Sparsity-Regularized Tensor Recurrent Network (SRTRN) for surface temperature prediction of large-format lithium-ion batteries

Lithium-ion batteries (LIBs) are ubiquitous in modern energy storage, valued for their high energy density and volumetric efficiency. Nevertheless, the deployment of large-format cells introduces significant thermal heterogeneity, complicating the design of robust Battery Thermal Management Systems (BTMS). Addressing this necessitates models that are simultaneously high-fidelity, data-efficient, and suitable for online deployment. This study proposes an improved spatiotemporal modeling framework in which a sparsity-regularized CP tensor decomposition with Bayesian automatic rank determination compresses the 2D surface temperature (ST) evolution into interpretable latent modes, and an attention-based bidirectional GRU (Attn-BiGRU) with exogenous current input predicts their temporal trajectory and reconstructs the full temperature field. The framework is evaluated on both numerical simulations covering multiple battery formats, C-rates, ambient conditions, and dynamic drive cycles, and on physical experiments using a calibrated infrared thermography (IRT) imaging system. Comparative results against CP-GRU, CNN-LSTM and ConvLSTM baselines demonstrate that the proposed Sparsity-Regularized Tensor Recurrent Network (SRTRN) approach achieves lower reconstruction error while requiring substantially fewer parameters, making it suitable for online BTMS deployment.