Foundation Models Knowledge Distillation For Battery Capacity Degradation Forecast

TL;DR

This work tackles cross-scale battery capacity degradation forecasting by leveraging a time-series foundation model (Timer) trained on open datasets and fine-tuned with LoRA to capture degradation dynamics. It then distills the foundation model into compact expert forecasters to enable edge deployment while preserving cross-capacity generalization. The approach yields state-of-the-art or competitive accuracy across multiple datasets and protocols, and distillation consistently improves lightweight models in both in-domain and zero-shot cross-protocol settings. Interpretability analyses via LIME reveal that the model emphasizes informative, physically meaningful degradation history, supporting trustworthy deployment and further extension to other long-horizon prognostics tasks.

Abstract

Accurate forecasting of lithium-ion battery capacity degradation is critical for reliable and safe operation, yet remains challenging under distribution shifts across scales and operating regimes. Here we investigate a time-series foundation model, that is, a large pre-trained time-series model for capacity degradation forecasting, and propose a degradation-aware fine-tuning strategy that aligns the model to capacity trajectories while retaining broadly transferable temporal structure. We instantiate this approach by fine-tuning the Timer model on 220,153 cycles of open-source charge-discharge records to obtain Battery-Timer. Using our released CycleLife-SJTUIE dataset, a real-world industrial collection from an energy-storage station with long-horizon cycling, we evaluate capacity generalization from small cells to large-scale storage systems and across varying operating conditions. Battery-Timer consistently outperforms specialized expert models. To address deployment cost, we further introduce knowledge distillation, a teacher-student transfer that compresses the foundation model's behavior into compact expert models. Distillation across several state-of-the-art time-series experts improves multi-condition capacity generalization while substantially reducing computational overhead, indicating a practical path to deployable cross-scale degradation forecasting by combining a foundation model with targeted distillation.

Figures (10)

• Figure 1: Process flow of our study. A TSFM is fine–tuned with LoRA on small–cell, cycle–indexed capacity sequences, evaluated in zero–shot on large energy–storage cells and different charging protocols, and distilled into compact expert models for onboard BMS.
• Figure 2: An overview of the datasets used in this study is provided. To evaluate the model's generalizability, experiments will be conducted on various battery models.
• Figure 3: Architectures of Timer: a Transformer-decoder-based forecasters.
• Figure 4: Our Method.(a) LoRA fine-tuning and cross-protocol deployment. A time-series foundation model is fine-tuned with Low-Rank Adaptation on degradation data from small-capacity cells to enable transfer to large-capacity energy-storage cells. (b) Response-Based Knowledge Distillation Framework for Time-Series Foundation Models. An expert model distilled under the CCCV protocol is then applied to CC operation to evaluate cross-protocol generalization.
• Figure 5: Performance comparison between Battery-Timer and recent time-series forecasters.(a) Scatter of predictions versus ground truth under window-based validation; points from our method cluster near the identity line (y = x), indicating higher accuracy. (b) Violin plots of average errors across eight cells with corresponding Friedman ranks; the TSFM (Battery-Timer), trained on large-scale data, demonstrates a consistent advantage.