Congestion-evolution aware road traffic prediction

Traffic congestion is a global concern due to continuous increase in traffic demand despite finite road capacity. Smart traffic management systems, supported with accurate Road Traffic Prediction (RTP), are essential for alleviating these concerns by maximizing roadway utilization. This research is concerned with the development of novel techniques for realizing accurate and efficient RTP in urban traffic by leveraging on congestion evolution patterns. Bus trajectories were recovered from traffic data provided by the Land Transportation Authority (LTA), Singapore in order to analyze the factors that majorly impact traffic conditions. The important intrinsic (i.e., spatial and temporal traffic patterns, periodicity of congestion evolution) and extrinsic (i.e., weather, social event, incident) features that impact the bus speed, and their significance in specific situations were then identified to propose an efficient Bus Travel Speed Prediction (BTSP) model. The proposed model utilizes attribute-driven attention network to dynamically assign significance to the heterogeneous feature components for traffic prediction. The proposed model was evaluated using real-world datasets. Compared to state-of-the-art methods, BTSP achieves error rate reduction of up to 57.07%, 36.48%, 48.26%, for MAE, MAPE, RMSE respectively. Next, new measurement techniques were introduced to model the spatial correlations among traffic data to show that the correlation patterns vary significantly under different traffic conditions. A Heterogeneous Spatial Correlation (HSC) model was proposed to capture these spatial multi-fold correlations. The multi-fold correlation model was further extended to model multiple temporal correlations using Long/Short Term Memory (LSTM) networks. The proposed Multi-fold Correlation Attention Network (MCAN) framework relies on the HSC model and LSTM networks to achieve accurate RTP. Our detailed evaluations confirm that MCAN outperforms all baselines including our BTSP model. In particular, it leads to an error rate reduction of up to 69.17%, 71.03%, 61.71%, for MAE, MAPE, RMSE respectively. A novel network embedding framework (AE-LPGT) is also proposed to characterize and predict congestion evolution among different road segments. AE-LPGT learns the representation of each road segment, which incorporates various realistic properties of congestion evolution, such as asymmetric transitivity, local proximity, and global propagation tendency of congestion evolution. A probability model has been proposed to determine the likelihood of congestion propagation/decay between any pair of road segments to overcome the limitation of existing works that make predictions by relying only on propagation patterns of historical traffic data. Experiments on real-world datasets show that AE-LPGT outperforms all the baselines with increased accuracy of up to 78.83%. A Deep Meta-Learning Model (DMLM) has been proposed to further improve the accuracy of AE-LPGT by considering time-varying influences of both dynamic and static impact factors. DMLM relies on dynamic attributed networks (DAN) to model congestion evolution by representing dynamic and static impact factors as node attributes, while preserving dynamic topological structures using dynamic links. DMLM relies on matrix factorization techniques and meta-LSTM modules to exploit temporal correlations at multiple scales, and employs meta-Attention modules to merge heterogeneous features, while learning time-varying impacts of both dynamic and static features. It has been shown that DMLM achieves significantly better prediction performance on congestion evolution behaviors than all state-of-the-art methods. Two methods for integrating congestion evolution behaviors into the road traffic prediction model were investigated to determine the most effective approach. The first method relies on a diffusion Graph Convolutional Network (GCN) to directly incorporate congestion propagation tendencies into road prediction model. The second method trains the evolution prediction model and traffic prediction model separately in advance, and dynamically combines them based on the congestion states. Noting that the second method is notably superior in terms of accuracy, it was incorporated into a unified framework together with other contributions made in this thesis to achieve congestion evolution-aware road traffic prediction system. Finally, the effectiveness of the proposed framework has been evaluated using large-scale Singapore traffic data.Doctor of Philosoph