A dynamic spatial-temporal attention ensemble model for traffic flow prediction

Abstract

With the advent of the era of big data, a large number of sensors are scattered on traffic roads, generating traffic flow data, posing huge challenges to traffic flow prediction. Therefore, capturing the spatial features of traffic flow data is greatly significant to improve the accuracy of traffic flow predicted models. Analysis of spatial and temporal characteristics of traffic flow data, this paper proposes a Spatial-Temporal Attention (MA-STLGAT) model to predict the traffic speed in the traffic network. The Graph Attention Network (GAT) assigns different weights to neighbor nodes to capture the dynamic topology structure. Learning the temporal correlation of traffic flow data through Long ShortTerm Memory network (LSTM), and Multi-Head Attention is introduced to learn the contribution of nodes, improve the convergence speed and save computing costs. The experimental results on real data set PeMSD7 show that the predict performance of MA-STLGAT is better than other baseline models.