This is an implementation of DSTCN:[Exploiting Dynamic Spatio-Temporal Correlations For Origin-Destination Demand Prediction]

We design a novel Dynamic Spatio-Temporal Correlation Network (DSTCN) to address the OD demand prediction problem. Our model comprises three modules. The first module fully considers the OD demand tendency correlations to capture the OD demand changes in terms of the OD and DO (Destination-Origin) orientations; the second module grabs the dynamic spatio-temporal characteristics of each area; and the third module integrates temporal knowledge to make next-step predictions. DSTCN can capture the complex spatial-temporal correlations between the areas in the OD demand prediction.

Framework of DSTCN, which comprises three modules: the first module, (Glstm2D) fully considers the OD demand tendency correlations; the second module (Simformer) captures the dynamic spatial features of each area; and the third module (F-GRU) integrates the dynamic demand correlations extracted from Glstm2D and Simformer, and makes next-step predictions via a well-designed multi-head temporal convolution layer.

Experimental Results of New York Taxi Dataset in Manhattan in 2018 and 2019