Pack Transformer: Transformer-based hybrid model for lithium-ion battery packs

Accurate modeling of lithium-ion battery (LiB) packs is crucial for enhancing their performance, longevity, and safety. This requires capturing the complex electrochemical and thermal dynamics within individual pack segments (e.g., cells) and the dynamic interactions across the pack. In this work, we propose a novel hybrid modeling approach that integrates a structurally identifiable distributed equivalent circuit model (ECM) coupled with a lumped thermal (LT) model for each pack segment with a Transformer-based residual that corrects ECM< model predictions. By leveraging the Transformer’s self-attention mechanism, the residual enables high-resolution modeling of both temporal dependencies and spatial interactions between pack segments. To reduce the computational burden and enable LiB dynamics predictions over long sequences, we evaluate efficient self-attention mechanisms for our Transformer-based hybrid model (Pack Transformer), including Longformer, Factorized Encoder, and Factorized Attention architectures. In numerical experiments using a high-fidelity LiB pack model as a reference, we find that the Pack Transformer successfully predicts state trajectories for each pack segment over extended periods, achieving very low RMSE values of 0.008V for voltage and 0.1 °C for temperature. Compared to simpler hybrid model approaches that use varying levels of lumping, the transformer model is able to capture the spatio-temporal dynamics of the pack at a much higher resolution. Our approach paves the way for more accurate and safer predictive control of LiB packs, including applications in battery balancing and thermal management.