A noise-immune and attention-based multi-modal framework for short-term traffic flow forecasting

Accurately forecasting short-term traffic flow is essential for intelligent transportation systems. However, current methods often struggle to fully exploit implicit variation patterns and heterogeneous correlations in traffic flow data, and can be sensitive to non-Gaussian noise. In this paper, we propose a novel noise-immune and attention-based multi-modal model (NIAMNet) for short-term traffic flow forecasting. Inspired by the success of computer vision techniques, NIAMNet transforms one-dimensional traffic flow into images and embeds residual dual-attention blocks (RDB) to extract in-deep features. Besides, we introduce a dynamic noise-immune loss to address the impact of noise and outliers on model performance. Experimental results on four real-world benchmark datasets demonstrate the superiority of NIAMNet over existing methods, achieving the lowest MAPE (10.43, 9.79, 10.51, and 11.01) and RMSE (247.13, 192.36, 208.40, and 150.01). Additional ablation experiments are carried out to provide insight into the significance of each component. Our approach contributes to the development of more accurate and robust short-term traffic flow forecasting models.