Distributed Coordination Control Of Traffic Network Flow And Parallel Optimization Based On Cloud Computing

The rapid development of artificial intelligence,cloud computing,communication technology and Internet of Things,has pushed the transportation systems to enter a new stage of big data-driven intelligent development.It is required to explore the traffic network flow intelligent control and decision-making methods under the background of big data,so as to alleviate traffic congestion and improve travelling satisfaction.To achieve the requirements of control and real-time performances for the analysis processing and optimization decision based on big data,this dissertation studies the theory and method of distributed coordination control of traffic network flow.The responsive control method and parallel solution strategy based on cloud computing are studied using macroscopic traffic flow evaluation model.Further,the predictive control method and parallel solution strategy based on cloud computing are studied using the microscopic traffic flow evaluation model.In order to improve the intelligent level of traffic network flow control,this dissertation continues to study the parallel learning method of spatiotemporal characteristics of traffic network flow based on deep learning models,and to extend this parallel method to the distributed deep reinforcement learning control based on the deep learning evaluation model.By establishing the edge computing solution of traffic network flow control,the cloud-edge collaborative processing strategy of offline big data learning and online decision-making applications is studied.The effectiveness of the control methods and parallel optimization algorithms is verified by simulation experiments based on the Beijing regional road network.The main research contents and innovations of this dissertation are summarized as follows:(1)For the sake of improving the control and real-time performances,the distributed coordination control approach and parallel optimization algorithm based on macroscopic traffic flow evaluation model are proposed.The distributed control of the regional road network is realized by optimizing the green ratio parameters of each signalized intersection on the basis of the macroscopic traffic flow delay model.The coordinated control of regional road network is achieved,which promotes the formation of green wave control situation,by optimizing the offset between adjacent intersections.A coarsegrained parallel adaptive genetic algorithm with the migration ratio and substitution probability is utilized to solve the dual-objective optimization problem of distributed coordination control,and further to accelerate the solution speed through cloud computing.Simulation results show that the proposed method is effective in reducing traffic delay and meeting the real-time requirements of big data processing.(2)In order to make up the shortcomings of the macroscopic evaluation method that may not match random microscopic traffic flow fluctuations and of the lack of predictability of responsive control,the distributed coordination predictive control approach and parallel optimization algorithm of traffic network flow based on microscopic traffic flow evaluation model are developed.A rule-based non-analytical microscopic traffic flow model is applied to traffic network flow predictive control,which can predict future traffic situations and evaluate candidate control schemes more accurately than the existing macroscopic prediction model.In order to reduce the solution time of optimal control sequences in the prediction horizon,a two-level hierarchical parallel genetic algorithm based on Spark cloud computing is utilized to accelerate the solution speed in the rolling horizon.Simulation results illustrate that the proposed approach achieves preferable control effectiveness and accelerates the solution speed under unsaturated and oversaturated traffic flow conditions,respectively.(3)In order to improve the intelligent learning ability of traffic flow prediction model,the learning mechanism of spatial-temporal characteristics of traffic network flow based on deep learning models and the parallel training method are proposed.Since the current learning objects of traffic flow features are mostly limited to local road sections,this dissertation adopts a hybrid deep learning model based on deep convolutional neural networks and long short-term memory neural networks,as well as establishes a learning model of traffic network flow features towards big data processing.It not only mines the spatial correlation characteristics among multiple road sections,but also extracts the dynamic evolution law of traffic flow time series.For the sake of decreasing the training time of deep learning under big data,this dissertation studies the theoretical foundation of parallel training with convergence guarantee based on dataset decomposition,and designs the parallel algorithm based on Spark cloud computing.Simulation results demonstrate that the proposed deep learning model and the parallel training method not only enhance the accuracy of feature learning,but also greatly reduce the training time.The parallel learning of spatial-temporal evolution characteristics of traffic network flow based on deep learning models is the research foundation of distributed deep reinforcement learning control of traffic network flow.(4)For the sake of improving the intelligent level of control decisions,this dissertation proposes the distributed deep reinforcement learning control of traffic network flow based on the deep learning evaluation model and the implementation algorithm based on edge computing.The value decomposition method is extended to the actor-critic algorithm framework,and the continuous control of traffic network signals is solved by introducing the green ratio adjustment method considering multiple constraints in the output layer of the action network.Through the adaptive allocation mechanism of policy contribution weights,the policy contribution weights with great influences on the global objective are continuously strengthened to achieve the adaptive distributed collaborative decision making.Finally,the distributed deep reinforcement learning method is deployed to the edge computing architecture to realize the collaborative processing of online decisions and offline learning.Simulation results verify the effectiveness of the intelligent control method and the cloud edge collaborative solution algorithm.