KAN-Therm: A Lightweight Battery Thermal Model Using Kolmogorov-Arnold Network

TL;DR

The paper addresses core-temperature estimation for lithium-ion batteries under resource constraints in BMSs. It introduces KAN-Therm, a lightweight Kolmogorov-Arnold Network that predicts the core temperature $\widehat{T}_1$ from four measurable inputs using a compact two-layer architecture and spline-based activations, optimized via grid search with L-BFGS. By incorporating sparsity and self-entropy penalties in the loss, KAN-Therm achieves a small parameter footprint while maintaining high accuracy, outperforming MLP, RNN, and LSTM baselines in RMSE and inference speed. The approach enables fast, memory-efficient deployment in safety-critical battery systems, offering a practical alternative to heavier neural networks and physics-based observers.

Abstract

Battery management systems (BMSs) rely on real-time estimation of battery temperature distribution in battery cells to ensure safe and optimal operation of Lithium-ion batteries (LIBs). However, physical BMS often suffers from memory and computational resource limitations required by highfidelity models. Temperature prediction using physics-based models becomes challenging due to their higher computational time. In contrast, machine learning based approaches offer faster predictions but demand larger memory overhead. In this work, we develop a lightweight and efficient Kolmogorov-Arnold networks (KAN) based thermal model, KAN-Therm, to predict the core temperature of a cylindrical battery. We have compared the memory overhead and computation costs of our method with Multi-layer perceptron (MLP), recurrent neural network (RNN), and long shortterm memory (LSTM) network. Our results show that the proposed KAN-Therm model exhibit the best prediction accuracy with the least memory overhead and computation time.

Figures (7)

• Figure 1: Flowchart shows how the dataset is split into training, validation, and testing subsets.
• Figure 2: Data distribution of current ($I$) vs core temperature ($T_1$).
• Figure 3: Block diagram illustration of the proposed KAN-Therm model for prediction of battery core temperature ($\widehat{T}_1$).
• Figure 4: Parallel coordinate plot illustrates the exploratory findings of the hyperparameter tuning process for our proposed KAN-Therm model.
• Figure 5: Figure shows the training and validation loss with epochs for the KAN-Thermal model.