Baozhen Yao's research while affiliated with Dalian University of Technology and other places

In an urban environment, the accurate prediction of congestion levels is a prerequisite for formulating traffic demand management strategies reasonably. Current traffic forecasting studies mostly focus on the road topological network and assume that the spatial linkages of road segments is fixed, thus ignoring associated congestion level changes between road segments. This study introduces a fusion-based graph convolutional network called the F-GCN. The model aims to capture the dynamic correlations of the congestion levels among road segments while constructing the static topology of the road network. In particular, the entropy-based grey relation analysis method is first implemented in the dynamic graph convolutional network (GCN) module to measure the potential correlations among the observed congestion levels. Then, spatio-temporal blocks that integrate GCN layers, spatial attention mechanisms, and long short-term memory layers are built for both the static and dynamic GCN modules. Finally, the F-GCN model is developed by fusing the static GCN and dynamic GCN module. The numerical experiments on Beijing taxies demonstrated that the proposed F-GCN model achieved a significant 5.47%, 5.64%, and 8.14% improvements for the 15-, 30-, and 45-min prediction tasks in the weighted mean absolute percentage error compared to the state-of-the-art baseline models. This improvement highlights the effectiveness of the model in predicting congestion levels, demonstrating its superiority and potential in capturing the dynamic potential correlations among the congestion levels of road segments.

The rebalancing of idle vehicles is critical to mitigating the supply–demand imbalance in on-demand ride services. Motivated by a ride service platform, this paper investigates a short-term vehicle rebalancing problem under demand uncertainty in the presence of contextual data. We deploy a novel data-driven robust optimization approach that takes a direct path from “Data” to “Decision” instead of the predict-then-optimize paradigm and leverages the prediction problem structure to seamlessly integrate demand predictions with optimization models. We further develop a risk-based uncertainty set to evaluate how well uncertain demand is estimated from contextual data by prediction models, and discuss the classes of prediction models that are highly compatible with robust optimization models. Based on the convex analysis and duality theory, we reformulate the original models into equivalent Mixed Integer Second Order Cone Programmings (MISOCPs) that are solvable via state-of-the-art commercial solvers. To solve large-scale instances, we utilize the affine decision rule technique to derive polynomial-sized reformulations. Extensive experiments are conducted on the instances based on a real-world on-demand ride service in Chengdu. The computational experiments demonstrate the promising performance of our rebalancing strategies and solution approaches.

Autonomous vehicles (AVs) have been made possible by advances in sensing and computing technologies. However, the high cost of AVs makes privatization take longer. Therefore, companies with autonomous vehicles can develop shared autonomous vehicle (SAV) projects. AVs with a high level of automation require high upgrade and use costs. In order to meet the needs of more customers and reduce the investment cost of the company, SAVs with different levels of automation may coexist for a long time. Faced with multiple travel modes (autonomous cars with different levels of automation, private cars, and buses), travelers' travel mode choices are worth studying. To further differentiate the types of travelers, this paper defines high-income travelers and low-income travelers. The difference between these two types of travelers is whether they have a private car. The differences in time value and willingness to pay of the two types of travelers are considered. Based on the above considerations, this paper establishes a multi-modal selection model with the goal of maximizing the total utility of all travelers and uses the imperial competition algorithm to solve it. The results show that low-income travelers are more likely to choose buses and autonomous vehicles with lower levels of automation, while high-income travelers tend to choose higher levels of automation due to their high value of travel time.

As mobile charging service has the advantages of flexible charging and simple operation, it is selected by more and more users of electric vehicles. However, due to the differences in road network density and traffic flow distribution, the uneven distribution of charging demand occurs in different regions. It reduces the service efficiency of mobile charging vehicles during the peak charging demand period, thus affecting the final revenue of operators. In order to solve this problem, this paper proposes a dynamic pricing strategy considering the spatiotemporal distribution of charging demand to induce users to transfer between different regions, which can alleviate the phenomenon that users wait too long during peak demand. In order to realize the city-level operation of mobile charging service, we divide the region into hexagons and make statistics on the charging demand in each region. The established demand updating model can reflect the impact of charging price on users’ charging behavior. Finally, we simulate the generation of charging demand in Haidian District, Beijing. According to the demand of each area, a thermodynamic diagram of charging demand is generated.KeywordsRegional divisionMobile chargingDynamic pricing

Accurately predicting the trajectory of a vehicle is a critical capability for autonomous vehicles (AVs). While human drivers can infer the future trajectory of other vehicles in the next few seconds based on information such as experience and traffic rules, most of the widely used Advance Driving Assistance Systems (ADAS) need to provide better trajectory prediction. They are usually only of limited use in emergencies such as sudden braking. In this paper, we propose a trajectory prediction network structure based on LSTM neural networks, which can accurately predict the future trajectory of a vehicle based on its historical trajectory. Unlike previous studies focusing only on trajectory prediction for highways without intersections, our network uses vehicle trajectory data from aerial photographs of intersections taken by Unmanned Aerial Vehicle (UAV). The speed of vehicles at this location fluctuates more frequently, so predicting the trajectory of vehicles at intersections is of great importance for autonomous driving.KeywordsTrajectory predictionDeep learningIntersections

How to accurately predict Short-term traffic travel time is an important problem in Intelligent Transportation Systems. However, the traffic data usually exhibit high nonlinearities and complex patterns. Predicting traffic travel time is a challenge. Most previous studies use the topological adjacency of road networks to explore the spatial correlations. However, as a real network, the road network contains higher-order connectivity patterns, which have different statistical significance. The topology adjacency cannot reflect these higher-order connectivity patterns. To obtain topological adjacency and higher-order connection pattern information, a novel deep learning framework was proposed: Multiple Motifs Graph Convolutional Recurrent Neural Networks, for traffic travel time prediction in this paper. The accuracy of travel time prediction can be improved by the proposed model. To be more specific, there are two meaning blocks in each unit of the model: (1) The spatial blocks captured spatial patterns information by the Multi-Motif graph convolution network and Motif Graph embedding; (2) The temporal blocks captured temporal patterns information by the combination of LSTM and the FC layer. To prove the effectiveness and accuracy of the prediction model, experiments were conducted on real world traffic travel time datasets.