对航班实际撤轮挡时刻（AOBT）的精准预测是优化机场运行效率、增强机场协同决策（A-CDM）系统性能的重要环节。针对AOBT预测面临的复杂作业时空依赖、时序变化吉布斯现象等挑战，提出了一种融入频率增强通道注意力的时间卷积网络（FTCA）与Transformer混合模型的航班撤轮挡时刻预测FTCA-Transformer方法。首先，将航班地面保障各里程碑事件数据编码为统一标准的时序数据，提取不同时间戳隐藏的时间季节性特征。其次，采用时间卷积网络（TCN）挖掘时间片段信息，通过Transformer中多头自注意力机制将时间点间的联系扩展到时间片段尺度，克服传统递归神经网络难以有效挖掘长时跨度下事件间长期依赖关系的问题。再次，引入基于离散余弦变换（DCT）的频率增强通道注意力机制，有效挖掘地面保障数据频域特征的同时避免传统频域变换带来的吉布斯现象。最后，在国内华北地区某大型枢纽机场真实地面保障运行数据上的实验结果表明：在单步预测下，FTCA-Transformer方法对比基线算法，实验结果的平均绝对误差、平均绝对百分比误差、均方根误差分别降低了48.0%、53.5%、44.7%。消融实验验证了各子模块在模型中的积极作用。

Accurate prediction of the Actual Off-Block Time (AOBT) is a crucial component for optimizing airport operational efficiency and enhancing the performance of Airport Collaborative Decision-Making (A-CDM) systems. To address the challenges posed by complex spatiotemporal dependencies and Gibbs phenomena in time-series variations, this study proposes a hybrid prediction model—FTCA-Transformer—integrating a Frequency-enhanced Channel Attention mechanism with a Temporal Convolutional Network (TCN) and Transformer architecture. First, milestone events in ground handling operations are encoded into unified sequential data, enabling the extraction of hidden temporal and seasonal features across timestamps. Next, the TCN captures local temporal segment information, while the multi-head self-attention mechanism in the Transformer extends temporal correlations from individual points to segment-level dependencies, overcoming the limitations of recurrent neural networks in modeling long-range dependencies. Furthermore, a Frequency-enhanced Channel Attention mechanism based on Discrete Cosine Transform (DCT) is introduced to effectively exploit frequency-domain features of ground operation data while mitigating Gibbs phenomena associated with conventional frequency transforms. Finally, experimental results using real ground handling operation data from a major hub airport in North China demonstrate that, under single-step prediction, the proposed FTCA-Transformer outperforms baseline methods, reducing the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) by 48.0%, 53.5%, and 44.7%, respectively. Ablation studies further confirm the positive contributions of each module within the model.