Spatial-temporal Multi-level Feature Fusion Traffic Flow Prediction Model Based on U-Net Architecture

Traffic flow prediction is important for intelligent transportation systems.Current models extract spatial-temporal features well but static adjacency graphs do not reflect the traffic's dynamics.In this study, we present a novel framework for the deep learning-based traffic flow prediction model, named Spatial-temporal Multi-level Feature Fusion to Traffic Flow Prediction Model (U-STMLFF) in which U-Net architecture is utilized.It employs multi-level pooling for feature extraction and builds dynamic attention adjacency matrices for temporal spatiotemporal learning on graphs with Graph CNNs incorporating a multi-temporal pattern capture module.Attension-based prediction fusion method on a progressive merge scheme improves the prediction accuracy.