Deep Learning Based Congestion Prediction Using PROBE Trajectory Data

Within transportation operations and research, the prediction of traffic congestion for a large-scale road network is always a challenge but is very useful. Different from traditional model driven approaches, this paper demonstrated an innovative data-driven approach that can effectively predict network-wide traffic congestion in short and long-term time spans. Based on the sanitized probe trajectory data, this paper proposed a hybrid deep learning architecture that combined 3-dimensional convolutional networks (C3D) with convolutional neuron networks (CNNs) and recurrent neuron networks (RNNs), which is called CRC3D. The prediction result of the CRC3D is further compared with a variety of recurrent neural network architectures. It is illustrated that the proposed model was successful in inheriting the advantages of C3D and CNN-RNN; and it could well reflect the trend and regularity of the traffic state with high accuracy, which can be used for large-scale transport network congestion prediction competitively.