Research On Taxi Spatio-temporal Feature Extraction And Travel Time Prediction Considering Road Network Topology

The massive taxi GPS track data mining is of great significance to urban traffic planning and residents'travel guidance.This study focuses on the analysis of taxi travel behavior patterns and prediction of road journey time,taking the GPS track data of Shenzhen taxis as the object of study.Through the process of track data pre-processing,map matching and taxi pick-up and drop-off location extraction,reliable data sets for behavior pattern analysis are established,spatio-temporal characteristics of taxis are analyzed,and a journey time prediction method based on LDA urban transportation theme pattern mining and map convolutional neural network GCN is proposed.(1)The spatial-temporal characteristics based on taxi GPS trajectory dataThe conclusion is as follows:1)temporal characteristic:early peak on holidays than weekdays and weekends 1 hour later,late peak than weekdays and weekends 1hour earlier;rest day early peak and weekday lunch peak within 40 minutes of the short-time travel is the main,in more than 60 minutes of travel time,early peak workdays than rest days,lunch peak rest days than workdays;workdays than holidays higher level of passenger carrying;workday travel is mainly within 15 kilometers,rest day travel is mainly 15-45 kilometers,the proportion of long-distance travel is greater than workdays.2)spatial characteristic:hot traffic is concentrated in several major commercial areas near the Futian port,the metro station,Union Square and Futian port is the hottest,rest day to Shenzhen People's Hospital metro,Guomao,Old Street is the hottest;hot roads on workdays to Huanggang Street,Huanggang Street,Shixia North Fourth Street,Fu Xiang Street most,rest day to Dongmen commercial walking street,Dongmen Road is the hottest.(2)The extraction of traffic theme area based on LDABased on the LDA theory,the complete spatio-temporal trajectory of the taxi is used as the document,and the passing road section is used as the document word,and the word-document-topic is transformed from road section-topic to road section-topic,and the road section that meets the threshold is visually displayed and the traffic topic area is extracted.The conclusions are as follows:1)the distribution of holiday traffic theme areas is obvious in the sections near the airport,railway station and tourist attractions;2)the traffic theme pattern is related to the importance of the sections,and the distribution of key traffic sections in Shenzhen such as Shennan Avenue,Riverside Avenue,North Ring Road and Beijing-Hong Kong-Macao Expressway is obvious;3)the distribution of traffic theme areas is regional in nature and closely related to road network,population and economy.(3)The prediction of travel time of road section based on GCNBased on the LDA transportation theme mode,a spatial-temporal graph convolutional network(T-GCN)time prediction method is integrated,and the average instantaneous speed of each road segment at each time period is obtained,and the travel time can be obtained by the ratio of the length of the road segment to the instantaneous speed,and finally compared with SVR,RFR,GBDT,DTR,XGBoost and other models.The experiment shows that the T-GCN model based on LDA and taking into account the spatial structure of the trajectory road network can be extracted from the road network.In the adjacency matrix information,the upstream and downstream conditions of urban roads are learned.The predicted MRE is 4.24%,MAPE is 8.19%,and R~2is 80.98%.The accuracy is higher than other comparison models,and the fitted curve is also better.From the experimental results,it can be seen that the prediction accuracy of the model based on the spatial topology of the road section is higher than other models.