batteries

Article
Battery Temperature Prediction Using an Adaptive Neuro-Fuzzy
Inference System
Hanwen Zhang 1 , Abbas Fotouhi 1, * , Daniel J. Auger 1 and Matt Lowe 2

1 School of Aerospace, Transport and Manufacturing, Cranfield University, Bedfordshire MK43 0AL, UK
2 Delta Cosworth, Unit 9, Silverstone Technology Park, Silverstone Circuit, Northamptonshire NN12 8GX, UK
* Correspondence: [email protected]; Tel.: +44-(0)1-2347-58092

Abstract: Maintaining batteries within a specific temperature range is vital for safety and efficiency,
as extreme temperatures can degrade a battery’s performance and lifespan. In addition, battery
temperature is the key parameter in battery safety regulations. Battery thermal management sys-
tems (BTMSs) are pivotal in regulating battery temperature. While current BTMSs offer real-time
temperature monitoring, their lack of predictive capability poses a limitation. This study introduces a
novel hybrid system that combines a machine learning-based battery temperature prediction model
with an online battery parameter identification unit. The identification unit continuously updates the
battery’s electrical parameters in real time, enhancing the prediction model’s accuracy. The prediction
model employs an Adaptive Neuro-Fuzzy Inference System (ANFIS) and considers various input
parameters, such as ambient temperature, the battery’s current temperature, internal resistance,
and open-circuit voltage. The model accurately predicts the battery’s future temperature in a finite
time horizon by dynamically adjusting thermal and electrical parameters based on real-time data.
Experimental tests are conducted on Li-ion (NCA and LFP) cylindrical cells across a range of am-
bient temperatures to validate the system’s accuracy under varying conditions, including state of
charge and a dynamic load current. The proposed models prioritise simplicity to ensure real-time
industrial applicability.

Keywords: Li-ion battery; temperature prediction; battery thermal management; ANFIS; neural
network; system identification; electric vehicle
Citation: Zhang, H.; Fotouhi, A.;
Auger, D.J.; Lowe, M. Battery
Temperature Prediction Using an
Adaptive Neuro-Fuzzy Inference
1. Introduction
System. Batteries 2024, 10, 85.
https://doi.org/10.3390/
Considering the drastic development of Li-ion battery technology in recent decades,
batteries10030085 energy density and safety have been two major topics of Li-ion batteries in various ap-
plications such as electric vehicles, mobile devices, power plants, etc. Batteries generate
Academic Editor: Carlos Ziebert
massive heat while discharging, particularly under high C-rates. Consequently, the heat
Received: 6 January 2024 causes the battery temperature to rise. Lithium-ion battery performance is susceptible to
Revised: 20 February 2024 temperature [1,2]. The battery’s electrochemical characteristics dynamically shift following
Accepted: 26 February 2024 temperature change [3]. A high-density battery pack consisting of a large number of cells
Published: 1 March 2024 creates a massive quantity of heat that accumulates inside, which causes the temperature
to rise inside the battery. High temperatures can lead to a decline in battery performance,
shorten the battery life span, and even cause a thermal runaway with critical safety is-
sues [4,5]. Hence, heat generation must not be overlooked [6]. Failure to manage the
Copyright: © 2024 by the authors.
battery’s thermal characteristics causes a risk of performing safely as well as adversely
Licensee MDPI, Basel, Switzerland.
affecting the battery’s lifetime with degradation of power and capacity [7–9]. At low
This article is an open access article
temperatures, the battery electrochemical characteristics tend to decrease the ionic conduc-
distributed under the terms and
tion rate of the electrolyte and slow the chemical reaction at the anode and cathode [10],
conditions of the Creative Commons
Attribution (CC BY) license (https://
which causes battery capacity loss and efficiency drop in both charge and discharge [11,12].
creativecommons.org/licenses/by/
Therefore, battery thermal management is a crucial issue that requires extensive research
4.0/). and development.

Batteries 2024, 10, 85. https://doi.org/10.3390/batteries10030085 https://www.mdpi.com/journal/batteries

Batteries 2024, 10, 85 2 of 22

Thermal management of batteries is also critical in achieving the desired performance

in a wide range of temperature environments and ensuring the battery’s safe operation,
particularly in high-temperature environments. Due to the physical characteristics of the
conventional lithium-ion battery, heat dissipation starts with internal heat diffusing to the
battery shell surface [13], which means that any temperature data captured by sensors on
the battery surface will have a data lag before the temperature rise or drop is reflected.
As a result, it is essential to predict the temperature change in the battery well in
advance to ensure the efficiency of the cooling/heating system. A battery cooling system is
critical for battery safety, mainly to prevent the battery thermal runaway [14]. In addition,
keeping the battery working under a temperature threshold can prolong its lifetime [12,15].
On the other hand, lithium-ion batteries suffer from capacity loss when they work at low
ambient temperatures. Battery preheating is a common way to help the battery avoid such
capacity loss. Due to the structure of the conventional lithium-ion cells, the difference
between the battery’s inner temperature and its surface temperature could reach around
5 ◦ C [16] or even more when the battery experiences over-discharge [17]. By considering
the heat conduction mechanism inside a cylindrical battery, we better understand the time
lag between the heat generation inside the battery and its surface temperature. Hence,
a battery temperature prediction unit can help BTMSs maintain battery safety at high
temperatures by preventing the battery temperature from going beyond the thresholds.
Meanwhile, it can also help improve the preheating process when the battery might face
capacity loss at low temperatures. Machine learning models are a promising solution to
address this problem. The system can predict the battery’s temperature change in a finite
prediction horizon by identifying and using proper input information. This research aims
to design and validate such a system using experimental data.
Surprisingly, there are few studies in the literature in which the problem mentioned
above has been addressed. Feng Fei [18] proposes an electrochemical–thermal neural
network (ETNN) model for battery state estimation. It combines simplified sub-models
for predicting core temperature and approximate terminal voltage with a neural network
to enhance their performance. The ETNN model accurately estimates battery voltage
and core temperature under a wide range of ambient temperatures and discharge rates.
An unscented Kalman filter (UKF) is integrated with the ETNN to achieve reliable co-
estimation of state of charge (SOC) and state of temperature (SOT), and experimental
results demonstrate satisfactory co-estimation performance. Marui Li et al. [19] applied
a sophisticated algorithm involving multi-sequential neural networks and deep learning
for battery state estimation. Andreas et al. [20] adopted a Convolutional Neural Network
(CNN) model for battery temperature prediction, which is trained with cross-domain data
from the simulation, vehicle fleet, and weather stations. They considered the battery cooling
thresholds, practical data, and battery heat generation. Seho Park et al. [21] developed a
technique to predict the battery temperature change trend based on the driving pattern and
adopted the UDDC driving cycle and the cooling effects with empirical function. Mengyi
Wang et al. [22] applied a CNN model for a battery pack’s internal temperature prediction
with good visual analysis and soft sensor technology with big data size that counts the
thermal convection rate and thermal dissipation rate. Kaizheng Fang et al. [23] developed
an ANN model to predict the surface temperature of a Ni-MH battery during charging
under different ambient temperatures. The model was trained using data from charging an
8Ah cylindrical Ni-MH battery under different current rates and ambient temperatures.
Currently, the main function of Battery Thermal Management Systems (BTMSs) is
to monitor the battery state (particularly the temperature) at the present time and take
any necessary action. What the existing BTMSs lack is the ability to predict the battery
temperature in the near future. Such a battery temperature prediction unit can help the
battery cooling system to prepare for the battery temperature rising in the next few minutes,
for instance, due to a high charge/discharge rate at the present time. This research aims to
develop an AI-based battery temperature prediction method to contribute to the robust
and efficient use of battery systems, particularly in an EV application. For that purpose,
Batteries 2024, 10, 85 3 of 22

an Adaptive Neuro-Fuzzy Inference System (ANFIS) is trained to predict the battery’s

surface temperature within a finite time horizon using the history data of the battery’s
charge/discharge and the battery’s internal parameters. A system identification technique
is also used along with the ANFIS model to update the battery parameters continuously.
As a result, the proposed hybrid method is able to deal with any sudden changes in the
battery’s state, which distinguishes it from all other existing methods in the literature.

2. An Overview of the Proposed Concept

The primary objective of the proposed model in this study is to accurately forecast
the surface temperature of a battery in a finite time horizon (for example, 30 s ahead) by
using updated values of battery-related parameters in real time. This must be achieved
despite the limitations of measurement tools as well as the uncertainties around the ambient
temperature and fluctuations of the load current over the entire battery lifetime.
To tackle the complexity of the problem mentioned above, it is imperative to incorpo-
rate several critical variables into the problem formulation. However, the parameters that
can be directly measured from a battery during its operation are limited and only include
the load current, terminal voltage, and surface temperature. This is insufficient to gain
a comprehensive understanding of the battery’s performance and its characteristics. For
instance, identifying the surface temperature of a battery solely based on terminal voltage
and current is not feasible. One approach is to create a mathematical representation of
the battery’s electrical and thermal behaviours. Battery modelling enables us to obtain
additional battery parameters beyond the direct measurements, including the SOC, state of
health (SOH), internal resistance, and open-circuit voltage. Adding these factors makes it
possible to predict the battery’s surface temperature using the existing direct measurements.
An analysis of the battery parameters obtained from battery modelling enables us
to find the relationship between the battery’s surface temperature and internal parame-
ters. This opens up new possibilities for understanding the behaviour of batteries and
optimising their performance. However, there is currently no established method that
can directly demonstrate the correlation between the battery parameters and the battery’s
surface temperature. This has posed a significant challenge to researchers seeking to fully
comprehend the underlying physics of battery operation. Fortunately, advances in machine
learning techniques offer a promising avenue for tackling this complex problem. Machine
learning can identify the relationship between various physical parameters by analysing
historical experimental data [24,25]. By training a machine learning model using a dataset
containing both battery parameters and corresponding surface temperature readings, it
will be possible to develop a predictive model that can accurately forecast the battery’s
surface temperature based on its electrical and thermal parameters.
In this study, an innovative and highly effective system is developed for predicting
battery surface temperature based on a hybrid approach. The proposed system, shown
in Figure 1, includes an online identification unit that enables real-time monitoring of
battery internal parameters, allowing for more accurate and precise temperature prediction.
To ensure the system’s effectiveness and reliability, real-world usage simulation test data
are used.
The hybrid approach developed in this study combines analytical and empirical
models, using the strengths of each to provide the most accurate and reliable predictions
possible. The analytical model is based on fundamental principles of battery operation,
such as the electrochemical heat generation equation. In contrast, the empirical model is
developed using historical data and machine learning techniques. This approach allows
for identifying and integrating a wide range of physical and environmental factors that can
impact battery performance and temperature. Using this hybrid battery surface tempera-
ture prediction system, we can understand the complex interplay between the battery’s
electrical and thermal parameters. By accurately predicting battery surface temperature, the
system can support the development of more effective and efficient battery management
strategies, leading to longer battery life and overall performance improvement.
Batteries 2024, 10, x FOR PEER REVIEW
From Figure 1, the data splits into two groups: directed measured data (i.e
4 of 22
V, current, I, and temperature, T) and the data coming from the battery parame
fication unit (battery resistance, R, open-circuit-voltage, OCV, state of charge,
Batteries 2024, 10, 85 of health,
support SoH, and generated
the development heat, Q).
of more effective andThe direct
efficient measurements
battery managementare 4provided
of 22
strategies, b
tery sensors,
leading to longer which areand
battery life current, terminal voltage,
overall performance and surface temperature. T
improvement.
From Figureunit
identification 1, theprocesses
data splits the
intomeasured
two groups:data
directed
andmeasured
obtains thedatarequired
(i.e. voltage,
parame
From Figure 1, the data splits into two groups: directed measured data (i.e. volt-
V,ascurrent, I, and temperature, T) and the data coming from the battery parameter identi-
age,internal
V, current,resistance and open-circuit
I, and temperature, voltage,
T) and the data coming which arebattery
from the used parameter
for the calculat
fication unit (battery resistance, R, open-circuit-voltage, OCV, state of charge, SoC, state
battery heat
identification generation
unit rate, and
(battery resistance, then estimation OCV,
R, open-circuit-voltage, of the expected
state of charge,change
SoC, in th
of health,
state of SoH,SoH,
health, and generated
and heat,heat,
generated Q). The
Q). direct
The measurements
direct measurements areare
provided
providedbybythe
thebat-
temperature
tery within
sensors, which the prediction
are current, terminal horizon.
voltage, and surface temperature. The online
battery sensors, which are current, terminal voltage, and surface temperature. The online
identification unit processes
identification unit processesthe
themeasured
measureddata
dataand
andobtains
obtains the
the required
required parameters,
parameters, such
such
as
as internal resistance and open-circuit voltage, which are used for the calculation of thethe
internal resistance and open-circuit voltage, which are used for the calculation of
battery
battery heat generation rate,
heat generation rate, and
andthen
thenestimation
estimationofofthe
theexpected
expected change
change in in
thethe battery
battery
temperature within the prediction horizon.
temperature within the prediction horizon.

Figure 1. The proposed concept of battery temperature prediction using direct measure
real-time
Figure
system identification.
Figure 1. The proposed
proposed concept
conceptof
ofbattery
batterytemperature
temperatureprediction
prediction using
using direct
direct measurements
measurements andand
real-time system identification.
real-time system identification.
3. Battery Temperature Prediction Methodology
3. Battery
3. Battery Temperature
Temperature Prediction
PredictionMethodology
Methodology
AsAs mentioned
mentioned earlier,
earlier, the battery
the battery surfacesurface temperature
temperature prediction prediction system req
system requires
As mentioned earlier, the battery surface temperature prediction system requires bat-
tery test
battery testdata
data to to build electricaland
build electrical and thermal
thermal models
models and and
train train the ANFIS.
the ANFIS. So, the So, the
tery test data to build electrical and thermal models and train the ANFIS. So, the proposed
proposed
methodologymethodology
includes includes (1)battery
the battery experimentalsetup
setup and
and test
testprocedures;
methodology includes (1) the(1) the
battery experimental
experimental setup and test procedures; procedures;
(2) the battery (2) t
(2) the battery modelling; (3) the ANFIS model; and (4) the prediction model performance
modelling;
modelling; (3) (3)
the the ANFIS
ANFIS model; model;
and (4)and (4) the prediction
the prediction model performance
model performance evaluation. In eval
evaluation. In the following, these parts are discussed for every detail of the whole concept
thefollowing,
the
shown following,
in Figure 2.these
these partsparts
Then, are discussed
arebattery
the discussed for everyfor
temperature every
detail detail
of the
prediction whole
results ofconcept
are the whole
shown
demonstrated concept
in Fig- sho
and
ure 2. Then, the battery temperature prediction results are demonstrated
ure 2. Then, the battery temperature prediction results are demonstrated and disc
discussed. and discussed.

Figure 2. Samsung NCA cylindrical cells.

3.1. Battery
Figure
Figure Experimental
2. Samsung
2. Samsung NCANCA Setup and Test Procedures
cylindrical
cylindrical cells. cells.
3.1.1. NCA Battery
3.1. Battery Experimental Setup and Test Procedures
3.1.The
3.1.1. Battery
NCAfirst Experimental
battery used inSetup
Battery and Test
this study is a Procedures
Lithium Nickel Cobalt Aluminium Oxide
(LiNiCoAlO ), commonly known as an NCA battery. NCA batteries are known for their
3.1.1.
TheNCA Batteryused in this study is a Lithium Nickel Cobalt Aluminium Oxide
first battery
high specific energy and long lifespan, making them an ideal candidate for EV power-
(LiNiCoAlO2 ), commonly known as an NCA battery. NCA batteries are known for their
trains. The first battery
This commercial
high specific
used
product
energy and long isinwidely
lifespan,
this study
makingapplied
is aEVs
Lithium
them aninideal
Nickel Cobalt
and significantly
candidate
Aluminiu
advances
for EV powertrains. in
(LiNiCoAlO ), commonly known as an NCA battery. NCA batteries are known
high specific energy and long lifespan, making them an ideal candidate for E
trains. This commercial product is widely applied in EVs and significantly ad
 battery manufacturing [26]. Figure 2 displays five NCA cells used in the experim
ducted in this research. Table 1 contains the detailed specification of the NCA
specifications include critical information such as the battery’s capacity, volta
and charge/discharge rates. Understanding the cell’s technical specifications is
Batteries 2024, 10, 85 5 of 22
to evaluate its performance and reliability.

Table
This 1. Specifications
commercial product of the NCA
is widely cell. in EVs and significantly advances in battery
applied
manufacturing [26]. Figure 2 displays five NCA cells used in the experiments conducted
Parameter
in this research. Table 1 contains the detailed specification of the NCA cells. The spec-
Value
ifications include critical Rated capacity
information such (mAh)
as the battery’s capacity, voltage limits, and4800 mA
charge/discharge rates. Understanding the cell’s
Nominal voltage (V) technical specifications is essential to 3.62 V
evaluate its performance and reliability.
Max charging voltage (V) 4.2 V
Cut-off
Table 1. Specifications of the NCA cell. voltage 2.5 V
Charge: 0 t
Operating
Parameter temperature (surface) (°C) Value
Rated capacity (mAh) 4800 mAh
Discharge: −3
NominalStandard
voltage (V)discharging current 3.62 V 0.2 C
Max charging voltage (V) 4.2 V
Standard charging current
Cut-off voltage 2.5 V
C/3
Maximum discharge rate
Operating temperature (surface) (◦ C)
Charge: 0 to 45 2C
Discharge: −30 to 60
Maximum charge rate
Standard discharging current 0.2 C
1C
Peak discharging
Standard current (30 s, 10 s) @
charging current SOC 50% C/3 42 A, 54
Maximum discharge rate 2C
Maximum charge rate 1C
3.1.2. LFP Battery
Peak discharging current (30 s, 10 s) @ SOC 50% 42 A, 54 A

The Lithium Iron Phosphate (LFP) battery has recently obtained a notabl
3.1.2. LFP Battery
ancy in the EV market [26]. The material of an LFP battery is made of lithium-
The Lithium Iron Phosphate (LFP) battery has recently obtained a notable ascendancy
phate (LiFePO ). LFP batteries are also known as lithium ferro-phosphate or p
in the EV market [26]. The material of an LFP battery is made of lithium-ion phosphate
batteries.
(LiFePO Compared with other types of lithium-ion batteries, LFP batteries off
4 ). LFP batteries are also known as lithium ferro-phosphate or phosphate batteries.
advantages
Compared withincluding
other typesaoflong cycle life
lithium-ion and outstanding
batteries, thermal
LFP batteries offer severalstability ensuring
advantages
including a long cycle life and outstanding thermal stability ensuring excellent
safety. LFP batteries are less likely to experience thermal runaway or catch fire, ma safety.
LFP batteries are less likely to experience thermal runaway or catch fire, making them a
a popular choice for EVs. Although LFP batteries have a lower potential, electrical,
popular choice for EVs. Although LFP batteries have a lower potential, electrical, and ionic
conductivity
conductivity compared
compared with with
ternaryternary lithium-ion
lithium-ion batteries,
batteries, they they have
have excellent excellent perfo
performance
terms
in termsofoflifespan
lifespanandand SOHSOH consistency, maintaining
consistency, maintaining their
their capacity
capacity over over numerous c
numerous
charge and discharge cycles. These reliable and cost-effective characteristics
discharge cycles. These reliable and cost-effective characteristics make LFP make LFPbatterie
batteries a widely used option for energy storage systems. In this study, additional tests
used option for energy storage systems. In this study, additional tests are conducted
are conducted on the RS Pro lithium-ion iron phosphate cells shown in Figure 3. The
Promodel
cell lithium-ion
is LIFePOiron phosphate cells shown in Figure 3. The cell model is LIFePO4-2
4 -26650 with a capacity of 3399 mAh, and other specifications are
a capacity
presented in of 3399
Table 2. mAh, and other specifications are presented in Table 2.

Figure
Figure 3. 3.
RSRS
ProPro
LFP LFP
cells. cells.
Batteries 2024, 10, 85 6 of 22

Table 2. RS Pro LFP cell specifications.

Item Value Note

Standard capacity 3200 mAh 0.5 C (current value of 3200 mAh at 1 C)
Capacity range 3100~3300 mAh 0.5 C
Standard voltage 3.2 V
Alternating internal resistance ≤30 mΩ With PTC
Cut-off
3.65 ± 0.05 V Constant current charge to 3.65 V at 0.5 C
Charge conditions voltage
Constant voltage charge to stop until 0.01 C mA
Cut-off current 0.01 C
Discharge cut-off voltage 2.5 V
100% DOD, the residual capacity is no less than
Cycle characteristic 2000 times
80% of rated capacity at 1 C-rate
Max. continuous discharge current 9.6 A
Pulse discharge current 15 A, 10 s
Charge: 0~55 ◦ C
Working temperature
Discharge: −20~60 ◦ C
Storage temperature −20–45 ◦ C
Battery weight 86 g (approx.)

3.1.3. Experimental Setup

To meet the demands of our battery testing objectives, we have developed a compre-
hensive battery testing system. Figure 4 shows the configuration of the battery testing
system, which consists of several key components that work together to achieve our exper-
imental goals. The central component of the system is a personal computer (PC), which
serves as the terminal control centre. The PC is equipped with the necessary software, i.e.
MATLAB/Simulink 2020b, and communication cables, which allow it to communicate with
the other components of the system, such as the power supply and the thermocouple data
logger. The PC sends charge/discharge commands to the power supply, which then applies
the required load current to the battery. The battery is inside a thermal chamber to control
the ambient temperature during all tests. The battery fixture is connected to the power
supply via wires and is designed to hold the battery in place during the testing process. To
monitor the battery’s performance during the test, the power supply measures the battery
voltage and current every second. This information is sent back to the terminal control
centre for recording and monitoring purposes, allowing us to track the performance of the
battery over time. In addition to measuring the battery voltage and current, temperature
signals are collected during the testing process. Thermocouples are attached to the surface
of the battery, allowing us to monitor the temperature of the battery during each test. The
thermocouples are connected to a data logger, which sends the temperature signal back
to the PC for recording and analysis. In general, the battery testing system is designed to
collect accurate and reliable data related to the performance of the test batteries during
Batteries 2024, 10, x FOR PEER REVIEW 7 of 22
the experiments.

Figure 4. Configuration of the battery test setup.

Figure 4. Configuration of the battery test setup.
3.1.4. Experimental Test Procedures
To successfully apply battery research, it is important to conduct performance tests and
identify the characteristics of the battery being developed. This helps to ensure that the battery
performs as intended and meets the requirements for its intended use. The battery testing plan
is typically divided into four stages mentioned below, each serving a specific purpose in as-
sessing the battery’s performance and identifying any potential issues.
1. The battery undergoes constant current discharge and mixed pulse discharge tests to
Batteries 2024, 10, 85 7 of 22

Figure 4. Configuration of the battery test setup.

3.1.4. Experimental Test Procedures
3.1.4.To
Experimental Test
successfully Procedures
apply battery research, it is important to conduct performance tests
and identify the characteristics
To successfully apply batteryofresearch,
the battery
it is being developed.
important to conductThis helps to ensure
performance that the
tests and
battery
identify performs as intended
the characteristics of theand meets
battery the
being requirements
developed. for its
This helps tointended
ensure thatuse. The battery
the battery
performs
testing as intended
plan and meets
is typically dividedtheinto
requirements
four stages for its intended use.
mentioned The each
below, battery testing aplan
serving specific
is typically
purpose in divided intothe
assessing four stages mentioned
battery’s performance below,
andeach serving aany
identifying specific purpose
potential in as-
issues.
sessing the battery’s performance and identifying any potential issues.
1. The battery undergoes constant current discharge and mixed pulse discharge tests to
1. The battery
assess undergoes constant current discharge and mixed pulse discharge tests to
its performance.
2. assess its performance.
The battery experiences the WLTP test to evaluate its performance under a realistic
2. The battery
driving experiences the WLTP test to evaluate its performance under a realistic
condition.
driving condition.
The Worldwide Harmonised Light Vehicles Test Procedure (WLTP) [27] is a globally
The Worldwide Harmonised Light Vehicles Test Procedure (WLTP) [27] is a globally
recognised standard for testing the fuel economy and emissions of light-duty vehicles.
recognised standard for testing the fuel economy and emissions of light-duty vehicles. This
This
standard definesdefines
standard a driving
a driving cycle
cycle that that of
consists consists of four low-speed
four segments: segments:urban low-speed
driving,urban
driving,
medium medium urbanhighway
urban driving, driving,cruising,
highway cruising,
and extra-highandspeed
extra-high
[28], asspeed [28],in
illustrated asFigure
illustrated
in Figure
5. The 5. The simulated
simulated WLTP batteryWLTPloadbattery
currentload current
profile, shownprofile, shown
in Figure 6, isinobtained
Figure 6, is obtained
after the
after the simulation
simulation of EV
of a typical a typical EVdriving
over the over thecycle.
driving cycle.
In this In this
profile, profile, acurrent
a negative negative current
value
value indicates that battery discharge occurs during vehicle acceleration.
indicates that battery discharge occurs during vehicle acceleration. The magnitude of the The magnitude
of the current
current value corresponds
value corresponds to the
to the rate rate of acceleration.
of acceleration. Conversely,Conversely,
a positive acurrent
positive current
value
value indicates that battery charging occurs during energy regeneration.
indicates that battery charging occurs during energy regeneration. This process is known as This process is
known as regenerative braking, which captures the energy generated during
regenerative braking, which captures the energy generated during braking and stores it in braking and
stores it in for
the battery thelater
battery
use.for later use.
Velocity [m/s]

Batteries 2024, 10, x FOR PEER REVIEW 8 of 22

Figure 5.
Figure 5. WLTP
WLTPdriving
drivingcycle.
cycle.

-2

-4

-6

-8

-10
0 200 400 600 800 1000 1200 1400 1600 1800
Time [s]

Figure 6.
Figure 6. Simulated
SimulatedWLTP
WLTPbattery
batteryload current
load profile.
current profile.

The WLTP
The WLTPtest testused
usedininthis research
this is intended
research to provide
is intended a more
to provide accurate
a more picturepicture
accurate of
real-world EV applications and consequently contribute to the development
of real-world EV applications and consequently contribute to the development of a more of a more practi-
cal temperature
practical prediction
temperature model formodel
prediction such anforindustrial
such an application. As the tests were
industrial application. per- tests
As the
formed on single cells, the current requirement was scaled down to
were performed on single cells, the current requirement was scaled down to match match the capabilities of the
each cell. Each WLTP test was conducted for repeated cycles until the battery was
capabilities of each cell. Each WLTP test was conducted for repeated cycles until the battery depleted.
To further simulate real-life usage, the tests are carried out at various ambient tem-
was depleted.
peratures ranging
To further from −10
simulate °C to 40
real-life °C. This
usage, thehelps
tests to
areidentify
carriedthe outimpact of theambient
at various ambient tem-
temperature on battery performance. The results from these tests can be used to optimise
peratures ranging from −10 ◦ C to 40 ◦ C. This helps to identify the impact of the ambient
the design and performance of batteries, especially for the EV application.
temperature on battery performance. The results from these tests can be used to optimise
As mentioned earlier, the battery test is primarily controlled using a MATLAB code,
the design and performance of batteries, especially for the EV application.
which sends the pre-programmed commands to the power supply, monitors the battery volt-
age, and controls the charge and discharge current during each test. An example WLTP test is
shown in Figure 7. In that figure, the cell is fully charged before starting the WLTP discharge
profile and the whole charge/discharge cycle is repeated several times in this case. As men-
tioned above, the required data for battery temperature prediction consists of two parts. The
first part is provided by direct measurements, the data recorded by the power supply and the
 The WLTP test used in this research is intended to provide a more accurate picture of
real-world EV applications and consequently contribute to the development of a more practi-
cal temperature prediction model for such an industrial application. As the tests were per-
formed on single cells, the current requirement was scaled down to match the capabilities of
Batteries 2024, 10, 85 each cell. Each WLTP test was conducted for repeated cycles until the battery was depleted. 8 of 22
To further simulate real-life usage, the tests are carried out at various ambient tem-
peratures ranging from −10 °C to 40 °C. This helps to identify the impact of the ambient
temperature on battery
As mentioned performance.
earlier, Thetest
the battery results from these
is primarily tests can be
controlled useda to
using optimise code,
MATLAB
the design and performance of batteries, especially for the EV application.
which sends the pre-programmed commands to the power supply, monitors the battery
As mentioned
voltage, and controlsearlier, the battery
the charge test is primarily
and discharge current controlled
during eachusing a MATLAB
test. An examplecode,WLTP
which sends the pre-programmed commands to the power supply, monitors the battery volt-
test is shown in Figure 7. In that figure, the cell is fully charged before starting the WLTP
age, and controls the charge and discharge current during each test. An example WLTP test is
discharge profile and the whole charge/discharge cycle is repeated several times in this
shown in Figure 7. In that figure, the cell is fully charged before starting the WLTP discharge
case. As mentioned above, the required data for battery temperature prediction consists
profile and the whole charge/discharge cycle is repeated several times in this case. As men-
of two parts. The first part is provided by direct measurements, the data recorded by the
tioned above, the required data for battery temperature prediction consists of two parts. The
power supply and the thermocouples. These data include the battery terminal voltage,
first part is provided by direct measurements, the data recorded by the power supply and the
the load current,
thermocouples. and
These theinclude
data battery
thesurface
batterytemperature.
terminal voltage, The thethermocouples
load current, andconnect
the bat- with
the temperature data logger, which converts the voltage signal from
tery surface temperature. The thermocouples connect with the temperature data logger, whichthe thermocouple to
the digital temperature data and sends it to the PC for recording. All
converts the voltage signal from the thermocouple to the digital temperature data and sends measurements are
recorded
it to the PC infor
therecording.
form of time series with a sampling
All measurements are recordedrateinofthe
1 Hz
formforofboth
time the power
series with supply
a
and the temperature data logger.
sampling rate of 1 Hz for both the power supply and the temperature data logger.

2
0
-2
-4
-6
-8
0 5 10 15 20 25 30 35 40
time (h)
4

3.5

2.5
0 5 10 15 20 25 30 35 40
time (h)

24

22

20
0 5 10 15 20 25 30 35 40
time (h)

Figure 7.7.Battery
Figure load load
Battery current, terminalterminal
current, voltage, and temperature
voltage, signals recorded
and temperature duringrecorded
signals a WLTP test.
during a
WLTP test.

3.2. Battery Modelling for Temperature Prediction

3.2.1. Battery Electrical Modelling
In this study, we focused on battery-equivalent circuit models due to their compatibil-
ity with the online identification process used in this research. An equivalent circuit model
comprises a network of resistors and capacitors that represent the electrical properties
of the battery. By analysing the voltage and current measurements of the battery, the
equivalent circuit model can accurately predict the battery’s performance. Although the
electrochemical models are not used in this study for their complexity, it is important to
acknowledge their accuracy, which makes them more suitable for offline applications. A
battery electrochemical model takes into account the internal chemical processes of the
battery, providing a more detailed analysis of its behaviour. However, due to the need
for light computational effort in the online identification process used in this study, the
equivalent circuit modelling approach sounds like a more practical and effective option.
After a detailed analysis of the battery test data, it was found that there is no direct
correlation between the battery surface temperature and the terminal voltage and/or the
load current signals. The mission of this study is to build a battery electrical model to
obtain additional battery parameters, which can help us to better understand the battery’s
performance and, specifically, the heat generation inside it. According to the literature,
the two primary parameters of interest are the internal resistance and the open-circuit
voltage (OCV). These parameters cannot be measured directly using sensors during the
tests. However, they can be calculated from the measured current (I) and terminal voltage
 After a detailed analysis of the battery test data, it was found that there is no direct
correlation between the battery surface temperature and the terminal voltage and/or the
load current signals. The mission of this study is to build a battery electrical model to
obtain additional battery parameters, which can help us to better understand the battery’s
Batteries 2024, 10, 85 performance and, specifically, the heat generation inside it. According to the literature, 9 of 22
the two primary parameters of interest are the internal resistance and the open-circuit
voltage (OCV). These parameters cannot be measured directly using sensors during the
tests. However, they can be calculated from the measured current (I) and terminal voltage
(V). The internal resistance represents the battery’s resistance to the flow of current, while
(V). The internal resistance represents the battery’s resistance to the flow of current, while the
the OCV represents the voltage of the battery when it is not connected to any load. These
OCV represents the voltage of the battery when it is not connected to any load. These param-
parameters are critical in evaluating the battery’s performance and predicting its behaviour
eters are critical in evaluating the battery’s performance and predicting its behaviour under
under different conditions.
different conditions.
The Rint model, shown in Figure 8, is a widely used mathematical model for recharge-
The Rint model, shown in Figure 8, is a widely used mathematical model for re-
able batteries that has proven to be an effective tool for optimising battery performance [29].
chargeable batteries that has proven to be an effective tool for optimising battery perfor-
The Rint
mance [29].battery
The Rintmodel hasmodel
battery only has
twoonly
parameters: an ideal
two parameters: voltage
an ideal to characterise
voltage to charac- the
open-circuit voltage and one resistance [30], as shown in Figure 8. The variable
terise the open-circuit voltage and one resistance [30], as shown in Figure 8. The variable IL is the
load current, whereas the parameter V represents the battery’s terminal voltage.
𝐼 is the load current, whereas the parameter 𝑉 represents the battery’s terminal voltage.
t

Figure
Figure 8.8.Rint
Rintmodel.
model.

Polarisation
Polarisation is is a special
a special effecteffect observed
observed in batteries,
in batteries, especiallyespecially duringwhich
during discharge, discharge,
can affect
which can theaffect
accuracy
the of the battery’s
accuracy electrical
of the parameters
battery’s electrical [31]. Adding one[31].
parameters RC model
Adding to the
one RC
Rint model can improve it by considering the effect of polarisation and
model to the Rint model can improve it by considering the effect of polarisation and this this model is called the
Thevenin
model is model
called[29],
the which
Theveninis illustrated in Figure
model [29], which 9. Inisthis figure, 𝑉 inis Figure
illustrated the battery
9. In terminal
this figure,
voltage, 𝑉 represents the battery open-circuit voltage, 𝐼 is the
Vt is the battery terminal voltage, Voc represents the battery open-circuit voltage, charge/discharge load cur- IL is
rent, 𝑅 is the ohmic resistance, and 𝑅 and 𝐶 stand for equivalent
the charge/discharge load current, Ro is the ohmic resistance, and R p and C p stand for polarisation resistance
and capacitance,
equivalent respectively
polarisation [32]. In and
resistance several studies, therespectively
capacitance, Thevenin model [32]. isInused for battery
several studies, the
electrical
Theveninparameter
model isanalysis
used forbecause
batteryofelectrical
its acceptable level of analysis
parameter accuracy because
[33,34]. Inofthis
itsstudy,
acceptable
the Thevenin
level model[33,34].
of accuracy is only used
In thisforstudy,
offlinethe
battery parameter
Thevenin model identification
is only used to generate
for offlinea ref-
battery
erence benchmark. The mathematical presentation of this model is as
parameter identification to generate a reference benchmark. The mathematical presentationfollows [35]:
of this model is as follows [35]: 𝑉 =𝑉 −𝑅 𝐼 −𝑉
(𝑑𝑉 1 1 (1)
=− −R 𝑉 o+IL 𝐶− 𝐼Vp
𝑑𝑡Vt = V𝑅oc 𝐶
Batteries 2024, 10, x FOR PEER REVIEW dVp 1 1 10 of 22 (1)
dt = − R p C p Vp + C p IL

Figure 9.
Figure 9. Thevenin
Theveninmodel.
model.

One of
One ofthethemost
mostcritical aspects
critical of the
aspects of Rint
the model is its ability
Rint model is its to predict
ability to battery
predictper-
battery
formance under a dynamic load profile accurately. When a battery
performance under a dynamic load profile accurately. When a battery is subjected is subjected to a dy- to a
namic load,
dynamic it experiences
load, it experiencesquick voltage
quick drops
voltage andand
drops jumps due due
jumps to itstoohmic resistance.
its ohmic The The
resistance.
Rint model can incorporate such voltage drops/jumps into its calculation.
Rint model can incorporate such voltage drops/jumps into its calculation. By incorporating By incorporat-
ing effect
the the effect of ageing
of ageing on Rint
on the the Rint model’s
model’s internal
internal resistance,
resistance, it can
it can easily
easily consider
consider thethe
battery
battery degradation as well.
degradation as well.
For the proposed battery temperature prediction methodology, the battery character-
For the proposed battery temperature prediction methodology, the battery characteris-
istic data are crucial for the heat generation model, where the battery parameter identifi-
tic data are crucial for the heat generation model, where the battery parameter identification
cation unit plays a key role. In this study, a 90 s time window is considered as battery
unit plays a key role. In this study, a 90 s time window is considered as battery usage
usage history to perform the parameter identification process in real time. Longer
history to perform the parameter identification process in real time. Longer timeframes
timeframes (5 min or more) do not make sense because very old data are almost irrelevant
(5
in min or more)
this case. On thedo other
not make
hand,sense
very because
short timevery old data
windows (a are
few almost
seconds) irrelevant
might also in face
this case.
On the other hand, very short time windows (a few seconds) might
some issues such as not being able to identify the battery model due to lack of information.also face some issues
such as not being able to identify the battery model due to lack
Therefore, something in between was needed. Our suggestion is a time window between of information. Therefore,
something
1 and 3 min. inInbetween
a dynamicwashigh
needed.
C-rate Our suggestion
current is a as
profile such time
the window
WLTP driving between cycle1 and
3scenario,
min. Inthe a dynamic highresults
identification C-ratemight
current profile
show such
higher as the
errors. WLTP driving
However, cycle scenario,
it is still acceptable
after applying an averaging filter as shown in Figure 10. The figure illustrates identified
values for the NCA cell’s internal resistance during a WLTP test, from a fully charged state
until depletion. The resistance values are averaged over a short time window (i.e., 90 s in
this case). According to the result, the resistance remains almost constant between 30%
and 100% SOC and it gradually increases at lower SOCs.
 istic data are crucial for the heat generation model, where the battery parameter identifi-
cation unit plays a key role. In this study, a 90 s time window is considered as battery
usage history to perform the parameter identification process in real time. Longer
timeframes (5 min or more) do not make sense because very old data are almost irrelevant
in this case. On the other hand, very short time windows (a few seconds) might also face
Batteries 2024, 10, 85
some issues such as not being able to identify the battery model due to lack of information.10 of 22
Therefore, something in between was needed. Our suggestion is a time window between
1 and 3 min. In a dynamic high C-rate current profile such as the WLTP driving cycle
the identification
scenario, results results
the identification might might
showshowhigher errors.
higher However,
errors. However, it it
isisstill
stillacceptable
acceptable after
applying
after an averaging
applying an averaging filterfilter
as shown
as shownin Figure 10. 10.
in Figure TheThe
figure illustrates
figure illustrates identified
identified values
for thefor
values NCA cell’s cell’s
the NCA internal resistance
internal during
resistance a WLTP
during a WLTP test, from
test, froma afully
fullycharged
charged state
state until
until depletion.
depletion. The The resistance
resistance values
values areare averagedover
averaged overaashort
shorttime
time window
window (i.e., (i.e.,90
90ssinin this
this
case).case). According
According to the
to the result,
result, thethe resistanceremains
resistance remainsalmost
almost constant
constant between
between 30% 30% and
and 100%
100% SOCSOC andand it gradually
it gradually increases
increases at at lower
lower SOCs.
SOCs.
Internal Resistance(ohm)

Figure
Figure 10.
10.The
Theinternal
internalresistance of of
resistance thethe
NCA cellcell
NCA identified during
identified the WLTP
during test. test.
the WLTP

In
Inthe
thefollowing
followingtwotwosubsections,
subsections, thethe
internal resistance
internal and and
resistance the open-circuit voltage
the open-circuit voltage
are
are explored
exploredwithwiththe
thetwo
twodifferent
differentbattery
batteryelectrical models,
electrical and and
models, thenthen
the outcomes are are
the outcomes
compared
comparedand andanalysed.
analysed.While therethere
While are some
are differences betweenbetween
some differences the two models
the twoin terms
models in
of their electrical parameters, that difference is not significant enough to affect the
terms of their electrical parameters, that difference is not significant enough to affect thesubsequent
stages of the battery
subsequent stages temperature
of the battery prediction system.prediction
temperature However, given the complexity
system. However,ofgiven the the
system and the computational requirements, the algorithm used for this analysis
complexity of the system and the computational requirements, the algorithm used for must be able
to perform the necessary calculations quickly and respond promptly. After careful considera-
this analysis must be able to perform the necessary calculations quickly and respond
tion and online identification, the Rint model was selected as the primary algorithm for battery
promptly. After careful consideration and online identification, the Rint model was selected
electrical parameter analysis. This model satisfies the criteria of the hybrid battery temperature
as the primary algorithm for battery electrical parameter analysis. This model satisfies the
prediction system by providing fast and accurate responses to changes in the battery’s electri-
criteria of the hybrid battery temperature prediction system by providing fast and accurate
cal parameters. Therefore, the Rint model is used throughout the remainder of this paper for
responses to changes in the battery’s electrical parameters. Therefore, the Rint model is
the online identification unit.
used throughout the remainder of this paper for the online identification unit.
For a dynamic load profile such as the WLTP current profile, the battery OCV is not
easy to identify. Similar to the battery resistance identification, a 90 s historical time window
is considered, and an averaging filter is applied to obtain the OCV in real time during a
dynamic load profile. In the laboratory environment, pulse tests can be applied to identify
the correct values of the battery model’s parameters called offline identification. The
mixed pulse discharge test and constant discharge test are used in the offline identification
process. Then, battery characterisation is performed by analysing the battery data to
determine the internal resistance, polarisation effect, and open-circuit voltage. The offline
identification results can also be used as a reference to validate the online identification
outcomes. It should be noted that the offline identification method is not feasible in
real-time applications.
Figures 11 and 12 show a comparison of the OCV values between the online and offline
identification obtained at three different temperatures. The online identification curves
have more fluctuations due to the variations in the WLTP current profile. Both test data are
collected in the same battery status and the same ambient temperature. Generally speaking,
the two curves have the same trends and the exact start and end values, which can verify
the online identification results to be used later for the battery temperature prediction.
 Figures 11 and 12 show a comparison of the OCV values between the online and offline
Figures 11 and 12 show a comparison of the OCV values between the online and offline
identification obtained at three different temperatures. The online identification curves have
identification obtained at three different temperatures. The online identification curves have
more fluctuations due to the variations in the WLTP current profile. Both test data are col-
more fluctuations due to the variations in the WLTP current profile. Both test data are col-
lected in the same battery status and the same ambient temperature. Generally speaking,
lected in the same battery status and the same ambient temperature. Generally speaking,
the two curves have the same trends and the exact start and end values, which can verify
Batteries 2024, 10, 85 the two curves have the same trends and the exact start and end values, which can verify
the online identification results to be used later for the battery temperature prediction. 11 of 22
the online identification results to be used later for the battery temperature prediction.

(V)(V)
Voc
Voc

Figure 11. OCV online identification results at various temperatures.

Figure 11.
Figure 11. OCV
OCVonline
onlineidentification results
identification at various
results temperatures.
at various temperatures.
(V)(V)
Voc
Voc

Figure
Figure 12.
12.OCV
OCVoffline
offlineidentification results
identification at various
results temperatures.
at various temperatures.
Figure 12. OCV offline identification results at various temperatures.

For battery
For batteryheat
heatgeneration
generationcalculation,
calculation,the
thebattery
batteryinternal
internalresistance
resistanceis is
a avital
vitalpa-
param-
For battery heat generation calculation, the battery internal resistance is a vital pa-
rameter to consider
eter to consider inin additionto
addition tothe
the current
currentandandvoltage.
voltage.Also, battery
Also, internal
battery resistance
internal resistance
rameter to consider in addition to the current and voltage. Also, battery internal resistance
changes
changes with
withthetheambient
ambient temperature
temperature and and
charge/discharge current,current,
charge/discharge especially under theunder
especially
changes with the ambient temperature and charge/discharge current, especially under the
WLTP
the test. The battery resistance changes as the discharge current fluctuates. In fluctuates.
addition,
WLTP test. The battery resistance changes as the discharge current fluctuates. In addition, In
WLTP test. The battery resistance changes as the discharge current
considering
addition, the battery
considering ageing effects, the battery heat generation model cannot use a fixed
considering the battery the battery
ageing ageing
effects, effects,
the battery thegeneration
heat battery heat
modelgeneration
cannot use model
a fixedcannot
resistance
use value.
a fixedvalue. A fixed
resistance internal resistance might work for heat generation calculation in
resistance A fixedvalue.
internalA resistance
fixed internal
mightresistance might
work for heat work for
generation heat generation
calculation in
acalculation
standard situation like normal
in a standard ambient
situation temperature
like normal and temperature
ambient a non-dynamic andload current
a non-dynamic
a standard situation like normal ambient temperature and a non-dynamic load current
load current profile. However, it cannot handle extreme temperatures as well as dynamic
load current profiles, which have remarkable impacts on battery internal resistance. The
online identification unit can solve that issue by regularly updating the resistance value to
be used more precisely for the heat generation calculation.
The battery’s thermal behaviour varies at different SOC levels impacting the battery’s
surface temperature prediction. It is crucial to highlight that an estimated SOC is not
employed as an input for the battery surface temperature prediction in this study. Although
adding a SOC estimator to the proposed temperature prediction system is possible, it is not
essential because the system can identify the battery model’s parameters automatically. The
proposed system is designed to work in a way that is independent of any state estimator.
This is an advantage because no additional error will be added to the system due to the state
estimation accuracy limitations. It should be noted that despite excluding SOC estimation
from the temperature prediction system, the proposed system considers the impacts of any
changes in the battery’s states by continuously identifying its parameters in real time. This
includes the changes in battery SOC and the battery’s state of health, which are discussed
in the next section.

3.2.2. Battery Thermal Modelling

We should consider the time required for the heat transfer from the battery’s core to its
surface for battery temperature prediction. The heat generation inside a lithium-ion battery
Batteries 2024, 10, 85 12 of 22

contains three main parts: (i) reaction heat Qr ; (ii) polarisation heat Q P ; and (iii) Joule heat
Q J [36]. So, there is always a lag in temperature change on the cell’s surface compared
with its core temperature. Considering this potential impact on the battery’s temperature
prediction, we need the battery thermal model to calculate the heat generation to be used
for surface temperature prediction. Two battery heat generation models have been applied
and tested in this research.
Battery heat generation is a complicated process with many discrete phases involving
the battery energy content. The battery heat generation process can be assumed to combine
these processes: chemical reactions, phase changes, mixing processes, electrical work, and
heat transfer with the ambient air.
Bernardi electro-chemical heat generation is a type of heat generation that occurs
in batteries during the charge and discharge processes. This type of heat generation is
down to the internal resistance of the cells and the resistance of the connecting wires, as
well as the chemical reactions that take place during charge and discharge. Bernardi heat
generation depends on the load current through the battery, the battery SOC, and the
power consumption. It can be calculated by measuring the cell’s voltage and current during
both charge and discharge. Regarding the battery heat generation theory, presented in
reference [36], in typical circumstances, the factors influencing thermal generation can be
broken down into four elements:
• Heat value Qr due to entropy changes from electrochemical reactions;
• Polarisation heat value Q p ;
• Battery side reactions, self-discharge, and such due to electrolyte decomposition,
which are named Qs ;
• Joule heat value Q j , which is caused by the ions shift between the anode and the
cathode through the electrolyte with resistance called battery internal equivalent
resistance, is also overpotential.
Therefore, the total generated heat Q could be written as follows:
 
∂Eε
Q = Qr + Q p + Qs + Q j = nFT + Q p + Qs + Q j (2)
∂T

In short, in the scenario of a Li-ion battery, the Qs is necessarily zero. During the
charging phase, if the charge current rate is very low, the polarisation heat Q p becomes
relatively small, and the battery is in a heat absorption condition at the beginning of the
charge, which could be kept at a low temperature. But when the charge rate increases,
both the Q p and Q j will elevate at the same time. In the discharge process, the battery
temperature will rise with a non-linear trend.
Generally, the heat generation rate of a single cell during discharge can be divided into
two groups, reversible heat and irreversible heat, as defined by the researcher Bernardi [37]:

∂OCV
QGen = Qirr + Qrev = I ·(U − OCV ) + I · T · (3)
∂T
where parameter Qirr is the irreversible heat generation caused by the losses, which happen
in the current collectors as well as the active materials. On the other hand, Qrev is the
reversible heat generation, which happens because of a change in entropy during the inter-
calation and deintercalation of the lithium ions. The parameter U is the terminal voltage in
volt (V), T is the core temperature in Kelvin (K), and ∂OCV
∂T is the entropy coefficient (V/K),
which depends on the SOC and has a significant influence on the temperature behaviour of
the model [38].
The entropy coefficient ∂OCV/∂T = −0.14 mV/K was determined by measuring the
open-circuit voltage as a function of temperature at SOC of 20% for the NCA battery [38,39].
 reversible heat generation, which happens because of a change in entropy during the in-
tercalation and deintercalation of the lithium ions. The parameter U is the terminal voltage
in volt (V), T is the core temperature in Kelvin (K), and is the entropy coefficient
(V/K), which depends on the SOC and has a significant influence on the temperature be-
Batteries 2024, 10, 85 haviour of the model [38].
13 of 22
The entropy coefficient 𝜕𝑂𝐶𝑉/𝜕𝑇 = −0.14 𝑚𝑉/𝐾 was determined by measuring the
open-circuit voltage as a function of temperature at SOC of 20% for the NCA battery [38,39].

3.3. ANFISModel
3.3. ANFIS Model
Based on
Based on the
thewhole
wholeconcept
conceptpresented
presented in in Figure
Figure 1, the
1, the hybrid
hybrid battery
battery temperature
temperature
prediction systemneeds
prediction system needsmachine
machine learning
learning techniques
techniques to identify
to identify the relationship
the relationship between
between
the
the battery’s futuretemperature
battery’s future temperature and
and itsits current
current state.
state. Figure
Figure 13 presents
13 presents a schematic
a schematic of of
the machinelearning
the machine learningmodelmodelrequired
required in in
thethe hybrid
hybrid battery
battery temperature
temperature prediction
prediction system.
system.
As
As shown
shown in inthe
thefigure,
figure,three
three inputs
inputs areare considered
considered here:here:
A) the(A) the ambient
ambient temperature,
temperature, B)
(B) battery’s
battery’s surface
surface temperature
temperature at present
at the the presenttime,time,
and and
C) the(C)amount
the amount of heat
of heat generated
generated
inside thebattery.
inside the battery.The The only
only output
output is shown
is shown by T, which
by T, which is the temperature
is the battery battery temperature
pre-
dicted within a finite time horizon. This study uses an Adaptive Neuro Fuzzy
predicted within a finite time horizon. This study uses an Adaptive Neuro Fuzzy Inference Inference
System (ANFIS)
System (ANFIS)to toachieve
achievethis
thisgoal.
goal.

Figure Machinelearning
13. Machine
Figure 13. learninginput/output
input/output configuration
configuration for battery
for battery temperature
temperature prediction.
prediction.

ANFIS isisaapowerful
ANFIS powerfuland andflexible
flexible tool
tool that
that cancan be used
be used to model
to model and and predict
predict complex
complex
systems
systems with
with several
several input
input and
andoutput
outputvariables
variablesand andunknown
unknownrelationships
relationshipsbetween
betweenthem.
It combines
them. the advantages
It combines of fuzzy
the advantages of logic
fuzzyandlogicneural networks
and neural to build
networks toabuild
hybrid system that
a hybrid
can handle
system that crisp and fuzzy
can handle crisp data. ANFIS
and fuzzy was
data. born in
ANFIS was 1993
bornbased
in 1993on based
the first-order Takagi–
on the first-
Sugeno fuzzy model fuzzy
order Takagi–Sugeno proposed
modelbyproposed
Jang [40].by After
Jangseveral years
[40]. After of development,
several ANFIS is
years of develop-
ment,aANFIS
now is nowartificial
well-known a well-known artificial
intelligence intelligence
technique usedtechnique
in a wide used in aofwide
range range of [41].
applications
applications [41]. Generally, ANFIS uses either backpropagation
Generally, ANFIS uses either backpropagation or a combination of least-squares or a combination of estimation
least-
squares estimation and backpropagation for membership function
and backpropagation for membership function parameter estimation [42]. Fuzzy logic parameter estimation
[42].
is Fuzzy logic is atechnique
a mathematical mathematicalthattechnique
deals with that deals with and
uncertainty uncertainty and imprecision
imprecision in data. Using
linguistic variables and fuzzy sets allows for approximate reasoning andreasoning
in data. Using linguistic variables and fuzzy sets allows for approximate inference. andOn the
inference. On the other hand, the neural network is a machine learning
other hand, the neural network is a machine learning algorithm that can learn complex algorithm that can
learn complex
patterns patterns andin relationships
and relationships the data. ANFIS in the data. ANFIS
combines these twocombines these by
techniques twousing
fuzzy if–then rules to represent the input–output relationships and neural networks to
learn the parameters of the rules. For example:
• Rule 1: if x is A1 and y is B1, then f1 = p1 x + q1 y + r1.
• Rule 2: if x is A2 and y is B2, then f2 = p2 x + q2 y + r2.
Each rule consists of a set of antecedent and consequent parts, where the antecedent
part represents the input variables x and y. The consequent part f represents the output
variable. The antecedent part is usually a fuzzy set, and the consequent part is a linear or
nonlinear function of the input variables.
The learning algorithm ANFIS focuses on adjusting the parameters of the fuzzy if–then
rules to minimise the error between the predicted and the actual outputs. This is carried
out using a combination of gradient descent optimisation and the least-squares method.
The gradient descent algorithm adjusts the parameters of the network in the direction of
the negative gradient of the error function. At the same time, the least-squares method
finds the optimal values of the parameters that minimise the sum of the squared errors [43].
In this study, as shown in Figure 13, three main inputs are used by the ANFIS model.
The first one is the ambient temperature, a directly measurable variable. The second input
is the battery’s temperature at present. This is also directly measurable in real time. The
third input, which is the most complicated one, estimates the heat generated inside the
battery within a predefined historical window. Both the electrical and thermal models,
discussed in the previous section, are used to estimate the heat generation in real time
 gradient of the error function. At the same time, the least-squares method finds the optimal
values of the parameters that minimise the sum of the squared errors [43].
In this study, as shown in Figure 13, three main inputs are used by the ANFIS model.
The first one is the ambient temperature, a directly measurable variable. The second input
Batteries 2024, 10, 85 is the battery’s temperature at present. This is also directly measurable in real time. The
14 of 22
third input, which is the most complicated one, estimates the heat generated inside the
battery within a predefined historical window. Both the electrical and thermal models,
discussed in the previous section, are used to estimate the heat generation in real time
according
according to tothe
thestate
stateofofthethebattery
battery asas well
well as as
thethe history
history of charge/discharge.
of charge/discharge. HavingHaving
all all
the inputs, the ANFIS model is trained to learn how to predict
the inputs, the ANFIS model is trained to learn how to predict the temperature change the temperature change
within the prediction
within the prediction horizon.horizon.
Forthe
For thetraining
training of of
thethe ANFIS
ANFIS model,
model, it is crucial
it is crucial to separate
to separate the datatheintodata intoand
training training
and
testing sets to evaluate the algorithm’s performance on new unseen data. Figure 14 shows 14
testing sets to evaluate the algorithm’s performance on new unseen data. Figure
shows an example
an example of data into
of data division division into the following
the following three groups:three groups:
training, training,
checking, andchecking,
testing. and
testing. Several
Several tests weretests were performed
performed at differentatambient
different ambient temperatures
temperatures and then
and then a training a training
dataset
dataset
was builtwas
usingbuilt using
all of them.all of them.
Around 50%Around 50% data
of the whole of the whole
were useddata were used
for training and for training
check-
and
ing tochecking
design and to design
prepareand prepare
the ANFIS the ANFIS
model. model.
As a result, theAs a result,
ANFIS model the ANFIS
could havemodel
accesscould
have access
to a wide rangeto of
a wide range
test data of testatdata
recorded recorded
various ambientattemperatures.
various ambient temperatures.
For testing, we used For
the remaining
testing, we used datathe
case by case asdata
remaining presented
case byin case
the following sections.
as presented in the following sections.

Figure 14.
Figure 14.An
Anexample
exampledivision
divisionofof
thethe
data into
data thethe
into three groups
three of training,
groups checking,
of training, and testing.
checking, and testing.

Figure15
Figure 15demonstrates
demonstrates the
the training
training and and testing
testing datadata in one
in one place place to highlight
to highlight their their
similarities and
similarities anddifferences.
differences.AllAll datasets
datasetsareare
independent
independent andandentirely different
entirely from from
different each each
other, ensuring
ensuringthatthatthe
themodel
modelis is
unbiased
unbiased towards
towards thethe
training data.data.
training During the training
During the training
phase, ANFIS
ANFIS learns
learnsthethepatterns
patternsand
andrelationships
relationshipsinin the
thetraining
trainingdata
datato to
make
makethethepre-
predic-
dictions
tions later.
later. However,
However, it is
it is essential
essential totoevaluate
evaluatethethemodel’s
model’sperformance
performance on new new unseen
unseen data
data
to to determine
determine its generalisation
its generalisation abilityunder
ability underanyanyunique
unique situation.
situation. Performing
Performing rigor-
rigorous
ous testing
testing
Batteries 2024, 10, x FOR PEER REVIEW andand evaluation
evaluation on on
thisthis diverse
diverse dataset
dataset cancan ensure
ensure thethe accuracyand
accuracy andreliability
reliability
15 of 22of the
of the ANFIS
ANFIS modelmodel for predicting
for predicting the battery
the battery temperature,
temperature, as discussed
as discussed in in Section
Section 4.4.

Figure15.
Figure Investigation
15.Investigation ofof
thethe training
training andand testing
testing datadata
usedused for ANFIS
for ANFIS training
training and validation.
and validation.

3.4. Model
3.4. ModelPerformance
Performance Evaluation,
Evaluation, Test
Test Reliability,
Reliability, andand Consistency
Consistency Assessment
Assessment
Thebattery
The batterytemperature
temperatureprediction
prediction topic
topic does
does notnot
havehave
anyany standard
standard for for accuracy.
accuracy. It It is
vital
is toto
vital demonstrate
demonstratea aproper
properway wayofofprocessing
processingtesttest data
data and
and propose
propose a
a performance
performance eval-
uation method
evaluation methodforfor
thethe
proposed
proposed battery
batterytemperature
temperatureprediction
prediction system. First,the
system. First, thesys-
system’s
performance
tem’s is evaluated
performance is evaluatedusing thethe
using testtest
data collected
data collectedatatvarious
various ambient temperatures
ambient tempera-
ranging
tures ranging 10 ◦ C
fromfrom ◦ C for a fresh cell under the WLTP test profile.
10to°C30to 30 °C for a fresh cell under the WLTP test profile.
Standardstatistical
Standard statistical measures
measures areare used
used to evaluate
to evaluate the accuracy
the accuracy and reliability
and reliability of theof the
proposedbattery
proposed battery temperature
temperature prediction
prediction system, system,
such assuch
meanas mean error
absolute absolute
(MAE), error (MAE),
maxi-
mum error, and root-mean-square error (RMSE). The MAE provides an average absolute dif-
ference between the predicted and the actual temperature values, which can be used to deter-
mine how well the model performs overall. The maximum error provides the highest absolute
difference between the predicted and the actual temperatures, which can help to identify any
potential outliers or errors in the predictions. In addition, RMSE provides a more comprehen-
Batteries 2024, 10, 85 15 of 22

maximum error, and root-mean-square error (RMSE). The MAE provides an average
absolute difference between the predicted and the actual temperature values, which can
be used to determine how well the model performs overall. The maximum error provides
the highest absolute difference between the predicted and the actual temperatures, which
can help to identify any potential outliers or errors in the predictions. In addition, RMSE
provides a more comprehensive picture of the model’s accuracy by taking into account
the squared differences between the predicted and the actual temperature values, which
can help identify any biases in the predictions. The MAE and RMSE formulations are
presented below:
1 m  i
m i∑
MAE = h x − yi (4)
=1
s
1 m
2
∑ h x (i ) − y (i )


RMSE( X, h) = (5)
m i =1
By evaluating the performance of the ANFIS model using these statistical methods, it is
possible to identify any areas where the model is underperforming and make improvements
to increase its accuracy and reliability. This can involve adjusting the parameters of the
fuzzy if–then rules, collecting more data to improve the training of the model, or using
different input variables to capture the underlying patterns in the data better.
Test Uncertainty Analysis (TUA) is a crucial process used to estimate and evaluate
any uncertainty’s potential sources and magnitudes in a measurement or testing process.
In this research, the accuracy of the test data is essential, as the data will be used in the
machine learning model for training and testing, which highly depends on the quality of
the data. TUA identifies and qualifies the sources of uncertainty that can arise and affect
the accuracy of the measurements.
In battery testing, potential sources of uncertainty can be attributed to various fac-
tors, including the accuracy of the measurement devices, environmental conditions like
changes in the ambient temperature, sample variability due to the noise, and inevitable
operator errors. Once the uncertainty sources have been identified, the potential impact
on measured data can be estimated and analysed using statistical methods. TUA aims to
provide a comprehensive understanding of the uncertainties associated with the battery
test measurement process and provide a robust characteristic for the battery temperature
prediction model in this research.
This analysis makes every effort to control the sources of uncertainty. The test opera-
tion procedure is refined, and the test rig is calibrated to ensure the accuracy of the mea-
surements. An analysis of all the uncertainty sources and their solutions is provided below:
• Data measurement accuracy: the power supply is calibrated before the test to ensure
it provides accurate power to the battery. Multi-thermocouples are attached to the
battery surface to avoid thermocouple failure or inconsistencies in quality. To minimise
computational effort, the average temperature data value is taken from all thermo-
couples, with a temperature difference threshold of ∆ ≤ 0.2 ◦ C between them. This
approach helped to reduce computational complexity while still providing accurate
temperature data. Furthermore, all data acquisition equipment is synchronised at a
1 Hz sampling rate, ensuring that the data collection is consistent and precise.
• Noise reduction: the battery and the rack are kept in the thermal chamber with a metal
shell to reduce electromagnetic interference. To minimise the impact of vibration from
the thermal chamber during testing, fire-resistant foam is added under the battery
rack to absorb any potential vibration. Maintaining a consistent ambient temperature
during testing is essential for obtaining accurate and reliable data. Keeping the battery
in the thermal chamber allows it to control the ambient temperature and minimise the
impact of external factors on the battery’s performance.
• Improved environmental conditions: while the thermal chamber can precisely control
the ambient temperature with a quick response, the battery temperature may not shift
Batteries 2024, 10, 85 16 of 22

as efficiently as expected. To ensure accurate temperature measurements, the battery is

left in the thermal chamber for at least two hours before each test to allow the battery
temperature to reach the same level as the ambient temperature. One factor that could
not be controlled is the humidity level inside the chamber. However, the thermal
chamber is equipped with a ventilation system to maintain relatively dry conditions
inside the chamber.
Test repeatability is essential to ensure consistency and reproducibility in the test
results obtained from multiple measurements or trials of the same item or system under
the same test condition. In this research, test repeatability is crucial as the test data are
applied for a machine learning-based model to avoid the model being trained explicitly for
a single cell.
To ensure test repeatability, two identical NCA cells are tested under the same test
conditions, including the same current profile, ambient temperature, and test equipment.
Figure 16 compares the voltage responses between two identical cells under the same test
profile. Figure 17 shows a similar comparison of the two cells’ temperatures under the same
test. According to the results, both figures demonstrate significant overlaps between the
curves. Statistical methods are used for repeatability evaluation, and the comparison results
are presented in Table 3 as a percentage of values. By analysing the repeatability of the test
results, it is possible to ensure the consistency and reliability of the data used in the machine
learning-based models. The test repeatability analysis is not only essential to ensure the
consistency and reproducibility of the test results but also to prove the robustness of the
hybrid battery temperature prediction model. If the prediction result accuracy remains
high level despite various battery characteristics, it indicates that the model is robust and
reliable in real-world applications. As battery cell consistency is a frequent issue in battery
Batteries 2024, 10, x FOR PEER REVIEW 17 of 22
pack applications, it is crucial to have a robust temperature prediction model
Batteries 2024, 10, x FOR PEER REVIEW 17 ofthat
22 can
handle various battery characteristics.

Figure 16. Voltage comparison between two identical NCA cells under the same test.
Figure Voltagecomparison
16.Voltage
Figure 16. comparisonbetween
between two
two identical
identical NCA
NCA cellscells
underunder the same
the same test. test.

Figure 17. Temperature comparison between two identical NCA cells under the same test.
Figure 17.
Figure 17.Temperature
Temperaturecomparison
comparisonbetween two
between identical
two NCA
identical cellscells
NCA under the same
under test. test.
the same
Table 3. Statistical evaluation of the voltage and temperature signals recorded from two identical
Table 3. Statistical evaluation of the voltage and temperature signals recorded from two identical
NCA cells under the same test.
NCA cells under the same test.
Cell Group Current Profile RMSE of Voltage RMSE of Temperature
Cell Group Current Profile RMSE of Voltage RMSE of Temperature
Cell 1 and Cell 2 WLTP 0.0599 0.2651
Cell 1 and Cell 2 WLTP 0.0599 0.2651
4. Battery Temperature Prediction Results
4. Battery Temperature Prediction Results
To evaluate the proposed battery temperature prediction system, experimental tests
To evaluate the proposed battery temperature prediction system, experimental tests
were conducted at different ambient temperatures. Each test consisted of several cycles
were conducted at different ambient temperatures. Each test consisted of several cycles
Batteries 2024, 10, 85 17 of 22

Table 3. Statistical evaluation of the voltage and temperature signals recorded from two identical
NCA cells under the same test.

Cell Group Current Profile RMSE of Voltage RMSE of Temperature

Cell 1 and Cell 2 WLTP 0.0599 0.2651

4. Battery Temperature Prediction Results

To evaluate the proposed battery temperature prediction system, experimental tests
were conducted at different ambient temperatures. Each test consisted of several cycles
from fully charged to fully discharged states. The WLTP current profile was utilised during
the discharge stage to mimic the battery’s real-world usage. To better see the results
obtained from the tests, only one cycle is shown in each figure. It should be noted that each
discharge cycle might include several repeats of the WLTP profile until the battery is fully
depleted. The following sections present and discuss the prediction results for both cases
of the NCA cells and the LFP cells separately.
An important point here is that the ANFIS model is not retrained each time when
switching between different ambient temperatures. The model is trained once for each cell
chemistry (i.e., NCA and LFP), and then it is validated under different scenarios.

4.1. NCA Battery Temperature Prediction Results

The first set of validation results comprises a comparison between the NCA battery’s
predicted temperature and the reference temperature 30 s ahead at an ambient temperature
of 10 ◦ C. As shown in Figure 18, there is a high degree of overlap between the predicted
and the reference temperatures. Furthermore, Table 4 presents statistical measures obtained
in this case. The MAE, which measures the average magnitude of the errors, confirms the
overall accuracy of the prediction model, while the max error and RMSE provide insights
into the degree of deviation between the predicted and actual temperatures. Notably, the
Batteries 2024, 10, x FOR PEER REVIEW 18 of 22
RMSE indicates that the temperature prediction closely follows the real battery surface
change with no significant outliers.

15.5
Reference temperature
15 Temperature prediction
14.5

14

13.5

13

12.5

12

11.5

11

10.5
0 0.5 1 1.5 2 2.5 3 3.5 4
Time(h)

Figure
Figure18.
18.The
TheNCA
NCAbattery temperature
battery prediction
temperature validation
prediction 30 s ahead
validation during during
30 s ahead the WLTP
thetest at test
WLTP
10 °C.◦
at 10 C.
Table 4. Statistical evaluation of the NCA battery temperature prediction errors 30 s ahead during
Table
the 4. Statistical
WLTP evaluation
test at various of the NCA battery temperature prediction errors 30 s ahead during
temperatures.
the WLTP test at various temperatures.
Test Scenario MAE Maximum Error RMSE
WLTP at 10 °C
Test Scenario 0.0369
MAE 0.7708 Error
Maximum 0.0655
RMSE
WLTP at 20 °C 0.0374 0.5729 0.0862
WLTP at 10 ◦ C 0.0369 0.7708 0.0655
WLTP at 30 °C
◦ 0.0488 0.6390 0.0922
WLTP at 20 C 0.0374 0.5729 0.0862
WLTP at 30 ◦ C 0.0488 0.6390 0.0922
The second validation case contains the same results at 20 °C. As depicted in Figure
19, the predicted and the reference temperatures exhibit a high degree of overlap, indicat-
The
ing the second of
accuracy validation case contains
the prediction the same
model. The results
statistical at 20 ◦ C.presented
measures, As depicted in Figure
in Table 4, 19,
the predicted
further and
reinforce thisthe reference
finding undertemperatures exhibit a high
this ambient temperature degree of overlap, indicating
too.

25.5
Reference temperature
25
Temperature prediction
24.5

24

23.5

23
 the WLTP test at various temperatures.
Test Scenario MAE Maximum Error RMSE
Test Scenario
WLTP at 10 °C MAE
0.0369 Maximum
0.7708Error RMSE
0.0655
WLTP
WLTP at at 10
20 °C
°C 0.0369
0.0374 0.7708
0.5729 0.0655
0.0862
WLTP
WLTP at at 20
30 °C
°C 0.0374
0.0488 0.5729
0.6390 0.0862
0.0922
Batteries 2024, 10, 85 WLTP at 30 °C 0.0488 0.6390 0.0922 18 of 22
The second validation case contains the same results at 20 °C. As depicted in Figure
The
19, the second validation
predicted case contains
and the reference the same
temperatures results
exhibit at 20 degree
a high °C. As of
depicted inindicat-
overlap, Figure
19,
ing the
thepredicted
accuracy and
of the
the reference
prediction temperatures
model. The exhibit a
statistical high degree
measures, of overlap,
presented in indicat-
Table
the accuracy of the prediction model. The statistical measures, presented in Table 4, further 4,
ing the
further accuracy
reinforce of the
this prediction
finding undermodel.
this The
ambientstatistical measures,
temperature
reinforce this finding under this ambient temperature too. too. presented in Table 4,
further reinforce this finding under this ambient temperature too.
25.5
Reference temperature
25
25.5 Temperature prediction
Reference temperature
24.5
25
Temperature prediction
24
24.5

23.5
24

23
23.5

22.5
23

22
22.5

21.5
22

21
21.5

20.5
21
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
20.5 Time(h)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Time(h)
Figure 19.
Figure 19.The
TheNCA
NCAbattery temperature
battery prediction
temperature validation
prediction 30 s ahead30
validation during the WLTP
s ahead testthe
during at 20 °C. test
WLTP
Figure
at 20 ◦ 19.
C. The NCA battery temperature prediction validation 30 s ahead during the WLTP test at 20 °C.
The third validation case similarly shows the effectiveness of the proposed model at
The
30 °C.
The third
Thethirdvalidation
results case similarly
are demonstrated
validation case shows
in Figure
similarly theand
20
shows effectiveness
the Table 4. of the
effectiveness ofproposed model
the proposed at
model at
30 °C.
◦ The results are demonstrated in Figure 20 and Table
30 C. The results are demonstrated in Figure 20 and Table 4. 4.
34.5
Reference temperature
34.5
34
Temperature prediction
Reference temperature
34
33.5 Temperature prediction

33.5
33

33
32.5

32.5
32

32
31.5

31.5
31

31
30.5

30.5
30
0 0.5 1 1.5 2 2.5 3 3.5 4
30 Time(h)
0 0.5 1 1.5 2 2.5 3 3.5 4
Time(h)
Figure 20. The NCA battery temperature prediction validation 30 s ahead during the WLTP test at 30 °C.
Figure
Figure20.
20.The NCA
The battery
NCA temperature
battery prediction
temperature validation
prediction 30 s ahead30
validation during the WLTP
s ahead testthe
during at 30 °C. test
WLTP
at 30 ◦ C.

The prediction results obtained from the three above-mentioned validation cases
demonstrate that the proposed battery surface temperature prediction system is highly
effective for a fresh NCA cell under various ambient temperatures. All the results were
related to a 30 s prediction horizon. Although the 30 s time window is sufficient for the EV
application safety concerns because this study is readily applicable to other applications
such as aerospace, further validation case studies were conducted to challenge the proposed
prediction model with longer prediction horizons. Table 5 includes the prediction errors for
the cases of 60 s and 90 s prediction horizons. Looking at the results, although the errors
are slightly higher than in the case of the 30 s prediction horizon, they are still deficient in
terms of system reliability. The MAE is around 0.12 ◦ C even at 90 s ahead.

Table 5. Statistical evaluation of the NCA battery temperature prediction errors 60 and 90 s ahead
during the WLTP test 20 ◦ C.

Test Scenario MAE Maximum Error RMSE

20 ◦ C,
60 s ahead 0.0765 1.1115 0.1203
20 ◦ C, 90 s ahead 0.1275 1.1812 0.1954

4.2. LFP Battery Temperature Prediction Results

To check if the proposed battery temperature system works for any battery, the same
analysis was performed here on the LFP cells introduced in Section 3.1. It should be noted
that the ANFIS model is trained again for the LFP cell; however, it was performed only
 Test Scenario MAE Maximum Error RMSE
20 °C, 60 s ahead 0.0765 1.1115 0.1203
20 °C, 90 s ahead 0.1275 1.1812 0.1954

4.2. LFP Battery Temperature Prediction Results

Batteries 2024, 10, 85 19 of 22
To check if the proposed battery temperature system works for any battery, the same
analysis was performed here on the LFP cells introduced in Section 3.1. It should be noted
that the ANFIS model is trained again for the LFP cell; however, it was performed only
once, which
once, whichmeans
meanswe wedid didnot
notneed
needtotoretrain
retrainit itwhen
when switching
switching between
between different
different ambient
ambi-
temperatures or various duty cycles,
ent temperatures or various duty cycles, etc. etc.
Thefirst
firstvalidation
validationtest test ◦ C, where the goal was
The ononthethe
LFPLFP cells
cells waswas performed
performed at 0 at
°C,0 where the goal was
to predict
to predictthethecell’s
cell’ssurface
surfacetemperature
temperature 30 30 s ahead.
s ahead. According
According to thetoresults
the results presented
presented in in
Figure 21, there is a good agreement between the predicted temperature
Figure 21, there is a good agreement between the predicted temperature and the reference and the reference
temperatureininthis
temperature this case.
case. SomeSome
pointspoints slightly
slightly deviate deviate from
from the the reference
reference valuethe
value around around
the highest
highest temperature
temperature area and area
theand the outcome
outcome is verytosimilar
is very similar to theinstances
the previous previousofinstances
the
of thecell
NCA NCA cell temperature
temperature prediction.
prediction. Upon conducting
Upon conducting statistical
statistical analysis, analysis,
presented presented
in Table
in itTable
6, 6, it was
was found thatfound
both MAEthat both MAE measures
and RMSE and RMSE measures
show showlevel
an excellent an excellent level of
of accuracy
on averageon
accuracy whereas
average thewhereas
maximum theerror reaches error
maximum 2.47 °C 2.47 ◦Therefore,
in this case.
reaches these Therefore,
C in this case. find-
ings
theseprove that prove
findings the proposed
that thetemperature prediction method
proposed temperature is generally
prediction methodreliable even for
is generally reliable
other types
even for of cells
other typesif itofhas already
cells been
if it has trainedbeen
already withtrained
the relevant
with data.
the relevant data.

10
Reference temperature
9
Temperature prediction

1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Time(h)

Figure
Figure 21.
21.The
TheLFP
LFPcell’s
cell’stemperature prediction
temperature 30 s30ahead
prediction during
s ahead the WLTP
during test at
the WLTP 0 °C.
test at 0 ◦ C.

Table 6. Statistical evaluation of the LFP cell’s temperature prediction errors 30 s ahead during the
Table 6. Statistical evaluation of the LFP cell’s temperature prediction errors 30 s ahead during the
WLTP test at various temperatures.
WLTP test at various temperatures.
Test Scenario Average Error Maximum Error RMSE
Test
WLTP Scenario
at 0 °C Average
0.1327Error Maximum
2.4744 Error RMSE
0.2947
WLTP
WLTP at 10 0 ◦C
°C 0.0686
0.1327 0.6954
2.4744 0.1137
0.2947
WLTP
WLTP at 20◦°C
at 10 C 0.0686
0.0559 0.6954
0.6266 0.1137
0.0904
Batteries 2024, 10, x FOR PEER REVIEW WLTP at 20 ◦ C 0.0559 0.6266 20 of
0.0904 22

The second and third validation cases of the LFP cell’s surface temperature prediction
underThethesecond
WLTPandtestthird
werevalidation
performedcases
atof
10the
◦ CLFP
andcell’s surface
20 ◦ C, temperature
respectively. prediction
Figures 22 and 23
under the WLTP test were performed at 10 °C and 20 °C, respectively. Figures
show the predicted temperature versus the reference temperature values in the 22 and 23 show
two cases,
the predicted
which again temperature
demonstrates versus the reference
an excellent temperature
match betweenvalues in the
the two two cases,
curves. which to
According
again demonstrates an excellent match between the two curves. According to Table 6, the max-
Table 6, the maximum error is less than 1 ◦ C in both cases, which is much better than in the
imum error is less than 1 °C in both cases, which is much better than in the previous case.
previous case. These comparative results are also similar to what was already observed for
These comparative results are also similar to what was already observed for the NCA cells.
the NCA cells.
16
Reference temperature
15.5
Temperature prediction
15

14.5

14

13.5

13

12.5

12

11.5

11
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time(h)

Figure 22.
Figure 22. The
TheLFP
LFPcell’s
cell’stemperature
temperatureprediction 30 30
prediction s ahead during
s ahead the the
during WLTP test at
WLTP 10at
test °C.
10 ◦ C.

25

Reference temperature
24.5 Temperature prediction

24

23.5

23

22.5
 13

12.5

12

11.5

11
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time(h)
Batteries 2024, 10, 85 20 of 22
Figure 22. The LFP cell’s temperature prediction 30 s ahead during the WLTP test at 10 °C.

25

Reference temperature
24.5 Temperature prediction

24

23.5

23

22.5

22

21.5

21
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time(h)

Figure
Figure 23.
23.The
TheLFP
LFPcell’s
cell’stemperature prediction
temperature 30 s30ahead
prediction during
s ahead the WLTP
during test at
the WLTP °C.20 ◦ C.
20 at
test

Conclusions
5. Conclusions
AAnew
newbattery
battery temperature
temperature prediction
prediction methodology
methodology was developed
was developed by combining
by combining an
online
an onlinebattery modelmodel
battery identification technique
identification with an ANFIS
technique with an model.
ANFIS Temperature predic-
model. Temperature
tion results results
prediction were compared against against
were compared experimental data obtained
experimental from the NCA
data obtained from and LFP and
the NCA
cells. The validation
LFP cells. The validation tests were repeated
tests were at different
repeated ambient
at different temperatures
ambient to ensure
temperatures the the
to ensure
system
system worked
workedunder undervarious
various conditions.
conditions.In addition,
In addition,the the
prediction horizon
prediction was changed
horizon was changed
from
from 30 30ssto
to90 90sstotomeasure
measurethe thesensitivity
sensitivityof the method
of the method to that factor.
to that As isAs
factor. theiscase
the for
case for
all types of predictors, a longer prediction horizon leads to an increase
all types of predictors, a longer prediction horizon leads to an increase in the prediction in the prediction
errors;
errors; however,
however,according
accordingtoto the results,
the thethe
results, prediction
predictionaccuracy remained
accuracy remainedrelatively low low
relatively
(i.e., MAE = 0.12 °C)
◦ even at 90 s ahead. Another observation for
(i.e., MAE = 0.12 C) even at 90 s ahead. Another observation for both the NCA and LFPboth the NCA and LFP
cells
cells was
wasthat
thatthetheprediction
prediction errors
errorswere
werehigher
higherat lower ambient
at lower temperatures.
ambient temperatures. This is a is a
This
good result because the risk of high battery temperatures usually becomes more critical
good result because the risk of high battery temperatures usually becomes more critical
when the ambient temperature is also high. According to the results, the proposed method
when the ambient temperature is also high. According to the results, the proposed method
showed its best performance at the highest temperatures.
showed its best performance at the highest temperatures.
Overall, the proposed battery temperature prediction system has proven to be a
Overall, the proposed battery temperature prediction system has proven to be a highly
highly practical and versatile tool. It not only performs exceptionally well for a particular
practical and versatile tool. It not only performs exceptionally well for a particular cell but
cell but also maintains a consistently high level of accuracy across other battery types. This
also maintains
is a crucial feature,a consistently
as batteries comehighinlevel of accuracy
various shapes and across
sizes,other battery
and their types.
thermal This is a
behav-
crucial feature, as batteries come in various shapes and sizes,
iour can differ significantly depending on their composition and environmental condi-and their thermal behaviour
can differ
tions. significantly
Therefore, depending
the proposed systemʹson their composition
potential to significantly and environmental
improve conditions.
the performance
Therefore, the proposed system’s potential to significantly improve the performance and
safety of the batteries was demonstrated, making it an up-and-coming tool for future
research and development in the field of energy storage and advanced battery management
systems. Accurate temperature prediction is essential for battery safety and effective battery
operation in various industries, such as electric vehicles, renewable energy storage, and
consumer electronics. The proposed methodology, which includes a combination of battery
model identification and an ANFIS model, provides a robust and efficient way of predicting
battery temperature in real-world applications. Although the effectiveness of the proposed
system was demonstrated for an automotive application, the method is expected to work
in other applications, too. The findings of this study have potential implications in all
industries that rely on battery thermal management systems. By accurately predicting
battery temperature, the proposed system could enhance the safety and functionality of a
battery system, ensuring that it operates within its safe temperature range to reduce the
risk of thermal runaway.

Author Contributions: H.Z. and A.F.; Methodology, A.F.; Software, H.Z.; Validation, H.Z.; Formal
analysis, H.Z. and A.F.; Investigation, H.Z.; Resources, M.L.; Data curation, M.L.; Writing—original
draft, H.Z.; Writing—review & editing, A.F. and D.J.A.; Visualization, H.Z.; Supervision, A.F.;
Funding acquisition, A.F. and D.J.A. All authors have read and agreed to the published version
of the manuscript.
Funding: This work was funded by the UKRI Faraday Battery Challenge project called Next Genera-
tion LFP Cathode Material (NEXLFP). In addition, Abbas Fotouhi acknowledges funding from the
Faraday Institution (Industrial Fellowships FIIF-003 and FIIF-014).
Batteries 2024, 10, 85 21 of 22

Data Availability Statement: The data presented in this study will be available on request from the
corresponding author.
Conflicts of Interest: Author Matt Lowe was employed by the company Delta Cosworth. The
remaining authors declare that the research was conducted in the absence of any commercial or
financial relationships that could be construed as a potential conflict of interest.

References
1. Ma, S.; Jiang, M.; Tao, P.; Song, C.; Wu, J.; Wang, J.; Deng, T.; Shang, W. Temperature effect and thermal impact in lithium-ion
batteries: A review. Prog. Nat. Sci. Mater. Int. 2018, 28, 653–666. [CrossRef]
2. Lu, Z.; Yu, X.L.; Wei, L.C.; Cao, F.; Zhang, L.Y.; Meng, X.Z.; Jin, L.W. A comprehensive experimental study on temperature-
dependent performance of lithium-ion battery. Appl. Therm. Eng. 2019, 158, 113800. [CrossRef]
3. Yi, F.; Jiaqiang, E.; Zhang, B.; Zuo, H.; Wei, K.; Chen, J.; Zhu, H.; Zhu, H.; Deng, Y. Effects analysis on heat dissipation
characteristics of lithium-ion battery thermal management system under the synergism of phase change material and liquid
cooling method. Renew. Energy 2022, 181, 472–489. [CrossRef]
4. Xu, C.; Feng, X.; Huang, W.; Duan, Y.; Chen, T.; Gao, S.; Lu, L.; Jiang, F.; Ouyang, M. Internal temperature detection of thermal
runaway in lithium-ion cells tested by extended-volume accelerating rate calorimetry. J. Energy Storage 2020, 31, 101670. [CrossRef]
5. Zhang, X.; Li, Z.; Luo, L.; Fan, Y.; Du, Z. A review on thermal management of lithium-ion batteries for electric vehicles. Energy
2022, 238, 121652. [CrossRef]
6. Pesaran, A.A.; Vlahinos, A.; Burch, S.D. Thermal Performance of EV and HEV Battery Modules and Packs; National Renewable Energy
Laboratory: Golden, CO, USA, 1997.
7. Motloch, C.G.; Christophersen, J.P.; Belt, J.R.; Wright, R.B.; Hunt, G.L.; Sutula, R.A.; Duong, T.; Tartamella, T.J.; Haskins, H.J.;
Miller, T.J. High-Power Battery Testing Procedures and Analytical Methodologies for HEV’s; SAE Technical Paper; SAE International:
Warrendale, PA, USA, 2002.
8. Zhao, R.; Zhang, S.; Liu, J.; Gu, J. A review of thermal performance improving methods of lithium ion battery: Electrode
modification and thermal management system. J. Power Sources 2015, 299, 557–577. [CrossRef]
9. Bandhauer, T.M.; Garimella, S.; Fuller, T.F. A Critical Review of Thermal Issues in Lithium-Ion Batteries. J. Electrochem. Soc. 2011,
158, R1. [CrossRef]
10. Wu, W.; Ma, R.; Liu, J.; Liu, M.; Wang, W.; Wang, Q. Impact of low temperature and charge profile on the aging of lithium-ion
battery: Non-invasive and post-mortem analysis. Int. J. Heat Mass Transf. 2021, 170, 121024. [CrossRef]
11. Na, Y.; Sun, X.; Fan, A.; Cai, S.; Zheng, C. Methods for enhancing the capacity of electrode materials in low-temperature
lithium-ion batteries. Chin. Chem. Lett. 2021, 32, 973–982. [CrossRef]
12. Jaguemont, J.; Boulon, L.; Dubé, Y. A comprehensive review of lithium-ion batteries used in hybrid and electric vehicles at cold
temperatures. Appl. Energy 2016, 164, 99–114. [CrossRef]
13. Zhang, Z.; Yu, W.; Li, H.; Wan, W.; Zhang, W.; Zhuo, W.; Liu, Q. Heat transfer characteristics and low-temperature performance of
a lithium-ion battery with an inner cooling/heating structure. Appl. Therm. Eng. 2023, 219, 119352. [CrossRef]
14. Shahid, S.; Chaab, M.A. A review of thermal runaway prevention and mitigation strategies for lithium-ion batteries. Energy
Convers. Manag. X 2022, 16, 100310. [CrossRef]
15. Rao, Z.; Wang, S. A review of power battery thermal energy management. Renew. Sustain. Energy Rev. 2011, 15, 4554–4571.
[CrossRef]
16. Ludwig, S.; Steinhardt, M.; Jossen, A. Determination of Internal Temperature Differences for Various Cylindrical Lithium-Ion
Batteries Using a Pulse Resistance Approach. Batteries 2022, 8, 60. [CrossRef]
17. Leising, R.A.; Palazzo, M.J.; Takeuchi, E.S.; Takeuchi, K.J. A study of the overcharge reaction of lithium-ion batteries. J. Power
Sources 2001, 97–98, 681–683. [CrossRef]
18. Feng, F.; Teng, S.; Liu, K.; Xie, J.; Xie, Y.; Liu, B.; Li, K. Co-estimation of lithium-ion battery state of charge and state of temperature
based on a hybrid electrochemical-thermal-neural-network model. J. Power Sources 2020, 455, 227935. [CrossRef]
19. Li, M.; Dong, C.; Li, X.; Dong, X.; Mu, Y.; Jia, H. A Sequential Network-model Alliance Module for Lithium-ion Battery
Temperature Prediction. In Proceedings of the 2022 IEEE Energy Conversion Congress and Exposition (ECCE), Detroit, MI, USA,
9–13 October 2022.
20. Billert, A.M.; Frey, M.; Gauterin, F. A Method of Developing Quantile Convolutional Neural Networks for Electric Vehicle Battery
Temperature Prediction Trained on Cross-Domain Data. IEEE Open J. Intell. Transp. Syst. 2022, 3, 411–425. [CrossRef]
21. Park, S.; Ahn, C. Stochastic Model Predictive Controller for Battery Thermal Management of Electric Vehicles. In Proceedings of
the 2019 IEEE Vehicle Power and Propulsion Conference (VPPC), Hanoi, Vietnam, 14–17 October 2019.
22. Wang, M.; Hu, W.; Jiang, Y.; Su, F.; Fang, Z. Internal temperature prediction of ternary polymer lithium-ion battery pack based on
CNN and virtual thermal sensor technology. Int. J. Energy Res. 2021, 45, 13681–13691. [CrossRef]
23. Fang, K.; Mu, D.; Chen, S.; Wu, B.; Wu, F. A prediction model based on artificial neural network for surface temperature simulation
of nickel–metal hydride battery during charging. J. Power Sources 2012, 208, 378–382. [CrossRef]
24. Willard, J.; Jia, X.; Xu, S.; Steinbach, M.; Kumar, V. Integrating Scientific Knowledge with Machine Learning for Engineering and
Environmental Systems. ACM Comput. Surv. 2022, 55, 66. [CrossRef]
Batteries 2024, 10, 85 22 of 22

25. Wu, B.; Zhang, B.; Deng, C.; Lu, W. Physics-encoded deep learning in identifying battery parameters without direct knowledge of
ground truth. Appl. Energy 2022, 321, 119390. [CrossRef]
26. E. Board. EV Markets Reports. August 2023. Available online: https://evmarketsreports.com/lfp-batteries-transforming-ev-
market-dynamics-globally/ (accessed on 22 November 2023).
27. EUR-Lex. Worldwide Harmonised Light-Duty Vehicles Test Procedure (WLTP) and Real Driving Emissions (RDE). 1 September
2023. Available online: https://eur-lex.europa.eu/EN/legal-content/summary/worldwide-harmonised-light-duty-vehicles-
test-procedure-wltp-and-real-driving-emissions-rde.html (accessed on 1 December 2023).
28. Pavlovic, J.; Marotta, A.; Ciuffo, B. CO2 emissions and energy demands of vehicles tested under the NEDC and the new WLTP
type approval test procedures. Appl. Energy 2016, 177, 661–670. [CrossRef]
29. Johnson, V.H. Battery performance models in ADVISOR. J. Power Sources 2002, 110, 321–329. [CrossRef]
30. Fotouhi, A.; Auger, D.J.; Propp, K.; Longo, S. A Study on Battery Model Parametrisation Problem—Application-Oriented
Trade-offs between Accuracy and Simplicity. IFAC PapersOnLine 2016, 49, 48–53. [CrossRef]
31. Fotouhi, A.; Auger, D.J.; Propp, K.; Longo, S.; Wild, M. A review on electric vehicle battery modelling: From Lithium-ion toward
Lithium–Sulphur. Renew. Sustain. Energy Rev. 2016, 56, 1008–1021. [CrossRef]
32. Shateri, N.; Auger, D.J.; Fotouhi, A.; Brighton, J. An Experimental Study on Prototype Lithium–Sulfur Cells for Aging Analysis
and State-of-Health Estimation. IEEE Trans. Transp. Electrif. 2021, 7, 1324–1338. [CrossRef]
33. Fotouhi, A.; Auger, D.J.; Propp, K.; Longo, S. Accuracy Versus Simplicity in Online Battery Model Identification. IEEE Trans. Syst.
Man Cybern. Syst. 2018, 48, 195–206. [CrossRef]
34. Shateri, N.; Shi, Z.; Auger, D.J.; Fotouhi, A. Lithium-Sulfur Cell State of Charge Estimation Using a Classification Technique. IEEE
Trans. Veh. Technol. 2021, 70, 212–224. [CrossRef]
35. Shateri, N.; Auger, D.J.; Fotouhi, A.; Brighton, J.; Du, W. Investigation of the Effect of Temperature on Lithium-Sulfur Cell Cycle
Life Performance Using System Identification and X-ray Tomography. Batter. Supercaps 2022, 5, e202200035. [CrossRef]
36. Sato, N. Thermal behavior analysis of lithium-ion batteries for electric and hybrid vehicles. J. Power Sources 2001, 99, 70–77.
[CrossRef]
37. Bernardi, D.; Pawlikowski, E.; Newman, J. A General Energy Balance for Battery Systems. J. Electrochem. Soc. 1985, 132, 5.
[CrossRef]
38. Schuster, E.; Ziebert, C.; Melcher, A.; Rohde, M.; Seifert, H.J. Thermal behavior and electrochemical heat generation in a
commercial 40 Ah lithium ion pouch cell. J. Power Sources 2015, 286, 580–589. [CrossRef]
39. Bai, Y.; Li, L.; Li, Y.; Chen, G.; Zhao, H. Reversible and irreversible heat generation of NCA/Si–C pouch cell during electrochemical
energy-storage process. J. Energy Chem. 2019, 29, 95–102. [CrossRef]
40. Jang, J.S. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 1993, 23, 665–685. [CrossRef]
41. Keller, J.M.; Krishnapuram, R.; Rhee, F.H. Evidence aggregation networks for fuzzy logic inference. IEEE Trans. Neural Netw. 1992,
3, 761–769. [CrossRef] [PubMed]
42. Jang, J.S.R.; Sun, C.T.; Mizutani, E. Neuro-Fuzzy and Soft Computing-A Computational Approach to Learning and Machine
Intelligence [Book Review]. IEEE Trans. Autom. Control 1997, 42, 1482–1484. [CrossRef]
43. Şişman Yilmaz, N.A. A Temporal Neuro-Fuzzy Approach for Time-Series Analysis. Master’s Thesis, Middle East Technical
University, Ankara, Turkey, 2003.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.