北京交通大学学报2025，Vol.49Issue(4)：84-93,10.DOI:10.11860/j.issn.1673-0291.20240138
基于时空UNet图卷积模型的交通流预测
Traffic flow prediction based on a spatiotemporal UNet-graph convolutional model
摘要
Abstract
To address the complex and dynamic spatiotemporal features in traffic flow prediction,this paper proposes an Attention-based Spatiotemporal UNet with Graph Convolutional Network(AST-UNet-GCN)model for long-term traffic flow forecasting.First,for temporal modeling,a U-Net-based feature pyramid architecture is introduced to capture multi-scale temporal features.The encoder-decoder structure enables effective extraction of features across different time scales.To enhance adaptability to sudden events,a short-term temporal feature extraction module based on attention mechanisms is designed.Furthermore,a global temporal feature extraction module utilizing deform-able attention mechanisms is constructed to capture long-term temporal dependencies.Then,a Squeeze-and-Excitation(SE)feature fusion module is developed to improve the expressiveness of the convolutional neural network,enabling dynamic weighting of features at different time scales and en-hancing the fusion of multi-scale temporal information.This approach effectively highlights key features while suppressing redundant information.Finally,for spatial modeling,a Graph Convolutional Network(GCN)is employed.By constructing the topological graph structure of the traffic network,the model captures spatial dependencies.A sigmoid-based feature fusion mechanism is further designed to explore the intricate dynamic relationships between spatial and temporal features,enabling comprehensive spatio-temporal modeling.Experimental results demonstrate that compared to other mainstream models,the AST-UNet-GCN model achieved reductions of 8.5%and 9.4%in MAE and RMSE metrics for short-term prediction,while reductions of 10.4%and 6.5%were observed for long-term prediction,respec-tively.These results demonstrate the model's strong performance in traffic flow forecasting,particularly in immediate prediction accuracy and the stability of long-term trend forecasting.关键词
交通流预测/时空特征/特征融合/UNet/注意力机制Key words
traffic flow prediction/spatiotemporal features/feature fusion/UNet/attention mechanism分类
交通工程引用本文复制引用
项新建,林贤鑫,陈田冬,刘力,宋雷鹏,袁天顺..基于时空UNet图卷积模型的交通流预测[J].北京交通大学学报,2025,49(4):84-93,10.基金项目
浙江省尖兵研发计划(2023C03015)Zhejiang Provincial Pioneering R&D Program(2023C03015) （2023C03015）
