To address the challenges of accurately reconstructing passenger travel chains and predicting the spatiotemporal distribution of passenger flow in transportation hubs, this paper proposes a deep learning architecture based on mobility-driven spatial embedding and extended long short-term memory networks (MDSE-sLSTM). Initially, we conducted route choice behavior experiments in a virtual reality (VR) hub scenario to reconstruct passenger mobility chains. Then, using Skip-gram, we analyzed the semantic correlations between nodes representing functional areas to achieve a spatially embedded representation of the hub area, obtaining spatial correlation features. In the temporal dimension, an extended bidirectional long short-term memory network with multi-head attention mechanism was employed to jointly learn the time-varying patterns of passenger flow data. Validation on the passenger flow distribution dataset of Tianjin West Station shows that the MDSE-sLSTM model outperforms traditional prediction models. Ablation experiments further confirm that the method of extracting spatial features based on mobility-driven approaches significantly improves passenger flow prediction compared to extracting spatial features based on the actual physical network.