Research And Application On Real-time Traffic Prediction In Urban Areas

The intelligent traffic management and control play the basic and important role in Intelligent Transportation Systems(ITS)and the research on it has attracted widespread attention from researchers all over the world.Proper traffic control and guidance are inseparable from real-time and accurate traffic flow prediction,and real-time and reliable traffic prediction is also the key technology to evolve the transportation system from "passive control" to "active control".Nevertheless,traffic system in urban areas is a complex dynamic system,and traffic data show different regularity in different periods.Therefore,it is necessary to predict traffic flow accurately in urban traffic application.The main work for short-term traffic prediction is to predict the current or future traffic flow in a certain time through exploring the relationship between historical and current or future traffic flow data and capturing traffic characteristics.However,the extraction and analysis of traffic characteristics cannot be done without the support of a complete and accurate dataset,and the obtained traffic data from detectors are imperfect.Currently,the detection range and precision of detectors cannot meet the demand of mass traffic data,and abnormal and missing data occurs from time to time.Therefore,in view of this phenomenon,this thesis focuses on the research on the prediction and application of short-term traffic flow in urban areas.On the one hand,aiming at the problem that traffic flow cannot be accurately predicted due to missing data of urban road network detector,considering the relationship between the missing data and available spatio-temporal data,an optimization method of missing data imputation for real-time traffic prediction was proposed with the goal of minimizing the imputation amount.In this thesis,the spatio-temporal correlation analysis between road segments in the road network is carried out by using the relevant road segment pairs and traffic flow correlation level,and the traffic flow propagation graph is generated.Secondly,the minimum number of imputing points searching is carried out to eliminate those missing data points that have few impacts on the prediction.Meanwhile,the traffic flow propagation probability is used to carry out partial imputing of missing points.Finally,the Probabilistic Principal Component Analysis(PPCA)is applied to impute the data points that cannot be inferred by using the spatio-temporal correlation tree,to obtain a complete predicted dataset.The modified Space-time Auto-regressive Integrated Moving Average(STARIMA)model is combined with the complete prediction dataset to predict the traffic flow.The prediction results show that the prediction time can be decreased by reducing the number of imputing points.At the same time,due to the selection of the most relevant road segments for traffic prediction,the proposed method can also improve the prediction accuracy,and finally achieve real-time and effective traffic prediction.On the other hand,in order to make up for the shortcomings of traditional data collection that cannot meet the growing demand for real-time traffic information of road network,and considering the road network traffic characteristics under urban scenarios,this thesis proposes a deployment method of new detectors based on spatio-temporal correlation prediction,aiming at the problem of road section observability,namely detector deployment location.Firstly,the observability of road segments is revised and a node-oriented deployment method is proposed.In other words,the proposed method does not need path enumeration,which is particularly suitable for complex road network to address observability problem.Secondly,based on the spatio-temporal correlation matrix and the idea of basis,the finite nodes(basis nodes)are identified to determine the minimum number of new detectors to be installed in the network to predict the traffic flow of the remaining road segments of the road network,so that the traffic flow of the entire road network can be all computed or inferred.Finally,taking the data of existing traditional detectors as samples,the proposed method is verified experimentally to show its advantages in coverage rate,observable degree,prediction error and other aspects under the condition of ensuring optimal deployment.