Research On Next Generation Intersection Control Systems In Stages Of CAV Penetration

Intersections as one of the most important parts of a traffic system,the impact of traffic issues such as congestion and crash risk at which are amplified as traffic conditions at intersections are complex.Recently,the emergence of connected and autonomous vehicles(CAVs)has provided a way to resolve the above issues at intersections.However,the 100%penetration rate of CAVs on road will not occur in the near future,as it will go through several stages.Hence,for the purpose of safety enhancement and operational efficiency,how to develop intersection control and management methods according to the traffic characteristics of different stages has become a challenge in academia.In order to resolve the aforementioned challenge,the thesis develops a general framework for intersection control and management,with the goal of arriving at an optimal trade-off among traffic efficiency,fairness and energy saving.In the framework,reinforcement learning technique is employed to control key components of the intersection(i.e.,agents),such as vehicles and traffic lights.Taking advantages of the Deep Q-Network algorithm embedded with convolutional neural networks,the proposed framework is able to achieve adaptive decision in complex and dynamic systems such as intersections.Compared with other modelling methods,the proposed framework has advantages of better generalization capability,adaptability and expansibility.Based on the general framework,the thesis further proposes intersection control systems which are suitable for the three stages of CAV penetration,namely,the stage of traditional humandriven vehicles(HVs),the stage of mixed HVs and CAVs,and the stage of full CAV implementation.For the stage of traditional HVs,the thesis proposes an adaptive traffic signal control system on intersections.The system is able to adaptively control the traffic light in regards to phases and duration,based on real-time traffic conditions.Moreover,it realizes the coordinative control of traffic lights at multiple intersections,by adopting state and decision sharing among agents.For the stage of full CAV implementation,the thesis divides the stage into two substages according to the sequence of technological development,and proposes an intersection control system for each substage,namely,INTEL-PLT and COOR-PLT.Herein,INTEL-PLT adopts adaptive CAV platooning strategy,as it identifies the optimal platoon size based on real-time traffic conditions.While COOR-PLT further implements CAV coordination strategy on the basis of INTELPLT,as it employs multi-agent coordination technique to identify the optimal passing priorities of CAV platoons.For the stage of mixed HVs and CAVs,the thesis proposes an adaptive signal control and CAV coordination system on intersections,which optimizes operational efficiency and safety of both HVs and CAVs.The system can be seen as the combination and modification of the systems in above two stages.By adopting efficient passing rules,the system maximizes the flexibility of CAVs while preserving the driving habits of HVs in the meanwhile.In the simulation,all the proposed systems mentioned above manifests high convergence speed,robustness and decision flexibility in several typical traffic conditions.Moreover,the superior performances on both throughput and fuel conservation of each system are demonstrated,as compared with state-of-the-art methods.In summary,the thesis not only makes prospective theoretical contributions to intersection control and management in various stages of CAV penetration,but also is of practical significance to the application of reinforcement learning in transportation domain.