Short-term Prediction of Local Traffic Conditions Reflecting the Relationship between Traffic Variables by Deep Learning

Abstract

Deep learning has been attracting attention as a prediction method of traffic condition. Deep learning, which can automatically extract the relationships inherent in the data, has shown high performance for many types of prediction problems, and its usefulness for prediction of traffic condition has also been confirmed. However, most previous prediction methods using deep learning do not consider the relationships between multiple traffic variables and only consider location-based predictions. However, when considering traffic control for mitigating traffic congestion as an example of application of the prediction results, the prediction of traffic condition for a region, which is the unit of control implementation, are highly useful, rather than a point-based one. In this study, the LSTM, a deep learning model, was used to examine short-term predictions aimed at representing the relationship between multiple macroscopic traffic condition indices aggregated on a regional basis, and its performance was confirmed through application to the observed data of urban roads.