Journal of Energy Storage 84 (2024) 110853

Contents lists available at ScienceDirect

Journal of Energy Storage

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

Research papers

AI-assisted reconfiguration of battery packs for cell balancing to extend

driving runtime
Yuqin Weng ∗, Cristinel Ababei
Electrical and Computer Engineering Department, Marquette University, Milwaukee, WI 53233, USA

ARTICLE INFO ABSTRACT

Keywords: State of charge (SoC) cell balancing is one of the most important roles of battery management systems
State of charge (BMS). The performance and lifespan of a battery pack can be significantly degraded and reduced by the
Cell balancing presence of imbalance in cells SoC. Recently, we have shown that using a machine learning driven battery
Reconfigurable battery pack
pack reconfiguration technique based on a network of controllable switches, one can periodically change the
Machine learning
battery pack topology to effectively achieve better cell SoC equalization. As a result, the driving runtime
achieved with a better balanced battery pack is increased. In this paper, we build on these promising results
and investigate novel machine learning models used for prediction of the best battery pack topologies used
during reconfiguration. In addition, to study the scalability of the proposed battery reconfiguration technique,
we conduct our study on a battery pack with double the number of cells. For validation, we developed an
in-house custom battery pack simulation tool that integrates state-of-the-art battery cell models and extended
Kalman filtering (EKF) algorithms for SoC state estimation. Simulation results using several battery discharging
workloads demonstrate that the machine learning algorithms can achieve better prediction accuracy compared
to previous work, thereby resulting in better cell balancing, which in turn translates into up to 22.4% longer
battery runtime.

1. Introduction densities, which helps storing larger amounts of energy in a relatively

small and lightweight battery pack. This is very important because
Adoption of electric vehicles (EVs) has increased significantly in re- weight and space are top priorities when choosing a battery type in
cent years. One of the most prominent components in EVs is the battery EVs [7]. However, one of the challenges of using Li-ion batteries is
pack. This is a crucial component because the amount of energy that it the battery cells imbalance with regard to the state of charge (SoC)
can store and its performance are directly related to the performance values of individual cells. SoC imbalance occurs when battery cells
of the vehicle. Therefore, it is important to pursue enhancements in have different values of cell voltage, internal resistance, and cell storage
battery technologies and battery pack operation and optimization in capacity during charging or discharging processes [8]. Such differences
order to improve EV performance, which will further speed up EV adop- can be caused by the environment, inconsistent manufacturing, and
tion. This in turn will help reduce the overall unwanted greenhouse cycle life (aging or degradation) [9] and can have adverse effects such
gas (GHG) emissions and global energy consumption [1]. Environment- as chemical explosion and permanent loss of capacity [10].
related departments are promoting appropriate modifications of tradi- One of the best ways to improve battery pack runtime is through
tional automobiles for a sustainable and renewable ecosystem [2]. In battery cell equalization or balancing techniques. The process of cell
this context, different countries proposed different plans to embrace the balancing usually involves a mechanism to equalize the charge or
adoption of EVs. For example, United States aims at reaching zeros discharge levels of all cells in the battery pack, i.e., all cells charge
emissions for 50% of new vehicles [3–5]. Likewise, China plans to or discharge uniformly with all cells having at all times the same SoC
continuously reduce GHG emissions after 2030. The highest ownership values. Cell equalization algorithms can maximize the battery pack
of EVs per capita in the world is Norway, and this is because generous performance, the longevity or lifetime of the battery pack, and the
incentives are given to EV owners, including exemption from taxes and runtime of the vehicle per charge. In general, most of the methods for
tolls, and free charging at public charging stations [6]. battery cell equalization can be categorized into two classes, passive
Li-ion batteries are the most popular types of batteries in EVs. One balancing and activate balancing. The difference between these two
of the reasons for that is that Li-ion technology offers high power

∗ Corresponding author.
E-mail addresses: [email protected] (Y. Weng), [email protected] (C. Ababei).

https://doi.org/10.1016/j.est.2024.110853
Received 10 October 2023; Received in revised form 7 January 2024; Accepted 5 February 2024
Available online 8 February 2024
2352-152X/© 2024 Elsevier Ltd. All rights reserved.
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

classes is that active balancing allows energy transported from cells

that have higher SoC values to cells that have lower SoC values, while
passive balancing enables the battery cell which has more energy to
dissipate energy as heat [11]. Pros and cons of these two categories of
methods are discussed in [11] as well as in recent review studies. For
example, the study in [12] shows that active cell balancing is usually
classified into three categories, depending on the type of hardware
components used: capacitors, converters, or inductors and transformers.
In the category of cell balancing by using capacitors, [13,14] propose
that by shifting energy between nearby cells, cell equalization can be
realized with the help of capacitors. In the category of cell balancing
by using converters, [15,16] show that a standard/modified DC–DC
converter or a PWM converter can be used for achieving cell equal-
ization. Lastly, [17,18] show that cell balancing can be achieved very
quickly by using inductors and transformers, but the disadvantage is
the high cost and high frequency of the transformer. Passive balancing
methods have also been proposed and studied in past literature. The
study in [19] provides a detailed survey of existing passive balancing
methods in various areas and compares them with each other. [20]
shows that passive balancing techniques are commonly implemented
in BMSs that support grid storage applications, and network stability
is the goal for those applications. Moreover, there are previous studies
that use deterministic control strategies. For example, the work in [21] Fig. 1. Reconfigurable battery pack designs with 1, 3, and 5 switches per cell.
formulated cell balancing as an optimal control problem and it pro-
posed to solve it with a network modeling and dynamic programming
approach — to arrive at global optimal energy equalization. [22]
In this paper, a given combination of battery cells connections is called
presents a general nonlinear model predictive control (NMPC) strat-
a pack topology. The main idea of the balancing (i.e., equalization of
egy on series-connected cells for obtaining a balancing-aware optimal
cells SoC) technique presented in this paper is to periodically switch
charging. In [23], a real-time model-predictive-controller was proposed
between different pack topologies in a way that achieves better balance
to balance battery SoC, which in turn extended the driving range,
Moreover, there have been previous studies that use machine learning compared to the case when such switching is not done. The challenging
(ML) and artificial intelligence (AI) algorithms as control strategies for aspect of this switching is how to identify in real-time the optimal
cell balancing. For example, the study in [24] proposed an ML based topology to switch to for the next control period. Another challenge
control algorithm to achieve balancing of both SoC and temperature. is to decide what type of network of reconfigurable switches to use.
The study in [25] also demonstrated that smarter controls and longer In previous literature, there have been studied several different re-
lifetimes can be achieved with emerging ML techniques. Nevertheless, configurable battery pack designs using from 1 up to 6 switches per
the use of AI-based cell balancing techniques is in its early beginning. battery cell, which may become costly when this number is large [28].
It is not clear yet how far such techniques can push cell balancing in Examples of reconfigurable pack designs that use 1, 3, and 5 switches
particular and battery pack optimization in general. per cell are depicted in Fig. 1.
Therefore, in this paper, we propose and study a novel ML-based The design with only 1–2 switches per cell is quite limited in the
cell balancing technique for reconfigurable battery pack systems. The ability to provide a variety of connection combinations; however they
proposed battery pack system is a smart system in line with recent allow access, exit, and bypass of each cell [29]. The advantages include
developments in reconfigurable battery packs as a special form of future low cost and easy assembly. Designs with 3–4 switches per cell are
smart batteries [26]. The proposed reconfigurable battery pack system much more flexible in their ability to facilitate a variety of connections.
and AI-based reconfiguration technique are verified in simulations Combinations of series, parallel, and parallel-series topologies are pos-
conducted with a simulation tool that we developed and reported in
sible with such pack designs. Furthermore, designs with 5 switches per
our recent conference paper [27]. In this paper, we extend upon this
cell can deal with any cell fault in any pack topology, while designs
preliminary work by investigating new machine learning models for
with 6 switches per cell can reliably supply loads with different volt-
prediction of best pack topologies and of larger battery packs, and for
age requirements [29]. While the flexibility of pack designs with 5–6
additional discharging workloads. One of the main advantages of using
switches per cell is the highest, the cost of manufacturing and mainte-
an ML/AI approach to determine the reconfiguration over deterministic
control strategies is that training can regularly be repeated based nance is also high, in addition to the more complex assembly processes.
on data gathered from the battery pack without the need to update In summary, the battery pack complexity and flexibility heavily depend
other models, such as equivalent circuit models, in order to improve on the number of switches. In this work, we use 3 switches per cell —
balancing. This is also a reason for which an ML/AI based approach is as a reasonable trade-off between pack cost and flexibility.
proposed in work over other more conventional approaches.
3. Simulation framework: Cell modeling, SoC estimation, and
2. Batteries can become smart through reconfiguration pack topology

Smart batteries usually rely on some form of reconfiguration of

the cells arrangement. Reconfiguration of the battery pack provides In this section, we describe the custom simulation framework that
a mechanism to achieve cell SoC equalization, which becomes very we developed for validation of the proposed ideas. We focus on the
challenging in the case of static battery packs. Reconfiguration means most important elements of the simulation framework: 1) the equiva-
that different combinations of battery cells connections (in series, or lent circuit model for the battery cell and SoC estimation techniques;
in parallel, or combinations of series and parallel) can be setup by and 2) the battery pack topology with a network of reconfigurable
means of a network of reconfigurable switches that connects the cells. switches connecting all sixteen cells studied in this paper.

2
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Fig. 2. Equivalent electric circuit for a battery cell used by the ESC model [31].
Fig. 3. Illustration of the ESC cell model in parallel with the true hardware cell. The
EKF algorithm solves the model at each time step 𝑘 by using the Kalman gain 𝐾𝑘 during
the state correction step or measurement update. The input 𝑢𝑘 is the instantaneous
3.1. Cell modeling current at each time 𝑘. The battery cell state 𝑥𝑘 and output 𝑦𝑘 are described by Eqs. (1)
and (2). EKF estimated state and output are 𝑥̂ 𝑘 and 𝑦̂𝑘 .

There are several popular equivalent circuit models for modeling

battery cells in the literature [30]. The most prominent one is the
enhanced self-correcting (ESC) model. Besides capturing cell SoC, the to record the cell testing data for 𝑂𝐶𝑉 (𝑧𝑘 ) in practical applications.
ESC model also captures the hysteresis voltage and diffusion processes The diffusion voltage is represented by 𝑅1 𝑖𝑅,𝑘 [32]. Similar to previous
from the battery cell. The equivalent circuit model for the ESC model studies [31], the ESC model uses one 𝑅1 𝐶1 pair in this work. The
is illustrated in Fig. 2, and the state space representation is described hysteresis state which is denoted as ℎ𝑘 , is the third state in the state
by the following equations [31]: equation. It also has a connection with the input 𝑖𝑘 in Eq. (1). Finally,
a small constant 𝛾 is used to model the voltage decay.
⎡ 𝑧𝑘+1 ⎤ ⎡1 0 0 ⎤ ⎡ 𝑧𝑘 ⎤
⎢𝑖 ⎥ = ⎢0 𝐴𝑅𝐶 0 ⎥ ⎢𝑖𝑅,𝑘 ⎥ +
⎢ 𝑅,𝑘+1 ⎥ ⎢ ⎥⎢ ⎥ 3.2. SoC estimation
⎣ ℎ𝑘+1 ⎦ ⎣0 0 𝐴𝐻,𝑘 ⎦ ⎣ ℎ𝑘 ⎦
⎡( −𝜂𝑘 𝛥𝑡 ) ⎤[ (1)
0 ] Because the SoC of a battery cell is not a quantity that can be
⎢ 𝑄 ⎥ 𝑖𝑘
⎢ 𝐵𝑅𝐶 0 ⎥ 𝑠𝑔𝑛(𝑖 ) directly measured, estimation methods are typically employed. One of
⎢ 0 (𝐴𝐻,𝑘 − 1)⎥⎦
𝑘 the most common used methods to estimate SoC of a battery cell is
⎣
Kalman filtering. Kalman filter theory is a classic technique developed
𝑦𝑘 = 𝑂𝐶𝑉 (𝑧𝑘 ) + 𝑀ℎ𝑘 − 𝑅1 𝑖𝑅,𝑘 − 𝑅0 𝑖𝑘 (2) in the 1960’s [33]. Kalman filters have been used in many application
areas including control systems, signal processing and image processing
⎡ 𝑧𝑘 ⎤ due to its accuracy and robustness [34]. As a variation of Kalman filter
𝑥𝑘 = ⎢𝑖𝑅,𝑘 ⎥ (3) theory, the extended Kalman filtering (EKF) is used for the estimation
⎢ ⎥
⎣ ℎ𝑘 ⎦ of non-linear state-space models. The idea of EKF is that a linearization
Where Eq. (1) represents the state equation. The output equation is can be used at around the current estimate by using partial derivatives
defined by Eq. (2). In these equations, the output of the system which of the state and output equations to compute estimates for each time
is also the terminal voltage of the battery cell is denoted as 𝑦𝑘 , and the step in discrete time [30]. Thus, EKF is a perfect choice for estimating
input of the system is 𝑖𝑘 , which is called the cell instantaneous current. battery cell SoC because the battery cell state-space model is a non-
The internal resistance of the battery cell is defined as 𝑅0 . There are linear representation at time 𝑘. Thus, by applying EKF to the battery
three states captured by this model and they are SoC: 𝑧𝑘 , diffusion cell equivalent circuit model (Eqs. (1) and (2)), the estimated state and
process current: 𝑖𝑅,𝑘 , and hysteresis voltage: ℎ𝑘 . They are kept in a state output can be generated at each time 𝑘. An illustration of how EKF is
vector in Eq. (3). Some entries in Eq. (1) are described by the following used for SoC estimation is shown in Fig. 3. The simulation tool, which
equations: we developed to be able to conduct the simulation experiments from
this paper, implements the EKF approach for estimating SoC for all cells
−𝛥𝑡
𝐴𝑅𝐶 = 𝑒𝑥𝑝( ) (4) in a battery pack, which are then used to estimate the SoC of the entire
𝑅 1 𝐶1
pack for the best possible accuracy of results.
−𝛥𝑡
𝐵𝑅𝐶 = 1 − 𝑒𝑥𝑝( ) (5)
𝑅 1 𝐶1
3.3. Battery pack topologies
𝜂 𝑖 𝛾𝛥𝑡
𝐴𝐻,𝑘 = 𝑒𝑥𝑝(−| 𝑘 𝑘 |) (6)
𝑄 The cell balancing technique proposed in this paper relies on a re-
When the cell is fully charged and fully discharged, the SoC 𝑧𝑘 is configurable battery pack topology. The inspiration for using different
defined as 100% and 0%. While in theory, SoC could not be larger pack topologies to equalize cell SoCs comes from [35]. In that study,
than 100% or smaller than 0%, during the process of charging or a network of programmable switches is controlled to change the con-
discharging in practice these limits are tighter; for example, it is typical nections inside a photovoltaic (PV) array. We adopt the same network
for the lowest SoC limit to be 5%–30%, during discharging because of reconfigurable switches, which uses 3 switches per battery cell as
that helps the battery reliability. The Coulombic efficiency, denoted shown in Fig. 4. In this structure, a set of three switches for every cell
as 𝜂𝑘 , is usually equal to one when the cell is discharging and less includes switches denoted as 𝑆𝑃 𝑇 ,𝑖 , 𝑆𝑃 𝐵,𝑖 , and 𝑆𝑆,𝑖 (shown on Fig. 4(a)).
than one when it is charging. To convert the continuous-time system These switches must be operated according to a simple rule: switches on
to a discrete-time system, a small sampling interval 𝛥t is used. The the top and bottom levels should remain in a consistent state, meaning
total capacity of the cell is denoted as 𝑄, which is the total amount of they should both be on or both be off; while the switch located on
charge when charging the cell from 0% to 100% SoC. The open circuit the middle level must be in a state that is opposite to the other two
voltage of the cell, denoted as 𝑂𝐶𝑉 (𝑧𝑘 ), is an ideal voltage source and is switches. Fig. 4(b) illustrates a battery pack topology changing from
dependent on SoC under a certain temperature. A lookup table is used [2,2,2,2,2,2,2] to [2,4,2,2,2,2,2] and then to [2,2,2,4,2,2,2], which can

3
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Fig. 5. Illustration of the Pareto optimality. 𝐶1 and 𝐶2 are calculated with Eqs. (7)
and (8).

simulation time or use. In addition, it was also observed that the level of
imbalance generally increases with time. This motivated us to develop a
topology-switching technique that allows the battery pack to switch be-
tween different topologies during different control or operation periods.
This technique helps reduce the imbalance in the next control period.
For instance, during pack discharging, it is desired to maintain all SoC
values high and as equal as possible among each other. In other words,
we want to reduce the difference between different cells SoC as much as
possible at any given time. Consequently, cells SoC in the battery pack
will be kept as uniform as possible. Results of our preliminary work
showed that the proposed topology switching algorithm can improve
battery cell equalization even more, beyond what one can achieve by
continuously using just one best topology. To identify which topology
is the best for the next control period, several ML models will be
investigated. To train these models, rich datasets with inputs and labels
need to be created. In this section, we discuss how we generate such
datasets using our simulation tool.
During dataset collection, we need to define what inputs will be
Fig. 4. (a) Reconfigurable battery pack with 16 cells. (b) Example of a pack topology
changing from [2,2,2,2,2,2,2] to [2,4,2,2,2,2,2] and then to [2,2,2,4,2,2,2]. used and what the output labels are. For our problem, the labels
must represent identifiers of the topologies that are included in the
set to work with for the purpose of switching among them. When a
switch from a topology to another happens, we are interested in two
easily be achieved by appropriately closing or opening switches shown
objectives: 1) Maximize the summation of all SoC values of battery
in Fig. 4(a).
cells which is equivalent to minimizing the inverse of the summation
Please note that our work assumes that the battery pack is connected
of all SoC values; and 2) Minimize the range of SoC values, which is
to the load through a power converter, which will make sure that
calculated as the difference between the maximum and minimum SoC
the voltage level delivered to the load is kept constant — the role
values among all cells. This can be regarded as a multi-objective op-
of the converter is to regulate the output voltage to the necessary
timization problem. Consequently, two objective functions are defined
level. This will ensure the stability of the voltage level supplied to
to determine the next best topology to switch to. All possible solutions
the load and will address the issue of potential changes in the output
of this optimization problem represent a solution space bordered by
voltage of the battery pack due to switching between different pack
a Pareto frontier [36]. Fig. 5 shows an example of such a solution
topologies during the reconfiguration process. Therefore, the amount
space where each dot represents a possible next pack topology. The
of power provided to the load should remain the same. Nevertheless, it
dashed line shows the Pareto frontier. Prior research [37] showed that
is the power converter’s efficiency and performance that may however
the Pareto optimality of these solutions ensures that it is not possible
introduce small difference in the output power supplied by different
to enhance one objective without compromising the other. The cost
pack configurations; but, in this work we consider that to be negligible.
functions for two objectives 𝐶1 and 𝐶2 are defined as:
4. Collection of datasets and input features of ML model 𝑚𝑎𝑥
𝐶1 = 𝑆𝑜𝐶𝑑𝑖𝑓 𝑓 = max 𝑆𝑜𝐶𝑖 − min 𝑆𝑜𝐶𝑖 (7)
𝑖∈[𝑛] 𝑖∈[𝑛]

4.1. Collection of datasets ∑

𝑛
𝐶2 = 1∕ 𝑆𝑜𝐶𝑖 (8)
The SoC balancing algorithm proposed in this paper and discussed 𝑖=1

later relies on machine learning models, which require training. Ex- Where we define [𝑛] = {1, 2, … , 𝑛} with 𝑛 representing the total number
tensive SoC simulations showed that different initial SoC conditions of cells in the battery pack.
and different cell parameter values can lead to widely different SoC Because we work with a total of 𝑁 different topologies only, the
variation profiles when different battery pack topologies are used. optimization problem is essentially identifying the topology that should
In other words, different pack topologies lead to different imbalance be used in the next control period. This identification is done by
levels starting from the same initial conditions and during the same evaluating the following expression that combines the two individual

4
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Fig. 6. Pseudocode of the simulation process used to generate datasets.

cost functions 𝐶1 and 𝐶2 . The topology that minimizes this expression

is the one identified as the best [38]:
√ Fig. 7. Flowchart of the proposed balancing algorithm; topology switching via
2 2
𝑑 = arg min (𝐶1𝑘 ) + (𝐶2𝑘 ) (9) reconfiguration is done periodically, with a control period of 5 min.
𝑘

Where we define [𝑁] = {1, 2, … , 𝑁} with 𝑁 representing the total num-

ber of topologies that are in the solution space. The topologies which
information about the current state of the battery pack while the output
are acceptable in terms of output voltage levels can be considered in
label is the ID of the best topology to use next. In our implementation
the solution space, more details about pruning existing topologies are
for the case of the battery pack with 16 cells, the final number of
described in section slowromancapvi@. In Eq. (9), the best topology
different topologies that we kept to work with is seven. Therefore, the
which is recorded as topology 𝑘 has the minimum Euclidean distance
output level is the ID of the next best topology as a number between
to the origin in Fig. 5. The minimum Euclidean distance is calculated
0 and 6. The input features include the following: the current state
with the help of Eq. (7), Eq. (8), and Eq. (9). In addition, 𝐶1𝑘 and 𝐶2𝑘
of charge (SoC) values of all 16 cells; their mean value, maximum
are the values of 𝐶1 and 𝐶2 for topology 𝑘 in Eq. (9).
value, minimum value, standard deviation, summation, and maximum
Our process of generating datasets is similar to what we used in
absolute difference (i.e., distance between the largest and the smallest
our previous preliminary study and it relies essentially on the above
SoC values of all cells; in other words, this is the range or span or spread
equation. The difference here is that we are working with a battery
of all values of SoCs) among all 16 cells; and, the currently being used
pack with a double number of cells, which makes the problem more
topology ID. Therefore, there is a total of 23 input values recorded as
challenging due to the much larger number of battery pack topolo-
input features. This is also indicated in Fig. 8.
gies that one can potentially form by means of the reconfigurable
There are several reasons for which we chose these features. We
switch network. The simulator is instrumented to conduct exhaustive
are optimizing two objectives, C1 and C2, as described by Eqs. (7)
simulations of the type what-if scenarios — during which dataset is
and (8). The calculation of expressions of C1 and C2 requires most
generated for all possible next topologies. For each control period that
of the input features discussed above. In addition, the mean, standard
is simulated and added to the dataset, we identify the Pareto optimal
deviation, and span are included as input features to capture statistical
topology that is labeled as the next best topology. Looking again at
Fig. 5, during the simulations for dataset generation, at the end of a characteristics in a compressed format. In our extensive empirical
simulated control period, there are 𝑁 topologies evaluated via the cost experiments, we found that the accuracy of the ML models increased if
functions 𝐶1 and 𝐶2 ; these topologies represent the dots in the figure. we included the current SoC values of all 16 cells compared to the case
The best next topology will be identified as the one that results in the when they were not included.
minimization of Eq. (9). That topology is recorded as the best one to
switch the pack to and then continue the simulation for the next control 5. Proposed cell balancing algorithm
period during the dataset generation process. Multiple simulations are
run to collect data for what the best next topology is for different states The idea of the proposed cell balancing algorithm is to use machine
of the battery cells. The process of conducting simulations for dataset learning models to identify the next best battery pack topology to
generation is described with the pseudocode from Fig. 6. switch to for the next control period. The models are trained with
the datasets generated as explained before. The simulation tool that
4.2. Input features and output label for the ML model we instrumented to generate the datasets is also used to implement
and verify the proposed cell balancing algorithm. Fig. 7 shows the
The dataset collected as described in the previous section will be flowchart describing the main steps involved in the proposed algorithm.
used for training of the machine learning models. Because we deal Identifying the next topology to switch to is done using a machine
with supervised training, each entry into the recorded dataset will be learning model for predictions. Once the ID of the next topology
a pair of the form Input features, Label. The input features represent is identified, all configurable switches in the reconfigurable network

5
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Table 1
Prediction accuracy of studied models.
Models Testing accuracy Training accuracy
𝐷𝑁𝑁 72.0% 73.5%
𝑆𝑉 𝑀 66.8% 72.1%
𝐴𝑑𝑎𝐵𝑜𝑜𝑠𝑡 55.9% 60.7%
𝐾𝑁𝑁 50.1% 53.8%
𝑅𝐹 40.9% 40.6%

are controlled to the correct on/off state to implement the desired

topology. The information about the state of all switches for a given
topology ID is easily stored and retrieved from a small loop-up table
in our implementation. We investigated several models in order to
find out which one would be best for our application: Adaptive Boost
(AdaBoost) algorithm, support vector machine (SVM), and deep neural
network (DNN).
Fig. 8. Structure of the DNN model.
In our preliminary work, we investigated random forest (RF) and
k-nearest neighborhood (KNN) models. Here, we investigate AdaBoost
as an ensemble learning technique, which is similar to RF. We selected
AdaBoost because it has been shown that it is more accurate than RF
for predictions generally [39]. The study in [40] showed that although
RF has higher overall accuracy, AdaBoost has better class-specific ac-
curacy. The second ML model we investigate is SVM, which is efficient
and known providing good results in terms of maximizing margins
between different classes. Therefore, SVM can handle high-dimensional
data, and is easily applied to both binary and multi-class classification
problems. The last model we investigate is DNN — because it has
become very popular and widely used in many problem settings from
engineering, data science, and biomedical fields. In addition, we found
previous literature that used DNNs to estimate the state of health for
battery cells in EVs [41,42]. Next, we describe the performance of these
models.
In order to limit the amount of time put into model development
and training, we restrict the analysis to only ten different topologies.
During the extensive simulations used for dataset generation, it turned
out that only seven were dominant and consistently used to switch
between. As such, the number of classes or labels we work with is
seven. All the investigated models are tested for those seven labels.
The process of initially selecting a reduced number of topologies will
be described in the next section; here, we focus on datasets generation
and models performance.
The datasets were generated as described earlier (Fig. 6) — where
each topology was simulated 400 times starting from different initial
SoC values drawn randomly from ranges between 90% to 30% with
a step of 15%. Hence, the total number of datapoints in datasets is
Fig. 9. Loss and accuracy plots for the DNN model.
10 × 400 × 5 = 20000. Datasets are split into train and test portions
using a 60%/40% split. A grid search technique [43] is implemented for
both AdaBoost and SVM algorithms to select suitable parameters. For presented in Fig. 9 — where the 𝑥-axis represents the epochs number.
AdaBoost, n estimators and learning rate parameters are set to 3000 and We find that the DNN model has the highest testing accuracy among
1. For SVM, C and gamma parameters are set to 15000 and 0.015. The all studied models, with the SVM model coming in close as the second.
DNN model has the structure presented in Fig. 8, and it has two hidden AdaBoost, KNN, and RF models are the poorest performing models.
layers — with 128 and 64 neurons, respectively. Other parameters used The confusion matrices for the testing portion of the DNN, SVM, and
include: loss function as categorical cross entropy, optimizer as Adam, AdaBoost models are shown in Fig. 10, which again shows that the
learning rate set to 0.00015 during training. The input to the DNN DNN model has the best performance (i.e., largest numbers of correct
model from Fig. 8 is a combination of: current topology, 16 cell SoC predictions are shown on the main diagonal).
values at the endpoint of previous control period, and several statistical
features of the cells SoC. A 10% dropout regularization is applied on 6. Simulation results
each hidden layer, meaning each node has a 10% chance of being
dropped out at every epoch when training. To avoid overfitting, an 6.1. Description of experiments
early-stopping technique is also implemented with a patience value
of 200. Both dropout regularization and early-stopping techniques are Simulations are conducted for a battery pack with 16 cells. All
useful tools to avoid overfitting [44,45]. simulations were run on a computer with an Intel(R) Core(TM) i5-
The training and testing accuracy for each model including the ones 10400 CPU running at 2.90 GHz using 16MB of RAM, running on
used in our preliminary work are reported in Table 1. In addition, Windows 11. The GPU used for training DNN model is the NVIDIA
for example, the loss and accuracy variations for the DNN model are GeForce GTX 1660 with 6 GB memory. The total simulation time for

6
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

the list represents the number of cells is connected in parallel on that

row; and, all rows are connected in series. Even when the number
of rows is restricted like that, there are still more than 100 possible
topologies that one can create. To further prune the topologies, a max-
imum cell difference between each row should be set. The maximum
cell difference in the 7-row topology listed is 2. Through simulations,
we found out that any topology that has a maximum cell difference
larger than 3 tends to be more unbalanced than those topologies that
have a maximum cell difference smaller than 3. Thus, pruning those
unbalanced topologies beforehand is necessary.
However, even after such pruning of topologies, simulations for
datasets generation is still computationally expensive and inefficient.
The datasets generation is the most time-consuming process in this
work, yet it is needed for training the machine learning models. There-
fore, we keep only the top 10 topologies from this subset efficiency rea-
sons. During the process of generating the datasets, three least-favorite
topologies were chosen less than ten times in the whole datasets
generating process. Thus, to further eliminate this redundancy, we have
selected as final topologies only seven out of the ten topologies to
continue with model training.
Simulations are carried out using three different workloads: con-
stant workload, step workload, and a combination of multiple urban
dynamo-meter drive schedule (UDDS) drive cycles (UDDS workload).
For each workload, we conduct a comparison of two scenarios: dis-
charging with one fixed topology (base case) and discharging with
switching between different topologies as dictated by the proposed
balancing algorithm. Only the DNN model, which turned out to be the
best, is implemented in these simulations. The base case topology is
selected to be [2, 2, 2, 2, 2, 2, 2] because it was found through empiri-
cal simulations to be the most balanced topology when no topology
switching is used. The discharging rate for UDDS workload was set to
0.1𝐶. The discharging rate for step workload (where the battery pack
discharges for 3–5 control periods and rests for one control period)
is 1.5 times as much as the UDDS workload and discharging rate for
the constant workload is 2 times as much as the UDDS workload.
These rates are similar to those used by previous work, in the range
of 0.02-1𝐶 [22,23].
Finally, using the final selected seven topologies, the average ex-
ecution time for the proposed algorithm is about 1.6 s during a full
simulation (SoC from 90% down to 10%). We believed that such a small
computational runtime would be similar when the proposed algorithm
is deployed on real battery management systems (BMS), which have
powerful enough processors. Also, we would like to note that we
Fig. 10. Confusion matrices for DNN, SVM, and AdaBoost models — obtained on the
test portion of the datasets.
assumed ideal switches in our models in all simulations as a simplifica-
tion. In reality, switches themselves have a non-zero resistance, which
will affect the output of the battery pack. However, high-performance
generating all datasets was approximately 30 h. The training time
power MOSFET transistors (that it is envisioned to be used as switches
for DNN model was about 60 min. Numerical values of various cell
in a future prototype, e.g., IRFP4668PBF N-channel power MOSFETs)
parameters including 𝑧, 𝑄, 𝑅1 , 𝐶1 , and 𝑅0 are adopted from previous
literature [31,32]. To mimic realistic variations of these parameters, have on resistance as small as 9.7 mΩ [46]. Hence, our simplification
their values are drawn from Gaussian probability distributions, charac- is acceptable.
terized by standard deviations of 5% and mean values also reported
in [31,32]. Also, the initial SoC values for all cells in the pack and 6.2. Results and discussion
generated from Gaussian distributions characterized by a mean of 0.9
(representing 90%) and 5% standard deviation — and used as starting The results of the simulations are shown in Fig. 11, 12, and 13.
SoC values for both the reference and proposed algorithm cases. There In each of these figures, the 𝑥-axis represents the time measured in
are thousands of topologies that can be formed for a battery pack with seconds. This time is the runtime of the battery pack before its overall
16 cells; however, most of them are not realistic for practical use. For SoC drops below a preset threshold (10%). Of course, it is desired
example, when all cells are connected in series or all cell are connected for this time to be the longest possible because that translates into
in parallel can have limitations in terms of required levels of current longer driving distances. The limits of 10%–90% are based on what
and voltage levels. Therefore, such topologies are excluded from being past literature recommends typically [47]. Again, as mentioned earlier,
considered useful topologies that the proposed balancing technique can the starting SoC values for each of the two compared simulations are
use during reconfiguration of the battery pack. the same, and generated from Gaussian distributions characterized by
To restrict the number of different topologies that we work with – a mean of 0.9 (representing 90%) and 5% standard deviation. Each
and thus to make the proposed technique realistic and efficient – we control period is 300 s (5 min) and is shown as the distance between
restrict our attention to topologies with 6, 7, and 8 rows. For instance, two consecutive vertical lines in these figures. From these plots, we
a 7-row topology could be [2, 4, 2, 2, 2, 2, 2], where each number inside see that the proposed balancing algorithm significantly improves the

7
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Table 2
Summary of improvements in terms of runtime length and SoC values span.
Discharging Improvement in Improvement in
Workload Battery Runtime SoC Values Span
Increase (%) Reduction (×)
Constant 21.5% 3.25×
Step 22.4% 2.75×
UDDS 19.8% 2.33×

those times. These non-perfect topologies will make for the overall SoC
values to follow a path that is possibly less optimal than the absolute
best. Currently, we do not have an efficient way to estimate how far
the solution achieved by the proposed algorithm is from the solution
one would get with a perfect predictor. The best way to address this
difference is to improve the ML model accuracy by refining further the
model or by employing other, more sophisticated models — which we
are currently exploring and be reporting in our future work.

7. Conclusion and future work

In this paper, we presented a novel and enhanced cell balancing

technique for reconfigurable battery packs that are integrated with
networks of reconfigurable switches, which can be controlled to create
different series, parallel or combinations of such connections. The
objective of the balancing technique is to keep all battery cells at
SoC values that are close as possible to each other during the dis-
charging process. This is achieved by periodically switching between
different battery pack topologies, which turn out to help improve cell
equalization. Decisions as of what topology to switch to during each
reconfiguration is done with the help of a machine learning based
prediction. We studied several machine learning models among which
deep neural networks seemed to offer the best prediction accuracy. Ex-
tensive simulation experiments on a 16 cell battery pack with a custom
Fig. 11. (a) Variation of cells SoC for base case topology [2, 2, 2, 2, 2, 2, 2, 2],
built simulation tool showed that battery runtime can be increased by
constant workload. (b) Variation of cells SoC for a pack with topology reconfigured
using the proposed balancing algorithm, constant workload. to up to 22.4%.
While the results in this paper are very encouraging and demon-
strate that AI based balancing approaches have great potential, there
is need for more research in this direction. That is why, in our future
runtime of the battery pack on all workloads at different discharging
work, we plan to first investigate scalability of the proposed technique
rates. For constant workload, the runtime of the battery pack is im-
by studying additional battery pack topologies formed with larger
proved from 15,600 to 18,900 s, i.e., 21.5%. Furthermore, compared
number of cells. More specifically, we will look into divide-and-conquer
to our previous work – in the case of constant workload – results are
approaches that partition large numbers of cells into partitions of fewer
improved with 10.9%.
cells and apply the technique discussed in this paper to those partitions.
Another figure of merit that we use to assess the differences between
This means an ML model is developed only once (for a partition) and
the base case and the proposed algorithm, is the span of the range
then applied during the reconfiguration algorithm to all partitions.
of all SoC values at the end of the simulation. We observe that also
It would also be interesting to investigate other types of workloads.
in terms of this metric, the proposed balancing algorithm improves
results from a span of SoC values of 0.13 to 0.04 in the case of the Finally, we plan to develop a hardware prototype for a battery pack
constant workload. In the case of step workload, results show that the with 8 cells and validate in practice the benefits of battery topology
improvement in pack runtime is from 25,500 to 31,200 s, i.e., 22.4%. reconfiguration and switching versus actual costs.
In terms of SoC values span, the improvement is from 0.11 to 0.04 for
the step workload. Finally, in the case of UDDS workload, the pack CRediT authorship contribution statement
runtime is improved from 30,300 to 36,300 s, i.e., 19.8% — which is
also with 5.2% better than our previous work. The cells SoC values span Yuqin Weng: Conceptualization, Investigation, Methodology, Soft-
is decreased from 0.07 to 0.03 for the UDDS workload. Although SoC ware, Validation, Writing – original draft. Cristinel Ababei: Conceptu-
is still not perfectly balanced (i.e., zero span of SoC values), the span alization, Supervision, Writing – review & editing.
of the final SoC values has significantly been reduced by the proposed
technique, and as a result the runtime of the battery pack increased. A Declaration of competing interest
summary of all the improvements in terms of runtime length and SoC
values span are listed in Table 2. The authors declare that they have no known competing finan-
Finally, we would like to mention that while the proposed balancing cial interests or personal relationships that could have appeared to
algorithm provides significantly longer battery runtime, it is still not influence the work reported in this paper.
the best one could possibly achieve. The reason for that is because the
DNN model’s prediction accuracy is only around 73%, which is far from Data availability
perfect. This results in turn to a certain fraction of reconfigurations to
be done to pack topologies that are not the ‘‘perfect’’ or ideal ones at The data that has been used is confidential.

8
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

Fig. 12. (a) Variation of cells SoC for base case topology [2, 2, 2, 2, 2, 2, 2, 2], step workload. (b) Variation of cells SoC for a pack with topology reconfigured using the proposed
balancing algorithm, step workload.

Fig. 13. (a) Variation of cells SoC for base case topology [2, 2, 2, 2, 2, 2, 2, 2], UDDS workload. (b) Variation of cells SoC for a pack with topology reconfigured using the
proposed balancing algorithm, UDDS workload.

9
Y. Weng and C. Ababei Journal of Energy Storage 84 (2024) 110853

References [25] B. Wu, W.D. Widanage, S. Yang, X. Liu, Battery digital twins: perspectives on the
fusion of models, data and artificial intelligence for smart battery management
[1] M. Iqbal, A. Benmouna, M. Becherif, S. Mekhilef, Survey on battery technologies systems, Energy AI 1 (2020).
and modeling methods for electric vehicles, Batteries 9 (3) (2023) 185. [26] Z.B. Wei, J.Y. Zhao, H.W. He, G.L. Ding, H.Y. Cui, L.C. Liu, Future smart
[2] M.-K. Tran, S. Panchal, T.D. Khang, K. Panchal, R. Fraser, M. Fowler, Concept battery and management: advanced sensing from external to embedded
review of a cloud-based smart battery management system for lithium-ion multi-dimensional measurement, J. Power Sources 489 (2021) 229462.
batteries: feasibility, logistics, and functionality, Batteries 8 (2) (2022) 19. [27] Y. Weng, C. Ababei, Battery pack cell balancing using topology switching and
[3] K. Fang, C. Li, Y. Tang, J. He, J. Song, China’s pathways to peak carbon machine learning, in: IEEE Vehicle Power and Propulsion Conf., VPPC, 2022, pp.
emissions: new insights from various industrial sectors, Appl. Energy 306 (2022). 1–6.
[4] N. Abhyankar, P. Mohanty, A. Phadke, Illustrative Strategies for the United [28] W.J. Han, T. Wik, A. Kersten, G.Z. Dong, C.F. Zou, Next generation battery
States to Achieve 50% Emissions Reduction by 2030, Lawrence Berkeley National management systems: dynamic reconfiguration, IEEE Ind. Electron. Mag. 14 (4)
Laboratory, Berkeley, CA, USA, 2021. (2020) 20–31.
[5] P. Mock, S. Díaz, Pathways to decarbonization: the European passenger car [29] K. Liu, Z. Wei, C. Zhang, Y. Shang, R. Teodorescu, Q.-L. Han, Towards long
market in the years 2021–2035, Communications 49 (2021). lifetime battery: AI-based manufacturing and management, IEEE/CAA J. Autom.
[6] H. Scott, Understanding the impact of reoccurring and non-financial incentives Sinica 9 (7) (2022) 1139–1165.
on plug-in electric vehicle adoption – A review, Transp. Res. A 119 (2019) 1–14. [30] G. Plett, Extended Kalman filtering for battery management systems of LiPB-
[7] S.S. Rangarajan, S.P. Sunddararaj, A. Sudhakar, C.K. Shiva, U. Subramaniam, based HEV battery packs. Part 1. Background, J. Power Sources 134 (2) (2004)
E.R. Collins, T. Senjyu, Lithium-ion batteries—the crux of electric vehicles with 252–261.
opportunities and challenges, Clean Technol. (2022) 908–930. [31] G. Plett, ECE5720: Battery Management and Control, University of Colorado Col-
[8] A. Samanta, S. Chowdhuri, Active cell balancing of lithium-ion battery pack orado Springs, 2015, [Online]. Available: http://mocha-java.uccs.edu/ECE5720/
using dual dc-dc converter and auxiliary lead-acid battery, J. Energy Storage 33 index.html.
(2021) 102109. [32] G. Plett, ECE4710/5710: Modeling, simulation, and identification of battery
[9] J. Chen, Z. Zhou, Z. Zhou, X. Wang, B. Liaw, Impact of battery cell imbalance dynamics, University of Colorado Colorado Springs, 2018, [Online]. Available:
on electric vehicle range, Green Energy Intell. Transp. 1 (3) (2022) 100025. http://mocha-java.uccs.edu/ECE5710/index.html.
[10] Z.B. Omariba, L. Zhang, D. Sun, Review of battery cell balancing methodologies [33] W. Greg, B. Gary, An introduction to the Kalman filter, 2006, [On-
for optimizing battery pack performance in electric vehicles, IEEE Access 7 line]. Available: chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/http://
(2019) 129335–129352. www.cs.unc.edu/~welch/media/pdf/kalman_intro.pdf.
[11] X. Wang, S. Li, L. Wang, Y. Sun, Z. Wang, Degradation and dependence analysis [34] Y. Weng, Detection and Characterization of Actuator Attacks Using Kalman Filter
of a lithium-ion battery pack in the unbalanced state, Energies 13 (22) (2020) Estimation (M.S. thesis), Marquette University, 2019.
5934. [35] Y. Wang, X. Lin, Y. Kim, N. Chang, M. Pedram, Architecture and control
[12] A.K.M.A. Habib, M.K. Hasan, G.F. Issa, D. Singh, S. Islam, T.M. Ghazal, Lithium- algorithms for combating partial shading in photovoltaic systems, IEEE Trans.
ion battery management system for electric vehicles: constraints, challenges, and Comput.-Aided Des. Integr. Circuits Syst. 33 (6) (2014) 917–930.
recommendations, Batteries 9 (3) (2023) 152. [36] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms: An
[13] A.K.M.A. Habib, M.K. Hasan, M. Mahmud, S.M.A. Motakabber, M.I. Ibrahimya, S. Introduction, KanGAL Report Number 2011003, 2011.
Islam, A review: energy storage system and balancing circuits for electric vehicle [37] Y. Akishita, Y. Ohsita, M. Murata, Network power saving based on Pareto optimal
application, IET Power Electron. 14 (2021) 1–13. control with evolutionary approach, in: Int. Conf. on Computing, Networking and
[14] T. Hein, A. Ziegler, D. Oeser, A. Ackva, A capacity-based equalization method Communications, ICNC, 2017.
for aged lithium-ion batteries in electric vehicles, Electr. Power Syst. Res 191 [38] C.E.A. Bundak, M.A.A. Rahman, M.K.A. Krim, N.H Osman, Fuzzy rank cluster
(2021). top k Euclidean distance and triangle based algorithm for magnetic field indoor
[15] F. Eroglu, M. Kurto, M. Vural, Bidirectional dc–dc converter based multilevel positioning system, Alex. Eng. J. 61 (5) (2022) 3645–3655.
battery storage systems for electric vehicle and large-scale grid applications: [39] A. Kumar, Differences between random forest vs AdaBoost, 2022, [Online]. Avail-
a critical review considering different topologies, state-of-charge balancing and able: https://vitalflux.com/differences-between-random-forest-vs-adaboost/.
future trends, IET Renew. Power Gener. 15 (2021) 915–938. [40] R. Saini, Integrating vegetation indices and spectral features for vegetation
[16] A. Turksoy, A. Teke, A. Alkaya, A comprehensive overview of the dc-dc mapping from multispectral satellite imagery using AdaBoost and random forest
converter-based battery charge balancing methods in electric vehicles, Renew. machine learning classifiers, Geomat. Environ. Eng. 17 (1) (2023) 57–74.
Sustain. Energy Rev. 133 (2020) 110274. [41] D. Yang, Y.J. Wang, R. Pan, R.Y. Chen, Z.H. Chen, A neural network based state-
[17] Y.-H. Park, R.-Y. Kim, Y.-J. Choi, An active cascaded battery voltage balancing of-health estimation of lithium-ion battery in electric vehicles, Energy Procedia
circuit based on multi-winding transformer with small magnetizing inductance, 105 (2017) 2059–2064.
Energies 14 (5) (2021) 1302. [42] Y.N. Sun, J.L. Zhang, K.F. Zhang, H.H. Qi, C.J. Zhang, Battery state of health es-
[18] G. Noh, J. Lee, J.-I. Ha, Design and analysis of single-inductor power converter timation method based on sparse autoencoder and backward propagation fading
for both battery balancing and voltage regulation, IEEE Trans. Ind. Electron. 69 diversity among battery cells, Int. J. Energy Res. 45 (5) (2021) 7651–7662.
(2021) 2874–2884. [43] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
[19] A. Khalid, A. Stevenson, A.I. Sarwat, Performance analysis of commercial passive Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D.
balancing battery management system operation using a hardware-in-the-loop Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine learning
testbed, Energies 14 (23) (2021) 8037. in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
[20] Z.D. Taylor, H. Akhavan-Hejazi, H. Mohsenian-Rad, Power hardware-in-loop sim- [44] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout:
ulation of grid-connected battery systems with reactive power control capability, a simple way to prevent neural networks from overfitting, J. Mach. Learn. Res.
in: North American Power Symposium, NAPS, 2017. 15 (56) (2014) 1929–1958.
[21] N. Bouchhima, M. Schnierle, S. Schulte, K.P. Birke, Active model-based balancing [45] Y. Wei, F. Yang, M.J. Wainwright, Early stopping for kernel boosting algorithms:
strategy for self-reconfigurable batteries, J. Power Sources 322 (2016) 129–137. a general analysis with localized complexities, IEEE Trans. Inform. Theory 6 (10)
[22] A. Pozzi, M. Zambelli, A. Ferrara, D.M. Raimondo, Balancing-aware charging (2019).
strategy for series-connected lithium-Ion cells: a nonlinear model predictive [46] Infineon, IRFP4668, 2023, [Online]. Available: https://www.infineon.com/cms/
control approach, IEEE Trans. Control Syst. Technol. 28 (2020) 1862–1877. en/product/power/mosfet/n-channel/irfp4668/.
[23] F.S.J. Hoekstra, H.J. Bergveld, M.C.F. Donkers, Optimal control of active cell [47] O.L. Gallaga, How to take care of your electric vehicle battery, 2023, [Online].
balancing: extending the range and useful lifetime of a battery pack, IEEE Trans. Available: https://www.wired.com/story/how-to-take-care-electric-vehicle-
Control Syst. Technol. 30 (2022) 2759–2766. battery/.
[24] R.D. Fonso, X. Sui, A.B. Acharya, R. Teodorescu, C. Cecati, Multidimensional
machine learning balancing in smart battery packs, in: Annual Conf. of the IEEE
Industrial Electronics Society, 2021.

10