Spatial-Temporal Feature Extraction and Evaluation Network for Citywide Traffic Condition Prediction

Abstract

Traffic condition prediction is crucial for executing traffic control and scheduling tasks within intelligent transportation systems. With diversified data sources, effectively modeling the complex spatial-temporal dependencies in the whole traffic network and predicting nonlinear general traffic condition changes become primary challenges for intelligent transportation systems. In this paper, a double-layer spatial-temporal feature extraction and evaluation (DL-STFEE) model is proposed, aimed at accurately predicting the traffic condition transferring of the whole traffic network. Firstly, a public traffic dataset is processed to extract the spatial and temporal features of vehicles, as well as the clusters of traffic conditions of the entire traffic network. Secondly, a double-layer deep learning traffic condition predictor is proposed. The model incorporates a spatial-temporal feature extraction layer, which leverages the graph convolution network (GCN) and attention mechanism to obtain spatiotemporal features across the whole traffic network. Additionally, a spatial-temporal combination layer employs a high-dimensional self-attention mechanism to integrate features across spatial-temporal combinations, bolstering prediction accuracy. Finally, through rigorous experiments, the contributions of neural network structures on the spatial-temporal feature extraction are comprehensively analyzed and the experiments also validate the effectiveness of traffic condition prediction by the proposed DL-STFEE model.