Accurate battery degradation trajectory forecasting is important for early warning and reliable health management. However, data-driven models trained under limited operating conditions often generalize poorly to unseen conditions. To address this issue, this paper proposes the Relative Degradation Trajectory Forecaster (RDTF) for domain-generalized battery state-of-health (SOH) trajectory prediction. RDTF decomposes future SOH prediction into current health state localization and relative degradation forecasting, combining them to reconstruct the future trajectory. A lightweight late-stage weight averaging strategy is further introduced to improve model robustness. Experiments on six operating conditions of the XJTU dataset are conducted under a strict leave one domain out protocol. Results show that RDTF consistently outperforms conventional baselines and representative domain generalization methods under unseen conditions. Ablation and individual component analyses further confirm the effectiveness of the relative forecasting formulation and the weight averaging strategy.

Lithium-ion battery,state of health,degradation trajectory forecasting,domain generalization,weight averaging