Mining And Prediction On Expressway User Travel Behavior Patterns Based On ETC Data

With the development of science and technology,China’s expressway has greatly shortened the time and space distance between provinces and cities after years of rapid construction.As an important traffic corridor,it has assumed the main force of transportation and provided important support for the development of China’s social economy.In order to vigorously promote the development of intelligent transportation systems,a large number of ETC gantry systems have been installed on 161,200 kilometers of expressways across the country and massive user travel data have been collected and accumulated.How to use big data mining technology to deeply mine ETC data? How to mine user travel behavior patterns based on ETC data? How to realize accurate prediction of expressway user travel behavior? Answering these questions will be of great significance to traffic network planning,travel guidance,personalized travel recommendation and other aspects.Therefore,in view of the above problems,this paper mines and predicts the travel behavior patterns of expressway users based on ETC data.The main work is as follows:(1)The spatiotemporal characteristics of expressway traffic flow and the statistical characteristics of expressway user travel behavior are analyzed.Aiming at massive ETC data,using big data processing technology and descriptive statistics,the temporal and spatial regularity and differences are analyzed in depth.The time dimension includes the distribution of traffic flow and travel time of different users under different date types and time scales.The spatial dimension includes the spatial distribution of traffic flow in different driving directions and road sections,as well as the spatial distribution of repetitive movement trajectories in user travel space,and the spatial distribution of frequent departure or destination.(2)This paper proposes a mining method of expressway user travel behavior patterns based on K-means++ and GMM hybrid clustering model.Firstly,features are constructed based on three aspects of time,space and user attributes.Low-variance filtering and Spearman correlation analysis are used to filter features without information value.For travel feature clustering,Kmeans++ algorithm is used for pre-clustering,and K value is judged by Silhouette Coefficient and Sum of the Squared Error index to provide GMM with approximate parameter range.Then,GMM algorithm is used for fine clustering,and EM algorithm is used to estimate parameters.The mining results of pre-clustering patterns are further optimized to obtain three travel patterns.Finally,with the qualitative and artificial judgment of various indicators,the travel behavior patterns of expressway users are classified into three categories: commuting activity mode,passenger and freight transport mode and short-term activity mode.(3)User Spatial Long Sequence Time-Series Forecasting(US-LSTF)model is constructed to predict the travel destination of expressway users.The model adopts modular design,in which the user travel mode discrimination module uses multi-layer perceptron to realize the first judgment of users.In order to solve the sparsity of expressway in space and time,the time representation module uses the local and global timestamp for coding;And in the spatial representation module,the distance matrix,adjacency matrix and the historical OD traffic flow matrix under different travel modes are constructed,and the convolution layer and hierarchical attention mechanism is used to fully mine the spatial features.The spatiotemporal features are fused into the codec module containing the multi-layer attention mechanism.In the encoder,the feature map with focus is obtained by stacking the multi-head ProbSparse self-attention mechanism and self-attention distillation operation,and then the feature map is input into the decoder.It is noteworthy that the long sequence received by the decoder uses the mask matrix to hide the real information to be predicted.This generative prediction method solves the problems of the traditional iterative prediction method,and the prediction results can be obtained only by decoding once.The effectiveness of the model was well demonstrated in several comparative experiments conducted on the ETC dataset.