| 摘要: |
| 针对传统4D航迹预测方法在数据单一和特征选择上的局限，提出了一种基于时域融合Transformer(Temporal Fusion Transformer,TFT) 模型的4D 航迹预测方法。引入下降率、时序分量等多元特征，并将数据按是否随时间变化及数值属性进行分类，以体现飞行过程中不同阶段的差异；采用TFT模型有效捕捉各特征之间的隐式相关性，从而提高了预测精度；同时，结合分位数回归实现不确定性量化，提供了具有置信区间的航迹预测结果。实验表明，所提方法在真实数据上优于传统模型：与CNNLSTM模型和LSTM模型相比，平均距离误差分别减少了22.7%和50.9%，纵向、横向和垂直误差分别为305.01 m、177.91 m和25.23 m，验证了模型在解决航迹预测问题上的有效性，能够为管制精细化调控提供有效支持。 |
| 关键词: 空中交通管制 4D航迹预测 自动相关监视系统数据 时域融合Transformer 时间序列预测 |
| DOI：10.20079/j.issn.1001-893x.240819002 |
|
| 基金项目:国家重点研发计划(2021YFF0603904)；中央高校基本科研业务费专项资金(PHD2023-035)；2024年度四川省级大学生创新创业项目(S202410624084) |
|
| A Method of 4D Trajectory Prediction Based on Temporal Fusion Transformer |
| KONG Jianguo,MA Kexin,LIANG Haijun,ZHANG Xiangwei,CHANG Hanwen |
| (College of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618307,China) |
| Abstract: |
| To address the limitations of traditional 4D trajectory prediction methods in data singularity and feature selection,a 4D trajectory prediction method based on the Temporal Fusion Transformer(TFT) model is proposel.The method introduces multiple features,such as descent rate and temporal components,and classifies the data based on whether it varies over time or is numerical,to reflect the differences at various stages of flight.The TFT model is employed to effectively capture the implicit correlations between features,thereby improving prediction accuracy.Additionally,quantile regression is combined to quantify uncertainty and provide trajectory predictions with confidence intervals.Experimental results show that the proposed method outperforms traditional models on real-world data:compared with that of the CNNLSTM model and the LSTM model,the mean distance error is reduced by 22.7% and 50.9%,respectively.The longitudinal,lateral,and vertical errors are 305.01 m,177.91 m,and 25.23 m,respectively.The results validate the effectiveness of the model in solving trajectory prediction problems and demonstrate it can provide valuable support for refined air traffic control. |
| Key words: air traffic control 4D trajectory prediction automatic dependent surveillance-broadcast data temporal fusion Transformer time series prediction |
《电讯技术》编辑部 