Research on Spatiotemporal Prediction Model of Urban Traffic Flow Based on Graph Neural Networks and Attention Mechanism

Traffic flow pblueiction is one of the hardest (and most practical) problems in smart city work. If you get the forecasts wrong,everything downstream suffers: signal timing,incident response,routing,even basic city operations. The trouble is that a lot of existing methods still don`t do a great job with the messy reality of road networks,where conditions change fast and influences ripple across space and time in non-obvious ways.This paper proposes a model called the Dynamic Spatiotemporal Graph Attention Network (DSTGAT). In plain terms,it mixes graph neural networks with attention to better match how traffic actually behaves. The key idea is a dynamic traffic graph that can change as road conditions change,instead of treating the network as fixed. A graph convolutional network is then used to learn spatial relationships from that graph,while a spatiotemporal attention module focuses on the temporal side of the problem.Tests on three real-world datasets show consistent gains over common baselines. DSTGAT cuts MAE by 15.3% and RMSE by 12.7%,which is a meaningful jump if you care about day-to-day traffic control rather than just pretty benchmark numbers. Overall,the results suggest DSTGAT is a solid step toward more reliable,real-world traffic pblueiction for urban management.