Article

A Physics-Based Electrochemical Model for Lithium-

Ion Battery State-of-Charge Estimation Solved by an
Optimised Projection-Based Method and Moving-
Window Filtering
Wei He 1, Michael Pecht 1, David Flynn 2 and Fateme Dinmohammadi 2,*
1 Center for Advanced Life Cycle Engineering (CALCE), University of Maryland, College Park, MD 20742,
USA; weihe@calce.umd.edu (W.H.); pecht@calce.umd.edu (M.P.)
2 Smart Systems Group, School of Engineering and Physical Sciences, Heriot-Watt University, Edinburgh

EH14 4AS, UK; d.flynn@hw.ac.uk

* Correspondence: F.Dinmohammadi@hw.ac.uk

Received: 25 July 2018; Accepted: 9 August 2018; Published: 14 August 2018

Abstract: State-of-charge (SOC) is one of the most critical parameters in battery management
systems (BMSs). SOC is defined as the percentage of the remaining charge inside a battery to the
full charge, and thus ranges from 0% to 100%. This percentage value provides important
information to manufacturers about the performance of the battery and can help end-users
identify when the battery must be recharged. Inaccurate estimation of the battery SOC may cause
over-charge or over-discharge events with significant implications for system safety and
reliability. Therefore, it is crucial to develop methods for improving the estimation accuracy of
battery SOC. This paper presents an electrochemical model for lithium-ion battery SOC estimation
involving the battery’s internal physical and chemical properties such as lithium concentrations.
To solve the computationally complex solid-phase diffusion partial differential equations (PDEs)
in the model, an efficient method based on projection with optimized basis functions is presented.
Then, a novel moving-window filtering (MWF) algorithm is developed to improve the
convergence rate of the state filters. The results show that the developed electrochemical model
generates 20 times fewer equations compared with finite difference-based methods without losing
accuracy. In addition, the proposed projection-based solution method is three times more efficient
than the conventional state filtering methods such as Kalman filter.

Keywords: lithium-ion battery; state-of-charge (SOC); prognostic and health management (PHM);
projection-based method; moving-window filtering (MWF); reliability

1. Introduction
The size of the market for lithium-ion (Li-ion) batteries has grown considerably in recent years.
According to recent statistics provided by Grand View Research, Inc.
(https://www.grandviewresearch.com/), the global Li-ion battery market is expected to grow at an
average annual rate of 17% and reaches $93.1 billion by 2025. Li-ion batteries are electrochemical
energy storage devices that are used in a variety of applications, ranging from Bluetooth
headphones to mobile phones, laptops, electric vehicles (EVs), and satellites. These batteries consist
of an anode (negative electrode) and a cathode (positive electrode) separated by an ion-conductive,
electronically insulating solution called electrolyte [1]. When the battery is connected to an external
load, the anode supplies a current of electrons that flow through the load and are accepted by the
cathode, while the ions move across the electrolyte from the cathode to the anode because of
Energies 2018, 11, 2120; doi:10.3390/en11082120 www.mdpi.com/journal/energies
Energies 2018, 11, 2120 2 of 24

electroneutrality. If the electrochemical process of the battery is able to be reversed by applying an

external current, then the battery is called a rechargeable or secondary battery.
State-of-charge (SOC) is one of the most critical parameters in rechargeable batteries. SOC is
defined as the percentage of the charge left in the battery divided by the battery rated capacity, and
thus ranges from 0% to 100% [2]. SOC percentage provides useful information to manufacturers
about the performance of the battery and also can help end-users identify when the battery must be
recharged. In addition, the battery management system (BMS) requires SOC information for power
management, because battery power is a function of impedance, which itself is dependent on SOC
percentage. Therefore, inaccurate estimation of SOC may cause over-charge or over-discharge
events which could lead to permanent damage to battery cells. For instance, EV drivers need to
know the battery SOC level in order to determine the remaining driving range. Some cases have
been reported of EV drivers halted in the highway due to underestimation or overestimation of the
true driving range as a result of erroneous assessment of the battery power capability. To avoid
such negative impact, a common practice is to overdesign the batteries to account for uncertainties.
However, the accurate estimation of SOC will help to reduce over-engineering of batteries in terms
of size, weight and cost, whilst maintaining safety and reliability.
Estimating a battery’s SOC level is a challenging task due to variable loading conditions, noisy
sensors, and biases in input data and measurements. To date, many approaches have been
developed for the estimation of battery SOC, among which Coulomb counting is the most popular
technique (for more see [3,4]). In this technique, the current is integrated over time to estimate the
battery SOC. Although Coulomb counting is simple and easy to implement, measurement and
calculation errors may be accumulated by the integration function, thus reducing accuracy of
estimation. In addition, the SOC estimation obtained by this technique is highly dependent on the
quality of input initial conditions. Another commonly used approach for the estimation of battery
SOC is the voltage-based method [5], where the value of SOC is determined based on a voltage-
SOC lookup table. However, voltage-based methods are not very suitable for Li-ion batteries
because of their flat plateau of discharge characteristics. Recently, physics-based methods have been
applied to Li-ion battery SOC estimation [6,7]. These methods have intrinsic advantages over
traditional Coulomb counting and voltage-based methods. For example, physics-based methods are
robust to sensor error and are less dependent on accurate initial SOCs.
In physics-based methods, the major problem for real-time SOC estimation is the
computational complexity of the coupled partial differential equations (PDEs) which are used to
describe the physical processes inside the battery. Simplification of the diffusion PDEs in the solid-
phase particles has been found to be the key to reduce the computation time and memory
requirement of the physics-based models. The existing methods for diffusion PDE simplification
suffer from lots of drawbacks. For instance, the commonly used volume-averaging method can
generate unstable equations, which may lead to unacceptable error. Projection-based methods are
promising for the model reduction of diffusion PDEs. However, the basis function used in the
projection can greatly influence the performance, and how to construct an optimal basis function
has not been discussed in the literature. In addition, state filters are required when using a physics-
based model for battery SOC estimation. Available state filter algorithms are effective to handle
measurement noise and modeling uncertainties. However, these algorithms converge slowly if the
initial error is high, which can cause inconvenience to battery users.
The current study aims to develop a physics-based electrochemical model for Li-ion battery
SOC estimation involving the battery’s internal physical and chemical properties such as lithium
concentrations. To solve the computationally complex solid-phase diffusion equations in the model,
an efficient method based on projection with optimized basis functions is presented. Then, a novel
moving-window filtering (MWF) algorithm is proposed to improve the convergence rate of the
state filters. The results show that the developed electrochemical model generates 20 times fewer
equations compared with finite difference-based methods without losing accuracy. In addition, the
proposed projection-based solution method is three times more efficient than the conventional state
filtering methods such as Kalman filter.
Energies 2018, 11, 2120 3 of 24

The remainder of this paper is structured as follows: Section 2 provides an overview of battery
SOC estimation algorithms, including electrochemical models and state filter algorithms. Section 3
develops an electrochemical model to apply for the real-time SOC estimation of Li-ion batteries.
Section 4 proposes an optimized projection-based method to reduce computational complexity of
the PDEs in the model. Section 5 presents an improved state filtering algorithm based on moving-
window strategy to infer SOC estimation. Experimental validations of the simplified
electrochemical model and the moving-window filter are performed in Section 6. Section 7
summarizes the contributions of this study and proposes future research directions.

2. Research Background

2.1. Electrochemical Model

The equivalent circuit model (ECM) is the most commonly used battery model in a BMS (see
[8–13]). It is formed by an open-circuit voltage source connected with a set of electric elements, such
as resistors and capacitors, to model the electrical behavior of a battery. The ECM is widely used for
battery SOC estimation because it is computationally efficient. The technique estimates battery
voltages based on current inputs, and different resistor-capacitor (RC) branches can be used to
capture different time constants inherent in the battery system. However, because the ECM is
empirical in nature, it provides little insights into the electrochemical process inside the battery, and
it cannot provide highly accurate results. In contrast, the electrochemical model uses partial
differential equations (PDEs) to model the physics of the battery; it can be used to compute the
electrochemical states inside a battery and provides accurate information about lithium
concentrations and over-potentials, which can be used to prevent harmful side-reactions. For
example, the electrochemical model developed by Doyle and Newman [14] included the key
chemistry and physics in a Li-ion battery, such as diffusion, migration, and reaction kinetics. The
model consisted of two scales, namely, a macro x-scale along the thickness of the electrodes, and a
micro r-scale along the radius direction of the small solid-phase electrode particles.
The electrochemical processes are usually modeled by a set of coupled partial differential
equations (PDEs). To solve PDEs, spatial discretization methods such as finite element, finite
difference, and finite volume are used. When applying such methods, the electrochemical model
has to be discretized in both the x and r directions, resulting in thousands of states (equations).
Implementing a real-time BMS using the spatial discretization method is computationally
prohibitive and on-board memory demanding. Therefore, in order to develop an electrochemical
model for real-time battery management, deriving a simplified and computationally efficient
mathematical model is crucial. The single-particle model (SPM) is a typical example of a simplified
model with less than 10 states [15], in which each electrode is modeled by a single spherical particle.
As a result, the spatial variations of potentials cannot be determined by SPM. In addition, the
electrolyte concentration is generally assumed to be a constant in time and space. However, this
assumption does not hold for high-current applications [16].
In a full electrochemical model, the diffusion PDE in solid particles is a major part of the
model. It determines the lithium concentration in the electrode particles and provides information
on the level of lithium available for producing an electrical charge. In other words, it provides
information to users about the remaining level of electrical energy in the battery. When solving the
electrochemical model, the solid-phase diffusion PDE generates more than 80% of the states
(equations) after discretizing the model. That is because the solid-phase diffusion not only varies in
the macro x-scale along the thickness of the electrode but also in the micro r-scale along the radius
direction of the solid-phase particles. Other PDEs, e.g., electrolyte diffusion, only change in the
macro x-scale. For this reason, the solid-phase diffusion is inherently more computationally
intensive than other PDEs in the electrochemical model. For example, if we discretize the battery
model in the x-scale with n nodes and in the r-scale with m nodes, then the total number of
generated equations from the solid-phase diffusion is n × m, while other PDEs in the model will
Energies 2018, 11, 2120 4 of 24

only generate n equations. Therefore, simplifying the diffusion PDEs in the solid phase is essential
for providing a real-time capability.
Subramanian et al. [17] developed an efficient solution for the diffusion PDEs by
approximating the solid-phase lithium concentration using even-order monomials. The coefficients
for the polynomials were obtained by volume-averaging of the PDEs and the derivatives of the
PDEs. However, Mayhew et al. [18] found that when using more than three monomials to
approximate the lithium concentration, the volume-averaging technique generated an unstable
reduced-order model (ROM). In another study, Chaturvedi et al. [19] proposed a generalized
projection framework to reduce the PDEs to a set of differential algebraic equations (DAEs).
However, the method did not address the issue on constructing an optimal basis function, which
resulted in generation of numerically unstable equations and led to erroneous results. To the best of
the authors’ knowledge, there is no available literature addressing the construction of an optimal
basis function for the diffusion PDE reduction that is independent of loading conditions.

2.2. State Filtering

In many practical cases, the physical state of a system cannot be determined by direct
observation. For example, the SOC or lithium concentration cannot be directly measured with
sensor information. A state filter is an algorithm to infer the internal state of a given system from
the measurements of the input. A system model is required to estimate system states. Figure 1
presents a flowchart for battery SOC estimation based on a state filter. As shown, the algorithm
compares the actual measured output and the model output, and then an optimal state value (in
this case, SOC) is found such that the difference between the measurement and the model output is
minimized.

Inputs: I Outputs: V

Sensor
Sensor

Battery State filter State: SOC

model

Figure 1. Flowchart of model-based battery SOC estimation algorithm.

State filter algorithms possess multiple advantages. For example, they can handle noisy
measurements and tolerate the modeling error due to unit-to-unit variations and unmodeled
physics. In this algorithm, the system model is expressed in a state-space form as below:
State function: 𝑥𝑘+1 = 𝑓(𝑥𝑘 , 𝑢𝑘 ) + 𝑤𝑘 (1)
Measurement function: 𝑦𝑘 = ℎ(𝑥𝑘 , 𝑢𝑘 ) + 𝑣𝑘 (2)
where 𝑤𝑘 and 𝑣𝑘 represent process noise and measurement noise, respectively.
The commonly used state filter algorithms include Kalman filter, extended Kalman filter
(EKF), and unscented Kalman filter (UKF). Kalman filter is used for linear systems and is the
optimal filter for Gaussian measurement and Gaussian process noise. For nonlinear systems, EKF
and UKF are mostly used. Battery SOC estimation is a nonlinear problem, because the open circuit
voltage (OCV) of the battery is nonlinear. Corno et al. [20] developed a SOC estimation method
based on EKF and an electrochemical model. Han et al. [21] proposed a method to estimate SOC
using a physics-based single particle model and EKF. Santhanagopalan et al. [22] used UKF and an
electrochemical model to estimate SOC of high-power Li-ion batteries. UKF has been shown to
provide better results compared with EKF. However, conventional Kalman filter-based approaches
converge slowly when initial estimation error is high. Attempts to increase the convergence rate by
Energies 2018, 11, 2120 5 of 24

changing initial covariance values can result in noisy estimations for SOC. Figure 2 shows an
example of using UKF for SOC estimation with an electrochemical model. The initial SOC error is
40%. The UKF can eventually converge to true values. However, it takes around half an hour to
converge. Therefore, the state filter algorithm has to be improved to ensure faster convergence.

Figure 2. Slow convergence of UKF given 40% initial SOC estimation error.

3. An Electrochemical Battery SOC Estimation Model

In this Section, a physics-based electrochemical SOC estimation model is developed for Li-ion
batteries. Li-ion batteries are electrochemical systems whose performance is determined by their
electrochemical states. For example, the battery output voltage and capacity are dependent on the
lithium concentration in the solid electrochemical particles. For SOC estimation, the lithium
concentration in the solid phase can be used to calculate electrode SOC as given in Equation (3):
𝐿 𝑅
3 𝑐𝑠 (𝑥, 𝑟, 𝑡)
SOC(𝑡) = 3
∫ ∫ 𝑟2 𝑑𝑟𝑑𝑥 (3)
𝐿𝑅 0 0 𝑐𝑠,𝑚𝑎𝑥
where 𝑐𝑠 (𝑥, 𝑟, 𝑡) is the solid-phase concentration, 𝑐𝑠,𝑚𝑎𝑥 is the max concentration in the solid phase, L
is the thickness of the electrode, and R is the particle radius.
Electrochemical models determine the state of a battery based on its degradation level. The
degradation of a Li-ion battery is often caused by harmful side-reactions inside the battery. Lithium
plating is a typical example that consumes cyclable lithium. The occurrence of side-reactions
depends on the overpotential. If the overpotential can be estimated, some algorithms can be
developed to minimize the side-reactions inside the cell, which results in improved battery
reliability. To estimate the electrochemical states, an accurate and computational efficient
electrochemical model is required.
The electrochemical model consists of coupled PDEs to describe the diffusion, migration, and
reaction kinetics inside the battery. Solving the electrochemical models is challenging. Doyle and
Newman [14] developed an electrochemical model that takes mass transfer, diffusion, migration,
and reaction kinetics into account. In the literature, this model is also called the “dualfoil” model. It
is the most commonly used model for simulating the electrochemical process of the Li-ion batteries,
and it has been extensively validated by experiments. In this model, the battery is divided into three
domains, namely, negative electrode, separator, and positive electrode. Figure 3 shows the structure
of a Li-ion battery and the governing equations of the physics process in each part of the battery.
The lattice structures of the electrode active materials are modeled by small spherical particles. The
transport of Li-ions in the solid electrode partials (solid phase) and electrolyte phase is modeled by
diffusion equations and coupled by Butler–Volmer reaction kinetics. Using this model, the lithium
concentrations and overpotentials inside the battery can be calculated. The major equations
describing the physics of the battery are introduced below.
Energies 2018, 11, 2120 6 of 24

Figure 3. Dualfoil battery electrochemical model.

We define an x-axis starting from the negative current collector to the positive current
collector, and an r-axis along the radius direction of the solid electrode particle. The transportation
of the Li-ions in the solid-phase particles can be modeled by a diffusion equation as follows [14]:
∂𝑐𝑠 (𝑥, 𝑟, 𝑡) 1 ∂ ∂𝑐𝑠 (𝑥, 𝑟, 𝑡)
= 2 (𝐷𝑠 𝑟 2 ) (4)
∂𝑡 𝑟 ∂𝑟 ∂𝑟
with boundary conditions and initial conditions given by:
∂𝑐𝑠
| =0 (5)
∂𝑟 𝑟=0
∂𝑐𝑠 1
| = − 𝑗𝑛 (6)
∂𝑟 𝑟=𝑅𝑝 𝐷𝑠

𝑐𝑠 (𝑥, 𝑟, 0) = 𝑐𝑠0 (7)

where cs is the solid-phase concentration, jn is the molar flux, Ds is the diffusion coefficient, and Rp is
the radius of the solid-phase particle. The transportation of lithium ions in the electrolyte can be
modeled as follows:
∂𝑐𝑒 (𝑥, 𝑡) ∂ ∂𝑐𝑒 (𝑥, 𝑡) 1 0
= (𝐷𝑒 + 𝑡 𝑖 (𝑥, 𝑡)) (8)
∂𝑡 ∂𝑥 ∂𝑥 𝐹𝜀𝑒 𝑎 𝑒
where ce is the lithium concentration in the electrolyte, De is the effective diffusion coefficient, 𝜀𝑒 is
the volume fraction of the electrolyte, F is Faraday’s constant, and 𝑡𝑎0 is the transference number for
the anion. On the surface of solid-phase particles, an electrochemical reaction occurs when lithium
ions are transferred from (to) the solid phase to (from) the electrolyte phase. The reaction flux jn
depends on the overpotential 𝜂𝑠 , and this relationship can be described by the Butler-Volmer
equation as
𝑖0 (𝑥, 𝑡) 𝛼𝑎 𝐹 𝛼𝑐 𝐹
𝑗𝑛 (𝑥, 𝑡) = [exp ( 𝜂𝑠 (𝑥, 𝑡)) − exp ( 𝜂 (𝑥, 𝑡))] (9)
𝐹 𝑅𝑇 𝑅𝑇 𝑠
Energies 2018, 11, 2120 7 of 24

where 𝛼𝑎 and 𝛼𝑐 are transport coefficients, and 𝑖0 is the exchange current density, which is defined
as
𝛼𝑎
𝑖0 (𝑥, 𝑡) = 𝑟𝑒𝑓𝑓 𝑐𝑒 (𝑥, 𝑡)𝛼𝑎 × (𝑐𝑠,𝑚𝑎𝑥 − 𝑐𝑠𝑠 (𝑥, 𝑡)) 𝑐𝑠𝑠 (𝑥, 𝑡)𝛼𝑐 (10)

where css and cs,max represent, respectively, the surface concentration and the maximum possible
concentration in the solid particles of the electrode, and reff is a reaction rate constant. As shown in
Equation (9), the overpotential 𝜂𝑠 determines the rate of electrochemical reactions or flux jn and
varies with time t and location x. The overpotentials are calculated by the solid-phase potential Φ𝑠 ,
electrolyte potential Φ𝑒 , open-circuit potential 𝑈(𝑐𝑠𝑠 (𝑥, 𝑡)), and the flux jn, as follows:

𝜂𝑠 (𝑥, 𝑡) = Φ𝑠 (𝑥, 𝑡) − Φ𝑒 (𝑥, 𝑡) − 𝑈(𝑐𝑠𝑠 (𝑥, 𝑡)) − 𝐹𝑅𝑓 𝑗𝑛 (𝑥, 𝑡) (11)

where 𝐹𝑅𝑓 𝑗𝑛 (𝑥, 𝑡) is the potential drop due to the film resistance 𝑅𝑓 at the solid electrolyte interface.
In each electrode, the flux jn is dependent upon the divergence of current density in the electrolyte
ie,, as follows:
∂𝑖𝑒 (𝑥, 𝑡)
= 𝑎𝐹𝑗𝑛 (𝑥, 𝑡) (12)
∂𝑥
where a is the specific interfacial area of the particle. The ie should be zeroes at the current collectors
and should be equal to the applied current density, I, at the separator.

4. An Optimised Projection Method to Solve Solid-Phase Diffusion PDEs

In order to solve the electrochemical model, the traditional finite element and finite difference
methods usually generate thousands of states, making these methods computationally intensive for
real-time applications. To implement the electrochemical model in a real-time BMS where the
memory and computational power are limited, an efficient numerical solution for the diffusion
PDEs must be developed. In this Section, a projection-based method is developed to reduce the
PDEs to a low-order DAE system. The projection-based method approximates the concentration
𝑐𝑠 (𝑟, 𝑡) by a set of basis functions with time-varying coefficients as follows:

𝑐𝑠 (𝑟, 𝑡) = 𝚽 T (𝑟) ∙ 𝐚(𝑡) (13)

where 𝚽(𝑟) = [𝜙1 (𝑟), 𝜙2 (𝑟), … , 𝜙𝑁 (𝑟)]T and 𝐚(𝑡) = [𝑎1 (𝑡), 𝑎2 (𝑡), … , 𝑎𝑁 (𝑡)]T . Once 𝑎𝑖 (𝑡) is
determined, 𝑐𝑠 (𝑟, 𝑡) can be found accordingly. Therefore, the problem is converted into a problem of
finding time-varying coefficients 𝑎𝑖 (𝑡). First, N equations are needed, as there are N unknown 𝑎𝑖 (𝑡).
The projection-based method is used to generate the N equations based on the diffusion equation.
The optimized basis function is constructed in three steps: (i) definition of elementary basis
functions, e.g., even-order monomials, (ii) orthonormalization of elementary basis function to
ensure the numerical stability of the generated equations, and (iii) linear transformation of the
orthonormal basis function. In what follows, these three steps are explained in detail:

4.1. Definition of Elementary Basis Functions

A function 𝑃: 𝑈 → 𝑈 is called a projector if it is: (i) linear, i.e., P(𝑎 ∙ 𝑓1 + 𝑏 ∙ 𝑓2 ) = 𝑎 ∙ P(𝑓1 ) + 𝑏 ∙
P(𝑓2 ) ; and (ii) idempotent, i.e., P 2 (𝑓) = P(𝑓) . For example, the Galerkin projector for the
function 𝑓(𝑟) is given by:
𝑅
∫0 𝑝 𝑟 2 𝑓(𝑟)∙𝛽𝑖 𝑑𝑟
P𝛽𝑖 (𝑓(𝑟)) = 𝑅 ∙ 𝛽𝑖 (14)
∫0 𝑝 𝑟 2 𝛽𝑖 ∙𝛽𝑖 𝑑𝑟

where 𝛽𝑖 is called the test function of the projector. With different test functions, the projection
results are different. This property is useful for constructing different equations to calculate the
time-varying coefficients in Equation (13). In this paper, we design a basis function such that the
first boundary condition at the center of the particle in Equation (5) is fulfilled. An example of such
Energies 2018, 11, 2120 8 of 24

a basis function is the even-order monomials (i.e., 1, r2, r4…). By substituting the approximation to
the boundary condition at the surface of the particle, we obtain the first equation as follows:

∂𝚽 T (𝑥)
𝐷𝑠 ( ∂𝑟
| )∙ 𝐚(𝑡) = −𝑗𝑛 (15)
𝑟=𝑅𝑝

The remaining N – 1 equations are obtained by applying the projection N – 1 times to both
sides of the solid-phase diffusion PDEs using different basis functions 𝛽𝑖 , 𝑖 = 1, … , 𝑁 − 1,
R𝑝
𝜕𝑐𝑠 (𝑟, 𝑡) 1 𝜕 𝜕𝑐𝑠 (𝑟, 𝑡) 𝜕𝑐𝑠 (𝑟, 𝑡)
P𝛽𝑖 ( ) = P𝛽𝑖 ( 2 (𝐷𝑠 𝑟 2 )) ⇒ ∫ 𝑟 2 𝛽𝑖 (𝑟)𝑑𝑟
𝜕𝑡 𝑟 𝜕𝑟 𝜕𝑟 0 𝜕𝑡
R 1 𝜕 𝜕𝑐𝑠 (𝑟,𝑡)
= ∫0 𝑝 𝑟 2 (𝑟2 𝜕𝑟 (𝐷𝑠 𝑟 2 𝜕𝑟
)) 𝛽𝑖 (𝑟)𝑑𝑟 (16)

Integrating by parts for the right-hand side of Equation (16) and applying the boundary
condition in Equations (5) and (6), we have:
𝑅 𝜕𝑐𝑠 (𝑟,𝑡) 𝑅 𝜕𝑐𝑠 (𝑟,𝑡) 𝜕𝛽𝑖 (𝑟)
∫0 𝑝 𝑟 2 𝜕𝑡
𝛽𝑖 (𝑟)𝑑𝑟 = − ∫0 𝑝 𝐷𝑠 𝑟 2 𝜕𝑟 𝜕𝑟
𝑑𝑟 − 𝑅𝑝 2 𝛽𝑖 (𝑅)𝑗𝑛 , i = 1,…,N − 1 (17)

By substituting the approximation of Equation (13) into Equation (17), we obtain:

𝑅𝑝
(∫ 𝑟 2 𝚽 T (𝑟)𝛽𝑖 (𝑟)𝑑𝑟) ∙ 𝐚̇ (𝑡) =
0 (18)
𝑅 𝜕𝚽 T (𝑟) 𝜕𝛽𝑖 (𝑟) 2
(− ∫0 𝑝 𝐷𝑠 𝑟 2 𝜕𝑟 𝜕𝑟
𝑑𝑟) ∙ 𝐚(𝑡) + (−𝑅𝑝 𝛽𝑖 (𝑅𝑝 )) 𝑗𝑛 , = 1, … , 𝑁 − 1

Then, the obtained differential algebraic equations are summarized as follows:

𝐌 ∙ 𝐚̇ (𝑡) = 𝐀 ∙ 𝐚(𝑡) + 𝑩 ∙ 𝑗𝑛 ; 𝐄 ∙ 𝐚(𝑡) = −𝑗𝑛 ; css (𝑡) = 𝐂 ∙ 𝐚(𝑡) (19)

where
𝑅 𝑅 𝜕𝛃(𝑟) 𝜕𝚽 T (𝑟)
𝐌 = [∫0 𝑝 𝑟 2 𝛃(𝑟)𝚽 T (𝑟)𝑑𝑟] ; 𝐀 = [− ∫0 𝑝 𝐷𝑠 𝑟 2 𝜕𝑟 𝜕𝑟
𝑑𝑟]
𝑁−1×𝑁 𝑁−1×𝑁
𝜕𝚽 T (𝑟)
𝐁 = [−𝑅𝑝 2 𝛃(𝑅𝑝 )]𝑁−1×1 ; 𝐄 = [𝐷𝑠 𝜕𝑟
| ] ; 𝐂 = [𝚽 T (𝑅𝑝 )]1×𝑁 (20)
𝑟=𝑅𝑝
1×𝑁

4.2. Orthonormalization of the Basis Function

To ensure numerical stability and avoid large condition number, 𝚽 and 𝛃 have to be
orthonormal basis functions, such that
1 𝑖=𝑗
Proj𝛽𝑖 (𝜙𝑗 ) = { (21)
0 𝑖≠𝑗
and then the matrix M becomes
1 0 0 0
⋯
𝐌 = [0 1 0 0] (22)
⋮ ⋱ ⋮
0 0 ⋯ 1 0
Let us assume that the original elementary basis set is 𝑺 = {𝛷1 , 𝛷2 , … , 𝛷𝑘 }. The Gram-Schmidt
algorithm is used to generate a set of orthonormal basis functions ̂
𝑺 = {𝑢1 , 𝑢2 , … , 𝑢𝑘 } that spans the
𝑛
same k-dimensional space of ℝ as S, and
1 𝑖=𝑗
〈𝑢𝑖 , 𝑢𝑗 〉 = { (23)
0 𝑖≠𝑗
The steps of the Gram-Schmidt algorithm are given in Table 1.
Energies 2018, 11, 2120 9 of 24

Table 1. Steps of the Gram-Schmidt algorithm.

Steps Orthogonalizing Normalizing

𝑢1
𝑢1 =
Step 1: 𝑢1 = 𝜑1 ‖𝑢1 ‖

𝑢2
𝑢2 =
Step 2: 𝑢2 = 𝜑2 − Proj𝑢1 (𝜑2 ) ‖𝑢2 ‖

𝑢3
𝑢3 =
Step 3: 𝑢3 = 𝜑3 − Proj𝑢1 (𝜑3 ) − Proj𝑢2 (𝜑3 ) ‖𝑢3 ‖

. . .
. . .
. . .
𝑘−1
𝑢𝑘
Step k: 𝑢𝑘 = 𝜑𝑘 − ∑ Proj𝑢𝑗 (𝜑𝑘 ) 𝑢𝑘 =
‖𝑢𝑘 ‖
𝑗=1

4.3. Linear Transformation of the Orthonormal Basis Function

The selection of basis function has a significant effect on the approximation accuracy. In this
study, we develop a method to construct an optimal basis function from a set of elementary basis
functions. First, we define a set of high-order orthonormal elementary basis functions, U, formed by
applying the Gram–Schmidt algorithm to even-order monomials (i.e., 1, r2, r4…):

𝐔 = [𝑢1 , 𝑢2 , … , 𝑢𝑁 ]𝑇 (24)
We also choose the same U as the elementary test functions in the projection to satisfy the
orthonormality between the basis and test functions in Equation (21). Then the M, A, B, E, and C
matrices in Equation (20) become:
𝑅 𝑅 𝜕𝐔(𝑟) 𝜕𝐔T (𝑟)
𝐌 = [∫0 𝑝 𝑟 2 𝐔(𝑟)𝐔T (𝑟)𝑑𝑟] ; 𝐀 = [− ∫0 𝑝 𝐷𝑠 𝑟 2 𝜕𝑟 𝜕𝑟
𝑑𝑟]
𝑁−1×𝑁 𝑁−1×𝑁

𝜕𝐔T (𝑟)
𝐁 = [−𝑅𝑝 2 𝐔(𝑅𝑝 )]𝑁−1×1 ; 𝐄 = [𝐷𝑠 𝜕𝑟
| ] ; 𝐂 = [𝐔 T (𝑅𝑝 )]1×𝑁 (25)
𝑟=𝑅𝑝
1×𝑁

The construction of the optimal basis function 𝚿 and test function 𝛽 is accomplished by using a
linear transformation given by:

𝚿 = 𝐕×𝐔;= 𝐖×𝐔 (26)

where V and W are k by N matrices, and N and k are the original and reduced dimensionality with
k ≪ N. To ensure the orthonormality of the 𝚿 and 𝛃, we must have 𝐖 𝐓 𝐕 = 𝑰. The concentration
𝑐𝑠 (𝑟, 𝑡) can then be approximated by:

𝑐𝑠 (𝑟, 𝑡) = 𝚿𝑻 𝐚𝑟 (𝑡) (27)

By substituting the basis and test functions in Equation (26) into Equations (19) and (20), the
DAEs generated by the optimized basis and test functions are:

𝐌𝒓 ∙ 𝐚̇ 𝒓 (𝑡) = 𝐀𝒓 ∙ 𝐚𝒓 (𝑡) + 𝑩𝒓 ∙ 𝑗𝑛 ; 𝐄𝒓 ∙ 𝐚𝒓 (𝑡) = −𝑗𝑛 ; 𝑐𝑠𝑠 (𝑡) = 𝐂𝒓 ∙ 𝐚𝒓 (𝑡) (28)

where:
𝑅
𝐌𝒓 = [𝐖 (∫0 𝑝 𝑟 2 𝐔(𝑟)𝐔T (𝑟)𝑑𝑟) 𝐕 𝑻 ]
𝑘−1×𝑘

𝑅 𝜕𝐔(𝑟) 𝜕𝐔T (𝑟)

𝐀𝒓 = [𝐖 (− ∫0 𝑝 𝐷𝑠 𝑟 2 𝑑𝑟) 𝐕 𝑻 ]
𝜕𝑟 𝜕𝑟 𝑘−1×𝑘
𝑩𝒓 = [𝐖 (−𝑅𝑝 2 𝐔(𝑅𝑝 ))]
𝑘−1×1
Energies 2018, 11, 2120 10 of 24

𝜕𝐔T (𝑟)
𝐄𝒓 = [(𝐷𝑠 𝜕𝑟
| ) 𝐕𝑻] ; 𝐂𝒓 = [(𝐔 T (𝑅𝑝 )) 𝐕 𝑻 ] (29)
𝑟=𝑅𝑝 1×𝑘
1×𝑘
Therefore, the number of DAEs is further reduced to k.
Suppose the transfer function of the DAEs based on the high-order elementary basis function is
H(s), and the transfer function of the DAEs generated by the low-order optimized basis function is
𝛨𝑟 (𝑠). The objective of the optimization is to find the transform matrices V and W to minimize the
2-norm error between H(s) and Hr(s):
+∞
‖𝐻 − 𝐻𝑟 ‖2 = ∫−∞ |𝐻(𝑖𝜔) − 𝐻𝑟 (𝑖𝜔)|2 𝑑𝜔 . (30)

The Meier–Luenberger condition [23] provides the first-order necessary optimality condition
for the optimization problem in Equation (30), which is stated as follows:

Theorem 1. Let H(s) be the full-order system, and 𝛨𝑟 (𝑠) be the minimizer of ‖𝐻(𝑠) − 𝐻𝑟 (𝑠)‖2 with simple
poles 𝜆̂1 , … , 𝜆̂𝑟 . Then 𝐻(−𝜆̂𝑘 ) = 𝐻𝑟 (−𝜆̂𝑘 ) and 𝐻′ (−𝜆̂𝑘 ) = 𝐻𝑟′ (−𝜆̂𝑘 ) for 𝑖 = 1, … , 𝑟.

To satisfy this necessary condition, we need to interpolate H(s) and its derivative at the mirror
image of the poles of the reduced-order system. For linear stable ODEs, the rational Hermite
interpolation provides a way to determine the W and V matrices to obtain 𝛨𝑟 , and below is its
definition:
Given 𝐻(𝑠) = 𝐂 𝑇 (𝑠𝐈 − 𝐀)−1 𝐁, and r distinct points σ1 , … , σ𝑟 ∈ ℂ, let

𝑅𝑎𝑛(𝐕) = 𝑠𝑝𝑎𝑛{(σ1 𝐈 − 𝐀)−1 𝐁, … , (𝜎𝑟 𝐈 − 𝐀)−1 𝐁}

−1 −1
𝑅𝑎𝑛(𝐖) = 𝑠𝑝𝑎𝑛 {(σ1 𝐈 − 𝐀𝑻 ) 𝐂, … , (𝜎𝑟 𝐈 − 𝐀𝑻 ) 𝐂}. (31)

Define 𝐀𝐫 = (𝐖 𝐓 𝐕)−𝟏 𝐖 𝐓 𝐀𝐕 ; 𝐁𝐫 = (𝐖 𝐓 𝐕)−𝟏 𝐖 𝐓 𝐁; 𝐂𝐫 = 𝐂𝐕, then 𝐻𝑟 (𝑠) = 𝐂𝐫𝐓 (𝑠𝐈𝐫 − 𝐀𝐫 )−1 𝐁𝐫 is a

rational Hermite interpolate to 𝐻(𝑠) at σ1 , … , σ𝑟 , and 𝐻(𝜎𝑖 ) = 𝐻𝑟 (𝜎𝑖 ) and 𝐻 ′ (𝜎𝑖 ) = 𝐻𝑟′ (𝜎𝑖 ), for 𝑖 =
1, … , 𝑟. The rational Hermite interpolate can be directly used to find V and W to construct 𝐻𝑟 , if the
poles of 𝐻𝑟 are known. However, this information is not known as a prior. In this case, the iterative
rational Krylov algorithm (IRKA) can be used to iteratively correct the interpolation points until the
local minimum is found [24]. The steps of IRKA are defined below:
1 Make an initial selection of 𝜎𝑖 for 𝑖 = 1, … , 𝑟 that is closed under conjugation.
2 Choose V and W such that:
𝑅𝑎𝑛(𝐕) = 𝑠𝑝𝑎𝑛{(𝜎1 𝐈 − 𝐀)−1 𝐁, … , (𝜎𝑟 𝐈 − 𝐀)−1 𝐁} ;
𝑅𝑎𝑛(𝐖) = 𝑠𝑝𝑎𝑛{(𝜎1 𝐈 − 𝐀)−1 𝐂, … , (𝜎𝑟 𝐈 − 𝐀)−1 𝐂} ; where 𝐖 𝐓 𝐕 = 𝐈.
3 While relative change in {𝜎𝑖 } > 𝑡𝑜𝑙
a. 𝐀𝐫 = 𝐖 𝐓 𝐀𝐕
b. Assign 𝜎𝑖 ← −𝜆𝑖 (𝐀𝐫 ) for i = 1,…,r
c. Update V and W, such that
𝑅𝑎𝑛(𝐕) = 𝑠𝑝𝑎𝑛{(𝜎1 𝐈 − 𝐀)−1 𝐁, … , (𝜎𝑟 𝐈 − 𝐀)−1 𝐁};
𝑅𝑎𝑛(𝐖) = 𝑠𝑝𝑎𝑛{(𝜎1 𝐈 − 𝐀)−1 𝐂, … , (𝜎𝑟 𝐈 − 𝐀)−1 𝐂} ; where 𝐖 𝐓 𝐕 = 𝐈.
4 𝐀𝐫 = 𝐖 𝐓 𝐀𝐕, = 𝐖 𝐓 𝐁, 𝐂𝐫 = 𝐂𝐕
The IRKA was developed originally for asymptotically stable systems, i.e., all of the
eigenvalues of the pencil (A, I) must be in the left half of the complex plane. However, our system is
unstable, therefore the IRKA is not directly applicable. To solve this problem, we first convert the
DAEs into ODEs. Let us assume that R = [M; E]. If R is invertible, then the DAEs in Equation (19)
can be transformed to ODEs [15]. Let:

𝐳 = 𝐌 ∙ 𝐚 ; 𝑹−𝟏 = [𝑸1 𝑸2 ] (32)

where 𝑸1 ∈ ℝ𝑁−1×𝑁 and 𝑸2 ∈ ℝ𝑁−1×1 , then we have

Energies 2018, 11, 2120 11 of 24

𝐌 −𝟏 𝐳 𝐳
𝑸2 ] [−𝑗 ] = 𝑸1 𝐳 − 𝑸2 𝑗𝑛
𝐚 = [ ] [−𝑗 ] = [𝑸1 (33)
𝐄 𝑛 𝑛

By substituting Equation (33) into Equation (19), we obtain the ODEs:

𝐳̇ = 𝐀(𝑸1 𝐳 − 𝑸2 𝑗𝑛 ) + 𝐁𝑗𝑛 = 𝐀𝑸1 𝐳 + (𝐁 − 𝐀𝑸2 )𝑗𝑛

css (𝑡) = 𝐂(𝑸1 𝐳 − 𝑸2 𝑗𝑛 ) = 𝐂𝑸1 𝐳 − 𝐂𝑸2 𝑗𝑛 (34)
Since [𝑢1 , 𝑢2 , … , 𝑢𝑁 ] are orthonormal, [M; E] is:
1 0 0 ⋯ 0 0
0 1 0 ⋯ 0 0
𝐌 0 0 1 ⋯ 0 0
[ ]= ⋮ ⋮ ⋮ ⋱ ⋮ ⋮ (35)
𝐄
0 0 0 ⋯ 1 0
′ ′ ′ ′
[0 𝐷𝑠 𝑢2 𝐷𝑠 𝑢3 ⋯ 𝐷𝑠 𝑢𝑁−1 𝐷𝑠 𝑢𝑁 ]
It is straightforward to prove that the inverse of Equation (35) is:
1 0 0 ⋯ 0 0
0 1 0 ⋯ 0 0
0 0 1 ⋯ 0 0
𝐌 −1
[ ] = ⋮ ⋮ ⋮ ⋱ ⋮ ⋮ . (36)
𝐄 0 0 0 ⋯ 1 0
′ ′ ′
𝑢2 𝑢3 𝑢𝑁−1 1
0 − ′
− ′
⋯ − ′ ′
[ 𝑢𝑁 𝑢𝑁 𝑢𝑁 𝐷𝑠 𝑢𝑁 ]

Because u1 = 1, the first column and first row of A in Equation (20) are zeros. Therefore, the first
row and first column of 𝐀𝑸1 are zeros, and the ODEs are obtained as follows:
ż 1 (𝑡) 0 𝟎 ̅1
B
[ ̇ ]=[ ] 𝐳(𝑡) + [ ∙ 𝑗 (𝑡)
𝐙𝟐 (𝑡) 𝟎 ̅𝟐
𝐀 ̅ ] 𝑛
𝐁 𝟐

𝑐𝑠𝑠 (𝑡) = [C̅1 ̅ ∙ 𝑗𝑛 (𝑡)

𝐂̅𝟐 ] ∙ [z1 (𝑡), 𝐙𝟐 (𝑡)]T + D (37)
where 𝐙𝟐 = [z2 , z3 , … , z𝑁 ]T . Then, we decompose the system into a stable part and an unstable part,
and the IRKA is only applied to the stable part of the system. Therefore,
̅1 ∙ 𝑗𝑛 (𝑡)
ż 1 (𝑡) = B
{
𝐙̇𝟐 (𝑡) = 𝐀 ̅ 𝟐 ∙ 𝐙𝟐 (𝑡) + 𝐁
̅ 𝟐 ∙ 𝑗𝑛 (𝑡)
𝑦(𝑡) = C̅1 ∙ z1 (𝑡) + 𝐂̅𝟐 ∙ 𝐙𝟐 (𝑡) + D
̅ ∙ 𝑗𝑛 (𝑡) (38)
The transfer function of Equation (38) is:

𝐻(𝑠) = C̅1 𝑠 −1 B 𝐂̅𝟐 (𝑠 ∙ 𝐈 − 𝐀

̅1 + ⏟ ̅ 𝟐 )−𝟏 𝐁
̅𝟐 + D
̅
(39)
to be reduced by IRKA

In the IRKA, only 𝐂̅2 (𝑠𝐈 − 𝐀

̅ 𝟐 )−1 𝐁
̅ 2 will be reduced, and the remaining part stays the same.
Supposing that the transformation matrices generated in IRKA are V2 and W2, then the
transformation matrices V and W are:
1 𝟎 0 1 𝟎 0
𝐕 = [𝟎 𝐕𝟐 𝟎] , 𝐖 = [𝟎 𝐖𝟐 𝟎] . (40)
0 𝟎 1 0 𝟎 1
Then we can use Equation (26) to obtain the optimized basis function. Figure 4 summarizes the
steps of the optimal projection method developed in this paper for solving diffusion PDEs.
Energies 2018, 11, 2120 12 of 24

Figure 4. Flowchart of the optimal projection method for solving diffusion PDEs.

5. Moving-Window Filter for Electrochemical SOC Estimation

Conventional state filtering methods such as the Kalman filter are sequential estimation
algorithms which utilize the measurements at a single time point to provide the state estimation.
Figure 5a shows a schematic plot for conventional sequential state filters. As mentioned earlier, the
conventional approaches are computationally efficient, however, they may converge very slowly
and may not be robust to outliers in the measurement. Therefore, this Section develops a new state
filter algorithm based on the moving window to ensure the fast convergence of the proposed
electrochemical SOC estimation model. In this method, more data points are used for estimation of
the battery SOC. As shown in Figure 5b, a moving window of N data points (from T–N + 1 to T) is
used to provide state estimation at T–N + 1. The estimation accuracy and robustness should be
improved as more data points are involved in the estimations. The key question is how to calculate
filter gain in this method.

(a) (b)
Figure 5. (a) Conventional sequential estimation scheme, (b) moving-window state filter.

The filter gain of the moving-window filter (MWF) is derived based on the maximum
likelihood theory. The derivation of the filter gain assumes: (1) the measurement noise and process
noise are Gaussian distributed with zero mean; (2) the system equations are continuously
differentiable. Let us consider a maximum likelihood estimation problem:
𝑎𝑟𝑔𝑚𝑚𝑎𝑥 𝑃(𝑥𝑇−𝑁+1 |𝑦0 , 𝑦1 , … , 𝑦𝑇 ) (41)
𝑥𝑇−𝑁+1

Based on Bayes’ theorem [25], we have the following:

𝑃(𝑦0 , 𝑦1 , … , 𝑦𝑇 |𝑥 𝑇−𝑁+1 )𝑃(𝑥𝑇−𝑁+1 )
𝑎𝑟𝑔𝑚𝑚𝑎𝑥 𝑃(𝑦0 ,𝑦1 ,…,𝑦𝑇 )
𝑥𝑇−𝑁+1
Energies 2018, 11, 2120 13 of 24

𝑃(𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 |𝑥 𝑇−𝑁+1 ) ∏𝑇 𝑘=𝑇−𝑁+1 𝑃(𝑦𝑘 |𝑥𝑘 )𝑃(𝑥𝑇−𝑁+1 )

= argm𝑚𝑎𝑥 𝑃(𝑦0 ,𝑦1 ,…,𝑦𝑇 )
𝑥𝑇−𝑁+1
𝑃(𝑥𝑇−𝑁+1 |𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 )𝑃(𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 ) ∏𝑇𝑘=𝑇−𝑁+1 𝑃(𝑦𝑘 |𝑥𝑘 ) 𝑃(𝑥𝑇−𝑁+1 )
= argm𝑚𝑎𝑥
𝑥𝑇−𝑁+1 𝑃(𝑦0 , 𝑦1 , … , 𝑦𝑇 )𝑃(𝑥𝑇−𝑁+1 )
= argm𝑚𝑎𝑥 𝑃(𝑥𝑇−𝑁+1 |𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 ) ∏𝑇𝑘=𝑇−𝑁+1 𝑃(𝑦𝑘 |𝑥𝑘 ) (42)
𝑥𝑇−𝑁+1

Equation (42) is equivalent to a quadratic estimation problem as follows:

𝑇
𝑎𝑟𝑔𝑚𝑖𝑛 𝛷(𝑥𝑇−𝑁+1 ) = (𝒚 − 𝑯(𝑥̂𝑇−𝑁+1 )) 𝑹−𝟏 (𝒚 − 𝑯(𝑥̂𝑇−𝑁+1 ))
𝑥𝑇−𝑁+1
+(𝑥𝑇−𝑁+1 − 𝑥̂𝑇−𝑁+1 )𝑇 ∏−1 (43)
𝑇−𝑁+1 (𝑥 𝑇−𝑁+1 − 𝑥
̂𝑇−𝑁+1 )
−𝟏
In Equation (43), ∏𝑻−𝑵+𝟏 𝑪𝑻 (𝑹𝑻 + 𝐶 𝑇 ∏𝑻−𝑵+𝟏 C) is the gain of the filter. The gain has a similar
structure as the Kalman filter, however the definition of the matrices in the gain is different from
the Kalman filter. Based on first-order optimality given by:
∂Φ(xT−N+1 )
∂xT−N+1
=0 (44)

we have:
−1
𝑥𝑇−𝑁+1 = 𝑥̂𝑇−𝑁+1 + ∏ 𝑇−𝑁+1 𝐶 T (𝑅T + 𝐶 T ∏ 𝑇−𝑁+1 C) (y − H(𝑥̂𝑇−𝑁+1)) (45)

where:

𝒚 = [yT−N+1 , y1 , … , yT ]T (46)

ℎ(𝑥)
ℎ(𝑓(𝑥)
𝑯(𝒙) = ℎ(𝑓(𝑓(𝑥)) (47)
…
[ℎ(𝑓(𝑓(… 𝑓(𝑥)))]𝑁×1
𝜕𝑯(𝑥𝑇−𝑁+1 )
𝐶= , (48)
𝜕𝑥𝑇−𝑁+1

𝑥̂𝑇−𝑁+1 ∶ 𝑚𝑒𝑎𝑛 𝑜𝑓 𝑃(𝑥𝑇−𝑁+1 |𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 ) (49)

∏𝑻−𝑵+𝟏 : 𝑐𝑜𝑣. 𝑜𝑓 𝑃(𝑥𝑇−𝑁+1|𝑦0 , 𝑦1 , … , 𝑦𝑇−𝑁 ) (50)
The implementation of the MWF algorithm for SOC estimation using the proposed
electrochemical model is illustrated in Figure 6.
Energies 2018, 11, 2120 14 of 24

Figure 6. Moving-window filter implementation.

6. Case Studies
In this study, the optimal basis function is constructed from 𝚽(𝑥) = [1, 𝑟 2 , 𝑟 4 , … , 𝑟 20 ]T . First, the
Gram–Schmidt algorithm is applied to 𝚽 to obtain the orthonormal elementary basis and test
functions, 𝐔 = [𝑢1 , 𝑢2 , … , 𝑢𝑁 ]𝑇 . The model parameters, including the diffusion coefficient Ds = 3.9 ×
10–10 cm2/s and particle radius Rp = 12.5 × 10–4 cm, were obtained from [26]. These model parameters
have shown good agreement with experimental data of LixC6\LiyMn2O4 cells in [9].
To compare the developed approach with available methods, the frequency responses of
different methods are calculated, including finite difference (FD), volume averaging (VA),
projection with even-order monomials (P), and our proposed method, i.e., projection with
optimized basis functions (OP). The actual transfer function of the diffusion equation is given by
[27]:

𝑅𝑝 tanh(𝑅𝑝 √𝑠⁄𝐷𝑠 )
𝐻𝑇 (𝑠) = (51)
𝐷𝑠 (tanh(𝑅𝑝 √𝑠⁄𝐷𝑠 ) − 𝑅𝑝 √𝑠⁄𝐷𝑠 )

Figure 7 shows the frequency responses of different methods compared with the true transfer
function. At low frequency (e.g., <10–2 rad/s), all methods well match the true transfer function. As
the frequency increases, we see different performances for different methods. Overall, the FD
method with 500 nodes (FD500) shows the best performance. However, it requires solving 500
equations, which is not feasible for real-time applications. Our proposed projection method with
optimized basis functions (OP) shows the best balance between accuracy and computational burden
among those methods.
Energies 2018, 11, 2120 15 of 24

150

VA3

100

|H(i )| (dB)
VA2

True

True
FD500 FD100
50 FD100 FD500
FD5
OP5 (Proposed) P5 OP5
P5
VA2
VA3
FD5

0
10 -4 10 -2 10 0 10 2 10 4
Freq(rad/sec)

Figure 7. Frequency responses of different methods compared to the true transfer function (FD5:
finite difference with 5 nodes; FD100: finite difference with 100 nodes; FD500: finite difference with
500 nodes; P5: projection with un-optimized basis; OP5: our proposed projection method with
optimized basis; VA2: volume averaging with 2 basis functions; VA3: volume averaging with 3 basis
functions).

As can be seen in Figure 7, OP5 outperforms all other methods except for FD500. For example,
OP5 is better than FD100 in the high-frequency domain, while it only requires solving about 20
times fewer equations compared to FD100. As expected, OP5 also outperforms VA and the
projection with un-optimized basis functions (P5). The VA method developed in [17] performs well
in the low-frequency domain but generates large errors in the high-frequency domain. In addition,
VA3 with 3 basis functions is worse than VA2 with 2 basis functions, although more basis functions
are used in VA3 because the system equations generated by VA become unstable when using high-
order monomials.
More simulations were conducted to illustrate how the error changes with the number of states
(or the convergence rate). Based on this investigation, we can determine how many states are really
necessary. Figure 8 shows a convergence rate comparison of the different methods. The relative
error is calculated by the following equation in the frequency domain:
𝑤 ‖𝐻(𝑖𝜔)
− 𝐻𝑇 (𝑖𝜔)‖2
𝐸𝑟𝑟𝑜𝑟 = ∫ 𝑑𝜔 (52)
0 ‖𝐻𝑇 (𝑖𝜔)‖2
The OP shows an exponential convergence rate, while the finite difference and the projection
with original un-optimized basis present a linear convergence rate. The advantage of the proposed
OP method is obvious. From Figure 8, the accuracy of the proposed OP with three states is similar
to that of the unoptimized basis with 10 states and is two times better than FD with 10 states. With
five states, the OP is almost converged, and there is no need to add more states.
Energies 2018, 11, 2120 16 of 24

100
NP
IRKA
90 FD

Projection with unoptimized basis

80

Error measure
70
Projection with unoptimized basis

60

50 Projection with optimized basis

40

30
2 3 4 5 6 7 8 9 10
Number of states (equations)

Figure 8. Convergence comparison of different methods.

To compare the performance in the time domain, three simulations were conducted with the
time-varying flux jn(t). The first case was a 50 s discharge step followed by a 950 s rest step as
shown in Figure 9. This is to evaluate the algorithms’ performance in slow dynamics and steady-
state simulations. The benchmarking result is obtained by using a FD method with 500 nodes,
which is an accurate solution due to the high number of nodes used. The time-domain outputs
using different methods are calculated and plotted in Figure 10. In the 50 s discharge step, the
results of FD100, P5, and OP5 are on top of that of FD500, meaning they are accurate for slow
dynamics. VA2 and FD5 are less accurate as they show obvious discrepancy from FD500 in the
concentration estimation. In the rest step, after 950 s rest, the system dynamics should set down and
converge to steady state. In this step, most algorithms, including FD500, FD100, P5, OP5, and VA2,
are converged to the same steady-state solution. However, FD5 shows a different steady-state
solution, which can cause issues in actual applications, for example, errors in SOC estimation.

10 -3
12

10

8
Flux ( mole/cm 2 s)

-2

-4

0 100 200 300 400 500 600 700 800 900 1000
Time (s)

Figure 9. 50 s flux and 950 s rest applied to the surface of the particle.
Energies 2018, 11, 2120 17 of 24

4
10
2.65
FD500
FD100
2.6
FD5
OP5 (Proposed)
P5
2.55
VA2

2.5

Concentration( mole/cm3)
2.45

2.4

4
2.35 10
2.6

2.3 2.5

2.4
2.25
2.3
2.2 2.2
0 20 40 60 80
2.15
0 100 200 300 400 500 600 700 800 900 1000
Time(s)

Figure 10. Time domain outputs of different methods for slow dynamics.

The second simulation uses a pulse series with frequency of 5 Hz and duration of 100 s, which
is to evaluate the algorithms’ accuracy in the simulation for fast dynamics. The waveform of the
pulses is shown in Figure 11 (only the first second of data is plotted in order to show the data
clearly).

0.15

0.1
Flux ( mole/cm 2 s)

0.05

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time (s)

Figure 11. 5 Hz pulses applied to the surface of the particle.

The time-domain results and the errors comparing to baseline (FD500) are illustrated in Figure
12.

4 4000
10
2.8
FD100
FD500 FD5
FD100 OP5 (Proposed)
FD5 2000 P5
OP5 (Proposed) VA2
2.6
P5
VA2
0
Concentration( mole/cm3)

2.4
Error( mole/cm3)

-2000

2.2

-4000

2
-6000

1.8
-8000

1.6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-10000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time(s)
Time (s)

(a) (b)
Energies 2018, 11, 2120 18 of 24

Figure 12. Time domain outputs of different methods for 5 Hz pulses (a), errors of different methods
compared to FD with 500 nodes (b).

Like the frequency domain results presented in Figure 7, in the high-frequency simulations,
our proposed OP5 shows obvious advantages comparing to P5, VA2, and FD5 as it better matches
the results of FD500. The VA2 and FD5 perform rather poorly and show large swings from the
baseline. The third case is based on federal driving schedule (FDS) to simulate the actual loading
condition of an EV battery. The time-domain current waveform and its Fourier transform are
shown in Figure 13. As can be seen, the main frequency components of FDS are less than 2 Hz. In
Figure 7, the transfer function of FD500 overlaps with the true transfer function for frequencies < 2
Hz. Hence, here we can still use FD500 as a baseline for the comparison.

1 0.3

0.25
0
0.2
-1

Magnitude
I [A]

0.15
-2
0.1
-3
0.05
-4
0 500 1000 1500 2000 2500 3000 3500 4000 0
0 1 2 3 4 5
Time [s] Frequency(Hz)
(a) (b)
Figure 13. Federal driving schedule (FDS) profile (a), Fourier transform of FDS (b).

Figure 14 shows the simulation results of FDS based on different algorithms, and Figure 15
shows the simulation errors. Similar to the previous case, FD100 and OP5 are still the best among all
methods.

Figure 14. Time-domain simulation of different algorithms based on FDS.

Energies 2018, 11, 2120 19 of 24

Figure 15. Time-domain simulation errors of different algorithms based on FDS.

Table 2 summarizes the comparison of the root mean square error (RMSE), computational
time, and memory requirements (for the equation storage) of difference algorithms in the
simulation of the FDS. As shown, OP5 has the least RMS error among all algorithms. OP5 reduces
the computational time and memory by over 20 and 200 times, respectively, compared to FD100
and is more accurate. With the same number of states, the proposed OP5 is also much better than
the conventional P5 method, where the RMS error is reduced from 92 to 14.
The simulations above demonstrate that FD100, P5, and the proposed OP5 performs similarly
for low-frequency inputs in terms of accuracy. For high-frequency inputs, our developed method is
superior to FD100 and P5. For applications where the current profile is stationary (e.g., constant
current discharge), either FD100, P5, or OP5 should provide good performance. For dynamical
applications, for instance, EVs, where the current consists of various transients due to starting,
braking, and accelerating, the proposed OP5 method is the best among all the methods investigated
in terms of accuracy and computational efficiency.

Table 2. Comparison of algorithm performance for diffusion PDE calculation.

Algorithm Simulation Time (s)# EquationsRMS Error (µmol/cm3)Memory (Byte)

FD100 12 100 18 81,680
FD5 0.44 5 271 288
OP5 (PROPOSED) 0.42 5 14 288
P10 1.21 10 16 1768
P5 0.43 5 92 488
VA2 0.41 2 382 88

To validate the proposed MWF approach, experiments were conducted to collect battery data.
Figure 16 shows the experimental setup. Five LiFePO4 batteries were tested in this research. Arbin
BT 2000 was used to control the battery charging discharge, and the voltage, current, and
temperature data of the battery were collected.
Energies 2018, 11, 2120 20 of 24

Figure 16. Experimental setup.

Figure 17 shows the current profile used in the test, which is converted from the speed profile
of federal driving schedule (FDS) based on an EV model. This is in order to simulate what the
battery can experience under the actual driving conditions.

2 100

Speed (mph)
0
Current

50
-2

-4
0
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Time (s) Time (s)
(a) (b)
Figure 17. FDS current profile (a), FDS velocity profile (b).

The collected data is used to test the performance of the MWF with the electrochemical model.
The electrochemical model parameters were obtained from [28]. First, we examined the effects of
the window size on the SOC accuracy. Figure 18 shows the results. We can see that the SOC
estimation error reduces with the increasing of window size, especially when the window size
increases from 2 to 4. After 4, there is not much improvement. Hence, in the next paragraphs, we
will use a window size of 4 for the state estimation and compare the results with unscented Kalman
filter (UKF).

15
SOC RMS Error (%)

10

0
0 2 4 6 8 10
Window size

Figure 18. Effect of window size on the filter accuracy.

Figure 19 shows the comparison results of Cell #1 to Cell #5. The black line is the actual SOC,
which was obtained by Coulomb counting because we know the initial condition and the sensor is
accurate in the lab. The red line is the MWF, which converges to the true value in less than 2 min.
The UKF result is represented by the blue line. It takes about 30 min for the UKF to converge.
Therefore, the MWF outperforms the UKF and shows consistent improvements over the UKF.
Energies 2018, 11, 2120 21 of 24

(a) (b)

(c) (d)

(e)
Figure 19. SOC estimation comparison of cells 1 to cell 5 (a–e).

Table 3 summarizes the results of all 5 cells. For convergence analysis, only the first few
minutes are important as we want to see which filter converges faster. Therefore, in Table 3, only
the first 10 min of errors were estimated. We can see the MWF’s improvement over the UKF is
significant as the SOC error decreases from >10% to 3%. The computationally efficient
electrochemical model is used here with state filters for SOC estimation. The MWF not only
converges fast, but also demonstrates good capability to handle unit-to-unit variations, because
here the model parameters are learned from Cell #1 and then the same parameter set is applied to
different cells.

Table 3. Summary of SOC estimation results based on electrochemical model and MWF.

Root Mean Square Error (first 10 min)

Cell ID
UKF MWF
Cell 1 12.1% 2.4%
Cell 2 19.7% 2.5%
Energies 2018, 11, 2120 22 of 24

Cell 3 12.4% 1.5%

Cell 4 18.1% 2.1%
Cell 5 12.1% 1.6%

7. Contributions and Suggestions for Future Research

In this paper, an electrochemical model was presented for the state-of-charge (SOC) estimation
of lithium-ion batteries involving their internal physical and chemical properties such as lithium
concentrations. To reduce the computational complexity of solid-phase diffusion partial differential
equations (PDEs)—which is the major bottleneck for the application of any electrochemical model
in a real-time battery management system—a projection-based method with optimized
orthonormal basis functions was developed. The optimized basis function was constructed by
linear transforming a set of orthonormal elementary basis functions. The linear transformation
matrices were found by using the iterative rational Krylov algorithm. Simulation studies were
conducted to compare the performance of the developed method with other commonly used
algorithms, including finite difference, volume average, and the projection method with
conventional basis functions. The results showed that the developed method with optimized basis
functions was the best in terms of accuracy and computational complexity among all the
investigated algorithms. Particularly, in the simulation of high-frequency inputs, which is common
for EVs, the developed method reduced the root mean square error by more than five times
compared to the conventional basis functions. In addition, compared with the finite difference
method, the developed method can cut the computational time and memory requirement by over
20 times and 200 times, respectively, without losing any accuracy. The developed method can be
applied in the future to solve electrolyte diffusion PDEs, and with proper modification, it can be
extended to solve 2-D or 3-D diffusion PDEs by adding basis functions in other dimensions.
This study also presented a moving-window filter (MWF) algorithm which converged faster
than conventional state filtering methods such as Kalman filter. The filter gain was derived based
on maximum likelihood theory. The MWF was implemented with the computationally efficient
electrochemical model developed in the first part, and it was validated using experimental data
based on the federal driving schedule. It showed that the MWF reduced the convergence time from
30 min to less than 2 min compared with unscented Kalman filter for the SOC estimation.
Future research can focus on how to take the battery aging into account for SOC estimation.
For electrochemical models, it means that the model parameters have to be updated with aging, for
example, the cyclable lithium will diminish and the film resistance will grow with aging. In
addition, to avoid out-of-range estimations, constraints must be implemented for the optimization,
as a result, iterative optimizations may be necessary. How to reduce the computational complexity
of the optimization and how to ensure the convergence of the optimization for on-line parameter
estimation can be topics for future research.

Author Contributions: W.H. designed and performed the experiments; F.D. built and wrote the majority of the
manuscript text; M.P. and D.F. reviewed the manuscript and made contributions to its structure.

Funding: The authors from Heriot-Watt University acknowledge the funding of the EPSRC Offshore Robotics
for Certification of Assets hub (https://orcahub.org) (EP/R026173/1).

Conflicts of Interest: The authors declare no conflict of interest.

Energies 2018, 11, 2120 23 of 24

Abbreviations

cs solid-phase concentration
ce lithium concentration in the electrolyte
𝑐𝑠,𝑚𝑎𝑥 max concentration in the solid phase
𝑐𝑠 (𝑥, 𝑟, 𝑡) solid-phase concentration
cs,max maximum possible concentration in the solid particles of the electrode
css surface concentration in the solid particles of the electrode
De effective diffusion coefficient
Ds diffusion coefficient
𝜀𝑒 volume fraction of the electrolyte
F Faraday’s constant
𝐹𝑅𝑓 𝑗𝑛 (𝑥, 𝑡) potential drop due to the film resistance 𝑅𝑓 at the solid electrolyte interface
jn molar flux
L thickness of the electrode
reff reaction rate constant
R particle radius
Rp radius of the solid-phase particle
𝑡𝑎0 transference number for the anion
𝑣𝑘 measurement noise
𝑤𝑘 process noise

References
1. Taracson, J.; Armand, M. Issues and challenges facing lithium ion batteries. Nature 2001, 414, 359–367.
2. Breitfelder, K.; Messina, D. The Authoritative Dictionary of IEEE Standards Terms, 7th ed.; Institute of
Electrical and Electronics Engineers (IEEE) Press: New York, NY, USA, 2000; ISBN 978-0-7381-2601-2.
3. Ng, K.S.; Moo, C.-S.; Chen, Y.-P.; Hsieh, Y.-C. Enhanced coulomb counting method for estimating state-
of-charge and state-of-health of lithium-ion batteries. Appl. Energy 2009, 86, 1506–1511,
doi:10.1016/j.apenergy.2008.11.021.
4. Zhao, L.; Lin, M.; Chen, Y. Least-squares based coulomb counting method and its application for state-of-
charge (soc) estimation in electric vehicles. Int. J. Energy Res. 2016, 40, 1389–1399, doi:10.1002/er.3530.
5. Pop, V.; Bergveld, H.J.; Notten, P.H.L.; Regtien, P.P.L. State-of-the-art of battery state-of-charge
determination. Meas. Sci. Technol. 2005, 16, R93–R110, doi:10.1088/0957-0233/16/12/r01.
6. Xiong, R.; Cao, J.; Yu, Q.; He, H.; Sun, F. Critical review on the battery state of charge estimation methods
for electric vehicles. IEEE Access 2018, 6, 1832–1843.
7. Meng, J.; Luo, G.; Ricco, M.; Swierczynski, M.; Stroe, D.-I.; Teodorescu, R. Overview of lithium-ion battery
modeling methods for state-of-charge estimation in electrical vehicles. Appl. Sci. 2018, 8, 659.
8. Xing, Y.; He, W.; Pecht, M.; Tsui, K.L. State of charge estimation of lithium-ion batteries using the open-
circuit voltage at various ambient temperatures. Appl. Energy 2014, 113, 106–115,
doi:10.1016/j.apenergy.2013.07.008.
9. Xiong, R.; Yu, Q.; Wang, L.Y.; Lin, C. A novel method to obtain the open circuit voltage for the state of
charge of lithium ion batteries in electric vehicles by using H infinity filter. Appl. Energy 2017, 207, 346–
353.
10. He, H.; Xiong, R.; Fan, J. Evaluation of lithium-ion battery equivalent circuit models for state of charge
estimation by an experimental approach. Energies 2011, 4, 582–598, doi:10.3390/en4040582.
11. Lee, S.; Kim, J.; Lee, J.; Cho, B.H. State-of-charge and capacity estimation of lithium-ion battery using a
new open-circuit voltage versus state-of-charge. J. Power Sources 2008, 185, 1367–1373,
doi:10.1016/j.jpowsour.2008.08.103.
12. Cho, S.; Jeong, H.; Han, C.; Jin, S.; Lim, J.H.; Oh, J. State-of-charge estimation for lithium-ion batteries
under various operating conditions using an equivalent circuit model. Comput. Chem. Eng. 2012, 41, 1–9,
doi:10.1016/j.compchemeng.2012.02.003.
13. Feng, T.; Yang, L.; Zhao, X.; Zhang, H.; Qiang, J. Online identification of lithium-ion battery parameters
based on an improved equivalent-circuit model and its implementation on battery state-of-power
prediction. J. Power Sources 2015, 281, 192–203, doi:10.1016/j.jpowsour.2015.01.154.
14. Doyle, M.; Newman, J. The use of mathematical modeling in the design of lithium/polymer battery
systems. Electrochim. Acta 1995, 40, 2191–2196, doi:10.1016/0013-4686(95)00162-8.
Energies 2018, 11, 2120 24 of 24

15. Newman, J.; Thomas-Alyea, K.E. Electrochemical Systems, 3rd ed.; John Wiley Press: California, CA, USA,
2004.
16. Ramadesigan, V.; Northrop, P.W.C.; De, S.; Santhanagopalan, S.; Braatz, R.D.; Subramanian, V.R.
Modeling and simulation of lithium-ion batteries from a systems engineering perspective. J. Electrochem.
Soc. 2012, 159, R31–R45, doi:10.1149/2.018203jes.
17. Subramanian, V.R.; Diwakar, V.D.; Tapriyal, D. Efficient macro-micro scale coupled modeling of
batteries. J. Electrochem. Soc. 2005, 152, A2002–A2008.
18. Mayhew, C.; He, W.; Kroener, C.; Klein, R.; Chaturvedi, N.; Kojić, A. Investigation of projection-based
model-reduction techniques for solid-phase diffusion in li-ion batteries. In proceedings of the American
Control Conference, Portland, OR, USA, 4–6 June 2014.
19. Chaturvedi, N.A.; Christensen, J.F.; Klein, R.; Kojic, A. Approximations for partial differential equations
appearing in li-ion battery models. In Proceedings of the ASME Dynamic Systems and Control
Conference, Palo Alto, CA, USA, 21–23 October 2013.
20. Corno, M.; Bhatt, N.; Savaresi, S.M.; Verhaegen, M. Electrochemical model-based state of charge
estimation for Li-ion cells. IEEE Trans. Control Syst. Technol. 2015, 23, 117–127.
21. Han, J.; Kim, D.; Sunwoo, M. State-of-charge estimation of lead-acid batteries using an adaptive extended
kalman filter. J. Power Sources 2009, 188, 606–612, doi:10.1016/j.jpowsour.2008.11.143.
22. Santhanagopalan, S.; White, R.E. State of charge estimation using an unscented filter for high power
lithium ion cells. Int. J. Energy Res. 2009, 34, 152–163, doi:10.1002/er.1655.
23. Meier, L.; Luenberger, D.G. Approximation of linear constant systems. IEEE Trans. Autom. Control 1967,
12, 585–588.
24. Gugercin, S.; Antoulas, A.C.; Beattie, C.A. H2 model reduction for large-scale linear dynamical systems.
SIAM J. Matrix Anal. Applications 2008, 30, 609–638.
25. Lee, P.M. Bayesian Statistics: An Introduction; John Wiley & Sons: West Susses, UK, 2012.
26. Jacobsen, T.; West, K. Diffusion impedance in planar, cylindrical and spherical symmetry. Electrochim.
Acta 1995, 40, 255–262, doi:10.1016/0013-4686(94)e0192-3.
27. He, W. Battery State of Charge Estimation Based on Data-Driven Models with Moving Window Filters
and Physics-Based Models with Efficient Solid-phase Diffusion PDEs Solved by the Optimized Projection
Method. Ph.D. Thesis, University of Maryland, College Park, MD, USA, 2018.
28. Zhang, L.; Wang, L.; Hinds, G.; Lyu, C.; Zheng, J.; Li, J. Multi-objective optimization of lithium-ion
battery model using genetic algorithm approach. J. Power Sources 2014, 270, 367–378,
doi:10.1016/j.jpowsour.2014.07.110.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons
Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).