Session: 09-01-04: Electrochemical Energy Storage and Conversion Systems IV

Paper Number: 168112

Deep Neural Network-Based Modeling of Electro-Thermal Lithium-Ion Batteries Responses Leveraging Hybrid Pulse Power Characterization 

Approved, DCN# 2025-3-3-688, 03/04/2025

 

Accurate modeling of lithium-ion batteries is critical for optimizing performance, improving efficiency, and extending lifespan in applications such as electric vehicles, aerospace systems, and microgrids. However, direct battery testing under extreme conditions, such as high C-rates and prolonged cycling, presents significant safety hazards, including thermal runaway, fire risk, and cell degradation. Additionally, battery testing is expensive, requiring specialized equipment, controlled environments, and long testing durations. These challenges necessitate the development of reliable predictive models that can emulate the electro-thermal behavior of a battery without the risks and costs associated with physical testing. Traditional physics-based models, while effective, are computationally intensive and require detailed knowledge of electrochemical properties, making them difficult to adapt to changing operating conditions, battery aging effects, or batteries of unknown origins (i.e. recycled batteries). To address these limitations, this paper presents a data-driven battery modeling approach using deep neural networks (DNNs) trained on Hybrid Pulse Power Characterization (HPPC) data. The proposed model aims to predict both electrical and thermal responses of lithium-ion batteries under dynamic load conditions, offering a scalable and efficient alternative to physics-based electro-thermal battery modeling. The deep learning model is trained using HPPC datasets, where input current loads are used to predict key battery parameters, including voltage response, state-of-charge (SOC), and temperature evolution. Unlike physics-based models, which rely on predefined equations to describe battery behavior, the DNN autonomously learns complex nonlinear relationships through supervised learning. In this work, various activation functions are explored to assess their impact on predictive accuracy, stability, and generalization performance. The trained model is validated against unseen load profiles to ensure robustness across different operating conditions. The expected results are that the DNN-based model will provide accurate predictions of battery voltage, SOC, and temperature while requiring significantly less knowledge about the cell’s electro-thermal characteristics and thermal behaviors compared to traditional physics-based models. It is anticipated that the model will generalize well to different load conditions, making it a suitable candidate for real-time implementation as a forecasting tool in battery management systems. Additionally, the exploration of activation functions is expected to demonstrate variations in model performance, with some functions improving convergence rates and predictive accuracy, while others may be more effective at capturing subtle variations in battery behavior. Beyond static training, this research introduces a framework for online learning, allowing the model to continuously update its parameters as new operational data is acquired. By dynamically adapting to battery aging, environmental variations, and evolving operating conditions, the model is expected to maintain long-term predictive accuracy, reducing reliance on costly and hazardous real-world testing. Future work will focus on implementing and evaluating the online learning framework, refining the neural network architecture, and expanding the dataset to include high C-rate conditions. Additionally, hybrid modeling approaches that combine physics-based constraints with data-driven learning will be explored to further enhance predictive accuracy. This research advances deep learning applications in battery modeling, providing a robust, adaptive, and scalable alternative to traditional techniques for optimizing energy storage systems.

The views expressed are those of the author and do not reflect the official policy or position of the Department of Defense of the U.S. Government.

DISTRIBUTION STATEMENT A. Approved for public release distribution unlimited.

Presenting Author: Connor Madden University of South Carolina

Presenting Author Biography: Connor Madden is a graduate research assistant at the University of South Carolina. His main research focus is on the advanced multi-domain modeling of lithium-ion batteries.

Authors:

Connor Madden University of South Carolina
Jarrett Peskar University of south carolina
Austin Downey University of south carolina
Kerry Sado University of South Carolina
Jamil Khan University of South Carolina