To address the shortcomings of existing aircraft taxiing path selection models in integrating dynamic traffic state information under real-world operational scenarios, an improved Transformer-based Route Choice Model (RCM) is proposed in this paper. Historical taxiing trajectories were used as the foundation of the model. Spatiotemporal relationships between known and unknown road segments in discontinuous sequences were modeled by the encoder under masking conditions. A dual cross-attention mechanism was introduced into the decoder as a key improvement. One branch aggregated global path features from the encoder, and the other integrated traffic flow information of busy road segments. These features were fused under reachable segment mask constraints to guide the selection of the current road segment. Validation experiments were conducted using real operational data from a major airport in western China. The results show that, compared with the baseline Transformer, the Bilingual Evaluation Understudy (BLEU), Jensen–Shannon Divergence (JSD), and Path Completion Rate (PCR) are improved by 4.14%, 0.62%, and 2.57%, respectively. Path generation quality and overall accuracy are enhanced. Ablation experiments further show that the effective and busy road segment masking strategy, the dual cross-attention structure, and the cross-layer connection design play key roles in improving model stability and predictive performance.