ss (t; θ1 ) = − + + css (t) + + +
εs L εs L A
" #T
    1 1
2 1   log qεˆ 2 θ = nLi , p , p ,Rf (43)
1 log ε̂   a+ AL+ k+ c0e a− AL− k− c0e
1 =  log (qε)
ˆ  (39)
log q̂
1 0 log (ε̂)
We have selected the elements of θ because diminishing nLi phys-
ically models capacity fade and increasing values for the other
which we re-write into compact notation as
parameters capture various forms of internal resistance.
The following sensitivity analysis is performed in discrete
Aεq log(θ̂εq ) = log(θ̂ pde ) (40) time, since the required data is supplied in discrete time. Let
k index time such that t = k∆T , k ∈ 1, 2, ..., nT . The sensitiv-
ity of the output with respect to variations in the parameter θi
where log(θ) = [log(θ1 ), log(θ2 ), ...]T . The parameter vector θ̂εq at time index k is defined as Si,k = ∂h(k∆T ;θ)
. For each param-
∂θi
can be uniquely solved from (39) via the Moore-Penrose pseu- eter θi , stack the sensitivities at time indeces k = 1, 2, ...nT , i.e.
doinverse. That is: Si = [Si,1 , Si,2 , . . . , Si,nT ]T . Denote S = [S1 , S2 , S3 , S4 ], such that
S ∈ RnT ×4 . A particular decomposition of ST S reveals useful
log(θ̂εq ) = (ATεq Aεq )−1 ATεq log(θ̂ pde ) (41) information about linear dependence between parameters. Let
ST S = DT CD where

This method works well in practice with respect to feeding param-  

kS1 k 0 0 0
eter estimates into the adaptive observer (lower-left-hand block in  0 kS2 k 0 0 
Fig. 2), since the pseudoinverse ultimately involves computation- D=
 0

0 kS3 k 0 
ally efficient matrix algebra.
0 0 0 kS4 k
 hS1 ,S2 i hS1 ,S3 i hS1 ,S4 i 
1 kS1 kkS2 k kS1 kkS3 k kS1 kkS4 k
5 OUTPUT FUNCTION PARAMETER IDENTIFICATION  hS2 ,S1 i hS2 ,S3 i hS2 ,S4 i 
 kS kkS k 1 
The greatest difficultly in battery estimation arguably stems C= 2 1
 hS3 ,S1 i hS3 ,S2 i
kS2 kkS3 k kS2 kkS4 k
hS3 ,S4 i

 (44)
from the nonlinear relationship between SOC and voltage [9].  kS kkS k kS kkS k 1 kS3 kkS4 k 
3 1 3 2
hS4 ,S1 i hS4 ,S2 i hS4 ,S3 i
We directly address this difficulty by developing an identifica- kS4 kkS1 k kS4 kkS2 k kS4 kkS3 k 1
tion algorithm for the uncertain parameters in the nonlinearly
parameterized output function (9). First, we analyze parameter
where k · k denotes the Euclidian norm and h·, ·i is the inner
identifiability to assess which subset of parameters are uniquely hSi ,S j i
identifiable. Second, we apply nonlinear least squares to this product. By the Cauchy Schwarz inequality −1 ≤ kSi kkS jk
≤ 1.
subset. i j hS ,S i
This has the interpretation that values of kSi kkS jk
near -1 or 1
imply strong linear dependence between parameters θi and θ j ,
5.1 Identifiability whereas values near zero imply orthogonality.
A necessary first step in nonlinear parameter identification is An example for the matrix C is provided in (45). This exam-
a parameter sensitivity analysis. We specifically apply the ranking ple analyzes parameter sensitivity for a UDDS drive cycle data
procedure outlined in [23] to assess linear dependence. Consider set applied to the SPM battery model.
the output function (9) written in parametric form:
 
1 −0.3000 0.2908 0.2956
   −0.3000 1 −0.9801 −0.9805 
RT θ2 I(t) C=
 0.2908 −0.9801
 (45)
h(t; θ) = V (t) = sinh−1  q  1 0.9322 
αF 2 css (t; θ1 )(c+
+ + 0.2956 −0.9805 0.9322 1
s,max − css (t; θ1 ))
 
RT θ 3 I(t) Note that strong linear dependence exists between θ2 , θ3 , θ4 . This
− sinh−1  q 
αF 2 c− (t)(c− − c− (t))
ss s,max ss property is uniformly true across various drive cycles (e.g. US06,
+ − −
SC04, LA92, naturalistic micro-trips). This means it is difficult to
+U (c+
ss (t; θ1 )) −U (css (t)) − θ4 I(t) (42) determine how each individual parameter value changes, amongst
these three parameters. As a result, we identify only two pa- We now choose a least-squares parameter update law:
rameters, nLi and R f , which capture capacity and power fade,
respectively. θ̂˙ h = Ph enl Φ (51)
ΦΦT T
5.2 Nonlinear Least Squares Ṗh = −Ph Ph , Ph (0) = Ph0 = Ph0 >0 (52)
m2h
Now our immediate goal is to identify the parameter vec-
tor θh = [nLi R f ]T via a nonlinear least squares identification m2h = 1 + γh ΦT Φ, γh > 0 (53)
algorithm. Define θ̃h = θh − θ̂h and write (42) in terms of θ̃h

" # 6 ADAPTIVE OUTPUT FUNCTION INVERSION

RT −1 I(t) In Section 3 we discussed how a linear state observer can be
V (t; θh ) = sinh designed by using boundary values of the PDE. These boundary
αF 2a+ AL+ i+ +
0 (css (t; θ̃h1 + θ̂h1 ))
  values must be processed from measurements by inverting the
RT −1 I(t) nonlinear output function. In this section we design an adaptive
− sinh (46)
αF 2a− AL− i− −
0 (css (t)) output function inversion scheme which utilizes the parameter
+U + (c+ − − estimate θh generated from Section 5.
ss (t; θ̃h1 + θ̂h1 )) −U (css (t)) − (θ̃h2 + θ̂h2 )I(t)
Our goal is to solve g(c− −
ss ,t) = 0 for css , where

Next we take the Maclaurin series expansion with respect to θ̃h " #
RT I(t)
g(c−
ss ,t) = sinh −1
" # αF 2a+ AL+ i+ +
0 (css (t; θ̂h1 ))
RT −1 I(t)  
V (t; θh ) = sinh RT I(t)
2a+ AL+ i+ + − sinh−1 (54)
0 (css (t; θ̂h1 ))
αF
  αF 2a− AL− i− −
0 (css (t))
RT I(t) +U + (c+ − −
ss (t; θ̂h1 )) −U (css (t)) − θ̂h2 I(t) −V (t)
−1
− sinh
αF 2a− AL− i− −
0 (css (t))
− −
+U + (c+
ss (t; θ̂h1 )) −U (css (t)) − θ̂h2 I(t) The main idea is to construct an ODE whose equilibrium satisfies
∂h g(c−
ss ,t) = 0 and is locally exponentially stable. This can be
+ (t; θ̂h )θ̃h1 − I(t)θ̃h2 + O(θ̃Th θ̃h ) (47) viewed as a continuous-time version of Newton’s method for
∂θh1
solving nonlinear equations [21]. Consider the ODE
Truncate the higher order terms and re-arrange the previous ex- d  − 
pression into the matrix form g(čss ,t) = −γg(č−
ss ,t) (55)
dt

enl = θ̃Th Φ (48) whose equilibrium satisfies g(c− ss ,t) = 0. We expand and re-
arrange this equation into the familiar Newton’s update law
where the nonlinear error term enl depends on the parameter  −1  
estimates θ̂h d − ∂g − − ∂g −
čss = − (č ,t) γg(č ,t) + (č ,t) (56)
dt ∂c−
ss
ss ss
∂t ss
" #
RT −1 I(t) One can prove Lyapunov stability of this ODE, given appropriate
enl =V (t) − sinh
αF 2a+ AL+ i+ +
0 (css (t; θ̂h1 )) bounds ∂g/∂c− −
ss and ∂g/∂t. The state čss of ODE (56) provides a
  recursive estimate of the surface concentration css (t) from mea-
RT −1 I(t)
+ sinh sured current and voltage data, adapted according to the parameter
αF 2a− AL− i− −
0 (css (t))
estimate θ̂h . The processed surface concentration č− ss supplies the
− −
−U + (c+
ss (t; θ̂h1 )) +U (css (t)) + θ̂h2 I(t) (49) “measured output” for the state estimator in Section 3.
Note that analytically proving stability of the fully-composed
and the regressor vector Φ is defined as: estimator is difficult. Consequently, we utilize simulation studies
described in the following section.
h iT
Φ= ∂h
∂θh1 (t; θ̂h ), −I(t) (50)
7 SIMULATIONS
In this section we present numerical experimental results,
which depends upon measured signals and parameter estimates. which demonstrate the adaptive PDE observer’s performance.
 3.5
I V
V̂

Voltage [V]
3.4
Current [A]

10
3.3
0
3.2

−10 3.1
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
60

Voltage Error [mV]

0.8 S O C b−u l k V − V̂
40
Bulk SOC

−
0.6 Sd
O Cb u l k
20

0.4 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
15
0.3 − ε̂
Bulk SOC Error

S O C b−u l k− Sd
O Cb u l k

PDE Params
0.2 10 q̂
0.1 Tru e Val u e
5
0
−0.1 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
10
Surface Concentration

3
Output Fcn Params

c− −
s s − ĉ s s n̂ L i/n L i
Error [kmol]

c s s − č −
−
ss
5
2 R̂ f/R f

0 Tru e Val u e
1
−5
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [min] Time [min]

Figure 4. STATE AND PARAMETER ESTIMATION RESULTS FOR UDDSX2 DRIVE CYCLE. ZERO MEAN GAUSSIAN NOISE WITH A 2 mV VARIANCE
WAS ADDED TO THE VOLTAGE MEASUREMENT. THE SPM PROVIDES THE “MEASURED” PLANT DATA. STATE AND PARAMETER ESTIMATES
WERE INITIALIZED WITH INCORRECT VALUES.

Specifically, we apply the observer to the original SPM (includes ilar for various other initial conditions and drive cycle inputs,
diffusion in both electrodes). The model parameters used in this including US06, SC04, LA92, and naturalistic micro trip data.
study originate from the genetic algorithm-based parameter iden-
tification study performed on commercial lithium-iron phosphate Z 1
− 3
cells in [19]. For these parameters the characteristic diffusion SOCbulk (t) = r 2 c−
s (r,t)dr (57)
times (R2s /Ds ) are 745 sec and 0.32 sec for the anode and cath- c−
s,max 0

ode, respectively, which supports the argument for approximating

cathode diffusion as instantaneous.
The state and parameter estimates are initialized at incor-
rect values: ĉ− −
s (r, 0) = 0.6cs (r, 0), ε̂(0) = 2, q̂ = 0.5,
n̂Li (0) = 1.25nLi , R̂ f (0) = 3R f . Moreover, zero mean normally
distributed noise with a standard deviation of 2 mV is added to
the voltage measurement. Figure 4 portrays the evolution of the
state and parameter estimates. The state estimates are represented
by the bulk SOC, defined in (57), and surface concentrations.
The surface concentration error includes the state estimation error
c− − − −
ss − ĉss and output function inversion error css − čss . The PDE
parameter estimates ε̂, q̂ and output function parameter estimates
n̂Li , R̂ f also converge near their true values. The results are sim-
8 CONCLUSION
This paper reports on the first combined SOC/SOH estimator
for electrochemical battery models. The adaptive observer uti-
lizes concepts from PDE estimation and adaptive control theory
to generate various new concepts for battery systems and control.
These are summarized by four key ideas: First, a backstepping
PDE state estimator is designed in previous work [20]. Second, a
Padé approximation of the transfer function for lithium diffusion
is used to identify the diffusion coefficient. Third, parameter
sensitivity analysis is applied to elucidate the linear dependence
between physically meaningful parameters related to capacity and
power fade. Fourth, an adaptive output function inversion tech-
nique enables linear state estimation designs. Finally, we present
simulations which demonstrate how the adaptive observer suc-
cessfully identifies the states and parameters from measurements
of voltage and current, only. The composition of these unique
ideas provides a combined SOC/SOH estimation algorithm for
battery systems using electrochemical models.
Our next steps include state/parameter estimation schemes
for the complete Doyle-Fuller-Newman model [5]. This work
will enable output feedback controllers for optimal charging and
discharging.
REFERENCES
[1] Chaturvedi, N. A., Klein, R., Christensen, J., Ahmed, J.,
and Kojic, A., 2010. “Algorithms for advanced battery-
management systems”. IEEE Control Systems Magazine,
30(3), pp. 49 – 68.
[2] Moura, S., Fathy, H., Callaway, D., and Stein, J., 2011. “A
Stochastic Optimal Control Approach for Power Manage-
ment in Plug-In Hybrid Electric Vehicles”. IEEE Transac-
tions on Control Systems Technology, 19(3), pp. 545 –555.
[3] Siegel, J. B., Lin, X., Stefanopoulou, A. G., Hussey, D. S.,
Jacobson, D. L., and Gorsich, D., 2011. “Neutron imaging
of lithium concentration in LFP Pouch cell battery”. Journal
of the Electrochemical Society, 158(5), pp. A523 – A529.
[4] Liu, P., Wang, J., Hicks-Garner, J., Sherman, E., Soukiazian,
S., Verbrugge, M., Tataria, H., Musser, J., and Finamore, P.,
2010. “Aging Mechanisms of LiFePO4 Batteries Deduced
by Electrochemical and Structural Analyses”. Journal of
the Electrochemical Society, 157(4), pp. A499–A507.
[5] Thomas, K., Newman, J., and Darling, R., 2002. Advances
in Lithium-Ion Batteries. Kluwer Academic/Plenum Publish-
ers, New York, NY USA, ch. 12: Mathematical modeling of
lithium batteries, pp. 345–392.
[6] Plett, G. L., 2004. “Extended Kalman filtering for battery
management systems of LiPB-based HEV battery packs.
Part 3. State and parameter estimation”. Journal of Power
Sources, 134(2), pp. 277–92.
[7] Verbrugge, M., and Tate, E., 2004. “Adaptive state of charge
algorithm for nickel metal hydride batteries including hys-
teresis phenomena”. Journal of Power Sources, 126(1-2),
pp. 236–249.
[8] Verbrugge, M., 2007. “Adaptive, multi-parameter battery
state estimator with optimized time-weighting factors”. Jour-
nal of Applied Electrochemistry, 37(5), pp. 605 – 616.
[9] Hu, Y., and Yurkovich, S., 2012. “Battery cell state-of-
charge estimation using linear parameter varying system
techniques”. Journal of Power Sources, 198, pp. 338 – 350.
[10] Santhanagopalan, S., and White, R. E., 2006. “Online esti-
mation of the state of charge of a lithium ion cell”. Journal
of Power Sources, 161(2), pp. 1346 – 1355.
[11] Domenico, D. D., Stefanopoulou, A., and Fiengo, G., 2010.
“Lithium-Ion Battery State of Charge and Critical Surface
Charge Estimation Using an Electrochemical Model-Based
Extended Kalman Filter”. Journal of Dynamic Systems,
Measurement, and Control, 132(6), p. 061302.
[12] Smith, K. A., Rahn, C. D., and Wang, C.-Y., 2008. “Model-
based electrochemical estimation of lithium-ion batteries”.
In 2008 IEEE International Conference on Control Applica-
tions, pp. 714–19.
[13] Klein, R., Chaturvedi, N. A., Christensen, J., Ahmed, J.,
Findeisen, R., and Kojic, A., 2012. “Electrochemical Model
Based Observer Design for a Lithium-Ion Battery”. IEEE
Transactions on Control Systems Technology, PP, pp. 1–13.
[14] Santhanagopalan, S., Guo, Q., Ramadass, P., and White,
R. E., 2006. “Review of models for predicting the cycling
performance of lithium ion batteries”. Journal of Power
Sources, 156(2), pp. 620 – 628.
[15] Chen, C., 1998. Linear System Theory and Design. Oxford
University Press, Inc.
[16] Krstic, M., and Smyshlyaev, A., 2008. Boundary Control
of PDEs: A Course on Backstepping Designs. Society for
Industrial and Applied Mathematics, Philadelphia, PA.
[17] Delacourt, C., Poizot, P., Levasseur, S., and Masquelier,
C., 2006. “Size effects on carbon-free LiFePO4 powders”.
Electrochemical and Solid-State Letters, 9(7), p. A352.
[18] Derrien, G., Hassoun, J., Panero, S., and Scrosati, B., 2007.
“Nanostructured Sn-C composite as an advanced anode ma-
terial in high-performance lithium-ion batteries”. Advanced
Materials, 19(17), p. 2336.
[19] Forman, J. C., Moura, S. J., Stein, J. L., and Fathy, H. K.,
2012. “Genetic identification and Fisher identifiability anal-
ysis of the Doyle-Fuller-Newman model from experimental
cycling of a LiFePO4 cells”. Journal of Power, 210, pp. 263–
275.
[20] Moura, S. J., Chaturvedi, N., and Krstic, M., 2012. “PDE
Estimation Techniques for Advanced Battery Management
Systems - Part I: SOC Estimation”. In Proceedings of the
2012 American Control Conference.
[21] Ioannou, P., and Sun, J., 1996. Robust Adaptive Control.
Prentice-Hall.
[22] Smyshlyaev, A., and Krstic, M., 2010. Adaptive Control of
Parabolic PDEs. Princeton University Press.
[23] Lund, B. F., and Foss, B. A., 2008. “Parameter ranking
by orthogonalization - Applied to nonlinear mechanistic
models”. Automatica, 44(1), pp. 278 – 281.