关键词: 交通工程； 交通流预测； 双尺度； 注意力机制； 时间序列聚类

Abstract: Traffic flow prediction, as a core technology of intelligent transportation systems (ITS), faces the fundamental challenge of effectively modeling complex spatio-temporal dependencies in traffic data. Current mainstream models (based on graph neural networks and attention mechanisms) have two key limitations. Firstly, node similarity computation is affected by temporal misalignment in traffic fluctuations, causing misjudgment of similar nodes with delayed propagation characteristics. Secondly, spatial feature extraction fails to jointly capture macro-level patterns (e.g., periodic travel patterns) and micro-level dynamics (e.g., sudden congestion, traffic accidents) in traffic flows. To address these issues, LDFormer model was proposed, which introduced dynamic time warping (DTW) algorithm to reconstruct node similarity measurement, eliminating spatio-temporal deviations caused by propagation delays. Meanwhile, a dual-path spatial modeling mechanism was designed. Macro-micro characteristics of the attention-generated spatial dependencies were respectively captured by learnable mask matrices (Mglo and Mmic). Experiments on three benchmark datasets demonstrate that LDFormer model significantly outperforms existing spatio-temporal prediction methods.

Key words: traffic engineering; traffic flow prediction; dual-scale; attention mechanism; time series clustering