【摘要】 为提升航班延误预测的精度与稳定性，提出了一种融合双向长短期记忆网络（BiLSTM）与Transformer的混合深度学习模型，该模型通过BiLSTM有效捕获局部时序特征，并利用Transformer的多头注意力机制建模数据中的长期依赖关系。基于南京禄口国际机场2024年共计115 734条离港航班及相应逐时气象观测数据开展实证验证。结果表明，该混合模型的平均预测误差为4.18分钟，与随机森林（RF）、BiLSTM和Transformer模型相比，平均绝对误差（MAE）分别降低了24.1%、12.2%和7.9%，显示出更优的预测性能。研究结果为航空运行管理与智能化决策提供了有效的技术路径与理论支撑。更多还原

【Abstract】 To improve the accuracy and robustness of flight delay time prediction, a hybrid deep learning model was proposed. That combines a bidirectional long short-term memory(BiLSTM) network with a Transformer, the BiLSTM module captures local temporal patterns in historical flight operation and meteorological data, while the Transformer module uses multi-head self-attention to model long-range dependencies. The proposed model is evaluated on 115,734 departure flights and corresponding hourly meteorological observations from Nanjing Lukou International Airport in 2024. The hybrid model achieves a mean absolute error(MAE) of 4. 18 minutes and reduces MAE by 24. 1%, 12. 2%, and 7. 9% relative to Random Forest, BiLSTM, and Transformer baselines, respectively, demonstrating superior predictive performance. These results indicate that the proposed approach provides an effective technical framework and theoretical basis for intelligent decision-making and operational management in civil aviation.更多还原