A Traffic Flow Dependency and Dynamics based Deep Learning Aided Approach for Network-Wide Traffic Speed Propagation Prediction

The information of network-wide future traffic speed distribution and its propagation is beneficial to develop proactive traffic congestion management strategies. However, predicting network-wide traffic speed propagation is non-trivial. This study develops a traffic flow dependency and dynamics based deep learning aided approach (TD2-DL), which predict network-wide high resolution traffic speed propagation by explicitly integrating temporal-spatial flow dependency, traffic flow dynamics with deep learning method techniques. Specifically, we first develop a graph theory-based method to identify the local temporal-spatial traffic dependency of each road among neighboring roads adaptive to the prediction horizon and traffic delay. Then, traffic speed propagation on every road is mathematically described by v-CTM based on traffic initial and boundary conditions. Next, the long short-term memory (LSTM) model is employed to predict boundary conditions factoring the traffic temporal-spatial dependency and historical data predicted by v-CTM. In this way, we well couple the physical models (traffic dependency and v-CTM) with the deep learning approach, and further make them coevolution under this framework. Last, an EKF is used to assimilate predicted traffic speed predicted by v-CTM coupled with the LSTMs and the field traffic data; an FNN is introduced to impute missing and corrupted data for improving the traffic speed prediction accuracy. The numerical experiments indicated that the TD2-DL predicted the network-wide traffic speed propagation in 30 minutes with accuracy varying from 85%-98%. It outperformed the tested models recently developed in literature. The ablation experimental results confirmed the significance of factoring traffic dependency and integrating data imputation and assimilation techniques for improving the prediction accuracy.