Sun, Zhanbo (Southwest Jiaotong University), Liu, Zhuo (Southwest Jiaotong University), Zhang, Chao (The Intelligent Policing Key Laboratory & Sichuan Police College), Kong, Mingming (Xihua University), Ji, Ang (Southwest Jiaotong University)

Dynamic Spatio-Temporal Adaptive Convolutional Neural Networks for Traffic Flow Prediction

Scheduled for presentation during the Regular Session "Traffic prediction and estimation II" (WeBT6), Wednesday, September 25, 2024, 14:30−14:50, Salon 14

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Traffic flow prediction is crucial for enhancing transportation systems' efficiency and safety, providing references for intelligent traffic management and road resource allocation. However, capturing the intricate spatial and temporal relationships between various areas and time intervals poses a significant challenge. To address the sophisticated spatio-temporal correlations encountered in traffic flow forecasting, we develop an adaptive convolutional neural network based on time attention. For spatial considerations, we design deformable convolutional kernels integrated with an attention mechanism to extract road information and adapt it to the road network structure. Furthermore, we deploy the attention mechanism to dynamically aggregate multiple parallel convolutional kernels instead of the traditional single-layer convolution on each input, focusing precisely on key areas within the network. From a temporal perspective, a dynamic time attention mechanism is applied based on weekly, daily, and nearest temporal data to grasp critical moments in periodic traffic flow. By integrating essential temporal and spatial elements, the Dynamic Spatial Temporal Adaptive Convolutional Networks (DSTACN) are constructed. The proposed model is validated on three real-world datasets. The results demonstrate its forecasting accuracy surpassing other baseline models by up to 11.98%, highlighting the proposed approach's capability in handling the spatio-temporal dependencies inherent in traffic flow data.