基于时空多视野注意残差网络的城市区域交通流量预测

doi:10.16182/j.issn1004731x.joss.23-1313

摘要/Abstract

摘要：

为高效、全面提取城市中复杂的时空相关性，提出一种新的端到端的深度学习框架—时空多视野注意残差网络（spatiotemporal multi-view attention residual network， ST-MVAR），用于城市区域交通流量预测。整合交通流量的临近性、周期性、趋势性和外部因素作为网络输入，该网络通过跳跃连接，形成多层嵌套残差网络结构；设计多视野扩展模块，用于捕获交通流量对不同距离的空间依赖；引入坐标注意力机制，有效建立交通流量的时空相关性；通过K-Means聚类方法获取各时段交通流量所属模式，作为额外特征，进一步提高模型的预测精度。实验结果表明：ST-MVAR使用更少的参数获得更高的性能，相比之前的方法RMSE降低14.2%。

关键词: 交通流量预测, 残差网络, 视野扩展, 坐标注意力, K-Means聚类

Abstract:

. However, efficiently and comprehensively capturing the complex spatiotemporal correlations within urban traffic flow presents a key challenge. Existing research methods struggle to fully capture these spatiotemporal dependencies. To address these issues, we propose a novel end-to-end deep learning framework called the spatiotemporal multi-view attention residual network (ST-MVAR) for predicting traffic flow in urban areas. we integrate the proximity, periodicity, trend, and external factors of traffic flow as inputs to the network. This network employs skip connections to form a multi-layer nested residual network structure. Additionally, we design a Multi-View Extension module to capture spatial dependencies of traffic flow at various distances and introduce a coordinate attention network to effectively establish the spatiotemporal correlations within traffic flow. Furthermore, we use the k-means clustering method to obtain patterns for each cross-sectional time traffic flow and incorporate them as additional features to further enhance the model's predictive accuracy. Experimental results demonstrate that ST-MVAR achieves higher performance with fewer parameters, 14.2% lower RMSE compared to the best previous methods.

Key words: traffic flow prediction, residual network, view extension, coordinate attention, K-Means clustering

中图分类号:

TP183

陈静,杨国威,张昭冲等 . 基于时空多视野注意残差网络的城市区域交通流量预测[J]. 系统仿真学报, 2025, 37(3): 607-622.

Chen Jing,Yang Guowei,Zhang Zhaochong,et al . City Regional Traffic Flow Prediction Based on Spatiotemporal Multi-view Attention Residual Network[J]. Journal of System Simulation, 2025, 37(3): 607-622.

图/表 22

图1

图2

图3

图4

图5

图6

图7

图8

图9

表1

北京出租车数据集和纽约共享单车数据集描述

数据集		TaxiBJ	BikeNYC
交通流	数据类型	出租车GPS数据	单车租赁数据
	时间范围	2013-07-01—10-30	2014-04-01—09-30
		2014-03-01—06-30
		2015-03-01—06-30
		2015-11-01—2016-04-10
	时间间隔/min	30	60
	网格尺寸	(32, 32)	(16, 8)
	时间间隔总数	22 459	4 392
	出租车/自行车数量	34 000+	6 800+
外部因素	假期	41种节假日	20
	天气状况	16种天气类型	/
	温度/℃	[ $-$ 24.6,41.1]	/
	风速/mph	[0,48.6]	/

表1

表2

图10

图11

图12

图13

表3

图14

图15

表4

表5

表6

图16

参考文献 22

1	Zheng Yu, Capra L, Wolfson O, et al. Urban Computing: Concepts, Methodologies, and Applications[J]. ACM Transactions on Intelligent Systems and Technology, 2014, 5(3): 38.
2	Song Xuan, Zhang Haoran, Akerkar Rajendra, et al. Big Data and Emergency Management: Concepts, Methodologies, and Applications[J]. IEEE Transactions on Big Data, 2022, 8(2): 397-419.
3	Li Shuzhe, Chen Wei, Yan Bingqi, et al. Self-supervised Contrastive Representation Learning for Large-scale Trajectories[J]. Future Generation Computer Systems, 2023, 148: 357-366.
4	Yao Huaxiu, Tang Xianfeng, Wei Hua, et al. Revisiting Spatial-temporal Similarity: A Deep Learning Framework for Traffic Prediction[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2019: 5668-5675.
5	Li Kuilin, Zhai Chunjie, Xu Jianmin. Short-term Traffic Flow Prediction Using a Methodology Based on ARIMA and RBF-ANN[C]//2017 Chinese Automation Congress (CAC). Piscataway: IEEE, 2017: 2804-2807.
6	Fu Rui, Zhang Zuo, Li Li. Using LSTM and GRU Neural Network Methods for Traffic Flow Prediction[C]//2016 31st Youth Academic Annual Conference of Chinese Association of Automation (YAC). Piscataway: IEEE, 2016: 324-328.
7	Liu Hongjie, Xu Hongzhe, Yan Yu, et al. Bus Arrival Time Prediction Based on LSTM and Spatial-Temporal Feature Vector[J]. IEEE Access, 2020, 8: 11917-11929.
8	Zhang Junbo, Zheng Yu, Qi Dekang, et al. DNN-based Prediction Model for Spatio-temporal Data[C]//Proceedings of the 24th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. New York: ACM, 2016: 92.
9	Zhang Junbo, Zheng Yu, Qi Dekang. Deep Spatio-temporal Residual Networks for Citywide Crowd Flows Prediction[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2017: 1655-1661.
10	Lin Ziqian, Feng Jie, Lu Ziyang, et al. DeepSTN+: Context-aware Spatial-temporal Neural Network for Crowd Flow Prediction in Metropolis[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2019: 1020-1027.
11	Wang Hongnian, Su Han. STAR: A Concise Deep Learning Framework for Citywide Human Mobility Prediction[C]//2019 20th IEEE International Conference on Mobile Data Management (MDM). Piscataway: IEEE, 2019: 304-309.
12	Dai Genan, Kong Weiyang, Liu Yubao, et al. Multi-perspective Convolutional Neural Networks for Citywide Crowd Flow Prediction[J]. Applied Intelligence, 2023, 53(8): 8994-9008.
13	Ding Fei, Zhu Yue, Yin Qi, et al. MS-ResCnet: A Combined Spatiotemporal Modeling and Multi-scale Fusion Network for Taxi Demand Prediction[J]. Computers and Electrical Engineering, 2023, 105: 108558.
14	Yao Huaxiu, Wu Fei, Ke Jintao, et al. Deep Multi-view Spatial-temporal Network for Taxi Demand Prediction[C]//Proceedings of the Thirty-second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence. Palo Alto: AAAI Press, 2018: 2588-2595.
15	Guo Shengnan, Lin Youfang, Li Shijie, et al. Deep Spatial-Temporal 3D Convolutional Neural Networks for Traffic Data Forecasting[J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 20(10): 3913-3926.
16	Chen Cen, Li Kenli, Teo Sin G, et al. Citywide Traffic Flow Prediction Based on Multiple Gated Spatio-temporal Convolutional Neural Networks[J]. ACM Transactions on Knowledge Discovery from Data, 2020, 14(4): 42.
17	Zhou Qianqian, Chen Nan, Lin Siwei. FASTNN: A Deep Learning Approach for Traffic Flow Prediction Considering Spatiotemporal Features[J]. Sensors, 2022, 22(18): 6921.
18	陈静, 张昭冲, 王琳凯, 等. 基于卷积长短时记忆网络的短时公交客流量预测[J]. 系统仿真学报, 2024, 36(2): 476-486.
	Chen Jing, Zhang Zhaochong, Wang Linkai, et al. Short-term Bus Passenger Flow Prediction Based on Convolutional Long-short-term Memory Network[J]. Journal of System Simulation, 2024, 36(2): 476-486.
19	Hu Jie, Shen Li, Sun Gang. Squeeze-and-excitation Networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 7132-7141.
20	Woo Sanghyun, Park Jongchan, Lee J Y, et al. CBAM: Convolutional Block Attention Module[C]//Computer Vision-ECCV 2018. Cham: Springer International Publishing, 2018: 3-19.
21	Hou Qibin, Zhou Daquan, Feng Jiashi. Coordinate Attention for Efficient Mobile Network Design[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 13708-13717.
22	Humaira H, Rasyidah R. Determining the Appropiate Cluster Number Using Elbow Method for K-means Algorithm[C]//Proceedings of the 2nd Workshop on Multidisciplinary and Applications (WMA) 2018. Nitra Slovakia: EAI, 2018: 2292388.

参数列表	具体参数
操作系统	Windows10
处理器及	i9-12900K 3.20 GHz
运行内存	64 G
显卡	NVIDIA GeForce RTX 3090
Cuda版本	12.2
编程语言	Python3.9

模型	TaxiBJ		BikeNYC
模型	RMSE	MAE	RMSE	MAE
HA	45.36	22.47	20.33	10.22
ARIMA	22.78	12.72	10.22	5.36
CNN+LSTM	17.29	9.87	5.09	2.61
GCN+LSTM	16.79	9.84	5.28	2.73
ST-ResNet	16.69	9.85	5.18	2.61
DeepSTN+	16.58	9.93	4.52	2.34
ST-3DNet	16.04	9.38	4.94	2.43
STAR	15.85	9.27	4.69	2.38
MS-ResCnet	15.74	9.33	4.61	2.36
ST-MVAR	13.78	8.72	4.38	2.20

模型	参数量	RMSE	MAE
ST-MVAR-w/o-CA	0.410×10⁶	14.06	8.92
ST-MVAR-SE	0.412×10⁶	13.91	8.82
ST-MVAR-CBAM	0.413×10⁶	13.93	8.82
ST-MVAR-w/o-TP	0.429×10⁶	14.58	9.03
ST-MVAR-w/o-VE	0.429×10⁶	13.90	8.83
ST-MVAR	0.429×10⁶	13.78	8.72

ST-ResNet	DeepSTN+	ST-3DNet	STAR	ST-ResCnet
ResUnit = 4	ResPlus unit = 1	3D layer = 3	ResUnit = 6	ResUnit = 6
Channels = 64	Channel = 64	Kernel size = (3,3,3)	Channels = 64	Channels = 64
Kernel size = (3,3)	Resplus channel = 8	ResUnit = 4	Kernel size = (3,3)	Feature ratio = 5:5
	Pooling rate = 2	ResUnit kernel size = (3,3)		Kernel size = (3,3)
	Kernel size = (3,3)

模型	参数量	单个epoch 运行时间/s	收敛 epoch	占用显存/GB
ST-ResNet	0.717×10⁶	15.27	65	5.33
DeepSTN+	130.8×10⁶	30.56	34	12.39
ST-3DNet	1.02×10⁶	11.72	175	6.44
STAR	0.47×10⁶	8.65	79	3.66
MS-ResCnet	0.51×10⁶	9.74	73	4.90
ST-MVAR	0.429×10⁶	8.52	85	3.57