Traffic Flow Forecasting with Dynamic Graph Neural Networks and Incident-Aware Attention

Abstract

Traffic flow forecasting constitutes a pivotal component of Intelligent Transportation Systems (ITS), enabling proactive congestion management and optimized urban planning. Traditional approaches typically model traffic networks as static graphs, relying on fixed adjacency matrices determined by Euclidean distances or physical connectivity. However, such static representations fail to capture the dynamic spatial-temporal dependencies that evolve rapidly, particularly under non-recurrent events such as traffic accidents, road closures, or adverse weather conditions. This paper introduces a novel framework, the Dynamic Graph Neural Network with Incident-Aware Attention (DGNN-IA), designed to address these limitations. The proposed model integrates a dynamic graph learning module that infers time-varying network topologies from data, coupled with a specialized attention mechanism that explicitly encodes incident information to modulate internode influence weights. By fusing traffic state tensors with incident embedding vectors, the model dynamically adjusts the information propagation path, allowing for robust prediction even in the presence of abrupt network perturbations. Extensive experiments on real-world traffic datasets augmented with incident logs demonstrate that DGNN-IA achieves state-of-the-art performance, significantly outperforming baseline models in both short-term and long-term forecasting horizons.