EADGN: an event-adaptive dynamic graph network for robust traffic flow forecasting under anomalous conditions

Abstract

Traffic flow forecasting across diverse scenarios is critical yet challenging. Existing methods often exhibit substantial performance degradation under anomalous conditions, such as traffic accidents, because of severe spatiotemporal disruptions. To address this limitation, we propose the Event-Adaptive Dynamic Graph Network (EADGN), a framework for accurate forecasting in both normal and anomalous scenarios. EADGN mitigates the effects of anomalous events through a dynamic graph structure learning module that adaptively updates the adjacency matrix to model their propagation and influence. It further incorporates a hierarchical Efficient GAT-based spatial modeling mechanism to capture both local and long-range dependencies. Extensive experiments on two large-scale real-world datasets, NE-BJ and PeMS-BAY, demonstrate that EADGN outperforms state-of-the-art methods, improving 60-minute prediction accuracy by 2.7%–6.4%. In anomalous scenarios, EADGN reduces prediction errors by 10.3%, confirming its robustness and adaptability to complex and dynamic traffic conditions.