| With the development of network communication technology and global satellite positioning technology,the impact of the"Internet plus"trend bring great changes to residents’transportation and travel,especially the popularity of smart phones,which add strong impetus to the rapid development of online car-hailing industry.With its fast,comfortable and convenient features,online car-hailing occupies a large number of traditional taxi market shares in recent years.At the same time,due to the accelerating process of urbanization,traffic congestion,taxi difficulty,difficult to find passengers and other traffic problems are still serious,online car-hailing supply and passenger travel demand mismatch problem is still plaguing both the supply and demand.In view of this series of practical problems,the academic circle and city managers are faced with the problems of what kind of temporal and spatial characteristics of residents’online car-hailing trips,what are the factors affecting the built environment of online car-hailing trips,and how to accurately predict residents’online car-hailing travel demands.Under the background of big data and artificial intelligence,this paper takes the area within the third Ring road of Chengdu as the research area,integrates online car-hailing order data and other relevant spatio-temporal big data,studies the spatio-temporal characteristics of online car-hailing travel under the small-scale grid,constructs a rich and effective system of explanatory variables,and deeply analyzes the global and local impact of the built environment on residents’online car-hailing travel.The spatial heterogeneity among explanatory variables is revealed,and then the significant influence variables and time series data are used to build and train the ride-hailing travel demand prediction model,and the performance of the prediction model is compared,verified and evaluated,so as to improve the accuracy and efficiency of online car-hailing travel demand prediction.This study is committed to providing reference and decision-making basis for relieving urban traffic congestion,reducing environmental pollution,improving the management and scheduling level of e-hailing cars,and providing residents with convenient travel.The main research contents and achievements are as follows:(1)Pre-processing and spatial statistics are carried out on the order data of online car-hailing.Under a small grid of 200m,the distribution characteristics of online car-hailing trips are studied in time and space,and the morning peak,afternoon peak,evening peak and night peak hours of online car-hailing trips in a day are accurately divided,and the travel characteristics of online car-hailing trips on weekdays and weekends are compared and analyzed.On this basis,this paper proposes a density-based clustering(DBSCAN)algorithm with noise combined with kernel density estimation(KDE)to effectively identify the hot areas of ride-hailing starting and finishing points under different date periods.(2)In order to study the global and local impacts of built environment variables on ride-sharing trips,we collect and preprocessed POI data,population,roads,buildings and other built environment variables.Through correlation analysis and collinearity test,we construct a built environment explanatory variable system containing 21 types of variables.Through the ordinary least squares(OLS)model regression,the global influence degree of the built environment explanatory variables of the model at different time periods of working days and weekends on ride-hailing trips is obtained.The results show that the midday peak model at the starting point of working day has the highest goodness of fit,and the adjusted R~2 was 0.352.Different variables have different effects on origin and destination models at different times.Then,in the study of the local influence of the built environment variables,the explanatory variables with significant performance in the OLS model are inserted into the geographically weighted regression(GWR)model and the multi-scale geographically weighted regression(MGWR)model for comparative regression verification.Compared with GWR and OLS,MGWR has smaller model error and higher fitting accuracy due to its multi-bandwidth characteristics and its ability to capture local features,and the adjusted R~2 reaches the highest 0.729 in the average all-day starting point model.At the same time,MGWR model captures the significant spatial heterogeneity of the influence intensity of various variables at different time periods on ride-hailing trips,among which the spatial heterogeneity distribution of road variables,accommodation service variables and financial insurance variables is obvious in the starting time of evening peak on weekdays and weekends and the average starting time of the whole day.(3)A multi-layer convolutional neural network-long and short term memory(CNN-LSTM)model based on multi-source spatial-temporal big data fusion is established by combining the best fitting significant explanatory variables in the analysis of the impact of the built environment with weather information and date type timing information data.The model not only accurately captures the spatial characteristics of ride-hailing demand,but also,The time dependence between training data sets is fully considered.After training and testing,the mean square error(MSE),mean absolute error(MAE)and R~2 of the accuracy evaluation indexes of the prediction model are 0.00087,0.0167 and 0.813,respectively.In order to ensure the effectiveness and accuracy of the model,it is verified that CNN-LSTM has better prediction effect compared with the mainstream traditional neural network model.Then,the performance of the model under different time granularity,different time step and different spatial grid scale is compared to determine the effectiveness of the model’s important superparameters.Under the setting of time granularity of 60 minutes,time step of 4 and spatial grid scale of 700m,the prediction model is relatively optimal in accuracy and operation efficiency.The results show that:From the perspective of small-scale and multi-period,the spatial analysis and hot area combination extraction algorithm can effectively capture the spatial and temporal characteristics of residential online car-hailing trips.Chengdu’s residential online car-hailing trips have obvious characteristics of different date types and time periods,and the distance attenuation of spatial online car-hailing trips and the structure of circle trips are significant.This method is suitable for analyzing the travel characteristics of urban traffic participants.OLS model and MGWR model are used to deeply explore the global and local influences of the variables of the built environment and the residents’online car-hailing trips.The spatio-temporal heterogeneity of influencing factors of ride-hailing trips is fully analyzed in the models of different time periods,and significant related built environment factors affecting ride-hailing trips are effectively explored.The CNN-LSTM model is built with built environment variables and time series variables as training characteristic data.Through multi-model,multi-scale and multi-parameter comparative experiments,the accuracy and performance of the model for residents’online ride-hailing travel prediction are verified,the prediction ability is improved,and new ideas and method support are provided for the improvement of the supply-demand relationship of residents’ride-hailing travel. |
PDF Full Text Request