基于动态图自注意力的车流参数预测方法

doi:10.3969/j.issn.1674-8484.2024.05.006

摘要/Abstract

摘要：

为了提高异常事件常发地段中智能车辆行驶的效率和安全性，以提升车流参数预测的准确度为出发点，该文设计了一种基于动态节点自注意力的车流参数预测方法, 在多个时间步中利用空间注意力聚合邻域节点的特征，沿着时间维度通过时间注意力机制预测交通参数。结果表明：该文设计的动态图自注意力（DGSA）模型的1 h预测结果平均绝对误差（MAE）、均方根误差（RMSE）和平均绝对百分误差（MAPE）指标分别下降了3.75%、3.45%、11.63%；测算的路段平均碰撞时间（TTC）更长，达到2.8 s。该方法能够在异常事件情况下有效预测车流演化态势并提升车辆的安全性。

关键词: 智能车辆, 车流参数预测, 异常事件, 动态图, 深度学习

Abstract:

In order to improve the driving efficiency and safety of intelligent vehicles in the areas with frequent abnormal events, a traffic flow parameter prediction method was designed based on dynamic node self-attention to improve the accuracy of traffic flow parameter prediction. The spatial attention was used to aggregate the features of neighborhood nodes in multiple time steps, and then the traffic parameters were predicted by the temporal attention mechanism along the time dimension. The results show that the Mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) of the one-hour prediction of Dynamic Picture Self-Attention (DGSA) model decrease by 3.75%, 3.45% and 11.63%, respectively. The simulated road average collision time (TTC) is longer, reaching 2.8 s. The proposed method can effectively predict the evolution trend of traffic flow and improve the safety of vehicles under abnormal events.

Key words: intelligent vehicles, traffic flow parameter prediction, abnormal events, dynamic graphs, deep learning

中图分类号:

TP391

石天京, 李旭. 基于动态图自注意力的车流参数预测方法[J]. 汽车安全与节能学报, 2024, 15(5): 680-688.

SHI Tianjing, LI Xu. Traffic flow parameter prediction method based on dynamic graphs self-attention[J]. Journal of Automotive Safety and Energy, 2024, 15(5): 680-688.

图/表 10

参考文献 27

[1]	张亚勤, 李震宇, 尚国斌, 等. 面向自动驾驶的车路云一体化框架[J]. 汽车安全与节能学报, 2023, 14(3): 249-273.
	ZHANG Yaqin, LI Zhenyu, SHANG Guobin, et al. A unified framework for vehicle-infrastructure-cloud autonomous driving[J]. J Autom Safe Energ, 2023, 14(3): 249-273. (in Chinese)
[2]	杨帅, 张金换, 钱占伟, 等. 汽车安全多领域融合的研究与展望[J]. 汽车安全与节能学报, 2022, 13(1): 29-47.
	YANG Shuai, ZHANG Jinhuan, QIAN Zhanwei, et al. Research and prospect of multi domain integration of automobile safety[J]. J Autom Safe Energ, 2022, 13(1): 29-47. (in Chinese)
[3]	YU Bing, YIN Haoteng, ZHU Zhanxing. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting[C]// Proc 27th Int’l Joint Conf Artif Intel (IJCAI). Stockholm: AAAI Press, 2018: 3634-3640.
[4]	耿彦斌. 基于事件惩罚的公路应急交通动态分配模型[J]. 交通运输系统工程与信息, 2016, 16(5): 129-135.
	GENG Yanbin. An emergency penalty-based dynamic highway traffic assignment model[J]. J Transport Syst Engi Info Tech, 2016, 16(5): 129-135. (in Chinese)
[5]	秦严严, 王昊, 王炜. 智能网联环境下的混合交通流LWR模型[J]. 中国公路学报, 2018, 31(11): 147-156.
	QIN Yanyan, WANG Hao, WANG Wei. LWR model for mixed traffic flow in connected and autonomous vehicular environments[J]. Chin J Highw Transport, 2018, 31(11): 147-156. (in Chinese)
[6]	梁海军, 骆春阳, 程万臻, 等. 基于元胞传输模型的终端区交通流特性研究[J]. 航空计算技术, 2024, 54(2): 16-19+25.
	LIANG Haijun, LUO Chunyang, CHENG Wangzhen, et al. Analysis of terminal traffic flow characteristics based on CTM[J]. Aeronaut Comput Tech, 2024, 54(2): 16-19+25. (in Chinese)
[7]	邢吉平. 基于多源数据融合的城市路网交通流量估计方法研究[D]. 南京: 东南大学, 2021.
	XING Jiping. Study on the urban network volume estimation based on multi-source data fusion[D]. Nanjing: Southeast University, 2021. (in Chinese)
[8]	Williams B M, Hoel L A. Modeling and forecasting vehicular traffic flow as a seasonal ARIMA process: Theoretical basis and empirical results[J]. J Transport Engi, 2003, 129(6): 664-672.
[9]	樊路, 钱雪忠, 姚琳燕. 基于K近邻的增量式聚类算法[J]. 传感器与微系统, 2019, 38(2): 136-139.
	FAN Lu, QIAN Xuezhong, YAO Linyan. Incremental clustering algorithm based on K-nearest neighbor[J]. Transducer Microsyst Tech, 2019, 38(2): 136-139. (in Chinese)
[10]	HU Wenbin, YAN Liping, LIU Kaizeng, et al. A short-term traffic flow forecasting method based on the hybrid PSO-SVR[J]. Neural Proc Letts, 2016, 43(1): 155-172.
[11]	ZHAO Zheng, CHEN Weihai, WU Xingming, et al. LSTM network: a deep learning approach for short-term traffic forecast[J]. IET Intel Transport Syst, 2017, 11(2): 68-75.
[12]	王体迎, 时鹏超, 刘蒋琼, 等. 基于门限递归单元循环神经网络的交通流预测方法研究[J]. 重庆交通大学学报(自然科学版), 2018, 37(11): 76-82. doi: 10.3969/j.issn.1674-0696.2018.11.13
	WANG Tiying, SHI Pengchao, LIU Jiangqiong, et al. Research on traffic flow prediction method based on gated recurrent unit recurrent neural network[J]. J Chongqing Jiaotong Univ (Nat Sci), 2018, 37(11): 76-82. (in Chinese)
[13]	WU Zonghan, PAN Shirui, LONG Guodong, et al. Graph wavenet for deep spatial-temporal graph modeling[C]// The 28th Int'l Join Conf Artif Intel (IJCAI). Macau, China, 2019: 1907-1913.
[14]	LI Yaguang, YU Rose, Shahabi C, et al. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting[C]// Int’l Conf Learn Represent (ICLR). Vancouver, Canada, 2018: 1-16
[15]	ZHENG Chuanpan, FAN Xiaoliang, WANG Chengheng, et al. Gman: A graph multi-attention network for traffic prediction[C]// Proc AAAI Conf Artif Intel. 2020, 34(1): 1234-1241.
[16]	WANG Jingcheng, ZHANG Yong, WANG Lixun, et al. Multitask hypergraph convolutional networks: A heterogeneous traffic prediction framework[J]. IEEE Trans Intel Transport Syst, 2022, 23(10): 18557-18567.
[17]	YANG Wang, ZHENG Jin, DU Yuqi, et al. Traffic-GGNN: Predicting traffic flow via attentional spatial-temporal gated graph neural networks[J]. IEEE Trans Intel Transport Syst, 2022, 23(10): 18423-18432.
[18]	YANG Leshanshui, Adam S, Chatelain C. Dynamic graph representation learning with neural networks: A survey[J]. IEEE Access, 2024, 12: 43460-43484.
[19]	Gehring J, Auli M, Grangier D, et al. Convolutional sequence to sequence learning[C]// Int’l Conf Mach Learn. PMLR, 2017: 1243-1252.
[20]	Petar V, Guillem C, Arantxa C, et al. Graph attention networks[C]// Int’l Conf Learn Represent (ICLR). Vancouver, Canada, 2018.
[21]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]// Proc 31st Int’l Conf Neural Info Proc Syst. Long Beach: ACM, 2017: 6000-6010.
[22]	Sutskever I, Vinyals O, Le Q V. Sequence to sequence learning with neural networks[J]. Advan Neural Info Proc Syst, 2014, 27(2): 3104-3112.
[23]	Cho K, Van Merriënboer B, Gulcehre C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]// Proceed 2014 Conf Empir Meth Nat Langu Proc (EMNLP). Doha, Qatar, 2014: 1724-1734.
[24]	Wikipedia contributors. Mean absolute error[G/OL]. Wikipedia, 2022. [2022-06-24]. https://en.wikipedia.org/wiki/Mean_absolute_error.
[25]	Wikipedia contributors. Root mean square error[G/OL]. Wikipedia, 2022. [2022-09-29]. https://en.wikipedia.org/wiki/Root-mean-square_deviation.
[26]	Wikipedia contributors. Mean absolute percentage error [G/OL]. Wikipedia, 2023. [2023-11-22]. https://en.wikipedia.org/wiki/Mean_absolute_percentage_error.
[27]	黄新朝, 张毅. 基于百度地图API的纯电动汽车未来行驶能耗预测[J]. 汽车安全与节能学报, 2023, 14(6): 715-722.
	HUANG Xinchao, ZHANG Yi. Prediction of future driving conditions for electrical vehicles based on Baidu maps API[J]. J Autom Safe Energ, 2023, 14(6): 715-722. (in Chinese)

预测模型		MAE			RMSE			MAPE / %
预测模型	15 min	30 min	60 min	15 min	30 min	60 min	15 min	30 min	60 min
ARIMA	2.369 4	2.660 4	3.737 0	4.149 7	5.004 7	7.300 5	4.890 2	5.418 9	6.413 7
LSTM	2.056 7	2.139 2	2.881 7	3.893 1	4.557 3	6.378 7	4.147 5	4.736 9	6.070 8
GRU	1.717 4	2.071 3	3.177 2	3.426 3	4.455 4	6.700 7	3.526 6	4.190 2	6.294 2
ST-GCN	1.983 6	2.330 2	3.069 8	3.820 5	4.348 4	6.082 0	3.193 6	3.912 7	5.513 7
WaveNet	1.964 3	2.283 1	2.761 4	3.367 4	4.002 6	5.620 3	3.105 2	3.803 2	5.127 5
DGSA	1.671 0	1.946 2	2.657 9	3.104 5	3.771 2	5.426 3	2.713 4	3.448 9	4.531 0

预测模型		MAE			RMSE			MAPE / %
预测模型	15 min	30 min	60 min	15 min	30 min	60 min	15 min	30 min	60 min
ARIMA	2.369 4	2.660 4	3.737 0	4.149 7	5.004 7	7.300 5	4.890 2	5.418 9	6.413 7
LSTM	2.056 7	2.139 2	2.881 7	3.893 1	4.557 3	6.378 7	4.147 5	4.736 9	6.070 8
GRU	1.717 4	2.071 3	3.177 2	3.426 3	4.455 4	6.700 7	3.526 6	4.190 2	6.294 2
ST-GCN	1.983 6	2.330 2	3.069 8	3.820 5	4.348 4	6.082 0	3.193 6	3.912 7	5.513 7
WaveNet	1.964 3	2.283 1	2.761 4	3.367 4	4.002 6	5.620 3	3.105 2	3.803 2	5.127 5
DGSA	1.671 0	1.946 2	2.657 9	3.104 5	3.771 2	5.426 3	2.713 4	3.448 9	4.531 0

预测模型	TTC / s
ARIMA	1.3
LSTM	1.9
GRU	2.6
ST-GCN	2.2
WaveNet	2.3
DGSA	2.8

预测模型	TTC / s
ARIMA	1.3
LSTM	1.9
GRU	2.6
ST-GCN	2.2
WaveNet	2.3
DGSA	2.8