Taxi Trajectories Destination Prediction Based On Deep Learning

The taxi trajectory reflects the taxi’s driving route and the city’s traffic status.Statistical methods or deep learning methods are deployed to predict the trajectory to determine the location of future trajectory in advance.It is of great significance to the intelligent transportation systems.Based on the deep learning theory,this thesis focuses on destination prediction in taxi trajectory prediction with neural network model.The main research contents are shown as follows:(1)Auto-encoder based prediction of taxi trajectories destinationThe traditional methods only utilize original information of the trajectory when predicting the destination of the taxi trajectory.There is a lack of potential temporal and spatial information in the sequence,which leads to low prediction accuracy.In response to the above problems,this thesis proposes an Auto-Encoder based Prediction of taxi Trajectories destination method(AEPT).The taxi trajectory is expressed as a sequence containing spatio-temporal information,and the self-learning ability of autoencoder is used to obtain spatio-temporal features’ deep representation.The problem of temporal spatio-temporal feature missing is solved.Experiments show that compared with the MLP that won the ECML/PKDD2015 competition,AEPT improves the prediction accuracy by 10.79%,and the average distance error is reduced by 1.34 km.(2)Heterogeneous graph embedding to predict taxi trajectories destinationThe spatio-temporal characteristics of the sequence data are considered only when utilizing the auto-encoder to extract the deep spatio-temporal features of the taxi trajectory.The lack of spatial semantic information extraction affects prediction effect.Aiming at the lack of spatial semantic information,a Heterogeneous Graph Embedding to Predict taxi Trajectories destination method(HGEPT)is proposed.Convert the original trajectory into a regional trajectory with taxis as the main body,capture the dependency between different taxi trajectories,and make up for the limitations of sequence data.Experiments show that compared to the MLP model is increased by13.98%,and the average distance error is reduced by 1.45 km.(3)Multi-Features fusion taxi destination prediction with hierarchical attentionAEPT learns spatio-temporal features with auto-encoder,and HGEPT extracts spatial semantic features with heterogeneous graph embedding.Combining the two will better integrate temporal and spatial features and spatial semantic features,thereby improving the prediction effect.Most traditional feature fusion methods are directly splicing features,ignoring the interactive information between different features.Therefore,this thesis proposes Multi-Features Fusion taxi trajectories destination prediction with Hierarchical Attention(MFFHA).A hierarchical attention mechanism is proposed to enhance the interactive information between features and improve the prediction accuracy.Experiments show that compared with the AEPT method and the HGEPT method,the prediction accuracy of MFFHA is improved by 11.20% and 8.01%,respectively.This thesis has 29 pictures,10 tables and 95 references.