Cloud and Digital-Twin Enhanced Thermal Safety Framework for E-

Mobility Battery Management Systems

by

Akash Samanta

A thesis submitted to the

School of Graduate and Postdoctoral Studies in partial
fulfillment of the requirements for the degree of

Doctor of Philosophy in Electrical and Computer Engineering

Department of Electrical, Computer, and Software Engineering

University of Ontario Institute of Technology (Ontario Tech University)

Oshawa, Ontario, Canada

April 2025

© Akash Samanta, 2025

 THESIS EXAMINATION INFORMATION

Submitted by: Akash Samanta

Doctor of Philosophy in Electrical and Computer Engineering

Thesis title: Cloud and Digital-Twin Enhanced Thermal Safety Framework for E-Mobility
Battery Management Systems

An oral defense of this thesis took place on March 12, 2025 in front of the following
examining committee:

Examining Committee:

Chair of Examining Committee Dr. Akramul Azim

Research Supervisor Dr. Sheldon Williamson

Examining Committee Member Dr. Tarlochan Sidhu

Examining Committee Member Dr. Walid Morsi Ibrahim

University Examiner Dr. Xianke Lin, Associate Professor, Department of

Automotive and Mechatronics Engineering, Ontario
Tech University, ON, Canada

External Examiner Dr. Balakumar Balasingam, Associate Professor,

University of Windsor, ON, Canada

The above committee determined that the thesis is acceptable in form and content and that
a satisfactory knowledge of the field covered by the thesis was demonstrated by the
candidate during an oral examination. A signed copy of the Certificate of Approval is
available from the School of Graduate and Postdoctoral Studies.

ii
ABSTRACT

Lithium-ion batteries (LIBs) are the preferred energy storage for electric vehicles (EVs)
due to their high energy density and efficiency. However, frequent EV fires expose critical
shortcomings in current battery management systems (BMS) and thermal controls. LIBs
exhibit nonlinear characteristics sensitive to operating and environmental conditions, but
existing BMS technologies rely solely on surface temperature measurements, neglecting
core temperatures, which can be up to 10°C higher during fast charging or high-load
conditions. This limitation contributes to overheating, accelerated degradation, and thermal
runaway. This thesis develops advanced core temperature estimation techniques, including
hybrid equivalent circuit models (ECM), long short-term memory (LSTM) networks,
bidirectional LSTM (Bi-LSTM), and Kolmogorov-Arnold network–LSTM fusion
networks. These methods achieve real-time core temperature estimation with errors as low
as 0.16°C across a wide range of ambient temperatures and charging/discharging C-rates.
A digital twin (DT)-based core temperature prediction model forecasts temperatures three
minutes ahead with a prediction error under 0.4°C. Validation includes LIB cells of varying
form factors, such as 18650 and 21700, and chemistries like nickel cobalt aluminum oxide
(NCA) and nickel manganese cobalt oxide (NMC). To enhance scalability and
functionality, the research integrates cloud computing and DT technologies for real-time
monitoring, predictive control, and efficient thermal management. Internal temperature-
informed closed-loop control demonstrated effective overheating prevention, improving
response time by two minutes. Unlike previous studies, this research validates the
framework using a 14-cell LIB module, rather than single cells. A key innovation is the
DT-based BMS, which predicts thermal behavior up to three minutes in advance, aligning
with IEC 62933-2-1:2017 standards for proactive risk mitigation. Cloud-based monitoring
and data storage support predictive maintenance and second-life applications for retired
batteries. The integration of machine learning, Internet of Things (IoT), and DT
technologies ensures adaptability to dynamic conditions. Computational cost and latency
were analyzed, with latencies ranging from 5 ms locally to 85 ms for cloud-based systems.
By addressing critical gaps in BMS capabilities and introducing predictive thermal
management strategies, this research advances EV battery safety and management.

iii
Keywords: electric vehicle; thermal management system; machine learning; artificial

intelligence; lithium-ion battery

iv
AUTHOR’S DECLARATION

I hereby declare that this thesis consists of original work of which I have authored.

This is a true copy of the thesis, including any required final revisions, as accepted by my

examiners.

I authorize the University of Ontario Institute of Technology (Ontario Tech

University) to lend this thesis to other institutions or individuals for the purpose of scholarly

research. I further authorize University of Ontario Institute of Technology (Ontario Tech

University) to reproduce this thesis by photocopying or by other means, in total or in part,

at the request of other institutions or individuals for the purpose of scholarly research. I

understand that my thesis will be made electronically available to the public.

Akash Samanta

v
STATEMENT OF CONTRIBUTIONS

Books
1. A. Samanta and S. Williamson, “Battery Technologies for E-Mobility-Innovations,
Applications, and Advanced Management Systems,” ISBN: 978-87-7004-086-0. River
Publishers, Denmark, 2024.
2. S. Williamson, A. Samanta, C. Chetri, L. Anekal, and A. Huynh, Energy management
strategies for electric and plug-in hybrid electric vehicles, In Press, Springer Nature,
Switzerland, 2024.
3. Williamson, S., Samanta, A., Anekal, L., “Principles of Electric Energy Storage
Systems for E-Mobility”. In Press. Wiley, New Jersey, 2025.

Book Chapters
1. A. Samanta, A. Huynh, L. Anekal, C. Chetri, and S. Williamson, “Advanced charging
and battery management systems for e-mobility,” in Power Electronic Converters and
Systems Applications, The Institution of Engineering and Technology, 2024.
2. A. Samanta, L. Anekal, C. Chetri, and S. Williamson, “Vehicle electrification and
energy storage systems in modern power grids,” in Vehicle Electrification in Modern
Power Grids: Disruptive Perspectives on Power Electronics Technologies and Control
Challenges, Elsevier, 2024.
3. C. Chetri, A. Samanta, L. Anekal, and S. Williamson, “Energy storage systems for
transportation electrification,” in Electric Vehicles and Distributed Generation -
Microgrids, River Publishers, Denmark, 2024.

Journal Articles
1. A. Samanta and S. Williamson, “Cloud-Integrated Adaptive Deep Learning
Framework for Real-Time Battery Core Temperature Estimation and Enhanced
Thermal Safety,” IEEE Transactions on Transportation Electrification, 2025,
Accepted.
2. A. Samanta, M. Sharma, W. Locke, and S. Williamson, “Cloud-enhanced battery
management system architecture for real-time data visualization, decision making, and
long-term storage,” IEEE Journal of Emerging and Selected Topics in Industrial
Electronics, 2025, Accepted.
3. D. Karnehm, A. Samanta, C. Rosenmuller, A. Neve, and S. Williamson, “Core
temperature estimation of lithium-ion batteries using long short-term memory (LSTM)
network and Kolmogorov-Arnold network (KAN),” IEEE Transactions on
Transportation Electrification, 2024, Accepted.
4. D. Karnehm, A. Samanta, L. Anekal, S. Pohlmann, A. Neve, and S. Williamson,
“Comprehensive comparative analysis of deep learning-based state-of-charge
estimation algorithms for cloud-based lithium-ion battery management systems,” IEEE
Journal of Emerging and Selected Topics in Industrial Electronics, 2024.
5. D. Karnehm, A. Samanta, M. Hohenegger, N. Tashakor, S. M. Goetz, M. Kuder, A.
Neve, and S. Williamson, “Universal data specification and real-time data streaming
architecture for cloud-based battery management systems,” IEEE Journal of Emerging
and Selected Topics in Power Electronics, 2024.

vi
6. A. Samanta and S. Chowdhuri, “Active cell balancing of lithium-ion battery pack
using dual DC-DC converter and auxiliary lead-acid battery,” Journal of Energy
Storage, vol. 33, p. 102 109, 2021.
7. S. Surya, A. Samanta, V. Marcis, and S. Williamson, “Hybrid electrical circuit model
and deep learning-based core temperature estimation of lithium-ion battery cell,” IEEE
Transactions on Transportation Electrification, vol. 8, no. 3, pp. 3816–3824, 2022.
8. A. Samanta, S. Chowdhuri, and S. S. Williamson, “Machine learning-based data-
driven fault detection/diagnosis of lithium-ion battery: A critical review,” Electronics,
vol. 10, no. 11, p. 1309, 2021.
9. A. Samanta and S. S. Williamson, “A survey of wireless battery management system:
Topology, emerging trends, and challenges,” Electronics, vol. 10, no. 18, p. 2193,
2021.
10. A. Samanta and S. S. Williamson, “A comprehensive review of lithium-ion cell
temperature estimation techniques applicable to health-conscious fast charging and
smart battery management systems,” Energies, vol. 14, no. 18, p. 5960, 2021.
11. S. Surya, A. Samanta, V. Marcis, and S. Williamson, “Smart core and surface
temperature estimation techniques for health-conscious lithium-ion battery
management systems: A model-to-model comparison,” Energies, vol. 15, no. 2, p. 623,
2022.
12. A. Samanta and S. Williamson, “Machine learning-based remaining useful life
prediction techniques for lithium-ion battery management systems: A comprehensive
review,” IEEJ Journal of Industry Applications, vol. 12, no. 4, pp. 563–574, 2023.

Conference Proceedings
1. A. Samanta and S. Williamson, “Introducing time-lag and Bi-LSTM neural network
for in operando surface temperature estimation in lithium-ion batteries,” in IEEE
Energy Conversion Congress Expo (ECCE) Asia, 2025, Accepted.
2. M. Sharma, A. Samanta, W. Locke, and S. Williamson, “Critical role of individual
cell temperature monitoring in mitigating thermal runaway and reducing accelerated
degradation in lithium-ion batteries,” in IEEE Applied Power Electronics Conference
and Exposition (APEC), IEEE, 2025.
3. A. Samanta, C. Chetri, and S. Williamson, “Crucial examination of thermal behavior
of solid-state battery for intelligent gray box model-based automotive battery
management systems,” in IECON 2024- 50th Annual Conference of the IEEE
Industrial Electronics Society (IECON), IEEE, 2024.
4. A. Huynh, A. Samanta, and S. Williamson, “Temporal sensitivity analysis for
enhanced dynamic equivalent circuit modeling of lithium-ion batteries in on-board
battery management systems,” in IECON 2024- 50th Annual Conference of the IEEE
Industrial Electronics Society (IECON), IEEE, 2024.
5. A. Samanta, C. Chetri, D. Karnehm, A. Neve, and S. Williamson, “Temporal
sensitivity analysis of internal temperature informed charging algorithms and rapid
thermal management system for e-mobility,” in 2024 IEEE Energy Conversion
Congress and Exposition (ECCE), IEEE, 2024.
6. M. Sharma, A. Samanta, C. Chetri, and S. Williamson, “Real-time can data acquisition
and visualization: Synerging physical-to-virtual (p2v) twinning of automotive battery

vii
 management systems,” in IECON 2024- 50th Annual Conference of the IEEE
Industrial Electronics Society (IECON), IEEE, 2024.
7. D. Karnehm, L. Anekal, A. Samanta, A. Neve, and S. Williamson, “Towards digital
twining of lithium-ion battery management systems: An extended Kalman filter for
state-of charge estimation in cloud-platform,” in 2023 IEEE Energy Conversion
Congress and Exposition (ECCE), IEEE, 2023, pp. 6524–6525.
8. D. Karnehm, A. Samanta, A. Neve, and S. Williamson, “Five-layer IoT and fog
computing framework towards digital twinning of battery management systems for e-
transportation,” in 2024 4th International Conference on Smart Grid and Renewable
Energy (SGRE), IEEE, 2024, pp. 1–7.
9. A. Samanta, A. Huynh, N. Shrestha, and S. Williamson, “Combined data driven and
online impedance measurement-based lithium-ion battery state of health estimation for
electric vehicle battery management systems,” in 2023 IEEE Applied Power
Electronics Conference and Exposition (APEC), IEEE, 2023, pp. 862–866.
10. A. Samanta, D. Karnehm, A. Neve, and S. Williamson, “Data-informed healthy cell
clustering technique for second-life applications of retired electric vehicle batteries,”
in 2024 IEEE Transportation Electrification Conference and Expo (ITEC), IEEE,
2024, pp. 1–5.
11. A. Samanta, D. Karnehm, A. Neve, and S. Williamson, “Voltage relaxation pattern
recognition for efficient sorting of healthy cells for second-life applications of retired
electric vehicle batteries,” in 2024 IEEE International Conference on Industrial
Technology (ICIT), IEEE, 2024, pp. 1–6.
12. A. Samanta, A. Huynh, E. Rutovic, and S. Williamson, “Rapid thermal modeling and
discharge characterization for accurate lithium-ion battery core temperature
estimation,” in IECON 2022–48th Annual Conference of the IEEE Industrial
Electronics Society, IEEE, 2022, pp. 1–6.
13. A. Samanta, A. Huynh, M. Sharma, V. Marcis, and S. Williamson, “Supercapacitor
and bidirectional dc-dc converter-based active charge balancing scheme for lithium-
ion batteries,” in 2022 IEEE Energy Conversion Congress and Exposition (ECCE),
IEEE, 2022, pp. 1–7.
14. C. Chetri, A. Samanta, and S. Williamson, “Critical understanding of temperature
gradient during fast charging of lithium-ion batteries at low temperatures,” in IECON
2023-49th Annual Conference of the IEEE Industrial Electronics Society, IEEE, 2023,
pp. 1–6.
15. L. Anekal, A. Samanta, and S. Williamson, “Wide-ranging parameter extraction of
lithium-ion batteries to estimate state of health using electrochemical impedance
spectroscopy,” in 2022 IEEE 1st Industrial Electronics Society Annual On-Line
Conference (ONCON), IEEE, 2022, pp. 1–6.
16. N. Shrestha, A. Samanta, F. C. Fietosa, and S. Williamson, “State-of-the-art wireless
charging systems for e-bikes: Technologies and applications,” in 2023 IEEE 14th
International Conference on Power Electronics and Drive Systems (PEDS), IEEE,
2023, pp. 1–6.
17. A. Huynh, A. Samanta, C. Chetri, and S. Williamson, “Online determination of
lithium-ion battery state of health based on normalized change of state of temperature
for e-mobility applications,” in 2023 IEEE Transportation Electrification Conference
& Expo (ITEC), IEEE, 2023, pp. 1–6.

viii
18. L. Anekal, A. Samanta, and S. Williamson, “Rapid parameterization of lithium-ion
batteries using frequency window identification technique for on-board charge control
and battery management,” in 2024 IEEE Applied Power Electronics Conference and
Exposition (APEC), IEEE, 2024, pp. 1330–1337.
19. A. Huynh, A. Samanta, and S. Williamson, “Effects of regenerative braking on hybrid
battery balancing,” in 2024 IEEE Applied Power Electronics Conference and
Exposition (APEC), IEEE, 2024, pp. 815–821.
20. L. Anekal, A. Samanta, U. Deshpande, and S. Williamson, “Real-time point-to-point
parameter tracking for fault prognosis of lithium-ion batteries using electrochemical
impedance spectroscopy,” in 2023 IEEE 8th Southern Power Electronics Conference
(SPEC), IEEE, 2023, pp. 1–5.
21. A. Huynh, A. Samanta, C. Chetri, L. Anekal, and S. Williamson, “Health-conscious
dual stage hybrid lithium-ion battery balancing strategy for e-mobility,” in 2023 IEEE
Energy Conversion Congress and Exposition (ECCE), IEEE, 2023, pp. 1683–1690.
22. M. Sharma, A. Samanta, A. Huynh, and S. Williamson, “Rapid PCB development
using CO2 laser and galvo scanner for modular battery management systems,” in
IECON 2023-49th Annual Conference of the IEEE Industrial Electronics Society, IEEE,
2023, pp. 1–6.

Patents
1. A. Samanta, and S. Williamson, “Next Generation EV Battery Safety Framework using
a Digital Twin-based Smart Battery Management Systems (BMS)”, Status-Pending.

2. A. Samanta, and S. Williamson, “Fusion Framework for Real-time Core Temperature

Estimation in Lithium-ion Batteries using Bidirectional LSTM and Kolmogorov-Arnold
Networks in Cloud-based Battery Management Systems”, Status-Pending.

ix
ACKNOWLEDGEMENTS

First and foremost, I would like to express my deepest gratitude to my supervisor, Dr.

Sheldon Williamson, for his unwavering support, guidance, and encouragement throughout

my doctoral journey. Dr. Williamson’s expertise and dedication have been instrumental in

shaping my research and my personal growth as a researcher. His ability to inspire

innovative thinking and his commitment to excellence have left an indelible mark on me,

and I will always be grateful for his mentorship and belief in my abilities.

I extend my heartfelt thanks to my colleagues and friends at the Smart Transportation

Electrification and Energy Research (STEER) lab. The collaborative environment,

insightful discussions, and camaraderie we shared have made this journey both

intellectually stimulating and personally rewarding. Working alongside such talented and

passionate individuals has been an incredible privilege, and I am honored to have been a

part of this exceptional team.

To my parents and family, words cannot fully capture my appreciation for your

unconditional love, sacrifices, and steadfast belief in me. Your support has been my

foundation, and your encouragement has been my strength. To my son, Trihaan, you are

my greatest joy and my constant reminder of why I strive for excellence. To my spouse,

Ajanta, your patience, and understanding have been my pillars of support, and I am

eternally grateful for your presence in my life.

Finally, I dedicate this work to my parents, family and to all those who have inspired

me to dream big and pursue my goals with relentless determination. This thesis is a

testament to your love and encouragement, and I hope it serves as a reflection of the values

you have instilled in me. Thank you all for being part of this journey.

x
TABLE OF CONTENTS

THESIS EXAMINATION INFORMATION................................................................. ii

ABSTRACT ...................................................................................................................... iii
AUTHOR’S DECLARATION ........................................................................................ v
STATEMENT OF CONTRIBUTIONS ......................................................................... vi
ACKNOWLEDGEMENTS ............................................................................................. x
TABLE OF CONTENTS ................................................................................................ xi
LIST OF TABLES .......................................................................................................... xv
LIST OF FIGURES ....................................................................................................... xvi
LIST OF ABBREVIATIONS AND SYMBOLS .......................................................... xx
Chapter 1. Introduction ............................................................................................... 1
1.1 Motivation ................................................................................................................. 4
1.2 Research Problem ...................................................................................................... 4
1.3 Objectives .................................................................................................................. 6
1.4 Structure of the Thesis............................................................................................... 7
Chapter 2. Literature Review ..................................................................................... 8
2.1 Automotive Batteries................................................................................................. 8
2.2 Safety Consideration in Automotive Lithium-ion Batteries ..................................... 9
2.3 Lithium-ion Battery Pack for Automotive Applications ......................................... 12
2.4 Existing BMS Architectures .................................................................................... 14
2.5 Typical Functionality of Automotive Battery Management Systems ..................... 15
2.5.1 Voltage, Current and Temperature Monitoring ................................................ 15
2.5.2 Charge and Discharge Control.......................................................................... 16
2.5.3 State Estimation ................................................................................................ 17
2.5.4 Cell Voltage and SOC Balancing ..................................................................... 18
2.5.5 Thermal Management System Control ............................................................. 18
2.5.6 Isolation Detection ............................................................................................ 19
2.5.7 Communication and Data Logging ................................................................... 19
2.6 Classification of Automotive Battery Management System ................................... 20
2.6.1 Centralized BMS .............................................................................................. 20
2.6.2 Modular and Scalable BMS .............................................................................. 21
2.6.3 Wireless BMS ................................................................................................... 22

xi
 2.6.5. Cloud-Based Battery Management System (CWBMS) .................................. 25
2.7 Core Temperature Estimation of Lithium-ion Battery ............................................ 26
2.7.1 Classification of Core Temperature Estimation Strategies ............................... 27
2.7.2 Core Temperature Estimation Strategies .......................................................... 30
2.7.3 Limitations of Conventional Temperature Monitoring Techniques ................. 53
2.8 Real-Time Core Temperature Estimation ............................................................... 55
2.9 Cloud and Digital-twin enhanced Battery Management Systems ........................... 56
2.10 Introduction to Digital Twin Technology ............................................................. 57
2.10.1 General Architecture of Digital Twin Technologies ...................................... 59
2.10.2 State-of-the-art application of DT in BMS for E-Transportation ................... 61
2.10.3 Importance of Digital Twin-Based Thermal Management Control Strategy
Informed by Core Temperature Estimation ............................................................... 64
Chapter 3. Research Methodology ........................................................................... 67
3.1 Proposed Framework ............................................................................................... 67
3.1.1 Proposed Framework and Key Components .................................................... 68
3.2 Temperature Estimation Models ............................................................................. 70
3.2.2 Hybrid ECM and LSTM -based Core Temperature Estimation ....................... 78
3.2.2.3 Architecture of 2D-GLSTM Neural Network ................................................... 81
3.2.3 LSTM Network and Kolmogorov-Arnold Network (KAN) for Core and
Surface Temperature Estimation ............................................................................... 85
3.2.4 Bi-LSTM-based Real-time Core Temperature Estimation ............................... 89
3.3 Cloud-enhanced Battery Management System ....................................................... 93
3.3.1 Data Acquisition and Processing ...................................................................... 94
3.3.2 Data Processing and Visualization ................................................................... 95
3.5 Digital-twin enabled Battery Thermal Management Control System ..................... 97
3.6 Deep Learning-based SOC Estimation Algorithms for Cloud-based BMS ............ 99
3.6.1 Extended Kalman filter (EKF) for SOC Estimation in Cloud ........................ 100
3.6.2 Feedforward Neural Network (FNN) ............................................................. 100
3.6.3 Long Short-Term Memory (LSTM) ............................................................... 101
3.6.4 Gated Recurrent Unit (GRU) .......................................................................... 102
3.7 Experimental Setup ............................................................................................... 103
3.7.1 Experimental Setup for Core Temperature Estimation of 18650 Cell ........... 103

xii
 3.7.3 Experimental Setup for Samsung INR 21700 Cell and Module Level Testing
and Real-time Data Logging in Cloud ..................................................................... 108
3.7.4 Experimental Setup for Samsung INT 21700 Cell and Module Level Testing
and Real-time Core Temperature Estimation and Logging in Local Machine and
Cloud ....................................................................................................................... 111
Chapter 4. Results and Discussion .......................................................................... 118
4.1 Validation of Proposed Models and Core Temperature Estimation Techniques .. 118
4.1.1. Model validation 1-RC and 2-RC EETM for Core Temperature Estimation 118
4.1.2 Validation of Hybrid ECM and LSTM-based Core Temperature Estimation 127
4.1.3 Validation of LSTM Network and KAN-based Surface and Core Temperature
Estimation Techniques............................................................................................. 136
4.2.1 Real-time Data Acquisition Efficiency ........................................................... 146
4.2.3 Grafana Visualization Capabilities ................................................................. 147
4.2.4 System Integration, Data Flow, and Data Security......................................... 149
4.3 Implementation of cloud-based SOC Estimation and Comparative Analysis ...... 162
4.3.1 Description of data and model training .......................................................... 162
4.4 Implementation of Cloud-enhanced Real-time Core Temperature Estimation and
Control ......................................................................................................................... 171
4.4.1 Data Acquisition for training and validation of Bi-LSTM ............................. 171
4.4.2 Model Training and Validation ...................................................................... 173
4.4.3 Realtime Core Temperature Estimation using Pre-Trained Bi-LSTM in Local
Machine ................................................................................................................... 176
4.4.4 Realtime Core Temperature Estimation using Pre-Trained Bi-LSTM in Cloud
................................................................................................................................. 178
4.4.5 Data Storage and Visualization ...................................................................... 179
4.4.6 Real-Time Updates and Decision-Making ..................................................... 180
4.5 Integration of Digital-twin-based Battery Thermal Management Control System 182
4.6 Challenges and Limitations ................................................................................... 185
4.6.1 Cloud-enhanced BMS Architecture ................................................................ 185
4.6.2 Digital-twin-based BMS ................................................................................. 185
Chapter 5. Conclusions and Future Work ............................................................. 187
5.1. Advanced Core Temperature Estimation Techniques ............................................. 187
5.2 Cloud-Enhanced Battery Management Systems ................................................... 188
5.3 Deep Learning-Based Estimation and Thermal Management Strategies .............. 190

xiii
 5.4 Integration of Digital-twin-based Battery Thermal Management Control System 190
5.5 Cloud-Based SOC Estimation and Comparative Analysis ................................... 191
5.7 Limitations of the Study ........................................................................................ 193
5.8 Future Research Directions ................................................................................... 193
REFERENCES .............................................................................................................. 195

xiv
LIST OF TABLES

Table 2.1: Safe operating regions of most widely used LIB chemistries. .......................... 9
Table 2.2: Most popular electric vehicle lithium-ion battery capacity ............................. 14
Table 2.3: Summary of electrochemical thermal modelling-based temperature estimation
strategies ........................................................................................................................... 33
Table 2.4: Summary of EECM-based temperature estimation strategies ......................... 39
Table 2.5: Summary of numerical methods-based temperature estimation strategies ...... 42
Table 2.6: Summary of direct impedance measurement-based strategies ........................ 46
Table 2.7: Summary of ML-based temperature estimation techniques ............................ 49
Table 2.8: Temperature Estimation Models and Performance ........................................... 56
Table 3.1: Specifications of 18650 LIB Cell Under Test ............................................... 105
Table 3.2: Specification of Molicel P42A ...................................................................... 106
Table 4.1: MAE and RMSE in the Prediction of Vt and Ts Estimation .......................... 128
Table 4.2: Hyperparameters of the 2D-GLSTM Model ................................................. 130
Table 4.3: Hyperparameters of KAN and LSTM models ............................................... 138
Table 4.4: MAE, RMSE, and R2 Core and Surface Temperature per Model of drilled cells
in °C. ............................................................................................................................... 139
Table 4.5: MAE Surface and Core Temperature per Model in °C ................................. 142
Table 4.6: MAE, and RMSE Core and Surface Temperature per Model at UDDS dataset
at variations of training in °C .......................................................................................... 143
Table 4.7: Computational cost comparison of models and different hardware setups in ms
......................................................................................................................................... 145
Table 4.8: Bandwidth allocation ..................................................................................... 155
Table 4.9: Data sampling rate and memory allocation ................................................... 155
Table 4.10: Average roundtrip latency based on the physical location of the cloud server
......................................................................................................................................... 158
Table 4.11: Architecture of the models ........................................................................... 163
Table 4.12: Machine learning Hyperparameters ............................................................. 163
Table 4.13: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of FNN,
LSTM, GRU, and EKF of 0.75 h (discharge of SOC 100% − 91.81%) UDDS driving
cycle at 0°C and a measurement frequency of 10 Hz. .................................................... 166
Table 4.14: Approximate distance between the cloud deployment locations and the client,
located in Oshawa ........................................................................................................... 168
Table 4.15: MAE of estimated SOC by FNN, and EKF compared to the capacity-based
SOC at different loss rates. ............................................................................................. 169
Table 4.16: Model Hyperparameters .............................................................................. 174

xv
LIST OF FIGURES

Figure 1.1: Single cell failure to thermal runaway and propagation ................................... 2
Figure 1.2: Surface and core temperature of 18650 LIB at 4C discharge with 25°C
ambient temperature ............................................................................................................ 3
Figure 2.1: Operating temperature region and effect in LIB ............................................ 11
Figure 2.2: EV battery: Cell, Module, Pack (a) Graphical representation (b) Actual Tesla
Model-S Battery Pack ....................................................................................................... 13
Figure 2.3: Functionalities of automotive battery management systems .......................... 15
Figure 2.4: Architecture of a centralized BMS ................................................................. 21
Figure 2.5: Modular- Master- Slave- Daisy Chain Configuration .................................... 21
Figure 2.6: Modular- Master- Slave- Multi Drop Configuration ...................................... 22
Figure 2.7: Schematic layout of a generic WBMS ........................................................... 23
Figure 2.8: Classification of WBMS topologies. .............................................................. 25
Figure 2.9: Schematic layout of a generic temperature estimation strategy for a LIB cell.
........................................................................................................................................... 27
Figure 2.10: Family of (A) Heat generation model, (B) Heat transfer model, (C)
Temperature estimation strategy ....................................................................................... 29
Figure 2.11: Thermal model of a LIB cell (a) First-order model, (b) Second-order model
........................................................................................................................................... 35
Figure 2.12: Fundamental steps in direct impedance measurement-based temperature
estimation .......................................................................................................................... 44
Figure 2.13: Schematic layout of ML-based temperature estimation scheme .................. 48
Figure 2.14: Structure of Physics Informed Neural Network for Cell Temperature
Estimation ......................................................................................................................... 51
Figure 2.15: Three different kinds of digital representation and the corresponding data
flow ................................................................................................................................... 58
Figure 2.16: Basic functionalities of digital twin of a physical system ............................ 59
Figure 2.17:Six-layer digital twin architecture ................................................................. 60
Fig. 2.18: A cloud of things framework ............................................................................ 62
Figure 2.19: Cyber-physical elements of a battery digital twin ........................................ 63
Figure 2.20: Key technologies for DT-based BMS .......................................................... 64
Figure 2.21: Surface and core temperature of 18650 LIB at 4C discharge with 25°C
ambient temperature .......................................................................................................... 65
Figure 2.22: Single-cell failure to thermal runaway and propagation .............................. 66
Figure 3.1: Schematic layout of a digital twin-based BMS for EVs................................. 67
Figure 3.2: Proposed framework of the cloud and digital twin-enhanced core temperature
estimation with feedback control ...................................................................................... 68
Figure 3.3: 1-RC ECM (Thevenin’s equivalent) model of a LIB cell .............................. 71
Figure 3.4: First-order heat resistor-capacitor-based thermal model of LIB .................... 73
Figure 3.5: Second-order equivalent circuit thermal model of LIB .................................. 75
Figure 3.6: Fundamental building blocks of KF-based estimation scheme. ..................... 77
Figure 3.7: Hybrid ECM+2D-GLSTM-based core temperature estimation technique .... 78

xvi
Figure 3.8: (a) 2-RC ECM of a LIB cell, (b) Second-order EETM of LIB cell ............... 79
Figure 3.9: Architecture of Single LSTM. ........................................................................ 82
Figure 3.10: Architecture of 2D-GLSTM. ........................................................................ 84
Figure 3.11: LSTM Architecture and cell structure .......................................................... 86
Figure 3.12: Structure for estimating surface and core temperature using interconnected
LSTM network .................................................................................................................. 87
Figure 3.13: Structure of a Kolmogorov-Arnold Network (KAN) .................................. 88
Figure 3.14: Schematic layout of the proposed framework. ............................................. 90
Figure 3.15: Architecture of Bi-LSTM for core temperature estimation. ......................... 92
Figure 3.16: System level architecture of the proposed cloud-based BMS architecture.94
Figure 3.17: Standard CAN data frame with 11 bits identifier (CAN 2.0A). ................... 95
Figure 3.18: Pseudo code of Python script for CAN data decoding. .............................. 96
Figure 3.19: A framework of a digital twin based BMS................................................... 97
Figure 3.20: Schematic layout of the considered extended Kalman Filter for SOC
Estimation ....................................................................................................................... 100
Figure 3.21: FNN architecture ........................................................................................ 101
Figure 3.22: LSTM architecture...................................................................................... 102
Figure 3.23: Structure of a GRU cell .............................................................................. 103
Figure 3.24: Schematic layout of 18650 cell testing setup ............................................. 104
Figure 3.25: Actual experimental setup for LG 18650 cell testing ................................. 105
Figure 3.26: Installation of core temperature sensors ..................................................... 107
Figure 3.27: Experimental setup: battery cycler, cells and data acquisition system ....... 107
Figure 3.28: Photograph of the 14-cell module. ............................................................. 108
Figure 3.29: Experimental setup for 14-cell module testing and data logging. .............. 109
Figure 3.30: Schematic layout the cloud-enhanced BMS. .............................................. 109
Figure 3.31: Internal Architecture of NXP MCU, MC33771 BMS IC and isolated
communication channels. ................................................................................................ 111
Figure 3.32: Experimental setup with data visualization interface. .............................. 112
Figure 3.33: Communication and control scheme. ......................................................... 114
Figure 3.34: Cylinder volumes used for the estimation of energy densities for 18650 and
21700 cells [268]. ............................................................................................................ 115
Figure 3.35: Evaluation of CT cross-sections of 21700 cell. (e) 2D cross-sections of CT
data at mid-height of the cylinders, (d) line profiles along the dashed lines in (e), f)
digitalization of spiral in (e) [267]. ................................................................................. 115
Figure 3.36: Comparison of Nyquist plot of a sample Cell before and after installing core
temperature sensor .......................................................................................................... 116
Figure 3.37: Installation of core temperature sensors and module assembling ............. 117
Figure 4.1: The pattern of the discharging current applied to the cell. ........................... 118
Figure 4.2: The plot of Tc, Ts and Tamb without using KF. ............................................. 119
Figure 4.3: The pattern of the discharging current applied to the cell. ........................... 120
Figure 4.4: The plot of the estimated Tc and measured Ts. ............................................. 120
Figure 4.5: Variation of the difference between the estimated Tc and measured Ts. ...... 120
Figure 4.6: The pattern of the discharging current applied to the cell. ........................... 121

xvii
Figure 4.7: The plot of the estimated Tc and measured Ts. ............................................. 121
Figure 4.8: Variation of the difference between the estimated Tc and measured Ts. ...... 122
Figure 4.9: The pattern of the discharging current applied to the cell. ........................... 123
Figure 4.10: The plot of the estimated Tc and measured Ts. ........................................... 123
Figure 4.11: Variation of the difference between the estimated Tc and measured Ts. .... 123
Figure 4.12: The pattern of the discharging current applied to both the models. ........... 124
Figure 4.13: Difference between Tc and Ts obtained from the second-order thermal model.
......................................................................................................................................... 125
Figure 4.14: Difference between Tc and Ts obtained from the first-order thermal model.
......................................................................................................................................... 125
Figure 4.15: Comparison between Tc - Ts for higher C discharge. ................................. 127
Figure 4.16: Plot of (a) Actual and predicted Vt with prediction error (b) Measured and
estimated Ts. .................................................................................................................... 128
Figure 4.17: Training dataset (a) HWFET drive cycle I, (b) Tc of cell at different Tamb (c)
Vt of cell. ......................................................................................................................... 130
Figure 4.18: Training dataset (a) US6 drive cycle I, (b) Tc of cell at different Tamb (c) Vt
of cell. ............................................................................................................................. 130
Figure 4.19: Actual and Estimated Tc in UDDS Cycle (a) 273K, (b) 293K, (c) 323K .. 131
Figure 4.20: Performance comparison of core temperature estimation methods at 293K in
UDDS (a) Estimated Tc by different method (b) Estimation error. ............................... 132
Figure 4.21: The plot of (a) current profile (b) Tc under HWFET. ............................... 134
Figure 4.22: The plot of (a) current profile (b) Tc under UDDS..................................... 135
Figure 4.23: The plot of (a) current profile (b) Tc under LA92. ..................................... 135
Figure 4.24. The plot of (a) current profile (b) Tc under UDDS+LA92. ........................ 135
Figure 4.25: Data processing steps. ................................................................................ 137
Figure 4.26: Measured voltage, current and temperature data at 20°C ambient
temperature. .................................................................................................................... 137
Figure 4.27: Measured surface temperature (green) compared with KAN (brown) and
LSTM (brown) estimation models at 10°C ambient temperature ................................... 140
Figure 4.28: Measured core temperature (green) compared with KAN (brown) and LSTM
(red) estimation models at 10°C ambient temperature. .................................................. 140
Figure 4.29: Validation of surface temperature estimation for the validation cell over the
wide range of ambient temperatures. .............................................................................. 141
Figure 4.30: Change of discharge capacity indicating battery degradation .................... 142
Figure 4.31: Measured core and surface temperature (green) compared with KAN
(brown) and LSTM (red) estimation models at UDDS 25°C. ........................................ 144
Figure 4.32: Individual cell voltages of 14 cells during CC-CV charging and CC
discharging at 1C rate with 10 Hz sampling rate. ........................................................... 148
Figure 4.33: Individual cell temperature of all 14 cells during CC-CV charging at 1C rate
with 10 Hz sampling rate. ............................................................................................... 148
Figure 4.34: Incremental pack voltage level with increasing numbers of cells in series.149
Figure 4.35: Data communication protocol ensuring security and privacy. ................... 151

xviii
Figure 4.36: Individual cell balancing profile of the 14-cell battery module with 10 Hz
sampling rate. .................................................................................................................. 153
Figure 4.37: Round trip data transfer latency from Oshawa, Canada. ............................ 158
Figure 4.38: Training loss of FNN, LSTM, and GRU models. ...................................... 164
Figure 4.39: Request time depending on the model between Oshawa (ON) and Montreal
(QC) (a) Oshawa (ON) and (b) Frankfurt (DE). ............................................................. 167
Figure 4.40: Request time measurement using Long-Term Evolution (LTE) as network.
......................................................................................................................................... 170
Figure 4.41: SOC of EKF, and FNN compared to the capacity-based SOC. ................. 171
Figure 4.42: Measured temperatures during 4C discharge at 25°C. ............................... 172
Figure 4.43: Measured temperatures during 2C charging and discharge at 40°C. ......... 173
Figure 4.44: Measured temperatures during 2C charging and discharge at -20°C. ........ 173
Figure 4.45: Training and validation loss. ...................................................................... 174
Figure 4.46: Measured and estimated core temperature of Cell#1 during 1C charge and
discharge at varied ambient temperature. ....................................................................... 176
Figure 4.47: Measured and estimated core temperature of Cell#1 during UDDS at 25°C
ambient temperature. ....................................................................................................... 176
Figure 4.48: Visualization of real-time core temperature estimation in Grafana. .......... 178
Figure 4.49: Real-time decision making to prevent overheating and thermal runaway
during charging. .............................................................................................................. 180
Figure 4.50: Real-time decision making to prevent overheating and thermal runaway
during discharging. ......................................................................................................... 181
Figure 4.51: Real-time decision making showing faster reaction time........................... 182
Figure 4.52: Measured and predicted core temperature (3 minutes ahead of time) by the
digital twin during 1C-charging at 25°C. ........................................................................ 183
Figure 4.53: Measured and predicted core temperature (3 minutes ahead of time) by the
digital twin during 1C-discharging at 25°C. ................................................................... 184

xix
LIST OF ABBREVIATIONS AND SYMBOLS

AC Alternating Current
ANN Artificial Neural Networks
AFE Analog Front-End
BMS Battery Management System
BESS Battery Energy Storage Systems
BPNN Backpropagation Neural Network
CAN Controlled Area Network
CAN FD Controller Area Network Flexible Data-Rate
CWBMS Cloud-Based Battery Management Systems
DC Direct Current
CCCV Constant Current Constant Voltage
COTA Configuration Over-The-Air
DEKF Dual Ensemble Kalman Filter
DUKF Dual Unscented Kalman Filter
DKF Dual Kalman Filter
DEKF Dual Ensemble Kalman Filter
DC Direct Current
DT Digital Twin
DOD Depth of Discharge
DOTA Data Over-The-Air
EV Electric Vehicles
EIS Electrochemical Impedance Spectroscopy
EECM Equivalent Electric Circuit Model
ECM Equivalent Circuit Model
EUKF Extended Unscented Kalman Filter
ESO Extended State Observer
ETNN Electrochemical Thermal Neural Network

xx
EMI Electromagnetic Interference
FEM Finite Element Method
FVM Finite Volume Method
FFNN Feed Forward Neural Network
FEA Finite Element Analysis
FOTA Firmware Over-The-Air
GRU Gated Recurrent Unit
GLSTM Gated Long Short-Term Memory
IoT Internet-of-Things
ITD Impedance-Temperature Detection
IaaS Infrastructure-as-a-Service
JKF Joint Kalman Filter
KAN Kolmogorov-Arnold Network
KF Kalman Filter
LIB Lithium-ion Battery
LPV Linear Parameter Varying
LSTM Long Short-Term Memory
LFP Lithium-iron Phosphate
MNPT Magnetic Nanoparticles Thermometer
MNP Magnetic nanoparticles
MLP Multi-Layer Perceptron
MAE Mean Absolute Error
MSE Mean Squared Error
NN Neural Network
NFC Near Field Communication
OTA Over-The-Air
PDE Partial Differential Equation
PINN Physics-Informed Neural Network

xxi
PCB Printed Circuit Board
PaaS Platform-as-a-Service
RUL Remaining Useful Life
ROM Reduced-Order Model
RLS Recursive Least Squares
RBFNN Radial Basis Function Neural Networks
RNN Recurrent Neural Networks
SOH State of Health
SOC State of Charge
SOP State of Power
SOT State of Temperature
SEI Solid Electrolyte Interphase
SVM Support Vector Machine
SOE State of Energy
SOTA Software Over-The-Air
TMS Thermal Management System
UKF Unscented Kalman Filter
VTS Virtual Thermal Sensor
V2G Vehicle-To-Grid
WBMS Wireless battery management systems

xxii
Chapter 1. Introduction

Transportation electrification holds immense potential in addressing global energy and

environmental challenges. Electric vehicles (EVs) offer a sustainable alternative to internal
combustion engine vehicles by reducing greenhouse gas emissions and reliance on fossil
fuels. According to the Canada Energy Regulator, zero-emission vehicles accounted for
11% of all new motor vehicle registrations in Canada in 2023, marking a 49% increase
from 2022 [1]. This growth highlights the potential of EVs in achieving emission reduction
goals and advancing sustainable mobility. Energy storage systems are an integral
component of e-transportation as it largely determines their performance and commercial
viability. Lithium-ion batteries (LIBs) are widely accepted energy storage system for EVs
due to its high energy density, long cycle life, and superior performance. However, as the
demand for EVs grows, so does the need to ensure the safety, reliability, and efficiency of
their battery systems.

The widespread adoption of EVs hinges significantly on advancements in battery life,

safety, and thermal management system (TMS) control [2]. Effective TMS ensures optimal
battery performance across diverse climates, reduces degradation, and prevents
overheating. These benefits contribute to longer battery life, enhanced safety, and lower
ownership costs, making EVs more reliable, cost-efficient, and appealing for broader
market adoption. Experimental studies and user experiences reveal that LIB performance,
safety, and longevity are highly influenced by operating temperatures [3], [4] . Deviations
from the manufacturer-specified safe operating range result in accelerated degradation [5],
reduced performance, and severe safety risks such as thermal runaway [6], [7]. These
thermal challenges are particularly pronounced in EV battery packs, which comprise
hundreds or thousands of closely packed cells, leading to thermal imbalances exacerbated
by inhomogeneous heating or cooling during fast charging, discharging, and dynamic
operating conditions [8]. Now, as the cells, and modules are very closely packed within a
battery pack, thus any kind of failure in a single cell propagates very fast and leads to
cascaded failure and eventually catastrophic failure as depicted in Figure 1.1.

1
 Figure 1.1: Single cell failure to thermal runaway and propagation

Significant temperature gradients often exist between the surface and core of LIB cells,
with differences of up to 10°C in cylindrical cells, especially under high charging or
discharging rates [9] as shown in Figure 1.2. Monitoring the core temperature of individual
cells is crucial for effective TMS, ensuring longer battery life and safer operation.
However, existing battery management systems (BMS) are limited by their reliance on
surface temperature measurements, which fail to capture critical thermal dynamics within
the battery. Furthermore, the nonlinear behavior of LIBs, influenced by temperature, state
of charge (SOC), and state of health (SOH), complicates accurate monitoring. The lack of
adaptive models and comprehensive monitoring leads to inefficiencies in capacity
utilization, cell balancing, and predictive maintenance. Studies indicate that 32% and 37%
of thermal runaway and overheating incidents occur during fast charging and dynamic
driving conditions, respectively [10]. During high-rate charging, accelerated internal
reactions cause significant heat accumulation, particularly when the cooling system is
insufficient, leading to rapid temperature rise and potential thermal runaway [11]. At this
stage, exothermic reactions such as electrolyte decomposition and instability of the anode
and cathode materials further elevate the temperature [12]. Additionally, the degradation
of the solid electrolyte interphase (SEI) layer at high temperatures exposes the anode to
direct contact with the electrolyte, triggering irreversible reactions and accelerating thermal
runaway [13]. These coupled thermodynamic and chemical processes continuously
escalate the temperature, increasing the risk of catastrophic events such as fires or
explosions.

2
 Figure 1.2: Surface and core temperature of 18650 LIB at 4C discharge with 25°C ambient temperature

Despite the critical need for monitoring, physical sensors for core temperature
measurement are impractical due to installation challenges and cost constraints. Current
approaches rely on sparse sensor deployment, as seen in the Chevy Volt and Ford C-Max
Hybrid, which employ 16 and 10 sensors for 288 and 76 cells, respectively [14]. Such
limited observability hinders precise thermal management. Moreover, in cylindrical
batteries with larger diameters (e.g., 18650 and 26650 cells), significant thermal gradients
further complicate core temperature estimation, with Biot numbers Bi (Equation 1) greater
than 0.1 indicating substantial disparities between surface and core temperatures [15], [16].

𝑘
𝐵𝑖 = ℎ𝐿 (1)

Here, h is the heat transfer coefficient, L is the characteristic length, and k is the thermal
conductivity.

The limitations of traditional BMS underscore the need for innovative solutions.
Emerging technologies, such as artificial intelligence (AI), machine learning (ML), and
digital twins, hold promise for addressing these challenges. Data-driven approaches enable
adaptive cell modeling and state estimation, while IoT and cloud computing facilitate real-
time data processing and remote monitoring. These technologies provide the foundation

3
for intelligent and scalable BMS capable of predictive control and advanced thermal
management.

The primary aim of this thesis is to develop an enhanced safety framework for
automotive batteries by leveraging advanced AI, ML, Internet of Things (IoT), and digital
twin (DT) technologies. Central to this research is the development of core temperature
estimation techniques, including hybrid equivalent circuit models (ECM), long short-term
memory (LSTM) networks, bidirectional LSTM (Bi-LSTM), and a novel Kolmogorov-
Arnold network–LSTM fusion to enable real-time temperature estimation and prediction.
These advancements aim to enhance TMS control by utilizing real-time core temperature
data to implement adaptive and predictive strategies that align with global safety standards.
Furthermore, the research addresses scalability and cost efficiency by reducing wire
harness complexity through the integration of cloud based BMS. Robust digital-twin-based
predictive models are also developed to mitigate safety risks associated with fast charging
and dynamic operating conditions. The subsequent chapters present a comprehensive
review of the relevant literature, detailed methodologies, experimental setups, and key
results, offering a pathway toward safer, more reliable, and cost-efficient battery
management solutions.

1.1 Motivation
Recent incidents of EV fires and battery failures underscore the limitations of current
BMS architectures. Studies reveal that during fast charging or high-power discharge, the
internal temperature of a battery cell can exceed its surface temperature by up to 10°C. This
core-to-surface temperature gradient, if unmonitored, can lead to severe consequences,
including thermal runaway. Current BMS technologies primarily rely on surface
temperature measurements and sparse sensor networks, which fail to capture the critical
thermal and electrochemical dynamics occurring within the battery.

1.2 Research Problem

Despite advancements in thermal modeling and state estimation techniques, several

critical gaps persist in existing BMS technologies:

4
 • Core Temperature Monitoring: Lithium-ion battery (LIB) cells are highly
sensitive to operating temperature. Studies show that the core temperature is often
7–10°C higher than the surface temperature, especially under high current loads
and rapid charge/discharge cycles. Such temperature gradients, if unmonitored, can
lead to accelerated aging, reduced performance, and potential thermal runaway.
However, current BMS solutions do not provide accurate real-time core
temperature estimates due to the impracticality and cost of installing physical
sensors inside each cell.

• Dynamic Adaptability: Existing thermal models lack the robustness to account for
varying ambient conditions, fluctuating charging and discharging rates, and battery
aging effects. This lack of adaptability undermines their effectiveness in real-world
dynamic operating environments.

• Data Integration and Scalability: Conventional BMS systems face significant

constraints in data processing and storage, which limit their ability to handle large-
scale operational data required for predictive insights. Advanced state estimation
strategies, particularly those driven by machine learning, require substantial
computational resources for real-time processing and historical data storage.
However, the onboard computational power of traditional BMS systems is
insufficient to meet these demands. The integration of cloud computing and DT-
based architectures offers a promising solution to overcome these limitations.

• Real-time Estimation and Predictive Capabilities: Current BMS technologies

lack real-time core temperature estimation and predictive mechanisms to anticipate
and mitigate risks such as thermal runaway or optimize battery utilization
proactively. This reactive approach falls short in ensuring the safety and efficiency
of LIB systems, particularly under fast charging and dynamic driving conditions.

These limitations underscore the urgent need for an intelligent, adaptive, and predictive
BMS framework. Such a framework should incorporate real-time core temperature
monitoring, dynamic adaptability, and predictive capabilities to ensure the safe, efficient,
and long-term operation of LIBs in electric vehicles and other applications.
5
1.3 Objectives

This thesis aims to address these challenges by developing an advanced BMS framework
that leverages hybrid modeling, machine learning, and digital twin (DT) technologies. The
specific objectives are as follows:

1. Develop Accurate Core Temperature Estimation Techniques:

o Design hybrid equivalent circuit models (ECM) and deep learning
architectures (e.g., LSTM, Bi-LSTM, and Kolmogorov-Arnold networks)
to estimate core temperatures in real-time with minimal real-time data.

o Validate these techniques under diverse operating conditions, including

varying ambient temperatures and C-rates.

2. Integrate Predictive Control Mechanisms:

o Implement a real-time core temperature estimation technique and DT based
BMS to predict thermal behavior up to three minutes in advance, ensuring
preemptive action against overheating and thermal runaway.

o Align predictive strategies with global safety standards such as IEC 62933-
2-1:2017.

3. Enhance Data Processing and Scalability:

o Utilize cloud computing and IoT frameworks to enable real-time
monitoring, large-scale data storage, and remote accessibility.

o Assess the impact of computational cost and latency on system

performance, considering cloud servers at varying geographical locations.

4. Facilitate Long-Term Battery Insights:

o Develop mechanisms for long-term data storage, predictive maintenance
and second-life applications of retired EV batteries, ensuring sustainable
lifecycle management.

6
1.4 Structure of the Thesis

The thesis is structured as follows:

• Chapter 2 reviews the state-of-the-art in BMS technologies, core temperature

estimation techniques, and digital twin frameworks.
• Chapter 3 outlines the proposed methodologies, including hybrid ECM-LSTM,
LSTM-KAN, Bi-LSTM models and the architecture of the digital twin-based BMS.
• Chapter 4 presents experimental setups, datasets, and the validation of proposed
models under various test conditions. Discusses the integration of local computing,
cloud computing and digital-twin architecture for real-time BMS operations, along
with an analysis of latency and computational efficiency.
• Chapter 5 concludes with key findings, contributions, and recommendations for
future research.

7
Chapter 2. Literature Review

2.1 Automotive Batteries

The automotive industry has embraced lithium-ion batteries (LIBs) as the preferred
energy storage solution due to their superior performance metrics compared to traditional
battery technologies like Nickel-Metal Hydride (NiMH) and lead-acid batteries [17]. LIBs
offer a high specific capacity of up to 3,860 mAh/g, lightweight construction, extended
lifespan, and rapid charging capabilities, making them ideal for electric vehicles (EVs) and
other demanding applications [18], [19], [20]. Unlike lithium-metal batteries (LMBs),
which use lithium metal anodes but suffer from challenges like dendrite growth and low
coulombic efficiency, LIBs employ graphite anodes and lithium-ion salts as electrolytes.
This configuration reduces plating on the anode surface and promotes the formation of a
stable solid electrolyte interface (SEI) layer, which is critical for safety and performance
[21]. LIBs typically achieve specific energies of around 200 Wh/kg and specific power of
300 W/kg, though these values vary depending on the chemistry used. Their versatility is
further demonstrated by their wide operating voltage (2.5–4.2 V) and temperature range (-
30°C to 55°C), which makes them adaptable to diverse automotive and energy storage
needs.

LIB chemistries are tailored to optimize specific performance characteristics, such as

energy density, safety, and cost. Lithium iron phosphate (LFP) batteries are prized for their
high safety, thermal stability, and long cycle life, making them suitable for EVs despite
their moderate energy density (110 Wh/kg) [22]. Nickel-manganese-cobalt oxide (NMC)
batteries balance energy density and stability, with formulations like NMC 111 prioritizing
stability and NMC 622 delivering higher capacity [23]. Lithium titanate oxide (LTO)
batteries, with their exceptional cycle life (up to 7,000 cycles) and wide temperature range
(-50°C to 65°C), excel in fast-charging applications and high-power environments [24]. On
the higher-performance end, lithium nickel cobalt aluminum oxide (NCA) batteries provide
outstanding energy density, supporting extended driving ranges in premium EVs [25].
Emerging technologies like solid-state batteries (SSBs) aim to further revolutionize the
industry by replacing liquid electrolytes with solid materials, enhancing safety, thermal

8
stability, and energy density [26]. Despite current challenges like cost and scalability, SSBs
hold immense potential for future applications. As advancements in materials science and
BMS continue, LIBs remain at the forefront of sustainable automotive technology, driving
the transition toward cleaner, more efficient mobility solutions. Studies presented in this
thesis is limited to NMC and NCA chemistry of two different form factors including 21700
and 18650 cells of three popular LIB manufacturers namely, Samsung SDI, LG Chem, and
Molicel.

2.2 Safety Consideration in Automotive Lithium-ion Batteries

While advancements in LIB technology have enhanced energy density and

performance, their unsafe operation can lead to catastrophic outcomes such as fires,
explosions, or thermal runaway [27]. Due to the nonlinear and aging dependent electrical
and thermal characteristics of LIBs, their use in automotive applications need robust
thermal safety measures for safety and optimum utilization. Factors such as temperature,
humidity, vibration, and shock significantly impact LIB degradation and safety [28]. In
automotive scenarios, the risks are amplified during normal operations or crash events,
where damaged battery cells can experience rapid temperature increases. Effective
temperature monitoring and management are therefore critical to maintaining battery safety
and functionality, especially as LIBs operate within a defined safe region of voltage,
current, and temperature [29]. Different LIB chemistries as discussed in section 2.1 exhibit
distinct thermal behaviors, making precise thermal management essential for optimizing
performance and preventing hazardous conditions. The safe operating regions of LIB
chemistries are summarized in Table 2.1.

Table 2.1: Safe operating regions of most widely used LIB chemistries.
NMC LFP NCA LTO LCO SSB
Nominal 3.7 3.3 3.7 2.4 3.7 3
Voltage (V)
Operating 3.0 - 4.2 2.5 - 3.65 2.5 - 4.2 1.8 - 2.85 3.0 - 4.2 2.5 - 4.2
Voltage
Range (V)
Specific 140 -180 80 -130 200 - 300 150 - 240 50 - 80 260 - 400
Energy
(Wh/kg)
Specific 500-3,000 1,400- 500 - 3,000-5,100 200 to 400 374 -500
Power 2,400 1,500
9
 (W/kg)
Charge (C- 0.7C-1C 1C 0.7C, 1C typical; 0.7–1C 1C
rate) typical charges to 5C
4.20V, fast maximum,
charging charges to
possible 2.85 V
with some
cells
Discharge (C- 1C to 2C 1C 1C 10C 1C
rate) typical, possible, 1C
25C on 30C 5s
some pulse
cells
Cycle Life 1000-2000 5000- 500 3,000– 500–1000 3000-
Cycles 7000; up 7,000
to 12000
possible
in some
cells
Safe -10 to 45 -20 to 50 -20 to 50 -30 to 55 -10 to 45 -30 to 150
Operating
Temperature
(°C)
Thermal 210 ◦C 270°C 150°C 200°C + 150°C
Runaway (410◦F) (518°F) (302°F) (302°F)
Safety Prone to Lowest Prone to Low- Dangerous Thermally
thermal rate of thermal temperature chemistry stable battery
runaway, thermal runaway, operation Limited
flame, and runaway flame, and and charge Lifespan
explosion and self- explosion
ignition
Environmental Cobalt and Use safer Heavy Heavy Rare More
impact Nickel have material metal metal material: environmentally
a negative extraction extraction Cobalt and friendly
impact on and and improper compared to
the improper improper disposal other Li-ion
environment disposal disposal batteries
Cost 151-177 118 -152 156 -172 240–220 135-150 80-90
(USD/kWh)

Thermal safety becomes even more crucial when considering global safety standards
for LIBs. Standards such as IEC 62660, ISO 12405, and SAE J2929 outline stringent
requirements for battery performance under varying thermal and mechanical conditions
[29]. These include battery testing under wide range of operating temperatures, overcharge,
short circuits, and thermal endurance, ensuring LIBs can withstand real-world scenarios.
Standards like UN/DOT 38.3 further highlight the need for advanced thermal monitoring
to prevent safety hazards during transportation. Adherence to these regulations is not only
necessary for compliance but also for ensuring consumer confidence in the safety of LIB-
10
powered vehicles. Advanced temperature monitoring systems and a deeper understanding
of thermal dynamics are indispensable for meeting these standards and supporting the
widespread adoption of LIBs in automotive applications. By prioritizing thermal safety,
manufacturers can unlock the full potential of LIB technology while safeguarding users
and the environment. Different temperature ranges and the effect on LIB is summarized in
Figure 2.1.

Figure 2.1: Operating temperature region and effect in LIB

Typically, the performance of LIBs generally improves with increasing temperature

within their safe operating range. Higher temperatures enhance electrochemical reactions
and accelerate ion mobility within the electrolyte and electrode materials, enabling the
battery to deliver significant power. However, higher temperatures also intensify side
reactions, leading to rapid capacity loss and potential safety hazards due to the
decomposition of the solid electrolyte interface (SEI) and electrode materials. Conversely,
at low temperatures, LIBs suffer from reduced rate capability and lithium plating on the

11
anode, which affects performance but typically does not pose immediate safety risks.
Overcharging, particularly beyond the upper voltage limit, can cause the cathode material
to deteriorate and the electrolyte to oxidize, significantly compromising safety. Similarly,
over-discharging, beyond the lower voltage limit, can result in SEI breakdown, copper
collector oxidation, and further cell damage. These scenarios highlight the critical
importance of maintaining LIBs within their prescribed operational boundaries to ensure
both performance and safety.

Worth noting that the safety profile of LIB varies significantly with their SOC. Fully
charged batteries pose a higher risk due to the elevated stored energy and increased
likelihood of thermal runaway under abusive conditions, such as overheating or mechanical
damage. In contrast, fully discharged batteries, particularly those that fall below critical
voltage thresholds (~2.0 V per cell), may suffer from copper dissolution and subsequent
dendrite formation, leading to internal short circuits upon recharge. Additionally,
prolonged deep discharge can degrade the SEI, increasing cell impedance and gas
generation. Both conditions necessitate careful SOC monitoring and protective battery
management strategies to ensure safe operation and longevity along with the temperature
of each cell.

2.3 Lithium-ion Battery Pack for Automotive Applications

In automotive applications, a single LIB cell cannot provide the voltage and power
required to drive an EV. To meet these demands, hundreds or even thousands of individual
cells are interconnected in series and parallel configurations to form a battery pack capable
of powering an EV powertrain. This modular design allows manufacturers to scale the
battery pack to suit the voltage and current requirements of various vehicle types, from
compact EVs to heavy-duty electric trucks. An EV battery pack typically consists of three
hierarchical components: cells, modules, and packs. Individual cells, which are the basic
units of energy storage, are assembled into modules. Each module groups a set number of
cells to provide a manageable intermediate structure. Multiple modules are then connected
to form the complete battery pack. Tesla's Model S employs a well-engineered battery pack
design, where clusters of cylindrical cells are grouped into modules, and the modules are
integrated into a robust pack that powers the vehicle's drivetrain [30]. This modular
12
architecture not only facilitates easier assembly and maintenance but also allows for
redundancy and safety measures to be implemented at the module level, ensuring the pack
continues to operate even if some modules fail.

The integration of these cells into modules and packs is crucial for managing thermal
behavior, ensuring safety, and optimizing performance. TMS are often integrated into the
pack to maintain temperature uniformity, prevent thermal runaway, and enhance the
lifespan of the battery. The modular approach also simplifies the application of BMS,
which monitor and regulate cell voltages, currents, and temperatures to ensure safe and
efficient operation. Figure 2.2 illustrates the hierarchy of cells, modules, and packs, showcasing how this scalable
architecture meets the diverse energy requirements of automotive applications. A summary of LIB pack capacity of
most popular EVs are provided in

Table 2.2.

Figure 2.2: EV battery: Cell, Module, Pack (a) Graphical representation (b) Actual Tesla Model-S Battery
Pack

13
Table 2.2: Most popular electric vehicle lithium-ion battery capacity

Brand Battery Battery Travel range Efficiency

Chemistry Capacity (mile) (kWh/mile)
(kWh)
Tesla model S Li-ion 75 249 0.33
VW e-Golf Li-ion 35 120 0.26
Renault Zoe Li-ion 41 250 0.26
Kia Soul EV LiPo 30.5 111 0.27
Hyundai Ioniq LiPo 28 124 0.23
Nissan Leaf Li-ion 30 107 0.28
Jaguar i-Pace Li-ion 90 234 0.36
BMW i3 Li-ion 33 114 0.27
Smart Fortwo Li-ion 17.6 65 .25

2.4 Existing BMS Architectures

The battery pack serves as the fuel tank for EVs, and its management is entrusted to the
BMS. The automotive BMS is analogous to the engine management system in internal
combustion engines. The BMS plays a critical role in EVs by continuously monitoring and
controlling the battery pack to ensure safe and efficient operation, and optimal utilization
[31], [32]. Besides monitoring important battery parameters including individual cell
voltage, current, temperature, the BMS addresses the control and safety management
requirements by estimating key battery states, including state of charge (SOC), state of
power (SOP), state of health (SOH), and remaining useful life (RUL), using the measured
parameters for effective monitoring and control. In addition to state estimation, the BMS
ensures the safe, reliable, and long-lasting performance of the battery pack under the
dynamic conditions of EV operation. It operates the battery pack within the safe limits
defined by manufacturers, preventing overcharging or undercharging by monitoring
voltage, current, and temperature [33], [34], [35], [36]. This task becomes more complex
in EVs, where battery packs comprise numerous individual cells, each requiring
independent monitoring to maintain optimum safety and efficiency. The BMS not only
tracks the voltage, current, and temperature of individual cells but also generates necessary

14
control signals for efficient operation and protection. Figure 2.3 illustrates a schematic of a
typical BMS, with the functionalities of its components discussed in subsequent sections.

Figure 2.3: Functionalities of automotive battery management systems

2.5 Typical Functionality of Automotive Battery Management Systems

2.5.1 Voltage, Current and Temperature Monitoring
The core function of a BMS is to monitor key parameters of the battery pack, including
individual cell voltage, temperature, and current. Effective monitoring not only helps
prevent hazardous situations like thermal runaway but also supports accurate state
estimations which are crucial for maintaining the performance and longevity of the battery
pack.

2.5.1.1 Voltage Measurement

Voltage measurement is critical as LIBs operate within a specific voltage range
depending on the chemistry, typically between 2.3 V to 4.2 V for most of the LIB cells. A
BMS must be capable of monitoring individual cell voltages in a module, module-level
voltages, and the overall pack voltage. Since battery packs are constructed by connecting
cells in series and parallel configurations, accurate voltage monitoring ensures that all cells
remain within their safe operating limits. This prevents overcharging or deep discharging,
both of which can significantly degrade cell performance and safety. Moreover,
measurement of individual cell voltage is essential for cell balancing.

15
2.5.1.2 Temperature Measurement
Temperature is a vital parameter that directly influences battery performance, state
estimations, and safety. Maintaining cell temperatures within the prescribed range is
essential to avoid thermal runaway and achieve reliable estimations of SOC and SOH.
Temperature monitoring is typically achieved using thermocouples, which can measure
individual cell temperatures. However, implementing thermocouples for each cell in a large
battery pack can be costly, complex, and adds weight due to the wiring harness. Despite
this, monitoring cell-level temperature is invaluable for fault prognosis and diagnostics, as
well as for ensuring consistent performance across all cells.

2.5.1.3 Current Measurement

Current measurement is equally important for SOC estimation and understanding the
charge and discharge behavior of the battery. For series-connected modules, the current is
uniform across all cells, allowing a single measurement for SOC estimation and protection.
However, in parallel configurations, the current splits among cells, necessitating individual
current measurements for cell-level diagnostics and prognostics. Current monitoring at the
pack level is essential for energy prediction and overall pack diagnostics. This can be
achieved using current shunts, which offer high accuracy but pose insulation challenges in
high-voltage systems, or hall effect sensors, which provide electrical isolation but are less
accurate and prone to electromagnetic interference (EMI).

2.5.2 Charge and Discharge Control

One of the primary responsibilities of a BMS is to ensure that each cell operate within
their safe working range to maintain performance, longevity, and safety. For example,
lithium iron phosphate (LFP) cells have a typical lower cut-off voltage rating of 2.5 V. The
BMS prevents any cells in the battery pack from discharging below this threshold. This
control mechanism protects the cells from damage due to over-discharge, which can lead
to capacity loss or potential safety hazards. In addition to discharge management, the BMS
monitors and regulates the charging process, which is critical for maintaining battery
health. Improper charging can significantly shorten a battery’s lifespan or result in cell
damage. Commonly, LIBs use a two-stage charging process involving constant current

16
(CC) and constant voltage (CV) stages. During the CC stage, the charger delivers a steady
current to rapidly charge the battery. As the battery approaches full charge (reached upper
cut-off voltage), the CV stage takes over, providing a constant voltage with a significantly
reduced current to top off the charge safely. This gradual reduction in current minimizes
stress on the battery and ensures the cells reach their full capacity without being
overcharged. The BMS plays an important role in managing voltage and current during
both stages of the charging process, ensuring they remain within the manufacturer-defined
limits specified in the battery datasheet. By carefully controlling these parameters, the BMS
prevents overcharging, fast charging beyond recommended levels, and other conditions
that could degrade the battery or compromise safety. This level of precision and control is
vital for optimizing the performance and reliability of LIBs in demanding applications like
EVs and grid-tied energy storage systems.

2.5.3 State Estimation

While voltage, temperature, and current can be directly measured in real time, crucial
parameters like SOC, SOH, RUL, and state of energy (SOE) must be estimated based on
these physical measurements. SOC acts as the fuel gauge for EVs, providing critical
information about the battery useful capacity. By monitoring the voltage and the charging
or discharging current, the depth-of-discharge (DOD) and capacity of the battery pack can
be tracked. Numerous algorithms are available to estimate SOC once voltage and current
measurements are obtained. The most employed method is the Coulomb counting
technique, which calculates SOC based on the integration of current over time. Advanced
methods, including Kalman filters and machine learning techniques, offer enhanced
accuracy and adaptability under dynamic operating conditions. In addition to SOC, the
BMS is responsible for calculating and monitoring the SOH of the battery. SOH reflects
the overall condition of the battery, accounting for factors such as aging, operating
temperature, and the voltage and current profiles of each cell and battery pack. This metric
provides insights into the remaining capacity of the battery and performance compared to
its original state. Estimating RUL based on SOH and other operational data allows for
proactive health-conscious battery management, enabling predictive maintenance and

17
timely replacement planning. Together, these advanced estimations ensure the battery
operates reliably, efficiently, and safely throughout its lifecycle.

2.5.4 Cell Voltage and SOC Balancing

Ensuring voltage uniformity among all cells in a battery pack is a critical function of
the BMS. In a series-connected module or pack, the voltage of each cell should remain
equal. If one cell deviates from the others such as being at 3.5 V while the others are at 4
V, it creates inefficiencies and compromises the overall performance of the battery pack.
The cell with the lower voltage would discharge to the lower cut-off voltage before the
others during discharging. Similarly, during charging, the weak cell would lag, reaching to
upper cut-off voltage while the other cells already reach the upper cut-off voltage (typically
4.2 V). This imbalance prevents the pack from utilizing the full capacity of the cells,
reducing efficiency and potentially causing premature wear on the weaker cell. Cell
balancing is the mechanism the BMS employs to address this issue, ensuring all cells
maintain consistent voltage levels and SOC. There are two primary strategies for cell
balancing: passive balancing and active balancing. Passive balancing involves discharging
cells with higher voltages through a resistor or similar load until their voltage matches that
of the lower-voltage cells. Active balancing, on the other hand, redistributes energy from
stronger cells to weaker ones during the balancing process, equalizing their potential and
improving overall pack utilization. These balancing techniques not only maximize the
pack’s efficiency but also prolong its lifespan by mitigating the stress caused by voltage
imbalances.

2.5.5 Thermal Management System Control

Temperature plays a pivotal role in determining the performance and lifespan of LIBs.
Higer operating temperatures above the recommended limits, often encountered in hot
climates, accelerate battery degradation, leading to faster depletion of capacity. Excessive
current usage further compounds the issue, causing the temperature to rise beyond optimal
levels. To address these challenges, an effective TMS is essential for maintaining the
temperature of the battery pack within the safe operating range. In colder climates, the TMS
must not only cool the battery pack when necessary but also have the capability to raise the
temperature to ensure optimal operation. The BMS continuously monitors the temperature
18
of individual cells and makes real-time adjustments to the TMS as needed. This control
ensures that the overall temperature of the battery pack remains balanced and within the
recommended range, enhancing both safety and operational efficiency while extending the
battery’s lifespan.

2.5.6 Isolation Detection

In high-voltage LIB applications, ensuring the battery pack is isolated from the chassis
is a critical safety requirement. Active monitoring of isolation levels is essential to detect
and address potential insulation breakdowns. Isolation faults can occur between the
positive and negative terminals of the battery, as well as between these terminals and the
vehicle’s chassis. A breakdown in insulation between high-voltage terminals creates a low-
resistance path between the positive and negative terminals, posing an extremely hazardous
condition. Similarly, isolation failure between high-voltage terminals and the chassis can
lead to current leakage through the metal chassis, which is regularly in contact with
humans, creating a significant safety risk. In EVs, where power and communication
interactions between low-voltage and high-voltage circuits are common, maintaining
proper isolation is particularly crucial. Components such as traction inverters, onboard
chargers, DC-DC converters, and communication systems must be designed with robust
isolation mechanisms to ensure safe operation. The BMS continuously monitors isolation
resistance levels to detect any faults and takes appropriate action to mitigate risks. By
actively ensuring proper isolation, the BMS not only safeguards the vehicle’s electrical
systems but also protects occupants from potential electrical hazards.

2.5.7 Communication and Data Logging

Effective communication and data logging are vital functions of a BMS, enabling
seamless integration of the battery pack with other vehicle components. The BMS collects
critical data, such as voltage, temperature, and current, and processes it to estimate
parameters such as SOC, SOP, SOE, SOH and so on. This information is then
communicated to other systems; for instance, the motor control system uses SOP and SOE
to meet acceleration demands, while the SOC is displayed on the instrument cluster for
user reference. Communication within the vehicle is facilitated by the controlled area
network (CAN) protocol, which enhances system reliability and reduces wiring complexity
19
by allowing multiple devices to communicate over a single network. When designing a
CAN network for an EV, factors such as data rate, network size, reliability, cost, and
industry standards must be considered. High-speed data transport applications may require
the CAN flexible data-rate (CAN-FD) protocol, while network size should accommodate
the required number of nodes. Reliability is paramount for communication between BMS
components and external devices, such as charging stations, ensuring dependable
operation. Additionally, cost implications of hardware and software requirements should
be evaluated, and industry-specific protocols like J1939 or CANopen are often preferred
for automotive applications.

Data logging is another critical aspect of a BMS, encompassing both real-time and
historical data related to battery usage. High-resolution data collection can impose
significant memory burdens on onboard systems. However, advancements in cloud
computing and the internet-of-things (IoT) have revolutionized data storage and analysis
in the EV sector. IoT-enabled BMS solutions can offload extensive data to cloud platforms,
where it is analyzed using sophisticated diagnostic algorithms to assess the health and
performance of individual cells. Furthermore, machine learning techniques integrated with
IoT systems enable predictive analysis and optimization, improving battery lifecycle and
performance while reducing computation efforts. This fusion of communication and data
logging capabilities is pivotal for the efficient and reliable operation of modern EVs.

2.6 Classification of Automotive Battery Management System

Automotive BMS can be classified based on their configuration and architecture,
primarily into centralized BMS and modular and scalable BMS. Each type has unique
advantages and limitations, making them suitable for specific battery pack designs and use
cases.
2.6.1 Centralized BMS
A centralized BMS integrates all monitoring and control functions onto a single
electronic circuit board. In this configuration, all battery cells in the pack are directly
connected to a central BMS, as illustrated in Figure 2.4. The BMS performs vital tasks such
as monitoring cell voltage, temperature, and current, and processing the data using an
onboard microprocessor on the same printed circuit board (PCB). The centralized
20
architecture offers several benefits. It is compact, which simplifies installation, and it is the
most cost-effective solution as it consolidates all functionalities into a single unit. However,
this design also has significant drawbacks. The need to connect every cell in the pack
directly to the BMS results in an excessive number of wires, complicating maintenance and
increasing the risk of wiring issues. For large battery packs with numerous cells, centralized
BMS may not be practical due to the complexity of wiring and the associated maintenance
challenges.

Figure 2.4: Architecture of a centralized BMS

2.6.2 Modular and Scalable BMS

Modular and scalable BMS have become increasingly popular due to their flexibility
and adaptability to varying battery pack configurations. This type of BMS consists of a
master controller board and multiple sub (slave) controller boards, as shown in Figure 2.5.

Figure 2.5: Modular- Master- Slave- Daisy Chain Configuration

Each battery module in the pack is equipped with its own sub-module that monitors
individual cell voltage, temperature, and current within that specific module. The sub-
21
modules communicate with the master controller through designated communication
channels, which collectively monitor and control the entire battery pack. The modular
design offers several advantages. Sub-modules are connected in either a daisy chain or
multi-drop configuration as shown in Figure 2.6, allowing for a highly scalable system.
This enables seamless expansion by adding more sub-modules without the need for
additional master controllers. The use of analog front-end (AFE) chips like LTC6804 from
Analog Devices® or BQ76PL536 from Texas Instruments® ensures efficient data
communication and system scalability. Furthermore, modular BMS minimize wiring
complexity by reducing the length and number of wires required. This not only simplifies
maintenance but also mitigates risks associated with long wires, such as signal degradation
or wiring errors. The localized nature of sub-modules allows them to remain closer to their
respective battery modules, enhancing reliability and efficiency. Modular BMS also
provide straightforward scalability, making them an excellent choice for larger or more
complex battery packs. In summary, the centralized BMS are cost-effective and suitable
for smaller battery packs, modular and scalable BMS offer superior flexibility,
maintenance ease, and scalability, making them the preferred choice for modern EV
applications.

Figure 2.6: Modular- Master- Slave- Multi Drop Configuration

2.6.3 Wireless BMS

Wireless battery management systems (WBMS) represent a significant advancement
over traditional wired-BMS, addressing many of the limitations associated with wiring
22
complexity and physical connectivity in large battery packs. Unlike wired systems that rely
on communication protocols such as CAN-bus and I2C/SPI, WBMS replaces the extensive
wiring network with wireless communication channels, offering several key benefits. The
elimination of extensive wiring in WBMS reduces system weight, cost, and design
complexity, while improving reliability by mitigating issues such as connection failures
caused by vibrations in EVs. The absence of physical connectors and galvanic isolation
requirements enhances safety and simplifies troubleshooting. WBMS also enables greater
flexibility in the placement of sensors and BMS modules, making it particularly
advantageous for high-capacity LIB packs with thousands of individual cells.

A typical WBMS architecture mirrors the functional design of a wired-BMS, with

sensor nodes gathering data such as voltage, current, and temperature for estimation of
states. However, instead of wired communication, these data are transmitted wirelessly to
the master controller, onboard display units, or cloud-based systems. This wireless
approach facilitates scalable and fault-tolerant designs, allowing for easy component
replacement without requiring extensive reconfiguration of the system. A schematic layout
of a typical WBMS is shown in Figure 2.7.

Figure 2.7: Schematic layout of a generic WBMS

Recent developments [37] in WBMS have explored innovative features such as in-node
memory, enabling data recovery after temporary connection failures and ensuring
enhanced data integrity. Despite promising progress and satisfactory real-world
demonstrations, WBMS technology remains in its early stages. Further research is needed
to address challenges related to cost, implementation, data transfer rates, and system
23
reliability, paving the way for industry-ready WBMS solutions. As the automotive industry
continues to embrace LIB-powered systems, WBMS offers a compelling path toward
lighter, more efficient, and reliable battery management architectures. One of the first
industrial development of WBMS is LTC6811 developed by Linear Technology® (now
Analog Devices®) [38]. However, it is only capable of multicell battery stack monitoring.
Wei et al. [39] provided a detailed review of smart cell technology and advanced BMS
technologies.

WBMS can be classified based on the type of wireless communication technologies

used for data transmission and control. In the literature, WBMS topologies are commonly
grouped into five main categories: Bluetooth-based WBMS (BWBMS), Zigbee-based
WBMS (ZWBMS), Wi-Fi-based WBMS (Wi-Fi-WBMS), IoT-based WBMS (IoT-
WBMS), and Cloud-based WBMS (CWBMS) [37]. Each topology offers unique
advantages and trade-offs in terms of cost, complexity, scalability, and data transfer
capabilities, making them suitable for different applications. Figure 2.8 illustrates the
detailed classification of WBMS topologies, capturing the breadth of approaches explored
by researchers.

In addition to these widely adopted topologies, other wireless communication

technologies, such as near field communication (NFC) and Wi-Fi HaLow, hold potential
for WBMS applications. However, these technologies have yet to be successfully
implemented in WBMS due to challenges like over-specification, limited generalizability,
and higher system costs. While they may not currently be practical for WBMS, these
technologies offer possibilities for future exploration. Some studies in the field focus on
general wireless communication principles or propose enhancements without adhering to
a specific WBMS topology. These contributions, while not directly classified into existing
categories, provide valuable insights for the advancement of WBMS. This classification
highlights the diverse technological pathways being pursued to optimize WBMS
performance and integration in modern battery systems.

24
 Figure 2.8: Classification of WBMS topologies.

2.6.5. Cloud-Based Battery Management System (CWBMS)

Cloud-based battery management systems (CWBMS) leverage cloud platforms for
advanced monitoring, diagnostics, and management of LIBs. Tanizawa et al. [40]
introduced a cloud-based system for battery status monitoring in EVs, while Kim et al.
proposed a CWBMS for fault diagnosis in large-scale stationary battery energy storage
systems (BESS) [41]. These systems address the challenges of integrating advanced
functionalities, such as aging prognostics and optimization strategies, by utilizing the
cloud's computational resources. Li et al. [42] enhanced this approach by employing IoT
technology to create a DT of the LIB system on the cloud, allowing for real-time estimation
of SOC and SOH using adaptive extended H-infinity filters and particle swarm
optimization. The DT architecture of CWBMS offers precise parameter monitoring, online
state estimation, and fault prediction, enabling a significant improvement over traditional
onboard BMS. Moreover, the DT technology in BMS can be utilized for state prediction
ahead of time based on real-time battery operating parameters and preventive actions can
be taken. However, existing studies did not demonstrate the actual implementation of DT
in BMS and lacks in proving detailed insights into software and hardware design and have
limited real-world validation. CWBMS shows immense potential for scalability and
reliability by offloading computational tasks to the cloud, thus overcoming the constraints
of onboard BMS in managing large datasets and running sophisticated diagnostic
algorithms. As CWBMS technology continues to evolve, its integration with IoT and

25
advanced optimization techniques is expected to revolutionize BMS for automotive and
stationary applications. The integration of cloud computing, DT, IoT in BMS applications
will not only provide optimum control and safety utilizing state predictions while at the
same time the long-term data storage capability will significantly aid in cell selection and
capacity estimation for second-life applications of retired EV batteries.

2.7 Core Temperature Estimation of Lithium-ion Battery

Research has shown that key battery states, including SOC [43], SOH [44], and battery
degradation [45], are significantly influenced by temperature, as discussed in the above
section. However, current BMS technology only utilizes sample surface temperature
measurements from a few cells in a LIB module or pack. Measuring the temperature of
each cell using physical sensors increases complexity and costs. Furthermore, surface
temperature readings fail to provide a comprehensive view of the internal thermal state of
a cell, unlike core temperature measurements. Direct measurement of core temperature
using physical sensors is impractical, as embedding sensors into the core of a cell can
damage its structure and adversely affect battery health.

Consequently, the estimation of core temperature in LIBs has become an open research
area. Accurate core temperature information is essential for rapid and effective thermal
management, safety, and optimal operation. It is also critical for accurate estimation of
other states and for enabling key BMS functionalities such as cell balancing [45] and fault
detection/diagnosis [46]. Coulombic efficiency, a vital performance metric, is particularly
sensitive to temperature variations during charging and discharging. This highlights the
importance of precise thermal data in ensuring reliability, safety, and extended battery life.

Temperature information is especially vital in extreme climates. In cold conditions,

battery capacity degrades sharply, necessitating preheating to achieve optimal performance
[47], [48]. Studies indicate that deviations of just 0.1°C beyond the safe operating range
can degrade battery capacity by approximately 5% [49], an effect exacerbated during rapid
discharge due to increased heat generation [50]. While surface temperature is measurable
via sensors, direct measurement of core temperature is impractical, particularly in EV
battery packs containing thousands of cells, due to cost and complexity. To address this,
26
researchers are exploring multi-dimensional sensing and self-healing functions in smart
batteries [39], [51], [52], [53]. However, even with such advancements, precise
temperature estimation remains critical, prompting the development of computationally
efficient and cost-effective strategies for onboard BMS.

2.7.1 Classification of Core Temperature Estimation Strategies

Core temperature estimation in LIBs can be broadly classified into model-based

techniques and data driven techniques. Model-based core temperature estimation typically
involves two models: a heat generation model and a heat transfer model [54]. Heat is
generated in LIBs during charging, discharging, and idle conditions due to exothermic
reactions and transport processes. If heat dissipation is insufficient, core and surface
temperatures rise, increasing the risk of thermal runaway, especially in hard-cased batteries
under fast charging or in hot environments. Techniques such as Bernardi’s heat generation
model [55] and adaptive estimation strategies are used to calculate heat generation and
predict cell temperature, incorporating battery states. Closed-loop estimation schemes
further enhance accuracy by using temperature feedback. These methodologies aim to
create reliable, cost-effective solutions for low-cost BMS, addressing the unique challenges
of core temperature estimation in LIBs. A typical functional block diagram of temperature
estimation scheme for LIB is shown in Figure 2.9.

Figure 2.9: Schematic layout of a generic temperature estimation strategy for a LIB cell.

As depicted in Figure 2.9, a standard temperature estimation scheme incorporates two

key components: a heat generation model and a heat transfer model. Heat generation
models can be broadly classified based on their modeling strategies and the sources of heat
generation. From a modeling perspective, these models are categorized into three primary
27
types: physics-based electrochemical models [34], [56], [57], [58], equivalent circuit
models (ECM) [59], [60], [61], and black-box models [62], [63], [64]. In terms of heat
generation sources, models are divided into concentrated models, distributed models [65],
and heterogeneous models [59], [66]. Concentrated heat generation models assume all heat
originates from the core, simplifying the modeling process but limiting detail. Distributed
models, by contrast, assume heat is generated uniformly across the entire cell geometry,
providing a balanced trade-off between complexity and accuracy. Heterogeneous models,
the most detailed of the three, account for non-uniform heat generation across various cell
layers, resulting in temperature and current density gradients. These models deliver highly
accurate predictions but require extensive experimental data, making them the most
complex to implement.

Heat transfer models are further classified into finite element analysis (FEA)-based
models [61], [67], [68], [69], [70], heat capacitor-resistor models (either lumped or
distributed parameters) [62], [71], [72], [73], [74], and data-driven techniques. Heat
capacitor-resistor models use an analogy between electrical and thermal systems and are
categorized into lumped and distributed parameter models, as shown in Figure 2.10. Lumped
parameter models are straightforward, making them suitable for online applications.
However, they are limited to predicting one or two average temperatures and do not
account for the spatial non-uniformity of temperature distribution, particularly in larger
cylindrical LIB cells. Distributed models, although computationally demanding, provide
detailed insights into temperature variations across the cell [75], [76].Comprehensive LIB
models that incorporate the thermal characteristics of different cell layers have been
developed and analyzed in studies such as [16], [77], [78], [79], [80], [81]. These models
range from one-state/node models, which estimate only the core temperature, to two-
state/node models that provide data on both core and surface temperatures, offering a
nuanced understanding of thermal dynamics in LIBs.

The term battery thermal model refers to a framework in which total heat generation,
derived from a heat generation model, is used as an input parameter to estimate heat
dynamics in a battery. While thermal modeling of LIBs is a distinct research field, this
study focuses specifically on temperature estimation strategies. Nevertheless, because most
28
temperature estimation methods rely heavily on thermal models, an overview of key
thermal modeling techniques relevant to temperature estimation is included. Researchers
have employed various heat generation and transfer models to create temperature
estimation schemes, leading to diverse and complex classification criteria. In general,
temperature estimation approaches can be categorized into six primary groups:
electrochemical thermal modeling, equivalent electric circuit models (EECM), ML
techniques, numerical models, direct impedance measurements, and methods involving
magnetic nanoparticles, as depicted in Figure 2.10. These categories reflect the wide array of
strategies developed to enhance the accuracy and efficiency of temperature estimation in
LIBs.

Figure 2.10: Family of (A) Heat generation model, (B) Heat transfer model, (C) Temperature estimation
strategy

29
2.7.2 Core Temperature Estimation Strategies
2.7.2.1 Electrochemical Thermal Modelling-Based Temperature Estimation
Researchers initiated the thermal modeling of batteries in the early 1990s by
mathematically representing the electrochemical system of the battery to emulate actual
battery thermal characteristics under various operating conditions, geometries, or cooling
rates. These models vary in complexity, ranging from simple one-dimensional (radial
direction) models [15], [71], [82], [83], [84], [85], [86] to intricate three-dimensional
thermal models [87], [88], [89], [90], [91]. To mathematically represent the
electrochemical behavior of cells, researchers typically employ different analytical
techniques.

One-dimensional models are commonly employed for fundamental temperature

prediction in LIBs. These models typically operate under the assumption of isothermal,
constant-current conditions with lumped thermophysical properties and consistent heat
generation rates. Their simplicity makes them suitable for basic applications; however, they
are limited in addressing complex thermal behaviors. On the other hand, advanced three-
dimensional models demand an in-depth understanding of the thermodynamic properties
of battery materials and components. These models incorporate the effects of various
phenomena, such as heat generation due to ohmic resistance, chemical reactions,
polarization, mixing processes, and electrode kinetic resistance. While these
comprehensive models yield exceptionally accurate temperature predictions, they are
generally developed for battery design and optimization rather than real-time temperature
estimation. Their high computational demands render them impractical for onboard BMS
that operate under constrained computational resources. These detailed models excel in
capturing the nonlinear and dynamic behaviors of LIBs over time. However, their
implementation often necessitates extensive experimental efforts to accurately characterize
system properties and operational parameters. Certain intrinsic characteristics, such as
transport properties, thermodynamic behavior, and heat generation, remain challenging to
quantify with precision, adding to the complexity of deploying such models in practical
applications.

30
 Thomas and Newman [92] introduced a pioneering heat generation model through
electrochemical analysis, which has become a cornerstone for understanding thermal
dynamics in LIBs. This model was designed to estimate the total heat generated during
charging and discharging cycles, offering valuable insights into the thermal behavior of
battery systems. The governing equation (Equation 2.1) proposed by Thomas and Newman
remains a critical tool for researchers and engineers striving to enhance the thermal
performance and safety of LIBs.

𝜕𝑈 𝑎𝑣𝑔 ̇ 𝑎𝑣𝑔 ̅𝑗𝑎𝑣𝑔 ) 𝜕𝐶𝑗 𝑑𝑣

̅𝑗 − 𝐻
𝑄 = 𝐼 (𝑉 − 𝑈 𝑎𝑣𝑔 ) + 𝐼𝑇 − ∑𝑖 ∆𝐻𝑖 𝑟𝑖 − ∫ ∑𝑗(𝐻 (2.1)
𝜕𝑇 𝜕𝑡

In Equation (2.1), Q represents the rate of heat generation or consumption within the
cell, V and U denote the terminal voltage and equilibrium or open circuit voltage,
respectively, I is the current associated with charging or discharging, and T signifies the
cell temperature. Additional parameters include ∆𝐻𝑖 , the enthalpy change of reaction i, 𝑟𝑖 ,
̅𝑗 , the partial molar enthalpy of species j, 𝑐𝑗 , the concentration of
the rate of reaction I, ∆𝐻
species j, and t and v, representing time and cell volume, respectively. These properties are
volume-averaged, as indicated by the superscript ‘avg’. This model emphasizes accurate
heat generation estimation during charging and discharging processes, but it does not
directly address temperature estimation. Despite its accuracy, the computational intensity
of the model makes it unsuitable for real-time applications, especially within resource-
constrained onboard BMS.

One widely adopted framework is the Doyle–Fuller–Newman model [71], [93],

commonly known as the pseudo-two-dimensional (P2D) model. This model uses nonlinear
partial differential-algebraic equations to describe the internal dynamics of LIBs. While
robust and detailed, its computational demands restrict its feasibility for online state
estimation in embedded BMS. A key challenge lies in the extensive experimental data,
expertise, and specialized equipment required to determine accurate model parameters.
Furthermore, these parameters are highly sensitive to operating conditions, leading to
diminished accuracy in practical applications compared to controlled laboratory
environments.

31
 To address these limitations, Al Hallaj et al. [15] proposed a simplified one-
dimensional transient thermal model with lumped parameters, adequate for cell design
purposes. This approach simulates the thermal behavior of scaled-up LIBs without
requiring detailed component-level heat generation contributions. Such simplified models
are particularly beneficial for applications that prioritize accessibility and ease of use.
Meanwhile, other researchers have leveraged detailed electrochemical models to
investigate pulse power limitations, mitigate thermal runaway, and design TMS [94], [95],
with a focus on LIB cell and pack design.

Fang et al. [96] utilized a lumped electrochemical-thermal-coupled model to predict the

thermal performance of LIBs under varying operating temperatures, including the
performance of individual electrodes. Despite validation against experimental data under
constant and pulsed current conditions, the model's applicability to real-world scenarios
remains limited. Similarly, studies such as [97] have explored the effects of charging
currents on internal temperature behavior, yielding valuable insights into thermal
dynamics. Gerver et al. [98] introduced a multi-dimensional electrochemical thermal
model incorporating specific cell characteristics, enabling precise thermal performance and
heat generation analysis. However, the high complexity and computational burden of such
models limit their utility in real-time applications, confining their use to design and
simulation tasks.

Recognizing the need to balance accuracy and computational feasibility, Zhang et al.
[99] developed a two-state thermal model employing discretization and inverse modeling
techniques. This method estimates total heat generation without requiring detailed
knowledge of individual heat sources or boundary conditions. Similarly, Wang et al. [100]
employed a high-fidelity electrochemical model integrated with onboard measurements
such as terminal voltage and current to estimate cell temperature across various C-rates
during charging and discharging. By incorporating a dual ensemble Kalman filter (DEKF)
with enhanced single-particle dynamics, the model correlates terminal voltage with cell
temperature and lithium-ion (Li+) concentration. However, the complexity of determining
Li+ concentration and the high computational costs restrict its real-time application. Marelli
and Corno [101] extended this approach by employing a pseudo-2D electrochemical model
32
combined with a soft-constrained dual unscented Kalman filter (DUKF) to estimate the
spatial distribution of internal temperature within LIBs. Although primarily intended for
Li+ concentration estimation, this approach holds potential for adaptation to temperature
prediction, albeit at a high computational expense. Smith et al. [94] developed a one-
dimensional electrochemical lumped thermal model to study pulse power limitations and
LIB pack thermal behavior. While this approach is effective for design purposes, its
complexity limits real-time feasibility. Table 2.3 summarizes various electrochemical
thermal modeling strategies for temperature estimation. In general, the major drawbacks of
electrochemical model-based techniques are their modeling intricacies and computational
demands, which make them unsuitable for cost-effective, real-time BMS applications.
Moreover, it needs in-depth knowledge of electrochemistry and mathematical background
to develop such models.

Table 2.3: Summary of electrochemical thermal modelling-based temperature estimation

strategies

Reference Types of Models Important Note

Al Hallaj et al.A transient one- The detailed properties of electrodes,
[15] dimensional thermal electrolytes and separator are required for
model with lumped developing the heat generation model.
parameters
Fang et al. Lumped parameter Experimentally validated model capable to
[96]. electrochemical- estimate cell average temperatures at dynamic
thermal-coupled model operating temperature and current.
Gu and Wang Thermal energy Temperature-dependent thermal model
[75] generation model, considered physicochemical properties able to
multiphase micro- predict average cell temperature as well as the
macroscopic temperature distribution inside a cell under
electrochemical model dynamic operating temperature. Used
volume-averaging technique, numerical
simulations for temperature estimation.
Kumaresan et One-dimensional Experimentally validated thermal model
al. [76] thermal model where effect of temperature variation on
model parameters are considered.

33
Kim et al. [97] Two-dimensional Experimentally validated model provides
modelling + Finite temperature distribution based on potential
element method (FEM) and current density distribution.
Gerver et al. A multi-dimensional Experimentally validated model considering
[98] electrochemical the variations of thermal properties of each
thermal model cell layer.
Wang et al. High-fidelity Experimentally validated model robust under
[100] electrochemical model wide range of C-rates during the
+onboard charging/discharging of cell.
measurements + dual
ensemble Kalman filter
(DEKF)
Marelli and Pseudo-2D Model can provide spatial distribution of
Corno [101] electrochemical model internal temperature, validated through
and soft-constrained MATLAB simulation studies.
dual unscented Kalman
filter (DUKF)
Smith et al. A one-dimensional Experimentally validated model adaptive to
[94] electrochemical lumped different drive-cycles including FUDS and
thermal model HWFET.

2.7.2.1 Equivalent Electric Circuit Model-Based Temperature Estimation

The EECM is a widely used framework for simulating the thermal dynamics of LIBs.
It utilizes electrical system parameters to construct a heat capacitor-resistor-based thermal
model, which can be employed for temperature estimation. EECMs are typically
categorized into two primary types: first order and second-order models. The first-order
model consists of a single heat capacitor representing the overall thermal energy storage
within the cell, while the second-order model includes two capacitors, one for the core and
another for the cell surface. This additional complexity allows the second-order model to
capture more intricate dynamic thermal behaviors compared to its first-order counterpart
[49].

In Figure 2.11, Q represents the heat generation rate, Cc and Cs denote the heat
capacitance of the core and surface, respectively, Tin and Tout refer to the core and surface
temperatures of the cell, and Tamb is the ambient temperature.
34
 (a) (b)
Figure 2.11: Thermal model of a LIB cell (a) First-order model, (b) Second-order model

EECMs are further classified into lumped-parameter and distributed-parameter models

based on their level of complexity [102]. Lumped-parameter models prioritize
computational efficiency and are often constructed with a single temperature input. For
improved accuracy, some researchers incorporate both surface and core temperatures and
consider correlations between cell geometry and physical properties [103]. However, these
simplifications can lead to a reduction in temperature estimation accuracy compared to
more detailed thermal models.

EECMs can also be distinguished based on their estimation scope: single-state/node

models estimate only the core temperature [104], whereas two-state/node models estimate
both surface and core temperatures [99]. Model parameters for these thermal models are
typically determined through experimental techniques, such as electrochemical impedance
spectroscopy (EIS), or derived from measurable quantities like voltage, current, and
temperature. To enhance robustness, modern approaches consider variations in the SOC
and SOH, along with the estimation of surface and core temperatures under dynamic
conditions.

These models are often employed at both the cell and pack levels for temperature
estimation, with configurations ranging from first order to second-order models. The
mathematical formulation for heat generation, as proposed by Bernardi et al. [55], is
expressed as Equation 2.2.

𝑑𝑉𝑜𝑐𝑣
𝑄 = 𝐼(𝑉 − 𝑉𝑂𝐶𝑉 ) + 𝐼𝑇𝑐 (2.2)
𝑑𝑇𝑐

𝑑𝑉𝑜𝑐𝑣
Here, 𝑉𝑜𝑐𝑣 is the open-circuit voltage, and represents the entropy coefficient. The
𝑑𝑇𝑐

core temperature (Tc) and surface temperature (Ts) are then estimated using the
35
mathematical frameworks of the first order and second-order thermal models, as illustrated
in Figure 2.11. The governing equations for temperature estimation in the first order and
second-order models are represented in Equations 2.3 and 2.4, respectively.

2.7.2.2 EECM-Based Cell Temperature Estimation

One of the primary challenges in EECM-based strategies is the accurate identification
of model parameters. Forgeze et al. [16] addressed this by conducting transient experiments
with current pulses of varying magnitudes to elevate the internal temperature and determine
the thermal model parameters. They calculated heat transfer coefficients and heat capacity
from these experiments to develop a lumped-parameter thermal model. Utilizing EIS for
parameter identification, they applied 2 Hz current pulses to estimate the Tc based on the
measured Ts. Their first-order thermal model also incorporated entropy changes, providing
a foundational mathematical framework for temperature estimation in LIBs.

𝑅 𝑅
𝑇𝑖𝑛 = 𝑇𝑠 (1 + 𝑅 𝑖𝑛 ) − 𝑇𝑎𝑚𝑏 𝑅 𝑖𝑛 (2.3)
𝑜𝑢𝑡 𝑜𝑢𝑡

Despite its utility, the model by Forgeze et al. [16] has limitations, particularly due to
the lack of quantitative analysis of heat generation effects. Model parameters derived from
EIS considering limited operating condition and aging profile often fail to align with the
dynamic operating conditions in real-world applications, resulting in inaccuracies in
capturing thermal dynamics. Additionally, their assumption of a uniform internal cell
temperature contradicts their findings, which showed significant temperature variations
exceeding 10°C at different internal points. Moreover, the need to measure surface
temperature using sensors on each cell limits its scalability for large-scale applications.

Maleki and Shamsuri [105] investigated thermal modeling for notebook computer LIB
packs, revealing that temperature increases during charging are mainly attributed to heat
dissipation from power electronics, while heat generated within the LIB cell dominates
during discharging. This insight is critical for designing efficient TMS, especially for fast-
charging applications. Surya et al. [49] extended this research by developing a second-
order thermal model using a Kalman filter to estimate core and surface temperatures.
However, the omission of environmental uncertainties reduced the real-world applicability
of the model. Lin et al. [106] validated a second-order thermal model and an EECM-based
36
two-state thermal model under a variety of conditions, demonstrating robust accuracy.
However, further testing with standard drive cycles is necessary. Recognizing that cell
aging significantly influences EECM parameters, Lin et al. [78], [107] proposed an
enhanced least squares (LS) algorithm that adapts to aging and other uncertainties.
Similarly, Dai et al. [108] employed joint Kalman filtering techniques to improve
temperature estimation accuracy by accounting for unmeasurable modeling errors and
time-varying external thermal resistance as mentioned in Equation 2.4.

𝑅
(1+ 𝑖𝑛 + 𝐶𝑠 𝑅𝑖𝑛 𝑠)
𝑅𝑜𝑢𝑡
𝑇𝑖𝑛 (𝑠) = 𝑅 𝐶𝑐 1
𝑄(𝑠) (2.4)
𝐶𝑠 𝐶𝑐 𝑠 +(𝐶𝑠 +𝐶𝑐 + 𝑖𝑛 )𝑠+
2
𝑅𝑜𝑢𝑡 𝑅𝑜𝑢𝑡

Here, s represents the Laplace operator, with other parameters defined as in Figure 2.11.

Dai et al. [108] enhanced the modeling accuracy by developing separate thermal
models for the core and battery shell, incorporating time-varying external heat exchange
coefficients. However, the authors did not detail computation times or hardware
requirements, leaving questions about the model’s practical feasibility. Doughty et al. [109]
and Park et al. [103] explored a middle ground between detailed and lumped-parameter
thermal modeling, developing two-state thermal models capable of predicting both core
and surface temperatures. These reduced-order models (ROMs) simplify complex thermal
problems into a manageable set of parameters, balancing accuracy and computational cost.
Chen et al. [104] proposed a hybrid approach combining a lumped two-state thermal model
with a second-order ECM using 2RC elements and a joint Kalman filter (JKF). This
strategy demonstrated enhanced prediction accuracy and adaptability to varying
temperatures and SOC conditions.

Xie et al. [110] developed a one-dimensional (radial) lumped parameter thermal model
with a dual Kalman filter (DKF), offering temperature distribution readings at three points
within cylindrical LIBs. By considering the anisotropy of thermal conductivity, this three-
node model improved accuracy and robustness. However, its reliance on a 1-RC ECM-
based heat generation model limits precision compared to 2-RC ECM approaches. Pan et
al. [54] employed online parameter estimation with particle swarm optimization during
pulse discharge experiments under varying ambient temperatures. Their hybrid model,
37
combining a 2RC ECM with a multi-node heat transfer model, produced results
comparable to finite element method (FEM) simulations while reducing computational
costs by approximately 90%. However, the study did not address the effects of cell aging
or extend the approach to pack-level thermal modeling. Sun et al. [111] integrated heat
dissipation through radiation into a lumped thermal model for core temperature estimation,
using an extended unscented Kalman filter (EUKF) to enhance accuracy and robustness.
While effective, the model assumed constant parameters, which limits its real-world
adaptability. Zhu et al. [112] introduced a lumped two-state thermal-electrical model,
accounting for adjacent cell thermal impacts during modeling. They used an extended state
observer (ESO) to address uncertainties and time-variant parameters, focusing on rapidly
self-heating batteries. Xiao Y. [113] proposed a virtual thermal sensor (VTS) approach,
combining a tuned thermal model with a Kalman filter and an online parameter
identification algorithm. While adaptive to environmental changes, the method still relies
on sensor feedback, limiting its sensorless potential. Surya and Mn [50] demonstrated the
impact of fast discharge on LIB core temperature using a combination of a 1-RC ECM, a
single-state thermal model, and a Kalman filter for core temperature estimation. While
these models provide valuable insights, their applicability to fast charging and discharging
scenarios remains limited.

2.7.2.3 EECM-Based Temperature Estimation of LIB Pack

While much of the research on temperature estimation has focused on individual LIB
cells, there has been growing interest in extending these studies to LIB packs. Ma et al.
[114] developed a ROM for LIB packs, emphasizing core temperature estimation by
incorporating the characteristic internal electrical resistance of the cells. They utilized the
recursive least squares (RLS) algorithm for thermal parameter identification. However,
their approach relied on several simplifying assumptions, such as uniform parameters
across all cells and identical thermal behavior within cell rows. Moreover, their model did
not account for heat transfer between cells via conduction through tabs and wires,
potentially introducing errors in temperature estimation. Ismail et al. [115] approached
pack-level thermal modeling by scaling up a single-cell thermal model using simulation
techniques. Although this method achieved notable accuracy, it was based on idealized

38
 assumptions, including uniform cell characteristics, constant ambient conditions, and a
fully efficient discharging process. Such assumptions diverge significantly from the
variability and complexities encountered in real-world applications, thereby limiting the
generalizability of their results.

These studies underscore the critical need for advancing pack-level temperature
estimation methods that can account for the diverse conditions present in practical
scenarios. One notable limitation of EECM-based temperature estimation techniques is
their reliance on model parameters which changes with aging and operating conditions.
Furthermore, adaptive models require online sensor feedback. The accuracy of these
methods depends heavily on precise knowledge of the cell's thermal properties, heat
generation rates, and thermal boundary conditions factors that are influenced by cell aging,
temperature fluctuations, and other operational uncertainties. Addressing these challenges
is essential to develop more robust and accurate pack-level thermal models that can support
real-world applications of LIB systems. A summary of EECM-based temperature
estimation strategies is mentioned in Table 2.4.

Table 2.4: Summary of EECM-based temperature estimation strategies

Reference Types of Models Important Note

Lumped Parameter heat The thermal model was validated using
Mahamud et
capacitance–resistance ANSYS-Fluent simulation as well as
al. [102]
thermal model experimental studies.
Forgeze et al. Lumped Parameter, Single- Experimentally validated thermal model,
[16] State, First-order model considered entropy changes in the model.
Lumped Parameter, Two- The model can predict both SOC and
Surya et al.
State, Second-order model surface temperature. Validated by
[49]
+Kalman Filter (KF) simulation study in MATLAB.
Experimentally validated using
Lumped Parameter, Two- electrochemical impedance spectroscopy.
Lin et al. [106]
State, Second-order model The model is robust to high current rate,
varying temperature and SOC.

Li and Yang Extended lumped parameter, The thermal model is robust against
[116] Two-state, Second-order temperature variation, cell ageing, and SOC
model + Forgetting factor variation. The heat transfer modes are
39
 Recursive Least Square developed in ANSYS Multiphysics
(FFRLS) Simulation. All models are also
experimentally validated.
Lumped parameter, Two- Model considered the effect of cell aging
Lin et al. [78], state model + Least square and uncertainties in practical operation.
[107] (LS) algorithm + Nonuniform Model validation is done experimentally as
forgetting factors (NUFF) well as through simulation studies.
The model can capture influence of
Lumped-parameter model +
Lin et al. [117] overpotential entropy changes. Validated
Closed-loop observer
using simulation study only.
The experimentally validated model can
Lumped parameter, Second- capture the effect of entropy changes,
Sun et al.
order, Single state thermal overpotential, surface and ambient
[118]
model + KF temperature changes, variation in
charge/discharge current profile.
Lumped parameter, Second- Experimentally validated model is built
Dai et al. [108] order, Two-state model using time dependent thermal resistance
+JKF+ LS algorithm and capable to address initialization error.
Doughty et al. The model is robust to the ambient
Lumped parameter, Two-
[109] and Park temperature and SOC variation. Validated
state model + Extended KF
et al. [103] using simulation study only.
The experimentally validated model is able
Lumped parameter, Two-
Chen et al. to capture constantly varying temperature,
state thermal model + Joint
[104] SOC, and surface thermal resistance due to
KF (JKF)
dynamic operating conditions.
The model considered battery geometry
Lumped Parameter, Second- and robust to the changes in
Pan et al. [54] order, multi-node model + charge/discharge profile. The performance
particle-swarm algorithm is compared with an FEA model with
experimental data.
The anisotropy of thermal conductivity and
One-dimensional (radial)
variation of SOC and ambient temperature
lumped parameter, Three
Xie et al. [110] are considered. The model is developed
node model + Dual KF
using FEM and computational fluid
(DKF).
dynamics and experimentally validated.

40
 The model considered heat radiation from
Lumped parameter, single- the surface of the cell, and it is robust to
Sun et al.
state model + Extended sensor measurement error and bias.
[111]
unscented KF (EUKF) Validated using experimental and
simulation studies.
The model is validated using EIS data. It
Lumped parameter, Two-
Zhu et al. can consider the thermal impact of
state model + extended state
[112] peripheral cells, model uncertainties and
observer (ESO)
time-variant parameters.
Lumped parameter, Single-
Model can consider the effect of fast
Surya and Mn state thermal model + KF +
discharge, validated using MATLAB
[50] Recursive Least Square
simulation
(RLS) algorithm
EECM-based virtual thermal Experimentally validated model can
Xiao Y. [113]
sensors (VTS) + KF capture environmental uncertainties.
Ma et al. [114] ROM of a LIB pack for a
Experimentally validated model, can
and Ismail et central temperature of LIB pack
capture variation of temperature and SOC.
al. [115] + Recursive least square (RLS)

2.7.2.4 Numerical Analysis-Based Temperature Estimation

Numerical methods have proven effective for estimating the temperature of various
LIB chemistries and configurations, including LIB packs. Approaches such as the FEM
[119], [120], [121], [122] and the finite volume method (FVM) [123] are widely employed
to model the thermal dynamics within batteries using nonlinear partial differential
equations (PDEs). For instance, Du et al. [121] applied FEM analysis with a three-
dimensional model and the Bernardi equation to estimate the internal heat generation rate.
These PDEs, while providing detailed insights, involve intricate boundary conditions, often
requiring infinite-dimensional modeling. The governing equation used by Du et al. [121]
is as in Equation (2.5).

𝜕𝑇 𝜕2𝑇 𝛿2 𝑇 𝛿2 𝑇
𝜌𝐶𝑝 𝜕𝑡 = 𝜆𝑥 𝜕𝑥 2 + 𝜆𝑦 𝛿𝑦 2 + 𝜆𝑧 𝛿𝑧 2 + 𝑄 (2.5)

41
 Here, Q represents the total heat loss within the cell, 𝜌, and 𝐶𝑝 denote the mean density
and mean specific heat of the cell, respectively, and 𝜆 is the thermal conductivity
coefficient of the cell surface material.

Dong Hyup Jeon [119] integrated a transient thermoelectric model with a porous
electrode model to numerically simulate the thermal behavior of a commercial LIB during
charging and discharging. The study observed that temperature increases more
significantly during discharging than charging and found that higher C-rates reduce the
temperature disparity between charging and discharging cycles. Similarly, Baba et al. [120]
employed an enhanced single-particle model for three-dimensional numerical simulations
to investigate the temperature distribution within a LIB cell. Yi et al. [122] used numerical
methods to analyze the transient thermal behavior of LIBs under dynamic driving cycles.
Their approach incorporated a double-layer thermal capacitance model to capture short-
term transients effectively. Fleckenstein et al. [123] applied the FVM to demonstrate the
influence of temperature gradients within cell layers, revealing how these gradients create
varying current densities and localized SOC inhomogeneities. Their findings emphasize
the importance of accounting for these factors in the design of TMS.

Although numerical methods provide highly accurate and detailed insights into cell
temperature gradients, their computational intensity makes them unsuitable for online
temperature estimation. Furthermore, the complexity of mathematical analysis requires
specialized expertise, and generalizing results across different chemistries and cell
configurations presents additional challenges. A summary of these numerical method-
based temperature estimation strategies is provided in Table 2.5.

Table 2.5: Summary of numerical methods-based temperature estimation strategies

Reference Types of Models Important Note

A transient thermoelectric The numerical models are developed
Dong Hyup model with a porous using COMSOL Multiphysics and tested
Jeon [119] electrode model + finite the performance using difference drive
element method (FEM) cycles.
Baba et al. Enhanced single-particle 3-D model is created using numerical
[120] model + FEM simulation to capture the internal
42
 temperature distribution inside the cell.
The variation of cell geometry and
current profile is considered as well.
COMSOL Multiphysics is used for
Three-dimensional model
Du et al. modeling. The impact of changes in
+ ECM based heat
[121] current profiles and temperature variation
generation model + FEM
are captured in modeling.
Experimentally validated model can
Transient thermoelectric
Yi et al. [122] consider the transient behaviors under
model + FEM
dynamic driving cycle condition.
The experimentally validated model can
Fleckenstein Three-dimensional model consider the different current density and
et al. [123] + FVM local SOC inhomogeneities at different
cell layers.

2.7.2.5 Direct Impedance Measurement-based Temperature Estimation

Estimating the internal temperature of LIB cells using lumped-parameter and
approximate distributed thermal models presents several challenges. Accurately
determining thermal model parameters, such as heat generation and cell thermal properties,
is complicated by the layered construction of cells and unknown thermal contact
resistances. Traditional methods estimate heat generation based on operating current,
voltage, and internal resistance; however, these approaches struggle to detect rapid internal
temperature changes. This limitation arises because heat conduction from the cell core to
the surface is a time-intensive process [124]. Embedding micro-temperature sensors in
large LIB packs is also impractical due to manufacturing complexities and high
implementation costs [125], [126].

Srinivasan et al. [127], [128] introduced an electrochemical impedance-based method

for internal temperature estimation. This technique leverages the sensitivity of
electrochemical impedance in the 40 to 100 Hz frequency range to temperature changes,
offering an alternative to direct thermal measurements. However, their approach assumes
a uniform internal temperature distribution and is limited to a temperature range of −20°C
to 66°C. Building on this work, Schmidt et al. [129] examined the effects of temperature

43
non-uniformity on impedance-based temperature estimation. While both methods provide
estimates of mean cell temperature, they fail to capture significant variations between the
maximum internal temperature, surface temperature, and mean temperature an issue
particularly pronounced in cylindrical batteries subjected to high currents.

Richardson et al. [130] addressed this limitation by developing a thermal-impedance

model that combines electrochemical impedance spectroscopy (EIS) measurements at a
single frequency with surface temperature data. This approach enables more accurate
determination of internal temperature distributions, bridging the gap between surface and
internal temperature variations and providing a more detailed understanding of thermal
behavior within the battery. Figure 2.12 illustrates the conceptual framework of the thermal-
impedance model, demonstrating its utility in capturing detailed temperature distributions
within LIB cells.

Figure 2.12: Fundamental steps in direct impedance measurement-based temperature estimation

Richardson et al. [130] provides a significant advantage by not requiring detailed

knowledge of cell thermal properties, heat generation mechanisms, or thermal boundary
conditions. Instead, it relies on the thermal-impedance model, which integrates EIS
measurements at a single frequency with surface temperature data for accurate internal
44
temperature distribution estimation. However, this approach faces several challenges. The
complexity of performing real-time impedance determination for each cell and the reliance
on surface-mounted temperature sensors make it less practical for certain scenarios. While
advancements in leveraging onboard power electronics for real-time impedance spectrum
determination are promising, these methods are not yet fully explored for real-time
temperature estimation. Furthermore, interpreting impedance measurements under
superimposed DC currents remains a significant research challenge.

Richardson and Howey [131] introduced the impedance-temperature detection (ITD)

strategy, an EIS-based temperature estimation method that adapts to cell aging and
operational uncertainties. Although ITD offers valuable insights, it is often insufficient on
its own and requires integration with surface-mounted sensors to ensure accurate online
temperature estimation [130]. The authors later combined ITD with an electric-thermal
model and a Dual Ensemble Kalman Filter (DEKF) for online core temperature estimation.
This enhanced method demonstrated performance comparable to ITD combined with
surface-mounted sensors. However, the approach still relies on online impedance
determination and the availability of an accurate electric-thermal model, inheriting similar
limitations to conventional thermal modeling. Additionally, the strategy has not been
extended to pack-level temperature estimation, leaving a critical gap in its applicability.

Zhu et al. [131] proposed an impedance response matrix analysis based on EIS
measurements to estimate internal cell temperature, incorporating factors like cell
temperature, SOC, and SOH. While this approach provides valuable insights, it does not
account for temperature non-uniformity within the cell. Moreover, the reliance on
extensive experimental data and the challenges associated with selecting appropriate
frequencies and accurately determining the real and imaginary parts of impedance
significantly hinder its suitability for online applications. Debert et al. [132] introduced a
hybrid approach combining a linear parameter varying (LPV) thermal model and a
polytopic observer-based algorithm for core temperature estimation. While this method
and other EIS-based strategies aim to achieve high accuracy, they depend heavily on
precise impedance-temperature characteristics, which require extensive preliminary testing

45
[130], [133], [134], [135], [136]. These characteristics are also subject to degradation with
cell aging, leading to inaccuracies in temperature predictions due to SOH deterioration.

Despite these challenges, EIS-based temperature estimation remains a promising area

of research, particularly for its ability to enhance accuracy and adaptability in
understanding thermal behavior. Major advancements in direct impedance measurement-
based temperature estimation strategies are summarized in Table 2.6.

Table 2.6: Summary of direct impedance measurement-based strategies

Reference Types of Models Important Note

Srinivasan et EIS data is used for modeling
EIS-based impedance measurement
al. [127], [128] and validation.
Experimentally validated
Schmidt et al. model, however temperature
EIS-based impedance measurement
[129] non-uniformity is not
considered.
Experimentally validated with
Thermal-impedance model + EIS EIS data. The model is
Richardson et
measurement at single frequency + independent of cell thermal
al. [130]
surface temperature feedback properties, heat generation, and
thermal boundary conditions.
Online EIS measurement
Richardson Experimentally validated model
(impedance-temperature detection
and Howey able to model unknown
(ITD) + dual-extended Kalman filter
[137] convection coefficient.
(DEKF)
Experimentally validated model
with EIS data. The influence
Zhu et al. Impedance response matrix analysis,
temperature, SOC, and SOH of
[131] developed using EIS measurements
cell is considered on the
impedance spectrum.

2.7.2.6 Magnetic Nanoparticles-Based Temperature Estimation

Magnetic nanoparticles (MNPs) exhibit nonlinear magnetization behavior when
subjected to an alternating current (AC) magnetic field. This property enables accurate

46
temperature estimation by analyzing the ratio of the third and fifth harmonic responses of
MNPs [138], [139], [140]. Zhong et al. [141] explored the temperature sensitivity of MNPs
under a direct current (DC) magnetic field and observed a decrease in sensitivity with
increasing DC magnetic field strength. Building on these findings, Zou et al. [142]
developed an advanced magnetic nanoparticles thermometer (MNPT) to estimate the core
temperature of LIBs. Their approach involved measuring the temperature of MNPs and
identifying an optimal range of DC magnetic field strength to enhance temperature
sensitivity while minimizing measurement errors.

Despite its potential, this method faces significant challenges. The MNPT system is
bulky and expensive, which limits its practicality for widespread application. Moreover,
the feasibility of using this approach for real-time temperature prediction remains
unassessed, raising concerns about its integration into BMS.

2.7.2.7 Machine Learning-Based Temperature Estimation

The intricate electrochemical reactions within LIBs, combined with their sensitivity to
environmental uncertainties, result in substantial variations in thermodynamic behavior
from the core to the surface of the battery. Existing distributed thermal models and lumped
parameter models struggle to accurately capture the spatial and temporal dynamics,
particularly in large-capacity battery packs. Developing a single physics-based model
capable of effectively representing these complex behaviors poses significant challenges.
To address these limitations, ML algorithms have emerged as a promising solution. ML
approaches excel in capturing localized dynamics and enhancing the modeling accuracy of
nonlinear systems like LIBs. By leveraging high-quality data, ML techniques can
effectively predict temperature distributions within batteries while accommodating the
complexities of spatial and temporal variations.

A typical functional block diagram of an ML-based temperature estimation scheme is

shown in Figure 2.13. This framework highlights the ability of ML models to process a wide
array of input features, such as current, voltage, SOC, and SOH, to provide accurate and
adaptive temperature predictions under dynamic operating conditions.

47
 Figure 2.13: Schematic layout of ML-based temperature estimation scheme

ML has shown immense potential in addressing the challenges associated with thermal
modeling of LIBs. Liu and Li [143] developed a hybrid model that integrates an EECM
with a neural network (NN) learning approach. This spatiotemporal thermodynamic model
was designed to estimate the internal temperature distribution of LIBs with high accuracy.
The data-driven NN component leveraged signals commonly measured by BMS to resolve
model-plant mismatches caused by spatial nonlinearities and other uncertainties. Sbarufatti
et al. [144] further explored LIB temperature estimation using NN and support vector
machine (SVM) techniques. In a related study, Liu et al. [145] employed a hybrid model
combining radial basis function neural networks (RBFNN) with an EKF. While their
approach focused primarily on estimating SOC and SOH, it highlighted the adaptability of
hybrid models for battery parameter estimation.

However, completely ML-based strategies often face challenges in generalization due

to the variability of real-world operating conditions. To overcome these limitations, Feng
et al. [146] proposed an electrochemical-thermal-neural-network (ETNN) model. This
model fused a lumped-parameter electrochemical thermal model with a feed-forward
neural network (FFNN) and an unscented Kalman filter (UKF). The ETNN demonstrated
promise in predicting the SOT across a wide range of temperatures and under dynamic
48
current conditions. Despite its potential, the ETNN’s complexity, reliance on underlying
electrochemical models, and concerns about computational efficiency and applicability in
online scenarios remain significant hurdles. Furthermore, like pure electrochemical
models, the ETNN inherits similar challenges regarding parameter determination and
computational demands.

While ML-based approaches are computationally efficient and adaptable, challenges

persist in collecting high-quality data for training and ensuring model generalization. The
need for extensive experimental validation in real-world battery testing environments
remains a critical gap. Addressing these challenges through further research is essential to
fully realize the potential of ML in LIB temperature estimation. A summary of ML-based
techniques for LIB temperature estimation is provided in Table 2.7.

Table 2.7: Summary of ML-based temperature estimation techniques

Reference Types of Models Important Note

Extensive battery test data is used for model
EECM + neural network
Liu and Li training and validation. Model-plant mismatch
(NN)-based learning
[143] caused by spatial nonlinearity and other model
approach
uncertainties are taken into consideration.
Battery test data is used for training and
validating NN and SVM-models. The
Sbarufatti Neural networks + Support
uncertainties of temperature,
et al. [144] vector machines
charging/discharging current variation are
considered in the modeling.
RBF neural network The impact of the changes in cell thermal
Liu et al.
(RBNN) and the extended behavior is considered and the model is
[145]
Kalman filter (EKF) validated using synthetic data.
Electrochemical-thermal-
Wide temperature and large current variations
Feng et al. neural-network (ETNN) +
are considered in modeling. Used
[146] Unscented Kalman filter
experimental data for training and validation.
(UKF).

49
2.7.2.8 Physics Informed Neural Network for Temperature Estimation
Physics-informed neural networks (PINNs) offer a promising hybrid approach by
combining the strengths of traditional physics-based models and data-driven machine
learning methods. PINNs leverage the insights of electrochemical and electrical models
with the predictive power of machine learning, addressing key challenges in battery
temperature estimation [147] [148], [149]. A PINN is a neural network trained for
supervised learning tasks while adhering to the underlying physical laws, which are
incorporated into its loss function through PDEs and automatic differentiation [147] [150].
Recent studies have demonstrated the effectiveness of PINNs in estimating battery states.
For example, Wang et al. proposed a PINN framework using LSTM networks to estimate
the surface temperature of battery cells [149]. Recurrent neural networks (RNNs),
particularly gated recurrent unit (GRU) networks, have also been widely employed for
time-series predictions due to their ability to capture dependencies in sequential data [151].
However, purely learning-based methods, including RNNs, often suffer from limitations
such as a lack of transparency, dependence on large datasets to avoid overfitting, and
degraded performance when encountering unseen features or operating conditions.
Moreover, these methods rarely account for aging-related factors, which are critical for
long-term temperature predictions as batteries generate more heat during later life cycles
[152]. PINNs overcome these challenges by embedding the residuals of physical equations
into the loss function, ensuring adherence to physical laws and improving interpretability
and reliability [153]. This makes PINNs particularly well-suited for solving inverse
problems involving uncertain parameters [154]. A typical architecture of PINN-based cell
temperature estimation techniques is shown in Figure 2.14.

50
 Figure 2.14: Structure of Physics Informed Neural Network for Cell Temperature Estimation

2.7.2.8 Deep-learning-based Core Temperature Estimation

The performance, safety, and lifespan of LIBs are strongly influenced by their operating
temperature [3], [4]. Operating outside the manufacturer-recommended safe temperature
range accelerates battery degradation and raises safety risks, including thermal runaway
and performance deterioration [6], [7]. These concerns are particularly pronounced in EVs,
where battery packs comprise hundreds or thousands of densely packed cells. Such
configurations often experience thermal imbalances caused by inhomogeneous heating or
cooling, especially during fast charging, discharging, or dynamic operating conditions.
Cylindrical LIB cells, for instance, frequently exhibit temperature differences of up to 10°C
between the core and surface [9]. Monitoring the core temperature of individual cells is
therefore critical for efficient TMS control, extending battery life and ensuring safety under
the demanding conditions of EV applications.

Given the limitations of traditional methods, deep learning has emerged as a promising
alternative for core temperature estimation by leveraging the spatiotemporal dynamics
inherent in battery systems [155]. Deep learning networks, such as LSTM networks and
GRUs, have demonstrated their suitability for time-series predictions, making them highly

51
effective for LIB thermal estimation [156], [157]. For example, Kleiner et al. [158]
employed a Nonlinear AutoRegressive model with eXogenous inputs (NARX) for internal
temperature estimation, while Wei et al. [159] utilized a Backpropagation Neural Network
(BPNN). Similarly, Wang et al. [160] applied LSTM networks, and Surya et al. [9]
implemented a 2D Grid LSTM for predicting core temperatures. These approaches rely on
key inputs, such as battery current, voltage, SOC, and heat loss data, enabling the networks
to effectively capture temporal dependencies. RNNs, including GRUs [161], have further
mitigated issues like vanishing and exploding gradients, enhancing the robustness of time-
series predictions.

Despite significant advancements, challenges remain. Many deep learning models

depend on surface temperature measurements as feedback, necessitating the installation of
physical sensors for each cell an impractical requirement for large EV battery packs.
Additionally, the use of heat generation as an input parameter presents difficulties in real-
world scenarios, where accurate heat generation data is challenging to measure or estimate
reliably due to dependence of model parameter or in appropriate measuring techniques in
addition to the effect of cell aging and changing operational conditions [162].

Yuan et al. [163] proposed a hybrid framework combining numerical models with
LSTMs, leveraging EIS data for feature extraction. While this approach achieved high
accuracy, with a mean absolute error (MAE) of less than 0.23°C, it was constrained by the
computational and operational challenges associated with numerical and EIS-based
methods. Zhang et al. [164] conducted a comparative analysis of RNN, LSTM, and GRU
models, optimizing their performance using Bayesian hyperparameter tuning and K-fold
cross-validation, leading to improved core temperature estimation accuracy. However,
these methods still relied on surface temperature sensors, which limited scalability and
increased system complexity.

Artificial neural networks (ANNs) have also demonstrated potential for thermal
estimation due to their high accuracy and computational efficiency [165], [166], [167],
[168]. For instance, Hussein et al. [168] applied ANNs to predict surface temperatures
across different chemistries and aging conditions. However, their approach required
52
separate networks for specific test scenarios, limiting generalizability. Moreover, most
studies focused on surface temperature estimation, overlooking critical thermal dynamics
within the battery core. While innovations such as introducing radiation terms in lumped
thermal models [111] and adaptive observers for parameter identification [78] have
improved modeling, they still depend on equivalent thermal models, which fail to account
for the nonlinear thermal behavior of LIBs. Liu et al. [169] proposed a dual-extended
Kalman filter combined with neural networks, but challenges such as reliance on equivalent
models and the complexity of parameter characterization persisted. Deep learning-based
strategies hold significant potential for advancing core temperature estimation, but further
research is needed to address challenges related to implementation of its real-time
estimation and prediction in addition to scalability, generalizability, and real-world
applicability.

2.7.3 Limitations of Conventional Temperature Monitoring Techniques

Traditionally, temperature monitoring in LIBs relies on surface temperature

measurements using sensors like thermistors or thermocouples. While these sensors are
effective for surface-level monitoring, equipping every cell in an EV battery pack with
such sensors is cost-prohibitive and significantly increases hardware complexity. For
example, the Chevy Volt employs only 16 sensors to monitor 288 cells, while the Ford C-
Max Hybrid uses 10 sensors for 76 cells [14]. This sparse distribution of sensors results in
limited coverage, often leaving large portions of the battery pack unmonitored.
Surface temperature measurements alone are insufficient for accurately determining
the core temperature of cells due to heat transfer delays between the core and surface and
the presence of significant thermal gradients. This limitation is especially critical in
cylindrical cells with high Biot numbers (Bi > 0.1), where thermal gradients are more
pronounced [9], [15], [16]. Such discrepancies between surface and core temperatures can
compromise the TMS, leading to suboptimal battery performance and safety risks.

High-fidelity thermal models [16], [170] are capable of accurately predicting

temperature distributions across the entire cell, including the core. These models, however,
are computationally intensive, requiring significant processing power, which renders them

53
unsuitable for real-time onboard BMS. Reduced-order models [171], [172] and lumped
thermal models [103], [170] provide computational efficiency and are often used in BMS
applications. However, these simplified models struggle to capture the intricate thermal
dynamics of LIBs, especially under challenging conditions such as high C-rate operations,
battery aging, and fluctuating ambient temperatures.

Numerical methods, such as FEA [121], [173], [174], are well-established for predicting
temperature distributions within LIBs. While highly accurate, these methods are
computationally expensive and impractical for real-time applications. EIS-based methods
[129], [127], [175] are hindered by high costs and the complexity of integrating EIS setups
into onboard BMS environments. Moreover, current research findings are only limited to
cell level analysis. Suitability of it for module and pack level analysis has not been
highlighted in literature.

Hybrid approaches combining lumped parameter models with data-driven techniques

[9], [116], [176] have been proposed to address some of these limitations. While these
methods improve computational efficiency and predictive capabilities, they remain
dependent on model parameters that evolve with battery aging and changing operational
conditions, reducing their long-term reliability and scalability for real-world applications.
Additional limitations stem from challenges in real-time implementation. Many traditional
and hybrid techniques rely on predefined parameters or experimental data that fail to
account for dynamic variations in battery behavior. For example, heat generation, thermal
conductivity, and other properties are influenced by SOC, SOH, and ambient conditions,
which evolve over the battery’s lifespan. Without continuous recalibration, these methods
can produce inaccurate results, undermining the effectiveness of the BMS.

In summary, while existing methods provide valuable insights into battery thermal
behavior, their limitations in scalability, computational demands, and adaptability to
dynamic conditions underscore the need for advanced, adaptive, and efficient temperature
estimation strategies to meet the growing demands of modern BMS applications.
Moreover, the practically of exiting techniques in real-time core temperature estimation
and prediction has not been explored.
54
2.8 Real-Time Core Temperature Estimation

Most existing core temperature estimation methods are validated using offline
experimental or simulated data, limiting their applicability to real-world EVs scenarios
[158], [177], [178]. While microsensor-based approaches [177], [179] and impedance-
based techniques have been explored, their practicality for large-scale EV deployments is
hindered by challenges such as sensor placement constraints, high implementation costs,
and long-term degradation issues. Deep learning-based temperature estimation provides
significant advantages over traditional and model-based methods by addressing nonlinear
thermal dynamics and adapting to evolving battery conditions. By capturing temporal
dependencies and leveraging BMS data, deep learning-based techniques enable precise
predictions of core temperatures even under varying loads and ambient conditions.
Moreover, the integration of cloud-based frameworks ensures enhanced computational
efficiency, allowing for seamless real-time processing and scalability across entire battery
modules. A summary of recently reported development of deep learning-based core
temperature estimation techniques is presented in Table 2.8. However, none of these
developments demonstrate the real-time estimation of core temperature, all authors used
offline data for training and validation.

To address these limitations, a real-time core temperature estimation strategy for cell
and battery module essential. Moreover, the approach should only require readily
accessible BMS data, including stack current, individual cell voltages, and sampled surface
temperature measurements enabling seamless integration with the existing on-board BMS
technologies. Integration of advanced deep learning technique in conjunction with cloud-
based frameworks is also essential to overcome the computational and scalability
challenges of traditional techniques. The proposed strategy should offer an efficient,
accurate, and scalable solution for monitoring core temperature in dynamic EV operating
conditions.

Despite these advancements, challenges remain in achieving universally reliable and

scalable core temperature estimation across diverse EV configurations and use cases.

55
Factors such as battery aging, fluctuating environmental conditions, and the need for robust
system integration must be addressed to ensure consistent performance.

Table 2.8: Temperature Estimation Models and Performance

Reference Model Model Input Estimated Performance
Parameters
Lin et al. [14] H∞ model I, inlet coolant Ts, Tcore Worst case ≤
temperature 3.1°C
Richardson et al. EKF-based I, V Temperature RMSE < 0.7°C
[173] estimator distribution
Chen et al. [176] EKF-based I, V Ts, Tcore /Ts ± ≤ 0.5°C;
estimator Tcore ± 1.5°C
Schmidth et al. Impedance- EIS- Bulk With SOC ±
[129] based Measurement, temperature 0.17°C; without
SOC SOC ± 2.5°C
Wang et al. LSTM with I, Ts, Tamb Tcore RMSE <
[160] transfer learning 0.3302°C; MAE
0.3302°C
Zheng et al. LSTM Impedance, I, V Volume-average RMSE 0.46°C
[175] temperature
Surya et al. [9] 2D grid LSTM I, V, Q Ts, Tcore RMSE 0.81°C
Xu et al. [155] Nonlinear I, V Temperature RMSE 0.1439°C
Spatiotemporal distribution
Modeling
Kleiner et al. ECM-NARX- SOCinit, I Tcore MSE < 0.5°C
[158] network
Wei et al. [159] LTNN-UKF I, V Temperature RMSE
distribution 0.4266°C; MAE
0.3180°C
Zhang et al. LSTM I, V, Ts Tcore RMSE ≤
[164] 0.171°C; MAE ≤
0.148°C
Zhang et al. GRU I, V, Ts Tcore RMSE ≤
[164] 0.131°C; MAE ≤
0.099°C
Chin et al. [162] Electro-Thermal I, Tamb Ts, Tcore Ts MSE ≤ 3°C
State-Space
Model

2.9 Cloud and Digital-twin enhanced Battery Management Systems

Onboard BMS in the industry are primarily designed to monitor real-time parameters
of battery packs, such as voltage, temperature, and current. While modern BMSs have
advanced capabilities to estimate battery states with high accuracy, they are often limited

56
by fixed model parameters. This limitation reduces their ability to maintain state estimation
accuracy over the entire lifecycle of the battery. Additionally, most onboard BMSs lack
predictive capabilities, which are crucial for forecasting the SOH, RUL, thermal runaway
events, safety prognosis, and enabling predictive maintenance [180].

Accurate state estimation and the ability to forecast safety failures require processing
vast amounts of data during vehicle operation [181]. However, onboard BMSs are
constrained in their capacity to store and process large datasets due to the computational
burden on onboard processors. Handling such large volumes of data can degrade
processing speed, undermining the primary functionality of BMS. Integrating cutting-edge
technologies such as big data analytics and AI into BMS architecture presents a promising
solution for achieving effective battery lifecycle data management [182].

Digital twin (DT) technology offers significant potential in addressing these challenges.
A DT creates a dynamic, interactive link between a physical object and its virtual
representation, enabling real-time synchronization and interaction through defined
interfaces. Initially developed in the aerospace industry for calculating RUL and managing
the health of aircraft systems [183], DT technology has proven its ability to predict and
diagnose faults in advance. This capability is critical for preventing catastrophic failures
and ensuring system reliability.

The LIB system, characterized by its inherent nonlinearity and complexity, presents
substantial challenges in achieving accurate state estimation, fast charging, effective
thermal management, and extended useful life. These challenges provide an opportunity to
leverage DT technology to address the intricacies of battery systems. By creating a
comprehensive DT framework for batteries, researchers can unlock new possibilities for
improved accuracy, fault diagnostics, and optimized management of LIBs throughout their
lifecycle. Furthermore, its prediction capabilities ahead of time and real-time update
features can be leveraged for predictive control, management and safety of battery system.

2.10 Introduction to Digital Twin Technology

The concept of the DT was first introduced by NASA in 2012 [184]. According to
NASA, a DT is a virtual simulation process or a digital representation of a physical system,

57
created using experimental data, physical measurements, scales, and probabilistic models.
A well-developed DT emulates the exact behavior of its physical counterpart, enabling
dynamic and adaptive control. Michael Grieves, a key pioneer of DT technology, along
with Singh et al., later defined DT as a dynamic virtual representation of a process or
system that leverages real-time data for system monitoring and control [185], [186].
Expanding on this, Semeraro et al. [187] described DTs as adaptive models designed to
replicate the performance of physical systems within a digital environment, facilitating
real-time control and failure prevention. The primary goal of DT technology is to provide
real-time feedback to optimize the physical system's behavior while simultaneously
predicting its future states to prevent failures and enhance safety [188].

A distinction should be made between DTs and related concepts like digital models and
digital shadows. In a digital shadow, data flows in a unidirectional manner from the
physical object to its digital counterpart. Conversely, a DT enables bidirectional
communication, where changes in the physical system trigger automatic updates in the
digital representation, and vice versa [189]. This seamless exchange of information allows
DTs to dynamically replicate and adapt to the behavior of physical systems, making them
indispensable across various industries. The different types of digital representations and
their respective data flows are illustrated in Figure 2.15.

Figure 2.15: Three different kinds of digital representation and the corresponding data flow

DTs have been successfully deployed in a range of industries, including transportation

[190], [191], [192], [193], operations [194], [195], manufacturing [196], [197], [198],
maritime [199], aerospace [200], [201], [202], and battery management [203], [205].
58
Current BMS technologies often fall short of providing optimal safety measures and
advanced fault diagnosis and prognosis capabilities through continuous system monitoring.
This limitation highlights the need for advanced digitalization techniques such as digital
twinning, remote monitoring, and real-time control. By integrating DT technology, BMS
can be transformed to deliver continuous battery health monitoring, predictive fault
diagnosis, and dynamic control, addressing the critical challenges faced by traditional
systems. This section critically reviews state-of-the-art DT technologies implemented in
BMS, focusing on their potential to meeting the modern requirements of safety, efficiency,
and reliability in automotive BMS.
2.10.1 General Architecture of Digital Twin Technologies
DT technology offers a comprehensive framework for creating virtual replicas of
physical systems, enabling real-time monitoring, predictive maintenance, and performance
optimization. A conceptual overview of the basic functionalities of DT is shown in Figure
2.16.

Figure 2.16: Basic functionalities of digital twin of a physical system

The six-layer DT architecture, as illustrated in Figure 2.17 and proposed by Redelinghuys

et al. [204], is particularly effective for integrating physical devices with their digital
counterparts in BMS for EVs. This architecture extends the 5-layer architecture introduced
by Lee et al. [205], adding an extra layer of intelligence to enhance real-time system control
and predictive capabilities.

59
 • Layer 1 and Layer 2: These layers represent the physical components, including
sensors and controllers, tasked with collecting critical data on battery parameters
such as voltage, temperature, and current.

• Layer 3: This layer facilitates seamless communication between the physical and
virtual layers, ensuring accurate and real-time data flow.

• Layer 4: Serving as the gateway to cloud services, this layer manages Internet of
Things (IoT) devices and securely transfers data for further processing.

• Layer 5: This layer stores historical data in cloud-based databases, forming a

foundation for advanced analysis and simulations.

• Layer 6: The final layer incorporates AI and ML to enable real-time analysis,

predictive modeling, and optimization of battery performance. It provides the
intelligence required to process historical and real-time data, ensuring dynamic and
adaptive system performance.

This layered architecture offers a robust, scalable, and efficient means of integrating
physical systems with their digital twins, providing comprehensive insights and predictive
capabilities for enhanced system management.

Figure 2.17:Six-layer digital twin architecture

60
 Cloud computing, IoT, and ML are essential enablers of DT technology for EV BMS
as discussed below.

• Cloud Computing: Cloud infrastructure supports the storage and processing of the
massive datasets generated by EV batteries. It offers scalability, flexibility, and
cost-efficiency through services such as Infrastructure-as-a-Service (IaaS) and
Platform-as-a-Service (PaaS), enabling advanced computational capabilities.

• IoT: IoT bridges physical and digital systems by using sensors to collect real-time
data, including cell temperature, voltage, and current. It enables remote monitoring
and control, providing a continuous data stream that ensures real-time
synchronization of the DT with its physical counterpart.

• Machine Learning: ML enhances the DT’s capabilities by analyzing historical and

real-time data to predict battery aging, optimize thermal management, and diagnose
faults. Advanced ML models can detect patterns, identify anomalies, and forecast
thermal behaviors, thereby ensuring safer and more efficient battery operations.

By combining these technologies, DT-based BMS solutions significantly enhance the

reliability, safety, and efficiency of EVs. They address the limitations of traditional
systems, enabling smart, data-driven advancements in battery management. This
innovative approach lays the groundwork for the future of intelligent and adaptive BMS
technologies.

2.10.2 State-of-the-art application of DT in BMS for E-Transportation

DT technology has emerged as a transformative solution in BMS for e-transportation,
providing a robust platform for real-time monitoring, fault diagnosis, and performance
optimization. Built upon its success in industries such as manufacturing and smart cities,
DT has become an integral part of the automotive industry, supporting applications like
intelligent driver assistance, navigation systems, power electronics, consumer-centered
development, health monitoring, and advanced BMS [188]. Its capability to emulate
physical systems with real-time feedback and predictive analytics has made it a critical tool
for addressing the complexities of managing LIBs in EVs.

61
 The integration of DT technology with advanced enablers such as cloud computing,
IoT, and ML has significantly expanded its applications in BMS. For example, cloud-based
frameworks like the CHAIN architecture [206] facilitate seamless BMS by enabling real-
time data sharing and optimization across the design, manufacturing, and operational stages
as shown in Fig. 2.18. These frameworks utilize cloud services to store and process extensive
datasets, enabling accurate SOC and SOH estimations through ECMs [42]. Adaptive
extended H-infinity filters and particle swarm optimization algorithms have been employed
within these systems to monitor battery performance effectively, with experimental
validation demonstrating their practicality in real-world applications [42], [207].

Fig. 2.18: A cloud of things framework [207].

62
 Figure 2.19: Cyber-physical elements of a battery digital twin

DT technology also addresses the computational and storage constraints of onboard

BMS by utilizing off-board systems for real-time data processing. Advanced architectures,
such as neural network-based models [208], have enabled cloud-assisted operations like
battery equalization, predictive maintenance, and enhanced SOC and SOH estimation
throughout the battery lifecycle. Frameworks proposed by Tang et al. [209] and Wu et al.
[210] integrate cyber-physical elements into DT-based BMS, optimizing thermal
management, charging strategies, and the overall lifespan of battery cells. These systems
aggregate data from diverse operational conditions to train robust machine learning models
for real-time predictions and vehicle-level optimization. The cyber-physical elements of a
battery DT is shown in Figure 2.19.

Recent advancements highlight the use of DT for improving LIB design and operation
under high-performance conditions. Zhang et al. [211] utilized a coupled electrochemical-
thermal DT model to optimize ultra-high-power LiFePO4/graphite LIBs. Their approach
enabled parameter refinement and performance prediction under high C-rate discharging,
facilitating efficient design iterations before construction. Similarly, DT-based smart
charging systems [212] utilize parallel intelligence to manage energy, monitor storage

63
systems, and optimize maintenance, significantly enhancing the efficiency and safety of
EV charging infrastructure.

The adoption of DT technology in BMS for e-transportation is revolutionizing BMS by

integrating real-time monitoring, predictive analytics, and system optimization. By
combining cloud computing, IoT, and ML, DT frameworks effectively address the
complexities of LIB management in EVs. These innovations lead to safer, more efficient,
and cost-effective solutions, while also extending battery life and improving overall system
performance. As the field evolves, DT frameworks are set to redefine the standards for
BMS, ensuring reliability and sustainability in electric transportation [213]. Key
technologies required for DT-based BMS is shown in Figure 2.20.

Figure 2.20: Key technologies for DT-based BMS

2.10.3 Importance of Digital Twin-Based Thermal Management Control Strategy

Informed by Core Temperature Estimation

Integrating DT-based TMS control strategies into LIB systems is essential to ensure
the safety, reliability, and longevity of battery packs in EVs. The nonlinear and highly
sensitive nature of LIB chemistry to operating conditions often results in significant
temperature gradients between the surface and core of cells, particularly under high
charging and discharging rates. Core temperatures can exceed surface temperatures by 10–
12°C as shown in Figure 2.21, leading to ineffective thermal management, accelerated
degradation, and increased safety risks such as cell over heating leading to thermal runaway
and catastrophic failure of LIB pack.

64
 Conventional BMS predominantly rely on limited surface temperature measurements
and fail to estimate internal core temperatures with accuracy. This shortcoming is
exacerbated in densely packed battery configurations, where thermal imbalances between
cells, compounded by a single-cell failure, can rapidly propagate and result in system-wide
failures as shown in Figure 2.22.

Figure 2.21: Surface and core temperature of 18650 LIB at 4C discharge with 25°C ambient temperature

DT-based BMS architectures offer a transformative solution by enabling real-time core

temperature estimation and prediction, cloud-integrated monitoring and control of core
temperatures without requiring impractical physical sensors inside cells. Leveraging AI
and ML algorithms, DT technology creates a virtual representation of the physical LIB
system, enabling accurate core temperature estimation and prediction even under dynamic
operating conditions. This facilitates predictive fault diagnosis and intelligent thermal
management, significantly improving system safety and operational efficiency.

By utilizing cloud-based data storage and processing, cloud enhanced BMS can off-
load the computational burden and long terms storage of onboard BMS to cloud platform
while simultaneously minimizing wiring complexity. The real-time synchronization of
physical and digital systems ensures precise thermal control, effectively mitigating the risk
of thermal runaway and extending battery lifespan. This integrated approach supports the
transition to safer and more reliable EV applications, addressing the inherent limitations of
traditional BMS.
65
 Figure 2.22: Single-cell failure to thermal runaway and propagation

In summary, DT-based TMS control strategies represent a critical advancement in LIB

technology, providing real-time insights and adaptive control mechanisms. These systems
address the challenges of thermal gradients, single-cell failures, and complex battery
configurations, paving the way for more efficient, safe, and market-ready EV battery
solutions.

66
Chapter 3. Research Methodology

3.1 Proposed Framework

A schematic layout of the DT-based BMS for EVs is shown in Figure 3.1. However, the
scope of this thesis is primarily limited to the following key aspects:

• Deep-learning-based real-time core temperature estimation.

• DT-based core temperature prediction three minutes ahead of time.
• Real-time feedback control implementation based on the core temperature of
individual cells to prevent overheating and thermal runaway propagation.

Figure 3.1: Schematic layout of a digital twin-based BMS for EVs

The proposed framework for a cloud and DT-enhanced BMS designed to prevent cell
overheating and thermal runaway propagation is illustrated in Figure 3.2. This framework
enables:

• Real-time core temperature estimation and prediction three minutes ahead of time.
• Identification of weak cells.
• Remote monitoring, data visualization, and long-term historical data storage for
predictive maintenance.
• Facilitating second-life applications of retired EV batteries.
The cloud-based storage system addresses the limitations of onboard BMS storage
capabilities. The primary goal of this framework is to prevent catastrophic failures, such as
thermal runaway and fires, through effective thermal management while simultaneously
ensuring maximum lifespan and reliable operation.

67
 Furthermore, the framework bridges research and development efforts with market-
ready solutions by integrating state-of-the-art DT techniques, predictive analytics, and real-
time control. Detailed methodologies for each key technology associated with the proposed
framework are discussed in the subsequent sections.

Figure 3.2: Proposed framework of the cloud and digital twin-enhanced core temperature estimation with
feedback control

3.1.1 Proposed Framework and Key Components

The framework connects physical LIB systems to their virtual counterparts through
IoT-enabled sensors, cloud computing, and AI-driven analytics. This setup ensures real-
time core temperature estimation, prediction and synchronization between physical and
cloud systems, allowing precise monitoring, prediction, and control of core temperatures.
The on-board computer (local machine) processes large datasets, facilitating advanced
analytics and predictive modeling while cloud infrastructure facilitate long term data
storage, remote visualization and access to the on-board BMS.

• Physical BMS and Sensor Network: A physical BMS and sensor network collect
critical battery parameters, including current, voltage, and surface temperature,
enabling the virtual model to estimate internal cell temperatures with high accuracy.
This enhances scalability and reduces system complexity.
• Safety and Predictive Maintenance: By leveraging DT's predictive capabilities,
the framework can identify potential safety risks, such as thermal imbalances or
overheating, and initiate preventive measures such as stopping or halting

68
 charging/discharging process. Predictive fault diagnosis and real-time feedback
mechanisms ensure that the battery operates within safe limits.

The methodology underpinning this framework integrates several key innovations and
research contributions as summarized below followed by detailed discussion.

• Comparison of First Order and Second Order Thermal Models: A detailed

comparative study of first order and second-order thermal models was conducted
for core temperature estimation. The second-order model demonstrated superior
accuracy in capturing thermal dynamics, while the first-order model offered
computational efficiency. These insights informed the limitations of considering
thermal model-based core temperature estimation techniques and the necessity of
AI and ML-based techniques for the proposed safety framework in the BMS.

• ECM+LSTM-Based Core Temperature Estimation: The ECM combined with a

LSTM network accurately estimates the core temperature of LIB cells. The ECM
captures the thermal and electrical behavior of the battery, while the LSTM models
temporal dependencies, enabling robust core temperature predictions under varying
operating conditions. This development worked as a baseline for developing more
advanced AI and ML-based techniques for core temperature estimation and
predictions.

• LSTM and KAN for Surface and Core Temperature Estimation: The LSTM
network was used for surface temperature estimation, while a Kalman Augmented
Network (KAN) provided accurate core temperature estimation. This hybrid
approach leveraged the strengths of both techniques, ensuring reliable temperature
estimation even under dynamic conditions without surface temperature
measurement using physical sensors.

• Bi-LSTM-Based Real-Time Core Temperature Estimation: Bidirectional

LSTM (Bi-LSTM) networks enabled real-time core temperature estimation by
considering both past and future data points. This approach provided enhanced
predictive capabilities, allowing the BMS to adapt dynamically to changes in

69
 operating conditions. The findings and understanding helped in developing the
baseline for developing the DT-based core temperature prediction ahead of time.

• Temperature-Based Feedback Control: A feedback control mechanism was

developed to prevent overheating by halting or stopping the charging/discharging
processes based on real-time core temperature data. This strategy significantly
improved safety and operational reliability.

• Digital Twin-Based Core Temperature Prediction (Three Minutes Ahead):

The DT framework enabled core temperature predictions up to three minutes in
advance, providing sufficient lead time for proactive safety measures. This feature
leveraged AI models to predict future thermal behavior, ensuring timely
interventions to prevent thermal runaway or other failures.

• Cloud-enhanced BMS Architecture: The proposed cloud-enhanced BMS

facilitates remote monitoring of real-time battery data including core temperature
to facilitate long-term storage and predictive maintenance. The architecture is
designed to be scalable and efficient, reducing the storage burden on onboard
systems.

3.2 Temperature Estimation Models

LIBs are available in various form factors, including prismatic, pouch [214], and
cylindrical cells. Among these, cylindrical cells are widely adopted for large-scale, high-
power applications due to their high energy density and robustness. However, cylindrical
cells such as 18650 and 21700 exhibit poorer thermal dissipation characteristics due to their
spiral-wound structure, which creates significant thermal gradients within the cell. These
gradients necessitate precise core temperature estimation for effective thermal management
and safe operation. This thesis focuses on core temperature estimation of 18650 and 21700
cylindrical LIB cells, addressing the challenges of maintaining thermal equilibrium under
dynamic operating conditions.

70
3.2.1 Equivalent Electro-thermal Model-based Core Temperature Estimation of
18650 Cell
The first step in this research involves thermal modeling and core temperature
estimation for an 18650 cylindrical LIB cell, as detailed in the experimental setup section
of this thesis. This section presents the mathematical analysis and integration of a KF with
thermal models to estimate core and surface temperatures. The objective is to understand
basic thermal modeling and preparing a baseline for hybrid thermal model and ML-based
estimation techniques for core temperature estimation.

3.2.1.1 Heat Generation Model

Thermal estimation models based on the ECM [215] have been widely utilized by
researchers to estimate the total heat generation within LIB cells. While electrochemical
modeling remains the most effective approach for capturing the nonlinearities of LIBs, it
is also the most complex to implement. High-order ECMs are often required to represent
these nonlinearities accurately; however, increasing the model order results in greater
computational cost and complexity. The key advantage of ECM lies in its ability to balance
modeling complexity and accuracy through optimization techniques such as model order
reduction [216], [217].

In this section, a 1-RC (first order) ECM is employed to quantify the total heat
generation inside the LIB. The 1-RC ECM, depicted in Figure 3.3, is chosen for its optimal
trade-off between performance, modeling simplicity, computational efficiency, and
accuracy when compared to higher-order RC models for calculating total heat generation
[218], [219].

Figure 3.3: 1-RC ECM (Thevenin’s equivalent) model of a LIB cell

In Figure 3.3, VOCV and V denote the open-circuit voltage and terminal voltage,
respectively. The steady-state DC series resistance (R0), representing the electrolyte
71
resistance during lithium-ion transport, is a key parameter of the model. The transient
response, caused by lithium-ion flow within the SEI layer and the anode electrode, is
represented by polarization resistance (R1) and capacitance (C1) [220]. These components
are active only during transient periods. The 1-RC model is selected for this study due to
its simplicity and suitability for online applications. Higher-order RC models [218], [219],
while offering greater accuracy, are less practical for real-time heat generation estimation
due to their computational demands. To address this, Bernardi et al. [55] proposed a
simplified equation for calculating heat generation in LIBs, which is well-suited for online
predictions. Compared to other methods, such as constant heat generation rates [221], curve
fitting techniques [222], and Joule’s Law-based calculations [223], the Bernardi equation
offers a practical and computationally efficient solution. The governing equation for total
heat generation (𝑄) in a LIB, as proposed by Bernardi et al. [55], is expressed in Equation
3.1.

𝑄 = 𝐼(𝑉 − 𝑉𝑂𝐶𝑉 ) (3.1)

In this equation, the parameters are functions of the charging/discharging current (𝐼),
SOC, and temperature, which are estimated using the ECM. The calculated 𝑄 serves as a
critical input for both first order and second-order thermal models, enabling accurate
temperature estimation. These models are further discussed in subsequent sections.

3.2.1.2 First Order Thermal Modelling

• Mathematical Analysis of First-order Thermal Model
In the first-order thermal model, the surface temperature is assumed to be uniform
across the cell's surface similar to other researchers. Heat transport is considered solely
along the radial direction, implying that the lateral surface temperature is identical to the
axial temperature (at the cell's two terminals), as reported in [16]. This simplification is
common in thermal modeling for LIBs to reduce computational complexity. Regarding
heat transfer, the model accounts only for heat conduction from the core to the surface of
the cell. Heat exchange between the cell's surface and the surrounding environment via
convection is excluded from this analysis. This assumption allows the model to focus on
internal thermal dynamics while maintaining computational efficiency. The first-order
thermal model is illustrated in Figure 3.4, which represents the cell as a simplified thermal
72
system. This modeling approach captures the essential thermal behavior of the cell,
providing a balance between accuracy and simplicity. The first-order model is particularly
suited for applications where rapid estimation of thermal states is required, such as in real-
time BMS. It forms the basis for subsequent enhancements, including higher-order models
that incorporate additional complexities for more precise thermal predictions.

Figure 3.4: First-order heat resistor-capacitor-based thermal model of LIB

In Figure 3.4, the thermal parameters, that is, heat capacity of the core is represented by
Cc (J/K), heat transfer resistance inside the cell by Rc (K/W) and heat transfer resistance
outside the cell is Ru (K/W), and total quantity of heat liberated concentrated from the core
is represented by Q (J). The unit of each respective quantity is mentioned in the parenthesis.
The temperature of the core, surface, and ambient is mentioned by Tc, Ts and Tamb
respectively measured in K. Core temperature at node Tc and surface temperature at node
Ts can be monitored using this model thus this type of model are also referred to as a two-
node or two-state thermal model [54], [104]. The heat resistor-capacitor model uses the
analogy between the thermal and electrical systems as discussed in the introduction section.
Thus, for mathematical analysis, the heat transfer rate is represented by electrical current
(i) and the branch currents are represented by ia, ib in the respective branch as shown in
Figure 3.4. Therefore, the governing equation of the model can be derived by applying
Kirchhoff’s current law at the Tc node. The current balance equation at node Tc reads;

𝑖 = 𝑖𝑎 + 𝑖𝑏 = 𝑄 (3.2)
Now, by rewriting Equation 3.2 in terms of thermal parameters, Equation 3.3 can be
found;

𝑑𝑇𝑐 𝑇𝑠 −𝑇𝑐 𝑇𝑎𝑚𝑏 −𝑇𝑠

𝑄 = 𝐶𝑐 + + (3.3)
𝑑𝑡 𝑅𝑢 𝑅𝑐

73
 Re-arranging Equation 3.3,
𝑑𝑇𝑐 𝑇𝑠 −𝑇𝑐 𝑇𝑎𝑚𝑏 −𝑇𝑠
𝐶𝑐 =𝑄+ + (3.4)
𝑑𝑡 𝑅𝑢 𝑅𝑐

Finally, the core temperature (Tc) can be determined by integrating Equation 3.4 over
the total heat transfer time, provided the surface temperature (Ts) and ambient temperature
(Tamb) are known. While Tamb can be easily measured using a single temperature sensor,
measuring Ts directly with physical sensors in high-power LIB packs presents significant
challenges due to accessibility and scalability issues. To address this limitation, an
alternative approach involves estimating the Ts using a temperature estimation scheme. One
such scheme, proposed in reference [224], enables the estimation of Ts from a known Tc,
offering a practical solution for scenarios where direct measurement of Ts is not feasible.

• KF for First Order Thermal Model

KF is used to estimate and predict an unknown parameter from known parameters. The
state model for a KF and the first-order model as developed in the reference [16], [225]
and in [226] respectively are considered for this study. Now, assuming the state as Tc,t,
output as Ts,t, and inputs as Q and Tamb. The state-space matrices are derived by linearizing
Equation 3.4 in the discrete domain. A linearized version of Equation 3.4 is shown in
Equation 3.5.

𝑄𝑡−1 𝑇𝑠,𝑡−1 −𝑇𝑐,𝑡−1 𝑇𝑎𝑚𝑏,𝑡 −𝑇𝑠,𝑡−1

𝑇𝑐,𝑡 − 𝑇𝑐,𝑡−1 = + + (3.5)
𝐶𝑐 𝐶𝑐 𝑅𝑐 𝐶𝑐 𝑅𝑢

As shown in reference [224], small changes in Ts can be neglected. Hence the term
𝑇𝑠,𝑡−1 can be considered as zero as in Equation 3.6.

𝑄𝑡−1 1 𝑇𝑎𝑚𝑏,𝑡−1
𝑇𝑐,𝑡 = + 𝑇𝑐,𝑡−1 (1 − )+ (3.6)
𝐶𝑐 𝐶𝑐 𝑅𝑐 𝑅𝑢 𝐶𝑐

The transfer matrices of the KF-based temperature estimation model can be found by
reducing Equation 3.6 in the form of state models as shown in Equation 3.7 through
Equation 3.9.

1
Hence, A =[ 1− ] (3.7)
Cc Rc

74
 1 1
B =[ ] (3.8)
Cc Cc Ru

C=D = 0 (3.9)
3.2.1.3 Second-Order Thermal Modelling
• Mathematical Analysis of Second-Order Thermal Model
The second-order thermal model accounts for the non-uniformity of Ts and radial heat
transport from the core to the surface via conduction. Additionally, it incorporates the heat
exchange between the cell surface and the ambient environment, which was not considered
in the first-order model. In this second-order approach, only convective heat exchange
between the cell surface and the ambient is included, providing a more detailed
representation of thermal dynamics. Beyond the thermal properties of the first-order model,
the second-order model also considers the thermal capacitance of the cell casing Cs, further
enhancing its ability to capture complex heat transfer behaviors. The equivalent circuit
representation of this second-order thermal model, based on heat resistor-capacitor
principles, is illustrated in Figure 3.5 This approach is consistent with methodologies
reported in several previous studies [49], [104], [107], [118].

Figure 3.5: Second-order equivalent circuit thermal model of LIB

The 𝑄 is estimated using the ECM-based strategy described in Section 3.2.1.1. To

develop the mathematical framework for the second-order thermal model, a heat balance
analysis is conducted for both the core and surface of the cell. The heat balance equations
for the core and surface are expressed in Equations 3.10 and 3.11, respectively [49].

𝑑𝑇𝑐 𝑇𝑠 −𝑇𝑐
𝐶𝑐 =𝑄+ (3.10)
𝑑𝑡 𝑅𝑐

𝑑𝑇𝑠 𝑇𝑎𝑚𝑏 −𝑇𝑠 𝑇𝑠 −𝑇𝑐

𝐶𝑠 = − (3.11)
𝑑𝑡 𝑅𝑢 𝑅𝑐

75
 • KF for Second-order Thermal Model
Tc could be estimated by re-arranging the coupled ordinary differential equations of the
second-order thermal model. Since the thermal model has two thermal energy storage
parameters (Cc and Cs), two governing equations are used to estimate Tc in terms of
measured Ts and Tamb.

1
A= [ 1 − ] (3.12)
Cc ( Rc + Ru )

1 1
B =[ ] (3.13)
Cc Cc ( Rc + Ru )

Ru
C =[ ] (3.14)
Rc + Ru

Rc
D =[ ] (3.15)
0 Rc + Ru

Worth noting that the Cc, Rc, and Ru in the second-order thermal model are the same as
Cp, Rin, and Rout of the first-order model respectively.
• Fundamentals of KF
A foundational understanding of the KF [227] is essential, as it serves as the
cornerstone of the temperature estimation scheme discussed in this study. The KF is a linear
quadratic estimator widely used in fields such as statistics and control engineering. It
estimates unknown states by accounting for noise and inaccuracies in measured outputs.
The general form of the KF is represented by the following equations:

𝑋𝑘 = 𝐴𝑘−1 𝑋𝑘−1 + 𝐵𝑘−1 𝑈𝑘−1 + 𝑊𝑘−1 (3.16)

𝑌𝑘 = 𝐶𝑘 𝑋𝑘 + 𝐷𝑘 𝑈𝑘 + 𝑉𝑘 (3.17)
Where Xk is the state of the system (Tc,t), Yk is the output of the system (Ts,t), Uk is the
input to the system ([Tamb,t Q]T), t presents the state of the system, and t-1 represents the
previous state of the system. The block diagram of a KF is shown in Figure 3.6. The KF
operates as a robust and efficient technique for estimating data based on input signals. It
utilizes a mathematical model of the system to predict the output. The measured output
from the actual system is then compared with the predicted output from the model, and the
76
resulting error is multiplied by the Kalman gain. This adjusted error is added to the
predicted state, producing an accurate estimated state.

Figure 3.6: Fundamental building blocks of KF-based estimation scheme.

In this thesis, the primary objective of using this heat generation model is to determine
the input parameters for the thermal model, including heat capacity and heat transfer
coefficients, through steady-state analysis and transient experiments using a nonlinear least
squares algorithm. LIB cells are tested at three different ambient temperatures, where
internal battery temperatures are elevated using standard current pulses. These pulses are
carefully controlled to remain within the permissible limits specified in the manufacturer’s
datasheet, ensuring no capacity degradation occurs during testing. A nonlinear least squares
algorithm is subsequently employed for online parameter estimation, aiding in the
development of the ECM and thermal model, as demonstrated by Surya et al. [49]. All
components of the model are implemented in MATLAB, utilizing three-dimensional
interpolated Look-Up Tables. Feature vectors such as SOC, battery current (Ibat), and
ambient temperature (Tamb) are used to enhance model accuracy. The heat generation
model, along with the first order and second-order thermal models, is developed in
MATLAB/Simulink and Simscape environments. An extensive simulation study is then
conducted to generate simulated core and surface temperature data for further analysis.
These simulation results are employed for model validation and to perform a detailed
comparison between the first order and second-order thermal models. Using the KF, Tc is
estimated under various current patterns. These patterns are selected to remain within
manufacturer-recommended limits, ensuring the battery remains uncompromised. The

77
simulation and analysis underscore the efficacy of the models, enabling accurate
temperature estimation and robust comparisons between the thermal modeling approaches.

3.2.2 Hybrid ECM and LSTM -based Core Temperature Estimation

Building upon the knowledge from section 3.2.1, a hybrid core temperature estimation
scheme integrating a second-order ECM with a 2D-Gated Long Short-Term Memory (2D-
GLSTM) neural network is developed. The hybrid ECM+LSTM-based core temperature
estimation technique is discussed in this section. To evaluate the performance of this
scheme, a comprehensive training and testing dataset is generated using a KF and an
equivalent electrothermal model (EETM)-based core and surface temperature estimation
approach, as derived from methodologies outlined in [78], [108] and discussed in the
previous section 3.2.1. The schematic layout of the proposed scheme, along with the
training and validation process, is depicted in Figure 3.7. This layout demonstrates the
seamless integration of the ECM and neural network components, highlighting the
workflow from data generation to model validation. The approach ensures robust
performance evaluation under diverse operating conditions, leveraging advanced machine
learning and modeling techniques for precise core temperature estimation.

Figure 3.7: Hybrid ECM+2D-GLSTM-based core temperature estimation technique

3.2.2.1 ECM-based Heat Generation Model

The ECM-based heat generation model enabled to capture the thermodynamic
properties of a LIB using equivalent electrical components. In this framework, thermal
78
resistance is represented by resistors, temperature by voltage and current, heat generation
by current sources, and thermal capacity by capacitors [16]. For this study, a 2RC ECM is
utilized to quantify the total Q within the cell. The core strategy involves mathematically
calculating the heat generation stemming from internal power losses. These losses are
primarily influenced by the cell’s internal resistance and the charging/discharging current
levels. By accurately modeling these factors, the ECM provides a reliable estimation of
heat generation, serving as an input to the LSTM-based core temperature estimation model.

Figure 3.8: (a) 2-RC ECM of a LIB cell, (b) Second-order EETM of LIB cell

The second-order ECM is shown in Figure 3.8(a) where R0 represents the steady-state
DC series resistance. The short transient response caused by the lithium-ion flow in the SEI
layer and the anode electrode is represented by polarization resistance (R1) and capacitance
(C1) respectively. R2 and C2 represent the long transient response caused by double-layer
capacitance in both electrodes. The mathematical expression of the terminal voltage (Vt) in
Figure 3.8(a) could be represented as Equation 3.18 where I is the charging/discharging
current from the battery cell.

𝑉𝑡 = 𝑉𝑂𝐶𝑉 − 𝐼𝑅0 − 𝐼𝑅1 − 𝐼𝑅2 (3.18)

A second-order ECM is considered here due to its optimum performance in terms of

accuracy, modeling complexity, and computational cost. The value of Q primarily depends
on the resistances and I. Again, the battery parameters are highly sensitive to SOC, I, and
the temperature of the cell. Therefore, extensive laboratory experiments are conducted for
ECM parameterization. Equation (3.19) represents the mathematical form of Q as proposed
by Forgez et al. [16] and Bernardi et al. [55].

79
 𝑑𝑉𝑂𝐶𝑉
𝑄 = 𝐼(𝑉𝑡 − 𝑉𝑂𝐶𝑉 ) + 𝐼𝑇𝑐 (3.19)
𝑑𝑇𝑐
𝑑𝑉𝑂𝐶𝑉
Here, is the entropy coefficient. 𝑉𝑡 , 𝑉𝑜𝑐𝑣 , 𝑇𝑐 represents the terminal voltage, open-
𝑑𝑇𝑐

circuit voltage, and core temperature respectively. The influence of the entropy term is
negligible as demonstrated by Dai et al. [108], thus it is not considered in this study.

3.2.2.2 Second-order EETM

Derived from [78], [108], an effective second-order EETM (Figure 3.8(b)) is adapted to
generate a large set of training and testing data for validating the proposed hybrid ECM
and 2D-GLSTM based core temperature estimation approach. The synthetic data is
primarily used for the proof of the concept. Mathematically EETM of a cell can be
expressed as Equation 3.20 and Equation 3.21.
𝑑𝑇𝑐 𝑇𝑠 −𝑇𝑐
𝐶𝑐 =𝑄+ (3.20)
𝑑𝑡 𝑅𝑖𝑛

𝑑𝑇𝑠 𝑇𝑎𝑚𝑏 −𝑇𝑠 𝑇𝑠 −𝑇𝑐

𝐶𝑠 = − (3.21)
𝑑𝑡 𝑅𝑜𝑢𝑡 𝑅𝑖𝑛

Here, resistors (𝑅𝑖𝑛 and 𝑅𝑜𝑢𝑡 (K/W)) and the capacitors (𝐶𝑐 and 𝐶𝑠 (J/K)) portray
thermal conductivity and thermal capacity respectively. 𝑇𝑐 , 𝑇𝑠 , and 𝑇𝑎𝑚𝑏 represent the core,
surface, and ambient temperature of the cell. During the data generation, KF [137] is used
with the EETM for improving model performance while estimating the Tc from known Ts
and Tamb. The discrete version of (3.19) and (3.20) as presented in (3.22) and (3.23) are
used during KF modeling.
𝑑𝑉𝑂𝐶𝑉
𝑄(𝑘) = 𝐼(𝑘)(𝑈(𝑘) − 𝑉(𝑘)) + 𝐼(𝑘) (𝑇𝑐 (𝑘) | ) (3.22)
𝑑𝑇𝑐 𝑘

𝑄(𝑘) 𝑇𝑠 (𝑘)−𝑇𝑐 (𝑘)

𝑇𝑐 (𝑘 + 1) = ∆𝑡 + (3.23)
𝐶𝑐 𝐶𝑐 𝑅𝑖𝑛

KF is widely used in data prediction and noise filtering applications. It is a robust and
simple technique to estimate data based on its input signal. In a KF, the measured output
from the actual system and predicated output from the mathematical model are compared
to determine the error. Then the error signal is multiplied with a Kalman gain and is added
to the predicted state for accurate state estimation.

80
3.2.2.3 Architecture of 2D-GLSTM Neural Network
The gated long short-term memory (GLSTM) network, introduced recently by
Kalchbrenner et al. [228], enhances the traditional LSTM architecture. LSTM, a
specialized type of RNN, is designed to handle complex sequence learning problems by
incorporating hidden internal states and recurrent connections. These features enable
LSTMs to retain memory and provide feedback, making them particularly well-suited for
analyzing sequential data such as time-series data. While traditional RNNs are effective for
sequence learning, they encounter the vanishing gradient problem when dealing with long-
term dependencies, particularly when using activation functions like the sigmoid function
[229], [230]. To overcome this limitation, Hochreiter and Schmidhuber [231] developed
the LSTM architecture, replacing standard neurons with interconnected memory blocks.
This innovation allows LSTMs to learn long-term dependencies and handle extensive time-
series data more effectively. Despite their success, simple LSTM networks face challenges
in meeting the increasing demands for accuracy and reliability. To address this, researchers
began developing deep recurrent architectures by stacking multiple LSTM layers, thereby
enhancing performance [232], [233]. However, these deeper architectures reintroduced the
vanishing gradient problem, particularly as the network depth increased. To mitigate this
issue, advanced deep LSTM architectures such as 2D-GLSTM were developed. The 2D-
GLSTM extends the traditional LSTM framework by addressing the limitations of
cascading LSTM layers, making it highly effective for handling complex, nonlinear, and
sequential data.

Core temperature estimation in LIBs is a time-series analysis task that involves

managing nonlinear dynamics and high sensitivity to operating and environmental
conditions. The 2D-GLSTM is employed in this study to leverage its superior capabilities
for analyzing time-series data and estimating states in complex nonlinear systems. As the
basic building block of 2D-GLSTM is a simple LSTM network, it is essential to first
understand the architecture and functionality of the LSTM framework before discussing
enhancements offered by 2D-GLSTM.

81
3.2.2.4 Architecture of a single LSTM layer
A single LSTM block comprises two internal recurrent states: the cell memory state
(mt) and the cell hidden state (ht), as illustrated in Figure 3.9.

Figure 3.9: Architecture of Single LSTM.

The cell memory state (mt) represents the long-term memory, while the cell hidden state
(ht) represents the short-term memory. At any time step t, the input (xt) is concatenated with
the hidden state from the previous time step (ht-1) and passed through three independent
gates: the input gate (it), the forget gate (ft), and the output gate (ot).
• Input Gate (it): Determines which values from the input sequence will update the
memory state.
• Forget Gate (ft): Decides which information to discard from the memory block
based on predefined conditions.
• Output Gate (ot): Determines the current output using the current input and the
previously acquired memory.
These gates regulate the flow of information using the sigmoid activation function (σ),
which generates outputs between 0 and 1, akin to digital logic gates. Additionally, the
concatenated input (xt and ht-1 ) is processed through a 𝑡𝑎𝑛ℎ layer, a nonlinear activation
function, to produce ct, the candidate input to the cell memory state (mt).
The LSTM block functions as a state machine, where weights and biases are adjusted
during training. The cell memory state (mt) is updated based on two inputs: the candidate
input (ct) from the input gate (it) and the memory state from the previous step (mt-1) as
determined by the forget gate (ft). Finally, the processed information in mt-1 passes through
a 𝑡𝑎𝑛ℎ layer and the output gate (ot), producing the updated hidden state (ht) and the new
output of the LSTM block. The internal processes of an LSTM block can be mathematically

82
represented as follows:
𝑐𝑡 = tanh (𝑊𝑥,𝑐 𝑥𝑡 + 𝑊ℎ,𝑐 ℎ𝑡−1 + 𝑏𝑐 ) (3.24)
𝑖𝑡 = 𝜎(𝑊𝑥,𝑖 𝑥𝑡 + 𝑊ℎ,𝑖 ℎ𝑡−1 + 𝑏𝑖 ) (3.25)
𝑓𝑡 = 𝜎(𝑊𝑥,𝑓 𝑥𝑡 + 𝑊ℎ,𝑓 ℎ𝑡−1 + 𝑏𝑓 ) (3.26)
𝑜𝑡 = 𝜎(𝑊𝑥,𝑜 𝑥𝑡 + 𝑊ℎ,𝑜 ℎ𝑡−1 + 𝑏𝑜 ) (3.27)
𝑚𝑡 = 𝑓𝑡 𝑚𝑡−1 + 𝑖𝑡 𝑐𝑡 (3.28)
ℎ𝑡 = 𝑜𝑡 tanh (𝑚𝑡 ) (3.29)

Here, W and b denote the weight matrix and bias of different layers, respectively.

3.2.2.4 2D-GLSTM Architecture

Deep recurrent architectures have shown superior performance compared to single
LSTM networks. However, constructing deep LSTM networks by merely stacking multiple
single LSTM blocks often exacerbates the vanishing gradient problem. To overcome this
limitation and further enhance performance, advanced frameworks such as the 2D-Gated
Long Short-Term Memory (2D-GLSTM) have been proposed [234], [235]. The 2D-
GLSTM extends the functionality of a single LSTM by incorporating memory
dependencies along both the depth and time dimensions. This dual-axis operation enables
the network to retain and propagate memory states not only across time but also between
adjacent LSTM blocks in the depth dimension. These interconnected memory
dependencies partially mitigate the vanishing gradient problem and contribute to improved
model performance. The fundamental architecture of the 2D-GLSTM is illustrated in
Figure 3.10. For further technical details about 2D-GLSTM, readers are referred to [228].
This architecture is particularly well-suited for capturing complex spatiotemporal
dependencies in sequential data, making it a valuable tool for advanced time-series analysis
tasks.

83
 Figure 3.10: Architecture of 2D-GLSTM.

The fundamental mechanism of the 2D-GLSTM involves processing inputs along both
the depth and time dimensions. At each layer, the depth dimension takes the input xt and
combines it with the outputs from the time dimension, producing results at the final layer.
The forward computation process of the 2D-GLSTM is mathematically represented by
Equations (3.24) to (3.29), while the basic LSTM functions, described in Equations (3.30)
to (3.34), remain consistent along the depth dimension.
In this study, the depth dimension is prioritized. Consequently, the computation begins
with the Time-LSTM, which processes the time dimension using HT, followed by the
Depth-LSTM, which processes the depth dimension using HL. Specifically:
• Time-LSTM (HT): Concatenates the last hidden states across both dimensions.
• Depth-LSTM (HL): Concatenates the last hidden state of the depth dimension with
the output hidden state of the time dimension.
Finally, the prioritized depth dimension is computed using HL, which concatenates the
last hidden state of the depth dimension with the output hidden state of the time dimension.
This hierarchical processing enables the model to effectively leverage spatiotemporal
dependencies for improved performance. The forward computational process of the 2D-
GLSTM is detailed in Equations 3.30 to Equation 3.33, where θT and θL denote all trainable
parameters for the time and depth dimensions, respectively.
𝑇 𝐿 𝑇
𝐻 𝑇 = [ℎ𝑡−1,𝑙 , ℎ𝑡,𝑙−1 ] (3.30)
𝑇 𝑇 𝑇
(ℎ𝑡,𝑙 , 𝑚𝑡,𝑙 ) = 𝑇𝑖𝑚𝑒 − 𝐿𝑆𝑇𝑀(𝐻 𝑇 , 𝑚𝑡−1,𝑙 , 𝜃𝑇 ) (3.31)
𝑇 𝐿 𝑇
𝐻 𝐿 = [ℎ𝑡,𝑙 , ℎ𝑡,𝑙−1 ] (3.32)
𝐿 𝐿 𝐿 𝐿
(ℎ𝑡,𝑙 , 𝑚𝑡,𝑙 ) = 𝐷𝑒𝑝𝑡ℎ − 𝐿𝑆𝑇𝑀(𝐻 , 𝑚𝑡,,𝑙−1 , 𝜃𝐿 ) (3.33)
84
 𝑇 𝐿
Here, ℎ𝑡,𝑙 and ℎ𝑡,𝑙 represents the hidden states of the time and depth dimension
𝑇 𝐿
respectively. Similarly, 𝑚𝑡,𝑙 and 𝑚𝑡,𝑙 represents the memory state of time and depth
dimension respectively. Finally, the output hidden state goes through a fully connected
layer at each time step yt to realize the final estimated results. The fully connected layer
can be mathematically represented as Equation 3.34.
𝑦𝑡 = 𝑊𝑓𝑐 ℎ𝑡 + 𝑏𝑓𝑐 (3.34)

Here, bfc and Wfc are the biases and weights of the fully connected final output layer
respectively. Mean squared error (MSE), as in Equation 3.35 is used in this study as the
loss function which indicates the error between the predicted and actual results after each
iteration (epoch) during the forward computation process.
1 𝑛𝑡
𝐿𝑜𝑠𝑠 = 𝑛 ∑𝑡=1(𝑦𝑡 − 𝑦𝑡∗ )2 (3.35)
𝑡

Here, 𝑦𝑡∗ and yt represent the target and predicted value respectively whereas nt is the
length of the time sequence in training data.

3.2.3 LSTM Network and Kolmogorov-Arnold Network (KAN) for Core and Surface
Temperature Estimation
The previously, discussed hybrid ECM+LSTM-based core temperature estimation
technique still relies on ECM model parameters which are subject to change with cell aging
and dynamic operating conditions and need domain specific knowledge on LIB modeling.
This in this section, a completely data driven core temperature estimation technique is
introduced which includes a self-learnable Kolmogorov-Arnold Network (KAN) and an
interconnected LSTM network architecture for estimating core and surface temperatures
of LIBs. Eliminating the need for a surface temperature sensor as feedback resulting in
reduced cost and wire-harness of BMS. The hybrid KAN and LSTM model is used for
estimating the core and surface temperature of a LIB cell respectively using the basic
operating parameters of the battery: voltage, current, and ambient temperature. The
prediction accuracy and reliability of the proposed core and surface temperature estimation
techniques are tested for a wide range of battery charging-discharging conditions and
operating temperatures (-20°C to 40°C) even with dynamic drive cycle discharging

85
conditions. The following section discusses the interconnected LSTM and the KAN
architecture in details.

3.2.3.1 Interconnected Long Short-Term Memory (LSTM)

LSTM is a popular recurrent neural network (RNN), having a feedback connection

within the model. The architecture makes LSTM well-suitable for processing time series
data. Compared to other variations of RNN-based models conventional LSTM is less
computationally intensive [236]. Moreover, as Lindemann et al. [237] demonstrated,
LSTM networks are highly efficient in detecting anomalies and learning the temporal
relationships within a time-dependent context. A single LSTM cell consists of an input
gate i, forget gate f and the output gate o. The structure of a standard LSTM cell is
illustrated in Figure 3.11. The LSTM cell utilizes an input gate i, forget gate f, and the
output gate o to obtain the desired estimated parameters based on the input given to the
model.

Figure 3.11: LSTM Architecture and cell structure

Previous studies have already presented estimation methods that use interconnected
LSTM networks to estimate two parameters that are related to each other. The internal
resistance of a battery cell is one of the most significant indicators of SOH [238]. In [238],
the authors estimated the SOH utilizing the input of temperature, current, and voltage
using LSTM. The interconnected network also employs the input of the first network and
the output, the SOH. In addition, Hu et al. [239] proposes a joined SOC and SOH estimation
method using two interconnected LSTM networks. The positive results of the two studies
86
indicate the potential of an interconnected LSTM for estimating interdependent states like LIB
temperatures. Therefore, in this thesis, an interconnected LSTM is designed to predict the TS
followed by Tcore using KAN. The first network takes voltage V, current I, and ambient
temperature Tamb as inputs, and its output is the surface temperature TS. The second network
uses the same inputs as the first and includes the surface temperature as an additional input
parameter to estimate the core temperature Tcore. Figure 3.12 illustrates the structure of the
method. Each estimation model consists of an LSTM layer followed by a feedforward neural
network (FNN) with one hidden layer and a final output layer. Battery current (It), voltage (Vt),
and ambient temperature (Tamb,t) of t time steps are used as an input vector x of the LSTM
network and as output vector y the estimated Ts and Tcore, at timestamp tn as presented in
Equation 3.36 and Equation 3.37.

Figure 3.12: Structure for estimating surface and core temperature using interconnected LSTM network

[ 𝐼𝑡𝑛 , 𝐼𝑡𝑛−1 , ⋯, 𝐼𝑡𝑛−𝑖 ,

𝑥= 𝑉𝑡𝑛 , 𝑉𝑡𝑛−1 , ⋯, 𝑉𝑡𝑛−𝑖 , (3.36)
𝑇𝑎𝑚𝑏,𝑡𝑛 , 𝑇𝑎𝑚𝑏,𝑡𝑛 , ⋯ , 𝑇𝑎𝑚𝑏,𝑡𝑛−𝑖 ]

𝑦 = [𝑇𝑠,𝑡𝑛 , 𝑇𝑐𝑜𝑟𝑒,𝑡𝑛 ] (3.37)

3.2.3.2 Kolmogorov-Arnold Network (KAN)

The KAN network is inspired by the Kolmogorov–Arnold representation theorem

[240]. In multi-layer perceptron (MLP) [241] during the learning process, the weights are
optimized, and the activation function on each node is fixed. Unlike MLP, the KAN learns
the activation function of each node individually [242]. The KAN does not have linear
weights; it has univariable spliced functions. When a neuronal network learns a high-
precision function, the model should learn to approximate the univariate functions and the
compositional structure. The KAN, as proposed by Liu et al. [242] combinations the splines
87
and MLP to achieve a high degree of robustness, accuracy, and adaptability. Figure 3.13
shows the structure of a KAN and the learnable activation functions as proposed in [242].

Figure 3.13: Structure of a Kolmogorov-Arnold Network (KAN)

In this thesis a similar architecture tailored to core and surface temperature estimation
of LIB is developed. The importance functions (Equation 3.38), and background calculus
of the KAN architecture are discussed below.

𝑓(𝑥) = 𝑓(𝑥1 , … … . , 𝑥𝑛 ) = ∑2𝑛+1 𝑛

𝑞=1 Φ𝑞 (∑𝑝=1 Φ𝑞,𝑝 (𝑥𝑝 )) (3.38)

Where, ϕq,p : [0, 1] → ℝ and Φq: ℝ → ℝ. A single KAN activation value of the (l + 1,
j) neuron is the sum of all incoming post activations, where l is the layer, j the neuron of
the (l + 1)th layer, and i the neuron of the lth layer.

𝑖𝑛
𝑥𝑙+1,𝑗 = ∑𝑖+1 Φ𝑙,𝑗,𝑖 (𝑥𝑙,𝑖 ), 𝑗 = 1, … … , 𝑛𝑙+1 (3.39)

In matrix form, it can be defined as:

𝚽𝒍,𝟏,𝟏 (∙) ⋯ 𝚽𝒍,𝟏,𝒏𝒊 (∙)

[ ⋮ ⋱ ⋮ ]
𝚽𝒍,𝒏𝒍+𝟏 ,𝟏 (∙) ⋯ 𝚽𝒍,𝒏𝒍+𝟏 ,𝒏𝒊 (∙)
𝒙𝒍+𝟏 =
𝚽𝒍
88
 Where, Φl is the function matrix of the lth KAN layer. A KAN network with L layers
and a given input vector x is defined as:

𝐾𝐴𝑁(𝑥) = (Φ𝐿−1 , … … . , Φ1 , Φ0 )𝑥 (3.40)

KAN as highly suitable for time-series data with a higher interdependency. These
studies demonstrate the network’s high interoperability and accuracy in predicting time-
series data. Furthermore, compared to similar architectures used in estimation and
prediction applications involving time-series data, such as MLP and LSTM, KAN requires
fewer parameters while maintaining higher or comparable accuracy. This work uses a
tailored made KAN model to estimate the core temperature of LIB cells and LSTM-based
surface temperature esimtation within one network. Battery current It, voltage Vt, and
ambient temperature Tamb, t of three-time steps tn, tn−1, and tn−2 are used as an input vector
x of the network and as output vector y the estimated surface temperature TS, tn and core
temperature Tcore, tn at time step tn.

[ 𝐼𝑡𝑛 , 𝐼𝑡𝑛−1 , 𝐼𝑡𝑛−2 ,

𝑥= 𝑉𝑡𝑛 , 𝑉𝑡𝑛−1 , 𝑉𝑡𝑛−2 , (3.41)
𝑇𝑎𝑚𝑏,𝑡𝑛 , 𝑇𝑎𝑚𝑏,𝑡𝑛−1 , 𝑇𝑎𝑚𝑏,𝑡𝑛−2 ]

𝑦 = [𝑇𝑠,𝑡𝑛 , 𝑇𝑐𝑜𝑟𝑒,𝑡𝑛 ] (3.42)

3.2.4 Bi-LSTM-based Real-time Core Temperature Estimation

3.2.4.1 Proposed Framework
Built upon the knowledge obtained from Section 3.2.2 and Section 3.2.3, a
Bidirectional LSTM-based real-time core temperature estimation and feedback control
strategy is developed for practical implementation of the proposed the thermal safety
framework. The proposed framework integrates multiple layers, including a physical layer,
data flow layer, deep learning layer, and cloud layer as illustrated in Figure 3.14. The role of
each layer is discussed in the following section.

89
 Figure 3.14: Schematic layout of the proposed framework.

• Physical Layer: This layer consists of a battery module integrated with embedded
temperature sensors and an automotive-grade NXP® 14-channel BMS, which is
responsible for real-time acquisition of key parameters such as individual cell voltages,
stack current, and balancing voltage differences. Additionally, the physical layer
includes a custom-developed STM32® microcontroller-based real-time temperature
data acquisition system. A local central processing unit (CPU) is also incorporated
within this layer to enable on-site data processing and core temperature estimation for
individual cells, particularly in scenarios where local computation is preferred, or cloud
access is limited.
• Cloud Layer: The cloud layer is dedicated to remote data storage, model deployment,
and real-time visualization. This layer comprises a DigitalOcean cloud server for
hosting services, InfluxDB as a time-series database to log both raw sensor data and
predicted core temperatures, and Grafana for real-time dashboard visualization.
Grafana queries data directly from InfluxDB, enabling dynamic monitoring of battery
behavior. This layer ensures scalable data management and provides stakeholders with
comprehensive visibility into system performance.
• Deep Learning Layer: This layer integrates a Bi-directional Long Short-Term
Memory (Bi-LSTM) neural network that predicts the core temperature of each cell
based on multiple sensor inputs. The Bi-LSTM model processes time-series data, learns

90
 temporal dependencies, and generates accurate core temperature estimations. These
predictions are subsequently transmitted to the cloud for logging and visualization,
enabling intelligent thermal management and advanced diagnostics.
• Data Flow Layer: The data flow layer interconnects the physical, deep learning, and
cloud layers to enable seamless end-to-end system communication. During data
acquisition, physical sensors collect real-time measurements, which are processed
locally by a Python script running on a computer or embedded system. The processed
data is then transmitted either to the local CPU for edge computing or to the cloud for
centralized analysis and visualization. This layer also manages the real-time
transmission of control signals such as charging or discharging commands ensuring
tight system integration and responsive feedback control.

3.2.4.2 Architecture of Bi-LSTM network for core temperature estimation

The architecture of the proposed Bi-LSTM network, designed for real-time core
temperature estimation in lithium-ion battery systems, is depicted in Figure 3.15. This
architecture is specifically structured to capture temporal dependencies in both forward and
backward directions, enabling the model to analyze both past and future sequences
simultaneously. Such bidirectional learning is crucial for accurately modeling the complex,
nonlinear, and dynamic thermal behavior of battery cells. The network is composed of the
following layers.

91
 Figure 3.15: Architecture of Bi-LSTM for core temperature estimation.

• Input Layer: This layer ingests four essential sensor-derived features for each cell:
voltage, current, surface temperature, and ambient temperature. These input parameters
reflect the instantaneous electrical and thermal state of the battery and serve as the
foundational data for core temperature prediction.
• Bidirectional Layer: At the heart of the architecture lies the bidirectional LSTM layer,
composed of two parallel LSTM units—one processing the sequence in the forward
direction (chronological order), and the other in the reverse direction. By merging the
outputs from both directions, this layer effectively captures holistic temporal
dependencies, enhancing the model’s ability to learn nuanced patterns from sequential
data.
• Dense Layer: The dense layer acts as a fully connected neural network layer that
receives the features extracted by the bidirectional LSTM layers. It condenses the
learned temporal patterns and maps them to a single scalar value representing the
predicted core temperature.
• Activation Layer: This layer combines the hidden states produced by the forward and
backward LSTM cells. These combined hidden states encapsulate both short-term and
long-term temporal dependencies, which are essential for ensuring continuity and
robustness in prediction across time steps.
92
• Output Layer: The final output layer generates the predicted core temperature for the
cell, based on the transformed and aggregated information from preceding layers. This
output enables real-time insight into the internal thermal state of the battery, supporting
enhanced safety, control, and predictive thermal management.

3.3 Cloud-enhanced Battery Management System

To realize the cloud-based BMS framework, a robust architecture integrating both
physical components and cloud infrastructure was developed. The proposed system
facilitates real-time data acquisition and visualization through a combination of local
hardware and remote server capabilities. The complete system-level architecture is
depicted in Figure 3.16. The physical layer consists of a lithium-ion battery (LIB) module,
an NXP® automotive-grade BMS, a PCAN adapter, an STM32® microcontroller, and a
local computer server for initial data processing and transmission. The virtual/cloud layer
comprises a remote server configured with Telegraf® for data collection, InfluxDB® for
structured time-series data storage, and Grafana® for real-time dashboard visualization.
Grafana® was selected for its high-performance, low-latency visualization features,
support for Websockets®, and open-source flexibility. It offers efficient rendering of real-
time plots and customizable dashboards, making it well-suited for monitoring battery
parameters. Similarly, InfluxDB® is utilized due to its robust compatibility with time-series
datasets, lightweight footprint, and extensibility particularly beneficial for high-frequency
sensor data logging and querying. In this setup, the NXP® BMS provides a 14-channel
analog interface, allowing it to monitor individual cell voltages across the battery module.
The STM32® board complements this by capturing analog temperature data from NTC
sensors affixed to the surface of each cell. Together, these components enable accurate
real-time measurement of key parameters such as cell voltage, pack current, and
temperature. It is important to note that while this study employs specific hardware and
software platforms for implementation, the overall architectural design is modular and
adaptable making it suitable for a wide range of battery chemistries, pack configurations,
and cloud-service providers.

93
 Figure 3.16: System level architecture of the proposed cloud-based BMS architecture.

3.3.1 Data Acquisition and Processing

In the proposed system, raw data is transmitted from both the BMS and the STM32
microcontroller as 29-bit hexadecimal messages, each associated with a unique CAN
identifier. These messages are converted from the CAN protocol to USB format using a
PCAN-to-USB interface before being forwarded to a local server for processing. A custom
Python script decodes these raw messages into a human-readable format and organizes the
output into a structured .json file. This .json data is then relayed to a remote server using a
pipeline comprising Mosquitto® (for lightweight MQTT message brokering), Telegraf®
(for data parsing and preprocessing), and InfluxDB® (a time-series database for structured
logging of sensor metrics). The CAN communication protocol utilized adheres to the 11-
bit identifier format specified in CAN 2.0A, which is the standard in most automotive
systems and illustrated in Figure 3.17.

In the CAN 2.0A frame, the start of frame (SOF) is marked by a dominant logic '0',
signaling the initiation of message transmission. The Identifier (ID) field designates
message priority, where lower values have precedence. The Remote Transmission Request
(RTR) distinguishes between data frames and remote request frames. The Control field
includes the Identifier Extension Bit (IDE) which remains dominant for 11-bit identifiers
and a 4-bit Data Length Code (DLC) that specifies the number of data bytes (0 to 8) in the
payload. The Data field carries the actual payload, containing sensor readings and CAN
signals that can be decoded into meaningful battery parameters. To ensure data integrity,
the cyclic redundancy check (CRC) validates message content, while the Acknowledgment
94
(ACK) slot confirms successful reception by a receiving node. The End of Frame (EOF)
indicates the conclusion of the transmission. Although the CAN 2.0B standard allows for
extended 29-bit identifiers and is commonly used in more complex or high-priority
systems, the present study employs the conventional CAN 2.0A format due to its wide
adoption and simplicity. Once the sensor data is decoded and formatted, Telegraf® uses its
plugin ecosystem to extract relevant metrics such as individual cell voltages, total pack
current, and cell temperatures. These metrics are timestamped and stored in InfluxDB®,
from which Grafana® dynamically queries and visualizes the data in real time. This setup
enables continuous monitoring and supports timely decision-making based on up-to-date
battery operating conditions.

Figure 3.17: Standard CAN data frame with 11 bits identifier (CAN 2.0A).

3.3.2 Data Processing and Visualization

The proposed system functions as an integrated pipeline that facilitates the acquisition,
processing, storage, and visualization of real-time battery data. Initially, a custom Python
script is employed to collect raw sensor data, which is then formatted into a structured .json
file suitable for downstream transmission. The pseudocode of this script, used for decoding
CAN messages, is provided in Figure 3.18. Once formatted, the .json data is sent to a remote
server for further processing and visualization. On the server side, Telegraf® is configured
to subscribe to specific MQTT® topics, allowing it to receive incoming sensor data in real
time. It then parses this data and forwards it to an InfluxDB® database, which serves as the
central repository for time-series data. Grafana® connects directly to InfluxDB®, providing
users with the ability to generate customized dashboards and visualizations using
InfluxQL® or Flux query languages. Through this interface, users can explore both real-
time metrics and historical trends via dynamic graphs, charts, and tabular displays, enabling
in-depth analysis and data-driven decision-making. In addition to cloud visualization, the
system also supports parallel data transmission back to the local server, enabling localized
monitoring. While real-time feedback control is not implemented in the current study, the
95
system architecture is inherently capable of supporting closed-loop BMS control in future
applications. The accuracy and reliability of the measurements are anchored by the NXP®
BMS, which is designed to meet automotive-grade precision standards. All data is
transmitted via a TCP socket, ensuring robust communication. Timestamps are recorded
using Unix nanosecond format through Influx and Telegraf, allowing precise tracking of
data events with sub-microsecond accuracy. The system supports logging frequencies
ranging from 1/60 Hz up to 100 Hz, with only minimal deviation (approximately 1.5 mHz)
observed due to processing overhead. This ensures that even at high sampling rates, the
temporal precision of the system is maintained down to the nanosecond level, making it
highly suitable for high-fidelity battery diagnostics and future real-time control
applications.

Figure 3.18: Pseudo code of Python script for CAN data decoding.

96
3.5 Digital-twin enabled Battery Thermal Management Control System
A schematic layout of the proposed digital-twin enabled BMS and TMS control
strategy is shown in Figure 3.19. The Bi-LSTM is then extended to implement digital-twin
which can predict the core temperature 3 minutes ahead of time.

Figure 3.19: A framework of a digital twin based BMS

The proposed framework leverages real-time data acquisition, predictive modeling, and
adaptive feedback to enhance the operational safety and performance of LIB. As illustrated
in Figure 3.19, the system establishes a synergistic relationship between the physical battery,
the physical BMS, the digital twin, and the cloud-enabled IoT communication interface,
facilitating intelligent and proactive thermal control.
Physical System Components: The physical layer of the system includes the battery pack
embedded in an electric vehicle and its associated thermal management system. The TMS
ensures that the battery operates within safe temperature limits by responding to control
signals received from the Physical BMS. This BMS continuously monitors essential battery
parameters such as voltage, current, and temperature, and provides this information to both
the TMS and the digital twin via an IoT communication module. The physical BMS acts
as the interface between the physical battery environment and the digital twin, playing a
crucial role in real-time data acquisition and execution of control actions.

Digital Twin and Predictive Intelligence: The digital twin component of the system
functions as a high-fidelity virtual replica of the physical BMS. It is embedded with a Bi-
LSTM neural network model, which has been trained to learn temporal dependencies and
behavioral patterns in battery data. Once deployed, the Bi-LSTM model enables the digital
97
twin to predict the core temperature of the battery up to three minutes in advance. This
forward-looking capability is essential for preemptive thermal control actions and
significantly improves system response time, reducing the risk of thermal runaway and
extending battery lifespan.
Feedback Loop and Adaptive Control: A critical feature of the proposed framework is
the feedback loop between the digital twin and the physical system. As the digital twin
receives real-time data via the IoT communication layer, it updates its internal model to
reflect the current state of the physical battery system. This feedback mechanism supports
continuous model refinement and adaptive control. Based on the predicted core
temperature and operational trends, the digital twin can issue control recommendations,
which are relayed back to the physical BMS. These control signals are then executed by
the TMS, ensuring proactive temperature regulation.
IoT Communication and Cloud Integration: The IoT communication layer forms the
backbone of the data and control exchange between the physical and digital domains. It
enables real-time synchronization of data streams and supports remote monitoring,
diagnostics, and system updates. Through this communication interface, the digital twin
not only receives live data from the physical system but also transmits predictive insights
and control decisions. Furthermore, cloud connectivity allows for scalable deployment,
over-the-air model updates, and seamless integration with advanced analytics platforms
and dashboards.
In summary, the digital-twin enabled battery thermal management control system presents
a transformative approach to managing EV batteries. By integrating Bi-LSTM-based core
temperature prediction, real-time data acquisition, and adaptive feedback control, the
framework transitions from reactive to predictive battery management. This ensures
enhanced thermal safety, improved energy efficiency, and prolonged battery life. The use
of IoT for data exchange and cloud capabilities further strengthens the system’s scalability
and intelligence, making it a forward-looking solution for next-generation electric mobility
platforms.

98
3.6 Deep Learning-based SOC Estimation Algorithms for Cloud-based BMS
Besides, estimating the core temperature estimation in local and cloud server, an
addition work on cloud-based SOC estimation studies were also performed in this thesis
contributing towards developing a DT-based BMS capable of multiple state estimation
capability. As discussed in the previous sections, the integration of cloud computing into
LIB BMS has opened new avenues for advanced state estimation, including SOC, terminal
voltage, and SOH [243], [244], [245]. Previous studies predominantly employed equivalent
circuit models (ECM) combined with adaptive filters, such as the Kalman filter, for SOC
estimation [246], [247], [248], [249]. While effective, these approaches face limitations in
leveraging the full potential of distributed computing and connectivity for real-time
applications. Yassin et al. [250] identified significant challenges in connectivity speed and
data loss in distributed computing, particularly for DTs of power systems. To address these
gaps, this study investigates the application of AI and ML techniques for SOC estimation
in a cloud-based environment. Primary intension was to understand the possible integration
of AI and ML-based state estimation in cloud platform off-loading the computational
burden of on-board BMS to the cloud-platform. A comprehensive comparative analysis is
also conducted among different deep-learning based SOC estimation methods and state-
of-the-art EKF-based SOC estimation method. The impact of data loss on estimation
accuracy, emulating real-world network interruptions, is also closely analyzed.

Tran et al. [249] first introduced the concept of a hybrid cloud-supported BMS
architecture, wherein critical safety mechanisms are implemented onboard while
computationally intensive tasks are handled in the cloud. This approach effectively
overcomes the limitations of onboard BMS systems, such as restricted data storage and
processing power. To emulate real-world conditions, this thesis employs mobile network
Long-Term Evolution (LTE) to evaluate connectivity and performance under varying
scenarios. By advancing the understanding of deep learning-based SOC estimation in cloud
environments, this work aims to enhance the scalability, reliability, and efficiency of
modern BMS. One of the primary concerns related to the usability of cloud-based BMS is
the computational cost. Therefore, in this thesis a detailed comparative analysis of GRU,

99
FNN and LSTM are conducted in terms of computational cost, error and impact of physical
distance of the cloud-server.

3.6.1 Extended Kalman filter (EKF) for SOC Estimation in Cloud

Due to the non-linear characteristics of LIB cell, KF is not the best suitable for accurate
battery state estimation. Therefore, EKF has been used frequently for SOC estimation
[251], [252] instead of the basic KF. EKF uses partial derivatives and resistor-capacitor
(RC) model expansions to linearize the battery model [36]. It relies on a set of observations
of battery voltage yk and current uk for adaptive and accurate state estimation 𝑥̂𝑘−1 [253].
Figure 3.20 shows the detailed internal working steps of EKF [18] where uk and yk are the
inputs for the prediction and correction operation stages of the filter, respectively. Hannan
et al. [36] highlighted the EKF as a method with high prediction accuracy but limited
robustness, and linearization of errors, which could occur in a highly nonlinear system.

Figure 3.20: Schematic layout of the considered extended Kalman Filter for SOC Estimation

3.6.2 Feedforward Neural Network (FNN)

In Figure 3.21 the structure of an FNN is shown. It includes an input layer, multiple
hidden layers, and an output layer [254]. Each layer is fully connected through the next
layer, so the information flows from the input nodes through the hidden layers to the output
layer. The relation between the neuron i of the current and the neuron j of the previous
layer is defined as Equation 3.43:

𝑜𝑗 = 𝜎(∑𝑁
𝑖=1(𝑤𝑗,𝑖 𝑥𝑖 + 𝑏𝑗,𝑖 )) (3.43)

100
 𝑥 𝑖𝑓 𝑥 ≥ 0
𝜎(𝑥) = 𝑥 + = max(0, 𝑥) = { (3.44)
0 𝑖𝑓 𝑥 < 0
where σ(x) describes the activation function rectified linear unit (ReLU) as in Equation
3.44, N the number of inputs, wj,i the weight of the input i at neuron j, xi the input value,
and bj,i the bias.

Figure 3.21: FNN architecture

3.6.3 Long Short-Term Memory (LSTM)

Unlike the FNN, LSTM has a feedback connection which is a derived model of basic
RNN. It is ideal for processing and predicting time series data. The cell memory unit is the
key component of a LSTM [255]. Figure 3.22 shows the structure of a LSTM model, and a
typical LSTM cell. The LSTM cell utilizes an input gate i, forget gate f, and the output gate
o. The forward pass at time t of a LSTM cell is processed as follows [255].

101
 Figure 3.22: LSTM architecture

Where xt, and ht are the cell input and output at time t, ct denotes hidden cell memory.
W and U are weight matrices and b bias vectors learned during training. it, ft, and ot
represent the gates, while σg , σc, and σh are the corresponding activation functions. LSTM
networks are especially good at detecting contextual anomalies by learning temporal
relationships and capturing them in a compact state representation. They are particularly
effective in modeling multivariate time series and time-variant systems, accommodating
both stationary and non-stationary dynamics, as well as short- and long-term dependencies
[237].

𝑓𝑡 = 𝜎𝑔 (𝑊𝑓 𝑥𝑡 + 𝑈𝑓 ℎ𝑡−1 + 𝑏𝑓 ),

𝑖𝑡 = 𝜎𝑔 (𝑊𝑖 𝑥𝑡 + 𝑈𝑖 ℎ𝑡−1 + 𝑏𝑖 ),

𝑜𝑡 = 𝜎𝑔 (𝑊𝑜 𝑥𝑡 + 𝑈𝑜 ℎ𝑡−1 + 𝑏𝑜 ),

𝑐𝑡 = 𝑓𝑡 𝑜𝑐𝑡−1 + 𝑖𝑡 𝑜𝜎𝑐 (𝑊𝑐 𝑥𝑡 + 𝑈𝑐 ℎ𝑡−1 + 𝑏𝑐 ),

ℎ𝑡 = 𝑜𝑡 𝑜𝜎ℎ (𝑐𝑡 ) (3.45)
3.6.4 Gated Recurrent Unit (GRU)
Like LSTM, GRU is a kind of RNN. Compared to LSTM, it also addresses long-term
memory problems, is cheaper in the sense of computational cost, and has comparable
accuracy [256]. The structure of a GRU cell is shown in Figure 3.23. The GRU cell utilizes

102
an update gate z and a reset gate r. The forward pass at time t of a GRU cell is processed
as follows [257].

𝑟𝑡 = 𝜎(𝑊𝑟 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑟 ),
𝑧𝑡 = 𝜎(𝑊𝑧 [ℎℎ−1 , 𝑥𝑡 ] + 𝑏𝑧 ),

ℎ̂𝑡 = 𝑡𝑎𝑛ℎ(𝑊[𝑟𝑡 𝑜ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏ℎ̂ ),

ℎ𝑡 = 𝑧𝑡 𝑜ℎ̃𝑡 + (1 − 𝑧𝑡 )𝑜ℎ𝑡−1 (3.46)

Where xt represents the input at time t, ℎ̃𝑡 , and ht are information vectors, representing
the output, Wr , Wz , and W the weight matrices of the gates and the output, and br , bz, and
𝑏ℎ̂ represent the different biases corresponding to the different weight matrices.

Figure 3.23: Structure of a GRU cell

3.7 Experimental Setup

3.7.1 Experimental Setup for Core Temperature Estimation of 18650 Cell

At first, a battery automated test system is used to conduct a series of charging/

discharging tests of an LG 18650 NMC 3.6 V, 4.0 Ah cell under controlled temperature. A
schematic layout of the entire experimental setup and a photograph of the actual
experimental setup in the laboratory are shown in Figure 3.24 and Figure 3.25 respectively. As
shown in Figure 3.25 the major equipment used in the experimental setup include a
programmable power supply (Keysight E36313A), and programmable electronic load
(B&K Precision BK8601), a data acquisition system (Chroma DPM66204), and a
103
controlled thermal chamber (Associated Environmental System). Finally, to control the
BAS a MATLAB script-based program is used. At first, the LIB cell is charged with a
standard CCCV charging profile at a 1C rate followed by a one-hour rest. Then, the cell is
discharged with a pulse discharge current at C/20 under three different Tamb; 0°C, 20°C,
and 50°C. External battery parameters, Vt, I, Tamb, Ts are recorded for model validation. The
temperature sensor LM335 of Texas Instruments® is attached at central point of the cell
surface to measure the surface temperature. Thermal paste is also used between the cell
surface and the sensor to improve the heat transfer. For simplicity, a uniform surface
temperature is considered here. The built-in temperature sensor of the thermal chamber is
used for ambient temperature measurement in real-time.

Figure 3.24: Schematic layout of 18650 cell testing setup

104
 Figure 3.25: Actual experimental setup for LG 18650 cell testing

The specifications of the considered 18650 NMC of LG Chem (manufacturer) as

provided by the manufacturer are mentioned in Table 3.1.

Table 3.1: Specifications of 18650 LIB Cell Under Test

Specification Name Values

Manufacturer and Model LG Chem/ INR18650HG2
Cell Form Factor Cylindrical (18650)
Chemistry Lithium Nickel Manganese Cobalt Oxide
(LiNiMnCoO2)
Nominal Voltage 3.6V
Nominal Capacity 3Ah
Standard Charging (CC-CV) 1.5A, 4.2V, Cut-off: 50mA
Fast Charging (CC-CV) 4.0A, 4.2V, Cut-off: 100mA
Discharging Condition 20A (Max. Current), 2.5V (Cut Off
Voltage)
Operating Temperature Charge: 0 to 50°C, Discharge: −30 to 60°C
Pack Weight 48g

105
3.7.2 Experimental Setup for 21700 Molicell Testing for KAN-LSTM-based Core and
Surface Temperature Estimation

As shown in Figure 3.27, the experimental setup collects the training, testing, and
validation data for the proposed core and surface temperature estimation techniques. The
battery testbench consists of an Ivium (Model: OctoStat5000) battery cycler, an Ivium
(Model: OctoPDA-T) analogue data collector for core and surface temperature
measurement, a Binder environment control chamber (Model MKF 240), an MCP (model:
LBN-1990) digital power supply for providing power for temperature measurement
module. A Microsoft Surface Pro 7 is used for the overall system control and centralized
monitoring using the IviumSoft platform. The specification of the considered INR-21700
cell of Molicel (manufacturer) as provided in the manufacturer’s data sheet is given in
Table 3.2.

Table 3.2: Specification of Molicel P42A

Specification Names Values

Nominal capacity 4200 mAh
Nominal Voltage 3.6 V
Maximum Voltage 4.2 V
Discharge cut-off Voltage 2.5 V
Discharge cut-off SOC 30%
Nominal discharge 4A
Continuous Discharge Current Rating 45 A
Nominal charge 2A
Maximum charge current 8.4 A

At first, a physical temperature sensor (TE Connectivity, GA10K3MCD1) is embedded

at the core of each cell by drilling the cell vertically through the negative terminal of the
cell. Before drilling, the cells were discharged to 0% SOC and cooled down to 0°C. The
width of the drill is 0.55 mm, which is just sufficient for installing the temperature sensor.
The process of cell drilling and installation of core temperature sensor is shown in Figure
3.26. Then, the adhesive is provided carefully to seal the air gap and protect the battery
chemistry from degradation. The NTC thermistors provide a measurement accuracy of ±
0.2°C at 20°C. A K-type thermocouple sensor is also installed on the surface of each cell
for surface temperature measurement. The K-type sensor provide an accuracy of ± 2°C for
106
temperatures between −25°C to 40°C. Then, a series of battery cycling is performed with
four drilled cells and one validation cell in a wide range of ambient temperatures and C-
rate as mentioned in the list below:

• Ambient Temperatures: 20°C, 10°C, 0°C, −10°C, −20°C, 30°C, 40°C

• 3× each: CC-CV (0.3C / 0.3C); CC-CV (0.5C / 0.5C); CC-CV (0.8C / 0.8C); CC-CV (1C
/ 1C)

A break of at least 3 hour was provided after changing the ambient temperature to
ensure that the cell rested under the nee temperature. The experimental setup is shown in
Figure 3.27.

Figure 3.26: Installation of core temperature sensors

Figure 3.27: Experimental setup: battery cycler, cells and data acquisition system

107
3.7.3 Experimental Setup for Samsung INR 21700 Cell and Module Level Testing and
Real-time Data Logging in Cloud

The experiment involves a 14-channel NXP BMS with CAN compatibility, capable of
reading the voltage of each of the 14 cells, as well as the pack voltage and pack current of
a LIB module consisting of 14 Samsung 21700 40T cells. A photo of the battery module is
shown in Figure 3.28.

Figure 3.28: Photograph of the 14-cell module.

Each cell has an energy storage capacity of 4.0 Ah, a rated cycle life of 600 cycles, a
recommended charging rate of 1C, and a discharge rate of 4C within an operating
temperature range of 0°C to 45°C. This experimental setup is designed as per the proposed
cloud-enhanced BMS architecture discussed in Section 3.3. Considering the importance of
monitoring individual cell temperature which is not typical in commercially available BMS
and battery packs used in automotive applications, the proposed BMS architecture includes
comprehensive individual cell temperature monitoring. The architecture incorporates an
STM32 board configured to read 14 ADCs connected to 14 NTC thermistors (103JT)
placed on the surface of the cells, enabling to monitoring of temperature gradients during
different charging and discharging cycles. The data buses of the STM32-based temperature
module and the NXP BMS are connected to the same CAN bus, and a Python script is
programmed to decode and manage all CAN messages. Additionally, the EA Battery
Cycler, which is also CAN-compatible, is connected to the same CAN bus. This integration
enables the entire system to function as a fully compensated closed-loop system, allowing
to control of CCCV charging and discharging algorithms autonomously and remotely. A

108
photograph of the experimental setup is shown in Figure 3.29. The schematic layout of the
cloud-enhanced BMS architecture is shown in Figure 3.30. Worth mentioning for accurate
and precise temperature measurement all the temperature sensors are carefully calibrated
by applying the Steinhart-Hart equation [258].

Figure 3.29: Experimental setup for 14-cell module testing and data logging.

Figure 3.30: Schematic layout the cloud-enhanced BMS.

109
 • Advantages of the Python-based CAN Data Decoder
Since the official release of the CAN protocol by the Society of Automotive Engineers
(SAE) in 1986 and its initial implementation in the Mercedes-Benz W140, the automotive
industry has widely adopted CAN communication across most vehicle platforms [259]. It
is especially prevalent in modern vehicles including EVs where it plays a crucial role in
various systems. Given this widespread use, real-time extraction and analysis of raw data
from onboard BMS have become essential for implementing digital twin technology and
cloud-enabled BMS [260]. While several tools exist for BMS data acquisition and
visualization, significant gaps remain [261], [262]. None utilize automotive-grade BMS
with integrated CAN communication protocols. Instead, most research prototypes rely on
simple microcontrollers or single-board computers like the Raspberry Pi, equipped with
built-in analog-to-digital converters (ADC) for battery data reading. These systems
typically transmit data to local computers or remote servers for storage and visualization
[263]. However, they are limited by computational power and data transmission
capabilities, making them unsuitable for automotive applications due to issues like limited
longevity, robustness, noise compatibility, and fault tolerance [264]. Existing architectures
also lack support for bidirectional data transmission, a critical requirement for digital-twin-
based BMS. As highlighted by Karnehm et al. [265], bidirectional communication between
the physical and DTs including real-time data acquisition, transmission to the DT, and
feedback to the physical twin is essential for a system to be truly DT enabled. The real-
time processing of CAN data from an onboard BMS and its visualization in automotive
applications have not been reported in the literature [19]. Therefore, the Python-based CAN
data processing tool is developed in this thesis to handle raw data from the NXP® BMS.
The .json data format is considered here as it has proven highly effective for cloud and DT-
enabled BMS applications, particularly where fast data transmission and long-term storage
of battery historical data are critical [260]. The architecture demonstrated a satisfactory
level of decoding accuracy and speed with error rates as low as 0.001% and processing
times around 10 ns. The ability of the script to handle different CAN message types and
identifiers is confirmed by successfully decoding messages simultaneously from the NXP
BMS, STM32, and EA battery cycler integrated on the same CAN bus. The NXP MCU is

110
programmed to receive data from the MC33771 BMS IC via an isolated communication
channel for the power circuitry for safety, as shown in the detailed internal architecture of
the NXP BMS shown in Figure 3.31.

Figure 3.31: Internal Architecture of NXP MCU, MC33771 BMS IC and isolated communication channels.

3.7.4 Experimental Setup for Samsung INT 21700 Cell and Module Level Testing and
Real-time Core Temperature Estimation and Logging in Local Machine and Cloud

Then previous experimental setup is extended to implement the real-time core

temperature estimation, visualization and implementing feedback control system. The
experimental setup employed for this study integrates a comprehensive architecture for
real-time data acquisition, processing, and visualization, as illustrated in Figure 3.32. This
setup includes both physical hardware components and software platforms to enable
seamless data flow and facilitate core temperature estimation using a Bi-LSTM model as
detailed below.

111
 Figure 3.32: Experimental setup with data visualization interface.

• Battery Module, NXP BMS and Sensors: The setup includes a LIB module and an
automotive grade NXP BMS. The LIB module consists of 14 cells connected in series,
forming a 14S LIB module where, Cells#1, #5, #8, #11, and #14 are equipped with both
core and surface temperature sensors, while cells#3 and #6 are equipped with surface
temperature sensors only. One additional sensor is mounted on the mechanical structure
of the module to measure the ambient temperature. All 14 temperature readings are
captured using NTC thermistors (103JT). The voltage and temperature measurement
resolutions of the architecture are 152 µV and 0.1°C, respectively, whereas the cell
balancing voltage threshold is 19 mV of the NXP BMS. It is noteworthy that all
temperature sensors are carefully calibrated using the Steinhart-Hart equation to ensure
accurate and precise temperature measurements. It is worth mentioning that, unlike the
state-of-the- art approach where the thermal chamber temperature is considered the
ambient temperature, this study measures the temperature near the battery module as
the ambient temperature. This approach is more practical and realistic for real-world
applications.

112
• STM32 Microcontroller: An STM32F446RE board is configured to read analog
values from the 14 temperature sensors and convert them to digital data using its 14
ADC channels. This real-time temperature acquisition system communicates via serial
communication. The serial data reading is synchronized with CAN-bus-based BMS
data acquisition to ensure seamless data acquisition and processing by the Bi-LSTM
model and the real-time visualization.
• Data Storage and Visualization: The experimental setup incorporates InfluxDB® for
real-time data storage and Grafana® for interactive data visualization. Grafana is
configured to generate dynamic dashboards for monitoring key battery parameters,
including individual cell voltages, stack current, measured core and surface
temperatures, and predicted core temperatures. The low- latency data reporting
capabilities of Grafana, supported by InfluxDB’s time-series database, provide an
efficient platform for real-time monitoring.
• Bi-LSTM Model Training and Integration: The Bi-LSTM model, implemented
using TensorFlow, is trained offline using historical time-series data collected through
a series of battery test under dynamic ambient and loading conditions to predict the
core temperature of individual cells in real-time. The trained model is deployed on the
experimental platform, both in the local machine and in the cloud platform, where it
processes incoming real-time data and predicts core temperatures for proactive battery
management.
• Communication and Integration Architecture: Data acquisition and processing are
managed through the Python script and a CAN-bus communication framework. Raw
sensor data from the NXP BMS is transmitted via the CAN bus, while temperature data
from the STM32 is received through serial communication. The data is de- coded using
the Python script, prepared for input into the Bi-LSTM model, and visualized in
Grafana in real- time, ensuring seamless integration and efficient system operation.
• Battery cycler and thermal chamber: The EA PSB 10360-120 battery cycler, which
is CAN-compatible, is integrated into the same CAN bus, enabling the system to
operate as a fully compensated closed-loop setup. This integration facilitates
autonomous and remote control of the battery cycler depending on the user input and/or

113
 reaching a set threshold to protect the battery. A flowchart of the overall control
architecture is illustrated in Figure 3.33.

Figure 3.33: Communication and control scheme.

The data acquisition system is designed to operate with sampling frequency ranging
from 0.1 Hz to 1 kHz and achieves nanosecond-level timestamp precision for all recorded
data, ensuring reliable tracking and analysis. In this study, a data sampling frequency of 1
Hz considered which is sufficient for temperature estimation. However, the architecture is
flexible and scalable for various battery configurations and real-world automotive
applications.

3.7.4.1 Installation of core temperature sensor and module assembling

LIB cells are typically transported at 50% SOC or nominal voltage to ensure safe
handling and maintain battery health [266]. In this study, 14 new Samsung 21700 40T LIB
cells were first unpacked, tested for voltage, and conditioned at 25°C. Each cell was
charged individually at a C/30 rate to 4.2 V (upper cutoff voltage), followed by a 1-hour
rest (cooling period). Subsequently, the cells were discharged at a C/30 rate to 2.5 V (lower
cutoff voltage) or 0% SOC with another 1-hour rest period. The charging, resting, and
discharging data, along with corresponding plots, were carefully inspected for anomalies.
Healthy cells were then evaluated for internal anomalies using electrochemical impedance

114
spectroscopy (EIS) with the BioLogic BCS-815 battery cycling system. Detailed EIS
analysis was conducted to identify any abnormalities. Once confirmed healthy, five cells
were randomly selected from the 14 cells for core temperature sensor installation.
According to Waldmann et al. [267], and Quinn et al. [268] a circular gap of exactly 1 mm
exists at the center of each Samsung 21700 cell as shown in Figure 3.34 and Figure 3.35.

Figure 3.34: Cylinder volumes used for the estimation of energy densities for 18650 and 21700 cells [268].

Figure 3.35: Evaluation of CT cross-sections of 21700 cell. (e) 2D cross-sections of CT data at mid-height
of the cylinders, (d) line profiles along the dashed lines in (e), f) digitalization of spiral in (e) [267].

Therefore, a specialized drill with a diameter of 0.8 mm, was used to carefully drill five
randomly selected cells vertically at the center through the negative terminal. Then NTC
103JT thermistors were inserted precisely into the center of each cell. After installation,
each cell was immediately sealed with high-temperature-resistant, non-reactive thermal
epoxy. The cells were then placed vertically at 25°C for 1 hour to allow the epoxy to dry
115
completely, ensuring no epoxy entered the cell core. Following the successful installation
of internal temperature sensors, EIS testing was repeated on the five cells to verify that no
internal damage occurred during drilling and sensor installation. The Nyquist plots
obtained before and after sensor installation is shown in Figure 3.36.

Figure 3.36: Comparison of Nyquist plot of a sample Cell before and after installing core temperature
sensor

Increase in cell resistance due to installation of a core temperature sensor is noticed as

2 mΩ, 9 mΩ, 8 mΩ, 12 mΩ respectively for Cell#1, Cell#2, Cell#3, Cell#4 respectively.
The observed rise in internal resistance after installing a core temperature sensor through
the negative terminal of a 21700 Samsung lithium-ion cell is likely due to a combination
of mechanical, electrical, and electrochemical factors [269]. Penetrating the cell’s negative
terminal to insert the sensor may disrupt internal components such as the current collector
or separator, leading to increased contact resistance or partial internal shorts. Additionally,
the intrusion could alter the structural integrity of the jellyroll, interfere with uniform heat
distribution, and potentially introduce contaminants or disturb the SEI layer. These
disturbances can collectively hinder ion transport and electron flow, contributing to a
measurable increase in the cell’s internal resistance. Moreover, the primary intention of
performing EIS test in this study was to identify any internal short circuit or major damages
due to internal temperature sensor installation. The Nyquist plots as shown in Figure 3.36,
demonstrate not much differences in cell EIS plots, confirming that the cells sustained no
internal damage. In addition to the internal sensors, the same thermistors were installed on
116
the surface of the five drilled cells as well as two additional cells for surface temperature
measurement. Before assembling the cells into the module mounting structure, all cells
were re-tested for terminal voltages. The cells were then assembled as a 14S module at a
uniform voltage of 2.5 V to ensure proper balancing from the beginning of the battery
cycling tests. The temperature sensor installing and module assembling steps are shown in
Figure 3.37.

Figure 3.37: Installation of core temperature sensors and module assembling

117
Chapter 4. Results and Discussion

4.1 Validation of Proposed Models and Core Temperature Estimation Techniques

4.1.1. Model validation 1-RC and 2-RC EETM for Core Temperature Estimation
This section evaluates the validation of EETM-based core temperature estimation, and
the necessity of a second-order thermal model compared to a first-order model. Both
models were developed using battery test data discussed in Section 3.7.1 and MATLAB-
based parameter estimation, followed by simulations to assess the impact of varying current
profiles on Tc, Ts. Simulations, conducted in MATLAB-Simulink with fixed solvers and
step times, initially excluded KF to establish baseline results. The current profile (Figure
4.1) and estimated Tc, Ts , and measured Tamb (Figure 4.2) confirm the validity of both the
models. The comparison between measured and estimated Ts shows consistency within
acceptable limits, and the expected relationship (Tc>Ts>Tamb) validates the accuracy the
model in capturing the thermal of LIBs under dynamic conditions.

Figure 4.1: The pattern of the discharging current applied to the cell.

118
 Figure 4.2: The plot of Tc, Ts and Tamb without using KF.

In the subsequent sections, the Tc is first estimated using the combined first-order
thermal model and KF for three distinct current profiles and ambient temperature
conditions, referred to as Case 1, Case 2, and Case 3. A similar analysis is then performed
using the second-order thermal model, with the results from both models compared to
assess their relative accuracy and performance. All experiments are conducted with varying
current profiles, adhering to manufacturer recommendations to prevent battery health
degradation [58].

4.1.1.1 Case 1: Tamb = 293 K ((20 °C)

In Case 1, the Tc is initialized to the Ts in the simulation, as the cell is assumed to be in
thermal equilibrium at the start. The Tamb is set to 293 K (20°C), and a low discharging
current is applied to observe the rise in core and surface temperatures. The current profile
used for this case is shown in Figure 4.3, while the estimated Tc and measured Ts are plotted
in Figure 4.4, and the difference between them is illustrated in Figure 4.5. The results indicate
that both Tc and Ts closely follow the current pattern, with the maximum difference
observed being 6.8 K. Additionally, it is consistently noted that Tc>Ts, with the largest
discrepancy occurring at the peak current, validating the model’s accuracy in capturing the
thermal behavior.

119
 Figure 4.3: The pattern of the discharging current applied to the cell.

Figure 4.4: The plot of the estimated Tc and measured Ts.

Figure 4.5: Variation of the difference between the estimated Tc and measured Ts.

120
4.1.1.2 Case 2: Tamb = 323 K (50 °C)
In the second phase of the experiments, the Tamb of the thermal chamber is set to 323 K
(50°C). The current profile applied in Case 2 is depicted in Figure 4.6. As in Case 1, the Tc
is initialized to the Ts in the simulation. The estimated Tc and measured Ts values are
presented in Figure 4.7, demonstrating that the temperatures closely follow the applied
current pattern. The maximum observed difference between Tc and Ts is 7 K, as illustrated
in Figure 4.8. The results align with the trends observed in Case 1, confirming that Tc
consistently exceeds Ts with the greatest discrepancy occurring at peak current levels.

Figure 4.6: The pattern of the discharging current applied to the cell.

Figure 4.7: The plot of the estimated Tc and measured Ts.

121
 Figure 4.8: Variation of the difference between the estimated Tc and measured Ts.

4.1.1.3 Case 3: Tamb =273 K (0 °C)

In Case 3, the thermal chamber temperature, Tamb is set to 273 K (0°C), with the Tc
initially set equal to the Ts. The discharging current pattern applied to the battery is depicted
in Figure 4.9, while Figure 4.10 illustrates the estimated Tc and measured Ts. The difference
between Tc and Ts is shown in Figure 4.11. As seen in Figure 4.11, both Tc and Ts exhibit
high initial magnitudes due to the large discharging current at the start of the cycle. The
temperature rise occurs more gradually, influenced by the thermal resistances (Ru and Rc)
within the system. The difference between Tc and Ts increases as the discharging current
intensifies. These results indicate that the temperature rise closely correlates with the
current through the battery. At lower current values, the rate of increase in Tc matches that
of Ts, but under higher currents, Tc rises significantly faster than Ts. This highlights the
critical need for accurate core and surface temperature estimation and the implementation
of efficient thermal management strategies to maintain Tc within safe operating limits.

122
 Figure 4.9: The pattern of the discharging current applied to the cell.

Figure 4.10: The plot of the estimated Tc and measured Ts.

Figure 4.11: Variation of the difference between the estimated Tc and measured Ts.

123
4.1.1.4 Comparison between First order and Second order Thermal Models
This section presents a comparative analysis of the first order and second-order thermal
models based on estimation accuracy, parameter identification, and experimental
requirements. Both models are evaluated under identical conditions, with the same thermal
parameters, current profiles, Ts, Tamb, and Q applied to ensure a fair comparison. Current
profiles from Case 1, Case 2, and Case 3 for the first-order model are similarly applied to
the second-order model. Heat generation is computed using the same 1-RC ECM employed
in the first-order model.

The estimated Tc profiles from both models are compared to assess prediction accuracy.
Figure 4.12 illustrates the current profile used for the comparative study, while Figure 4.13 and

Figure 4.14 display the differences in Tc and Ts obtained from the first-order and second-
order thermal models, respectively. This analysis highlights the strengths and limitations
of each model, providing valuable insights into their applicability for real-time thermal
management in LIB systems.

Figure 4.12: The pattern of the discharging current applied to both the models.

124
 Figure 4.13: Difference between Tc and Ts obtained from the second-order thermal model.

Figure 4.14: Difference between Tc and Ts obtained from the first-order thermal model.

The analysis reveals that the temperature difference is larger in the first-order thermal
model, primarily due to changes in Tc rather than Ts. This discrepancy arises from the
decoupling between Ts and Tc as highlighted in Equation 3.6. Furthermore, a comparison
of Equations 3.7 and 3.9 with Equations 3.12 through 3.15 indicates that in the KF, the
output parameter Ts does not depend on the state Tc,t-1, contributing to the estimation error
observed in the first-order model. References [225] and [16] emphasize the critical role of
Cc and Ru in the second-order thermal model, which significantly affect Tc. These

125
parameters, absent from the C and D matrices of the first-order model, lead to a noticeable
increase in Tc. Thermal parameter sensitivity analyses from references [49], [225] confirm
that the first-order model's estimation accuracy diminishes as the discharge current
increases, resulting in a growing difference between Tc and Ts. Under dynamic current
conditions, the first-order model exhibits significant deviations from the second-order
model. While Cc primarily influences the transient behavior of Tc, even minor variations in
Rc and Ru lead to substantial changes in Tc.

Despite the increased complexity of the second-order thermal model, its experimental
requirements and computational costs are not significantly higher than those of the first-
order model. Given the tradeoff between modeling complexity and accuracy, implementing
a second-order thermal model is recommended for BMS.

4.1.1.5 Comparison between First order and Second order Thermal Models for
Higher C rates
As discussed earlier, the performance of thermal models is significantly influenced by
the magnitude of the charging or discharging current. Previous analyses have already
demonstrated that the second-order thermal model is more accurate than the first-order
model. However, it is crucial to evaluate the performance of the second-order model under
high discharge current conditions, as these are commonly encountered in practical
applications. To this end, a discharge current of 5 A is applied to both the first order and
second-order thermal models to examine the resulting changes in Tc and Ts. The difference
between estimated Tc and Ts for both models is illustrated in Figure 4.15. The results clearly
indicate that the error is significantly higher in the first-order model compared to the
second-order model. This finding confirms that the second-order thermal model not only
provides enhanced accuracy under standard operating conditions but also performs reliably
in high-current scenarios. Therefore, the second-order model is well-suited for practical
applications requiring precise temperature state predictions in LIBs.

126
 Figure 4.15: Comparison between Tc - Ts for higher C discharge.

4.1.2 Validation of Hybrid ECM and LSTM-based Core Temperature Estimation

The Experimental data from the experimental setup as discussed in Section 3.7.1 is
utilized to parameterize the second-order ECM-based heat generation model and the
second-order equivalent electrothermal model (EETM)-based heat transfer model,
employing a nonlinear least squares algorithm like the methodology outlined in [270]. The
validation of the ECM is performed by comparing the terminal voltage predicted by the
model with the actual experimental values.
Figure 4.16(a) illustrates the comparison between the actual and predicted terminal
voltage, along with the absolute error (prediction error) produced by the ECM at
Tamb=293 K. Additionally, Figure 4.16(b) presents the measured Ts and the Ts estimated by
the EETM under the same ambient working temperature. These comparisons validate the
accuracy and reliability of the hybrid ECM and LSTM-based framework for core
temperature estimation in LIBs.

127
 Figure 4.16: Plot of (a) Actual and predicted Vt with prediction error (b) Measured and estimated Ts.

Table 4.1: MAE and RMSE in the Prediction of Vt and Ts Estimation

Parameter Error Tamb = 0°C Error Tamb = 20°C Error Tamb = 50°C
MAE RMSE MAE RMSE MAE RMSE
Vt in mV 0.1008 0.11 0.116 0.105 0.1066 0.114
Ts in K 0.35 0.37 0.41 0.44 0.75 0.9

In addition to Figure 4.16 (a and b), the MAE and root mean square error (RMSE) values,
as presented in Table 4.1, demonstrate that the overall trends of the estimated Vt and Ts
closely align with the experimental values, achieving reasonable accuracy [270]. The MAE
and RMSE are calculated using Equation 4.1 and Equation 4.2, respectively:

1
𝑀𝐴𝐸 = 𝑁 ∑𝑁 ∗
𝑘=1|𝑥𝑘 − 𝑥𝑘 | (4.1)

1
𝑅𝑀𝑆𝐸 = √𝑁 ∑𝑁 ∗ 2
𝑘=1(𝑥𝑘 − 𝑥𝑘 ) (4.2)

128
 These plots and statistical indices confirm the suitability of the ECM and EETM models
for generating accurate data on Q, Ts, and Tc for training and testing the proposed hybrid
core temperature estimation technique. For data generation, three different drive cycles
HWFET, US6, and UDDS are simulated at three different Tamb. The corresponding current
profiles are input into the ECM to calculate Q. Subsequently, the Q and Tamb values are fed
into the EETM to estimate Tc and Ts. The simulation studies, conducted in MATLAB, yield
nine datasets (3 drive cycles × 3 Tamb) containing Vt, I, Tc, Ts, Tamb, and Q. These datasets
form the basis for validating and optimizing the hybrid core temperature estimation
method.

4.1.2.1 Training and Validation of the Proposed Estimation Method

The dataset derived from the HWFET and US6 drive cycle profiles is used to train the
proposed 2D-GLSTM model, while data from the UDDS drive cycle, which is unknown
to the model, is reserved for validation. Figure 4.17 and Figure 4.18 depict the training
datasets for the HWFET and US6 profiles, respectively. Upon completing the training
process, the inputs Vt, I, Tc, Ts, Tamb, and Q from the UDDS drive cycle are fed into the 2D-
GLSTM model. For validation, the Tc calculated by the EETM under the UDDS drive cycle
serves as the actual Tc. The performance of the 2D-GLSTM model is then evaluated by
comparing its estimated Tc with the actual Tc values.

The TensorFlow, an open-source ML framework is utilized for implementing the 2D-

GLSTM algorithms in this study. Training and validation are conducted on the Google
Colab platform, which provides free access to a graphics processing unit (GPU), enhancing
computational efficiency. The hyperparameters of the 2D-GLSTM model, summarized in
Table 4.2, are empirically determined using a trial-and-error approach, ensuring optimal
performance for the core temperature estimation task.

129
Table 4.2: Hyperparameters of the 2D-GLSTM Model

Hyper Parameters Value

Number of epochs 100
Learning rate 0.01
Window length 200
Batch size 50
Number of layers 3
Number of deep layer (neurons/layer) 128/64/64
Dropout 30%

Figure 4.17: Training dataset (a) HWFET drive cycle I, (b) Tc of cell at different Tamb (c) Vt of cell.

Figure 4.18: Training dataset (a) US6 drive cycle I, (b) Tc of cell at different Tamb (c) Vt of cell.

The plots of actual and estimated Tc at three different Tamb values during the UDDS
drive cycle, shown in Figure 4.19, demonstrate that the hybrid core temperature estimation
scheme effectively captures the variations in Tc throughout the drive cycle. The estimated
values are in close agreement with the actual Tc. In Figure 4.19 (a), although there are
instances where the estimated Tc does not perfectly follow the variation pattern of the actual
Tc, the deviation is minimal and well within an acceptable error margin of less than 1 K.
This level of accuracy is significant for battery system safety, as minor deviations over
short durations have a negligible impact on the performance of the battery TMS.
130
 To further evaluate the accuracy of the proposed scheme, the RMSE is calculated using
Equation 4.2. The RMSE values for Tc estimation are 0.81°C, 0.60°C, and 0.52°C at
ambient temperatures of 0°C, 20°C, and 50°C, respectively. These low RMSE values
confirm a high level of agreement between the estimated and actual Tc, with errors ranging
from 0.3% to 0.9%. This validates the reliability and effectiveness of the hybrid estimation
approach.

Moreover, the validation dataset from the UDDS drive cycle was not used during model
training, meaning it was completely unknown to the DL model. The satisfactory
performance of the DL model with this unseen dataset indicates its capability to generalize
across varying battery loading and operating conditions. Additionally, its robust
performance under complex current profiles, such as the UDDS, underscores the reliability
and adaptability of the proposed hybrid core temperature estimation technique for real-
world applications.

Figure 4.19: Actual and Estimated Tc in UDDS Cycle (a) 273K, (b) 293K, (c) 323K

To further highlight the superiority of the 2D-GLSTM model in core temperature

estimation, a comparative analysis was conducted with two other similar core temperature
estimation techniques reported in literature. The estimation results of the 2D-GLSTM
model under the UDDS drive cycle were compared with the result obtained from a linear
neural network with an extended Kalman filter (LNN_EKF) [145] and a linear neural
network with a particle filter (LNN_Particle Filter) [144]. The comparison at Tamb=293 K
(20°C) is illustrated in Figure 4.20, where Figure 4.20 (a) presents the core temperature
estimation results, and Figure 4.20 (b) displays the absolute estimation error, defined as the
difference between the actual Tc and the estimated Tc.

131
 As seen in Figure 4.20, while all models accurately estimate the core temperature, the
2D-GLSTM outperforms the others in terms of accuracy. The 2D-GLSTM achieves an
RMSE of 0.54%, compared to 1.25% for the LNN_EKF and 2.67% for the LNN_Particle
Filter. These results demonstrate that the 2D-GLSTM model not only effectively estimates
core temperature under dynamic conditions across a wide operating temperature range but
also delivers higher accuracy than others in its class. This establishes the 2D-GLSTM as a
superior method for core temperature estimation in complex and variable operating
scenarios.

Figure 4.20: Performance comparison of core temperature estimation methods at 293K in UDDS (a)
Estimated Tc by different method (b) Estimation error.

132
4.1.2.2 Discussions

The validation results demonstrate that the hybrid ECM and a 2D-GLSTM neural
network model can effectively estimate core temperature across a wide range of ambient
temperatures (0°C to 50°C) and dynamic loading profiles. The RMSE for core temperature
estimation remains within 0.3% to 0.9% (0.52°C to 0.81°C), indicating its suitability for
practical applications.

In addition to external battery parameters such as current, voltage, and surface

temperature, the estimated total heat generation within the cell, derived from the ECM-
based heat generation model, is included as an input feature for the hybrid model. Since
heat generation is influenced by charging/discharging current and internal resistances,
which vary with cell aging, incorporating this parameter enables the model to automatically
account for the impact of aging and variations in loading profiles. This inclusion enhances
the deep learning model’s accuracy and adaptability to changes in physical and operational
parameters. The hybrid model’s performance is further compared with similar core
temperature estimation techniques, confirming its superior prediction accuracy. These
findings underscore the effectiveness of the proposed hybrid ECM and 2D-GLSTM
approach in handling nonlinear mappings from external measurements to core temperature,
a task that is highly challenging using traditional mathematical modeling techniques.

Additionally, this method addresses the limitations and challenges of deploying

physical sensors within battery cells. Installing sensors directly inside the cell is intrusive,
potentially damaging the cell structure and causing malfunctions. Furthermore, thermal
sensors placed within the core require insulation to protect them from chemical reactions,
which can compromise measurement accuracy. In contrast, surface temperature is easily
accessible and can be measured accurately and safely using external sensors. The proposed
scheme leverages surface temperature and other external measurements to infer core
temperature, eliminating the need for physical sensors inside the cell. This makes the
approach safe, non-intrusive, and straightforward to implement. Therefore, it is noticed that
the hybrid ECM and 2D-GLSTM-based methodology offers a robust, accurate, and
adaptive solution for core temperature estimation, effectively addressing the complexities

133
and challenges associated with battery TMS control. However, as this approach still
significantly depends on ECM model parameters which is subjected to change with battery
aging and dynamic operating conditions and needs domain specific expertise and
knowledge for developing ECM, thus further analysis is conducted to develop completely
data driven core temperature estimation strategies.

4.1.2.3 Impact of EV Drive Cycle on Core Temperature

While the thermal behavior of LIBs under constant current conditions has been
extensively studied, such scenarios do not accurately reflect real-life operating conditions.
To simulate practical usage, four standard drive cycle current profiles are employed to
assess the thermal behavior of LIB cells. The experimentally validated EECM and ETM
are utilized to estimate core temperature under these dynamic current profiles. The
HWFET, UDDS, LA92, and combined UDDS+LA92 drive cycles are used for this
assessment. For each experiment, all model parameters and initial conditions are kept
constant, with only the drive cycle current profile is modified. The total heat generation
within the cell core, resulting from thermal losses in the internal resistances, is calculated
using the EECM and then input into the ETM to estimate the core temperature. The current
profiles of the input drive cycles and their corresponding core temperature plots for
HWFET, UDDS, LA92, and UDDS+LA92 are illustrated in Figure 4.21 through Figure
4.24, respectively. These simulations provide insights into the thermal dynamics of LIB
cells under realistic operating conditions, highlighting the impact of varying drive cycles
on core temperature.

Figure 4.21: The plot of (a) current profile (b) Tc under HWFET.

134
 Figure 4.22: The plot of (a) current profile (b) Tc under UDDS.

Figure 4.23: The plot of (a) current profile (b) Tc under LA92.

Figure 4.24. The plot of (a) current profile (b) Tc under UDDS+LA92.

Close observation of the drive cycle current and temperature plots reveals that core
temperature variations closely follow the current fluctuations in the battery cell. This
behavior occurs because heat losses in the internal resistances of the battery are
proportional to the current. Notably, the rate of core temperature rise is particularly high
during periods of rapid acceleration in EVs, which correspond to higher current loads over
shorter durations. Despite temperature fluctuations caused by varying current loads, the
135
overall core temperature increases steadily over time due to temperature accumulation
within the cell, as only natural cooling is considered in this analysis. This highlights the
necessity of designing a robust TMS for EV applications to maintain core temperatures
within the permissible limits specified by cell manufacturers. Across all drive cycles
studied, the average temperature increase remains within 9°C. However, the highest
temperature rise occurs during the HWFET drive cycle, even though its peak current
demand is lower than that of other cycles. This phenomenon is attributed to the frequent
and rapid changes in current loading. An interesting observation is noted in the LA92 drive
cycle, where the relatively consistent rate of current change across most of the cycle results
in a slower average temperature increase. Meanwhile, the combined UDDS and LA92 drive
cycle data reveal that changes in driving patterns significantly impact core temperature.
Between 30,000 and 40,000 seconds, a transition from the UDDS to the LA92 profile
causes abrupt changes in core temperature.

These findings emphasize the critical importance of implementing fast-responsive

TMS and obtaining accurate core temperature information for LIB packs, particularly in
EV applications. Such systems are essential to mitigate rapid temperature fluctuations and
maintain battery safety and performance under dynamic operating conditions.

4.1.3 Validation of LSTM Network and KAN-based Surface and Core Temperature
Estimation Techniques

The data obtained using the experimental setup discussed in Section 3.7.2 is used for
validating the proposed completely data driven LSTM and KAN based surface and core
temperature estimation techniques. The flow of data acquisition, processing, and the
neuronal network is illustrated in Figure 4.25. The measured core and surface temperature at
20°C, including cell voltage and current over 1 A to 4 A charging and discharging cycles,
is shown in Figure 4.26. Steps involved in data processing through model implementation
and validation are discussed in the following section.

136
4.1.3.1 Data Processing
To maintain data integrity, the experimental setup and acquisition software were designed
to prevent any loss of data points. Additionally, the measurement tool incorporated built-
in outlier correction. For feature scaling, Min-Max normalization was applied. Prior to
model training and validation, the dataset was pre-processed to exclude any potential data
leakage. Specifically, data were randomly partitioned based on cell type and ambient
temperature. For each ambient condition, one of the four cells was randomly assigned as
the test set, while the remaining three served as the training set. The time-series data were
segmented into 5-second windows, each comprising 50 data points.

Figure 4.25: Data processing steps.

Figure 4.26: Measured voltage, current and temperature data at 20°C ambient temperature.

4.1.3.2 Model Implementation

The modeling framework was developed using Python 3.10, with PyTorch version
2.3.1 for deep learning and pykan 0.0.5 for implementing the KAN models. To identify
optimal hyperparameters for both model architectures, a grid search approach was
employed. For the LSTM-based networks, variations in the number of hidden layers
(ranging from 1 to 3) and hidden unit sizes (25, 50, and 75) were systematically evaluated.

137
In the case of the KAN model, the search focused on the sizes of the two hidden layers,
assessing combinations such as 8/4, 16/8, 32/16, and 64/32. The final configurations
selected through this process are summarized in Table 4.3.

Table 4.3: Hyperparameters of KAN and LSTM models

Layer LSTM KAN

1 LSTM (input =3, hidden size =25, layers =1) KAN (9)
2 Dense (32) KAN (32)
3 Dense (1) KAN (16)
4 LSTM (input =4, hidden size =50, layers =1) KAN (2)
5 Dense (32)
6 Dense (1)

Hyperparameter tuning was carried out through a comprehensive comparative analysis

using the proposed LSTM architecture, followed by systematic training and validation.
Two distinct datasets were employed throughout the training and testing phases. Initially,
the experimental dataset acquired under constant current-constant voltage (CCCV)
charging and constant current (CC) discharging protocols was used to train and validate
the model. To assess the model's robustness under realistic drive conditions, an additional
dataset was used, which captures measurements from an Urban Dynamometer Driving
Schedule (UDDS) profile at ambient temperatures of −5°C, 25°C, and 45°C. This dataset,
developed by Zhang et al. [164], features readings from A123 ANR26650 cylindrical
lithium-ion cells and includes parameters such as core temperature, surface temperature,
voltage, current, and ambient temperature. For evaluation purposes, data corresponding to
25°C and 45°C ambient conditions were selected, ensuring the use of entirely unseen data
for testing. Model performance was quantitatively assessed using three standard error
metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the
coefficient of determination (R²), as defined in Equations 4.3 through 4.5.

(𝑦̂𝑖 −𝑦𝑖 )2
𝑅𝑀𝑆𝐸 = √∑𝑛𝑖=1 (4.3)
𝑛

1
𝑀𝐴𝐸 = ∑𝑛𝑖=1|𝑦̂𝑖 − 𝑦𝑖 | (4.4)
𝑛

∑𝑛 ̅)2
𝑖=1(𝑦𝑖 −𝑦
𝑅2 = ∑𝑛 ̂ 𝑖 )2
(4.5)
𝑖=1(𝑦𝑖 −𝑦

138
 1
Where, 𝑦̅ = ∑𝑛𝑖=1 𝑦𝑖 , n is the sample size, yi is the actual value, and 𝑦̂𝑖 is the predicted
𝑛
value.

4.1.3.3 Validation Results

• CC-CV dataset
The performance of the LSTM and KAN models for both surface and core temperature
estimation under the CC-CV dataset is summarized in Table 4.4. Quantitative evaluation
was carried out across the entire ambient temperature range for both architectures. Among
the two, the KAN model consistently demonstrated superior accuracy across all selected
performance metrics. As observed from Table 4.4, the KAN-based core temperature
estimation achieved an average RMSE below 0.75°C and MAE below 0.47°C. Similarly,
for surface temperature predictions, the KAN model maintained average RMSE and MAE
values below 0.47°C and 0.30°C, respectively. Further validation of both models'
robustness across varying ambient conditions is presented in Table 4.5. For surface
temperature estimation, the KAN model exhibited MAE values ranging from 0.230°C at
40°C to 0.422°C at −20°C. In core temperature prediction using KAN, the MAE spanned
from 0.207°C at 20°C to a peak of 1.40°C at 30°C. On the other hand, the LSTM model
showed its highest core temperature estimation error 0.355°C at −10°C. Interestingly, for
surface temperature, the LSTM produced its best result with a minimum MAE of 0.293°C
at 40°C.

Table 4.4: MAE, RMSE, and R2 of core and surface temperature at 10°C

KAN LSTM
Core 0.579 0.613
MAE
Surface 0.353 0.370
Core 0.814 0.875
RMSE
Surface 0.519 0.520
Core 0.99808 0.99779
R2
Surface 0.99911 0.99910

Figure 4.27 and Figure 4.28 depict the estimated surface and core temperatures,
respectively, at an ambient temperature of 10°C. As seen in these figures, the predicted
temperature profiles closely follow the measured values, demonstrating the effectiveness
of the models in capturing thermal behavior. Figure 4.29 presents the surface temperature
139
estimation for a separate validation cell using the same input features voltage, current, and
ambient temperature. A noticeable increase in estimation error is observed at lower ambient
temperatures. The zoomed-in section of the figure highlights the superior accuracy of the
KAN model over the LSTM, particularly during transient conditions. During short periods,
the predicted values deviate more significantly from the measured temperatures,
particularly with the LSTM model. Model performance for the validation cell is
summarized in Table 4.5. A decline in estimation accuracy is evident for the validation cell
when compared to the results from drilled cells. This discrepancy can be attributed to the
physical alteration caused by the insertion of a temperature sensor, which affects the
thermal dynamics of the cell.

Figure 4.27: Measured surface temperature (green) compared with KAN (brown) and LSTM (brown)
estimation models at 10°C ambient temperature

Figure 4.28: Measured core temperature (green) compared with KAN (brown) and LSTM (red) estimation
models at 10°C ambient temperature.

140
 Figure 4.29: Validation of surface temperature estimation for the validation cell over the wide range of
ambient temperatures.

Figure 4.30 illustrates the measured capacity of each cell throughout the testing period
under different ambient temperature conditions. A general decline in capacity is observed
across all cells over time. However, the validation cell without an internally inserted core
temperature sensor exhibits noticeably less capacity loss compared to the instrumented
cells. This disparity suggests that the insertion of physical temperature sensors may have
contributed to accelerated degradation, likely due to structural disturbances introduced
during sensor installation.

141
 Figure 4.30: Change of discharge capacity indicating battery degradation

Table 4.5: MAE Surface and Core Temperature estimation in °C

Ambient Surface Core

Temperature
KAN LSTM KAN LSTM
40 0.230 0.293 0.464 0.768
30 0.363 0.473 1.403 1.143
20 0.325 0.264 0.207 0.419
10 0.403 0.413 0.714 0.639
0 0.267 0.306 0.532 0.528
-10 0.375 0.362 0.285 0.355
-20 0.422 0.432 0.493 0.571

• UDDS dataset
The UDDS dataset [164], comprising 4,467 data points collected at ambient
temperatures of 25°C and 45°C, presents a relatively limited dataset for training deep
neural networks. Despite this constraint, it serves as a valuable benchmark for assessing
model performance under real-world dynamic driving conditions. To evaluate model
accuracy, training and validation were conducted using two configurations. In both cases,
data recorded at 45°C were utilized for training. For validation, 60% of the data (equivalent
to 21.7 minutes) at 25°C was reserved split into two variations: Variation I used the initial

142
21.7 minutes for testing, while Variation II used the final 21.7 minutes. The remaining data
from 25°C was included in the training set. Table 4.6 summarizes the results for both KAN
and LSTM models across these two variations. In Variation I, the KAN model achieved a
MAE of 0.711°C for core temperature and 0.325°C for surface temperature estimation. In
Variation II, LSTM demonstrated superior performance in surface temperature prediction.
It is important to note that due to the low variability in both core and surface temperature
within this dataset, the R² was not considered a meaningful metric, as its sensitivity
diminishes with low standard deviation in the target variable.

Table 4.6: MAE, and RMSE core and surface temperature with at UDDS dataset

Variation I Variation II
KAN LSTM KAN LSTM
Core 0.711 0.878 0.652 0.935
MAE
Surface 0.325 0.550 1.138 0.423
Core 0.895 1.349 1.373 1.782
RMSE
Surface 0.387 0.777 1.813 0.652

Figure 4.31 presents a comparison between measured and estimated core and surface
temperatures for Variation I. The actual temperature profiles (in green) are plotted
alongside predictions from the KAN (brown) and LSTM (pink) models. During this period,
the surface and core temperatures are observed to rise by approximately 2.3°C and 4.6°C,
respectively. The results highlight the influence of ambient conditions on model
performance, particularly for the LSTM model, which tends to produce nearly constant
predictions 27.85°C for surface and 29.00°C for core temperatures with minimal standard
deviations of 0.11°C and 0.30°C, respectively. This behavior suggests a limited
responsiveness to temperature dynamics in the LSTM architecture under this specific
configuration. In contrast, the KAN model captures the evolving thermal behavior more
effectively, demonstrating improved alignment with the measured values. Although a
deviation from the actual temperature still exists, the KAN model more accurately tracks
the increasing trend throughout the experiment, offering better representation of battery
thermal dynamics.

143
Figure 4.31: Measured core and surface temperature (green) compared with KAN (brown) and LSTM (red)
estimation models at UDDS 25°C.

• Computational Cost
Table 4.7, provides a comparison of the inference times for the LSTM and KAN models
at an ambient temperature of 20°C. The LSTM model demonstrates lower computational
latency, requiring only 0.55 ms to 0.58 ms on the current CPU setup. In contrast, the KAN
model incurs higher processing times, ranging from 2.92 ms to 3.28 ms. While KAN
demands more computational resources in its current implementation, this limitation is
expected to diminish with future hardware upgrades and software-level optimizations. As
KAN is a relatively new modeling framework, ongoing research and development efforts
are likely to significantly improve its efficiency. Despite its higher computational burden,
KAN offers enhanced predictive capabilities, particularly in modeling complex nonlinear
patterns. This flexibility, however, comes at the cost of increased processing time
compared to the more mature and optimized LSTM model. It is also worth noting that
validation using unaltered cells (i.e., without sensor insertion) introduces certain behavioral
differences, potentially influencing model performance. Future work will focus on refining
the sensor integration method to reduce its impact on cell integrity. Nonetheless, the results
suggest that such limitations can be proactively addressed during the early modeling stages,
contributing to more accurate and robust estimations in subsequent model iterations.

144
Table 4.7: Computational cost comparison of models and different hardware setups

Computational Cost in ms
CPU-Setup LSTM KAN
Intel Xeon Gold 6338 2G 0.55 ± 0.10 2.92 ± 0.45
Intel Core i9-10885H 0.58 ± 0.04 2.28 ± 0.34

4.1.3.4 Discussions

The proposed interconnected LSTM architecture demonstrates strong capability in

estimating correlated time-series parameters, specifically surface and core temperatures of
lithium-ion batteries. By capturing temporal dependencies and parameter
interrelationships, the model delivers accurate thermal state estimation. Furthermore, the
adoption of learnable activation functions as employed in the KAN framework presents a
promising evolution beyond traditional MLP approaches. The KAN architecture was
implemented to estimate core and surface temperatures using both experimentally acquired
data from Molicel P42A cells and a publicly available dynamic dataset from Zhang et al.
[164].

Several important insights emerged from this study:

• The insertion of internal temperature sensors has a notable impact on cell
degradation, contributing to accelerated capacity loss.
• In terms of computational efficiency, the LSTM model clearly outperforms the
current KAN implementation, achieving inference times between 0.55 ms and 0.58
ms over five times faster than KAN, which ranges from 2.92 ms to 3.28 ms.
• Despite its higher computational demands, the KAN model demonstrated strong
predictive accuracy, with MAE values of 0.353°C for surface temperature and
0.579°C for core temperature.
• Even with the limited sample size of the UDDS drive cycle dataset, the KAN model
achieved MAEs of 0.325°C and 0.711°C for surface and core temperature
predictions, respectively.
These findings underscore the potential of both deep learning frameworks, with KAN
offering a compelling alternative due to its flexibility and modeling capacity for complex
nonlinear behaviors. Notably, the proposed approach allows simultaneous estimation of
145
both surface and core temperatures without relying on surface measurements, which can
reduce sensor dependency in real-world applications. However, the current research is
limited to offline analysis at the cell level. To bridge the gap toward practical deployment,
future work will focus on enhancing prediction accuracy and transitioning to real-time
implementation. Emphasis will be placed on integrating the models within on-board BMS
and exploring cloud-based BMS architectures for scalable deployment.

4.2 Validation of the Practicality of the Proposed Cloud-enhanced BMS Architecture

4.2.1 Real-time Data Acquisition Efficiency

To evaluate the real-time efficiency of the proposed cloud-integrated BMS architecture,

key operational parameters including individual cell voltages, stack current and voltage,
individual cell temperatures, and cell balancing voltage differences were continuously
collected from a 14-cell battery module. A cloud server was configured for uninterrupted
data logging and visualization to facilitate near-instantaneous monitoring. The system
supported sampling rates ranging from 1/60 Hz to 100 Hz, enabling flexible data resolution
based on application needs. Coulomb counting-based SOC estimation was implemented on
the cloud server using real-time data transmitted from the physical battery module. The
data was stored remotely and visualized with a response time of approximately 10
milliseconds, ensuring minimal delay between measurement and display. Figure 4.32 to
Figure 4.34 illustrate the real-time visualization of voltage, current, and temperature across
all cells. To assess system stability and reliability, the setup was operated continuously
over a 30-day period. Performance benchmarks demonstrated exceptional results: event-
to-logging latency consistently remained below 5 ms, throughput remained high, and
accuracy deviations were confined to under 0.1% when compared with reference
instruments. The data acquisition process operated without interruption or errors,
validating the robustness of the architecture for prolonged deployment. Visualization of
the recorded data revealed consistent thermal and electrical behavior across all cells,
affirming the system’s capacity for precise real-time monitoring. Measurement resolutions
for voltage and temperature were 152 μV and 0.1°C, respectively. Additionally, the BMS
under consideration supports a cell balancing voltage of 19 mV, enabling the system not

146
only to track key battery parameters but also to facilitate and document real-time balancing
activities.

4.2.3 Grafana Visualization Capabilities

The proposed real-time monitoring architecture effectively supports the visualization

of both individual cell voltages and cumulative pack voltages across varying series
configurations. As illustrated in Figure 4.32, the voltage of each cell in a 14-cell battery
module is plotted using a line chart, offering clear visibility into individual cell behavior.
Additionally, Figure 4.34 displays the progressive increase in total pack voltage
corresponding to the number of series-connected cells rising from 24.1 V with 7 cells to
48.1 V with all 14 cells connected demonstrating the system’s flexibility in accommodating
dynamically reconfigurable battery pack configurations. This capability offers a distinct
advantage over conventional battery architectures with fixed series-parallel topologies. The
fidelity and reliability of the visualized data are ensured by the underlying NXP BMS,
which serves as the primary source of battery parameter measurements. Data
communication with the cloud-based visualization platform is handled via TCP socket
connections. While any transmission anomalies are directly related to the associated data
processing scripts and, consequently, to the Grafana dashboard, the system remains
transparent and traceable in its operation. Time-series data is managed through InfluxDB
and Telegraf, where event timestamps are recorded in Unix nanoseconds. Although
absolute timestamps may carry an error of up to one second, the relative timing of events
remains highly precise, with sub-nanosecond resolution within a given sampling session.
Furthermore, Figure 4.33 illustrates the temperature evolution of all 14 cells during a 1C CC-
CV charging cycle, sampled at a frequency of 10 Hz. This visualization underscores the
system’s capability to track high-resolution thermal dynamics in real-time, reinforcing its
applicability for advanced diagnostics and dynamic pack control.

147
Figure 4.32: Individual cell voltages of 14 cells during CC-CV charging and CC discharging at 1C rate
with 10 Hz sampling rate.

Figure 4.33: Individual cell temperature of all 14 cells during CC-CV charging at 1C rate with 10 Hz
sampling rate.

148
 Figure 4.34: Incremental pack voltage level with increasing numbers of cells in series.

4.2.4 System Integration, Data Flow, and Data Security

In conventional cloud-based BMS, data transmission and command communication are

typically handled using TCP/IP and MQTT protocols. While effective in basic
implementations, these protocols present limitations in terms of scalability, latency, and
security. In this study, a more robust and secure data communication architecture is
adopted, employing OpenVPN and WireGuard both open-source virtual private network
(VPN) solutions for end-to-end encryption and secure data tunneling. WireGuard, being a
more recent development, offers a leaner codebase and has gained attention for its superior
performance and security features compared to traditional VPN solutions. Although
attackers may still detect the presence of network activity, the use of VPN protocols

149
effectively encrypts and obscures the transmitted data, thereby mitigating risks of
interception and tampering. While the battery data itself is transmitted over the TCP
protocol, VPN management is handled using UDP, which improves performance and
stability under real-time cloud transmission conditions. This hybrid approach ensures low-
latency communication without compromising data security. Unlike MQTT, which is
optimized for low-bandwidth and resource-constrained environments, the proposed system
architecture is designed to support high-throughput, low-latency data transmission suited
for complex real-time analytics and cloud-based BMS applications. MQTT’s inherent
limitations including potential security vulnerabilities and reduced efficiency under high
data volumes make it less appropriate for the demands of this application. The implemented
communication protocol, illustrated in Figure 4.35, encapsulates these elements and ensures
secure, reliable, and scalable data exchange between the physical battery system and the
cloud infrastructure. This integration strategy supports not only real-time monitoring and
control but also future expansion toward predictive analytics and edge-cloud hybrid
deployments.

150
 Figure 4.35: Data communication protocol ensuring security and privacy.

4.2.5 Use Case Scenarios and Advantages

Due Conventional onboard BMS are typically constrained by limited computational
resources and storage capacity, which restricts them to employing relatively simple, open-
loop state estimation techniques. While sufficient for basic operations, these approaches
are increasingly inadequate in meeting the growing demands for advanced safety,
reliability, and predictive maintenance particularly in applications such as thermal runaway
prediction, aging diagnostics, and energy management optimization. Modern data-driven
and ML-based estimation algorithms offer significant improvements in adaptability and
accuracy. However, their deployment is often impractical on embedded BMS platforms
due to their computational complexity and reliance on large volumes of historical and real-
time data. Moreover, open-loop estimators are prone to accumulating errors over time,
exacerbated by sensor drift and the infrequent availability of battery relaxation periods,
which are critical for accurate state estimation.

151
 The proposed system architecture addresses these limitations by integrating a high-
resolution sensor and data acquisition system capable of capturing precise cell-level
parameters. All operational data is logged in a cloud-based environment, enabling the use
of data-driven algorithms for continuous learning, accurate state prediction, and long-term
analytics. This not only enhances estimation accuracy but also supports advanced
functionalities such as predictive diagnostics and adaptive control strategies. To ensure
data privacy and system integrity, all user interface endpoints are protected via HTTPS
encryption and require authenticated access. This security framework facilitates safe and
controlled access to the data, enabling secure remote monitoring, model updates, and
algorithm deployment.

4.2.6 Visibility of Real-time Capabilities in Enhanced Decision-making

This section highlights the ability of the proposed BMS architecture to enhance
operational decision-making, system efficiency, and safety through real-time monitoring
and analytics. A key advantage of this system is its ability to identify the weakest-
performing cells and enable instantaneous response to anomalies thereby minimizing
system downtime and reducing safety risks. The availability of live data facilitates
informed, data-driven decisions by providing up-to-date system insights. This improves
accuracy in control algorithms and fault management. Visualization tools such as Grafana
are leveraged for continuous trend analysis, strategic forecasting, and actionable insights.
These dashboards not only simplify the interpretation of complex datasets but also make
system behavior accessible to both technical personnel and non-experts. Furthermore, the
cloud-based implementation allows remote access from anywhere, promoting centralized
control for critical applications, including fault isolation and dynamic adjustment of SOC
thresholds an approach validated in real-world scenarios such as Tesla®'s response
strategies during natural disasters.

The architecture also supports predictive maintenance by identifying early signs of cell
degradation, thus enabling proactive intervention and reducing unplanned outages. The
ability to share real-time dashboards fosters collaboration and improves communication
between stakeholders, streamlining reporting and decision-making processes. To
demonstrate the real-time decision-making capability of the proposed system, a targeted
152
experimental scenario was conducted using a 14-cell lithium-ion battery module. A weak
cell (designated as C#12) was intentionally integrated to simulate a fault condition. Despite
the BMS’s active balancing functionality, C#12 exhibited significantly different thermal
behavior, its temperature rose 6°C higher than the module average and more than 6°C
above the coolest cell by the end of the discharge cycle. This behavior is captured in Figure
4.33, which illustrates the thermal divergence over time.

In addition, Figure 4.36 presents the delta voltage (cell balancing signal), showing that
C#12 also experienced the highest voltage imbalance, exceeding 200 mV compared to
healthy cells. Notably, the adjacent cell (C#13) began to heat up as well, indicating the
onset of a thermal propagation effect. These findings emphasize the critical need for
individual cell-level temperature monitoring, as relying solely on average or sample pack
temperatures can obscure the early signs of thermal issues. The temperature monitoring
setup operates at a 10 Hz sampling rate with a processing latency of 7 ms per sample,
ensuring high-frequency thermal tracking. This capability is essential for detecting
localized anomalies that may lead to cascading failures, including thermal runaway. The
results clearly illustrate that degradation often initiates in a single faulty cell and may
quickly spread, especially when temperature sensors are not placed near the problematic
cell. Overall, this study demonstrates the architecture’s robustness in real-time anomaly
detection, predictive decision-making, and thermal risk mitigation, marking a significant
step toward safer and smarter battery management systems.

Figure 4.36: Individual cell balancing profile of the 14-cell battery module with 10 Hz sampling rate.
153
4.2.7 Scalability, Flexibility, and Resource allocation
One of the defining strengths of the proposed cloud-integrated BMS architecture lies
in its scalability and flexibility. Platforms such as InfluxDB and Grafana are designed to
manage high-volume time-series data efficiently while seamlessly integrating with
multiple data sources. These systems can scale in response to evolving operational
demands, making them highly suitable for dynamic applications like electric vehicle
battery management. In addition to operational flexibility, they also offer cost-effective
solutions through automation and optimization, reducing the need for manual intervention
and minimizing labor costs.

The real-time data acquisition and visualization capabilities enhance overall system
performance by enabling more accurate monitoring, informed resource allocation, and
responsive decision-making. User-friendly dashboards simplify the interpretation of
complex datasets, making the system accessible to both technical and non-technical users.
For instance, Figure 4.33 underscores the importance of monitoring individual cell
temperatures in real-time, while Figure 4.32 and Figure 4.34 showcase Grafana's customizable
panel configurations, adaptable to specific user needs. These examples reinforce the
effectiveness of the proposed architecture in supporting real-time data acquisition,
decoding, and visualization across diverse operational contexts.

Moreover, the system architecture supports centralized control of computational and

storage resources. It enables optimization of sampling frequencies based on the
application’s requirements. Developers must make informed decisions about sampling
rates balancing the benefits of high-resolution data (greater accuracy and responsiveness)
against the cost of increased bandwidth, storage, and processing demands. The rise of 4G
and 5G wireless technologies further enables seamless deployment of cloud-based BMS
solutions, supporting extensive data storage and long-term performance analysis over a
battery’s lifecycle.

In a cloud-based BMS, computational tasks and long-term storage are handled by

remote servers, relieving the onboard system of resource constraints. However, this
introduces new challenges including limitations in data transfer rates, memory availability,

154
network bandwidth, and communication latency. While several cloud-enabled BMS
frameworks have been introduced, existing literature [271], [272], [273] often lacks
detailed analysis of critical system-level parameters particularly concerning sampling
frequency, bandwidth usage, and resource allocation strategies for automotive-grade
systems at the module or pack level. To address this gap, a comprehensive analysis was
conducted to evaluate the impact of sampling rate on resource utilization during the
monitoring of voltage, current, and temperature in a 14-cell lithium-ion battery module.
Sampling frequencies ranging from 1/60 Hz to 100 Hz were tested to assess bandwidth
usage, memory consumption, data transfer rates, and round-trip latency. The results
underscore the importance of selecting an optimal sampling rate based on application-
specific trade-offs between data fidelity and system efficiency.

A detailed comparative analysis of bandwidth and memory usage corresponding to

different sampling frequencies is presented in Table 4.8 and Table 4.9, respectively. These
findings provide practical guidance for engineers and developers in designing scalable,
responsive, and resource-efficient BMS architectures tailored for real-time cloud-based
applications.

Table 4.8: Bandwidth allocation

Communication Standard Data Rate

3G 144 kbps -2 Mbps
4G At most 50 Mbps
5G At most 20 Gbps, common 180 Mbps
Home Broadband 5 Mbps and up

Table 4.9: Data sampling rate and memory allocation

Sampling Sampling Rate Data Rate (kbps) Memory (Mb/day)

Frequency
one/minute 0.016667 0.006718 0.580436
ten/minute 0.166667 0.06718 5.80436
1 Hz 1 0.403081 34.82616
10 Hz 10 5.758293 497.5165
50 Hz 50 20.15403 1741.308
75 Hz 75 30.63412 2646.788
100 Hz 100 40.30805 3482.616

155
 In cloud-based BMS implementations, data transfer rates play a critical role in
determining system efficiency, cost-effectiveness, and operational feasibility. For service
providers, increased data rates result in higher bandwidth demands and infrastructure costs,
while for end-users, they directly affect responsiveness, real-time control, and the stability
of BMS functionality. As shown in Table 4.8, data transmission requirements scale
proportionally with sampling frequency and the number of monitored cells. For instance,
a 14-cell module operating at a 10 Hz sampling rate can maintain feasible transmission
rates over most modern wireless networks. However, expanding this to a full-scale EV
battery pack containing approximately 1,400 cells would generate an estimated 12 Mbps
of data comparable to streaming high-definition video in real-time. This load necessitates
careful consideration of available wireless infrastructure and system design.

Table 4.9 offers a comparison of common wireless communication standards. At the

lower end, 3G networks are barely sufficient to support 10 Hz reporting rates without data
loss or delay. Higher rates, such as 100 Hz, are only achievable using 4G and 5G
technologies under ideal conditions. This highlights the need for bandwidth-conscious data
management, particularly in large-scale deployments. Data optimization strategies are
essential to reduce transmission overhead. In the current architecture, voltage values are
transmitted as text strings with 12 significant digits, significantly increasing data size.
However, through optimization including rounding, binary encoding, and compression
data rates can be reduced substantially. Under optimized conditions, even large battery
packs can transmit essential parameters at 10 Hz while maintaining a data rate below 50
kbps. Storage considerations are equally critical. In this study, the online BMS operated
continuously for 60 days, collecting real-time voltage, current, temperature, and cell
balancing data at a 10 Hz frequency. This produced approximately 29 GB of raw,
uncompressed data. Compression techniques reduced this to 2 GB, and further data pruning
removing redundant or non-informative entries brought the storage requirement down to
around 200 megabits. Such optimized datasets preserve essential historical behavior
patterns and support state estimation while also enabling second-life applications of retired
batteries.

156
 Latency is another vital performance metric. A round-trip latency study (illustrated in
Figure 4.37) measured the time required to send data from the BMS in Oshawa, Canada, to
servers in various global locations and back. Results showed average latencies of:

• 5.0 ms for Toronto (55 km),

• 65.3 ms for San Francisco (3,600 km),
• 95.7 ms for Amsterdam (6,030 km),
• 229.3 ms for Bangalore (13,400 km).
These results, summarized in Table 4.10, affirm that data centers located closer to the
user ensure significantly lower latency, supporting real-time responsiveness in control
applications. The proposed BMS architecture was tested for long-term operation and
reliability. Despite continuous operation over two months and significant data volumes, the
system maintained zero data loss, demonstrating its robustness. Furthermore, it supports
scalable, long-term storage with future reusability of data for predictive analytics, model
updates, and adaptive control. In comparison to prior designs such as the architecture
proposed by Li et al. [9], which supports up to 24 slave BMS units and uses CAN bus for
communication the proposed system offers several distinct improvements. These include:

• Automatic assignment of sequential CAN IDs, removing the need for manual
firmware configuration.
• Full compatibility with cloud infrastructure without reliance on limited single-
board computers (e.g., Raspberry Pi).
• Seamless integration of real-time data acquisition, cloud-based model deployment,
and adaptive control logic.
While Yang et al. [206] introduced the CHAIN framework for cloud-based BMS
integration, their study lacked an assessment of its practical viability. In contrast, the
current work addresses this gap by focusing on real-world implementation, highlighting
the effectiveness of the proposed architecture in meeting the demands of modern, cloud-
enabled battery systems.

157
 Figure 4.37: Round trip data transfer latency from Oshawa, Canada.

Table 4.10: Average roundtrip latency based on the physical location of the cloud server

Location Aereal Distance (km) Roundtrip Latency (ms)

Toronto (Canada) 55 5.0025
San Francisco (USA) 3,600 65.3177
Amsterdam (Netherlands) 6,030 95.7380
Bangalore (India) 13,400 229.3086

4.2.8 Applications of the proposed cloud-enhanced BMS architecture

Beyond facilitating real-time battery data acquisition and visualization, the proposed
cloud-integrated BMS architecture serves as a pivotal enabler for advanced applications
such as DT-based battery management systems, predictive analytics, and second-life
battery use cases. By offering seamless access to high-resolution, continuous data streams
including voltages, currents, and temperatures this architecture lays the foundation for
intelligent battery monitoring, diagnostics, and control in automotive and energy storage
applications.

158
 A significant opportunity lies in the integration of this real-time data infrastructure with
DT frameworks. Doing so enables manufacturers to maintain a digital replica of each
battery pack from the moment of vehicle delivery. This persistent connection supports
continuous health monitoring, operational optimization, and feedback control, ultimately
benefiting both original equipment manufacturers (OEMs) and end-users.

Key future applications of the proposed architecture include:

• Additional Sensor Integration: While while this study primarily focuses on voltage,
current, and temperature measurements, the system architecture is designed to support
various sensor types. For example, vibration sensors can be integrated to monitor
structural integrity of inter-cell connections. If a welding strip becomes detached due
to mechanical stress or vibration, the system can detect anomalies and trigger a remote
spark prevention alert, notifying the OEM or service provider in real-time.
• Feedback Control: The architecture supports fully integrated closed-loop control over
charging and discharging operations by leveraging real-time measurements of current,
cell voltages, and temperature. This capability can be extended to implement
temperature-aware and health-conscious charging strategies, such as the ones discussed
in [270], [274], which require real-time feedback to adaptively regulate charging
current and prevent overheating or degradation.
• Long-Term Health Monitoring: By continuously logging detailed operational data,
the system enables long-term health tracking of individual battery cells. This includes
analysis of charging/discharging cycles, voltage relaxation behavior, and temperature
fluctuations, which can inform SOH and RUL estimations. Integration with cloud
computing and DT platforms, as demonstrated in [275], further supports complex
thermo-electrical and aging behavior analysis over a cell’s lifecycle.
• Reconfigurable Battery Pack Control: The architecture supports intelligent control
of reconfigurable battery packs via cloud-based decision-making. Predictive algorithms
can dynamically adjust series-parallel configurations based on usage trends, load
profiles, and cell-level aging, thereby improving thermal balance, extending service
life, and optimizing energy throughput.

159
• Universal BMS Applications: Flexibility is a core advantage of the proposed system.
Instead of replacing hardware when transitioning to a new battery type, manufacturers
can update configuration files and datasheets remotely via the cloud. This modular and
software-defined approach enables universal BMS compatibility across different
chemistries and pack sizes, minimizing downtime and reducing maintenance costs.
• Second-Life Applications: Unlike conventional BMS solutions limited by onboard
memory (typically 256 Kbits to 4 MB) the proposed cloud-based system supports
virtually unlimited historical data storage, constrained only by cloud subscription levels
and OEM interest. This data can be used to classify cells based on SOH and cycling
history, enabling efficient repurposing for second-life applications such as stationary
energy storage. This process ensures optimal reuse of cells and is supported by studies
such as [276], [277].
• Remote Upgrades and Operation: The enhanced connectivity of the system supports
advanced remote update mechanisms, including firmware-over-the-air (FOTA),
software-over-the-air (SOTA), configuration-over-the-air (COTA), and data-over-the-
air (DOTA). These capabilities allow manufacturers to deploy performance
improvements, bug fixes, and new features without requiring physical access.
Additionally, remote access to historical performance data enables advanced
applications like vehicle-to-grid (V2G) power-sharing, where healthy battery modules
can be intelligently selected to support grid services.

The proposed cloud-enhanced BMS architecture not only addresses current limitations in
onboard battery monitoring but also provides a scalable, intelligent platform for future
applications ranging from adaptive control and predictive diagnostics to reusability and
lifecycle optimization.

4.2.9 Discussion of cloud-enhanced BMS Architecture

The proposed cloud-based BMS architecture offers a transformative approach to real-

time monitoring, data acquisition, and predictive control in automotive applications. By
integrating an automotive-grade BMS with CAN communication, advanced data privacy
protocols, and closed-loop control, this architecture represents a novel contribution to the

160
field addressing current limitations of onboard systems while enabling scalable, secure,
and intelligent battery management solutions. The key findings and practical implications
derived from the implementation and testing of the architecture are summarized below.

• Efficient Data Decoding: A custom-designed Python-based data parser was

developed to decode raw CAN messages into human-readable formats with
negligible latency. The decoding process is agnostic to the specific BMS used and
retains full accuracy, provided that appropriate CAN message identifiers (IDs) are
configured. This modularity enables seamless integration with any standard CAN-
enabled battery system, making the solution highly adaptable for real-world
deployment.
• Real-Time Monitoring and Fault Detection: The architecture effectively
supports remote, real-time visualization of critical battery parameters including
individual cell voltage, current, and temperature regardless of the user's physical
location. A key demonstration involved identifying the weakest cell (Cell #12) in a
14-cell module. The system tracked its abnormal thermal behavior, which resulted
in increased temperature not only in Cell#12 but also in the adjacent Cell #13,
suggesting the early onset of thermal propagation. In addition to visualizing thermal
stress, the system’s ability to track delta voltage provided insights into cell
balancing performance, further supporting early fault diagnosis and predictive
maintenance.
• Resource Allocation Analysis: A detailed analysis was conducted to evaluate the
resource implications of the proposed architecture. The system's data generation,
bandwidth requirements, and computational load were quantified at different
sampling rates. For a 14-cell module collecting voltage, current, and temperature
data, it was observed that:
▪ At 1 Hz, daily data generation amounted to approximately 34 MB.
▪ At 100 Hz, this increased substantially to 3,482 MB per day.
These results emphasize the importance of selecting appropriate sampling rates
tailored to the application’s requirements to ensure efficient use of storage, network
bandwidth, and processing capabilities.

161
 • Impact of Server Location on Latency: The Latency analysis revealed a strong
dependence on the geographical distance between the BMS and the remote server.
For identical operations:
▪ A server located in Toronto, Canada (55 km away) resulted in a round-trip
latency of approximately 5.0 ms.
▪ A server in Bangalore, India (13,400 km away) produced a latency of
approximately 229 ms.
These findings underscore the critical role of infrastructure planning and server
placement in ensuring real-time responsiveness particularly for latency-sensitive
applications like closed-loop control and thermal event mitigation.

4.3 Implementation of cloud-based SOC Estimation and Comparative Analysis

The implementation of cloud-based SOC estimation utilizes Amazon Web Services

(AWS) as the cloud computing service provider, with servers located in Frankfurt,
Germany, and Montreal, Quebec. This setup leverages battery data obtained Kollmeyer et
al. [278] and for cloud based computing the experimental setup discussed in Section 3.7.3.
is utilized. The cloud-based approach enables efficient processing and storage of large
datasets, providing the computational resources necessary for accurate SOC estimation and
facilitating real-time analysis.

4.3.1 Description of data and model training

4.3.1.1 Dataset

The dataset of Kollmeyer et al. [278] includes measurements of current, voltage,

capacity, and temperature data recorded at a frequency of 10 Hz. The data encompasses
charge, discharge, and driving cycle measurements collected from a new Turnigy Graphene
5.0 Ah 65C cell, with detailed experimental procedures provided by the researchers. Tests
were conducted over a wide temperature range, including 40°C, 25°C, 10°C, 0°C, −10°C,
and −20°C. The temperature-dependent 2RC ECM parameters R0, R1, C1, R2, and C2 from
this dataset as documented by Khanum et al. [279] also used in this study. For this study,
data from the UDDS drive cycle at 0°C is utilized for experimental validation of the cloud-
based SOC estimation concept. The remaining data is randomly divided into training and
162
testing subsets in a 70:30 ratio, facilitating the training and evaluation of the ANN models
considered in this study.

4.3.1.2 Model Implementation and Data Preprocessing

This study focuses on evaluating and comparing state-of-the-art EKF-based SOC

estimation alongside FNN, LSTM, and GRU-based SOC estimation in the cloud
environment. FNN, LSTM, and GRU-based SOC estimation techniques are adopted from
Guo et al. [254] for the comparative analysis in cloud platform. These models are selected
for their varying computational costs and error rates, offering a balanced comparison
between efficiency and accuracy. The architectures of these models are summarized in
Table 4.11.

The models are implemented using the open-source deep learning library Keras
(version 2.12.0) and the Python programming language (version 3.8). Additionally, the
EKF is implemented in Python for comparison. During preprocessing, the input data for
the neural networks undergoes feature scaling. Each feature voltage (u), current (i), and
temperature (T) is scaled using a Min-Max Scaler, ensuring that all features are normalized
to an appropriate range for effective model training. The scaling is defined in Equation 4.5.
This preprocessing step ensures consistency across features, enhancing the robustness and
performance of the neural network models.

Table 4.11: Architecture of the models

Layers FNN LSTM GRU

Layer I Flatten LSTM (128) GRU (128)
Layer II Dense (512) Dropout (0.2) Dropout (0.2)
Layer III Dense (128) Dense (8) Dense (8)
Layer IV Dense (64) Dense (1) Dense (1)
Layer V Dense (32)
Layer VI Dense (1)

Table 4.12: Machine learning Hyperparameters

Parameter Value
Batch Size 64
Epochs 150
Learning Rate 10-4
163
 𝑥(𝑘)−min (𝑥)
𝑥𝑠𝑐𝑎𝑙𝑒𝑑 (𝑘) = max(𝑥)−min (𝑥) (4.5)

Here, x represents the set of elements for a single feature, and x(k) denotes an individual
element of this set. Such scaling is essential to address scale discrepancies among features,
ensuring that the neural network models process each feature effectively. The
hyperparameters used during model training batch size, number of epochs, and learning
rate are listed in Table 4.12. These hyperparameters remain consistent across the training
of all models. The training loss for each epoch during the training of the FNN, LSTM, and
GRU are illustrated in Figure 4.38. These loss plots provide insights into the convergence
behavior and performance of the models during the training process.

Figure 4.38: Training loss of FNN, LSTM, and GRU models.

To explore the applicability of SOC estimation in the cloud, various scenarios are
analyzed, including:
• Distance vs. Request Time: Examining how the physical distance between the
client and the cloud server impacts the request and response times.
• Network Dependency on Request Time: Assessing the influence of network
conditions, such as latency and bandwidth, on the time required for processing and
transmitting SOC estimation requests.

164
 • Model Dependency on Accuracy: Comparing the accuracy of different SOC
estimation models (EKF, FNN, LSTM, GRU) to evaluate their effectiveness in the
cloud environment.
• Model Accuracy vs. Network Stability: Investigating how variations in network
stability, such as packet loss or fluctuating latency, affect the performance and
accuracy of SOC estimation models.

These scenarios provide a comprehensive understanding to evaluate the feasibility,

reliability, and performance of cloud-based SOC estimation under various operational
conditions.
4.3.1.3 Analysis of SOC Estimation Accuracy
The accuracy of the different SOC estimation methods is evaluated using the first 10%
of measurement data from the UDDS driving cycle at 0°C. This dataset corresponds to an
SOC discharge from 100% to 91.81% over 2760 seconds. Table 4.13 presents the MAE
and RMSE of SOC estimation. Among the models, the EKF demonstrates the highest
accuracy, with an MAE and RMSE of 0.0002. The FNN model achieves the best accuracy
among the neural network models, with an MAE of 0.0067, outperforming the LSTM and
GRU models.

Interestingly, the results for FNN differ from those reported by Guo et al. [254], where
FNN exhibited significantly worse performance. This deviation may stem from differences
in the datasets used in the two studies. The high precision of EKF, compared to the neural
network models, is attributed to the use of highly accurate battery parameters, whereas the
neural networks were trained on a relatively small dataset. However, it is important to note
that EKF accuracy is susceptible to degradation over time if model parameters are not
adaptive to battery aging and other variabilities encountered in practical applications. The
primary focus of this study is not to develop highly accurate neural network models for
SOC estimation but rather to explore the feasibility of implementing various SOC
estimation methods in a cloud-based BMS environment. As such, the results of this
experiment provide a baseline for cloud and DT-based BMS implementation. Some of the
major concerns are computational cost, response time, and the impact of network

165
connectivity issues on the performance of cloud-based BMS which are also analyzed in
this study as discussed in the proceeding sections.

Table 4.13: MAE and RMSE of FNN, LSTM, GRU, and EKF under UDDS driving cycle
at 0°C and a measurement frequency of 10 Hz.

Model MAE RMSE

EKF 0.0002 0.0002
FNN 0.0067 0.01
LSTM 0.0744 0.0763
GRU 0.0455 0.0493

4.3.1.4 Location dependency on execution time

This scenario evaluates the feasibility of a cloud-based BMS in terms of computation

time and the infrastructure requirements. Two elastic cloud computing (EC2) instances
from AWS were deployed in different geographic regions: Canada (Central), located in
Montreal, Quebec, and Europe (Frankfurt), located in Frankfurt, Germany. The EC2
instances are of type t2.2xlarge, running the Amazon Linux operating system. The
execution requests were sent from Ontario Tech University in Oshawa, Ontario. The direct
distances between the locations are detailed in Table 4.14. To minimize network issues, at
first a wired internet connection was used for the experiments.

The request times for the four implementations based on server location are shown in
Figure 4.39. As illustrated in Figure 4.39 (a), the average request time from Oshawa to
Montreal is approximately 62.5 ms. In contrast, the mean request time to Frankfurt is
significantly higher, around 455 ms, as shown in Figure 4.39(b).

166
Figure 4.39: Request time depending on the model between Oshawa (ON) and Montreal (QC) (a) Oshawa
(ON) and (b) Frankfurt (DE).

The models deployed in Montreal consistently achieve computation times below 100
ms, allowing for a measurement frequency of 10 Hz, apart from a few outliers visible in
the box plot. To the best my knowledge, there is no established standard for SOC estimation
frequency. However, Wei et al. [280] highlighted a typical sampling frequency of 1 Hz to
10 Hz in current practices, while Rosewater [281] emphasized that faster measurement
frequencies generally improve accuracy. Based on this, a measurement frequency of 10 Hz
is defined as the baseline for this study. Under these conditions, none of the solutions
deployed in Frankfurt meet the requirement of a computation time below 100 ms. This
indicates that cloud-based BMS solutions for global fleet operations require careful
consideration of server locations and cloud infrastructure data orchestration.

The observed increase in computation time can be attributed to geographic distance,

leading to higher latency and potential packet loss, as all instances used the same
configuration. Additionally, LSTM and GRU models exhibit higher variability in
execution time when deployed in Montreal compared to FNN and EKF. This is likely due
to the increased complexity of LSTM and GRU models, their state-handling requirements,
and the resulting computational costs.

167
Table 4.14: Approximate physical distance between the cloud server and physical BMS

Client location Cloud location Cloud location

Oshawa, ON Montreal, QC ≈450 km
Oshawa, ON Frankfurt, DE ≈6300 km

4.3.1.5 Network dependency on request time

In real-world applications, LTE networks are widely used for mobile devices, making
them an important consideration for evaluating the feasibility of cloud-based BMS.
Experiments were conducted using LTE as the network connection, focusing on the EKF
and FNN, which demonstrated the highest accuracy in Table 4.13. During these tests,
physical BMS was statically located at Oshawa campus of Ontario Tech University.

The experiments yielded an average request time of 66.1 ms ± 29.4 ms for EKF and
54.8 ms ± 10.3 ms for FNN. Interestingly, a better request time was observed for FNN
under LTE compared to the wired connection. For EKF, a higher request time was initially
recorded, which can be attributed to the network interruption at 632.5 seconds, where a
1.1-second delay occurred, as highlighted in Figure 4.40. After this interruption, the mean
request time for EKF decreased to 53.5 ms. Potential causes for the interruption include
issues with the network provider, network reconfiguration, or a switch in the radio cell.
Despite the interruption, both measurement series demonstrated acceptable results, with a
maximum request time of 100 ms. Following the reconnection, EKF also exhibited
improved request times compared to the wired connection. This improvement could be due
to suboptimal configurations in the wired network. In summary, these experiments indicate
that cloud-based SOC estimation is feasible in real-world scenarios when using LTE
networks. However, careful consideration of execution time and network stability is
necessary to ensure consistent and reliable performance.

168
Table 4.15: MAE of estimated SOC by FNN, and EKF compared to the capacity-based
SOC at different loss rates.

Loss Rate MAE FNN MAE EKF

0% 0.0062 0.0002
1% 0.0062 0.0002
2.5% 0.0062 0.0002
5% 0.0062 0.0004
7.5% 0.0062 0.0005
10% 0.0062 0.0005

4.3.1.6 Accuracy depending on loss rate

Connectivity losses are a critical factor in assessing the feasibility of cloud-based BMS
implementation. Given the reliance on mobile networks and the challenges in ensuring
complete network coverage, this study evaluates the impact of connection loss rates on
SOC prediction accuracy over time. For this analysis, the only the FNN and EKF models
are chosen, as they demonstrated the highest accuracy levels among all considered models
as discussed in Section 4.3.1.2. To simulate connection loss, a synthetic packet loss with a
specified loss rate is randomly applied across all transmitted data packages. Figure 4.41
illustrates the estimated SOC over time with a 10% loss rate for FNN, EKF, and a capacity-
based SOC estimation method, which serves as the baseline for comparison. EKF, being
state-observation dependent, is affected by missing observations due to connection losses.

Cipral et al. [282] suggested mitigating this issue by substituting missing observations
with zero values to handle incomplete data. Table 4.15 presents the MAE as a function of
connection loss rate. The results show that EKF's accuracy decreases as the loss rate
increases, evidenced by a rise in MAE. Conversely, the FNN model exhibits no significant
change in error under varying connection loss rates. Despite this, EKF consistently
demonstrates higher overall accuracy compared to the FNN model. These findings suggest
that for cloud-based SOC estimation, a hybrid approach could be advantageous. By
integrating EKF with data-driven methods like FNN, GRU, or LSTM, EKF can calibrate
neural network parameters, improving and maintaining estimation accuracy even under
varying network conditions.

169
Figure 4.40: Request time measurement using Long-Term Evolution (LTE) as network.

170
 Figure 4.41: SOC of EKF, and FNN compared to the capacity-based SOC.

4.4 Implementation of Cloud-enhanced Real-time Core Temperature Estimation

and Control

4.4.1 Data Acquisition for training and validation of Bi-LSTM

As discussed in Section 3.2.4, a Bi-LSTM model is used for real-time core temperature
estimation. To train and validate the Bi-LSTM network for core temperature estimation, a
comprehensive dataset was generated subjecting the LIB module to a wide range of
ambient temperatures, spanning from -20°C to +40°C, and performing CCCV charge and
CC discharging cycle at 0.5C, 1C, 1.5C, 2C, 3C, and 4C. Here, charging was limited to a
maximum of 2C, in compliance with the manufacturer’s specifications, while 3C and 4C
rates were used exclusively for discharging. For each change in ambient temperature, the
cells were allowed to rest for a minimum of 1 hour to ensure thermal equilibrium before
cycling began. The charge-discharge cycles were repeated multiple times at each C-rate to

171
ensure the robustness and reliability of the dataset. The dataset includes core and surface
temperature measurements, cell voltages, and currents collected over several charge-
discharge cycles. This comprehensive dataset serves as the basis for training, validating,
and testing the Bi-LSTM network. This carefully curated dataset, encompassing a wide
range of thermal and electrical conditions, ensures that the Bi-LSTM model can learn the
temporal dependencies inherent to battery thermal dynamics. Moreover, it allows the
model to generalize effectively under varying ambient temperatures and current rates,
making it practical and reliable for real-world applications in BMS. The plots of measured
core and surface temperature during 4C discharge at 25°C, 2C charge and discharge 40°C,
and 2C charge and discharge at -20°C are shown in Figure 4.42, Figure 4.43, and Figure 4.44
respectively.

Figure 4.42: Measured temperatures during 4C discharge at 25°C.

172
 Figure 4.43: Measured temperatures during 2C charging and discharge at 40°C.

Figure 4.44: Measured temperatures during 2C charging and discharge at -20°C.

4.4.2 Model Training and Validation

• Training Dataset: The Bi-LSTM model was trained and validated using sensor data
collected during various load cycles as discussed above. To ensure unbiased
evaluation, all measured data except the charge and discharge cycles of Cell#1 over
varied ambient temperatures were used for model training. Cell#1 data was reserved
exclusively for testing and validation which is completely unknown to the model.
• Data Preprocessing: To prepare the data for training, feature columns (V, I, Ts, Ta)
were scaled using Min-Max-Scaler, normalizing them to a range of [0, 1]. The target
173
 variable (Tc) was scaled separately to maintain consistency. Sequential data for the
LSTM model was generated by segmenting the dataset into overlapping windows of
5 samples, corresponding to 5 seconds of measurements. Gaussian noise was added
to the training data to enhance generalization and reduce overfitting.
• Model Architecture and Training: The Bi-LSTM network comprises two
bidirectional LSTM layers, each with 100 neurons, followed by dropout layers for
regularization and a dense layer for predicting core temperature. The model was
optimized using the Adam optimizer with a learning rate of 0.0001 and the MSE loss
function. Early stopping was employed to prevent overfitting, halting training when
validation loss failed to improve for 10 epochs. In this study, the model was trained
through 50 epochs. The plot of loss function during training and validation is shown
in Figure 4.45.

Figure 4.45: Training and validation loss.

Table 4.16: Model Hyperparameters

Batch size 32
LSTM layers 124/64/32
Epochs 50
Loss function Adam
Learning rate 0.0001
Dropout 32

174
• Model Hyperparameters: The model’s performance is fine-tuned using specific
hyperparameters to optimize learning and inference. These hyperparameters, as
shown in Table 4.16. The batch size is 32, allowing efficient processing of data in
manageable chunks during training. Epochs of 50, ensuring sufficient iterations for
the network to learn from the dataset. The loss function MSE minimizes the squared
differences between actual and predicted core temperatures. Adam optimizer is
chosen for its adaptability and efficiency in adjusting learning rates. The learning
rate is kept at 0.0001, balancing convergence speed and stability during training.
• Validation and Results
Performance of Bi-LSTM in Varied Ambient Temperatures: The model was
validated on unseen Cell#1 data. Common validation metrics MAE and RMSE as
used in the other sections of this thesis were also used here to evaluate performance,
while actual and predicted values were plotted for visual comparison. The model
demonstrated high accuracy and reliability in predicting core temperatures, proving
its suitability for real-world integration into automotive BMS. The final trained
model was saved for real-time core temperature estimation in the local machine and
cloud server. The estimation performance during 1C charge and discharge at 0°C
through 40°C for Cell#1 is shown Figure 4.46. The MAE and RMSE are noticed
0.16°C and 0.26°C respectively.

175
 Figure 4.46: Measured and estimated core temperature of Cell#1 during 1C charge and discharge at
varied ambient temperature.

Performance of Bi-LSTM in Varied Ambient Temperatures: The performance of the Bi-

LSTM under UDDS drive cycle at 25°C shows a MAE of 0.18°C and RMSE of 0.27°C for
Cell#1 as shown in Figure 4.47.

Figure 4.47: Measured and estimated core temperature of Cell#1 during UDDS at 25°C ambient
temperature.

4.4.3 Realtime Core Temperature Estimation using Pre-Trained Bi-LSTM in Local

Machine

The pre-trained Bi-LSTM model is first deployed on the local machine for real-time
core temperature estimation. The local machine equipped with a 13th Gen Intel® Core™ i9-
13900 processor, 64 GB of RAM, and a 64-bit Windows 11 operating system. The
estimation process starts with raw sensor data collected through CAN-bus and serial
communication, which includes cell voltage, stack current, surface temperature, and
ambient temperature. This data is preprocessed using Min-MaxScaler to normalize it,
ensuring compatibility with the input structure of the pre-trained Bi-LSTM. To optimize
performance, handle missing data, and maintain low-latency operation, batch processing is
introduced, where real-time data is collected in batches of five samples before being passed
to the Bi-LSTM model. Batch processing significantly improves computational efficiency
176
by reducing the overhead associated with processing individual samples, allowing the
system to operate smoothly in real-time. Moreover, it enhances the stability and reliability
of predictions by providing the model with a temporal window of data, allowing it to better
capture sequential patterns and dependencies essential for accurate temperature estimation.
The scaled batch data after screening for missing data is fed into the Bi-LSTM model for
core temperature estimation based on the temporal dependencies in the input. These
predictions are logged into an InfluxDB database alongside measured values for
visualization in Grafana dashboard. A screenshot of the Grafana dashboard showing the
real-time core temperature estimation of Cell#1 is shown in Figure 4.48. The computational
delay for real-time core temperature estimation is measured using timestamp markers
inserted at critical workflow stages (data acquisition, preprocessing, inference, and
logging). This technique ensures precise tracking of each process, providing insights into
system performance and opportunities for optimization. The total delay of approximately
30 ms is noticed including starting from CAN Bus initiation to visualization of estimated
core temperature in Grafana. The estimation error MAE and RMSE in real-time estimation
is recorded as 0.31°C and 0.4°C respectively.

177
 Figure 4.48: Visualization of real-time core temperature estimation in Grafana.

4.4.4 Realtime Core Temperature Estimation using Pre-Trained Bi-LSTM in Cloud

The pre-trained Bi-LSTM model is then deployed on a remote server hosted on

DigitalOcean, located in Toronto, approximately 55 km from the physical BMS and battery
test setup in Oshawa. The server, configured with 2 AMD vCPUs, 4 GB of RAM, and 80
178
GB of disk space running Debian 12x64, receives raw sensor data transmitted from the
physical setup over a secure network. Battery data including cell voltage, stack current,
surface temperature, and ambient temperature, is preprocessed on the server using
MinMaxScaler to ensure compatibility with the Bi-LSTM model, followed by batch
processing like the local machine setup. The total computational delay for remote core
temperature estimation, measured using timestamp markers at key workflow stages (data
acquisition, transmission, preprocessing, inference, and logging), is approximately 85 ms.
This delay, including the data transmission from Oshawa to Toronto, remains within
acceptable limits for real-time applications, typically referred as 100 ms. The estimation
accuracy remains unchanged when transitioning to the cloud, although the total
computational delay increased by approximately threefold. This delay could be reduced
further by utilizing cloud servers with parallel computing capabilities. The primary goal of
testing real-time estimation in the cloud is to assess its feasibility and demonstrate its
potential for long-term data storage, predictive maintenance, and decision-making by
OEMs. Additionally, the cloud-based solution offers centralized monitoring capabilities
and easy accessibility over the internet, making it a scalable and practical option for
extended deployment scenarios.

4.4.5 Data Storage and Visualization

In both scenarios, all measured and estimated core temperature data are stored on the
cloud server. The current cloud server has a storage capacity of 80 GB, which is flexible
and can be extended with premium plans. Users can customize the computational power
and storage capacity based on specific application requirements. In this study, all measured
and estimated values were first logged to InfluxDB, with Grafana querying the database to
visualize real-time measured and estimated values. To test the robustness and reliability of
the cloud server, the system was operated continuously for 7 days, collecting raw data and
estimating core temperatures in real-time. Over this period, approximately 2 GB of data
was generated at a sampling rate of 1 Hz. However, through data compression and
optimization techniques, such as data pruning, the raw data size can be significantly
reduced by removing redundant entries and retaining only the critical information needed

179
for state estimation. This approach ensures efficient use of storage while preserving the
historical behavior of the cells essential for predictive analysis and decision making.

4.4.6 Real-Time Updates and Decision-Making

The estimated and measured core temperature data is used to monitor all 14 cells in the
battery module and for feedback control purposes. If the core temperature of any cell
exceeds the predefined safety threshold, the system immediately halts charging or
discharging operations by sending control signals to the EA battery cycler via the CAN
bus. This closed-loop control mechanism, implemented using threading and continuous
data monitoring, ensures proactive thermal management to mitigate thermal runaway risks.
To demonstrate this control function, the upper core temperature threshold was set to 45°C.
The module was charged at a 2C rate starting from an ambient temperature of 25°C. The
system began charging the LIB module at approximately 23:07 (local time) and
successfully identified that the core temperature of Cell#5 reached the upper temperature
threshold around 23:22, halting the charging process to prevent overheating and thermal
runaway. Similarly, during discharging at a 2C rate, the system initiated the operation at
15:18 (local time) and stopped at 15:32 when the temperature threshold was reached. A
screenshot of the Grafana dashboard during these events is shown in Figure 4.49 and Figure
4.50 respectively.

Figure 4.49: Real-time decision making to prevent overheating and thermal runaway during charging.

180
 Figure 4.50: Real-time decision making to prevent overheating and thermal runaway during discharging.

Additionally, as shown in Figure 4.51, during 2C discharging at an ambient temperature

of 25°C, the core temperature of Cell#5 reached 40°C (as an example) approximately 2
minutes earlier than the surface temperature. This highlights that a core-temperature-
informed TMS response 2 minutes faster compared to the surface temperature informed
control, which is critical for preventing overheating and thermal runaway. Such prompt
action is crucial as thermal runaway propagates rapidly from a single cell to the entire
battery pack, potentially leading to catastrophic failure.

181
 Figure 4.51: Real-time decision making showing faster reaction time.

4.5 Integration of Digital-twin-based Battery Thermal Management Control System

Finally, building on the successful implementation of the Bi-LSTM-based real-time

core temperature estimation method and its superior accuracy compared to other techniques
explored in this thesis, the Bi-LSTM-based core temperature estimation approach is
extended to create a DT of a LIB cell for predicting core temperature ahead of time. The
primary goal of the developed DT is to emulate the core thermal behavior of a single battery
cell by integrating real-time data with historical context.

By leveraging voltage, current, surface temperature, and ambient temperature as input

parameters, the DT predicts the core temperature of a cell up to three minutes into the
future. This predictive capability facilitates proactive thermal management, significantly
enhancing the safety and reliability of battery systems. The proposed DT architecture
includes essential components such as a data acquisition layer, a preprocessing module,
predictive analytics powered by a Bi-LSTM model, and a feedback mechanism. Real-time
data from sensors and CAN systems serve as the foundation for the predictive model, while

182
historical data provides temporal context, enabling the Bi-LSTM to capture intricate
thermal dynamics over time.

Using a sliding-window approach, the DT continuously captures sequences of

historical inputs and predicts the future core temperature, which is validated against real-
time measurements to ensure accuracy. Figure 4.52 and Figure 4.53 show the measured and
predicted core temperatures (three minutes ahead) during 1C charging and discharging at
25°C. The MAE of the predicted core temperature three minutes ahead is approximately
0.42°C during charging and 0.40°C during discharging, compared to the measured values
recorded at the same timestamps

Figure 4.52: Measured and predicted core temperature (3 minutes ahead of time) by the digital twin during
1C-charging at 25°C.

183
Figure 4.53: Measured and predicted core temperature (3 minutes ahead of time) by the digital twin during
1C-discharging at 25°C.

The integration of the DT within the battery thermal safety framework greatly enhances
its functionality. By offering a forward-looking view of core temperature dynamics, the
DT enables predictive control actions that mitigate risks associated with thermal runaway
or suboptimal performance. The visualization of both predicted and actual core
temperatures within the cloud-enhanced BMS ensures transparency and supports informed
decision-making. This integration establishes a foundation for future advancements in DT-
based control systems, emphasizing scalability, real-time adaptability, and improved
safety. The feedback loop, which leverages predicted core temperatures ahead of time,
enhances the reliability of the core temperature estimation model and facilitates adaptation
to evolving operational conditions. Consequently, the extension of the Bi-LSTM-based
core temperature estimation framework into a DT architecture marks a significant step
forward in achieving predictive and real-time battery thermal management.

184
4.6 Challenges and Limitations

4.6.1 Cloud-enhanced BMS Architecture

Implementing the proposed cloud-based BMS architecture presents several challenges.

One primary concern is the potential for latency and interruptions in data communication
in real-world applications. Additionally, the reliance on data-driven techniques such as AI
and ML for state estimation demands substantial computational power. Processing
significant historical data for state predictions can introduce computational delays, which,
when combined with network communication latency, may slow down the prediction
process. Furthermore, the use of the internet and centralized remote servers raises concerns
about data protection and cybersecurity, especially as these systems become increasingly
integrated and complex. This thesis not only presents a flexible, efficient, and robust cloud-
based BMS architecture but also provides detailed discussions on these challenges to offer
realistic guidelines for further research and development.

4.6.2 Digital-twin-based BMS

Some of the major challenges associated with the integration of DT-based BMS
includes the mix usage of various DT models and the nonuniformity in the semantic syntax
of data, resulting in surplus or deficits of knowledge, this would cause issues between tiers
as it would create obstacles to basic knowledge interoperability. This is also highlighted by
Wang et al. [213]. For varying applications and different domains, the repetition of
rebuilding DT models is required which will require substantial time and labor to perform.
Other difficulties to account for would be the exact model and data-sharing when migrating
DT models to other platforms. Furthermore, DT operations process large sets of
heterogeneous data, this will strain the storage as well as the communication capacity.
Some DT application scenarios require massive data storage and processing as the
applications focus on preventative maintenance and fault diagnosis, thereby slowing down
the data processing. These DT applications accumulating the vast amount of data and
utilizing big data source analysis pose large challenges for storage and computational
capabilities. Some of the research challenges as highlighted by Semeraro et al. [283]
includes, the development of DT is still very basic and significant research efforts are

185
required to improve the applicability of the DT in the energy storage sector. Also, the
performance of the DT needs to be improved in terms of the lifecycle stage and functions
of a DT.

Another, major concern is the DT is often misinterpreted as computer-aided design

(CAD) or IoT. CAD is utilized for representation of physical systems or entities in a digital
space [284] and IoT is for physical space. Comparatively, the concept of DT is designed
around the mutual dependence between digital and physical spaces. To develop the
interconnection, it necessitates the need to circulate online sensory data between the DT
and physical system to create a strong link. This relation can only be developed after
accurate replication of the physical system’s dynamic environment and all operational
factors of the digital space [284]. To create this bridge, it is necessary to develop a strong
link between the physical object and DTs. This can be achieved by circulating online
sensory data between the twins. A proper link between the physical object and DTs can
only be developed after precisely replicating the dynamic environment of the physical
system and all the factors that impact the operations of this system in the digital
space [285]. Also, often DT is confused with digital clones. Digital clones are the static
virtual representation of physical system. It does not have any adaptability with the changes
in operating conditions and self-aging. Whereas, DT has the adaptability functionality. In
short it can be stated that the digital clone with a feedback loop makes a DT. In summary,
proper understanding and in-depth knowledge is essential for practical implementation of
DT-based BMS.

186
Chapter 5. Conclusions and Future Work

This thesis has made substantial contributions to advancing LIB management systems
by addressing critical challenges in battery data monitoring (including voltage, current,
temperature, and balancing voltage) in local and cloud-server, rapid thermal management
control, real-time state estimation, and predictive analytics. Integrating cutting-edge
technologies, including deep learning, cloud computing, and digital twin architectures, the
proposed solutions improve the safety, reliability, and enhanced TMS control response
time for LIB management system. This chapter summarizes the key findings of the research
and outlines potential directions for future work.

5.1. Advanced Core Temperature Estimation Techniques

• Contributions and Key Findings

Accurate core temperature estimation is vital for LIB battery management to improve
safety, prompt TMS control and preventing from overheating and accelerated degradation
compared to state-of-the-art BMS technology. The thesis developed and validated first-
and second-order thermal models-based techniques and hybrid ECM-LSTM-based core
temperature estimation techniques.

• Key Findings:
1) The second-order thermal model provided superior accuracy by
incorporating thermal capacitance and resistances, reducing the core
temperature estimation error to 1–2 K compared to 7–8 K in the first-order
model.

2) The hybrid ECM-LSTM framework achieved RMSE values of 0.52–0.81°C

across diverse ambient temperatures and dynamic profiles, significantly
outperforming existing methods.

3) Incorporating total heat generation into the hybrid model enhanced

adaptability to cell aging and varying operational conditions to some extent.

187
 4) The core temperature of a battery is significantly influenced by the drive
cycle profile of an EV. Variations in driving patterns can lead to drastic
changes in the core temperature. To effectively capture these changes, the
BMS temperature measurement system must exhibit a rapid response.
Additionally, a core temperature-informed control system is critical for
ensuring the safety and protection of the battery pack by capturing this rapid
changes.

5) Despite, satisfactory estimation performance, both the modeling technique

still depends on the ECM model parameters which are subject to change
with battery aging and dynamic operating conditions. This requires adaptive
ECM modeling which poses significant challenges and complexity.
Moreover, developing such model requires domain specific expertise,
testing equipment, and time.

• Future Directions:
1) Extending the models to include long-term aging effects.

2) Developing real-time adaptive parameterization for diverse LIB chemistries

and pack configurations.

3) Investigate scalability to multi-cell systems, considering thermal gradients

and optimal sensor placement.
4) Developing completely data drive AI, ML based core temperature
estimation techniques free from ECM based model and requirement of
domain specific expertise and cell testing equipment.

5.2 Cloud-Enhanced Battery Management Systems

• Contributions and Key Findings

A novel cloud-based BMS architecture was proposed to enhance scalability, real-time

monitoring, remote monitoring, control and long-term data storage to facilitate predictive
maintenance and second life application of retired EV batteries. The proposed architecture
demonstrated its practicality for real-time operations and advanced data analytics.

188
• Key Findings:

1) Efficient data decoding and real-time monitoring were achieved with

negligible latency (<10 ms), enabling advanced fault detection and
predictive maintenance.

2) The cloud-based BMS framework provided real-time battery data including

voltage, current, temperature of individual cell monitoring and predictive
capabilities, accurately forecasting thermal behavior three minutes in
advance.

3) Experimental validations highlighted the architecture's potential for

centralized monitoring, fault detection, and predictive maintenance.

4) Analysis of resource allocation and latency demonstrated feasibility, with

latencies ranging from 5 ms (local) to 229 ms (remote).

5) Quantitative analyses of data storage and bandwidth requirements provide

insights into the system's scalability.

• Future Directions
1) Incorporate edge computing to reduce dependency on cloud infrastructure
and achieve ultra-low latency.

2) Expand the digital twin framework to include fault propagation modeling

and pack-level dynamics.

3) Address cybersecurity concerns in cloud-based BMS implementations.

4) Develop second-life applications for retired batteries using the proposed

architecture.

5) Incorporate advanced encryption and cybersecurity measures to enhance

system security.

189
5.3 Deep Learning-Based Estimation and Thermal Management Strategies
• Contributions and Key Findings

Deep learning algorithms, including Bi-LSTM and KAN-LSTM, were employed to

estimate core and surface temperature. Further, core temperature informed BMS enhance
the accuracy and adaptability of TMS control and protection system.

• Key Findings:

1) Bi-LSTM achieved MAE value as low as 0.16°C with off-line data and in
real-time MAE values as low as 0.31°C, providing actionable insights for
preemptive thermal management.

2) KAN-LSTM-based strategy reduced dependency on sensors and enabled

simultaneous core and surface temperature estimation, demonstrating
scalability and efficiency for practical applications without needing any
physical measurement sensor for surface temperature measurement.

3) Bi-LSTM-based real-time core temperature estimation and implementation

of the feedback control of TMS reduced the response time by 2 minutes to
prevent individual cell overheating and preventing from thermal runaway
compared to surface temperature information TMS control.

• Future Directions

1) Combine physics-informed neural networks (PINNs) with Bi-LSTM for

enhanced interpretability and accuracy.
2) Validate scalability of these models for large-scale battery systems in EVs
and grid-tied energy storage.

5.4 Integration of Digital-twin-based Battery Thermal Management Control System

• Contributions and Key Findings

The Bi-LSTM-based core temperature estimation framework was extended into a DT

architecture, enabling real-time predictive thermal management.

190
 • Key Findings:

1) The DT integrates real-time data and historical context to predict core

temperature up to three minutes into the future, achieving an MAE as low
as 0.4°C compared to measured values.

2) The architecture incorporates key components, including data acquisition,

preprocessing, predictive analytics powered by Bi-LSTM, and a feedback
mechanism, enabling continuous adaptation to evolving conditions.

3) The sliding-window approach ensures accurate, dynamic predictions,

validated against real-time measurements and demonstrated significant
accuracy in core temperature prediction.

4) Visualization of predicted and actual core temperatures in a Grafana-based

dashboard facilitates transparency and informed decision-making even
from remote location enabling centralized monitoring and control.

• Future Directions

1) Incorporate edge computing for ultra-low latency and enhanced scalability.

2) Extend the framework to battery pack applications, including dynamic

driving cycles and fault propagation.

3) Develop advanced fault detection modules to complement the thermal

management system.

5.5 Cloud-Based SOC Estimation and Comparative Analysis

• Contributions and Key Findings

A comparative analysis of SOC estimation methods was conducted using cloud-based

frameworks, including EKF and neural network approaches.

191
 • Key Findings:

1) The FNN model achieved the shortest average computing time of 54.8 ms
under LTE mobile network conditions.

2) EKF demonstrated the highest accuracy (MAE = 0.0002) among the

methods, while neural networks showed superior adaptability under
dynamic conditions.

3) There was no significant computational cost difference observed among the

methods (EKF, FNN, LSTM, and GRU), suggesting a negligible impact of
computational differences on local performance.

4) The computational times of 66.1 ms ± 29.4 ms for EKF and 54.8 ms ± 10.3
ms for FNN are acceptable, as they align with the typical sampling
frequency range for state estimation (1–10 Hz).

5) Latency analysis highlighted the need for geographically distributed cloud

infrastructure to optimize performance.

• Future Directions

1) Develop fusion models combining EKF and neural networks for enhanced
adaptability and long-term accuracy.

2) Validate SOC estimation techniques at the pack level to assess real-world

applicability.

This thesis presents a comprehensive approach to modernizing LIB management

systems by integrating advanced modeling, deep learning, and DT technologies. The
proposed cloud-enhanced BMS architecture is capable to provide cloud-based state
estimation, control and long-term data storage for predictive maintenance and second-life
applications of retired EV batteries. The findings offer transformative solutions for
improving battery safety, reliability, and operational efficiency. Future research should
focus on enhancing scalability, reducing latency, and addressing emerging challenges to
further advance the capabilities of next-generation BMS.

192
5.7 Limitations of the Study

While the proposed solutions addressed significant challenges, some limitations

remain:

• The models require extensive initial data collection for training and validation.

• The integration of cloud-based solutions introduces potential cybersecurity risks,

requiring robust data encryption and access control.

• The scalability of the framework for larger battery packs, such as those in heavy-
duty EVs, remains untested.

5.8 Future Research Directions

Building on the contributions of this thesis, the following directions are recommended for
future research:

1. Integration with Emerging Technologies:

o Incorporate solid-state batteries and hybrid chemistries into the proposed
framework.

o Leverage advancements in quantum computing and blockchain for

enhanced computational efficiency and data security.

2. Autonomous TMS Control:

o Develop DT enabled closed-loop control systems for autonomous TMS and
predictive safety intervention.

3. Applications Beyond EVs:

o Adapt the framework for grid-tied energy storage systems, aerospace, and
marine applications.

o Explore second-life applications for retired EV batteries in renewable

energy systems.

4. Scalability and Cost Optimization:

193
o Investigate cost-effective solutions for scaling the framework to large-scale
battery systems.

o Enhance computational efficiency to enable real-time deployment in

resource-constrained environments.

194
REFERENCES

[1] Canada Energy Regulator, “Market Snapshot: Zero emission vehicles now account
for over 10% of all new vehicles in Canada,” 2024. [Online]. Available:
https://www.cer-rec.gc.ca/en/data-analysis/energy-markets/market-
snapshots/2024/market-snapshot-zero-emission-vehicles-now-account-for-over-
10-percent-of-all-new-vehicles-in-canada.html#:~:text=Zero-emissions vehicles
(ZEVs),185%2C000 new registrations *
[2] B. Chidambaranathan, M. Vijayaram, V. Suriya, R. Sai Ganesh, and S. Soundarraj,
“A review on thermal issues in Li-ion battery and recent advancements in battery
thermal management system,” in Materials Today: Proceedings, 2020. doi:
10.1016/j.matpr.2020.03.317.
[3] F. Leng, C. M. Tan, and M. Pecht, “Effect of Temperature on the Aging rate of Li
Ion Battery Operating above Room Temperature,” Sci. Rep., 2015, doi:
10.1038/srep12967.
[4] C. Chetri, A. Samanta, and S. Williamson, “Critical Understanding of Temperature
Gradient During Fast Charging of Lithium-ion Batteries at Low Temperatures,” in
IECON Proceedings (Industrial Electronics Conference), 2023. doi:
10.1109/IECON51785.2023.10312304.
[5] K. Liu, Y. Li, X. Hu, M. Lucu, and W. D. Widanage, “Gaussian Process
Regression with Automatic Relevance Determination Kernel for Calendar Aging
Prediction of Lithium-Ion Batteries,” IEEE Trans. Ind. Informatics, 2020, doi:
10.1109/TII.2019.2941747.
[6] J. S. Edge et al., “Lithium ion battery degradation: what you need to know,”
Physical Chemistry Chemical Physics. 2021. doi: 10.1039/d1cp00359c.
[7] X. Feng, M. Ouyang, X. Liu, L. Lu, Y. Xia, and X. He, “Thermal runaway
mechanism of lithium ion battery for electric vehicles: A review,” Energy Storage
Materials. 2018. doi: 10.1016/j.ensm.2017.05.013.
[8] J. B. Robinson et al., “Non-uniform temperature distribution in Li-ion batteries
during discharge - A combined thermal imaging, X-ray micro-tomography and
electrochemical impedance approach,” J. Power Sources, 2014, doi:
10.1016/j.jpowsour.2013.11.059.
[9] S. Surya, A. Samanta, V. Marcis, and S. Williamson, “Hybrid Electrical Circuit
Model and Deep Learning-based Core Temperature Estimation of Lithium-ion
Battery Cell,” IEEE Trans. Transp. Electrif., pp. 1–1, 2022, doi:
10.1109/TTE.2022.3170359.
[10] Y.-M. Cheng, D.-X. Gao, F.-M. Zhao, and Q. Yang, “A data-driven early warning
method for thermal runaway during charging of lithium-ion battery packs in
electric vehicles,” Meas. Sci. Technol., vol. 36, no. 1, p. 016233, Jan. 2025, doi:
10.1088/1361-6501/ad9d68.

195
[11] J. Liu, S. Yadav, M. Salman, S. Chavan, and S. C. Kim, “Review of thermal
coupled battery models and parameter identification for lithium-ion battery heat
generation in EV battery thermal management system,” International Journal of
Heat and Mass Transfer. 2024. doi: 10.1016/j.ijheatmasstransfer.2023.124748.
[12] Q. Yu, C. Wang, J. Li, R. Xiong, and M. Pecht, “Challenges and outlook for
lithium-ion battery fault diagnosis methods from the laboratory to real world
applications,” eTransportation. 2023. doi: 10.1016/j.etran.2023.100254.
[13] Y. Yan et al., “Fault diagnosis of lithium-ion battery sensors based on multi-
method fusion,” J. Energy Storage, 2024, doi: 10.1016/j.est.2024.110969.
[14] X. Lin, H. E. Perez, J. B. Siegel, and A. G. Stefanopoulou, “Robust estimation of
battery system temperature distribution under sparse sensing and uncertainty,”
IEEE Trans. Control Syst. Technol., 2020, doi: 10.1109/TCST.2019.2892019.
[15] S. Al Hallaj, H. Maleki, J. S. Hong, and J. R. Selman, “Thermal modeling and
design considerations of lithium-ion batteries,” J. Power Sources, vol. 83, no. 1–2,
pp. 1–8, 1999, doi: 10.1016/S0378-7753(99)00178-0.
[16] C. Forgez, D. Vinh Do, G. Friedrich, M. Morcrette, and C. Delacourt, “Thermal
modeling of a cylindrical LiFePO4/graphite lithium-ion battery,” J. Power
Sources, vol. 195, no. 9, pp. 2961–2968, 2010, doi:
10.1016/j.jpowsour.2009.10.105.
[17] A. Masias, J. Marcicki, and W. A. Paxton, “Opportunities and Challenges of
Lithium Ion Batteries in Automotive Applications,” ACS Energy Letters. 2021.
doi: 10.1021/acsenergylett.0c02584.
[18] M. U. Ali, S. H. Nengroo, M. A. Khan, K. Zeb, M. A. Kamran, and H. J. Kim, “A
real-time simulink interfaced fast-charging methodology of lithium-ion batteries
under temperature feedback with fuzzy logic control,” Energies, 2018, doi:
10.3390/en11051122.
[19] S. Manzetti and F. Mariasiu, “Electric vehicle battery technologies: From present
state to future systems,” Renewable and Sustainable Energy Reviews. 2015. doi:
10.1016/j.rser.2015.07.010.
[20] B. Bilgin et al., “Making the Case for Electrified Transportation,” IEEE Trans.
Transp. Electrif., 2015, doi: 10.1109/TTE.2015.2437338.
[21] R. Xiong, Y. Pan, W. Shen, H. Li, and F. Sun, “Lithium-ion battery aging
mechanisms and diagnosis method for automotive applications: Recent advances
and perspectives,” Renewable and Sustainable Energy Reviews. 2020. doi:
10.1016/j.rser.2020.110048.
[22] L. H. Saw, K. Somasundaram, Y. Ye, and A. A. O. Tay, “Electro-thermal analysis
of Lithium Iron Phosphate battery for electric vehicles,” J. Power Sources, 2014,
doi: 10.1016/j.jpowsour.2013.10.052.
[23] I. Baccouche, S. Jemmali, B. Manai, A. Nikolian, N. Omar, and N. Essoukri Ben
196
 Amara, “Li-ion battery modeling and characterization: An experimental overview
on NMC battery,” International Journal of Energy Research. 2022. doi:
10.1002/er.7445.
[24] T. Nemeth, P. Schröer, M. Kuipers, and D. U. Sauer, “Lithium titanate oxide
battery cells for high-power automotive applications – Electro-thermal properties,
aging behavior and cost considerations,” J. Energy Storage, 2020, doi:
10.1016/j.est.2020.101656.
[25] R. Kumar and S. Chavan, “Numerical and Experimental Investigation of Thermal
Behaviour for Fast Charging and Discharging of Various 18650 Lithium Batteries
of Electric Vehicles,” Int. J. Heat Technol., 2022, doi: 10.18280/ijht.400618.
[26] L. Xu et al., “Interfaces in Solid-State Lithium Batteries,” Joule. 2018. doi:
10.1016/j.joule.2018.07.009.
[27] J. Duan et al., “Building Safe Lithium-Ion Batteries for Electric Vehicles: A
Review,” Electrochemical Energy Reviews. 2020. doi: 10.1007/s41918-019-
00060-4.
[28] C. R. Birkl, M. R. Roberts, E. McTurk, P. G. Bruce, and D. A. Howey,
“Degradation diagnostics for lithium ion cells,” J. Power Sources, vol. 341, pp.
373–386, 2017, doi: 10.1016/j.jpowsour.2016.12.011.
[29] Y. Chen et al., “A review of lithium-ion battery safety concerns: The issues,
strategies, and testing standards,” Journal of Energy Chemistry. 2021. doi:
10.1016/j.jechem.2020.10.017.
[30] A. K. M. A. Habib, M. K. Hasan, M. Mahmud, S. M. A. Motakabber, M. I.
Ibrahimya, and S. Islam, “A review: Energy storage system and balancing circuits
for electric vehicle application,” IET Power Electron., vol. 14, no. 1, pp. 1–13,
2021, doi: 10.1049/pel2.12013.
[31] A. Opitz, P. Badami, L. Shen, K. Vignarooban, and A. M. Kannan, “Can Li-Ion
batteries be the panacea for automotive applications?,” Renewable and Sustainable
Energy Reviews. 2017. doi: 10.1016/j.rser.2016.10.019.
[32] S. Ahmed et al., “Cost of automotive lithium-ion batteries operating at high upper
cutoff voltages,” J. Power Sources, 2018, doi: 10.1016/j.jpowsour.2018.09.037.
[33] Y. Xing, E. W. M. Ma, K. L. Tsui, and M. Pecht, “Battery management systems in
electric and hybrid vehicles,” Energies. 2011. doi: 10.3390/en4111840.
[34] L. Lu, X. Han, J. Li, J. Hua, and M. Ouyang, “A review on the key issues for
lithium-ion battery management in electric vehicles,” Journal of Power Sources.
2013. doi: 10.1016/j.jpowsour.2012.10.060.
[35] W. Waag, C. Fleischer, and D. U. Sauer, “Critical review of the methods for
monitoring of lithium-ion batteries in electric and hybrid vehicles,” Journal of
Power Sources. 2014. doi: 10.1016/j.jpowsour.2014.02.064.

197
[36] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, “A review of
lithium-ion battery state of charge estimation and management system in electric
vehicle applications: Challenges and recommendations,” Renewable and
Sustainable Energy Reviews. 2017. doi: 10.1016/j.rser.2017.05.001.
[37] A. Samanta and S. S. Williamson, “A survey of wireless battery management
system: Topology, emerging trends, and challenges,” Electronics (Switzerland).
2021. doi: 10.3390/electronics10182193.
[38] Analog Devices (Linear Tech.), “Wireless battery management systems highlight
Industry’s drive for higher reliability,” Analog Devices Inc., 2016.
[39] Z. Wei, J. Zhao, H. He, G. Ding, H. Cui, and L. Liu, “Future smart battery and
management: Advanced sensing from external to embedded multi-dimensional
measurement,” Journal of Power Sources. 2021. doi:
10.1016/j.jpowsour.2021.229462.
[40] T. Tanizawa, T. Suzumiya, and K. Ikeda, “Cloud-connected battery management
system supporting e-mobility,” Fujitsu Sci. Tech. J., 2015.
[41] T. Kim, D. Makwana, A. Adhikaree, J. S. Vagdoda, and Y. Lee, “Cloud-based
battery condition monitoring and fault diagnosis platform for large-scale lithium-
ion battery energy storage systems,” Energies, vol. 11, no. 1, pp. 1–15, 2018, doi:
10.3390/en11010125.
[42] W. Li, M. Rentemeister, J. Badeda, D. Jöst, D. Schulte, and D. U. Sauer, “Digital
twin for battery systems: Cloud battery management system with online state-of-
charge and state-of-health estimation,” J. Energy Storage, vol. 30, no. May, p.
101557, 2020, doi: 10.1016/j.est.2020.101557.
[43] T. R. Tanim, C. D. Rahn, and C. Y. Wang, “State of charge estimation of a lithium
ion cell based on a temperature dependent and electrolyte enhanced single particle
model,” Energy, 2015, doi: 10.1016/j.energy.2014.12.031.
[44] A. Farmann and D. U. Sauer, “A study on the dependency of the open-circuit
voltage on temperature and actual aging state of lithium-ion batteries,” J. Power
Sources, 2017, doi: 10.1016/j.jpowsour.2017.01.098.
[45] F. Zheng, J. Jiang, B. Sun, W. Zhang, and M. Pecht, “Temperature dependent
power capability estimation of lithium-ion batteries for hybrid electric vehicles,”
Energy, 2016, doi: 10.1016/j.energy.2016.06.010.
[46] A. Samanta, S. Chowdhuri, and S. S. Williamson, “Machine Learning-Based Data-
Driven Fault Detection/Diagnosis of Lithium-Ion Battery: A Critical Review,”
Electronics, vol. 10, no. 11, p. 1309, May 2021, doi: 10.3390/electronics10111309.
[47] S. S. Zhang, K. Xu, and T. R. Jow, “The low temperature performance of Li-ion
batteries,” J. Power Sources, 2003, doi: 10.1016/S0378-7753(02)00618-3.
[48] S. Mohan, Y. Kim, and A. G. Stefanopoulou, “Energy-Conscious Warm-Up of Li-
Ion Cells from Subzero Temperatures,” IEEE Trans. Ind. Electron., 2016, doi:
198
 10.1109/TIE.2016.2523440.
[49] S. Surya, V. Marcis, and S. Williamson, “Core Temperature Estimation for a
Lithium ion 18650 Cell,” Energies, 2020, doi: 10.3390/en14010087.
[50] S. Surya and A. Mn, “Effect of Fast Discharge of a Battery on its Core
Temperature,” Int. Conf. Futur. Technol. Control Syst. Renew. Energy, ICFCR
2020, 2020, doi: 10.1109/ICFCR50903.2020.9249999.
[51] S. Steinhorst, M. Lukasiewycz, S. Narayanaswamy, M. Kauer, and S. Chakraborty,
“Smart cells for embedded battery management,” in Proceedings - 2nd IEEE
International Conference on Cyber-Physical Systems, Networks, and Applications,
CPSNA 2014, 2014. doi: 10.1109/CPSNA.2014.22.
[52] X. Huang, X. Sui, D. I. Stroe, and R. Teodorescu, “A Review of Management
Architectures and Balancing Strategies in Smart Batteries,” in IECON Proceedings
(Industrial Electronics Conference), 2019. doi: 10.1109/IECON.2019.8926687.
[53] J. Fleming, T. Amietszajew, J. Charmet, A. J. Roberts, D. Greenwood, and R.
Bhagat, “The design and impact of in-situ and operando thermal sensing for smart
energy storage,” J. Energy Storage, 2019, doi: 10.1016/j.est.2019.01.026.
[54] Y. wei Pan et al., “A computational multi-node electro-thermal model for large
prismatic lithium-ion batteries,” J. Power Sources, vol. 459, no. March, p. 228070,
2020, doi: 10.1016/j.jpowsour.2020.228070.
[55] D. Bernardi, E. Pawlikowski, and J. Newman, “GENERAL ENERGY BALANCE
FOR BATTERY SYSTEMS.,” in Electrochemical Society Extended Abstracts,
1984. doi: 10.1149/1.2113792.
[56] F. Sun, R. Xiong, H. He, W. Li, and J. E. E. Aussems, “Model-based dynamic
multi-parameter method for peak power estimation of lithium-ion batteries,” Appl.
Energy, 2012, doi: 10.1016/j.apenergy.2012.02.061.
[57] M. Ghalkhani, F. Bahiraei, G. A. Nazri, and M. Saif, “Electrochemical–Thermal
Model of Pouch-type Lithium-ion Batteries,” Electrochim. Acta, 2017, doi:
10.1016/j.electacta.2017.06.164.
[58] X. G. Yang, Y. Leng, G. Zhang, S. Ge, and C. Y. Wang, “Modeling of lithium
plating induced aging of lithium-ion batteries: Transition from linear to nonlinear
aging,” J. Power Sources, 2017, doi: 10.1016/j.jpowsour.2017.05.110.
[59] M. Chen et al., “A multilayer electro-thermal model of pouch battery during
normal discharge and internal short circuit process,” Appl. Therm. Eng., 2017, doi:
10.1016/j.applthermaleng.2017.03.135.
[60] Y. Zhao, L. B. Diaz, Y. Patel, T. Zhang, and G. J. Offer, “How to Cool Lithium
Ion Batteries: Optimising Cell Design using a Thermally Coupled Model,” J.
Electrochem. Soc., 2019, doi: 10.1149/2.0501913jes.
[61] N. Damay, C. Forgez, M. P. Bichat, and G. Friedrich, “Thermal modeling of large
199
 prismatic LiFePO4/graphite battery. Coupled thermal and heat generation models
for characterization and simulation,” J. Power Sources, 2015, doi:
10.1016/j.jpowsour.2015.02.091.
[62] W. Allafi et al., “A lumped thermal model of lithium-ion battery cells considering
radiative heat transfer,” Appl. Therm. Eng., 2018, doi:
10.1016/j.applthermaleng.2018.07.105.
[63] J. Esmaeili and H. Jannesari, “Developing heat source term including heat
generation at rest condition for Lithium-ion battery pack by up scaling information
from cell scale,” Energy Convers. Manag., 2017, doi:
10.1016/j.enconman.2017.02.052.
[64] S. Arora, W. Shen, and A. Kapoor, “Neural network based computational model
for estimation of heat generation in LiFePO4 pouch cells of different nominal
capacities,” Comput. Chem. Eng., 2017, doi:
10.1016/j.compchemeng.2017.02.044.
[65] X. Hu, W. Liu, X. Lin, and Y. Xie, “A Comparative Study of Control-Oriented
Thermal Models for Cylindrical Li-Ion Batteries,” IEEE Trans. Transp. Electrif.,
2019, doi: 10.1109/TTE.2019.2953606.
[66] Y. Xie, W. Li, X. Hu, C. Zou, F. Feng, and X. Tang, “Novel Mesoscale
Electrothermal Modeling for Lithium-Ion Batteries,” IEEE Trans. Power
Electron., 2020, doi: 10.1109/TPEL.2019.2927014.
[67] A. Friesen, X. Mönnighoff, M. Börner, J. Haetge, F. M. Schappacher, and M.
Winter, “Influence of temperature on the aging behavior of 18650-type lithium ion
cells: A comprehensive approach combining electrochemical characterization and
post-mortem analysis,” J. Power Sources, 2017, doi:
10.1016/j.jpowsour.2016.12.040.
[68] Y. Fan, Y. Bao, C. Ling, Y. Chu, X. Tan, and S. Yang, “Experimental study on the
thermal management performance of air cooling for high energy density
cylindrical lithium-ion batteries,” Appl. Therm. Eng., 2019, doi:
10.1016/j.applthermaleng.2019.03.157.
[69] L. H. Saw et al., “Novel thermal management system using mist cooling for
lithium-ion battery packs,” Appl. Energy, 2018, doi:
10.1016/j.apenergy.2018.04.042.
[70] B. Liu, S. Yin, and J. Xu, “Integrated computation model of lithium-ion battery
subject to nail penetration,” Appl. Energy, 2016, doi:
10.1016/j.apenergy.2016.08.101.
[71] M. Doyle, “Modeling of Galvanostatic Charge and Discharge of the
Lithium/Polymer/Insertion Cell,” J. Electrochem. Soc., 1993, doi:
10.1149/1.2221597.
[72] Y. Xiao and B. Fahimi, “State-space based multi-nodes thermal model for lithium-

200
 ion battery,” in 2014 IEEE Transportation Electrification Conference and Expo:
Components, Systems, and Power Electronics - From Technology to Business and
Public Policy, ITEC 2014, 2014. doi: 10.1109/itec.2014.6861846.
[73] N. Tian, H. Fang, and Y. Wang, “3-D Temperature Field Reconstruction for a
Lithium-Ion Battery Pack: A Distributed Kalman Filtering Approach,” IEEE
Trans. Control Syst. Technol., 2019, doi: 10.1109/TCST.2017.2779434.
[74] H. Ruan, J. Jiang, B. Sun, W. Gao, L. Wang, and W. Zhang, “Online estimation of
thermal parameters based on a reduced wide-temperature-range electro-thermal
coupled model for lithium-ion batteries,” J. Power Sources, 2018, doi:
10.1016/j.jpowsour.2018.03.075.
[75] W. B. Gu and C. Y. Wang, “Thermal-Electrochemical Modeling of Battery
Systems,” J. Electrochem. Soc., 2000, doi: 10.1149/1.1393625.
[76] K. Kumaresan, G. Sikha, and R. E. White, “Thermal Model for a Li-Ion Cell,” J.
Electrochem. Soc., 2008, doi: 10.1149/1.2817888.
[77] D. W. Dees, V. S. Battaglia, and A. Bélanger, “Electrochemical modeling of
lithium polymer batteries,” J. Power Sources, 2002, doi: 10.1016/s0378-
7753(02)00193-3.
[78] X. Lin et al., “Online parameterization of lumped thermal dynamics in cylindrical
lithium ion batteries for core temperature estimation and health monitoring,” IEEE
Trans. Control Syst. Technol., vol. 21, no. 5, pp. 1745–1755, 2013, doi:
10.1109/TCST.2012.2217143.
[79] J. W. Choi and D. Aurbach, “Promise and reality of post-lithium-ion batteries with
high energy densities,” Nature Reviews Materials. 2016. doi:
10.1038/natrevmats.2016.13.
[80] F. Yang, D. Wang, Y. Zhao, K. L. Tsui, and S. J. Bae, “A study of the relationship
between coulombic efficiency and capacity degradation of commercial lithium-ion
batteries,” Energy, 2018, doi: 10.1016/j.energy.2017.12.144.
[81] Q. Wang, X. Zhao, J. Ye, Q. Sun, P. Ping, and J. Sun, “Thermal response of
lithium-ion battery during charging and discharging under adiabatic conditions,” J.
Therm. Anal. Calorim., 2016, doi: 10.1007/s10973-015-5100-4.
[82] Y. Chen and J. W. Evans, “Heat Transfer Phenomena in Lithium/Polymer‐
Electrolyte Batteries for Electric Vehicle Application,” J. Electrochem. Soc., 1993,
doi: 10.1149/1.2220724.
[83] T. F. Fuller, M. Doyle, and J. Newman, “Relaxation Phenomena in Lithium‐Ion‐
Insertion Cells,” J. Electrochem. Soc., 1994, doi: 10.1149/1.2054868.
[84] C. R. Pals and J. Newman, “Thermal Modeling of the Lithium/Polymer Battery: I .
Discharge Behavior of a Single Cell,” J. Electrochem. Soc., 1995, doi:
10.1149/1.2049974.

201
[85] C. R. Pals and J. Newman, “Thermal Modeling of the Lithium/Polymer Battery: II
. Temperature Profiles in a Cell Stack,” J. Electrochem. Soc., 1995, doi:
10.1149/1.2049975.
[86] P. M. Gomadam, R. E. White, and J. W. Weidner, “Modeling heat conduction in
spiral geometries,” in Proceedings - Electrochemical Society, 2005. doi:
10.1149/1.1605743.
[87] Y. Chen and J. W. Evans, “Three‐Dimensional Thermal Modeling of Lithium‐
Polymer Batteries under Galvanostatic Discharge and Dynamic Power Profile,” J.
Electrochem. Soc., 1994, doi: 10.1149/1.2059263.
[88] J. Newman and W. Tiedemann, “Temperature Rise in a Battery Module with
Constant Heat Generation,” J. Electrochem. Soc., 1995, doi: 10.1149/1.2044130.
[89] Y. Chen and J. W. Evans, “Thermal Analysis of Lithium‐Ion Batteries,” J.
Electrochem. Soc., 1996, doi: 10.1149/1.1837095.
[90] L. Rao and J. Newman, “Heat‐Generation Rate and General Energy Balance for
Insertion Battery Systems,” J. Electrochem. Soc., 1997, doi: 10.1149/1.1837884.
[91] G. H. Kim, A. Pesaran, and R. Spotnitz, “A three-dimensional thermal abuse
model for lithium-ion cells,” J. Power Sources, 2007, doi:
10.1016/j.jpowsour.2007.04.018.
[92] K. E. Thomas, J. Newman, and R. M. Darling, “Mathematical Modeling of
Lithium Batteries,” in Advances in Lithium-Ion Batteries, 2002. doi: 10.1007/0-
306-47508-1_13.
[93] W. van Schalkwijk and B. Scrosati, “Advances in Lithium Ion Batteries
Introduction,” in Advances in Lithium-Ion Batteries, 2002. doi: 10.1007/0-306-
47508-1_1.
[94] K. Smith and C. Y. Wang, “Power and thermal characterization of a lithium-ion
battery pack for hybrid-electric vehicles,” J. Power Sources, 2006, doi:
10.1016/j.jpowsour.2006.01.038.
[95] S. A. Khateeb, S. Amiruddin, M. Farid, J. R. Selman, and S. Al-Hallaj, “Thermal
management of Li-ion battery with phase change material for electric scooters:
Experimental validation,” J. Power Sources, 2005, doi:
10.1016/j.jpowsour.2004.09.033.
[96] W. Fang, O. J. Kwon, and C. Y. Wang, “Electrochemical-thermal modeling of
automotive Li-ion batteries and experimental validation using a three-electrode
cell,” Int. J. Energy Res., 2010, doi: 10.1002/er.1652.
[97] U. S. Kim, J. Yi, C. B. Shin, T. Han, and S. Park, “Modelling the thermal
behaviour of a lithium-ion battery during charge,” J. Power Sources, 2011, doi:
10.1016/j.jpowsour.2011.01.103.
[98] R. E. Gerver and J. P. Meyers, “Three-Dimensional Modeling of Electrochemical
202
 Performance and Heat Generation of Lithium-Ion Batteries in Tabbed Planar
Configurations,” J. Electrochem. Soc., 2011, doi: 10.1149/1.3591799.
[99] J. Zhang, X. G. Yang, F. Sun, Z. Wang, and C. Y. Wang, “An online heat
generation estimation method for lithium-ion batteries using dual-temperature
measurements,” Appl. Energy, vol. 272, no. May, p. 115262, 2020, doi:
10.1016/j.apenergy.2020.115262.
[100] P. Wang, L. Yang, H. Wang, D. M. Tartakovsky, and S. Onori, “Temperature
estimation from current and voltage measurements in lithium-ion battery systems,”
J. Energy Storage, vol. 34, no. September 2020, 2021, doi:
10.1016/j.est.2020.102133.
[101] S. Marelli and M. Corno, “Model-Based Estimation of Lithium Concentrations and
Temperature in Batteries Using Soft-Constrained Dual Unscented Kalman
Filtering,” IEEE Trans. Control Syst. Technol., vol. 29, no. 2, pp. 926–933, 2021,
doi: 10.1109/TCST.2020.2974176.
[102] R. Mahamud and C. Park, “Reciprocating air flow for Li-ion battery thermal
management to improve temperature uniformity,” J. Power Sources, vol. 196, no.
13, pp. 5685–5696, Jul. 2011, doi: 10.1016/j.jpowsour.2011.02.076.
[103] C. Park and A. K. Jaura, “Dynamic thermal model of Li-Ion battery for predictive
behavior in hybrid and fuel cell vehicles,” in SAE Technical Papers, 2003. doi:
10.4271/2003-01-2286.
[104] L. Chen et al., “Core temperature estimation based on electro-thermal model of
lithium-ion batteries,” Int. J. Energy Res., vol. 44, no. 7, pp. 5320–5333, 2020, doi:
10.1002/er.5281.
[105] H. Maleki and A. K. Shamsuri, “Thermal analysis and modeling of a notebook
computer battery,” J. Power Sources, 2003, doi: 10.1016/S0378-7753(02)00722-
X.
[106] X. Lin et al., “A lumped-parameter electro-thermal model for cylindrical
batteries,” J. Power Sources, 2014, doi: 10.1016/j.jpowsour.2014.01.097.
[107] X. Lin, A. G. Stefanopoulou, H. E. Perez, J. B. Siegel, Y. Li, and R. D. Anderson,
“Quadruple adaptive observer of the core temperature in cylindrical Li-ion
batteries and their health monitoring,” Proc. Am. Control Conf., vol. 578, no. 2,
pp. 578–583, 2012, doi: 10.1109/acc.2012.6315386.
[108] H. Dai, L. Zhu, J. Zhu, X. Wei, and Z. Sun, “Adaptive Kalman filtering based
internal temperature estimation with an equivalent electrical network thermal
model for hard-cased batteries,” J. Power Sources, vol. 293, pp. 351–365, 2015,
doi: 10.1016/j.jpowsour.2015.05.087.
[109] D. H. Doughty, P. C. Butler, R. G. Jungst, and E. P. Roth, “Lithium battery
thermal models,” J. Power Sources, 2002, doi: 10.1016/s0378-7753(02)00198-2.
[110] Y. Xie et al., “An Enhanced Online Temperature Estimation for Lithium-Ion
203
 Batteries,” IEEE Trans. Transp. Electrif., vol. 6, no. 2, pp. 375–390, 2020, doi:
10.1109/TTE.2020.2980153.
[111] L. Sun, W. Sun, and F. You, “Core temperature modelling and monitoring of
lithium-ion battery in the presence of sensor bias,” Appl. Energy, vol. 271, no.
May, p. 115243, 2020, doi: 10.1016/j.apenergy.2020.115243.
[112] C. Zhu, Y. Shang, F. Lu, Y. Jiang, C. Cheng, and C. Mi, “Core Temperature
Estimation for Self-Heating Automotive Lithium-Ion Batteries in Cold Climates,”
IEEE Trans. Ind. Informatics, vol. 16, no. 5, pp. 3366–3375, 2020, doi:
10.1109/TII.2019.2960833.
[113] Y. Xiao, “Model-Based Virtual Thermal Sensors for Lithium-Ion Battery in EV
Applications,” IEEE Trans. Ind. Electron., vol. 62, no. 5, pp. 3112–3122, 2015,
doi: 10.1109/TIE.2014.2386793.
[114] Y. Ma, Y. Cui, H. Mou, J. Gao, and H. Chen, “Core temperature estimation of
lithium-ion battery for EVs using Kalman filter,” Appl. Therm. Eng., vol. 168, no.
December 2019, p. 114816, 2020, doi: 10.1016/j.applthermaleng.2019.114816.
[115] N. H. F. Ismail, S. F. Toha, N. A. M. Azubir, N. H. Md Ishak, M. K. Hassan, and
B. S. Ksm Ibrahim, “Simplified heat generation model for lithium ion battery used
in electric vehicle,” IOP Conf. Ser. Mater. Sci. Eng., vol. 53, no. 1, 2013, doi:
10.1088/1757-899X/53/1/012014.
[116] D. Li and L. Yang, “Identification of spatial temperature gradient in large format
lithium battery using a multilayer thermal model,” Int. J. Energy Res., vol. 44, no.
1, pp. 282–297, Jan. 2020, doi: 10.1002/er.4914.
[117] X. Lin et al., “Parameterization and Observability Analysis of Scalable Battery
Clusters for Onboard Thermal Management,” Oil Gas Sci. Technol. – Rev. d’IFP
Energies Nouv., 2013, doi: 10.2516/ogst/2012075.
[118] J. Sun et al., “Online internal temperature estimation for lithium-ion batteries
based on Kalman filter,” Energies, vol. 8, no. 5, pp. 4400–4415, 2015, doi:
10.3390/en8054400.
[119] D. H. Jeon, “Numerical modeling of lithium ion battery for predicting thermal
behavior in a cylindrical cell,” Curr. Appl. Phys., vol. 14, no. 2, pp. 196–205,
2014, doi: 10.1016/j.cap.2013.11.006.
[120] N. Baba, H. Yoshida, M. Nagaoka, C. Okuda, and S. Kawauchi, “Numerical
simulation of thermal behavior of lithium-ion secondary batteries using the
enhanced single particle model,” J. Power Sources, vol. 252, pp. 214–228, 2014,
doi: 10.1016/j.jpowsour.2013.11.111.
[121] S. Du et al., “Study on the thermal behaviors of power lithium iron phosphate
(LFP) aluminum-laminated battery with different tab configurations,” Int. J.
Therm. Sci., vol. 89, pp. 327–336, 2015, doi: 10.1016/j.ijthermalsci.2014.11.018.
[122] J. Yi, J. Lee, C. B. Shin, T. Han, and S. Park, “Modeling of the transient behaviors
204
 of a lithium-ion battery during dynamic cycling,” J. Power Sources, vol. 277, pp.
379–386, 2015, doi: 10.1016/j.jpowsour.2014.12.028.
[123] M. Fleckenstein, O. Bohlen, M. A. Roscher, and B. Bäker, “Current density and
state of charge inhomogeneities in Li-ion battery cells with LiFePO4 as cathode
material due to temperature gradients,” J. Power Sources, vol. 196, no. 10, pp.
4769–4778, 2011, doi: 10.1016/j.jpowsour.2011.01.043.
[124] S. Santhanagopalan, P. Ramadass, and J. (Zhengming) Zhang, “Analysis of
internal short-circuit in a lithium ion cell,” J. Power Sources, 2009, doi:
10.1016/j.jpowsour.2009.05.002.
[125] Z. Li et al., “Examining temporal and spatial variations of internal temperature in
large-format laminated battery with embedded thermocouples,” J. Power Sources,
2013, doi: 10.1016/j.jpowsour.2013.04.117.
[126] M. S. K. Mutyala, J. Zhao, J. Li, H. Pan, C. Yuan, and X. Li, “In-situ temperature
measurement in lithium ion battery by transferable flexible thin film
thermocouples,” J. Power Sources, 2014, doi: 10.1016/j.jpowsour.2014.03.004.
[127] R. Srinivasan, B. G. Carkhuff, M. H. Butler, and A. C. Baisden, “Instantaneous
measurement of the internal temperature in lithium-ion rechargeable cells,”
Electrochim. Acta, 2011, doi: 10.1016/j.electacta.2011.03.136.
[128] R. Srinivasan, “Monitoring dynamic thermal behavior of the carbon anode in a
lithium-ion cell using a four-probe technique,” J. Power Sources, 2012, doi:
10.1016/j.jpowsour.2011.09.077.
[129] J. P. Schmidt, S. Arnold, A. Loges, D. Werner, T. Wetzel, and E. Ivers-Tiffée,
“Measurement of the internal cell temperature via impedance: Evaluation and
application of a new method,” J. Power Sources, 2013, doi:
10.1016/j.jpowsour.2013.06.013.
[130] R. R. Richardson, P. T. Ireland, and D. A. Howey, “Battery internal temperature
estimation by combined impedance and surface temperature measurement,” J.
Power Sources, vol. 265, pp. 254–261, 2014, doi:
10.1016/j.jpowsour.2014.04.129.
[131] J. G. Zhu, Z. C. Sun, X. Z. Wei, and H. F. Dai, “A new lithium-ion battery internal
temperature on-line estimate method based on electrochemical impedance
spectroscopy measurement,” J. Power Sources, vol. 274, pp. 990–1004, 2015, doi:
10.1016/j.jpowsour.2014.10.182.
[132] M. Debert, G. Colin, G. Bloch, and Y. Chamaillard, “An observer looks at the cell
temperature in automotive battery packs,” Control Eng. Pract., 2013, doi:
10.1016/j.conengprac.2013.03.001.
[133] X. Hu, H. Yuan, C. Zou, Z. Li, and L. Zhang, “Co-Estimation of State of Charge
and State of Health for Lithium-Ion Batteries Based on Fractional-Order
Calculus,” IEEE Trans. Veh. Technol., vol. 67, no. 11, pp. 10319–10329, 2018,

205
 doi: 10.1109/TVT.2018.2865664.
[134] Z. Wang, J. Ma, and L. Zhang, “State-of-Health Estimation for Lithium-Ion
Batteries Based on the Multi-Island Genetic Algorithm and the Gaussian Process
Regression,” IEEE Access, 2017, doi: 10.1109/ACCESS.2017.2759094.
[135] A. Hande, “Internal battery temperature estimation using series battery resistance
measurements during cold temperatures,” J. Power Sources, 2006, doi:
10.1016/j.jpowsour.2005.11.027.
[136] D. A. Howey, P. D. Mitcheson, V. Yufit, G. J. Offer, and N. P. Brandon, “Online
measurement of battery impedance using motor controller excitation,” IEEE Trans.
Veh. Technol., 2014, doi: 10.1109/TVT.2013.2293597.
[137] R. R. Richardson and D. A. Howey, “Sensorless Battery Internal Temperature
Estimation Using a Kalman Filter with Impedance Measurement,” IEEE Trans.
Sustain. Energy, vol. 6, no. 4, pp. 1190–1199, 2015, doi:
10.1109/TSTE.2015.2420375.
[138] Y. Li, W. Liu, and J. Zhong, “Comparison of noninvasive and remote temperature
estimation employing magnetic nanoparticles in DC and AC applied fields,” in
2012 IEEE I2MTC - International Instrumentation and Measurement Technology
Conference, Proceedings, 2012. doi: 10.1109/I2MTC.2012.6229437.
[139] J. Zhong, J. Dieckhoff, M. Schilling, and F. Ludwig, “Influence of static magnetic
field strength on the temperature resolution of a magnetic nanoparticle
thermometer,” J. Appl. Phys., 2016, doi: 10.1063/1.4964696.
[140] T. Yoshida and K. Enpuku, “Simulation and quantitative clarification of AC
susceptibility of magnetic fluid in nonlinear Brownian relaxation region,” Jpn. J.
Appl. Phys., 2009, doi: 10.1143/JJAP.48.127002.
[141] J. Zhong, W. Liu, Z. Du, P. César De Morais, Q. Xiang, and Q. Xie, “A
noninvasive, remote and precise method for temperature and concentration
estimation using magnetic nanoparticles,” Nanotechnology, 2012, doi:
10.1088/0957-4484/23/7/075703.
[142] D. Zou et al., “Temperature Estimation of Lithium-Ion Battery Based on an
Improved Magnetic Nanoparticle Thermometer,” IEEE Access, vol. 8, pp.
135491–135498, 2020, doi: 10.1109/ACCESS.2020.3007932.
[143] Z. Liu and H. X. Li, “A spatiotemporal estimation method for temperature
distribution in lithium-ion batteries,” IEEE Trans. Ind. Informatics, vol. 10, no. 4,
pp. 2300–2307, 2014, doi: 10.1109/TII.2014.2341955.
[144] C. Sbarufatti, M. Corbetta, M. Giglio, and F. Cadini, “Adaptive prognosis of
lithium-ion batteries based on the combination of particle filters and radial basis
function neural networks,” J. Power Sources, vol. 344, no. April 2000, pp. 128–
140, 2017, doi: 10.1016/j.jpowsour.2017.01.105.
[145] K. Liu, K. Li, Q. Peng, Y. Guo, and L. Zhang, “Data-driven hybrid internal
206
 temperature estimation approach for battery thermal management,” Complexity,
2018, doi: 10.1155/2018/9642892.
[146] F. Feng et al., “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, doi: 10.1016/j.jpowsour.2020.227935.
[147] M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural
networks: A deep learning framework for solving forward and inverse problems
involving nonlinear partial differential equations,” J. Comput. Phys., 2019, doi:
10.1016/j.jcp.2018.10.045.
[148] F. Wang, Z. Zhai, Z. Zhao, Y. Di, and X. Chen, “Physics-informed neural network
for lithium-ion battery degradation stable modeling and prognosis,” Nat.
Commun., vol. 15, no. 1, p. 4332, May 2024, doi: 10.1038/s41467-024-48779-z.
[149] Y. Wang et al., “Temperature state prediction for lithium-ion batteries based on
improved physics informed neural networks,” J. Energy Storage, 2023, doi:
10.1016/j.est.2023.108863.
[150] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang,
“Physics-informed machine learning,” Nat. Rev. Phys., vol. 3, no. 6, pp. 422–440,
May 2021, doi: 10.1038/s42254-021-00314-5.
[151] Q. Xu, M. Ma, T. Jiang, Q. Hu, and Z. Wang, “A Data-Driven Surface
Temperature Estimation Method for Lithium-ion Battery Based on GRU-RNN
with Attention Mechanism,” in PEAS 2023 - 2023 IEEE 2nd International Power
Electronics and Application Symposium, Conference Proceedings, 2023. doi:
10.1109/PEAS58692.2023.10394755.
[152] L. Spitthoff, P. R. Shearing, and O. S. Burheim, “Temperature, ageing and thermal
management of lithium-ion batteries,” Energies. 2021. doi: 10.3390/en14051248.
[153] M. Raissi, A. Yazdani, and G. E. Karniadakis, “Hidden fluid mechanics: Learning
velocity and pressure fields from flow visualizations,” Science (80-. )., 2020, doi:
10.1126/science.aaw4741.
[154] L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis, “DeepXDE: A deep learning
library for solving differential equations,” SIAM Rev., 2021, doi:
10.1137/19M1274067.
[155] K. K. Xu, H. X. Li, and H. D. Yang, “Local-Properties-Embedding-Based
Nonlinear Spatiotemporal Modeling for Lithium-Ion Battery Thermal Process,”
IEEE Trans. Ind. Electron., 2018, doi: 10.1109/TIE.2018.2818645.
[156] J. Tian, C. Chen, W. Shen, F. Sun, and R. Xiong, “Deep Learning Framework for
Lithium-ion Battery State of Charge Estimation: Recent Advances and Future
Perspectives,” Energy Storage Mater., 2023, doi: 10.1016/j.ensm.2023.102883.
[157] K. Li, Y. Wang, and Z. Chen, “A comparative study of battery state-of-health
estimation based on empirical mode decomposition and neural network,” J. Energy
207
 Storage, 2022, doi: 10.1016/j.est.2022.105333.
[158] J. Kleiner, M. Stuckenberger, L. Komsiyska, and C. Endisch, “Real-time core
temperature prediction of prismatic automotive lithium-ion battery cells based on
artificial neural networks,” J. Energy Storage, 2021, doi:
10.1016/j.est.2021.102588.
[159] Z. Wei et al., “Machine learning-based hybrid thermal modeling and diagnostic for
lithium-ion battery enabled by embedded sensing,” Appl. Therm. Eng., 2022, doi:
10.1016/j.applthermaleng.2022.119059.
[160] N. Wang et al., “Core Temperature Estimation Method for Lithium-Ion Battery
Based on Long Short-Term Memory Model With Transfer Learning,” IEEE J.
Emerg. Sel. Top. Power Electron., 2023, doi: 10.1109/JESTPE.2021.3136906.
[161] K. Cho et al., “Learning phrase representations using RNN encoder-decoder for
statistical machine translation,” in EMNLP 2014 - 2014 Conference on Empirical
Methods in Natural Language Processing, Proceedings of the Conference, 2014.
doi: 10.3115/v1/d14-1179.
[162] C. S. Chin, Z. Gao, and C. Z. Zhang, “Comprehensive electro-thermal model of
26650 lithium battery for discharge cycle under parametric and temperature
variations,” J. Energy Storage, 2020, doi: 10.1016/j.est.2020.101222.
[163] A. Yuan, T. Cai, H. Luo, Z. Song, and B. Wei, “Core temperature estimation of
lithium-ion battery based on numerical model fusion deep learning,” J. Energy
Storage, vol. 102, p. 114148, Nov. 2024, doi: 10.1016/j.est.2024.114148.
[164] X. Zhang, H. Xiang, X. Xiong, Y. Wang, and Z. Chen, “Benchmarking core
temperature forecasting for lithium-ion battery using typical recurrent neural
networks,” Appl. Therm. Eng., vol. 248, p. 123257, Jul. 2024, doi:
10.1016/j.applthermaleng.2024.123257.
[165] T. J. Kim, B. D. Youn, and H. J. Kim, “Battery pack temperature estimation model
for EVs and its semi-transient case study,” in Chemical Engineering Transactions,
2013. doi: 10.3303/CET1333160.
[166] S. Zhu, C. He, N. Zhao, and J. Sha, “Data-driven analysis on thermal effects and
temperature changes of lithium-ion battery,” J. Power Sources, 2021, doi:
10.1016/j.jpowsour.2020.228983.
[167] J. H. Kim, N. C. Seong, and W. Choi, “Cooling load forecasting via predictive
optimization of a nonlinear autoregressive exogenous (NARX) Neural network
model,” Sustain., 2019, doi: 10.3390/su11236535.
[168] A. A. Hussein and A. A. Chehade, “Robust Artificial Neural Network-Based
Models for Accurate Surface Temperature Estimation of Batteries,” in IEEE
Transactions on Industry Applications, 2020. doi: 10.1109/TIA.2020.3001256.
[169] Y. Liu, Z. Huang, Y. Wu, L. Yan, F. Jiang, and J. Peng, “An online hybrid
estimation method for core temperature of Lithium-ion battery with model noise
208
 compensation,” Appl. Energy, 2022, doi: 10.1016/j.apenergy.2022.120037.
[170] B. Fan, B. Zhang, Y. Shi, and Y. Chang, “A high-fidelity lithium-ion battery
emulator for electric vehicle application,” Sci. Rep., vol. 14, no. 1, p. 19742, Aug.
2024, doi: 10.1038/s41598-024-70445-z.
[171] R. Mahamud and C. Park, “Reciprocating air flow for Li-ion battery thermal
management to improve temperature uniformity,” J. Power Sources, 2011, doi:
10.1016/j.jpowsour.2011.02.076.
[172] P. T. Coman, E. C. Darcy, B. Strangways, and R. E. White, “A Reduced-Order
Lumped Model for Li-Ion Battery Packs during Operation,” J. Electrochem. Soc.,
2021, doi: 10.1149/1945-7111/ac2dcb.
[173] R. R. Richardson, S. Zhao, and D. A. Howey, “On-board monitoring of 2-D
spatially-resolved temperatures in cylindrical lithium-ion batteries: Part II. State
estimation via impedance-based temperature sensing,” J. Power Sources, 2016,
doi: 10.1016/j.jpowsour.2016.06.104.
[174] G. Guo, B. Long, B. Cheng, S. Zhou, P. Xu, and B. Cao, “Three-dimensional
thermal finite element modeling of lithium-ion battery in thermal abuse
application,” J. Power Sources, 2010, doi: 10.1016/j.jpowsour.2009.10.090.
[175] Y. Zheng et al., “Real-Time Sensorless Temperature Estimation of Lithium-Ion
Batteries Based on Online Operando Impedance Acquisition,” IEEE Trans. Power
Electron., vol. 39, no. 10, pp. 13853–13868, Oct. 2024, doi:
10.1109/TPEL.2024.3424267.
[176] A. Samanta, A. Huynh, E. Rutovic, and S. Williamson, “Rapid Thermal Modeling
and Discharge Characterization for Accurate Lithium-ion Battery Core
Temperature Estimation,” in IECON Proceedings (Industrial Electronics
Conference), 2022. doi: 10.1109/IECON49645.2022.9968451.
[177] P. Wang et al., “Real-time monitoring of internal temperature evolution of the
lithium-ion coin cell battery during the charge and discharge process,” Extrem.
Mech. Lett., vol. 9, pp. 459–466, 2016, doi: 10.1016/j.eml.2016.03.013.
[178] C. Y. Lee et al., “Integrated microsensor for real-time microscopic monitoring of
local temperature, voltage and current inside lithium ion battery,” Sensors
Actuators, A Phys., 2017, doi: 10.1016/j.sna.2016.10.011.
[179] Z. Xiao et al., “Review—Online Monitoring of Internal Temperature in Lithium-
Ion Batteries,” J. Electrochem. Soc., 2023, doi: 10.1149/1945-7111/acd586.
[180] H. Dai, B. Jiang, X. Hu, X. Lin, X. Wei, and M. Pecht, “Advanced battery
management strategies for a sustainable energy future: Multilayer design concepts
and research trends,” Renew. Sustain. Energy Rev., vol. 138, Mar. 2021, doi:
10.1016/J.RSER.2020.110480.
[181] Y. Wang et al., “A comprehensive review of battery modeling and state estimation
approaches for advanced battery management systems,” Renew. Sustain. Energy
209
 Rev., vol. 131, no. June, p. 110015, 2020, doi: 10.1016/j.rser.2020.110015.
[182] M. Syafrudin, G. Alfian, N. L. Fitriyani, and J. Rhee, “Performance analysis of
IoT-based sensor, big data processing, and machine learning model for real-time
monitoring system in automotive manufacturing,” Sensors (Switzerland), vol. 18,
no. 9, 2018, doi: 10.3390/s18092946.
[183] C. M. Ezhilarasu, Z. Skaf, and I. K. Jennions, “Understanding the role of a digital
twin in integrated vehicle health management (IVHM),” Conf. Proc. - IEEE Int.
Conf. Syst. Man Cybern., vol. 2019-Octob, no. October, pp. 1484–1491, 2019, doi:
10.1109/SMC.2019.8914244.
[184] E. H. Glaessgen and D. S. Stargel, “The digital twin paradigm for future NASA
and U.S. Air force vehicles,” in Collection of Technical Papers -
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials
Conference, 2012, pp. 1–14. doi: 10.2514/6.2012-1818.
[185] M. Grieves, “Digital Twin : Manufacturing Excellence through Virtual Factory
Replication,” White Pap., 2014.
[186] S. Singh, M. Weeber, and K. P. Birke, “Advancing digital twin implementation: A
toolbox for modelling and simulation,” in Procedia CIRP, 2021. doi:
10.1016/j.procir.2021.03.078.
[187] C. Semeraro, M. Lezoche, H. Panetto, and M. Dassisti, “Digital twin paradigm: A
systematic literature review,” Comput. Ind., 2021, doi:
10.1016/j.compind.2021.103469.
[188] G. Bhatti, H. Mohan, and R. Raja Singh, “Towards the future of ssmart electric
vehicles: Digital twin technology,” Renew. Sustain. Energy Rev., vol. 141, no.
February, p. 110801, 2021, doi: 10.1016/j.rser.2021.110801.
[189] A. Fuller, Z. Fan, C. Day, and C. Barlow, “Digital Twin: Enabling Technologies,
Challenges and Open Research,” IEEE Access, 2020, doi:
10.1109/ACCESS.2020.2998358.
[190] A. Martínez‐gutiérrez, J. Díez‐gonzález, R. Ferrero‐guillén, P. Verde, R. Álvarez,
and H. Perez, “Digital twin for automatic transportation in industry 4.0,” Sensors,
2021, doi: 10.3390/s21103344.
[191] M. Gonzalez, O. Salgado, J. Croes, B. Pluymers, and W. Desmet, “A Digital Twin
for Operational Evaluation of Vertical Transportation Systems,” IEEE Access,
2020, doi: 10.1109/ACCESS.2020.3001686.
[192] S. Liao, J. Wu, A. K. Bashir, W. Yang, J. Li, and U. Tariq, “Digital Twin
Consensus for Blockchain-Enabled Intelligent Transportation Systems in Smart
Cities,” IEEE Trans. Intell. Transp. Syst., 2022, doi: 10.1109/TITS.2021.3134002.
[193] Y. Gao, S. Qian, Z. Li, P. Wang, F. Wang, and Q. He, “Digital twin and its
application in transportation infrastructure,” in Proceedings 2021 IEEE 1st
International Conference on Digital Twins and Parallel Intelligence, DTPI 2021,
210
 2021. doi: 10.1109/DTPI52967.2021.9540108.
[194] W. Hofmann and F. Branding, “Implementation of an IoT- And cloud-based
digital twin for real-time decision support in port operations,” in IFAC-
PapersOnLine, 2019. doi: 10.1016/j.ifacol.2019.11.516.
[195] T. Y. Melesse, V. Di Pasquale, and S. Riemma, “Digital Twin models in industrial
operations: State-of-the-art and future research directions,” IET Collaborative
Intelligent Manufacturing. 2021. doi: 10.1049/cim2.12010.
[196] G. Shao and M. Helu, “Framework for a digital twin in manufacturing: Scope and
requirements,” Manuf. Lett., 2020, doi: 10.1016/j.mfglet.2020.04.004.
[197] C. Zhang, G. Zhou, H. Li, and Y. Cao, “Manufacturing Blockchain of Things for
the Configuration of a Data- And Knowledge-Driven Digital Twin Manufacturing
Cell,” IEEE Internet Things J., 2020, doi: 10.1109/JIOT.2020.3005729.
[198] Y. Wang, S. Wang, B. Yang, L. Zhu, and F. Liu, “Big data driven Hierarchical
Digital Twin Predictive Remanufacturing paradigm: Architecture, control
mechanism, application scenario and benefits,” J. Clean. Prod., 2020, doi:
10.1016/j.jclepro.2019.119299.
[199] Z. Wang and B. Lin, “A Digital Twin enabled Maritime Networking Architecture,”
in 2022 IEEE 95th Vehicular Technology Conference: (VTC2022-Spring), IEEE,
Jun. 2022, pp. 1–5. doi: 10.1109/VTC2022-Spring54318.2022.9860929.
[200] S. Liu, J. Bao, Y. Lu, J. Li, S. Lu, and X. Sun, “Digital twin modeling method
based on biomimicry for machining aerospace components,” J. Manuf. Syst., 2021,
doi: 10.1016/j.jmsy.2020.04.014.
[201] L. Li, S. Aslam, A. Wileman, and S. Perinpanayagam, “Digital Twin in Aerospace
Industry: A Gentle Introduction,” IEEE Access, 2022, doi:
10.1109/ACCESS.2021.3136458.
[202] F. G. Medina, A. W. Umpierrez, V. Martinez, and H. Fromm, “A Maturity Model
for Digital Twin Implementations in the Commercial Aerospace OEM Industry,”
in Proceedings - 2021 10th International Conference on Industrial Technology
and Management, ICITM 2021, 2021. doi: 10.1109/ICITM52822.2021.00034.
[203] Y. Wang, R. Xu, C. Zhou, X. Kang, and Z. Chen, “Digital twin and cloud-side-end
collaboration for intelligent battery management system,” J. Manuf. Syst., vol. 62,
no. August 2021, pp. 124–134, 2022, doi: 10.1016/j.jmsy.2021.11.006.
[204] A. Redelinghuys, A. Basson, and K. Kruger, “A six-layer digital twin architecture
for a manufacturing cell,” in Studies in Computational Intelligence, 2019. doi:
10.1007/978-3-030-03003-2_32.
[205] J. Lee, B. Bagheri, and H. A. Kao, “A Cyber-Physical Systems architecture for
Industry 4.0-based manufacturing systems,” Manuf. Lett., 2015, doi:
10.1016/j.mfglet.2014.12.001.

211
[206] S. Yang et al., “Implementation for a cloud battery management system based on
the CHAIN framework,” Energy AI, vol. 5, p. 100088, 2021, doi:
10.1016/j.egyai.2021.100088.
[207] A. P. Das, K. Thomas, A. Mohan, A. S. Akhil, and K. P. Ganesh, “A Battery
Digital Twin framework for Predictive Maintenance and State of Health
Estimation of Electric Vehicles,” Eliv 2021, no. 2384, pp. 579–592, 2021, doi:
10.51202/9783181023846-579.
[208] Y. Wang, R. Xu, C. Zhou, X. Kang, and Z. Chen, “Digital twin and cloud-side-end
collaboration for intelligent battery management system,” J. Manuf. Syst., 2022,
doi: 10.1016/j.jmsy.2021.11.006.
[209] H. Tang, Y. Wu, Y. Cai, F. Wang, Z. Lin, and Y. Pei, “Design of power lithium
battery management system based on digital twin,” J. Energy Storage, vol. 47, no.
November 2021, p. 103679, 2022, doi: 10.1016/j.est.2021.103679.
[210] B. Wu, W. D. Widanage, S. Yang, and X. Liu, “Battery digital twins: Perspectives
on the fusion of models, data and artificial intelligence for smart battery
management systems,” Energy AI, vol. 1, p. 100016, 2020, doi:
10.1016/j.egyai.2020.100016.
[211] H. Zhang et al., “Digital Twin Enables Rational Design of Ultrahigh‐Power
Lithium‐Ion Batteries,” Adv. Energy Mater., vol. 13, no. 1, Jan. 2023, doi:
10.1002/aenm.202202660.
[212] G. Yu, X. Ye, X. Xia, and Y. Chen, “Towards Cognitive EV Charging Stations
Enabled by Digital Twin and Parallel Intelligence,” pp. 2021–2024, 2021.
[213] Y. Wang, X. Kang, and Z. Chen, “A survey of Digital Twin techniques in smart
manufacturing and management of energy applications,” Green Energy Intell.
Transp., vol. 1, no. 2, p. 100014, 2022, doi: 10.1016/j.geits.2022.100014.
[214] S. J. An, J. Li, C. Daniel, S. Kalnaus, and D. L. Wood, “Design and Demonstration
of Three-Electrode Pouch Cells for Lithium-Ion Batteries,” J. Electrochem. Soc.,
2017, doi: 10.1149/2.0031709jes.
[215] L. Patnaik and S. Williamson, “A Five-Parameter Analytical Curvefit Model for
Open-Circuit Voltage Variation with State-of-Charge of a Rechargeable Battery,”
in Proceedings of 2018 IEEE International Conference on Power Electronics,
Drives and Energy Systems, PEDES 2018, 2018. doi:
10.1109/PEDES.2018.8707806.
[216] N. T. Tran, T. Farrell, M. Vilathgamuwa, S. S. Choi, and Y. Li, “A
Computationally Efficient Coupled Electrochemical-Thermal Model for Large
Format Cylindrical Lithium Ion Batteries,” J. Electrochem. Soc., 2019, doi:
10.1149/2.1241913jes.
[217] A. Subramaniam, S. Kolluri, S. Santhanagopalan, and V. R. Subramanian, “An
Efficient Electrochemical-Thermal Tanks-in-Series Model for Lithium-Ion

212
 Batteries,” J. Electrochem. Soc., 2020, doi: 10.1149/1945-7111/aba700.
[218] C. Lin, Q. Yu, R. Xiong, and L. Y. Wang, “A study on the impact of open circuit
voltage tests on state of charge estimation for lithium-ion batteries,” Appl. Energy,
2017, doi: 10.1016/j.apenergy.2017.08.124.
[219] S. Surya, J. Channegowda, S. D. Datar, A. S. Jha, and A. Victor, “Accurate battery
modeling based on pulse charging using MATLAB / Simulink,” in 9th IEEE
International Conference on Power Electronics, Drives and Energy Systems,
PEDES 2020, 2020. doi: 10.1109/PEDES49360.2020.9379617.
[220] T. S. Bryden et al., “Methodology to determine the heat capacity of lithium-ion
cells,” J. Power Sources, 2018, doi: 10.1016/j.jpowsour.2018.05.084.
[221] A. Greco, D. Cao, X. Jiang, and H. Yang, “A theoretical and computational study
of lithium-ion battery thermal management for electric vehicles using heat pipes,”
J. Power Sources, 2014, doi: 10.1016/j.jpowsour.2014.02.004.
[222] Z. Rao, S. Wang, and G. Zhang, “Simulation and experiment of thermal energy
management with phase change material for ageing LiFePO4 power battery,”
Energy Convers. Manag., 2011, doi: 10.1016/j.enconman.2011.07.009.
[223] Y. Xie, S. Shi, J. Tang, H. Wu, and J. Yu, “Experimental and analytical study on
heat generation characteristics of a lithium-ion power battery,” Int. J. Heat Mass
Transf., 2018, doi: 10.1016/j.ijheatmasstransfer.2018.02.038.
[224] S. Surya, V. Marcis, and S. Williamson, “Core Temperature Estimation for a
Lithium ion 18650 Cell,” Energies, vol. 14, no. 1, p. 87, 2020, doi:
10.3390/en14010087.
[225] S. Surya, A. Bhesaniya, A. Gogate, R. Ankur, and V. Patil, “Development of
thermal model for estimation of core temperature of batteries,” Int. J. Emerg.
Electr. Power Syst., 2020, doi: 10.1515/ijeeps-2020-0070.
[226] S. Surya and V. Patil, “Cuk Converter as an Efficient Driver for LED,” in 4th
International Conference on Electrical, Electronics, Communication, Computer
Technologies and Optimization Techniques, ICEECCOT 2019, 2019. doi:
10.1109/ICEECCOT46775.2019.9114648.
[227] G. Plett, Battery Management Systems, Volume II: Equivalent-Circuit Methods.
Artech, 2015.
[228] N. Kalchbrenner, I. Danihelka, and A. Graves, “Grid long short-term memory,” in
4th International Conference on Learning Representations, ICLR 2016 -
Conference Track Proceedings, 2016.
[229] Z. Hu, J. Zhang, and Y. Ge, “Handling Vanishing Gradient Problem Using
Artificial Derivative,” IEEE Access, 2021, doi: 10.1109/ACCESS.2021.3054915.
[230] M. Roodschild, J. Gotay Sardiñas, and A. Will, “A new approach for the vanishing
gradient problem on sigmoid activation,” Prog. Artif. Intell., 2020, doi:
213
 10.1007/s13748-020-00218-y.
[231] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Comput.,
1997, doi: 10.1162/neco.1997.9.8.1735.
[232] Y. Liu et al., “Deep C-LSTM Neural Network for Epileptic Seizure and Tumor
Detection Using High-Dimension EEG Signals,” IEEE Access, 2020, doi:
10.1109/ACCESS.2020.2976156.
[233] B. Majhi, D. Naidu, A. P. Mishra, and S. C. Satapathy, “Improved prediction of
daily pan evaporation using Deep-LSTM model,” Neural Comput. Appl., 2020,
doi: 10.1007/s00521-019-04127-7.
[234] Y. Zhang, G. Chen, D. Yu, K. Yaco, S. Khudanpur, and J. Glass, “Highway long
short-term memory RNNS for distant speech recognition,” in ICASSP, IEEE
International Conference on Acoustics, Speech and Signal Processing -
Proceedings, 2016. doi: 10.1109/ICASSP.2016.7472780.
[235] W. N. Hsu, Y. Zhang, and J. Glass, “A prioritized grid long short-term memory
RNN for speech recognition,” in 2016 IEEE Workshop on Spoken Language
Technology, SLT 2016 - Proceedings, 2017. doi: 10.1109/SLT.2016.7846305.
[236] S. Siami-Namini, N. Tavakoli, and A. S. Namin, “The Performance of LSTM and
BiLSTM in Forecasting Time Series,” in Proceedings - 2019 IEEE International
Conference on Big Data, Big Data 2019, 2019. doi:
10.1109/BigData47090.2019.9005997.
[237] B. Lindemann, B. Maschler, N. Sahlab, and M. Weyrich, “A survey on anomaly
detection for technical systems using LSTM networks,” Computers in Industry.
2021. doi: 10.1016/j.compind.2021.103498.
[238] C. N. Van and D. T. Quang, “Estimation of SoH and internal resistances of
Lithium ion battery based on LSTM network,” Int. J. Electrochem. Sci., 2023, doi:
10.1016/J.IJOES.2023.100166.
[239] P. Hu, W. F. Tang, C. H. Li, S. L. Mak, C. Y. Li, and C. C. Lee, “Joint State of
Charge (SOC) and State of Health (SOH) Estimation for Lithium-Ion Batteries
Packs of Electric Vehicles Based on NSSR-LSTM Neural Network,” Energies,
2023, doi: 10.3390/en16145313.
[240] A. N. Kolmogorov, “On the representation of continuous functions of many
variables by superposition of continuous functions of one variable and addition,”
1963. doi: 10.1090/trans2/028/04.
[241] S. Haykin, “Neural networks: a comprehensive foundation by Simon Haykin,” The
Knowledge Engineering Review. 1999.
[242] M. T. Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James
Halverson, Marin Soljačić, Thomas Y. Hou, “KAN: Kolmogorov-Arnold
Networks,” 2024. doi: https://doi.org/10.48550/arXiv.2404.19756.

214
[243] D. Shi et al., “Cloud-Based Deep Learning for Co-Estimation of Battery State of
Charge and State of Health,” Energies, 2023, doi: 10.3390/en16093855.
[244] R. Di Rienzo et al., “Cloud-Based Optimization of a Battery Model Parameter
Identification Algorithm for Battery State-of-Health Estimation in Electric
Vehicles,” Batteries, 2023, doi: 10.3390/batteries9100486.
[245] S. Li, H. He, Z. Wei, and P. Zhao, “Edge computing for vehicle battery
management: Cloud-based online state estimation,” J. Energy Storage, 2022, doi:
10.1016/j.est.2022.105502.
[246] L. Merkle, M. Pöthig, and F. Schmid, “Estimate e-golf battery state using
diagnostic data and a digital twin,” Batteries, vol. 7, no. 1, pp. 1–22, 2021, doi:
10.3390/batteries7010015.
[247] S. Yang et al., “Implementation for a cloud battery management system based on
the CHAIN framework,” Energy AI, 2021, doi: 10.1016/j.egyai.2021.100088.
[248] W. Li, M. Rentemeister, J. Badeda, D. Jöst, D. Schulte, and D. U. Sauer, “Digital
twin for battery systems: Cloud battery management system with online state-of-
charge and state-of-health estimation,” J. Energy Storage, vol. 30, no. June, p.
101557, 2020, doi: 10.1016/j.est.2020.101557.
[249] M. K. Tran, S. Panchal, T. D. Khang, K. Panchal, R. Fraser, and M. Fowler,
“Concept Review of a Cloud-Based Smart Battery Management System for
Lithium-Ion Batteries: Feasibility, Logistics, and Functionality,” Batteries. 2022.
doi: 10.3390/batteries8020019.
[250] M. A. M. Yassin, A. Shrestha, and S. Rabie, “Digital twin in power system
research and development: Principle, scope, and challenges,” Energy Reviews.
2023. doi: 10.1016/j.enrev.2023.100039.
[251] G. Pérez, M. Garmendia, J. F. Reynaud, J. Crego, and U. Viscarret, “Enhanced
closed loop State of Charge estimator for lithium-ion batteries based on Extended
Kalman Filter,” Appl. Energy, 2015, doi: 10.1016/j.apenergy.2015.06.063.
[252] Z. Chen, Y. Fu, and C. C. Mi, “State of charge estimation of lithium-ion batteries
in electric drive vehicles using extended Kalman filtering,” IEEE Trans. Veh.
Technol., 2013, doi: 10.1109/TVT.2012.2235474.
[253] A. Wadi, M. Abdel-Hafez, H. A. Hashim, and A. A. Hussein, “An Invariant
Method for Electric Vehicle Battery State-of-Charge Estimation Under Dynamic
Drive Cycles,” IEEE Access, 2023, doi: 10.1109/ACCESS.2023.3237972.
[254] S. Guo and L. Ma, “A comparative study of different deep learning algorithms for
lithium-ion batteries on state-of-charge estimation,” Energy, 2023, doi:
10.1016/j.energy.2022.125872.
[255] F. Yang, S. Zhang, W. Li, and Q. Miao, “State-of-charge estimation of lithium-ion
batteries using LSTM and UKF,” Energy, 2020, doi:
10.1016/j.energy.2020.117664.
215
[256] Z. Zhang et al., “An Improved Bidirectional Gated Recurrent Unit Method for
Accurate State-of-Charge Estimation,” IEEE Access, vol. 9, pp. 11252–11263,
2021, doi: 10.1109/ACCESS.2021.3049944.
[257] M. Jiao, D. Wang, and J. Qiu, “A GRU-RNN based momentum optimized
algorithm for SOC estimation,” J. Power Sources, 2020, doi:
10.1016/j.jpowsour.2020.228051.
[258] K. P. S. Rana, V. Kumar, A. K. Dagar, A. Chandel, and A. Kataria, “FPGA
Implementation of Steinhart-Hart Equation for Accurate Thermistor
Linearization,” IEEE Sens. J., 2018, doi: 10.1109/JSEN.2018.2795098.
[259] K. H. Johansson, M. Törngren, and L. Nielsen, “Vehicle Applications of
Controller Area Network,” in Handbook of Networked and Embedded Control
Systems, 2005. doi: 10.1007/0-8176-4404-0_32.
[260] D. Karnehm et al., “Universal Data Specification and Real-time Data Streaming
Architecture for Cloud-based Battery Management Systems,” IEEE J. Emerg. Sel.
Top. Power Electron., pp. 1–1, 2024, doi: 10.1109/JESTPE.2024.3413163.
[261] T. Kim, D. Makwana, A. Adhikaree, J. S. Vagdoda, and Y. Lee, “Cloud-based
battery condition monitoring and fault diagnosis platform for large-scale lithium-
ion battery energy storage systems,” Energies, 2018, doi: 10.3390/en11010125.
[262] H. Tang, Y. Wu, Y. Cai, F. Wang, Z. Lin, and Y. Pei, “Design of power lithium
battery management system based on digital twin,” J. Energy Storage, 2022, doi:
10.1016/j.est.2021.103679.
[263] W. Li, M. Rentemeister, J. Badeda, D. Jöst, D. Schulte, and D. U. Sauer, “Digital
twin for battery systems: Cloud battery management system with online state-of-
charge and state-of-health estimation,” J. Energy Storage, 2020, doi:
10.1016/j.est.2020.101557.
[264] G. Kour and R. Perveen, “Battery Management System in Electric Vehicle,” in
AIP Conference Proceedings, 2023. doi: 10.1063/5.0123645.
[265] D. Karnehm, L. Anekal, A. Samanta, A. Neve, and S. Williamson, “Towards
Digital Twining of Lithium-ion Battery Management Systems: An Extended
Kalman Filter for State-of-Charge Estimation in Cloud-platform,” in 2023 IEEE
Energy Conversion Congress and Exposition, ECCE 2023, 2023. doi:
10.1109/ECCE53617.2023.10362099.
[266] T. Joshi, S. Azam, C. Lopez, S. Kinyon, and J. Jeevarajan, “Safety of Lithium-Ion
Cells and Batteries at Different States-of-Charge,” J. Electrochem. Soc., 2020, doi:
10.1149/1945-7111/abc8c4.
[267] T. Waldmann, R. G. Scurtu, K. Richter, and M. Wohlfahrt-Mehrens, “18650 vs.
21700 Li-ion cells – A direct comparison of electrochemical, thermal, and
geometrical properties,” J. Power Sources, 2020, doi:
10.1016/j.jpowsour.2020.228614.

216
[268] J. B. Quinn, T. Waldmann, K. Richter, M. Kasper, and M. Wohlfahrt-Mehrens,
“Energy Density of Cylindrical Li-Ion Cells: A Comparison of Commercial 18650
to the 21700 Cells,” J. Electrochem. Soc., 2018, doi: 10.1149/2.0281814jes.
[269] B. Gulsoy, T. A. Vincent, J. E. H. Sansom, and J. Marco, “In-situ temperature
monitoring of a lithium-ion battery using an embedded thermocouple for smart
battery applications,” J. Energy Storage, 2022, doi: 10.1016/j.est.2022.105260.
[270] L. Patnaik, A. V. J. S. Praneeth, and S. S. Williamson, “A Closed-Loop Constant-
Temperature Constant-Voltage Charging Technique to Reduce Charge Time of
Lithium-Ion Batteries,” IEEE Trans. Ind. Electron., 2019, doi:
10.1109/TIE.2018.2833038.
[271] H. Li, M. Bin Kaleem, I. J. Chiu, D. Gao, and J. Peng, “A Digital Twin Model for
the Battery Management Systems of Electric Vehicles,” 2021 IEEE 23rd Int. Conf.
High Perform. Comput. Commun. 7th Int. Conf. Data Sci. Syst. 19th Int. Conf.
Smart City 7th Int. Conf. Dependability Sensor, Cl, pp. 1100–1107, 2022, doi:
10.1109/HPCC-DSS-SmartCity-DependSys53884.2021.00171.
[272] W. Wang, J. Wang, J. Tian, J. Lu, and R. Xiong, “Application of Digital Twin in
Smart Battery Management Systems,” Chinese J. Mech. Eng., 2021, doi:
10.1186/s10033-021-00577-0.
[273] N. G. Panwar, S. Singh, A. Garg, A. K. Gupta, and L. Gao, “Recent Advancements
in Battery Management System for Li-Ion Batteries of Electric Vehicles: Future
Role of Digital Twin, Cyber-Physical Systems, Battery Swapping Technology, and
Nondestructive Testing,” Energy Technology, vol. 9, no. 8. 2021. doi:
10.1002/ente.202000984.
[274] L. D. Couto et al., “Faster and Healthier Charging of Lithium-Ion Batteries via
Constrained Feedback Control,” IEEE Trans. Control Syst. Technol., 2022, doi:
10.1109/TCST.2021.3135149.
[275] O. Lizaso-Eguileta, E. Martinez-Laserna, M. Rivas, E. Miguel, U. Iraola, and I.
Cantero, “Module-Level Modelling Approach for a Cloudbased Digital Twin
Platform for Li-Ion Batteries,” in 2021 IEEE Vehicle Power and Propulsion
Conference, VPPC 2021 - ProceedingS, 2021. doi:
10.1109/VPPC53923.2021.9699271.
[276] A. Samanta, D. Karnehm, A. Neve, and S. Williamson, “Data-informed Healthy
Cell Clustering Technique for Second-life Applications of Retired Electric Vehicle
Batteries,” in 2024 IEEE Transportation Electrification Conference and Expo
(ITEC), IEEE, Jun. 2024, pp. 1–5. doi: 10.1109/ITEC60657.2024.10598850.
[277] A. Samanta, D. Karnehm, A. Neve, and S. Williamson, “Voltage Relaxation
Pattern Recognition for Efficient Sorting of Healthy Cells for Second-life
Applications of Retired Electric Vehicle Batteries,” in 2024 ICIT – 25th IEEE
International Conference on Industrial Technology, Bristol, UK: IEEE, 2024.
[278] P. Kollmeyer and M. Skells, “Turnigy Graphene 5000mAh 65C Li-ion Battery
217
 DataNo Title,” Mendeley Data, V1, vol. V1, 2020, doi: DOI:
10.17632/4fx8cjprxm.1.
[279] F. Khanum, E. Louback, F. Duperly, C. Jenkins, P. J. Kollmeyer, and A. Emadi,
“A Kalman filter based battery state of charge estimation MATLAB function,” in
2021 IEEE Transportation Electrification Conference and Expo, ITEC 2021, 2021.
doi: 10.1109/ITEC51675.2021.9490163.
[280] Z. Wei, C. Zou, F. Leng, B. H. Soong, and K. J. Tseng, “Online model
identification and state-of-charge estimate for lithium-ion battery with a recursive
total least squares-based observer,” IEEE Trans. Ind. Electron., 2018, doi:
10.1109/TIE.2017.2736480.
[281] D. Rosewater, “Battery Management System Standards.,” 2022. doi:
10.2172/2005394.
[282] T. Cipra and R. Romera, “Kalman filter with outliers and missing observations,”
Test, 1997, doi: 10.1007/BF02564705.
[283] C. Semeraro et al., “Digital twin application in energy storage: Trends and
challenges,” J. Energy Storage, vol. 58, no. December 2022, p. 106347, 2023, doi:
10.1016/j.est.2022.106347.
[284] F. Tao et al., “Digital twin-driven product design framework,” Int. J. Prod. Res.,
2019, doi: 10.1080/00207543.2018.1443229.
[285] W. Yang, K. Yoshida, and S. Takakuwa, “Digital Twin-Driven Simulation for a
Cyber-Physical System in Industry 4.0 Era,” 2017. doi:
10.2507/daaam.scibook.2017.18.

218