Abstract
Electric bikes (e-bikes) powered by lithium-ion batteries are essential for sustainable mobility in smart cities. However, limited range and slow charging remain significant hurdles. Shared Electric Bike (SEB) systems, utilizing battery swapping stations, offer a solution but require precise battery management. We present the Large Lithium-ion Battery Model (LLiM), a billion-parameter foundation model trained on over ten million battery time-series records. Through unsupervised pretraining and efficient finetuning, LLiM achieves state-of-the-art performance in anomaly detection, state-of-health (SoH) estimation, and remaining range (RR) prediction. Deployed on cloud servers, LLiM enhances the safety and efficiency of battery swapping services, accelerating the transition to green urban mobility.
Introduction: The Challenge of Shared Mobility
As smart cities transition toward zero-carbon transportation, e-bikes have become a cornerstone of urban "last-mile" delivery. Shared E-bike Battery (SEB) systems address the limitations of plug-in charging by allowing users to swap depleted batteries for fully charged ones at automated cabinets. However, managing these systems at scale is complex. Traditional Battery Management Systems (BMS) often fail to accurately predict health or detect safety anomalies because they are trained on small, lab-controlled datasets that lack the variability of real-world usage.
Methods: A Foundation Model for Batteries
LLiM adapts the success of Large Language Models (LLMs) to the domain of battery physics. The model features a Transformer-based architecture scaled to 1 billion parameters, designed to capture long-range dependencies in time-series data such as voltage, current, and temperature.
The training process involves two main stages:
- Unsupervised Pre-training: Using mask modeling on 10.1 million sequences, the model learns to reconstruct hidden portions of battery data. This forces the model to understand underlying electrochemical patterns without human-labeled data.
- Task-Specific Fine-tuning: The pretrained backbone is adapted for downstream tasks using Low-Rank Adaptation (LoRA), which allows for efficient updates without retraining the entire massive network.
Data Scale and Diversity
The model's robustness stems from an unprecedented dataset:
- Pre-training Data: 10.1 million sequences (6 billion data points) from 50+ cities in China.
- Diversity: Coverage of six distinct battery cell types and 71 features (numerical and categorical).
- Benchmarks: Evaluation on labeled datasets for anomaly detection (23,250 records) and SoH estimation (63,000 records).
Key Results and Performance
LLiM significantly outperforms existing models like TimesNet and PatchTST across three critical metrics:
- Anomaly Detection: Achieved a 0.975 F1-score and 98.4% accuracy in identifying safety risks like internal short circuits.
- State of Health (SoH): Mean Absolute Error (MAE) of only 0.62 Ah, representing an error rate of approximately 2%.
- Remaining Range (RR): Predicted riding range with an error of less than 1.2 km, providing high reliability for delivery riders.
Furthermore, the model demonstrates exceptional robustness, maintaining high accuracy across temperatures ranging from 10°C to 40°C, where traditional models typically degrade.
Conclusion and Future Directions
LLiM represents a paradigm shift in battery management, proving that scaling laws apply to physical systems as they do to language. While currently deployed on cloud servers, future work will focus on reducing computational overhead for edge deployment and adapting the model to emerging chemistries like Sodium-ion batteries. This advancement is a vital step toward safer, more reliable, and efficient green smart city infrastructure.
Summary updated