Generative pre-trained transformers (GPT) for lithium-ion battery charging state prediction in battery swapping station

Abstract

Battery swapping stations face considerable challenges due to heterogeneous battery types, inconsistent initial states, and complex spatiotemporal patterns in BMS data. To address these issues, we propose BatteryGPT, a frozen pre-trained Generative Pre-trained Transformer (GPT) framework tailored for lithium-ion battery energy forecasting. Instead of retraining the entire model, our approach fine-tunes only the input embedding and output projection layers, enabling efficient adaptation to varied battery conditions. To enhance temporal feature representation, we introduce a contrastive temporal embedding module that compresses multivariate sequences while retaining essential dynamic features. Furthermore, we design a temporal suffix alignment strategy to align time-series data with textual prompts, improving the model’s capacity for temporal reasoning. Experiments show that BatteryGPT achieves an average 55.52% improvement in prediction accuracy over LSTM and conventional deep learning baselines. Instruction-based evaluations further demonstrate its end-to-end applicability in dynamic charging management scenarios. These results highlight the potential of integrating large language models with time-series adaptation techniques for industrial energy forecasting tasks.