An Attention-based Approach for Traffic Conditions Forecasting Considering Spatial-Temporal Features

Accurate and efficient traffic conditions forecasting is a significant challenge in Intelligent Transportation System (ITS) applications. The conditions can be characterized by traffic speed. To forecast the traffic conditions accurately, we propose a novel attention-based GCN-GRU hybrid model named AGG which can capture the spatial and temporal features of traffic speed simultaneously. In this model, the Graph Convolutional Network (GCN) captures the topological features for modeling spatial correlations. The Gated Recurrent Unit (GRU) captures the temporal features for modeling temporal correlations. The attention mechanism is used to assign weights to features according to the degree of importance of speed data, further improving model's forecasting precision. Then the experiments in real-world traffic speed data reveal that the AGG model can efficiently capture the spatial and temporal features of traffic speed and realize the precise forecasting of traffic conditions.