Traffic Flow Prediction Based on Deep Neural Networks

Abstract

Traffic forecasting plays an important role in intelligent transportation system. Accurate forecasting enables appropriate travel suggestions for commuters, and can further benefit transportation management. Recent research has led to significant progress in spatial-temporal feature extraction with the help of deep learning methods, such as graph convolutional neural network (GCN) and long short-term memory (LSTM) network. However, simply applying these methods to the specific situation of road network is not enough, hence in this paper, we propose an optimized spatial-temporal traffic flow prediction model. With line-graph transformation introduced into the construction of road traffic topology, the application of GCN has a more efficient and effective ability to learn spatial features in the road network compared to conventional convolutional neural network (CNN) operations; and the utilization of LSTM with weather and periodic features introduced as additional information makes the model more powerful in temporal feature extraction. Experimental results on real-world large-scale dataset verify the proposed models ability to capture spatial and temporal properties, showing the advantages of our model beyond several baseline models.