基于Transformer-TCN-GRU的场面滑行轨迹预测模型

王兴隆; 李国祥; 张钊; 叶可; 苏婷; 葛京

doi:10.3963/j.jssn.1674-4861.2025.02.006

基于Transformer-TCN-GRU的场面滑行轨迹预测模型

doi: 10.3963/j.jssn.1674-4861.2025.02.006

王兴隆^1,,
李国祥²,
张钊^2, ,,
叶可³,
苏婷³,
葛京⁴

1.
中国民航大学民航飞联网重点实验室天津 300300
2.
中国民航大学计算机科学与技术学院天津 300300
3.
中国民航信息网络股份有限公司信息服务部北京 100011
4.
中国民航信息网络股份有限公司安全生产与质量管理部北京 100011

基金项目:

天津市教育委员会自然科学重点项目 2020ZD01

详细信息

作者简介:
王兴隆（1979—），硕士，研究员. 研究方向：空域运行安全、空中交通流量管理等. E-mail：xinglong1979@163.com

通讯作者:
张钊（1988—），博士，讲师. 研究方向：模式识别. E-mail：zhangzhao0807@sina.com

中图分类号: V351.11
计量
- 文章访问数: 58
- HTML全文浏览量: 26
- PDF下载量: 8
- 被引次数: 0
出版历程
- 收稿日期: 2024-11-05
- 网络出版日期: 2025-09-29

A Model for Aircraft Surface Taxiing Trajectory Prediction Base on Transformer-TCN-GRU

WANG Xinglong^1
,,
LI Guoxiang²,
ZHANG Zhao^{2
, ,},
YE Ke³,
SU Ting³,
GE Jing⁴

1.
Key Laboratory of Internet of Aircrafts, Civil Aviation University of China, Tianjin 300300, China
2.
College of Computer Science and Technology, Civil Aviation University of China, Tianjin 300300, China
3.
Information service department, TravelSky Technology Limited, Beijing 100011, China
4.
Safety production and quality Management Department, TravelSky Technology Limited, Beijing 100011, China

摘要

摘要: 对于航空器滑行轨迹预测，现有方法在实时推算中等时间尺度内的未来位置精度较低，为进一步提高中等时间尺度内轨迹预测的精度，并保证实时预测的高效性，将Transformer网络、交叉注意力机制、时间卷积网络（temporal convolutional network，TCN）与门控循环单元（gated recurrent unit，GRU）相结合，构建1种输出多条候选轨迹的地面滑行轨迹预测模型。引入Transformer编码器捕捉航空器历史轨迹数据中的时间依赖性和运动状态，获取轨迹特征序列的全局特征表示；结合机场矢量地图和管制系统给出的滑行路径指令计算航空器在未来计划的滑行路径坐标序列，使用交叉注意力机制，以轨迹序列的全局特征作为查询，关注路径坐标序列中对未来滑行影响最大的位置，将融合路径特征后的轨迹全局特征映射为多种模态，对应每条候选轨迹的特征；TCN-GRU轨迹解码器对每种模态的轨迹特征进行解码，捕捉轨迹序列中的长期时间依赖，输出多条预测轨迹及其概率。以国内某大型机场航空器真实滑行轨迹进行验证，未来8 s的位置轨迹预测最小平均位移误差（minimum average displacement error，minADE）为1.932 m，最小最终位移误差（minimum final displacement error，minFDE）为1.811 m，相较于单一的GRU、TCN模型，minADE降低14.10%、16.62%，minFDE降低30.88%、34.72%，测试样本平均耗时17.70 ms，可以准确、快速预测滑行轨迹，有利于保障飞行区的安全运行。
- 滑行轨迹 /
- 轨迹预测 /
- Transformer模型 /
- 时间卷积网络 /
- 门控循环单元
Abstract: For aircraft taxiing trajectory prediction, existing methods exhibit low accuracy in real-time estimation of future positions over medium-term time horizons. To enhance prediction precision within this temporal scope while maintaining computational efficiency, this study proposes a taxiing trajectory prediction model integrating transformer networks, cross-attention mechanisms, temporal convolutional networks (TCN), and gated recurrent units (GRU) to generate multiple candidate trajectories. The Transformer encoder captures temporal dependencies and motion patterns from historical trajectory data to derive global feature representations. Airport vector maps and taxiing route instructions from air traffic control systems are utilized to compute planned future taxiing path coordinates. A cross-attention mechanism then aligns the global trajectory features (as Query) with critical positions in the planned path sequence, mapping the fused path-enhanced features into multimodal representations corresponding to candidate trajectories. The TCN-GRU decoder processes each modality to capture long-term temporal dependencies and outputs multiple predicted trajectories with associated probabilities. Validation on real taxiing trajectories from a major Chinese airport demonstrates minimum average displacement error (minADE) of 1.932 m and minimum final displacement error (minFDE) of 1.811 m for 8-second predictions. Compared to individual GRU and TCN models, the proposed approach reduces minADE/minFDE by 14.10%/30.88% and 16.62%/34.72% respectively, while maintain an average runtime of 17.70 milliseconds per sample. The proposed method achieves accurate and efficient trajectory prediction, supporting enhanced safety management in airport maneuvering areas.
- taxiing trajectory /
- trajectory prediction /
- transformer /
- temporal convolutional network /
- gated recurrent unit

HTML全文

图 1 场面滑行轨迹预测模型整体架构

Figure 1. Overall architecture of surface taxiing trajectory prediction model

下载: 全尺寸图片幻灯片

图 2 轨迹解码器数据维度变化

Figure 2. Trajectory decoder data dimension change

下载: 全尺寸图片幻灯片

图 3 机场矢量地图

Figure 3. Airport vector map

下载: 全尺寸图片幻灯片

图 4 不同预测时长的预测结果

Figure 4. Prediction results for different prediction time horizons

下载: 全尺寸图片幻灯片

图 5 不同数量预测轨迹的预测结果Fig

Figure 5. Prediction results for different numbers of predicted trajectories

下载: 全尺寸图片幻灯片

图 6 测试样例可视化

Figure 6. Test example visualization

下载: 全尺寸图片幻灯片

表 1 对比实验结果

Table 1. Comparative experimental results

模型	minADE（N=6）	minFDE（N=6）	ADE	FDE
CVM			5.725	8.583
Conv_s2s	2.672	3.537	3.771	6.376
TCN	2.317	2.774	3.540	5.916
GRU	2.249	2.620	3.480	5.773
TF			4.087	6.732
WIMP	2.094	2.187	2.828	4.108
Our	1.932	1.811	2.596	3.717

下载: 导出CSV

表 2 消融实验结果

Table 2. Ablation study results

模型	minADE（N=6）	minFDE（N=6）	ADE	FDE
Our（0）	2.256	2.653	3.399	5.632
Our（1）	1.961	1.882	2.619	3.784
Our（2）	1.940	1.863	2.619	3.874
Our（3）	1.945	1.826	2.615	3.727
Our	1.932	1.811	2.596	3.717

下载: 导出CSV

表 3 不同预测时长实验的误差和平均耗时

Table 3. Errors and average runtime of experiments with different prediction time horizons

预测时长/s	minADE（N=6）	minFDE（N=6）	平均耗时/ms
3	1.071	1.185	16.89
5	1.448	1.535	17.03
8	1.932	1.811	17.70
11	2.312	2.201	17.68
15	2.797	3.455	17.58

下载: 导出CSV

表 4 不同数量预测轨迹实验的误差和平均耗时

Table 4. Errors and average runtime of experiments with different numbers of predicted trajectories

预测轨迹数量	minADE（N=6）	minFDE（N=6）	平均耗时/ms
1	2.596	3.717	17.70
3	2.075	2.353	17.61
6	1.932	1.811	17.70
9	1.871	1.593	17.79

下载: 导出CSV

表 5 不同机型在转弯位置的预测结果

Table 5. Prediction results of different aircraft types at turning positions

航空器机型	minADE（N=6）	minFDE（N=6）
Wide_Body	2.615	2.636
Narrow_Body	2.363	2.536
Regional_Jet	2.627	2.619
NAN	2.385	2.406
738	2.449	2.628
320	2.421	2.541
320NEO	2.302	2.491
321	2.157	2.421
319	2.081	2.247
737MAX8	2.484	2.831

下载: 导出CSV

参考文献(22)

[1]	廖超伟. 航空器跑道滑行轨迹预测方法研究[J]. 舰船电子工程, 2015, 35(8): 127-129, 166. LIAO C W. Prediction method of aircraft runway slide trajectory[J]. Ship Electronic Engineering, 2015, 35 (8) : 127-129, 166. (in Chinese)
[2]	周龙. 基于4D轨迹的航空器地面滑行动态路径规划及仿真系统研究[D]. 南京: 南京航空航天大学, 2015. ZHOU L. Research on system of dynamic route planning and simulation for taxiing aircraft based on 4D trajectory[D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2015. (in Chinese)
[3]	ZHANG X, YU W. Research on the application of kalman filter algorithm in aircraft trajectory analysis[C]. IEEE 7th International Conference on Intelligent Computing and Signal Processing, Xi'an: IEEE, 2022.
[4]	吕波, 王超. 改进的扩展卡尔曼滤波在航空器4D航迹预测算法中的应用[J]. 计算机应用, 2021, 41(增刊1): 277-282. LYU B, WANG C. Application of improved extended kalman filtering in aircraft 4D trajectory prediction algorithm[J]. Journal of Computer Applications, 2021, 41(S1): 277-282. (in Chinese)
[5]	陈明强, 傅嘉赟. 基于无迹卡尔曼滤波的飞行航迹预测方法研究[J]. 计算机仿真, 2021, 38(6): 27-30, 36. CHENG M Q, FU J Y. Flight track prediction method based on unscented kalman filter[J]. Computer Simulation, 2021, 38 (6): 27-30, 36. (in Chinese)
[6]	王新, 杨任农, 左家亮, 等. 基于HPSO-TPFENN的目标机轨迹预测[J]. 西北工业大学学报, 2019, 37(3): 612-620. WANG X, YANG R N, ZUO J L, et al. Trajectory prediction of target aircraft based on HPSO-TPFENN neural network[J]. Journal of Northwestern Polytechnical University, 2019, 37 (3): 612-620. (in Chinese)
[7]	MA L, TIAN S. A hybrid CNN-LSTM model for aircraft 4D trajectoryprediction[J]. IEEEAccess, 2020, (8): 134668-134680.
[8]	SHI Z Y, XU M, PAN Q. 4-D flight trajectory prediction with constrained LSTM network[J]. IEEE Transactions on Intelligent Transportation Systems, 2020, 22(11): 7242-7255.
[9]	ZENG W L, QUAN Z B, ZHAO Z Y, et al. A deep learning approach for aircraft trajectory prediction in terminal airspace[J]. IEEE Access, 2020, (8): 151250-151266.
[10]	MA Z M, YAO M F, HONG T, et al. Aircraft surface trajectory prediction method based on LSTM with attenuated memory window[J]. Journal of Physics: Conference Series, 2019, 1215(1): 012003. doi: 10.1088/1742-6596/1215/1/012003
[11]	杨春伟, 刘炳琪, 王继平, 等. 基于注意力机制的高超声速飞行器LSTM智能轨迹预测[J]. 兵工学报, 2022, 43(增刊2): 78-86. YANG C W, LIU B Q, WANG J P, et al. LSTM intelligent trajectory prediction for hypersonic vehicles based on attention mechanism[J]. Acta Armamentarii, 2022, 43(S2): 78-86. (in Chinese)
[12]	ZHANG X G, ZHONG S Q, MAHADEVAN S. Airport surface movement prediction and safety assessment with spa-tial-temporal graph convolutional neural network[J]. Transportation Research Part C: Emerging Technologies, 2022, 144: 103873. doi: 10.1016/j.trc.2022.103873
[13]	刘雨生, 汤新民, 任宣铭. 基于注意力机制的GRU-IKF场面滑行轨迹预测模型[J]. 北京航空航天大学学报, 2025, 51 (3): 1028-1036. LIU Y S, TANG X M, REN X M. Prediction of aircraft surface trajectory based on the GRU-IKF model with attention mechanism[J]. Journal of Beijing University of Aeronautics and Astronautics, 2025, 51(3): 1028-1036. (in Chinese)
[14]	王兴隆, 许晏丰. 基于AM-LSTM的飞行区航空器滑行轨迹预测与冲突识别[J]. 中国安全科学学报, 2024, 34(1): 116-124. WANG X L, XU Y F. Aircraft taxiing trajectory prediction and conflict risk identification in airfield area based on AM-LSTM[J]. China Safety Science Journal, 2024, 34(1): 116-124. (in Chinese)
[15]	殷萌暄, 胡明华, 尹嘉男, 等. 基于多元嵌入特征的航空器场面滑行轨迹预测[J]. 航空计算技术, 2024, 54(5): 58-63. YIN M X, HU M H, YIN J N, et al. Aircraft taxiing trajectory prediction based on multi-embedding features[J]. Aeronautical Computing Technique, 2024, 54(5): 58-63. (in Chinese)
[16]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: transformers for image recognition at scale[EB/OL]. (2021-06)[2024-11-04]. https://arxiv.org/abs/2010.11929.
[17]	中国民用航空局. 民用机场飞行区技术标准: MH 5001-2021[S]. 北京: 中国民航出版社, 2021. Civil Aviation Administration of China. Technical standards for airfield area of civil airports: MH 5001-2021[S]. Beijing: China Civil Aviation Publishing House, 2021. (in Chinese)
[18]	ZHOU Z K, YE L Y, WANG J P, et al. Hivt: hierarchical vector transformer for multi-agent motion prediction[C]. IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA: IEEE, 2022.
[19]	GEHRING J, AULI M, GRANGIER D, et al. Convolutional sequence to sequence learning[C]. International Conference on Machine Learning, Sydney, Australia: PMLR, 2017.
[20]	BAI S J, KOLTER J Z, KOLTUN V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling[EB/OL]. (2018-04)[2024-11-04]. https://arxiv.org/abs/1803.01271
[21]	GIULIARI F, HASAN I, CRISTANI M, et al. Transformer networks for trajectory forecasting[C]. 25th International Conference on Pattern Recognition, Milan, Italy: IEEE, 2021.
[22]	KHANDELWAL S, QI W, SINGH J, et al. What-if motion prediction for autonomous driving[EB/OL]. (2020-08)[2024-11-04]. https://arxiv.org/abs/2008.10587.