| 摘 要: 针对现有交通预测方法在捕捉交通数据的动态性、时间延迟传播特性以及复杂的时空依赖关系方面仍存在不足的问题,提出了一种基于时空动态增强自注意力(SDES)的交通预测方法。该方法通过结合局部增强注意力模块(LAAM)和时空动态特征增强模块(SDFEM),有效解决了交通数据中的动态性建模、时间延迟传播和聚焦重要信息的挑战。LAAM 通过通道注意力机制自动聚焦交通数据关键时空位置,SDFEM 通过平滑动态时间规整(Soft-DTW)算法捕捉交通数据的时间延迟传播特性。在3个真实交通数据集上进行了实验,SDES相较于次优方法在PEMS08数据集的长期预测中将平均绝对误差 MAE、均方根误差 RMSE和平均绝对百分比误差 MAPE分别降低了2.24%、2.22%和2.78%,达到了最优水平。 |
| 关键词: 深度学习 交通预测 自注意力机制 时空动态增强 智能交通系统 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 江苏省大学生创新创业训练计划项目(202410293178Y);江苏省研究生科研与实践创新计划项目 (KYCX24_1125,SJCX24_0279) |
|
| Traffic Forecasting Based on Spatio-temporal Dynamically Enhanced Self-attention |
|
YAO Hongbo, ZHANG Xuena, ZHENG Hanyu, LIU Zihan, CAI Yunfeng
|
(College of Science, Nanjing University of Posts and Telecommunications, Nanjing 210023, China)
shiro_2003@163.com; zxn2026@163.com; 3108239901@qq.com; 3121441956@qq.com; cyfslf@163.com
|
| Abstract: Existing traffic forecasting approaches still exhibit limitations in effectively capturing the dynamic nature of traffic data, temporal delay propagation characteristics, and complex spatio-temporal dependencies. To address these challenges, a Spatio-temporal Dynamic Enhanced Sel-f attention ( SDES) method is proposed for traffic forecasting. The method effectively addresses three key challenges: dynamic modeling of traffic data, temporal delay propagation, and focus on critical information through the integration of a Local Augmented Attention Module (LAAM) and a Spatio-temporal Dynamic Feature Enhancement Module ( SDFEM). The LAAM automatically concentrates on crucial spatio-temporal locations through channel attention mechanisms, while the SDFEM captures temporal delay propagation characte-ristics using Sof-t Dynamic Time Warping (Sof-t DTW) algorithm. Comprehensive
experiments conducted on three rea-l world traffic datasets demonstrate that SDES achieves state-o-f the-art performance, particularly showing significant improvements in long-term forecasting on the PEMS08 dataset where the Mean Absolute Error(MAE), Root Mean Square Error(RMSE), and Mean Absolute Percentage Error(MAPE) are reduced by 2.24% , 2.22% and 2.78% respectively compared with the suboptimal baseline method. |
| Keywords: deep learning traffic forecasting sel-f attention mechanism spatio-temporal dynamic enhancement intelligent transportation system |
