A spatial-temporal framework including traffic diffusion for short-term traffic prediction

Abstract

With the increasing popularity of Intelligent Transportation Systems, how to achieve accurate and real-time traffic prediction has become more and more important. In this paper, we intend to improve the accuracy of traffic prediction by appropriate integration of diffusion process. The spatial-temporal features of traffic flow are captured within an encoder-decoder framework. Specifically, (1) a 1-dimension Convolutional Network (CNN) is exploited to capture the spatial features when fed by the congestion matrix; (2) two long short-term memory methods (LSTMs) are applied to mine the temporal closeness and period properties; (3) external factors such as traffic diffusion, time characteristics are also considered to enhance prediction performance; (4) CNN, LSTMs and external factors are integrated into the final CNN-LSTM based encoder-decoder framework. Experiment results on a public dataset indicate that the consideration of traffic diffusion has advantage in short-term traffic prediction.