Title: BatteryMFormer: A Multi-Level Learning Framework for Forecasting Battery Degradation Trajectories

Abstract: Predicting the complete state-of-health trajectory of a battery using only early operational data—a task known as early battery degradation trajectory forecasting (BDTF)—is essential for optimizing battery manufacturing, deployment, and overall performance. Battery degradation data is defined by two primary features: a multi-level structure, which encompasses regularities common to specific aging conditions as well as trajectory patterns shared among different batteries, and localized variations in voltage-current profiles that are typically confined to particular state of charge (SOC) intervals. Current methods frequently overlook these specific characteristics. To address this limitation, we introduce BatteryMFormer, a multi-level Transformer architecture designed for early BDTF. This model incorporates three key components: first, an aging-condition-aware decoder that utilizes aging-condition-informed queries and attention mechanisms to inject priors related to aging conditions; second, a meta degradation pattern memory designed to learn and retrieve trajectory prototypes to assist in long-horizon forecasting; and third, a dual-view encoder that simultaneously captures temporal dynamics and SOC-localized variations from voltage and current time series. Comprehensive experiments across four distinct battery domains demonstrate that BatteryMFormer consistently surpasses state-of-the-art baselines, representing a major advancement in reliable BDTF. The source code is accessible at https://github.com/Ruifeng-Tan/BatteryMFormer.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC