LiRUL: A physics-constrained and lightweight learning framework for edge-based RUL prediction

Accurate prediction of the Remaining Useful Life (RUL) of lithium-ion batteries is essential for ensuring operational safety, reliability, and cost-effective lifecycle management in electric vehicles and energy storage systems. However, deploying RUL prediction models on real-world edge devices remains challenging due to limited computational resources, nonlinear degradation behaviours, and the need for physically interpretable yet efficient models. Existing physics-based approaches offer high fidelity but are computationally intensive, whereas deep learning models provide accuracy at the cost of transparency and scalability. To bridge this gap, this paper proposes LiRUL, a hybrid and lightweight learning framework specifically designed for RUL prediction under edge computing constraints. LiRUL integrates a compact feature engineering pipeline that fuses statistical compression with physics-informed domain variables, such as temperature and C-rate, to preserve degradation interpretability while minimising model complexity. Furthermore, a physics-constrained loss regularisation is introduced within temporal learning models to enforce monotonic degradation consistency, aligning predictions with electrochemical ageing behaviour. Experiments conducted on the LG M50 dataset demonstrate that LiRUL achieves up to 40% lower error and 60% faster inference latency compared to conventional LSTM and GRU models, while maintaining smooth and physically consistent degradation trajectories. These results highlight LiRUL’s capability to deliver an effective trade-off between accuracy, interpretability, and computational efficiency, establishing it as a deployable and generalisable framework for next-generation edge-based battery health management systems.