A Graph Convolutional Network-Based Model for Traffic Flow Prediction Using Multimodal Spatial and Temporal Data

Abstract

In this paper, with the continuous expansion of cities in China, ring expressways have been adopted in most cities, playing a positive role in diverting urban transit traffic, relieving urban congestion, and improving traffic efficiency. However, few scholars have conducted comprehensive and systematic studies on ring expressways. Traffic flow characteristics, situation assessment and prediction, and operation management measures are still in research. Due to the disadvantages of the traditional models using single data sources and their limitation to describe and forecast specific roads, we demonstrate an algorithm-optimized model (TK-GCN) based on graph convolutional networks (T-GCN) for capturing multimodal spatial and temporal datasets and Kalman filtering for correcting the evolution of phase points to describe the phenomena of the Chengdu ring expressway (National Expressway G4202) in China from May to October 2019, according to dynamic data observed from our intelligent transportation system. Experiments on traffic datasets show good performance of our deep architecture. Abundant experiments show that our approach achieved improvements over the state of the art. It is also presented that our model can improve the generalization performance of shared tasks. These positive results demonstrate that our model is promising in transportation research.