Probabilistic Graphical Model To Support Short-term Traffic Flow Prediction And Reasoning

With the modernization process of the army, more and more motorized drives have been adopted in the army's travel, especially when handling various emergencies. If we can forecast the future condition of the transportation, and choose a smooth path for travel, we will be able to deal with the emergencies more quickly and effectively. Modern traffic conditions provide the convenience for us in the forecast, but only using the traffic information with real-time acquisition technology is difficult to completely cover the transportation network of a certain regional, and can't reflect the future condition of the traffic network. The traffic flow is important information to reflect the traffic conditions by describing the number of vehicles of a particular road crossing. As the real-time traffic information can not reflect the future traffic changes, we should process, model, analyze, predict the real-time traffic flow data, and point out a rational, convenient and unobstructed path for people. This is exactly the short-term traffic flow prediction in Intelligent Transportation System, and it is also one of the critical problems of informative and intelligent management technology in traffic transportation paradigm.Aiming at the real-time and efficiency requirements in short-tern tfraffic prediction, in this thesis, we adopted Baysian network (BN), an important probabilistic graphical model (PGM) as the framework of representing and inferring uncertain knowledge in traffic flow data, We considered the time-series dependency among traffic junctions between adjacent time periods. We explored the algorithm for BN construction and inferences for the cases concerning any two adjacent periods, which is the basis for short-term traffic flow prediction.The main work and contribution of this thesis can be summarized as follows:■In order to construct the BN that reflects the dependency among traffic road crossings, in this thesis, we proposed the construction algorithm for TBN structure with time-series conditional dependence. As is known that Directed Acyclic Graph (DAG) construction is the critical and difficult problem when constructing a BN, we gave our DAG construction algorithm based on graph traversal taking as inputs the given starting and destination. TBN is the basis for traffic flow prediction.■On the basis of the DAG, in this thesis we adopted the likelihood estimation method to deal with the historical traffic flow data, and then calculated the Conditional Probability Table (CPT) of the nodes in TBN. Thius, the TBN was ultimately constructed to reflect the time-series dependency among road crossings in traffic network.■In order to realize real-time and efficient traffic flow prediction, we proposed an approximate probabilistic inference algorithm for TBN based the idea of Gibbs sampling. This can be used to predict the traffic flow data in the next period at the condition of given the current traffic flow data.■By using the real observed traffic flow data, we implemented and tested the performance of the methods for TBN construction, inferences and the corresponding traffic flow prediction. Experimental results show that the methods we proposed in this thesis are efficient, accurate, and practical.