Deep Learning Models for Traffic Prediction in Urban Transport Networks.

Abstract

With the development of smart technologies built on big data analysis, the vision of smart city, that aims to improve quality of life, reduce energy consumption and pollution and drive economic growth, has been enhanced. One important component of this vision is the urban transport network which is continuously challenged with increasing number of vehicles on roads. This results in issues including long travel time, increasingly persistent traffic congestion, accidents and road safety concerns. The research work in this thesis aims to predict traffic in urban transport networks which are capable of solving those issues. Understanding and control of traffic patterns are fundamental towards this aim. For this, we develop a short-term traffic flow prediction model, named EM, on linear roadways based on machine learning technology. EM is able to analyse and extract spatial and temporal features from original traffic data for the final prediction. This short-term traffic prediction could give drivers traffic situation in advance and guide them to avoid congested roads to reduce travel time and also relieve traffic congestion. Moreover, armed with accurate short-term traffic prediction models on linear roadways, we further develop a novel deep learning model, named ALLSCP, for short-term traffic flow prediction on intersections where traffic situations are more complex compared to linear roadways. The difference to EM is that ALLSCP can analyse more detailed features (i.e. global- and local-spatial and short-, medium- and long-term temporal features) for the final prediction. Due to the fact that traffic flows are almost controlled by traffic lights in most cities around the world, this could benefit the optimisation of traffic light strategies so that more vehicles are allowed to pass in a short duration. After analysing traffic data on linear roadways and intersections, we consider and solve traffic prediction problem on large-scale road networks and develop a deep learning model (named SAGCN-SST) based on Graph Convolution Network (GCN) and Attention mechanism for multiple traffic speed prediction on large-scale road networks. This SAGCN- SST is able to capture dynamic-spatial and temporal features for the final prediction. The predicted traffic speed on large-scale road networks could be used to find vulnerable roads and then optimise routing strategies for reducing traffic congestion and long travel time. Furthermore, we design another deep learning model based on Virtual Dynamic Graph Convolution Network and Transformer with Gate and Attention mechanisms (VDGC- NTGA) for multi-step traffic speed prediction on large-scale road networks, which can explore dynamic and hidden spatial and temporal features in network-wide for the final prediction. VDGCNTGA has ability to generate and update a virtual dynamic road graph each batch to represent dynamic and hidden spatial dependencies that are not described in the real road graph. Compared to our last SAGCN-SST model, VDGCNTGA can exploit more useful hidden spatial dependencies of road segments in network-wide. Finally, we develop a deep learning model based on Sequence-to-Sequence architecture with an embedded module using Graph Convolution Neural Network and Transformer, named GCNT-Seq2Seq, for long-term traffic speed prediction in large-scale road net- works. This model is able to analyse and extract local- and global-spatial and long-term temporal features for the final prediction, which benefits the optimisation of traffic strategies for improving road network efficiency. For this research, we use real transportation data collected from urban transport networks of different characteristics to evaluate our proposed models and also compare them against well-known existing traffic prediction models.