A multi-timescale dynamic graph attention network (MTDGAT) for short-term traffic prediction under special events

Abstract

Accurate traffic prediction is a key task in Intelligent Transportation Systems (ITS) and is crucial for proactive traffic control and management. Traffic prediction under special events (SEs) has long been challenging due to the uneven spatio-temporal distribution of traffic state and high traffic fluctuations. The present work proposes a Multi-Timescale Dynamic Graph Attention Network(MTDGAT) for short-term traffic prediction under SEs. Specifically, we design a multi-timescale block that employs attention mechanisms to model the relationships between historical traffic states and future traffic states across different time scales in a fine-grained manner, aiming to accurately capture the complex traffic evolution patterns under SEs. Our model adopts an encoder–decoder architecture, wherein the encoder combines historical data and SEs Encoding to construct a historical multi-timescale spatio-temporal graph, which is then transformed into a future multi-timescale spatio-temporal graph through a transform module. Extensive experiments were conducted on real-world traffic datasets to further validate the performance of the proposed model. First, ablation experiment results demonstrate the superiority of the proposed MTDGAT model in capturing short-term evolution of traffic states under SEs. Furthermore, through comprehensive comparison experiments, it is indicated that the MTDGAT outperforms other baseline models across all prediction steps under SEs. Specifically, MTDGAT achieves MAE of 3.21 and MAPE of 14.46 for 1-step prediction, alongside MAE of 4.15 and MAPE of 19.9 in 4-step prediction. The outcomes of the present work could provide deeper insights into spatio-temporal traffic evolution under the influence of SEs and contribute to the formulation of proactive traffic management and control strategies in such scenarios.