Intelligent transportation systems (ITS) have developed rapidly for urban road networks in recent years. Accurate and efficient short-term traffic flow speed prediction is the key to the realization of ITS. Traffic flow data usually perform stochastic and nonlinear characteristics, making short-term forecasting of large-scale urban road networks challenging. To extract the spatial and temporal correlations between traffic flows, we propose a novel short-term speed forecasting of large-scale urban road network based on the deep learning algorithm Transformer used in the field of natural language processing. We test the model using a real floating car dataset collected on a large-scale urban road network. It is found that the Transformer model shows high prediction accuracy and efficiency performance and outperforms the benchmark models.

CICTP