Long-term traffic flow prediction model based on SUTDGCN

Daily long-time traffic flow prediction is a crucial urban computing problem that aids in rational planning of traffic routes and efficient allocation of traffic resources. Existing models, while improving the accuracy of long-time traffic flow prediction, also increase computational complexity and additional space overhead. To address this, an optimization method named SUTDGCN is proposed, leveraging spatial upsampling and s graphs to enhance the real-time performance of long-time traffic flow prediction. In the spatial domain, spatial upsampling involves augmenting the original road network with K virtual nodes for upsampling, thereby constructing an upsampled road network to adequately capture both local and global spatial correlations. In the temporal domain, a simple directed graph is first constructed to represent local and global dependencies of time slices and then gated graph convolution is employed to further learn the underlying temporal dependencies. Experimental results on real datasets PeMS04 and PeMS08 demonstrate that, the SUTDGCN outperforms the STUGCN,it reduces MAE by 4.5% and 5.8 %, RMSE by 2.2% and 2.3%, and MAPE by 2.8% and 6.8%, respectively. Meanwhile, the SUTDGCN model reduces the time required for ssstraffic flow prediction by half compared to the STUGCN model.