Abstract: Accurate acquisition of Channel State Information (CSI) is a core prerequisite for Multiple-Input Multiple-Output (MIMO) systems to achieve reliable transmission and low-latency communication control. In tunnel scenarios, the channel is characterized by narrow spatial dimension and prominent multipath effect; traditional channel estimation methods, constrained by high signaling overhead and weak resistance to channel aging, can hardly meet the real-time and accuracy requirements of vehicular communication systems. To address this problem, this paper proposes an adaptive CSI prediction method for tunnel scenarios. Built upon the Pathway-Transformer framework, this method jointly models the local evolution features and long-term dependencies of CSI via multi-scale Patch representations, and devises a dual attention mechanism to capture intra-Patch local correlations and inter-Patch global dependencies. On this basis, an attention-inspired gating routing mechanism is introduced to implement adaptive selection and sparse activation of multi-scale branches, so as to improve channel representation capability while reducing computational overhead. Coupled with a lightweight decoder, it realizes future CSI prediction and enhances the online inference capability of the model. Experimental results show that under typical tunnel channel configurations, the proposed method outperforms mainstream methods in both prediction accuracy and robustness, and can provide effective support for communication and control in tunnel Internet of Vehicles (IoV) and autonomous driving systems.