A short-term traffic flow prediction model for road networks using inverse isochrones to determine dynamic spatiotemporal correlation ranges

Spatio-temporal mining neural networks have proven to be effective methods for predicting traffic flow in road networks. Existing research has designed numerous network structures but has often overlooked the impact of spatiotemporal correlation ranges on prediction results. To determine a reasonable spatiotemporal correlation range, we constructed a Inverse Isochrone (ISOv) model that considers the dynamic diffusion time and direction of traffic flow. The dynamic spatio-temporal correlation range defined by this model allows for the selection of highly relevant spatio-temporal features. We also designed the Dynamic Temporal Graph Convolutional Network (ISOv-DTGCN) method, which incorporates a graph pooling layer to adapt to the dynamically changing spatiotemporal correlation range and extract spatiotemporal correlations. Experimental results on a real dataset from the Wuhan road network show that the complete ISOv-DTGCN model improves prediction accuracy by approximately 15% in terms of RMSE compared to existing baseline models.