1 s2.0 S2352152X20319587 Main

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Journal of Energy Storage 34 (2021) 102133

Contents lists available at ScienceDirect

Journal of Energy Storage


journal homepage: www.elsevier.com/locate/est

Temperature estimation from current and voltage measurements in


lithium-ion battery systems
P. Wang a,c ,∗, L. Yang c , H. Wang c , D.M. Tartakovsky b , S. Onori b ,∗∗
a
School of Integrated Circuit Science and Engineering & Beijing Advanced Innovation Center for Big Data and Brain Computing, Beihang
University, 100191, China
b
Department of Energy Resources Engineering, Stanford University, CA 94305, USA
c
LMIB & School of Mathematical Sciences, Beihang University, 100191, China

ARTICLE INFO ABSTRACT

Keywords: Performance and safety of lithium-ion batteries depend on the ability to efficiently estimate their temperature
Lithium-ion battery during charge/discharge operations. We propose a novel algorithm to infer temperature in cylindrical lithium-
Temperature estimation ion battery cells from measurements of current and terminal voltage. Our approach employs a dual ensemble
Electrochemical modeling
Kalman filter, which incorporates the enhanced single-particle dynamics to relate terminal voltage to battery
Ensemble Kalman filter
temperature and Li-ion concentration. The numerical results and experimental validation from LGChem
LiNiMnCoO2 battery (INR21700 M50) cell data demonstrate the method’s ability to estimate temperature at
various charge/discharge C-rates.

1. Introduction To account for thermal phenomena in batteries, such electrochem-


ical models are supplemented with an energy conservation equation,
Lithium-ion battery (LIB) systems are widely used in modern elec- in which Li ion (Li+ ) concentration affects the rate of heat genera-
tronic devices. Their reliability and safety are of crucial importance tion. Depending on whether the well-mixed assumption is invoked,
for sustainable development and successful transition to carbon-neutral the resulting electrochemical–thermal models can be subdivided into
energy economy. However, safety concerns, mainly due to thermal lumped (averaged in space) [1,8] and spatially distributed [9–11]. Such
runaway, hinder large-scale LIB utilization for these purposes. Limit- electrochemical–thermal models have been used for state of charge and
ing temperature of LIBs to a restricted window of operation is also temperature estimation in battery management systems. The experi-
important for their longevity and satisfactory performance [1]. Prac- mental validation of thermal models aimed at predicting core tempera-
tical limitations preclude direct measurement of temperature of LIB ture is challenging and can pose safety hazards, as it requires a battery
cells within a pack, and accurate temperature estimation from readily to be drilled in order to insert a thermocouple into the cell, in the case
available sensors (e.g., current and terminal voltage) remains elusive. of cylindrical cells.
Electrochemical models of LIB systems rely on a number of sim-
These practical considerations make the ability to estimate the
plifying assumptions whose veracity is hard to ascertain a priori. For
core temperature from existing sensor layouts an appealing objective,
example, the single-particle model (SPM) [2] ignores concentration
which we tackle in this study. Prediction of a battery model, and its
gradients in the electrolyte (a well-mixed assumption) and approx-
use for indirect measurements of battery temperature, are inherently
imates the solid phases in both electrodes by spherical particles; it
uncertain due to a large number of unknown (fitting) parameters,
fails to accurately model the terminal voltage at high charge/discharge
ambient operating conditions, and sparse sensing [12]. This uncertainty
C-rates. The enhanced SPM model (eSPM) accounts for electrolyte
is typically quantified probabilistically by treating uncertain model
dynamics, yielding improved performance at high C-rates [3,4]. More
complex models, e.g., [5–7], incorporate additional physics, but are inputs and outputs as random variables and random processes. The
computationally prohibitive when used either for sampling-based pa- probabilistic framework replaces the inference of a unique parameter
rameter/state estimation or in micro-controllers of battery management value and a unique model prediction with identification of their (joint)
systems. probability density functions [13].

∗ Correspondence to: Beihang University, 37 Xueyuan Road, Beijing, 100191, China.


∗∗ Corresponding author.
E-mail addresses: wang.peng@buaa.edu.cn (P. Wang), sonori@stanford.edu (S. Onori).

https://doi.org/10.1016/j.est.2020.102133
Received 24 September 2020; Received in revised form 24 November 2020; Accepted 25 November 2020
Available online 15 December 2020
2352-152X/© 2020 Elsevier Ltd. All rights reserved.
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

the eSPM. Section 3 contains both a description of our general method-


Nomenclature
ology for indirect measurements of the temperature of LIBs and the
𝑗 Index, 𝑗 = 𝑛 for anode, 𝑗 = 𝑝 for cathode numerical algorithm for its implementation. In Section 4, we validate
𝐴 Cell cross-sectional area [m2 ] our methodology by using experimental data for LGChem LiNiMnCoO2
max
𝑐s,j Maximum electrode concentration battery (INR21700 M50). The key conclusions drawn from this study
are collated in Section 5.
[mol/m3 ]
𝐷e Electrolyte phase diffusion [m2 /s]
in 2. Electrochemical model
𝐶e,𝑖 Initial concentration in electrolyte
[mol/m3 ]
ref LIB cells consist of five major parts: positive electrode (cathode),
𝐷s,j Reference solid phase diffusion [m2 /s]
negative electrode (anode), electrolyte, separator and current collectors
act
𝐸𝐷,j Activation energy for solid phase diffusion (Fig. 1). The two electrodes and the separator have a thickness of
in electrode 𝑗 𝐿p , 𝐿n , and 𝐿S , respectively. The cathode is typically a metal ox-
act
𝐸𝑘,j Activation energy for reaction in electrode ide (LiMO2 ), whereas low-cost graphite (C6 ) is widely used as anode
𝑗 material due to low electrode potential [15].
𝐹 Faraday’s constant [C/mol] The electrolyte is a salt, generally hexafluorophosphate (LiPF6 ),
𝑘ref
j
Electrode reaction rate constant [m2.5 /s dissolved in an organic solvent that acts as a transport medium allowing
mol0.5 ] migration of Li+ between the two electrodes. The electrochemical
𝐿j Electrode thickness [m] reactions in a Li-ion cell involve Li+ moving in and out of the electrodes
𝐿s Separator thickness [m] via intercalation (insertion) and deintercalation (extraction) process,
𝑅 Lumped resistance [Ω] while electrons travel through an external load producing work.
𝑅g Universal gas constant [J/mol K] During discharge, an oxidation reaction at the anode produces
positively charged Li ions (Li+ ) and negatively charged electrons (e− ).
𝑅j Electrode particle radius [m]
This process is described by the anodic half-reaction,
𝑡+
0
Transference number
𝑇ref Reference temperature [◦ C] Li𝑦 C6 ⟶ 𝑦Li+ + 𝑦e− + C6 . (1a)
𝛼cell Electrode transfer coefficient
The Li+ ions are subsequently transported through the electrolyte and
𝜀j Electrode active volume fraction of solid
the separator towards the cathode. The electrons pass through the
phase
external circuit before reconnecting to the cathode, which undergoes
𝜃j,100% Reference stoichiometry ratio at 100% SOC reduction, while the migrated Li+ ions are intercalated into the cathode.
𝜃j,0% Reference stoichiometry ratio at 0% SOC This phenomenon is described by the cathodic half-reaction,
𝜔𝑖 Porosity of anode, cathode, and separator
(𝑖 = n, p, S) Li(1−𝑦) MO2 + 𝑦Li+ + 𝑦e− ⟶ LiMO2 . (1b)

During charging, the cathode undergoes oxidation and the anode un-
dergoes reduction [16].

When combined with sensor measurements, the latter task can


2.1. Enhanced single-particle model for Li+ concentration
be accomplished via sampling-based approaches to data assimilation,
such as extended Kalman filter [10]. The reliance on such generaliza-
We use the eSPM from [4] as a modeling platform to develop the
tions of Kalman filter is necessary because the relationship between
observer. The eSPM describes diffusion of Li+ in the cathode, anode
temperature and the measurable quantities is highly nonlinear in the
and electrolyte of a LIB. The eSPM represents the solid material in
battery models used and, hence, temperature is highly non-Gaussian.
each electrode as a spherical particle subjected to a uniform electrical
Alternatively, one could sacrifice model veracity by assuming a linear
current. Inside of each electrode particle with radius 𝑅𝑗 (𝑗 = n and p),
relationship between temperature and the observables in order to use
the Li+ concentration, 𝑐s,𝑗 (𝑟, 𝑡) with 𝑗 = n and p, obeys Fick’s law of
standard Kalman filter [14], but this – the use of an oversimplified
diffusion [3],
model – might lead to erroneous predictions. [ ]
𝜕𝑐s,𝑗 𝐷s,𝑗 𝜕 2 𝜕𝑐s,𝑗
Technical contribution and novelty. Temperature has a significant im- = 𝑟 , 0 < 𝑟 < 𝑅𝑗 , 𝑡 > 0, 𝑗 = n, p; (2)
𝜕𝑡 𝑟2 𝜕𝑟 𝜕𝑟
pact on the performance, safety and cycle lifetime of LIBs. Accurate
and boundary conditions at the sphere’s center (𝑟 = 0) and surface
estimation of available charge and energy during short- and long-
(𝑟 = 𝑅𝑗 ),
term operation as well as detection of safety-hindering events are
possible only through continuous battery temperature monitoring. Our 𝜕𝑐s,𝑗 𝜕𝑐s,𝑗 𝐽𝑗 𝐼
(0, 𝑡) = 0, −𝐷s,𝑗 (𝑅𝑗 , 𝑡) = = . (3)
contribution to the state-of-the-art in estimation of the temperature of 𝜕𝑟 𝜕𝑟 𝐹 𝑎𝑗 𝐴𝐿𝑗 𝐹
LIBs is fourfold. First, our approach utilizes readily available on-board
Here, 𝑡 [s] denotes time, 𝑟 [m] is the radial coordinate, 𝐹 is the Faraday
measurements of terminal voltage and current to estimate the temper-
constant, and 𝐽𝑗 [A m−2 ] is the intercalation current density at each
ature. Second, our reliance on a relatively high-fidelity electrochemical
electrode. The latter is defined as the input current 𝐼 [A] divided by
model, based on the experimentally validated eSPM, enables us to
the specific interfacial surface area 𝑎𝑗 [m−1 ], the cross-sectional area
estimate the battery cell temperature at a wide range of C-rates. Third,
𝐴 [m2 ] and the electrode thickness 𝐿𝑗 [m]. In our case, 𝑎𝑗 = 3𝜀𝑗 ∕𝑅𝑗 ,
we rely on dual ensemble Kalman filter to assimilate measurements where 𝜀𝑗 is the active volume fraction of solid phase [16, Section 3.3.2].
of terminal voltage into this complex nonlinear model. Finally, we The solid-phase diffusion coefficients 𝐷s,𝑗 [m2 s−1 ] and reaction rate
provide validation of our model predictions by using experimental constants 𝑘s,𝑗 for each electrode have an Arrhenius-type dependence
battery data collected at a different temperature than that used for
on temperature 𝑇 [17],
model calibration. [ act (
𝐸𝐷,𝑗 )]
The remainder of this paper is organized as follows. In Section 2, ref 1 1
𝐷s,𝑗 = 𝐷s,𝑗 exp − , (4a)
we formulate the problem and its mathematical conceptualization via 𝑅g 𝑇ref 𝑇

2
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

Fig. 1. A schematic diagram of the macroscopic (𝑙-direction) cell model with coupled microscopic (𝑟-direction) solid-diffusion model.

[ act ( )]
𝐸𝑘,𝑗 1 1 where 𝛥𝛷e ≡ 𝛷e (𝑥 = 𝐿) − 𝛷e (𝑥 = 0) is the potential difference
𝑘s,𝑗 = 𝑘ref
s,𝑗 exp − , (4b)
𝑅g 𝑇ref 𝑇 in the electrolyte phase, 𝛥𝑈oc is the open circuit potential difference
ref , 𝑘ref and 𝑇 (positive minus negative) between the electrodes, and 𝛥𝜂 is the over-
where 𝐷s,𝑗 s,𝑗 ref denote the reference values for the diffusion
act potential difference between the electrodes, and 𝐼𝑅 is the voltage
coefficients, reaction rate constants and temperature, respectively; 𝐸𝐷,𝑗
act are the activation energy; and 𝑅 is the universal gas constant.
and 𝐸𝑘,𝑗 drop caused by the lumped resistance. An approximate (linearized)
g
The Li+ concentration in the electrolyte phase, 𝑐e,𝑖 (𝑥, 𝑡), is governed analytical solution of (5) gives an expression for the potential difference
by the Planck–Nernst equations, in the electrolyte [18],
( ) ( )
𝜕𝑐e,𝑖 𝜕 𝜕𝑐e,𝑖 𝑎𝐽 𝑐e,p (𝐿, 𝑡) 𝐿n 2𝐿s 𝐿p
𝜔𝑖 = ef f
𝐷e,𝑖 + (1 − 𝑡+ ) 𝑖 𝑖,
0 𝐹
𝑖 = n, p, S, (5) ef f
𝛥𝛷e = 𝜅̃ ln −
𝐼
+ + . (9a)
𝜕𝑡 𝜕𝑥 𝜕𝑥 𝑐e,n (0, 𝑡) 2𝐴 𝜅nef f 𝜅sef f 𝜅pef f
where 𝑥 is the Cartesian coordinate along the thickness, 𝐿, of the cell;
𝜔𝑖 is porosity of the negative (𝑖 = n) and positive (𝑖 = p) electrodes Here, the effective ionic conductivities 𝜅𝑖ef f are estimated empiri-
and the separator (𝑖 = S); 𝑡+
0
is the transference number; the exchange cally [17],
current density at the separator is 𝐽S (𝑡) = 0; and the effective diffusion
ef f are assumed to follow the concentration-dependent
coefficients 𝐷e,𝑖 𝜅𝑖ef f = 𝑐 e,𝑖 [(−10.5 + 0.074 𝑇 − 6.96 × 10−5 𝑇 2 ) + 𝑐 e,𝑖 (0.668 − 0.0178 𝑇
Bruggeman relation [3], −2.8 × 10−5 𝑇 2 ) + 𝑐 2e,𝑖 (0.494 − 8.86 × 10−4 𝑇 )]2 , (9b)
ef f
𝐷e,𝑖 ≈ (𝛼1 𝑐e,𝑖 +𝛼2 )𝜔1.5 with 𝛼1 = −1.17×10 −13
, 𝛼2 = 3.97×10 −10
. (6)
𝑖 as functions of temperature 𝑇 and the average Li+ concentrations at the
Eq. (5) are defined on the segments 𝑥 ∈ (0, 𝐿n ), (𝐿n , 𝐿n + 𝐿S ) and anode, cathode or separator 𝑐 e,𝑖 ; and
(𝐿n + 𝐿S , 𝐿 = 𝐿n + 𝐿S + 𝐿p ) representing, the anode, the separator
and the cathode, respectively. The external (impermeable) boundary 2𝑅g 𝑇 (1 − 𝑡+
0
)(𝜅nef f + 𝜅sef f + 𝜅pef f ) { √
𝜅̃ ef f = 0.601 − 0.24 𝑐 e
conditions, 3𝐹
}
𝜕𝑐e,n 𝜕𝑐e,p + 0.982 [1 − 0.0052(𝑇 − 293)] 𝑐 1.5 . (9c)
(0, 𝑡) = (𝐿, 𝑡) = 0, (7a) e
𝜕𝑥 𝜕𝑥
are supplemented with continuity conditions at the anode–separator The open circuit potentials at the two electrodes, 𝑈p and 𝑈n , are
interface, 𝑥 = 𝐿n , functions of temperature, 𝑇 , and stoichiometric ratios, 𝜃p and 𝜃n , at the
𝜕𝑐e,n 𝜕𝑐e,S solid phase [2],
ef f ef f
𝑐e,n (𝐿n , 𝑡) = 𝑐e,S (𝐿n , 𝑡), (𝐿n , 𝑡) = 𝐷e,S
𝐷e,n (𝐿n , 𝑡), (7b)
𝜕𝑥 𝜕𝑥 max
n
𝛥𝑈oc (𝑇 ) = 𝑈p (𝜃p , 𝑇 ) − 𝑈n (𝜃n , 𝑇 ), 𝜃𝑗 = 𝑐s,𝑗 (𝑅𝑗 , 𝑡)∕𝑐s,𝑗 , (10)
and at the separator–cathode interface, 𝑥 = 𝐿S ≡ 𝐿n + 𝐿S ,
max is the theoretical maximum concentration. The difference
where 𝑐s,𝑗
𝜕𝑐e,S 𝜕𝑐e,p
𝑐e,S (𝐿nS , 𝑡) = 𝑐e,p (𝐿nS , 𝑡), ef f
𝐷e,S (𝐿nS ) = 𝐷e,p
ef f
(𝐿nS ). (7c) between the cathode and anode over-potentials is computed from
𝜕𝑥 𝜕𝑥
max ×
Eqs. (2) and (5) are also subject to initial conditions 𝑐s,𝑗 (𝑟, 0) = 𝑐s,𝑗 the Butler–Volmer kinetic equation [19], which describes the rate of
in
𝜃𝑗,100% with 𝑗 = n, p and 𝑐e,𝑖 (𝑥, 0) = 𝐶e,𝑖 with 𝑖 = n, p, S. Here, 𝜃j,100% intercalation and deintercalation of Li+ ,
is the electrode stoichiometry ratio at 100% state of charge (SOC) and [ ( ) ( )]
2𝑅g 𝑇 𝐽p 𝐽n
in is the initial concentration of Li+ in the electrolyte.
𝐶e,𝑖 𝛥𝜂 = 𝜂p − 𝜂n = sinh−1 − sinh−1 , (11a)
𝐹 2𝐽0,p 2𝐽0,n

2.2. Terminal voltage where



𝐽0,𝑗 = 𝐹 𝑘s,𝑗 surf (𝑐 max − 𝑐 surf ),
𝑐 e 𝑐s,𝑗 𝑗 = n, p (11b)
The terminal voltage, 𝑉 , is defined as [2] s,𝑗 s,𝑗

𝑉 = 𝛥𝛷e + 𝛥𝑈oc + 𝛥𝜂 − 𝐼𝑅, (8) is the exchange current density.

3
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

3. Methodology for core temperature estimation

We start by describing a numerical scheme used to solve the eSPM


(Section 3.1) and then discuss its observability from measurements
(Section 3.2). Our approach to data assimilation and a detailed algo-
rithm for its implementation are outlined in Section 3.3.

3.1. Numerical solution of the eSPM

A finite-difference method is employed to approximate the spatial


derivatives in the eSPM [4]. The simulation domain 0 ≤ 𝑟 ≤ 𝑅𝑗 is
discretized with 𝑀s + 1 nodes, and the boundary-value problem for the
Li+ concentration in the solid phase, (2) and (3), is discretized in space
as
𝑚
d𝑐s,𝑗 𝐷s,𝑗 𝑀s2 (𝑚 − 1 𝑚−1 𝑚 𝑚 + 1 𝑚+1
)
= 𝑐s,𝑗 − 2𝑐s,𝑗 + 𝑐s,𝑗 , 𝑚 = 1, … , 𝑀 − 1;
d𝑡 𝑅 2 𝑚 𝑚
𝑗 Fig. 2. Experimental data showing the temperature behavior from 1-C rate discharge
for the LGChem INR21700 battery cell.
and
0 1 𝑀 𝑀 −1 𝑅𝑗 𝐼
𝑐s,𝑗 = 𝑐s,𝑗 and 𝑐s,𝑗 s = 𝑐s,𝑗 s − .
𝐷s,𝑗 𝐹 𝑀s 𝑎𝑗 𝐴𝐿𝑗 Consider a differentiable manifold 𝑀 and let 𝑋 be a differentiable
For the electrolyte phase, the simulation domains 0 ≤ 𝑥 ≤ 𝐿n , vector field on 𝑀. Given a function 𝑓 ∶ 𝑀 → R and a set 𝑥 ∈ 𝑀, the
𝐿n < 𝑥 ≤ 𝐿n +𝐿S and 𝐿n +𝐿S < 𝑥 ≤ 𝐿 are discretized, respectively, with Lie derivative is defined as 𝐿𝑋 [𝑓 ] = 𝜕𝑥 𝑓 (𝑥)𝑋(𝑥). With this definition,
𝑀e,n , 𝑀e,S and 𝑀e,p nodes. To simplify the presentation, we assume the first 𝑙 = 2𝑀𝑠 + 𝑀𝑒 + 4 − 1 Lie derivatives of 𝑔m with respect to 𝐅 are
these nodes to be equally distant for all of these subdomains, such that
the distance between any two adjacent nodes in the whole simulation ⎧ 𝐿0𝐅 [𝑔m ] ≡ 𝑔m
domain 0 ≤ 𝑥 ≤ 𝐿 is 𝛥𝑥 = 𝐿∕𝑀e , where 𝑀e = 𝑀e,n + 𝑀e,S + 𝑀e,p . Then ⎪
⎪ 𝐿1𝐅 [𝑔m ] ≡ ∇𝑔m ⋅ 𝐅
a finite-difference approximation of (5) on the domain 0 ≤ 𝑥 ≤ 𝐿 is ⎪
⎨ 𝐿2𝐅 [𝑔m ] ≡ ∇(𝐿1𝐅 [𝑔m ]) ⋅ 𝐅 (14)
d𝑐e𝑘 (𝑐e𝑘−1 )2 − 𝑐e𝑘−1 𝑐e𝑘 − (𝑐e𝑘 )2 + 𝑐e𝑘 𝑐e𝑘+1 ⎪
𝜔𝑖 = 𝛼1 ⎪ ⋮
d𝑡 (𝐿∕𝑀e )2 ⎪ [ ]
⎩ 𝐿𝑙𝐅 [𝑔m ] ≡ 𝐿𝐅 𝐿𝑙−1
𝐅
[𝑔m ]
𝑐e𝑘−1 − 2𝑐e𝑘 + 𝑐e𝑘+1 𝑎𝑖 𝐽 𝑖
+ 𝛼2 + (1 − 𝑡+
0
) , 𝑘 = 1, … , 𝑀e − 1; where
(𝐿∕𝑀e )2 𝐹
( )
with the boundary and interfacial conditions (7a)–(7c) taking the form 𝜕 𝜕 𝜕
∇≡ , …, 𝑀 , .
0
𝑀e 𝑀e −1 𝜕𝑐s,n 𝜕𝑐e e 𝜕𝑇
𝑐e0 = 𝑐e1 , 𝑐e = 𝑐e ,
𝑀 𝑀 −1 𝑀 +1 𝑀 The local observability of this system, at a point (𝐰∗ , 𝐼 ∗ ), depends on
𝑐e e,n − 𝑐e e,p 𝑐e e,n − 𝑐e e,n
𝜔1.5
n = 𝜔1.5
S
, the rank of the matrix [20]
𝐿∕𝑀e 𝐿∕𝑀e
0
𝑀e,p +𝑀e,s 𝑀e,p +𝑀e,s −1 𝑀e,p +𝑀e,s +1 𝑀 +𝑀 ⎡ ∇(𝐿𝐹 [𝑔m ]) ⎤
𝑐e − 𝑐e 1.5 𝑐e − 𝑐e e,p e,s ⎢ ⎥
𝜔1.5 = 𝜔 . ∇(𝐿1𝐹 [𝑔m ])
S 𝐿∕𝑀e p 𝐿∕𝑀e 𝐎(𝐰∗ , 𝐼 ∗ ) = ⎢ ⎥ . (15)
⎢ ⋮ ⎥
⎢ ⎥
3.2. Observability of Li+ concentration from terminal voltage ⎣ ∇(𝐿𝑙−1 [𝑔m ]) ⎦
𝐹 𝐰=𝐰∗ ,𝐼=𝐼 ∗

Specifically, the system is locally observable if 𝚛𝚊𝚗𝚔(𝐎)|column = 2𝑀𝑠 +


Our goal is to estimate the battery temperature, 𝑇 (𝑡), from the 𝑀𝑒 + 4. We verified numerically that this condition holds at different
terminal output voltage 𝑉 (𝑡) and input current 𝐼(𝑡). Let us define the operating points (𝐰∗ , 𝐼 ∗ ) with 𝐼 ∗ ≠ 0. Therefore, for any unknown
0 , … , 𝑐 𝑀s , 𝑐 0 , … , 𝑐 𝑀s , 𝑐 0 , … , 𝑐 𝑀e )⊤ ∈ R2𝑀s +𝑀e +3
state vector 𝐜 = (𝑐s,n s,n s,p s,p e e initial state, there exists a finite time-step 𝑡𝑘 > 0, for which the
+
comprising the Li concentration in the anode, cathode and electrolyte. knowledge of the input current 𝐼 and the output terminal voltage 𝑉
Then, the eSPM and terminal voltage is rewritten as over the time period [0, 𝑡𝑘 ] suffices to uniquely determine the initial
d𝐜 state.
= 𝑓m [𝐜, 𝑇 , 𝐼], 𝑉 = 𝑔m [𝐜, 𝑇 , 𝐼], (12)
d𝑡
where 𝑓m [⋅] is the eSPM (forward) model that predicts the temporal 3.3. Dual ensemble Kalman filter
evolution of the system state 𝐜(𝑡), and 𝑔m [⋅] is the nonlinear oper-
ator that calculates 𝑉 from the concentrations and temperature in Various flavors of Kalman filter, e.g., a joint filtering approach [21]
accordance with (8). or a dual filtering structure [22], can be used to estimate the temper-
System observability refers to the ability to estimate 𝐜 and 𝑇 from ature 𝑇 . Many of them require the knowledge of the state-parameter
measurements of the input 𝐼 and output 𝑉 . Since the eSPM is highly matrix, which carries a significant computational cost proportional to
nonlinear, we investigate local observability by assuming that 𝑇 (𝑡) the dimensionality of the system. Instead, we deploy the dual ensemble
varies slightly within a small time window, such that d𝑇 ∕d𝑡 = 0 locally. Kalman filter [23] because of its efficiency and ease of implementation.
This assumption is justified by our experimental observations depicted A key feature of the proposed dual ensemble Kalman filter is the
in Fig. 2. addition of a new equation to describe the quasi-constant change in
Consider the vector 𝐰 = (𝐜⊤ , 𝑇 )⊤ ∈ R2𝑀𝑠 +𝑀𝑒 +4 , then, the eSPM in temperature over the sampling time window between the 𝑘th and
(12) is reformulated as (𝑘 + 1)st time steps,
d𝐰
= 𝐅[𝐰, 𝐼]. (13) 𝑇𝑘+1 = 𝑇𝑘 + 𝜖𝑘𝑇 . (16)
d𝑡

4
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

Fig. 3. Experiments were carried out in the Stanford Energy Control laboratory at the Energy Resources Engineering Department, Stanford University.

Here 𝑇𝑘 and 𝑇𝑘+1 are the temperature values at the 𝑘th and (𝑘 + 1)st
time steps, respectively; and 𝜖𝑘𝑇 ∈ R is a small quantity capturing the Algorithm 1 Dual state-parameter estimation with ensemble Kalman
temporal change of the temperature over the time window 𝑘 and 𝑘 + 1. filter
We treat 𝜖𝑘𝑇 as a Gaussian random variable with zero mean and variance (i) Initialization, 𝑘 = 0
𝜎𝑇2 ∈ R. By coupling (16) with (12), we obtain a new system for the thus
randomized state vector 𝐜. • Draw 𝑁 samples for the initial states of Li+ concentration
Subsequent assimilation of measurements of current 𝐼 and terminal and temperature, 𝐜(𝑛)
0
, 𝑇0(𝑛) , (𝑛 = 1, … , 𝑁), which are taken
voltage 𝑉 yields an estimate of battery temperature 𝑇 . Direct mea- as uniformly distributed random variables:
surements of 𝑉 can be noisy due to errors associated with numerical
𝐜0 ∼ U(𝐜min , 𝐜max ), 𝑇0 ∼ U(𝑇min , 𝑇max ).
roundoff and operating conditions. To account for this observation
error, we introduce a zero-mean Gaussian process in time, 𝜖𝑘d ∈ R with (ii) For 𝑘 ≥ 1 and for all samples (𝑛 = 1, … , 𝑁),
variance 𝜎d2 ∈ R. Thus, the observed data are modeled as 𝑑𝑘 = 𝑉𝑘obs +𝜖𝑘d .
With the model prediction, 𝐰(𝑡) = (𝐜⊤ , 𝑇 )⊤ , and observations, 𝑑(𝑡), (a) Forward prediction of system states:
in place, we formulate the state-parameter estimation problem as a
filtering problem. Probability of the discrete system state at an arbitrary • Predict temperature and terminal voltage from (12)
time step 𝐾 ≥ 1 is given by Bayes’ theorem, and (16):
(𝑛)
P(𝐰𝐾 |D𝐾 ) ∝ P(𝐰𝐾−1 |D𝐾−1 )P(𝑇𝐾 |𝑇𝐾−1 )P(𝐜𝐾 |𝐰𝐾−1 )P(𝑑𝐾 |𝐰𝐾 ), (17) 𝑇̂𝑘+1 = 𝑇𝑘(𝑛) + 𝜖𝑘+1
T (𝑛)
,

where D𝐾 = {𝑑𝑘 }𝐾 is the data set of terminal voltage measurements 𝐜̂(𝑛)


𝑘+1
= 𝑓m (𝐜(𝑛)
𝑘
(𝑛)
, 𝑇̂𝑘+1 , 𝐼𝑘+1 ),
𝑘=1
collected up to and including the 𝐾th time step. Filtering replaces (𝑛)
𝑉̂𝑘+1 = 𝑔m (𝐜̂(𝑛) , 𝑇̂ (𝑛) , 𝐼 ).
𝑘+1 𝑘+1 𝑘+1
this computationally expensive Bayesian update with an optimization
problem • Perturb the measurements of terminal voltage with
{[ ]2 } data noise:
𝑑𝐾 − 𝑔m (𝐰̄ 𝐾 , 𝐼𝐾 ) (𝐜̄𝐾 − 𝐜̂𝐾 )2 (𝑇̄𝐾 − 𝑇̂𝐾 )2
𝐰̄ ⋆
𝐾 = argmin + + , (18) (𝑛) obs d (𝑛)
𝐰̄ 𝐾 𝜎2 d
𝛴̂ 𝐾 𝐜 𝛴̂ 𝐾 𝑇
𝑑𝑘+1 = 𝑉𝑘+1 + 𝜖𝑘+1 .

where 𝐰̄ 𝐾 = (𝐜̄⊤ , 𝑇̄ )⊤ are the posterior means of the system states at


𝐾 𝐾 • Compute Kalman gain for temperature, 𝐊tem , and
𝑘+1
the 𝐾th time step; and 𝐰̂ 𝐾 = (𝐜̂⊤ , 𝑇̂ )⊤ and 𝛴̂ 𝐜𝐾 , 𝛴̂ 𝑇𝐾 are the prior means
𝐾 𝐾 update temperature:
and (co)variances predicted by the eSPM (13) at the 𝐾th time step. A
pseudo code of the dual ensemble Kalman filter is shown in Algorithm 𝐊tem
𝑘+1
= Cov(𝑇̂𝑘+1 , 𝑉̂𝑘+1 )[Cov(𝑉̂𝑘+1 , 𝑉̂𝑘+1 ) + 𝜎d2 ]−1 ,
1. (𝑛) (𝑛) (𝑛) (𝑛)
𝑇𝑘+1 = 𝑇̂𝑘+1 + (𝑑𝑘+1 − 𝑉̂𝑘+1 )𝐊tem
𝑘+1
.
3.4. Identification of system parameters
(b) Update terminal voltage and concentration states:

The eSPM presented in Section 2 incorporates a large number of • Update terminal voltage with temperature
parameters, whose values depend on both the battery type and factors
such as manufacturing process, duration of past discharge/charge, and 𝐜̂(𝑛)
𝑘+1
= 𝑓m (𝐜(𝑛)
𝑘
(𝑛)
, 𝑇𝑘+1 , 𝐼𝑘+1 ),
working environment. The parameters that have the largest impact on (𝑛)
𝑉𝑘+1 = 𝑔m (𝐜̂(𝑛) , 𝑇 (𝑛) , 𝐼 ).
𝑘+1 𝑘+1 𝑘+1
the overall model output are arranged in the vector
ref
𝛩 = (𝐷s,p ref
, 𝐷s,n , 𝜃n,100% , 𝜃p,100% , 𝐴, 𝑅). (19) • Compute Kalman gain matrix for concentration,
𝐊con
𝑘+1
, and update concentration:
These parameters are identified by minimizing the root mean square
error between the model-predicted terminal voltage 𝑉 (𝑡𝑘 ) and 𝑁obs 𝐊con
𝑘+1
= Cov(𝐜̂𝑘+1 , 𝑉𝑘+1 )[Cov(𝑉𝑘+1 , 𝑉𝑘+1 ) + 𝜎d2 ]−1 ,
observations 𝑉obs (𝑡𝑘 ),
𝐜(𝑛)
𝑘+1
= 𝐜̂(𝑛)
𝑘+1
(𝑛)
+ (𝑑𝑘+1 (𝑛)
− 𝑉𝑘+1 )𝐊con
𝑘+1
.
𝑁obs

𝛩⋆ = argmin |𝑉 (𝑡 ) − 𝑉 (𝐜 , 𝑇 , 𝛩, 𝑡 )|2 . (20)
𝛩∈R6
| obs 𝑘 𝑘 𝑘 𝑘 |
𝑘=1

5
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

Fig. 4. Left: 2 C-rate discharge profile from measurements (red line) and predicted terminal voltage using identified model parameters (blue dashed line). Right: The discrepancy
between the measurement and predicted terminal voltage.

A set of constraints for this minimization problem comes from Table 1


Values of the model parameters. The horizontal line separates the fixed parameters
geometric and physical considerations. The most obvious of these, 𝐴 >
from those identified via model calibration at 2 C-rate discharge.
0, 𝑅 > 0 and 0 < 𝜃𝑗,100% < 1, are supplemented with a constraint
Parameters Anode Cathode Separator
imposed by the long-time behavior of the Li+ concentrations in the
𝐿𝑗 3.655 × 10−5 4 × 10−5 2.5 × 10−5
cathode and anode, 𝑐s,𝑗 (𝑟, 𝑡) (𝑗 = n, p). Specifically, according to (2),
𝐷e 1.5 × 10−10 1.5 × 10−10 1.5 × 10−10
the diffusion time-scale (i.e., the time it takes to reach the equilibrium) 𝐶e,𝑖in
1200 1200 1200
in each of the spherical particles is 𝑡′ = 𝑅2𝑗 ∕(𝜋𝐷s,𝑗 ). Since the time (in 𝜔𝑖 0.4 0.4 0.4
seconds) it takes to discharge a battery at 1 C-rate is 𝑡′ ∈ [1, 3600], this 𝑐̄e,𝑖 1200 1200 1200
places a constraint on the solid-phase diffusion coefficient at the two 𝛼cell 0.5 0.5
𝑡+0 0.363 0.363
electrodes, 𝐷s,𝑗 ≤ 𝑅2𝑗 ∕(𝜋 𝑡′ ) (𝑗=n, p). max
𝑐s,j 31080 51830
In summary, the identification of system parameters in (19) is cast 𝑅𝑗 5 × 10−6 5 × 10−6
as a quadratic optimization problem (20) subject to the constraints 𝜀𝑗 0.662 0.58
𝑘ref 1 × 10−10 1 × 10−10
𝑅2𝑗 𝑗
𝐸𝐷,𝑗act
4 × 104 2.5 × 104
𝐴 > 0, 𝑅 > 0, 0 < 𝜃𝑗,100% < 1, 0 < 𝐷s,𝑗 ≤ for 𝑗 = n, p. (21)
𝜋𝑡′ 𝐸𝑘,𝑗act
3 × 104 3 × 104
𝑅g 8.314
4. Experimental setup, results and discussion 𝐹 96487
𝑇ref 25
ref
The experimental setup used for this work and shown in Fig. 3 𝐷s,j 3.825 × 10−14 3.196 × 10−14
includes two Arbin battery testing systems capable of applying diverse 𝜃𝑗,100% 0.836 0.375
𝐴 2.836 × 10−1
current profiles to cells and modules and measure the load current, cell
𝑅 4.967 × 10−2
voltage and surface temperature using highly accurate sensors. A ther-
mal chamber is used to conduct temperature controlled experiments
and Omega Type-T thermocouples are used to measure the battery
surface temperature. The measured current and voltage data from the parameter are reported in Table 1; the fixed parameters are taken
Arbin are used to identify the eSPM parameters. from [3,24].
Constant current discharge experiments are conducted on LGChem We used the dual ensemble Kalman filter with 20 samples to es-
INR21700 M50 cylindrical LIB cells whose chemical composition is timate the LIB temperature, 𝑇 , from 3408, 1648 and 1027 direct mea-
lithium nickel manganese cobalt oxide (LiNiMnCoO2 or NMC) for the surements of 𝑉 for 1, 2 and 3 C-rate discharge, respectively. The
cathode and mixed graphite and silicon at the anode. The atomic ratio measurement error variance is 𝜎d2 = 1 × 10−6 V2 , and the variance of
of NMC used in LGChem INR21700 M50 is Ni:Mn:Co:Al = the temperature model error is 𝜎𝑇2 = 0.001 ◦ C2 . In Fig. 5, the temporal
82.9:5.1:10.6:1.4, and the anode has 90% graphite and 10% SiO0.64 evolutions of temperature 𝑇 , reconstructed from direct measurements
in weight percentage [24]. The LIB has a nominal capacity of 4.85 A⋅ h and estimated via the proposed ensemble Kalman filter at 1, 2 and
and operates at a nominal voltage 3.63 V with maximum and minimum 3 C-rate discharge are shown. In all regimes, the algorithm yields
voltage of 4.20 V and 2.50 V, respectively. The cell was subjected to averaged (smoothed) temperature profiles that are in close agreement
three input discharged current at 1 C-rate, 2 C-rate and 3 C-rate at a with the observations. Moreover, in Fig. 5 the prediction errors, i.e., the
temperature of 25 ◦ C. We measured the terminal voltage and the cell difference between the observed and predicted values of voltage 𝑉 are
surface temperature. shown for the same C-rates. The maximum error at 2 C-rate discharge
The finite-difference method described in Section 3.1 is imple- reaches 2 ◦ C, whereas its counterparts at the two other rates are merely
mented on 𝑀s = 𝑀e = 31 nodes. The system of ordinary differential 0.6 ◦ C. The temperature root mean square error (RMSE) at 1, 2 and 3
Eqs. (12) is solved with the explicit fourth-order Runge–Kutta method C-rate discharge is 0.1085, 0.2550 and 0.3224, respectively.
using the MATLAB® solver ode45. The vector of model parameters 𝛩 The results shown demonstrate that the proposed method can ef-
in (19) is obtained as a solution of the constrained optimization prob- fectively forecast the LIB temperature at a constant current. Moreover,
lem (Section 3.4) using the MATLAB® prediction error minimization it is also applicable to high-rate currents, as discussed in Section 3.3
toolbox pem. Fig. 4 shows a close agreement between the terminal wherein the current varies with time. Application of our method to
voltage 𝑉 (𝑡) predicted with the model upon parameter identification discharge/charge cycles with varying ambient temperature is left for
and its observations at the 2 C-rate discharge. The value of identified future studies. For example, a working cycle of an LIB in an electric

6
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

Fig. 5. Left column: Temporal evolution of the LIB temperature 𝑇 (𝑡), predicted with dual ensemble Kalman filter (solid line) and observed experimentally (dashed line) at 1 C-rate
(Top), 2 C-rate (Middle), 3 C-rate (Bottom). The predicted results are averages over 20 realizations. Right column: The corresponding voltage 𝑉 (𝑡), predicted with dual ensemble
Kalman filter (dashed lines) and observed experimentally (solid lines).

vehicle consists of a sequence of charging (regenerative braking) and The proposed numerical framework provides accurate estimates of
discharging of the battery. In addition, since the definition of 𝑓𝑚 and battery temperature, terminal voltage and Li+ concentration in both
𝑔𝑚 is general, our method can potentially work for other battery types the solid phase and the electrolyte, over a wide range of C-rate dis-
as long as the state propagation function model is accurate. charge. Results have been validated against experimental data collected
from the LGChem NMC battery cells at the Stanford Energy Control
laboratory.
5. Summary and conclusions
The proposed algorithm can be used in current battery management
systems to enhance SOC/SOH estimation while guaranteeing battery
We proposed a novel algorithm to estimate the temperature of a safety operation.
LIB cell using the currently available on-board current and voltage
sensors by means of inversion of the electrochemical dynamics. It uses Declaration of competing interest
the enhanced single-particle model to relate the current and terminal
voltage to the temperature, and a dual ensemble Kalman filter to The authors declare that they have no known competing finan-
assimilate direct measurements of the former into predictions of the cial interests or personal relationships that could have appeared to
latter. influence the work reported in this paper.

7
P. Wang et al. Journal of Energy Storage 34 (2021) 102133

Acknowledgments [10] Y. Kim, S. Mohan, J.B. Siegel, A.G. Stefanopoulou, Y. Ding, The estimation
of temperature distribution in cylindrical battery cells under unknown cooling
conditions, IEEE Trans. Control Syst. Tech. 22 (6) (2014) 2277–2286.
P. Wang, L. Yang, and H. Wang were supported by the National Key
[11] M.V. Morganti, S. Longo, M. Tirovic, C.-Y. Blaise, G. Forostovsky, Multi-scale,
Research and Development Program of China (Grant No. electro-thermal model of NMC battery cell, IEEE Trans. Veh. Tech. 68 (11) (2019)
2017YFB0701700 and 2018YFB0703902). DMT was supported in part 10594–10606.
by the gift from Total. We thank Anirudh Allam, Ph.D. student at ERE [12] X. Lin, H.E. Perez, J.B. Siegel, A.G. Stefanopoulou, Robust estimation of battery
Department, Stanford University, for making the eSPM numerical code system temperature distribution under sparse sensing and uncertainty, IEEE
Trans. Control Syst. Tech. 28 (3) (2020) 753–765.
from [4] available. Experimental battery data were collected at the [13] D.M. Tartakovsky, P.A. Gremaud, Method of distributions for uncertainty quan-
Stanford Energy Control Lab, https://onorilab.stanford.edu/research. tification, in: R. Ghanem, D. Higdon, H. Owhadi (Eds.), Handbook of Uncertainty
Quantification, Springer, 2015, pp. 1–22.
References [14] J. Sun, G. Wei, L. Pei, R. Lu, K. Song, C. Wu, C. Zhu, Online internal temperature
estimation for lithium-ion batteries based on Kalman filter, Energies 8 (5) (2015)
4400–4415.
[1] C. Forgez, D.V. Do, G. Friedrich, M. Morcrette, C. Delacourt, Thermal modeling
[15] J. Kasnatscheew, T. Placke, B. Streipert, S. Rothermel, R. Wagner, P. Meister,
of a cylindrical LiFePO4/graphite lithium-ion battery, J. Power Sour. 195 (9)
I.C. Laskovic, M. Winter, A tutorial into practical capacity and mass balancing
(2010) 2961–2968.
of lithium ion batteries, J. Electrochem. Soc. 164 (12) (2017) A2479.
[2] M. Guo, G. Sikha, R.E. White, Single-particle model for a lithium-ion cell:
[16] C.D. Rahn, C.-Y. Wang, Battery System Engineering, John Wiley & Sons, Ltd,
Thermal behavior, J. Electrochem. Soc. 158 (2) (2011) A122–A132.
2013.
[3] T.R. Tanim, C.D. Rahn, C.-Y. Wang, State of charge estimation of a lithium ion
[17] L.O. Valøen, J.N. Reimers, Transport properties of LiPF6-based Li-ion battery
cell based on a temperature dependent and electrolyte enhanced single particle
electrolytes, J. Electrochem. Soc. 152 (5) (2005) A882–A891.
model, Energy 80 (2015) 731–739.
[18] E. Prada, D. Di Domenico, Y. Creff, J. Bernard, V. Sauvant-Moynot, F. Huet, Sim-
[4] A. Allam, S. Onori, An interconnected observer for concurrent estimation of bulk
plified electrochemical and thermal model of LiFePO4-graphite Li-ion batteries
and surface concentration in the cathode and anode of a lithium-ion battery, IEEE
for fast charge applications, J. Electrochem. Soc. 159 (9) (2012) A1508–A1519.
Trans. Ind. Electron. 65 (9) (2018) 7311–7321.
[19] Y. Ji, Y. Zhang, C. Wang, Li-ion cell operation at low temperatures, J.
[5] M. Doyle, T. Fuller, J. Newman, Modeling of the Galvanostatic charge and
Electrochem. Soc. 160 (4) (2013) A636–A649.
discharge of the lithium/polymer/insertion cell, J. Electrochem. Soc. 140 (6)
[20] R. Hermann, A. Krener, Nonlinear controllability and observability, IEEE Trans.
(1993) 1526–1533.
Automat. Control 22 (5) (1977) 728–740.
[6] H. Arunachalam, S. Onori, I. Battiato, On veracity of macroscopic lithium-ion
[21] L. Ljung, Asymptotic behavior of the extended Kalman filter as a parameter
battery models, J. Electrochem. Soc. 162 (10) (2015) A1940–A1951.
estimator for linear systems, IEEE Trans. Automat. Control 24 (1) (1979) 36–50.
[7] X. Zhang, D.M. Tartakovsky, Effective ion diffusion in charged nanoporous
[22] E.A. Wan, A.T. Nelson, Dual extended Kalman filter methods, in: Kalman Filtering
materials, J. Electrochem. Soc. 164 (4) (2017) E53–E61.
and Neural Networks, vol. 123, Wiley Online Library, 2001.
[8] X. Lin, H.E. Perez, J.B. Siegel, A.G. Stefanopoulou, Y. Li, R.D. Anderson, Y.
[23] H. Moradkhani, S. Sorooshian, H.V. Gupta, P.R. Houser, Dual state–parameter
Ding, M.P. Castanier, Online parameterization of lumped thermal dynamics in
estimation of hydrological models using ensemble Kalman filter, Adv. Water
cylindrical lithium ion batteries for core temperature estimation and health
Resour. 28 (2) (2005) 135–147.
monitoring, IEEE Trans. Control Syst. Tech. 21 (5) (2012) 1745–1755.
[24] C.-H. Chen, F.B. Planella, K. O’Regan, D. Gastol, W.D. Widanage, E. Kendrick,
[9] W.B. Gu, C.Y. Wang, Thermal-electrochemical modeling of battery systems, J.
Development of experimental techniques for parameterization of multi-scale
Electrochem. Soc. 147 (8) (2000) 2910–2922.
lithium-ion battery models, J. Electrochem. Soc. 167 (8) (2020) 080534.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy