Technologies Of Regional Road Network Traffic Status Recognition And Prediction Based On Sparse Data

With the rapid expansion of car ownership,total travel volume and travel range in China,traffic congestion is becoming more and more serious,gradually evolving from local bottleneck congestion to spreading congestion within the regional road network,which has seriously affected the operational efficiency of the transportation system,generated social problems such as traffic safety and environmental pollution.Traffic flow parameter forecasting can analyse the variation pattern of traffic parameters with randomness and uncertainty,and provide technical support to alleviate traffic congestion.In recent years,floating vehicle data,which contains more continuous temporal and spatial traffic flow information than traditional fixed detector data,has gradually become an important data source in the field of intelligent transportation.However,the problem of sparse sample size due to the low coverage of floating vehicles severely restricts the application of floating vehicle data in the field of traffic,how to reconstruct complete traffic flow parameters based on sparse floating vehicle data is one of the key research problems in traffic flow parameter prediction.The traffic network and distribution in cities have clear heterogeneity problems,predict traffic flow parameter based on the overall road network data will reduce the accuracy of the prediction model,meanwhile,increase the computational complexity.How to aggregate road sections with similar operational characteristics to achieve accurate delineation of regional road network subareas is the second problem that needs to be focused on in traffic.The evolution of traffic flow states has omplex non-linear characteristics,for the insufficient prediction accuracy of traffic flow parameter prediction methods,there are some limitations of the role of traffic flow parameter prediction in alleviating traffic congestion.How to achieve high accuracy and robust traffic flow prediction is the third of the problems that need to be focused on.To address the above problems,this paper takes sparse floating vehicle data as the basis,selects the road average travel speed,which is an important traffic flow parameter,as the research object,focuses on the connecting area of highway and urban expressway,and carries out research on its traffic state estimation and prediction methods.The specific research contents include:(1)Collecting,processing and analyzing road network floating car data.First,in view of the manifestations of various abnormal data in the original floating car trajectory data,formulate corresponding abnormal data processing rules and abnormal data cleaning procedures.Then,in order to extract traffic flow parameter information from GPS trajectory data,a map mapping method based on Hidden Markov Model(HMM)is used to combine GPS floating car data with electronic map data obtained from Open Street Map(OSM).Then,analyze the minimum sample requirement of the floating car data based on the coverage strength and coverage rate of the road network,and the sparse characteristics of the floating car data are analyzed.(2)Proposing a model for estimation of road section traffic flow parameters based on low-rank tensor decomposition algorithm.First,based on the correlation analysis of multi-mode traffic flow and low-rank analysis,the travel speed tensor is constructed based on the four dimensions of road section,week,day,and time period.Then,in view of the obvious sparse problem of the tensor model,a low-rank tensor decomposition model based on truncated kernel norm is proposed to fill in the traffic flow parameter data at the missing position,thereby realizing the missing road section traffic flow parameters in the tensor model estimate.Then,based on the massive GPS trajectory data of floating vehicles acquired from Beijing’s road network,validation of the algorithm was carried out.The results show that the filling accuracy of the proposed method in this paper is more than 2% better than the methods based on matrix and vector forms.(3)Proposing a normalized cut algorithm based on density peak optimization to realize the division of road network sub-areas.First,calculate the similarity of traffic flows on different road sections and construct the weighted map of the road network.Then,based on the NCut algorithm,the road sections with similar traffic flow characteristics are gathered,and then a road network with large heterogeneity is divided into multiple sub-regions with strong homogeneity;the results of the NCut algorithm are sensitive to the initial clustering center.For the problem,a NCut algorithm based on density peak optimization is proposed.This method determines the initial cluster center and the number of clusters by certain rules.Then,based on the two traffic flow parameters of average travel speed and unit time delay,the Fuzzy C-means(FCM)algorithm optimized by density peak is used to realize traffic state recognition.The validation results of the algorithm show that the improved algorithm has a 15.27% improvement in traffic status classification recognition.(4)Proposing a combinatorial optimization LSTM prediction algorithm for the average speed,which combines the time series decomposition model and the attention mechanism.First,the algorithm model takes the time series of traffic flow parameters and the traffic state of the sub-regions as input to predict the road segments in the sub-region where the road is located.For the nonlinearity and non-stationarity of traffic data can restrict the prediction accuracy,a time series decomposition method is proposed to decompose time series data,so as to obtain several components with weaker volatility and strong regularity,which are input into the LSTM model to achieve average Prediction of travel speed.Then,aiming at the local periodic transfer problem of the evolution of the traffic flow state,the attention mechanism is introduced to improve the LSTM model.Later,carry out the experimental road section verification.The validation results in multiple scenarios show that the proposed prediction method based on a combination of attention mechanism and time series decomposition to optimise the LSTM improves the prediction accuracy by more than 3% compared with other baseline methods,and has strong robustness when unfrequently traffic congestion occurs.