Research On Travel Time Estimation Method Based On Trajectory Data Mining

Travel time is an important parameter for transportation tasks and an important factor for residents’ travel considerations.It has important applications in intelligent transportation systems and location-based services.Accurate prediction of travel time will provide better help and support for urban intelligent transportation.The traditional route-based travel time estimation methods have certain limitations.They do not fully make full use of the machine learning methods to mine the value of trajectory big data.For this problem,the research on travel time estimation method based on trajectory data mining is studied in this thesis.First of all,this thesis implements a trajectory data preprocessing framework for taxi passenger trip extraction,including trajectory data compression,noise filtering and road network matching.And a trajectory outlier detection algorithm Tra LOF based on Local Outlier Factor is proposed.Experiments prove that the preprocessing framework can effectively complete the preprocessing of passenger trip trajectory data and Tra LOF algorithm can detect abnormal trips on the same OriginDestination dataset with a small number of trip features.Secondly,to overcome the shortcomings of the inadequate data sources and the one-sidedness of the single model used in the previous methods based on Origin-Destination data,this thesis proposes a travel time estimation model TTE-Ensemble based on ensemble learning by analyzing and digging the movement behavior of taxi Origin-Destination data.This model extracts spatial features from Origin-Destination data through clustering and meshing method and combines multi-source data features to predict travel time by integrating Gradient Boosting Decision Tree and Deep Neural Network.Experiments show that compared with the existing methods,the travel time estimation accuracy has been significantly improved on the Shanghai datasets containing 940,000 records and Chengdu datasets containing 1.17 million records.Finally,in view of the complex spatio-temporal correlation of trajectory data,this thesis proposes a travel time estimation model Tra2 Time based on feature extraction of trajectory data and spatiotemporal multi-task learning.In this model,the global correlation feature of trajectory data sequence is extracted by Multi-Head Self-Attention mechanism,and the sequential feature of trajectory data sequence is extracted by Bi-directional Long Short-Term Memory network.At last,it uses the spatiotemporal multi-task learning framework to train the model.Compared with the other existing methods,experiments on the Chengdu dataset with 1.2 million records show that the travel time estimation accuracy of Tra2 Time model has been further improved since our model could effectively extract the spatio-temporal features that affect travel time.