STL-LLM: A seasonal-trend decomposition-enhanced large language model for battery capacity aging trajectory prediction

Abstract

Reliable monitoring of lithium-ion battery health is critical for electric vehicles and energy storage systems. Accurate prediction of the remaining capacity aging trajectory remains essential for battery management, yet current machine learning approaches often fail to capture long-term temporal dependencies in degradation data or leverage heterogeneous datasets effectively. In particular, while Pre-trained Large Language Models (LLMs) exhibit powerful reasoning abilities, their application to time-series-based capacity aging trajectory prediction is hindered by a fundamental modality mismatch. To address this, we propose STL-LLM, a novel framework integrating Seasonal-Trend decomposition using LOESS (STL) with frozen LLMs. STL-LLM disentangles battery health sequences into seasonal and trend components, reprograms these temporal features into text-aligned prompts, and employs prefix-based prompting to enhance temporal reasoning. The LLM's output is projected to generate a capacity aging trajectory prediction. Evaluations demonstrate STL-LLM's state-of-the-art accuracy across three public battery datasets, with consistent superiority in ablation and sensitivity studies. From a methodological perspective, STL-LLM offers a principled cross-modal representation learning solution for time-series forecasting, enabling frozen LLM deployment in non-text domains with minimal tuning. Practically, the framework provides a scalable and generalizable approach for battery prognostics, with potential applications in predictive maintenance and cloud-based battery management systems. More broadly, this work bridges the modality gap between structured time-series signals and pre-trained language models. It introduces a transferable paradigm for leveraging LLMs, which holds significant potential for advancing scientific time-series analysis and sequence modeling. While the direct application lies in battery health monitoring for new energy vehicles, this framework creates a pathway for broader impacts across energy systems.