基于路段-路口联合建模的对偶图卷积网络的行程时间估计方法
金广垠(1996—),男,安徽芜湖人,博士生,主要从事时空数据挖掘研究。E-mail: jinguangyin96@foxmail.com |
收稿日期: 2022-10-10
修回日期: 2023-01-05
网络出版日期: 2023-06-30
基金资助
国家自然科学基金项目(62073333)
国家自然科学基金项目(72201029)
国家自然科学基金项目(72288101)
Travel Time Estimation Method based on Dual Graph Convolutional Networks via Joint Modeling of Road Segments and Intersections
Received date: 2022-10-10
Revised date: 2023-01-05
Online published: 2023-06-30
Supported by
National Natural Science Foundation of China(62073333)
National Natural Science Foundation of China(72201029)
National Natural Science Foundation of China(72288101)
估计给定路径的行程时间在许多城市交通系统中起着重要作用,例如导航、路线规划和拼车等。然而,现有的大多数工作都侧重于对路段或交叉路口进行单独建模,这并不能准确估计行驶时间,因为交叉路口和路段作为路径的基本要素不仅各自包含多样化的空间属性和时间动态,而且它们之间还具备较强的耦合相关性。为了解决上述问题,本论文提出了一种新颖的端到端深度学习框架,即面向行程时间估计的对偶图卷积网络(DGCN-TTE)来对交叉路口和路段进行联合建模。具体来说,这个模型采用对偶图卷积方法来捕获路口和路段的复杂关系,其中构建节点图来刻画路口之间的相关性,构建边图来表征路段之间的交互特征。为了捕捉空间和时间特征的联合关系,模型中还引入了一种在捕捉时间依赖性的同时结合了从多个邻域范围内整合多尺度空间关系的时空学习模块。本论文通过对3个真实世界的轨迹数据集上的充分的实验来评估提出的DGCN-TTE模型,结果表明该模型显著优于现有的方法,评估指标相比于次优方法最多可以获得超过10%的提升。
金广垠 , 沙恒宇 , 张金雷 , 黄金才 . 基于路段-路口联合建模的对偶图卷积网络的行程时间估计方法[J]. 地球信息科学学报, 2023 , 25(7) : 1500 -1513 . DOI: 10.12082/dqxxkx.2023.220768
Estimating travel time for a given route plays an important role in many urban transportation systems, such as navigation and route planning. However, most of the existing works focus on modeling road segments or intersections individually, which cannot accurately estimate travel time, because intersections and road segments, as the basic elements of paths, not only contain diverse spatial attributes and temporal dynamics, but also have strong coupling correlation between them. To address the above problems, this paper proposes a novel end-to-end deep learning framework, Dual Graph Convolutional Network for Travel Time Estimation (DGCN-TTE), to jointly model intersections and road segments. Specifically, we adopt a dual graph convolution method to capture the complex relationship between intersections and road segments, in which a node-wise graph is constructed to characterize the correlation between intersections, and an edge-wise graph is constructed to characterize the interaction features between road segments. To capture the joint relationship of spatial and temporal features, we also introduce a spatiotemporal learning module that integrates multi-scale spatial relationships from multiple neighborhood scales while capturing temporal dependencies. This paper evaluates the proposed DGCN-TTE model through experiments on three real-world trajectory datasets, the results show that the proposed model significantly outperforms existing methods, and the evaluation metrics can achieve more than 10% improvement compared with the suboptimal method.
表1 实验数据集的基本信息Tab. 1 Basic information of the experimental dataset |
数据集 | 成都 | 波尔图 | 北京 |
---|---|---|---|
轨迹数量/条 | 15 303 | 12 683 | 346 96 |
路段数量/条 | 873 | 544 | 714 |
路口数量/个 | 807 | 450 | 320 |
平均行程时间/s | 246.54 | 248.49 | 343.3 |
平均移动距离/m | 1435.34 | 1394.82 | 3929.7 |
表2 成都数据集在郊区、高峰和非高峰3个场景下的数值统计Tab. 2 Numerical statistics of Chengdu dataset in suburban, peak and off-peak scenarios |
场景 | 郊区 | 高峰时间 | 非高峰时间 |
---|---|---|---|
轨迹数量/条 | 8 156 | 2 867 | 12 436 |
路段数量/条 | 639 | 873 | 873 |
路口数量/个 | 291 | 807 | 807 |
平均行程时间/s | 242.42 | 249.81 | 243.57 |
平均移动距离/m | 1 105.42 | 1 435.47 | 1 435.31 |
图7 模型在不同参数组合下的性能变化Fig. 7 The performance of the model under different parameter combinations |
表3 DGCN-TTE及其变体在2个数据集上估计路径行程时间的性能Tab. 3 Performance of DGCN-TTE under different combinations of k and n |
参数组合 | 成都数据集 | 波尔图数据集 | |||||
---|---|---|---|---|---|---|---|
RMSE | MAE | MAPE | RMSE | MAE | MAPE | ||
(1,1) | 154.11 | 88.34 | 0.297 5 | 55.52 | 42.03 | 0.166 3 | |
(1,2) | 141.60 | 78.39 | 0.265 3 | 54.12 | 41.03 | 0.160 8 | |
(1,3) | 140.30 | 77.60 | 0.258 5 | 50.95 | 40.06 | 0.155 8 | |
(1,4) | 143.78 | 80.55 | 0.268 9 | 53.76 | 40.83 | 0.159 6 | |
(2,1) | 151.67 | 86.23 | 0.286 5 | 54.84 | 41.62 | 0.163 1 | |
(2,2) | 140.52 | 77.78 | 0.259 6 | 51.17 | 38.83 | 0.154 4 | |
(2.3) | 141.22 | 78.92 | 0.261 2 | 52.59 | 40.20 | 0.157 9 | |
(2,4) | 146.75 | 83.32 | 0.275 6 | 54.12 | 41.24 | 0.160 8 |
注:表中加粗数字表示评估指标的最佳性能。 |
表4 DGCN-TTE 和其他方法在3个数据集上估计路径行程时间的性能Tab. 4 Performance of DGCN-TTE and other methods for estimating travel time on three datasets |
算法 | 成都数据集 | 波尔图数据集 | 北京数据集 | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
RMSE | MAE | MAPE | RMSE | MAE | MAPE | RMSE | MAE | MAPE | |||
AVG | 202.43 | 132.86 | 0.640 4 | 111.67 | 88.75 | 0.4509 | 338.43 | 264.55 | 0.736 7 | ||
TEMP | 167.41 | 108.50 | 0.553 8 | 98.01 | 78.12 | 0.4127 | 342.08 | 272.46 | 1.642 5 | ||
GBDT | 165.03 | 101.68 | 0.542 8 | 91.24 | 71.79 | 0.3914 | 372.87 | 262.49 | 0.685 4 | ||
MlpTTE | 150.67 | 87.32 | 0.307 6 | 59.46 | 44.49 | 0.1783 | 124.92 | 79.75 | 0.253 2 | ||
RnnTTE | 155.16 | 89.37 | 0.301 1 | 56.45 | 43.32 | 0.1702 | 127.80 | 86.59 | 0.292 4 | ||
DeepTTE | 148.30* | 86.40 | 0.298 4 | 57.32 | 43.54 | 0.1686 | 147.60 | 92.64 | 0.250 0 | ||
T-GCN | 153.49 | 87.71 | 0.291 8 | 56.35 | 43.28 | 0.1678 | 129.56 | 85.66 | 0.275 8 | ||
DRCNN | 151.59 | 86.14* | 0.288 4 | 54.73* | 41.81* | 0.1673* | 116.30 | 76.71 | 0.256 1 | ||
ConSTGAT | 152.23 | 87.10 | 0.284 5* | 56.67 | 43.22 | 0.1684 | 107.58* | 69.90* | 0.228 8* | ||
GCNAttTTE | 152.95 | 87.41 | 0.289 8 | 55.88 | 42.79 | 0.1654 | 112.88 | 73.82 | 0.237 2 | ||
DGCN-TTE | 140.30 | 77.60 | 0.258 5 | 50.95 | 40.06 | 0.1558 | 96.27 | 63.02 | 0.214 5 | ||
提升比例/% | 5.39 | 9.91 | 9.14 | 6.90 | 4.19 | 4.83 | 10.51 | 9.84 | 6.25 |
注:表中加粗数字表示评估指标的最佳性能。 |
表5 DGCN-TTE 和其他方法在3个数据集上估计路径行程时间的性能Tab. 5 Performance of DGCN-TTE and other graph learning-based methods for estimating road segment travel time on three datasets |
算法 | 成都数据集 | 波尔图数据集 | 北京数据集 | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
RMSE | MAE | MAPE | RMSE | MAE | MAPE | RMSE | MAE | MAPE | |||
T-GCN | 57.84 | 24.85 | 0.553 7 | 22.48 | 13.03 | 0.552 1 | 47.06/- | 23.87/- | 0.588 6/- | ||
DRCNN | 57.82 | 24.76 | 0.551 3 | 20.59 | 11.68 | 0.460 5 | 49.14/- | 26.23/- | 0.630 6/- | ||
ConSTGAT | 58.17 | 24.55 | 0.550 3 | 21.76 | 12.23 | 0.486 5 | 47.26/- | 21.64/- | 0.507 4/- | ||
GCNAttTTE | 58.26 | 24.79 | 0.551 2 | 19.89 | 10.95 | 0.432 3 | 45.31/22.88 | 21.50/10.12 | 0.488 4/0.466 5 | ||
DGCN-TTE | 55.91 | 22.88 | 0.512 2 | 15.81 | 8.06 | 0.301 6 | 41.34/20.60 | 18.87/8.04 | 0.441 4/0.370 7 |
注:表中加粗数字表示评估指标的最佳性能。 |
表6 DGCN-TTE和其他方法在3种场景下估计路径行程时间的性能Tab. 6 Performance of DGCN-TTE and other methods for estimating travel time in three scenarios |
算法 | 郊区 | 高峰时段 | 非高峰时段 | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
RMSE | MAE | MAPE | RMSE | MAE | MAPE | RMSE | MAE | MAPE | |||
AVG | 159.14 | 110.85 | 0.517 7 | 177.83 | 137.24 | 0.688 7 | 159.25 | 122.93 | 0.630 7 | ||
TEMP | 148.26 | 101.43 | 0.458 9 | 156.97 | 112.18 | 0.512 0 | 150.12 | 115.24 | 0.493 6 | ||
GBDT | 136.19 | 83.77 | 0.373 9 | 142.58 | 91.48 | 0.397 5 | 138.75 | 82.15 | 0.356 8 | ||
MlpTTE | 125.49 | 78.48 | 0.293 8 | 131.78 | 87.64 | 0.311 5 | 127.11 | 79.67 | 0.311 9 | ||
RnnTTE | 121.75 | 76.64 | 0.281 7 | 126.43 | 84.87 | 0.296 8 | 122.50 | 83.69 | 0.302 0 | ||
DeepTTE | 114.47 | 72.55 | 0.287 9 | 129.45 | 87.67 | 0.312 1 | 123.36 | 83.97 | 0.296 8 | ||
T-GCN | 121.58 | 76.38 | 0.277 5 | 132.05 | 89.74 | 0.303 9 | 121.21 | 83.17 | 0.290 5 | ||
DRCNN | 120.34 | 75.40 | 0.274 8 | 131.22 | 89.10 | 0.297 6 | 120.76 | 82.78 | 0.285 3 | ||
ConSTGAT | 118.89 | 72.12 | 0.264 8 | 128.79 | 88.06 | 0.293 0 | 118.64 | 81.89 | 0.282 1 | ||
GCNAttTTE | 111.87 | 68.63 | 0.242 1 | 117.13 | 78.45 | 0.272 9 | 105.81 | 71.05 | 0.257 9 | ||
DGCN-TTE | 159.14 | 110.85 | 0.517 7 | 177.83 | 137.24 | 0.688 7 | 159.25 | 122.93 | 0.630 7 |
注:表中加粗数字表示评估指标的最佳性能。 |
表7 DGCN-TTE及其变体在2个数据集上估计路径行程时间的性能Tab. 7 Performance of DGCN-TTE and its variants in estimating travel time on two datasets |
变体模型 | 成都数据集 | 波尔图数据集 | |||||
---|---|---|---|---|---|---|---|
RMSE | MAE | MAPE | RMSE | MAE | MAPE | ||
DGCN-TTE w/o R-GCN | 146.78 | 82.03 | 0.273 2 | 51.64 | 40.82 | 0.158 8 | |
DGCN-TTE w/o RNN | 147.48 | 83.46 | 0.276 4 | 54.07 | 41.13 | 0.162 6 | |
DGCN-TTE w/o Intersection | 143.82 | 79.59 | 0.264 2 | 53.96 | 41.02 | 0.161 0 | |
DGCN-TTE Simple GCN | 144.56 | 80.21 | 0.270 9 | 53.87 | 40.95 | 0.162 2 | |
DGCN-TTE | 140.30 | 77.60 | 0.258 5 | 50.95 | 40.06 | 0.155 8 |
注:表中加粗数字表示评估指标的最佳性能。 |
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
/
〈 |
|
〉 |