Research On Traffic Flow Prediction And Incident Warning Method For Urban Road Network

The rapid growth of the numbers of drivers, vehicles and traffic volume has brought huge convenience for people, but also brings the traffic security risk that can not be ignored. In order to make sure the safety of resident trip, it is necessary to analyze the risk factors (such as human, vehicle, road and environment factors), the internal mechanism between the risk factors and the traffic incident, and construct efficient incident warning model. A considerable studies have shown that traffic flow factor is one of the key factors that affect the incident risk. Therefore, studying the influence of dynamic characteristics of traffic flow on traffic incident risk is one of the fundamental approaches to analyzing the traffic safety situation with high accurate rate and improving the safety level of the whole road traffic system. In addition, studying the evolution pattern of traffic flow and improving the performance of traffic flow prediction, such as real-time, reliability and self-adaptability, has become the research focuses and it also provides important data guarantee for developing traffic incident warning model. This paper mainly focuses on solving those key problems as follows: (1)how to study the spatial dependency characteristics of traffic flow for urban road network; (2) how to forecast traffic flow in short term based on the spatial dependency characteristics of traffic flow; and (3) how to utilize the traffic flow variables to study the traffic incident warning model. The issues mentioned above are studied deeply and verified by field data and numerical examples. The main contributions of the dissertation can be summarized in four aspects:(1) Based on the complex network theory, a method for modeling and analyzing urban road traffic network which reflects the spatial correlation of traffic flow is proposed.In order to interpret well the complexity of traffic flow distribution on the road network in a large spatial scope, the global road traffic system is represented by a network with interacting segments. The distinguishing feature of modeling method is that the connections among different segments are not dependent on the spatial proximity but rather on the spatial correlation calculated by the traffic temporal series. A geographical weight based PageRank algorithm (GWPA) is proposed to evaluate the importance of segments in the network. The evaluation results are the prerequisites for exploring the spatial correlation patterns of traffic flow for urban road traffic network.(2) The spatial deployment characteristics of traffic flow correlation are explored under a community detection perspective, the GWPA-k-means algorithm is proposed.The GWPA-k-means algorithm is put forward to identify the spatial correlation patterns of traffic flow for urban road network. The proposed method extends the traditional k-means algorithm in two steps: one is predefining the initial cluster centers by two properties of nodes (the GWPA value and the shortest path length) and the other one is utilizing the weight signal propagation process to transfer the topological information of the road traffic network into a node similarity matrix. The experimental results prove that the proposed modeling and analysis methods can reveal well the traffic flow spatial dependency patterns in a large spatial scale road network.(3) A method combining the community detection algorithm and long short-term memory neural network (CD-LSTM NN) is proposed to forecast traffic flow in short-term.CD-LSTM NN works on the basis of the spatial correlation patterns of traffic flow.The GWPA-k-means algorithm divides the whole road network into several sub regional networks, where segment roads are strongly correlated in the same region. For each sub regional nework, the traffic spatial-temporal series are transferred into a series of two-dimensional matrix, and the LSTM NN is applied to predict the traffic flow.Moreover, a self-adaptive orthogonal genetic algorithm is put forward to select the optimal parameter set for the CD-LSTM NN model. Experimental results show that considering the spatial correlation patterns of traffic flow can improve the accuracy of prediction result, and the proposed method can also work with good adaptability and strong robustness in case of sensor failure.(4) A traffic incident warning method based on traffic flow variables is put forward.On the basis of mastering the traffic states in real time, the most influencing traffic flow variables to traffic incident occurrence are selected. The research of traffic incident warning method is carried out from two aspects, i.e., traffic incident identification and traffic incident risk prediction: the relationship between traffic flow variables and traffic accident risk is explained from the point of view of domain division, and the concept of traffic safety region is given; a hybrid intelligent algorithm, combining sequential forward selection and principal components analysis, is proposed to select the most influencing traffic flow variables; least squares support vector machines is applied to estimate traffic safety region and classify the traffic safety states; moreover, according the similarity between the traffic safety region theory and reliability analysis theory, a traffic reliability model is presented. The proposed model not only comprehensive evaluating the traffic incident risk in a macro level, but also predict the urban road traffic incident risk in an individual level.