Spatial-temporal Traffic Congestion Prediction Based on Attention Mechanism

Abstract

Accurate traffic flow prediction is of great significance for intelligent transportation system in intelligent city. For the problem of traffic congestion, accurate prediction of traffic congestion is conducive to rational urban planning and efficient energy utilization. There are some problems in data-driven traffic flow congestion prediction, such as inaccurate prediction caused by complex spatiotemporal correlation characteristics. Facing this problem, this paper proposes a spatiotemporal attention combination network (STACN) based on attention mechanism for traffic congestion prediction. Firstly, this paper uses the conventional attention mechanism to capture the multi-dimensional time series correlation of the target road. Secondly, this paper uses the graph attention mechanism to capture the spatial dependence of all neighborhoods of the target road in the graph. In this paper, the prediction accuracy and consistency of congestion level are evaluated by using the actual traffic data set, and the experimental results verify the effectiveness of the model.