Traffic speed prediction using a deep neural network to accommodate citywide spatio-temporal correlations

摘要

The present invention relates to a method and apparatus for predicting a road network traffic speed using a deep neural network. The traffic speed of a given road section is affected by the current and past traffic speeds of the neighboring road sections, and the effect is further extended to the rest of the traffic network. Therefore, a successful predictive model must consider not only the effect of neighbors, but also the effect of distant road sections. Based on this principle, we propose an in-depth neural network structure for accommodating spatial correlations across cities as well as time dependence of the actual traffic network topology, and a model that extends the proposed prediction model in terms of traffic transition and propagation. . The present invention was performed using a large data set collected for more than 10 months and successfully predicted the traffic speed of 170 road sections in Gangnam, Seoul.