Traffic Density Based Travel-Time Prediction With GCN-LSTM

In recent years, data-driven travel-time prediction methods have been actively developed owing to the widespread availability of various observation data such as probe vehicle data. In most existing studies on large-scale networks, the time-series data of speed, which can be directly estimated from probe vehicle data, are used as input. However, in a free-flow regime, the change in speed does not depend much on the number of vehicles. Therefore, it cannot accurately represent the traffic states of the regime. Using traffic density as input, which can describe traffic states from the free-flow regime to the congested-flow regime, we expect to learn the fluctuation pattern of traffic states more efficiently in the free-flow regime before the occurrence of congestion and contribute to the early detection of congestion. In this study, we propose a new methodology for travel-time prediction using a combined model of graph convolutional networks and long short-term memory with spatially interpolated density as input. Empirical validation using real observation data from the Hanshin region, Japan, shows that the density input is superior to the speed input in achieving the early detection of traffic congestion and improving the accuracy of travel-time prediction.