A Multi‐Level Spatiotemporal and Abrupt‐change‐aware Fusion Model for Traffic Flow Forecasting

ABSTRACT

Traffic flow forecasting plays a critical role in urban traffic management. Recent graph neural network‐ and Transformer‐based methods have achieved strong performance in modelling complex spatiotemporal dependencies. Nevertheless, real‐world traffic flow data often contain local abrupt changes and nonstationary fluctuations, and the forecasting behaviour of existing models under such high‐variation segments remains less explicitly evaluated. To address this issue, this paper proposes a multilevel spatiotemporal and abrupt‐change‐aware fusion (ML‐STACAF) model. Built upon the Transformer architecture, ML‐STACAF consists of three key components. First, a dynamic multilevel embedding (DMLE) module enhances input representations by jointly fusing raw traffic flow series, periodic labels and adaptive vectors. Second, a spatiotemporal and local discontinuity‐sensing attention (ST‐LDSA) module employs three parallel attention branches to simultaneously model temporal dependencies, spatial dependencies and local discontinuity characteristics. Third, a trans‐dimensional interactive fusion module (TIFM) integrates the outputs of the parallel attention branches through cross‐attention and gating mechanisms. Extensive experiments conducted on multiple real‐world traffic datasets demonstrate that, compared with state‐of‐the‐art methods, ML‐STACAF achieves performance improvements of 2.1%, 2.6% and 1.63% in terms of MAE, RMSE and MAPE, respectively. These results indicate that ML‐STACAF provides consistent and competitive forecasting performance, while further evaluation under abrupt‐change scenarios is conducted to examine its effectiveness in high‐variation traffic segments.