Date of Award

12-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Automotive Engineering

Committee Chair/Advisor

Beshah Ayalew

Committee Member

Rong Ge

Committee Member

Jiangfeng Zhang

Committee Member

Yunyi Jia

Abstract

Lithium-ion batteries (LiBs) play a critical role in the automotive industry’s transition to electric mobility, offering a compelling combination of high energy and power density, extended cycle life, and strong safety performance. Realizing these advantages requires accurate modeling of LiB, which is inherently complex due to tightly coupled electrochemical and thermal dynamics that are sensitive to temperature, charge/discharge rates, and evolve over time as the battery ages. To address this challenge, hybrid modeling approaches that combine the explainability and generalizability of physics-based models with the expressive power of data-driven neural networks are gaining traction. These models are most effective when the decomposition between the two components is well-posed. In this dissertation, we propose a disciplined hybrid modeling scheme that prioritizes the physics-based component to capture the dominant system dynamics, while employing a neural network-based residual model to account for dynamics not captured by the physics model.

Our approach builds on a globally structurally identifiable first-order equivalent circuit model (ECM) as the physics component, ensuring the uniqueness of the estimated physics parameters. For a single LiB unit—either an individual cell or a group of cells treated as one—we employ a recurrent neural network (RNN) to model the residual dynamics, capturing temporal dependencies through its memory mechanism. Extending the hybrid modeling approach to a multisegment battery pack configuration requires capturing both the complex electro-thermal dynamics within individual segments (e.g., cells) and the interactions across the entire pack. To address this, we propose a Pack Transformer, a novel hybrid model that integrates distributed ECM models for each segment with a centralized Transformer-based residual. By leveraging the Transformer’s self-attention mechanism, this residual captures both temporal dependencies and spatial interactions among pack segment. The nonlinear dependence of the ECM parameters on LiB states, as well as the residual components of the hybrid model, are learned from real-world operational data, including voltage, current, and temperature. This enables the model to adapt to changes due to battery aging and varying operational conditions without relying on laboratory measurements. The proposed framework is validated through both simulation and experimental studies, demonstrating that the hybrid model provides highly accurate predictions by effectively capturing the dynamics of individual LiB units as well as the interactions among segments within the pack.

Continually learned hybrid models can serve as digital twins to support advanced control and optimization tasks for battery management functions. To this end, we formulate a Markov Decision Process (MDP) models and apply deep reinforcement learning (DRL) techniques to extract optimal policies through interaction with the hybrid model as a digital twin. This approach avoids safety risks and operational disruptions associated with DRL training via direct on-battery training. Since the DRL policy is trained on a continually updated hybrid model, it remains well-adapted to the current state of battery aging and varying operational conditions. To further ensure safety, we incorporate formal guarantees using a control barrier function (CBF) filter, which enforces constraints on state of charge (SOC), voltage, and temperature. Using the single LiB hybrid model, we demonstrate that the proposed framework can generate safe fast charging policies, addressing the persistent challenge of prolonged charging times in LiBs. Furthermore, by leveraging the Pack Transformer, the framework learns an adaptive and safety-aware DRL policy for concurrent charging and balancing across the battery pack.

Overall, this dissertation presents a unified framework for data-driven yet physically grounded modeling and control of LiBs at both the single unit and multi-segment pack configurations. The results demonstrate the potential of learning-based adaptive strategies to significantly outperform traditional conservative approaches in terms of operational performance and safety. This approach paves the way for more accurate modeling and reliable predictive control of LiBs, enabling next generation battery management systems (BMSs) that intelligently adapt to aging and varying operational conditions

Author ORCID Identifier

https://orcid.org/0009-0005-0794-7242

Download

Available for download on Thursday, December 31, 2026

Share

COinS