Zimeng Fan, Yuting Guo, Lei Song, Junrong Du, Hongfei Wang, Xuzhi Li
Machine learning has demonstrated remarkable breakthroughs in predicting the state of health (SOH) for lithium-ion batteries. However, conventional methods face critical challenges in cross-domain adaptation, inter-dataset generalization, and long-horizon forecasting due to variations in usage conditions and electrochemical characteristics. Inspired by the success of large language models (LLMs), time-series foundation models (TSFMs) offer an alternative solution to overcome the issues above. Nevertheless, studies to explore the generalization enhancement capability of TSFMs for battery SOH forecasting under different cross domain factors remain insufficient. Therefore, a novel TSFMs based framework named BatteryTSFM is proposed for SOH forecasting. First, we introduce backbone-aware temporal resampling that dynamically adapts preprocessing to structural characteristics of diverse TSFMs, enabling optimal cross-domain generalization through feature scaling. Second, Monte Carlo dropout is integrated into autoregressive inference to quantify the multi-step prediction errors. Across four public datasets, BatteryTSFM reduces RMSE by an average of 35% in cross-condition tasks and 88% in cross-chemistry tasks, indicating that foundation-model methods can deliver reliable long-horizon SOH forecasts for energy systems. We also conduct exploratory analyses that link generalization to fine-tuning dataset size and resampling granularity, yielding practical guidance for deployment. • Broadly pretrained time-series foundation models improve long-horizon battery state of health trajectory forecasting. • Backbone-aware resampling aligns temporal granularity, supporting plug-and-play backbones. • Next-segment regression unifies training/inference; robust multi-step extrapolation. • Monte Carlo dropout provides prediction intervals, quantifying uncertainty. • Demonstrated effectiveness across two cases over eight baselines.