KAN-Therm: A Lightweight Battery Thermal Model Using Kolmogorov-Arnold Network
For battery management systems, this provides a lightweight thermal model that balances accuracy and resource efficiency, though the improvement is incremental over existing neural approaches.
KAN-Therm uses Kolmogorov-Arnold networks to estimate battery core temperature with lower memory and computational overhead than existing neural network models, enabling deployment on resource-constrained battery management systems.
A battery management system (BMS) relies on real-time estimation of battery temperature distribution in battery cells to ensure safe and optimal operation of Lithium-ion batteries. However, physical BMS often suffers from memory and computational resource limitations required by high-fidelity models. Temperature estimation of batteries for safety-critical systems using physics-based models on physical BMS can potentially become challenging due to their higher computational time. In contrast, neural network-based approaches offer faster estimation but require greater memory overhead. To address these challenges, we propose Kolmogorov-Arnold network (KAN) based thermal model, KAN-therm, to estimate the core temperature of a cylindrical battery. Unlike traditional neural network architectures, KAN uses learnable nonlinear activation functions that can effectively capture system complexity using relatively lean models. We have compared the memory overhead and estimation time of our model with state-of-the-art neural network and tree-based models to demonstrate the applicability and potential scalability of KAN-therm on a physical BMS.