Alexandria Engineering Journal (2022) 61, 11277–11290

H O S T E D BY
Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

Data-Driven modeling for Li-ion battery using

dynamic mode decomposition
Mohamed A. Abu-Seif a,*, Ayman S. Abdel-Khalik a, Mostafa S. Hamad b,
Eman Hamdan c, Noha A. Elmalhy a

a
Department of Electrical Engineering, Alexandria University, Alexandria 21544, Egypt
b
Research and Development Center Arab Academy for Science, Technology and Maritime Transport, Al Alamein, Egypt
c
Department of Marine Engineering Technology, Arab Academy for Science, Technology and Maritime Transport, Alexandria
1029, Egypt

Received 26 February 2022; revised 8 April 2022; accepted 21 April 2022

Available online 14 May 2022

KEYWORDS Abstract Lithium-ion (Li-ion) batteries are the workhorse of energy storage systems in electric
Lithium-ion battery (Li-ion); vehicles (EVs) due to their high energy density and desirable characteristics. To obtain an optimized
Dynamic mode decomposi- and safe operation, a battery management system (BMS) should be implemented to provide the
tion (DMD); main safety features based on the estimation of different battery states, which entails an accurate,
Battery management system fast, and adaptive mathematical model. In this paper, a data-driven linear model using dynamic
(BMS); mode decomposition is proposed. The proposed modeling procedure bridges the gap between
State estimation abstracted models (e.g., circuit-based models) and empirical models (e.g., data-driven models) of
Li-ion batteries. Unlike the abstracted models, the proposed model does not impose any assump-
tions on what the model should be, and also considers the battery as a black box similar to the
empirical techniques, and yet gives an interpretable linear model in state space form. In order to
generate this model, only one discharge cycle that contains high dynamic content is required,
and no other specialized tests are required. Two models are proposed for the Li-ion battery, namely,
a model-based estimator based on Kalman filter for BMS purposes and a piecewise model for off-
line simulation. The results are verified experimentally via a lab scale prototype.
Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria
University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction

Due to the high energy density and long life of lithium-ion (Li-
* Corresponding author. ion) batteries, they become the cornerstone of energy storage
E-mail address: es-mohamedashrf2021@alexu.edu.eg (M.A. Abu-Seif). systems in electrical vehicles (EVs) [1], mobile devices [2,3], air-
Peer review under responsibility of Faculty of Engineering, Alexandria crafts power systems [4], and smart grid applications. In order
University. to optimize the battery usage and ensure a safe operation, a
https://doi.org/10.1016/j.aej.2022.04.037
1110-0168 Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
11278 M.A. Abu-Seif et al.

battery management system (BMS) needs to be implemented. necessarily considering the physical meaning of this model
BMS aims to extend the battery lifetime, ensures that the bat- [13,20,23]. An example of these models is the neural
tery is operating in the safe operation area (SOA), and moni- networks-based model introduced in [21,22]. Another example
tors the main battery states during operation, such as of using neural networks in modeling heat transfer treatment
voltage, current, temperature, state of charge (SoC), and avail- illustrated in [40]. These models require large volume of data
able capacity. Battery management techniques could be for training and validation. Despite their high accuracy, those
designed in various complexity levels from simple to compre- models are computationally and memory expensive [13].
hensive management [5]. Simple management only perform a Dynamic mode decomposition (DMD) has originated in
very basic functions like states monitoring, SoC estimation the fluid mechanics field as a method to decompose complex
and fault detection. On the other hand, a comprehensive man- flows into a simple representation based on spatiotemporal
agement necessitates taking a lot of factors into consideration, coherent structures called modes [24], with a corresponding
such as real time monitoring of the battery, preventing over eigenvalue for each mode that defines the growth/decay rates
charge/discharge, and extending battery life through implying and oscillation frequencies of this mode. DMD is closely
constrains on the usage of the battery. Another major usage of related to Koopman spectral analysis [25]. DMD can be con-
the mathematical models is to design the battery packs for a sidered as a numerical approximation for Koopman spectral
specific device or vehicle, whether to model each battery alone analysis. Therefore, DMD is applicable to nonlinear systems
as a single system or to deal with the battery pack as a single [26,27], which is the case for the Li-ion batteries.
dynamical system. In this paper, two linear discrete models for the li-ion bat-
In order to have an efficient model for either the BMS or tery using (DMD) are proposed. The first model is intended to
offline simulation, a mathematical model of the battery should be utilized by the BMS in a model-based configuration based
be obtained. This model needs to fulfill the following on optimum estimator represented by a linear Kalman filter,
requirements: and the second model is a piecewise linear one intended to
be used in offline simulation of battery packs. The proposed

Accurate states estimation. models meet the following requirements:

Computationally efficient to run on low-cost embedded
microprocessors.
Simple and accurate models in discrete state space form.

Adaptive and easy to be regenerated to compensate for the
No prior knowledge of the battery model is required since it
battery degradation. is a data-driven approach.

Requires only one dataset of dynamic discharge/charge
To the previous end, several models have been proposed in cycle.
[6–16], which can be classified into three main categories:
Computationally and memory efficient.

1. Physical/electrochemical models, Using DMD, a hybrid approach of abstracted and empiri-

2. Abstracted models, cal modeling techniques is presented by introducing a model
3. Empirical models. that is data driven and yet be represented in a state space form.
Consequently, the processes within the model are inter-
Physical models are the most accurate model but with a sig- pretable. DMD allows to obtain the best fit linear model for
nificant complexity. These models are a low-level representa- what is originally a nonlinear process, like the battery. To this
tion of the battery with a high accuracy. They describe the end, it is needed to augment the states of the battery (voltage
electrochemical process happening inside the cell, such as ther- and SoC) with kernel functions. These kernel functions are
modynamics, active species kinetics, and transport phenomena chosen to be rich enough, represented by linear model, and
[17]. Such models offer an insight on what really happens somehow uncorrelated to each other to avoid the redundancy
inside the battery but at the cost of high computational power in order to obtain a simpler model. The datasets in [28] are
[15]. Upon this fact, these models are not commonly used in used in order to generate the models and validate the proposed
the monitoring algorithms but in the development of the bat- approach. A test bench has also been constructed to validate
tery design itself [7]. the introduced models.
Abstracted models give an alternative methodology to This paper is organized as follows; In section II, the steps of
model the battery using abstracted elements that represents the dynamic mode decomposition algorithm and the choice of
the physical behavior of the battery. The most used abstracted system states have been introduced. In section III, the model
models are the circuit-based models [13,14], where the chemical training and validation results are given. Section IV introduce
processes like diffusion and polarization are modeled using RC the hardware setup for further validation of the proposed
pairs [19]. These models offer lower computational burden model.
than the physical models without missing much of the battery
dynamics. However, look-up tables are still required in order
to meet experimental data [20], while specific tests are essential 2. Dynamic mode decomposition with control
to find the relation between open circuit voltage and SoC [39].
Mostly, Kalman filters are used to estimate the states of the DMD is a data-driven, equation free architecture that is able
battery using the circuit-based models [18,39], this fusion to reconstruct the underlying dynamics of the system from
between give the optimum results in terms of the measurement the snapshots of the measurements [29]. DMD gives an
error and the model uncertainties [37]. approximation for the leading eigen values, eigen modes and
Empirical models on the other hand deals with the battery eigen functions of the Koopman operator [30]. Koopman
as a black box, modeling an input–output correlation without operator is a linear but infinite dimensional operator that
Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11279

governs the evaluation of scalar observables that define a state However, it can successfully model the SoC response as illus-
space for an autonomous system [25,31]. By using the DMD, trated in Fig. 1(a). The system state space model is represented
Koopman operator can be approximated to obtain a linear in (7) as:
model of the battery’s observables.      " #
The battery can be considered as an autonomous system Vkþ1 1 0 Vk 5:71  105
¼ þ Ik ð7Þ
subjected to external forcing ‘‘the current”. Since the battery SoCkþ1 0 1 SoCk 2:75  105
is subjected to an external forcing, it cannot be modeled
directly using the DMD. Since in actuated systems, DMD is It can be noticed that the states (voltage and SoC) are
incapable of producing input–output models. Moreover, the decoupled. From this output, it can be assumed that the
dynamics can be corrupted by the external forcing [25,31]. DMDc output is a model similar to the Rint model [36], which
Therefore, a variation of DMD called dynamic mode decom- models the battery as a voltage source and an equivalent inter-
position with control (DMDc) introduced by Proctor et. al nal resistance represented by the element in the first row of the
[31] is used. DMDc uses all the advantages of DMD and gives B matrix. The element in the second row represents the sam-
additional innovation of being able to differentiate between the pling time, which is 0.1 sec divided by 3600 that represents
actuation effect and the underlying dynamics, giving the possi- the capacity in the ampere hour unit, which is the same as cou-
bility of deriving the system on the following state space lomb counting technique.
model: As mentioned, the system cannot properly estimate the
voltage since the battery voltage dynamics are not modelled.
xkþ1 ¼ Axk þ Buk ð1Þ Accordingly, the model can be further optimized by augment-
where: ing the system states with the Kernels [32] of the main two
states (voltage and SoC). Referring to Koopman theorem,
x 2 Rn u 2 Rp A 2 Rnn B 2 Rnp
any nonlinear dynamical system can be represented as a linear
The subscript k represents the sampling instance of the sys- but infinite dimensional system of observables [33], these
tem, each measurement xk represents a snapshot of data. observables could be the states of the system in addition to
in time k, and uk represents the system input. To perform functions of these states. The new state space can be repre-
the proposed DMDc algorithm, the data is arranged in a sented as a new basis that allows us to model the nonlinear
0
sequence of snapshots of two state matrices X and X , where: dynamics of the battery in a linear fashion. Taking this into
2 3 account, a linear system of the states and the Kernels can be
j j  j constructed, which are able to fully model the battery nonlin-
6 7
X ¼ 4 x1 x2    xm1 5 ð2Þ earities. However, the choice of the augmented states is not an
j j  j easy task. Typically, the augmented basis needs to be propa-
gated linearly, stable, rich, as small as possible, and uncorre-
2 3
j j  j lated to avoid redundancy in the basis. The proposed
60 7 approach to check whether the selected basis is rich or redun-
X ¼ 4 x2 x3  xm 5 ð3Þ
dant is that an initial basis is constructed that contains all the
j j  j
proposed Kernel functions. The choice of those functions
and a sequence of the input data ;, where: comes from the physical intuition of the system. Basically,
2 3 polynomial and gaussian radial Kernels are used. Then, a
j j  j
6 7 covariance matrix of the basis is constructed to get the vari-
Y ¼ 4 u1 u2    um1 5 ð4Þ ance and the covariance of each state. Based on the obtained
j j  j matrix elements, it can be decided whether the chosen basis
is rich enough or not by comparing the diagonal values of
where m represents the total number of snapshots, X0 is one the matrix, which represent the variance of each state. Whereas
snapshot ahead of X, and ! represents the input sequence of
the data redundancy is checked by comparing the off-diagonal
the actuation. Hence, (2)–(4) can be written as: values. Typically, the diagonal element should have high vari-
0
X ¼ AX þ B! ð5Þ ance values, while low covariance values are required for the
off-diagonal elements. This statistical representation gives an
After constructing the data in the mentioned format, the insight about the basis. Then these new states basis are used
DMDc algorithm can be used and these modes and eigen values and fed to the DMDc algorithm to check for other criteria.
of the model are obtained. Finally, the state transition matrix Based on the mean square error (MSE) of each basis, it can
A and input matrix B are constructed. be assumed whether this basis is the optimum basis or not.
These steps are applied in an iterative sequence until the best
2.1. The choice of system states basis is obtained. The followed criteria are as follows:

If the battery is considered as a two-state system, it can be rep-
Basis stability.

resented as following:
Low dimensionality.
   
High variance values across the diagonal.
Vkþ1 Vk
¼A þ BIk ð6Þ
Low covariance values across the off diagonals.
SoCkþ1 SoCk

Low MSE of the generated model.
After training the model, the reconstructed data of this sys-
tem shown in Fig. 1 depicts that modelling the system using Table.1 illustrates some of the proposed basis and the val-
these states cannot properly capture the voltage dynamics. ues of the covariance matrix. The first column represents a
11280 M.A. Abu-Seif et al.

Fig. 1 Two variable system response. (a) Training voltage (black) and model response (red). (b) Training SoC (black), and model
response (red).

Table 1 Covariance colour map matrix for each basis.

2 3
proposed basis, and the second column contains the corre- V
sponding covariance matrix represented as a colour map to 6 SoC 7
6 7
6 1 7
aid the visualization of the covariance matrices. Based on the 6 2 7
mentioned criterion, it is noticed that system 4 has a high 4 V 5
:5
covariance value along the off-diagonal elements, so it was e V2 kþ1
2 3
not considered. And owing to the high dimensionality of sys- 0:936 1:15  102 0:587 0:305
tem 2 relative to the variance and covariance values, it was 6 3:63  105 1:98  105 7
6 0 1 7
not considered. For the rest of the basses, they meet the spec- ¼6 7
4 2:57  10 3
4:71  10 4
1 1:24  102 5
ified criterion. And it was found that coupling two different
state variables together in a single kernel as illustrated in sys- 1:23  103 2:26  104 1:14  102 1
2 3 2
tem 3 gives a high variance value, but it also leads to an unsta- 3 3
V 1:03  10
6 SoC 7 6
ble eigen values, so system_3 was not considered. Based on 6 7 6 2:77  105 7 7
those findings, system 1 and system 5 were considered, and sys- 6 7
6 12 7 þ 6 5 7
Ik ð8Þ
4 V 5 4 4:23  10 5
tem 1 resulted into lower training and validation MSE values,
:5
so system 1 was considered. e V2 k 2:03  105
Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11281

3. Building and validation the model 3.2. Model validation

After choosing the suitable basis for the data, the 25 °C data- The model Validation is evaluated on the mixed dataset under
sets will be considered so the model will assume a fixed temper- the same temperature of 25°. Table 2 shows the validation
ature. In order to compensate for the temperature change, MSE for each dataset using the evaluated model. Fig. 3 shows
multiple models can be generated for each temperature seg- the validation results of the Mixed_2 dataset. As seen in vali-
ment, or the model may be trained with a dataset that has a dation results in Fig. 2. (a)  3. (a), the model performs well
dynamic temperature change, then this change can be consid- except for the last region of the voltage profile. In order to
ered as an external input to the model. develop a model that fully cover the whole voltage, a piecewise
The model is constructed using the Urban Dynamometer model based on the value of the SoC value is introduced in the
Driving Schedule (UDDS) dataset [28] as it represents high following subsection. This approach gave the best results and
dynamical content throughout the loading profile of the bat- the lowest MSE, and it would be suitable to model separate
tery, which will allow the algorithm to discover the battery batteries in a battery pack in offline simulations.
dynamics. All singular values have been included in the pro-
posed model without any truncation, which yields better vali- 3.3. Piecewise model
dation results. The Mixed datasets in [28] are used to validate
the model since they contain high dynamical profile of data. Modeling the battery using DMD results into a linear model
The PyDMD package [34] in Python is used. The generated that describes the dynamics of the battery in a linear form.
system state space model is given by (8). And in order to enhance the accuracy of this linear model, a
piecewise model can be constructed by dividing the battery
3.1. Model evaluation model into a piecewise SoC-variant model. As illustrated in
Fig. 2, the error magnitude in voltage estimation is higher in
Since basis_1 is selected, the system dimension will be four. the begging and the ending of the SoC range, while the error
Therefore, the obtained state transition matrix A, given by is minimal in the middle of the SoC region. This can be inter-
(8), will have four eigen values and eigen vectors. The eigen preted that the generated model is biased towards the middle
values are real and do not have any imaginary component, of the SoC region, since it is the region that contains higher
which matches the physical intuition of the battery. The first number of data points, so the generated model will be biased
three eigen values are on the real axis inside the unit circles giv- to model this region with higher accuracy than the other
ing an exponential decay effect, the fourth mode is on the unit SoC regions. Another reason for this results can be obtained
circle. The results of the training and training results are shown by examining the open circuit voltage (OCV) of the battery
in Fig. 2, The MSE of the voltage is 0:0049, and the MSE of in Fig. 4, it can be noticed that the OCV is approximately flat
the SoC is 4:57  105 . across the mid SoC range, while the OCV mimics a non-linear

Fig. 2 Training profile (UDDS dataset) and model output. (a) Training voltage data (black) and the single model output (green). (b)
Training SoC data (black) and the model SoC output (green). (c) and (d) MSE for the voltage and SoC, respectively.
11282 M.A. Abu-Seif et al.

Table 2 MSE of single and piecewise models.

One state space model Three states space piecewise model Four states space piecewise model
Dataset Voltage MSE (mV) SoC MSE (Ah) Voltage MSE (mV) SoC MSE (Ah) Voltage MSE (mV) SoC MSE (Ah)
551_Mixed1 2.18 3.04 0.90 2.4 3.28 0.013

106 106 106

Fig. 3 Validation profile (Mixed_2 dataset) results. (a) Validation voltage data (black) and the single model output (green). (b)
Validation SoC data (black) and the model SoC output (green). (c) and (d) MSE for the voltage and SoC, respectively.

Fig. 4 Segments of the system.

Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11283

behavior across the begging and the ending of the SoC which given in appendix A. Each pair is generated using a subset of
goes along with the error results. Based on the model bias the data profile with an interaction segment from the next or
towards the mid SoC range and the non-linearity of the the last segment. For example, the first segment is trained
OCV, it was found that dividing the battery model into three using data in a segment from 0 Ah to 0.15 Ah SoC, and
separate SoC-variant models while each model will represent the transition happens at the 0.1 Ah SoC instance. Subse-
the battery in a certain SoC segment and specified with its quently, the second segment is trained from 0.1 Ah to 2.2
own SoC range resulted into higher accuracy than the single Ah, and the transition happens at 2.1 Ah SoC instance. This
state space model. This approach can even go further to divide overlap improves the transition from segment to segment lead-
the battery into more than three segments to gain a higher ing to a more robust piecewise model. Fig. 5 shows the perfor-
accuracy model, but at the cost of more state space matrices mance of the piecewise model comparing to the single states
to represent this segments. Table.2 shows the error of some system and Figs. 6-7 show the validation results and Table.3
of the proposed piecewise models that represents single, three, compares the results of the single state models and the piece-
and four segments. wise model. It is clear that the proposed piecewise model is
As illustrated in Fig. 4, the battery has three distinctive capable of providing higher accuracy over the whole range
regions. The first region is in the beginning of the SoC range of voltage with a lower error than the single state model. This
till 0.1Ah. Then, the mid SoC region extends up to 2.2Ah, will not deny that the single state model is very acceptable,
while the third region represents the SoC values higher than especially when the estimation is fused with the measurements
2.2Ah. The size and the boundary of each segment is chosen using a Kalman filter. The procedure of kalman filter is illus-
such that the validation results of the generated model have trated in Table 4.
the lowest MSE. Based on the constant discharge curve in
Fig. 4, initial guesses of each segment can be derived to deduce 3.4. Kalman filter formulation
the optimum boundaries as follows.
8 Kalman filters are used to estimate states based on linear
>
< fA1 ; B1 g; SoC < 0:1Ah
dynamical systems in state space format. In order to obtain
fA; Bg ¼ fA2 ; B2 g; 0:1Ah  SoC < 2:2Ah ð9Þ the best estimate of the states at low computational power,
>
:
fA3 ; B3 g; 2:2Ah  SoC Kalman filter is used to fuse the model estimates and the mea-
surements of the sensors [38]. If we considered the battery
The piecewise model consists of three pairs of state transi- mode given by,
tion matrices fA1 ; A2 ; A3 g and three input matrices
fB1 ; B2 ; B3 g as shown in (9). The derived piecewise system is xkþ1 ¼ Axk þ Buk þ wk ð10Þ

Fig. 5 Training profile (UDDS dataset) of single model and piecewise model results. (a) Dataset voltage (black) and the single model
output (green) and the piecewise model (red). (b) Dataset SoC (black) and the single model output (green) and the piecewise model (red).
(c) and (d) MSE for the voltage and SoC, respectively.
11284 M.A. Abu-Seif et al.

Fig. 6 Validation profile (Mixed_1 dataset) of single model and piecewise model results. (a) Dataset voltage (black) and the single model
output (green) and the piecewise model (red). (b) Dataset SoC (black) and the single model output (green) and the piecewise model (red).
(c) and (d) MSE for the voltage and SoC, respectively.

Fig. 7 Validation profile (Mixed_4 dataset) of single model and piecewise model results. (a) Dataset voltage (black) and the single model
output (green) and the piecewise model (red). (b) Dataset SoC (black) and the single model output (green) and the piecewise model (red).
(c) and (d) MSE for the voltage and SoC, respectively.
Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11285

Table 3 MSE of Mixed datasets.

Single model Piecewise model
Dataset Voltage MSE (mV) SoC MSE (Ah) Voltage MSE (mV) SoC MSE (Ah)
106 106
551_Mixed1 2.18 3.04 0.90 2.4
551_Mixed2 1.77 0.22 0.60 0.42
552_Mixed3 1.80 8.24 0.83 7.43
552_Mixed4 1.82 0.35 0.74 0.13
552_Mixed5 1.74 0.62 0.80 0.24
552_Mixed6 1.93 0.33 0.63 0.12
552_Mixed7 2.01 0.60 0.66 0.34
552_Mixed8 1.87 0.50 0.60 5.4

vk  Nð0; RÞ ð13Þ
Table 4 Kalman filter steps. Kalman filter consists of two main steps, prediction, and
Prediction update. The algorithm can be summarized in Table 3. Since
Predicted state estimation b
x k ¼ Axk1 þ Buk1 the SoC cannot directly be measured using a sensor, the SoC
Predicted error covariance Pk ¼ APk1 AT þ Q measurement is fed to Kalman filter by applying the coulomb
counting techniques given by (14). The equation is constructed
Update such that the SoC output is in the Ah unit.
Measurement residual yk ¼ zk þ Cbxk Z t
Computing Kalman gain  1
Kk ¼ Pk CT R þ CPk HT SoCðtÞ ¼ SoC0 þ Idt ð14Þ
Update state estimation xk ¼ b
x k þ kk yk t0
Update error covariance Pk ¼ ðI  Kk CÞPk
where SoC0 represents the initial SoC of the battery. The
results of the Kalman filter model are shown in Figs. 7-8. In
order to simulate the real conditions of the measurement
devices, the measurements of both the voltage and current
with an uncertainty w that follows gaussian distribution are injected with a zero mean gaussian noise. The voltage noise
with zero mean and variance Q: has a covariance of 0.05 and the current covariance is 0.01.
Also, we introduced error in the initial conditions to check
wk  Nð0; QÞ ð11Þ the convergence of the filter. As illustrated in the Figs. 9-11,
and by fusing the process model with the measurement states converge to the true values. The results of the two data-
model z½k : sets are shown in Table 5.

zk ¼ Cxk þ vk ð12Þ 3.5. Adaptive model

where, C represents the measurement matrix, and vk is a zero
mean gaussian distribution measurement noise with covari- Since the generated model is basically a data based one, it gives
ance R: the possibility to regenerate a new model based on the profile

Fig. 8 Kalman filter block diagram.

11286 M.A. Abu-Seif et al.

Fig. 9 Kalman filter Validation profile (mixed_5 dataset). (a) shows clean voltage (black) and the noisy voltage measurements (blue) and
Kalman filter voltage estimation (red), (b) shows clean SoC (black) and the noisy SoC measurements (blue) and Kalman filter SoC
estimation (red).

Fig. 10 Kalman filter Validation profile (mixed_8 dataset). (a) shows clean voltage (black) and the noisy voltage measurements (blue)
and Kalman filter voltage estimation (red), (b) shows clean SoC (black) and the noisy SoC measurements (blue) and Kalman filter SoC
estimation (red).

Fig. 11 Kalman filter Validation profile (HPPC dataset). (a) shows clean voltage (black) and the noisy voltage measurements (blue) and
Kalman filter voltage estimation (red), (b) shows clean SoC (black) and the noisy SoC measurements (blue) and Kalman filter SoC
estimation (red).
Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11287

Table 5 MSE results of Kalman filter.

Dataset Voltage MSE (mV) SoC MSE (Ah)105
551_Mixed1 0.87 5.90
551_Mixed2 0.93 5.93
552_Mixed3 0.98 5.94
552_Mixed4 0.94 5.95
552_Mixed5 0.86 5:91
552_Mixed6 0.98 6.01
552_Mixed7 0.98 5.95
552_Mixed8 0.92 5.96
549_HPPC 1.31 6.16

of the battery over time. The main advantage of this procedure Fig. 12 Hardware setup.
is to avoid the side effects of battery degradation [35]. This will
lead to a more robust model against changes over time, and
this technique can be generalized to generate an approximate The model generation data and Kalman filter estimations
degradation model to track capacity degradation of the based on this model are shown in Figs. 13 and 14, respectively.
battery. The battery used in the test bench and the one used to generate
In order to be able to generate a new model, mainly there the data in [28] are not of the same type. The main reason to
are two problems. The first problem is to decide whether the choose this approach is to give an adequate reason to leverage
data profile is dynamic enough for the DMDc algorithm to the flexibility and the ability of the proposed approach of gen-
generate a model that captures the battery dynamics. The sec- erating a model of the battery using only one single dynamic
ond problem is how to train the model online. data profile, giving the possibility for this technique to be gen-
To overcome the first problem, a function of the gradient of eralized to any other type of Li-ion batteries or even to be uti-
the states rxk is constructed, and any numerical differentia- lized to model battery packs as long as the training data is
tion introduced in [36] can be used especially if the data are clean and dynamic enough.
noisy. Based on the absolute sum of this function represented
by equation (15) over a charge/discharge profile which is called 5. Conclusion
the dynamic number of the set, it can be decided whether this
data is dynamic enough or not.
In this paper, a new approach to model the Li-ion batteries is
X
n
proposed based on one single profile of data using dynamic
jrxk j ð15Þ
k¼1
mode decomposition with control (DMDc). Two models were
proposed: a single state space model and a piecewise state
where n can be the maximum buffer size of the memory on the space model. These models provide the following advantages:
BMS.
Table 6 illustrates the dynamical number of some the data-
Do not require any prior knowledge of the battery model or
sets used in this paper. Since the UDDS results to the highest any specific test, only a dynamic profile data of the battery.
dynamic score, it was selected for the model training. The sec-
Data driven technique yet they give an interpretable linear
ond problem can be solved by broadcasting the data to a state space model.
cloud, and the training will be done on this cloud and the
Can be used for offline simulation (piecewise model) or to
model will be fed back to the BMS microprocessor. be used in a model-based architecture with Kalman filter
(single state space model), either way the model is computa-
4. Experimental implementation tionally and memory efficient.

Test bench setup was conducted in order to evaluate the pro- This approach can be used to model the battery in a more
posed technique. The test bench consists of two series Li-ion general context to include the temperature effect on the battery
batteries LG BM26 18650, electronic load, and digital oscillo- dynamics, the degradation of the battery over time, and online
scope. The current is measured using Fluke i30 current probe. estimation of battery states to develop an adaptive model, ad
Fig. 12 shows the complete hardware setup. those approaches will be discussed in a future work.

Table 6 Dynamic number for the voltage of the datasets.

551_UDDS 551_Mixed1 551_Mixed3 551_Mixed6 551_Mixed8
7325.1 6268.9 6605.7 6104.3 6371.9
11288 M.A. Abu-Seif et al.

Fig. 13 Model data and validation results. (a) Voltage (black), the single model output (green) and the piecewise model (red). (b) SoC
(black) and the model SoC output (green). (c) and (d) The MSE for the voltage and SoC, respectively.

Fig. 14 Kalman filter model validation results. (a) Voltage (black), the single model output (green) and the piecewise model (red). (b)
SoC (black) and the model SoC output (green). (c) and (d) The MSE for the voltage and SoC, respectively.

Financial Disclosure interest (such as honoraria; educational grants; participation in

speakers’ bureaus; membership, employment, consultancies,
stock ownership, or other equity interest; and expert testimony
ITIDA’s ITAC collaborative funded project Fund category:
or patent-licensing arrangements), or non-financial interest
Advanced Research Projects (ARP) ARP2020.R28.18.28.18.
(such as personal or professional relationships, affiliations,
knowledge or beliefs) in the subject matter or materials dis-
Declaration of Competing Interest
cussed in this manuscript.].
The authors declare the following financial interests/per-
sonal relationships which may be considered as potential com- Acknowledgements
peting interests: [The authors whose names are listed
immediately below certify that they have NO affiliations with ITIDA’s ITAC collaborative funded project Fund category:
or involvement in any organization or entity with any financial Advanced Research Projects (ARP) ARP2020.R28.18.28.18
Data-Driven modeling for Li-ion battery using dynamic mode decomposition 11289

Appendix A [9] Y. Zheng, M. Ouyang, L. Lu, J. Li, Understanding aging

mechanisms in lithium-ion battery packs: From cell capacity loss
to pack capacity evolution, J. Power Sources 278 (2015) 287–
The state space matrices of the three piecewise systems are 295.
given by, [10] C. Lin, H. Mu, R. Xiong, J. Cao, Multi-model probabilities
2 3 based state fusion estimation method of lithium-ion battery for
0:882 1:76  102 1:55 0:593
6 electric vehicles: State-of-energy, Appl. Energy 194 (2017) 560–
6 0 1 2:48  105 8:39  106 7
7 568.
A1 ¼ 6 7;
4 3:76  103 5:72  104 1 1:90  102 5 [11] X. Tang, X. Mao, J. Lin, B. Koch, Li-ion battery parameter
1:82  103 2:77  104 2:38  102 1 estimation for state of charge, in: Proceedings of the 2011
2 3 American Control Conference, 2011, pp. 941–946.
1:59  103 [12] Y. Zou, X. Hu, H. Ma, S.E. Li, Combined State of Charge and
6 2:77  105 7
6 7 State of Health estimation over lithium-ion battery cell cycle
B1 ¼ 6 7 ðA:1Þ
4 5:19  105 5 lifespan for electric vehicles, J. Power Sources 273 (2015) 793–
803.
2:51  105 [13] I.B. Espedal, A. Jinasena, O.S. Burheim, J.J. Lamb, Current
2 3 trends for state-of-charge (SoC) estimation in Lithium-ion
0:882 2:59  102 0:604 0:536 battery electric vehicles, Energies 14 (11) (2021) 3284.
6 4:39  105 3:26  105 7
6 0 1 7 [14] H. He, R. Xiong, J. Fan, Evaluation of lithium-ion battery
A2 ¼ 6 7;
4 5:04  10 3
1:12  104
1 2:30  102 5 equivalent circuit models for state of charge estimation by an
experimental approach, Energies 4 (4) (2011) 582–598.
2:24  103 5:38  104 1:23  102 1 [15] H. Hinz, Comparison of lithium-ion battery models for
2 3
2:07  103 simulating storage systems in distributed power generation,
6 2:77  105 7 Inventions 4 (3) (2019) 41.
6 7
B2 ¼ 6 7 ðA:2Þ [16] Y. Ye, H. Ma, J. Yang, Research on accurate model of lithium
4 8:99  105 5
battery, in 2020 Chinese Control And Decision Conference
4:32  105 (CCDC), 2020.
2 3 [17] W. He, M. Pecht, D. Flynn, F. Dinmohammadi, A Physics-
0:840 2:87  102 1:76 7:95 Based Electrochemical Model for Lithium-Ion Battery State-of-
6 3:43  104 1:66  104 7 Charge Estimation Solved by an Optimised Projection-Based
6 0 1 7
A3 ¼ 6 7; Method and Moving-Window Filtering, Energies 11 (8) (2018)
4 8:48  103 1:54  104 1 4:23  102 5 2120.
4:05  103 7:36  104 4:46  102 1 [18] B. Rzepka, S. Bischof, T. Blank, Implementing an extended
2 3
1:03  103 Kalman filter for SoC estimation of a Li-ion battery with
6 2:77  105 7 hysteresis: A step-by-step guide, Energies 14 (13) (2021) 3733.
6 7
B3 ¼ 6 7 ðA:3Þ [19] L.W. Sommer, P. Kiesel, A. Ganguli, A. Lochbaum, B. Saha, J.
4 4:23  105 5 Schwartz, C.-J. Bae, M. Alamgir, A. Raghavan, Fast and slow
2:03  105 ion diffusion processes in lithium ion pouch cells during cycling
observed with fiber optic strain sensors, J. Power Sources 296
(2015) 46–52.
References [20] F. Saidani, F.X. Hutter, R.-G. Scurtu, W. Braunwarth, J.N.
Burghartz, Lithium-ion battery models: a comparative study
[1] C. Fellner, J. Newman, High-power batteries for use in hybrid and a model-based powerline communication, Adv. Radio Sci.
vehicles, J. Power Sources 85 (2) (2000) 229–236. 15 (2017) 83–91.
[2] T. Iwahori et al, Development of lithium ion and lithium [21] M.A. Hannan, M.S.H. Lipu, A. Hussain, P.J. Ker, T.M.I.
polymer batteries for electric vehicle and home-use load leveling Mahlia, M. Mansor, A. Ayob, M.H. Saad, Z.Y. Dong, Toward
system application, Electrochim. Acta 45 (8–9) (2000) 1509– enhanced state of charge estimation of lithium-ion batteries
1512. using optimized machine learning techniques, Sci. Rep. 10 (1)
[3] G. Jeong, Y.-U. Kim, H. Kim, Y.-J. Kim, H.-J. Sohn, (2020).
Prospective materials and applications for Li secondary [22] D. Jiménez-Bermejo, J. Fraile-Ardanuy, S. Castaño-Solis, J.
batteries, Energy Environ. Sci. 4 (6) (2011) 1986. Merino, R. Álvaro-Hermana, Using dynamic neural networks
[4] T. Kim, W. Song, D.-Y. Son, L.K. Ono, Y. Qi, Lithium-ion for battery state of charge estimation in electric vehicles,
batteries: outlook on present, future, and hybridized Procedia Comput. Sci. 130 (2018) 533–540.
technologies, J. Mater. Chem. A Mater. Energy Sustain. 7 (7) [23] J. Rivera-Barrera, N. Muñoz-Galeano, H. Sarmiento-
(2019) 2942–2964. Maldonado, SoC Estimation for Lithium-ion Batteries:
[5] J. Jiang, C. Zhang (Eds.), Fundamentals and Applications of Review and Future Challenges, Electronics (Basel) 6 (4) (2017)
Lithium-ion Batteries in Electric Drive Vehicles, Wiley, 2015. 102.
[6] R. Rao, S. Vrudhula, D.N. Rakhmatov, Battery Modeling for [24] J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L, ‘‘Proctor
Energy-Aware System Design, Computer 36 (2003) 77–87. dynamic mode decomposition: Data-Driven Modeling of
[7] A. Barré, B. Deguilhem, S. Grolleau, M. Gérard, F. Suard, D. Complex Systems.”.
Riu, A review on lithium-ion battery ageing mechanisms and [25] M.O. Williams, M.S. Hemati, S.T.M. Dawson, I.G. Kevrekidis,
estimations for automotive applications, J. Power Sources 241 C.W. Rowley, Extending data-driven Koopman analysis
(2013) 680–689. to actuated systems, IFAC-PapersOnLine 49 (18) (2016) 704–
[8] T. Huria, M. Ceraolo, J. Gazzarri, and R. Jackey, ‘‘High fidelity 709.
electrical model with thermal dependence for characterization [26] M.O. Williams, I.G. Kevrekidis, C.W. Rowley, A data–driven
and simulation of high power lithium battery cells,” pp. 1–8, approximation of the Koopman operator: Extending dynamic
2012. mode decomposition, J. Nonlinear Sci. 25 (6) (2015) 1307–1346.
11290 M.A. Abu-Seif et al.

[27] C.W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, D.S. [34] N. Demo, M. Tezzele, G. Rozza, PyDMD: Python Dynamic
Henningson, Spectral analysis of nonlinear flows, J. Fluid Mode Decomposition, J. Open Source Softw. 3 (22) (2018) 530.
Mech. 641 (2009) 115–127. [35] M.M. Kabir, D.E. Demirocak, Degradation mechanisms in Li-
[28] M. Naguib, LG 18650HG2 Li-ion Battery Data, Mendeley ion batteries: a state-of-the-art review: Degradation Mechanisms
(2020). in Li-ion Batteries: A State-of-the-Art Review, Int. J. Energy
[29] E. Kaiser, J.N. Kutz, S.L. Brunton, Data-driven discovery of Res. 41 (14) (2017) 1963–1986.
Koopman eigenfunctions for control, Mach. Learn.: Sci. [36] R. Chartrand, Numerical differentiation of noisy, nonsmooth
Technol. 2 (3) (2021) 035023. data, ISRN Appl. Math. 2011 (2011) 1–11.
[30] J.H. Tu et al, On dynamic mode decomposition: Theory and [37] Y. Kim, H. Bang, ‘‘Introduction to Kalman filter and its
applications, J. comput. dyn. 1 (2) (2014) 391–421. applications”, in Introduction and Implementations of the
[31] J.L. Proctor, S.L. Brunton, J.N. Kutz, Dynamic mode Kalman Filter, in: F. Govaers (Ed.), Introduction and
decomposition with control, SIAM J. Appl. Dyn. Syst. 15 (1) Implementations of the Kalman Filter, IntechOpen, 2019.
(2016) 142–161. [38] X. Ding, D. Zhang, J. Cheng, B. Wang, P.C.K. Luk, An
[32] M. O. Williams, United Technologies Research Center, 411 improved Thevenin model of lithium-ion battery with high
Silver Lane, East Hartford, CT 06118, C. W. Rowley, I. G. accuracy for electric vehicles, Appl. Energy 254 (113615) (2019)
Kevrekidis, Dept. of Mechanical and Aerospace Engineering, 113615.
Princeton University, Princeton, NJ 08544, and Department of [39] C. Taborelli, S. Onori, State of charge estimation using extended
Chemical and Biological Engineering & PACM, Princeton Kalman filters for battery management system, in: 2014 IEEE
University, Princeton, NJ 08544, ‘‘A kernel-based method for International Electric Vehicle Conference (IEVC), 2014, pp. 1–8.
data-driven koopman spectral analysis,” J. comput. dyn., vol. 2, [40] M. Sheikholeslami, M. Barzegar Gerdroodbary, R. Moradi, A.
no. 2, pp. 247–265, 2015. Shafee, Zhixiong. Li, Application of Neural Network for
[33] S. L. Brunton, B. W. Brunton, J. L. Proctor, and J. N. Kutz, estimation of heat transfer treatment of Al2O3-H2O nanofluid
‘‘Koopman invariant subspaces and finite linear representations through a channel, Comput. Methods Appl. Mech. Eng. 344
of nonlinear dynamical systems for control,” PLoS One, vol. 11, (2019) 1–12.
no. 2, 2016.