Thermal effect has a significant impact on the performance of lithium-ion batteries (LIBs). Therefore, an accurate and effective temperature distribution model is of great significance for their thermal management. The complex thermal process often has different characteristics in local and global scales. To properly process these multi-scale dynamics, a two-scale manifold learning method is proposed for spatiotemporal modeling of the battery system. This new method starts with local neighbor graph construction and then using global geodesic distance as the supplementary graph. The complete modeling procedure can be summarized as follows: Firstly, a group of nonlinear spatial basis functions (BFs) for time/space separation is developed by the newly developed manifold learning method. Secondly, the low-order temporal model is derived using Galerkin’s method, where the low-dimensional representative of heat generation term model is approximated using extreme learning machine (ELM). Finally, the entire temperature distribution in the LIBs can be reproduced through time/space synthesis. Since the spatial information in both local and global scales is properly considered in the BFs optimal learning, the proposed method will have higher model accuracy than those existing methods, which only consider spatial information in one scale. Real-time experiment and simulations validate the efficiency and accuracy of the proposed method.