Traffic Volume Prediction Based on Multi-Sources GPS Trajectory Data by Temporal Convolutional Network

Abstract

Predicting urban traffic volume is of great significance to traffic management and urban construction. An accurate prediction model can help drivers optimize driving routes, allocate resources reasonably and reduce urban traffic congestion. Most of the existing studies do not consider the complex nonlinear spatio-temporal relationship. In the spatial dimension, they do not consider the impact of regional semantics and regional interactions. In the temporal dimension, they ignore the impact of long-term historical information and key time points. Aiming at the complexity of traffic data, in this paper, we design a ResNet-TCN model to predict the urban traffic volume. Firstly, we construct and extract features from the vehicle GPS tracking and external information, such as velocity, time, location and weather. Then, we obtain regional semantic information by the ResNet model and combine the weights of the regional division with the average vehicle velocity into a two-channel matrix. We extract the key features of the matrix sequence and predict the velocity by the TCN model. Finally, we estimate the traffic volume through a traffic volume inference model in the traffic field. We conduct a large number of experiments on the actual dataset of Chengdu and compare our model with the existing models. The experimental results show that our method has better performance on prediction accuracy.