Flow decomposition based spatial–temporal filtering self-attention networks for traffic flow forecasting

Abstract

Traffic prediction is essential for intelligent transportation systems but challenging due to complex spatial–temporal dependencies. Current methods often overlook data entanglement caused by stable flow sequences and traffic events, and attention mechanisms may introduce irrelevant spatial information. In this paper, we propose FDFSAN (Flow Decomposition based spatial–temporal Filtering Self-Attention Networks), which addresses these issues by decomposing traffic flow data into stationary and sudden components modeled through a dual-channel spatial–temporal network. Our filtering self-attention mechanism captures both temporal and spatial dependencies, integrating information from nearby and distant roads while minimizing irrelevant spatial noise. Extensive experiments on four real-world datasets show that FDFSAN outperforms state-of-the-art methods, making it suitable for urban traffic networks with dynamic spatial correlations and anomalies.