DST: A Deep Urban Traffic Flow Prediction Framework Based on Spatial-Temporal Features

Abstract

Traffic flow prediction is an interesting and challenging problem in transportation modeling and management. The complex topological structure of urban road network makes it more complicated. The performance of traditional traffic flow prediction models like time series models is not satisfactory, for those methods cannot describe the complicated nonlinearity and uncertainty of the traffic flow precisely. With the rapid development of deep learning, many researchers try to apply deep learning methods to traffic flow prediction. However, those deep learning models neither consider both spatial relation and temporal relation, nor do they combine spatial relation and temporal relation in an effective way. In this paper, we propose a deep urban traffic flow prediction framework (DST) based on spatial-temporal features. In our framework, we use a local convolutional neural network (CNN) method which only considers spatially nearby regions to extract the spatial features and a long short-term memory (LSTM) model to extract the temporal features. In addtion to the traffic flow data, we also use external context data when predicting traffic flow. The experiments on a large-scale taxi trajectory dataset TaxiCQ show that our proposed model significantly outperforms other comparison models.