基于Informer的客机长时4D航迹预测方法

冯霞; 孙琦琦; 左海超

doi:10.3963/j.jssn.1674-4861.2023.04.012

基于Informer的客机长时4D航迹预测方法

doi: 10.3963/j.jssn.1674-4861.2023.04.012

冯霞^{1, 2, ,},
孙琦琦^{1, 2},
左海超^{1, 2}

1.
中国民航大学计算机科学与技术学院天津 300300
2.
中国民航大学民航智慧机场理论与系统重点实验室天津 300300

基金项目:

国家重点研发计划项目 2021YFF0603902

中央高校基本科研业务费中国民航大学专项 3122021063

详细信息

通讯作者:
冯霞（1970—），博士，教授. 研究方向：民航智能信息处理、智慧机场. E-mail：caucfx@126.com

中图分类号: TP391
计量
- 文章访问数: 280
- HTML全文浏览量: 108
- PDF下载量: 27
- 被引次数: 0
出版历程
- 收稿日期: 2022-12-26
- 网络出版日期: 2023-11-23

A Method for Predicting Long-term 4D Trajectory of Airplanes Based on Informer

FENG Xia^{1, 2
, ,},
SUN Qiqi^{1, 2},
ZUO Haichao^{1, 2}

1.
College of Computer Science and Technology, Civil Aviation University of China, Tianjin 300300, China
2.
Key Laboratory of Smart Airport Theory and System, Civil Aviation University of China, Tianjin 300300, China

摘要

摘要: 客机长时4D航迹预测是基于航迹运行的重要基础，对于改善空中交通系统安全性能和优化空域结构有重要意义。针对现有长时4D航迹预测未充分考虑长序列航迹数据之间存在隐式关联信息等问题，借助Informer模型的自注意力机制，研究构建了基于Informer的长时4D航迹预测模型。为提取航迹数据的全局特征信息，增强数据独立性和时间序列特征学习能力，在数据嵌入层中增加全局时间戳模块，并利用航迹点序列等分层时间戳突破Informer模型固有的嵌入层时间刻度限制；为更好地捕捉非相邻时序序列点之间的隐式相关性，采用自注意力机制提取航迹数据特征，并运用概率稀疏方法降低自注意力机制的计算复杂度至O（LlogL），同时在编码器中增加蒸馏机制以减少计算维度和网络参数量；为避免传统的逐步预测输出方法造成的误差累积现象，提高航迹预测精度，采用全连接层对预测输出数据进行维度调整，完成一步生成式输出。对历史4D航迹数据进行三次样条插值等预处理后，与时序特征数据同时输入到航迹预测模型中，经过模型迭代训练，输出航迹预测结果。实验结果表明：在同时预测航迹的4D特征时，基于Informer模型的预测表现优于LSTnet方法，其均方根误差和欧氏距离误差分别为0.218 5和15.980 km，相较于LSTnet网络分别减少了1.48%和2.44%。此外，对于分别预测航迹特征的任务，Informer模型的欧氏距离误差为13.248 km，相较于LSTnet网络减少了3.11%，相较于传统LSTM网络减少了34.99%。
- 智能交通 /
- 4D航迹 /
- 航迹预测 /
- Informer /
- 时间序列 /
- 自注意力机制
Abstract: Prediction of long-term 4D trajectory is an important foundation for trajectory-based operation, which is significant for improving safety of air transportation system and optimizing airspace. The existing methods for predicting long-term 4D trajectory do not fully consider implicit association among trajectory data with a long sequence. To address this problem, a long-term 4D trajectory prediction model based on Informer model with the self-attention mechanism is developed. To extract the global feature from trajectory data, enhance data independence and the capability to learn the feature of time series, a global timestamp module is added into the data embedding layer. Moreover, the layered timestamps, such as trajectory point sequences, are utilized to overcome the inherent time scale limitation of the Informer model. To better capture the implicit correlations between non-adjacent temporal sequence points, a self-attentive mechanism is employed to extract the features of trajectory data, and a probabilistic sparse method is applied to reduce the computational complexity of the self-attentive mechanism to O(LlogL). Additionally, a distillation mechanism is incorporated into the encoder to reduce the computational dimensions and the number of network parameters. To avoid the error accumulation arising from traditional step-by-step prediction models and improve the accuracy of trajectory prediction, a fully connected layer is applied to adjust the dimensions of the predicted data, achieving one-step generative output. After three-time spline interpolation, the pre-processed historical 4D trajectory data are inputted to the trajectory prediction model along with the data presenting the feature of time-series. Through iterative training of the model, the trajectory prediction results are generated and output. Study results show that, the Informer-based model outperforms the LSTnet method when predicting the trajectory of 4D features simultaneously. The root mean square error and Euclidean distance error is 0.2185 and 15.980 km, respectively, which is a reduction of 1.48% and 2.44% compared to that of the LSTnet network. In addition, when predicting the trajectory features separately, the Euclidean distance error of the Informer-based model is 13.248 km, with a reduction of 3.11% compared to the LSTnet network and a reduction of 34.99% compared to the traditional LSTM network.
- intelligent transportation /
- 4D trajectory /
- trajectory prediction /
- Informer /
- time series /
- self-attention mechanism

HTML全文

图 1 航迹数据对齐预处理对比示例图

Figure 1. Data alignment pre-processing comparison example chart

下载: 全尺寸图片幻灯片

图 2 构造输入和航迹预测模型结构示例图

Figure 2. Example diagram for constructing input and the structure of trajectory prediction model

下载: 全尺寸图片幻灯片

图 3 输入嵌入层示例图

Figure 3. Example diagram of input embedding layer

下载: 全尺寸图片幻灯片

图 4 实验步骤

Figure 4. Experimental steps

下载: 全尺寸图片幻灯片

图 5 模型预测结果（同时预测）

Figure 5. Model prediction results(simultaneous prediction)

下载: 全尺寸图片幻灯片

图 6 航迹3各模型预测结果（分别预测）

Figure 6. Model prediction results for trajectory 3 (separate prediction)

下载: 全尺寸图片幻灯片

表 1 ADS-B原始数据示例

Table 1. Examples of ADS-B raw data

采样时间	航班号	高度/m	速度/（km/h）	航向/（°）	纬度/（°）	经度/（°）
2021-04-03 T12：37：01	MF8506	5 562.60	594.492	254	30.999 6	120.841 0
2021-04-03 T12：37：44	MF8506	6 111.24	598.196	255	30.979 2	120.759 1
2021-04-03 T12：38：44	MF8506	6 812.28	622.272	254	30.963 3	120.695 0
2021-04-03 T12：40：24	MF8506	7 757.16	657.460	254	30.916 7	120.509 5
2021-04-03 T12：41：20	MF8506	8 183.88	729.688	216	30.844 8	120.387 0
2021-04-03 T12：42：17	MF8506	8 770.62	709.316	235	30.767 9	120.295 2
2021-04-03 T12：42：33	MF8506	8 770.62	709.316	235	30.767 9	120.295 2

下载: 导出CSV

表 2 时序特征处理后的航迹数据示例

Table 2. Example of trajectory data after processing of temporal features

年	月	日	时	分	秒	周	航班号	航迹点序列	高度/m	速度/（km/h）	航向/（°）	纬度/（°）	经度/（°）
2021	4	3	12	37	44	6	MF8506	09	6 111.24	598.19	255	30.979	120.759
2021	4	3	12	39	00	6	MF8506	10	7 010.40	629.68	254	30.951	120.649
2021	4	3	12	40	16	6	MF8506	11	7 642.86	646.34	254	30.917	120.510
2021	4	3	12	41	32	6	MF8506	12	8 206.74	731.54	216	30.833	120.377
2021	4	3	12	42	48	6	MF8506	13	8 869.68	711.16	235	30.767	120.295
2021	4	3	12	44	04	6	MF8506	14	9 357.36	774.13	216	30.685	120.185
2021	4	3	12	45	20	6	MF8506	15	9 723.12	792.65	216	30.539	120.065

下载: 导出CSV

表 3 实验环境

Table 3. Experimental environment

硬件环境	处理器	Intel（R）Xeon（R）Silver4214@2.20GHz
	内存	64GB
	显卡	NVIDIA Quadro RTX 5000
软件环境	操作系统	Windows 10
	编程语言	Python 3.7
	基础框架	Pytorch 1.10
	编程软件	Pycharm

下载: 导出CSV

表 4 模型超参数

Table 4. Model hyperparameters

超参数	设置
编码器输入步数	240
解码器输入步数	120
预测步数	60
训练周期	20
学习率	0.000 1
批尺寸	32
损失函数	MSE
激活函数	gelu
优化器	Adam
防止过拟合	EarlyStop

下载: 导出CSV

表 5 对比实验模型超参数设置

Table 5. Hyperparameter settings of the model for the comparison experiment

超参数	LSTnet	LSTM	BiLSTM	Seq2Seq
输入步数	240	240	240	240
预测步数	60	60	60	60
隐藏层维度	100	64	64	64
卷积核大小	6	—	—	—
丢弃率	0.1	—	—	—
训练周期	20	20	20	20
学习率	0.000 1	0.000 1	0.000 1	0.000 1
批尺寸	32	32	32	32
损失函数	MSE	MSE	MSE	MSE
激活函数	gelu	gelu	gelu	gelu
优化器	Adam	Adam	Adam	Adam
注：“—”为该模型不具备该超参数，无需设置相关数据。

下载: 导出CSV

表 6 时间步数(N)参数敏感性实验

Table 6. Time step (N) parameter sensitivity experiments

时间步数	MAE	MSE	RMSE	欧氏距离/km
40	0.196 0	0.089 0	0.298 3	20.225
60	0.142 4	0.047 7	0.218 5	15.980
80	0.150 2	0.049 3	0.222 0	16.366
100	0.153 9	0.056 1	0.236 8	16.919

下载: 导出CSV

表 7 Seq_len与Label_len多组合参数敏感性实验

Table 7. Experiments on parameter sensitivity of various combinations of Seq_len and Label_len

Seq_len	Label_len	MAE	MSE	RMSE	欧氏距离/km
	2×N	0.157 8	0.056 0	0.236 6	17.870
2×N	4×N	—	—	—	—
	8×N	—	—	—	—
	2×N	0.142 4	0.047 7	0.218 5	15.980
4×N	4×N	0.143 1	0.049 4	0.222 3	16.032
	8×N	—	—	—	—
	2×N	0.144 9	0.049 4	0.222 3	16.307
8×N	4×N	0.145 0	0.050 7	0.225 2	16.524
	8×N	0.152 0	0.053 6	0.231 5	17.066
	2×N	0.149 6	0.053 1	0.230 4	17.118
12×N	4×N	0.161 1	0.067 4	0.259 6	18.408
	8×N	0.163 8	0.078 8	0.280 7	19.279
注：“—”为该模型因其网络搭建的特质，不能设置该参数组合。

下载: 导出CSV

表 8 同时预测实验模型评价指标对比

Table 8. Comparison of model evaluation metrics during uniform prediction experiments

预测模型	MAE	MSE	RMSE	欧氏距离/km
Informer	0.142 4	0.047 7	0.218 5	15.980
LSTnet	0.158 8	0.049 2	0.221 8	16.381

下载: 导出CSV

表 9 分别预测实验模型评价指标对比

Table 9. Comparison of model evaluation metrics in separate prediction experiments

预测模型	预测维度	MAE	MSE	RMSE	欧氏距离/km
	高度	0.216 6	0.080 2	0.283 2
Informer	经度	0.132 6	0.032 4	0.180 0	13.248
	纬度	0.054 2	0.005 1	0.071 9
	高度	0.274 0	0.140 8	0.375 2
LSTnet	经度	0.112 1	0.020 1	0.141 8	13.673
	纬度	0.072 1	0.009 4	0.097 0
	高度	0.236 1	0.101 9	0.319 2
LSTM	经度	0.312 4	0.297 7	0.545 6	20.378
	纬度	0.142 0	0.028 0	0.167 3
	高度	0.229 6	0.094 7	0.307 7
BiLSTM	经度	0.417 2	0.687 2	0.829 0	29.410
	纬度	0.146 9	0.035 7	0.188 9
	高度	0.236 9	0.098 0	0.313 0
seq2seq	经度	0.369 7	0.308 1	0.555 1	27.532
	纬度	0.150 7	0.040 9	0.202 2

下载: 导出CSV

表 10 使用和不使用飞行计划实验误差对比

Table 10. Comparison of experimental errors with and without the use of flight plans 单位: km

预测模型	飞行计划数据	逐点水平误差	逐点垂直误差
Informer	不使用	13.205 4	0.139 0
LSTM-高斯混合概率估计	使用	91.859 2	0.864 1

下载: 导出CSV

参考文献(22)

[1]	WANG Z Y, LIANG M, DANIEL D. A hybrid machine learning model for short-term estimated time of arrival prediction in terminal manoeuvring area[J]. Transportation Research Part C: Emerging Technologies, 2018, 95: 280-294. doi: 10.1016/j.trc.2018.07.019
[2]	GUAN X M, ZHANG X J, HAN D, et al. A strategic flight conflict avoidance approach based on a memetic algorithm[J]. Chinese Journal of Aeronautics, 2014, 27(1): 93-101. doi: 10.1016/j.cja.2013.12.002
[3]	徐正凤, 曾维理, 羊钊. 航空器轨迹预测技术研究综述[J]. 计算机工程与应用, 2021, 57(12): 65-74. https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202112008.htm XU Z F, ZENG W L, YANG Z. Survey of civil aircraft trajectory prediction[J]. Computer Engineering and Applications, 2021, 57(12): 65-74. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202112008.htm
[4]	王涛波, 黄宝军. 基于改进卡尔曼滤波的四维飞行航迹预测模型[J]. 计算机应用, 2014, 34(6): 1812-1815. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY201406065.htm WANG T B, HUANG B J. 4D flight trajectory prediction model based on improved Kalman filter[J]. Journal of Computer Applications, 2014, 34(6): 1812-1815. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY201406065.htm
[5]	AYHAN S, SAMET H. Aircraft trajectory prediction made easy with predictive analytics[C]. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York: Association for Computing Machinery, 2016.
[6]	ZHANG J F, WU X G, WANG F. Aircraft trajectory prediction based on modified interacting multiple model algorithm[J]. Journal of Donghua University(English Edition), 2015, 32: 180-184.
[7]	LEE J, LEE S, HWANG I. Hybrid system modeling and estimation for arrival time prediction in terminal airspace[J]. Journal of Guidance Control Dynamics, 2016, 39(4): 1-8.
[8]	WU X P, YANG H Y, CHEN H, et al. Long-term 4D trajectory prediction using generative adversarial networks[J]. Transportation Research Part C: Emerging Technologies, 2022, 136: 103554. doi: 10.1016/j.trc.2022.103554
[9]	赵子瑜. 基于深度LSTM的四维航迹预测方法及应用[D]. 南京: 南京航空航天大学, 2020. ZHAO Z Y. Four-dimensional trajectory prediction based on deep LSTM and application[D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2020. (in Chinese)
[10]	LIU Y L, HANSEN M. Predicting aircraft trajectories: a deep generative convolutional recurrent neural networks approach[R/OL]. (2018-12-31)[2023-05-11]. https://arxiv.org/abs/1812.11670.
[11]	PANG Y T, XU N, LIU Y M. Aircraft trajectory prediction using LSTM neural network with embedded convolutional layer[C]. Annual Conference of the PHM Society, Scottsdale: PHM Society, 2019.
[12]	PANG Y T, YAO H P, HU J M, et al. A recurrent neural network approach for aircraft trajectory prediction with weather features from sherlock[C]. AIAA Aviation 2019 Forum, Reston: AIAA, 2019.
[13]	SCHIMPF N, KNOBLOCK E J, WANG Z, et al. A generalized approach to aircraft trajectory prediction via supervised deep learning[J/OL]. IEEE Transactions on Aerospace and Electronic Systems. (2022-4-1)[2023-05-11]. https://ntrs.nasa.gov/citations/20220002176.
[14]	ZHOU H Y, ZHANG S H, PENG J Q, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[C]. 35th AAAI Conference on Artificial Intelligence. Online: AAAI, 2021.
[15]	GONG M J, ZHAO Y, SUN J W, et al. Load forecasting of district heating system based on Informer[J]. Energy, 2022, 253: 124179. doi: 10.1016/j.energy.2022.124179
[16]	张友, 李聪波, 林利红, 等. 数据不完备下基于Informer的离心鼓风机故障趋势预测方法[J]. 计算机集成制造系统, 2023, 29(1): 133-145. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJJ202301012.htm ZHANG Y, LI C B, LIN L H, et al. Centrifugal blower fault trend prediction method based on informer with incomplete data[J]. Computer Integrated Manufacturing Systems, 2023, 29(1): 133-145. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-JSJJ202301012.htm
[17]	宫淑丽, 吴红兰, 王旭辉, 等. 基于ADS-C的机场场面飞机监视系统研究[J]. 导航与控制, 2015, 14(2): 51-56. GONG S L, WU H L, WANG X H, et al. Aircraft surveillance system on airport surface based on ADS-C[J]. Navigation and Control, 2015, 14(2): 51-56. (in Chinese)
[18]	张小江, 高秀华. 三次样条插值在机器人轨迹规划应用中的改进研究[J]. 机械设计与制造, 2008, 9: 170-171. https://www.cnki.com.cn/Article/CJFDTOTAL-JSYZ200809073.htm ZHANG X J, GAO X H. The research on the cubic splines in robot' s trajectory planning[J]. Machinery Design and Manufacture, 2008, 9: 170-171. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-JSYZ200809073.htm
[19]	ZHANG S, ZHENG D Q, HU X C, et al. Bidirectional long short-term memory networks for relation classification[C]. 29th Pacific Asia Conference on Language, Information and Computation, Shanghai: Shanghai Jiao Tong University, 2015.
[20]	SUTSKEVER I, VINYALS O, LE Q V. Sequence to sequence learning with neural networks[J]. Advances in Neural Information Processing Systems, 2014, 27: 3104-3112.
[21]	LAI G K, CHANG W C, YANG Y M, et al. Modeling long and short-term temporal patterns with deep neural networks[C]. 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, Michigan: ACM, 2018.
[22]	夏添. 窄带组网雷达航迹贯序连接和目标匹配识别技术研究[D]. 哈尔滨: 哈尔滨工业大学, 2015. XIA T. Research on the technology of track sequential connection and target matching identification based on narrowband radar network[D]. Harbin: Harbin Institute of Technology, 2015. (in Chinese)