MIFNetwork: Multi-scale Information Fusion Network for Long Traffic Prediction

Traffic prediction is a critical task in the development of intelligent transportation systems (ITS) and smart cities. Current methods mainly rely on Spatial-Temporal Graph Neural Networks (STGNNs), which effectively capture spatial dependencies but require significant computational resources and limited to short-term forecasting. However, research on time-series data in the transportation field, particularly on sequence prediction across multiple time scales, remains underexplored. To leverage the multi-scale characteristics of time-series data, as well as its seasonal and periodic features, we introduce two solutions based on DLinear model: multi-scale decomposition and cross-scale information fusion, and propose a traffic prediction model named MIFNetwork. We conduct comparative experiments on several traffic datasets, selecting transformer-based models with strong time-series modeling capabilities for comparison. The results demonstrate that our model outperforms other approaches, especially in handling long time spans and complex trend changes.