The development of Intelligent and Connected Vehicles (ICV) has raised higher requirements for the real-time performance and reliability of In-Vehicle Networks (IVN). Time-Sensitive Networking (TSN), with its high bandwidth, low latency, and deterministic communication, has emerged as an ideal solution. However, TSN scheduling is an NP-hard problem, and existing methods face limitations in scalability and joint optimization of routing and scheduling. To address this, we proposes a GAT-Enhanced DRL decision framework, which leverages Graph Attention Networks (GAT) to dynamically capture topological dependencies and integrates Deep Reinforcement Learning (DRL) for efficient scheduling. Furthermore, a Single-Frame Incremental Scheduling (SFIS) method is proposed to simplify the decision-making process by eliminating redundant scheduling operations within hyperperiods. The proposed method has been validated for effectiveness and necessity through a TSN topology constructed based on real-world vehicle communication requirements.