Battery-Timer: A Foundation Model for Battery Discharge Capacity Degradation Forecasting

Battery-Timer is a domain-adapted time series foundation model tailored for lithium-ion battery capacity degradation prediction. Built upon the Transformer-based Timer architecture developed by Tsinghua University, Battery-Timer is fine-tuned using approximately 10GB of publicly available battery degradation datasets to encourage a degradation-aware representation learning. This enables the model to generalize well across varying battery chemistries, operational scenarios, and degradation profiles. Extensive validation on the in-house CycleLife-SJTUIE dataset demonstrates that Battery-Timer not only surpasses the performance of its untuned counterpart but also exhibits strong zero-shot learning (ZSL) capabilities across both constant current (CC) and constant current-constant voltage (CCCV) charging conditions. Despite its large parameter size, Battery-Timer serves as a powerful teacher model in the proposed knowledge distillation framework, significantly improving the generalization ability of lightweight expert models while preserving low computational costs.

Overview

Dataset

SJTUIE-cyclelife

📦 CycleLife-SJTUIE Dataset

The CycleLife-SJTUIE dataset contains real-world battery degradation data collected from eight lithium-ion cells manufactured by Gotion High Tech Co., Ltd. This dataset is designed to support battery health prediction research, especially under different charging strategies and temperature conditions.

---

🔋 Battery Specifications

Attribute	Value
Product Name	IFR32135
Rated Capacity	13Ah
Normal Voltage	3.2V
Shape	Cylindrical
Cathode	LiFePO₄
Anode	Graphite
Voltage Range	2.0V – 3.65V

---

🧪 Experimental Setup

• Tests conducted indoors without active temperature control.
• Each cell mounted in a Neware A708-4B-J-30A fixture.
• Charging/discharging handled by Neware CTE-4008D-5V30A battery tester.
• Data/control managed by Neware CT-ZWJ-4 ST-1U unit.
• Surface temperature monitored using thermocouples + Jinko JK5000-24.
• Ambient temperature recorded via Huahanwei TH42W-EX thermometer.

---

⚙️ Charging Profiles

• Cells #1–4: Constant Current (CC) charging
• Cells #5–8: Constant Current + Constant Voltage (CCCV) charging

---

🔁 Test Cycle Procedure

Each battery cell undergoes repeated degradation cycles until its capacity falls below 80% of its rated capacity. The cycle is as follows:

1. Discharge the cell completely.
2. Repeat:
   • CC profile: Charge at 1C (13A) until voltage reaches 3.9V.
   • CCCV profile: Charge at 1C (13A) to 3.65V, then hold at 3.65V until current drops below 0.65A (0.05C).
   • Rest for 30 minutes.
   • Discharge at 1C (13A) until voltage drops to 2.0V.
   • Rest for 30 minutes.
3. Stop when capacity / rated capacity < 0.8.

---

Please cite the dataset and associated papers when using this data in your research.

Open-source battery capacity degradation dataset for fine-tuning

@article{severson2019data,
  title={Data-driven prediction of battery cycle life before capacity degradation},
  author={Severson, Kristen A and Attia, Peter M and Jin, Norman and Perkins, Nicholas and Jiang, Benben and Yang, Zi and Chen, Michael H and Aykol, Muratahan and Herring, Patrick K and Fraggedakis, Dimitrios and others},
  journal={Nature Energy},
  volume={4},
  number={5},
  pages={383--391},
  year={2019},
  publisher={Nature Publishing Group UK London}
}

@misc{calce_battery_data,
  title        = {CALCE Battery Data Sets},
  author       = {Center for Advanced Life Cycle Engineering (CALCE)},
  year         = {2024},
  howpublished = {\url{https://calce.umd.edu/battery-data}},
  note         = {Accessed: 2025-04-29}
}

@article{preger2020degradation,
  title={Degradation of commercial lithium-ion cells as a function of chemistry and cycling conditions},
  author={Preger, Yuliya and Barkholtz, Heather M and Fresquez, Armando and Campbell, Daniel L and Juba, Benjamin W and Rom{\`a}n-Kustas, Jessica and Ferreira, Summer R and Chalamala, Babu},
  journal={Journal of The Electrochemical Society},
  volume={167},
  number={12},
  pages={120532},
  year={2020},
  publisher={IOP Publishing}
}

@article{wang2023large,
  title={Large-scale field data-based battery aging prediction driven by statistical features and machine learning},
  author={Wang, Qiushi and Wang, Zhenpo and Liu, Peng and Zhang, Lei and Sauer, Dirk Uwe and Li, Weihan},
  journal={Cell Reports Physical Science},
  volume={4},
  number={12},
  year={2023},
  publisher={Elsevier}
}

Quick start

quick_start_validation.ipynb

Citation

If you find this repo helpful, please cite our paper.

@misc{chan2025foundationmodelsknowledgedistillation,
      title={Foundation Models Knowledge Distillation For Battery Capacity Degradation Forecast}, 
      author={Joey Chan and Zhen Chen and Ershun Pan},
      year={2025},
      eprint={2505.08151},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2505.08151}, 
}

Contributors

If you have any questions or want to use the code, feel free to contact:

Chan.Joey ([email protected])

Zhen Chen ([email protected])

Wei Wu ([email protected])