A Risk‐Aware Multimodal Framework for Long‐Horizon Traffic Flow Forecasting Under External Disturbances

Long‐horizon traffic flow forecasting under external disturbances is challenging due to the combined effects of road‐network dependencies, temporal evolution, and heterogeneous environmental factors such as adverse weather, traffic incidents, and holidays. Existing spatiotemporal graph models mainly rely on historical numerical sequences, while direct concatenation of external features may cause scale mismatch and introduce disturbance‐induced noise. To address these limitations, this paper proposes RAMF, a risk‐aware multimodal traffic flow forecasting framework. RAMF adopts a graph‐language decoupled architecture, where a spatiotemporal graph learning branch captures normal traffic dynamics and road‐network dependencies, and a topological semantic prompting branch transforms weather, incident, and holiday information into structured prompts encoded by a lightweight pretrained language model. A physical‐risk‐aware gated fusion module is further designed to adaptively balance the two branches according to node‐level environmental risks, while gradient decoupling is used to mitigate the impact of localized anomalous disturbances on graph‐structure learning. Experiments are conducted on three real‐world datasets, PEMS03, METR‐LA, and PEMS‐BAY, for 4‐h traffic forecasting. RAMF achieves MAE/RMSE/MAPE values of 19.06/29.85/19.55%, 8.60/13.60/17.48%, and 2.30/4.43/5.61% on the three datasets, respectively. Compared with representative time‐series and spatiotemporal graph baselines, RAMF achieves competitive or lower forecasting errors, and stratified experiments show its improved robustness under severe weather, traffic incidents, holiday periods, and high‐risk scenarios.