Comparative Analysis of GenAI Models for Extensions and Validation of EV Battery Characterization Data

<div class="section abstract"><div class="htmlview paragraph">Rapid advancement of electric vehicle (EV) technology has propelled the need for reliable and efficient methods of battery data. This has vital importance – to ensure safety aspects and efficient design of EV system. Traditional data collection methods for battery characterization is a large subject for the design of experiments and is often expert’s skill intensive, time-consuming, and do not allow scalability. This study proposes an approach which bases on Generative Artificial Intelligence (GenAI) for two activities. First, to assist the DOE in characterizing cell/batteries at different C-rates and temperatures considering different degradation rates. Second, for manipulation of characterization data taking into account measurement and data recording errors. The study compares GenAI models like Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and transformer-based (Time-GPT) models in generating and validating EV battery characterization data. This is not a complete replacement for battery testing, as batteries must physically undergo cyclic aging and other tests. The paper explores ability of different GenAI models to accurately capture critical electro-chemical and thermal features. This enables better planning of next characterization experiments and assist in eliminating boundary and intermediate scenarios for cell characterization experiments by generating synthetic data from GenAI models. The model so tuned for cell characteristics can also be extended as data manipulator to re-establish battery characterization data generated from testing experiments, considering sensor issues, data logging issues, data transportation and synchronization issues, etc. The robustness of these models in handling diverse, heterogeneous, and asynchronous datasets sourced from different EV manufacturers, battery chemistries, and specifications are scrutinized. The performance of the models is compared across multiple attributes like execution times, computing resource requirements, accuracy and consistency of generated data, and volume of data required to optimize the models. This study contributes to improve the modelling, simulation, and optimization of EV batteries and enable rapid development of data-driven products specifically for battery health analysis.</div></div>