A Spatiotemporal Prediction Model for Urban Road Freight Flow Based on Residual Graph Convolution and Attention Mechanism

Abstract

To improve the foresight of urban logistics scheduling, this study proposes a freight flow prediction model that combines residual graph convolutional networks (ResGCN) with a multi-head spatiotemporal attention mechanism. The model constructs a traffic graph using the structure of the road network and integrates factors such as freight orders, road conditions, and holidays, aiming to capture non-Euclidean correlations between nodes and multi-dimensional temporal variations. Experimental results on a real-world freight platform dataset show that the proposed model achieves improvements of 14.5% and 11.8% in MAE and RMSE, respectively, compared with traditional LSTM and TCN methods.