Link Travel Time Prediction Based On Didi GPS Data

In recent years,with the continuous progress of urbanization,the number of vehicle has increased dramatically and traffic congestion has become an urgent "urban disease".In today's increasingly congested traffic,accurately predicting travel time is of great importance to both traffic managers and individuals.A travel route is composed of several links.Link travel time prediction is the basis of travel time forecasting,and it is also the most widely used traffic information data.It has been a research hotspot.In the traditional link travel time prediction research,most of the applied data are data collected by the toroidal coil detector,floating car data or taxi data.These data have the disadvantages of difficulty in obtaining,small amount of data,and incomplete data information.It is difficult to fully describe the temporal dynamics of actual traffic.Moreover,the existing methods mostly use only the historical travel time of the target link for prediction,ignoring the spatial correlation of adjacent links.In view of the above shortcomings,this paper establishes a link travel time prediction model based on PCA-PSO-GRNN,fully considers the time-space characteristics of travel time,and conducts an in-depth study of link travel time prediction based on the data published by DIDI online taxi-hailing service platform.First,considering the possible existence of noise data,preprocess the data,including the pre-processing of repeated data and velocity anomaly data.Using the hidden Markov model to project the GPS trajectory onto the road network,extract travel time of the target link and its adjacent links for travel time prediction.Secondly,the feature vector is constructed by considering the time-space characteristics of the link travel time.However,the feature vector may be too large or the feature space is redundant.So use the principal component analysis algorithm to extract the main features.In view of the problem that most neural network models have many parameters and complex network structure,this paper uses generalized regression neural network model to predict link travel time.The structure of the generalized regression neural network is determined based on the new eigenvectors obtained after principal component analysis.Finally,considering the smoothing factor of the generalized regression neural network will directly affect the prediction accuracy of the network and the better global convergence characteristics of the particle swarm optimization algorithm,find the optimal smoothing factor using particle swarm optimization.The optimal smoothing factor is assigned to the generalized regression neural network to determine the network model.Using the GPS trajectory data published by DIDI online taxi-hailing service platform for Chengdu from November 1st to 30 th,the effectiveness of the proposed algorithm is verified by experiment and comparative analysis.