Learning And Prediction Over Mass Taxi Trajectory Data

With the rapid development of the urbanization,the urban population both permanent residents and migration groups are increasing year by year,the demand in public transportation is rising.However,the pace that city's road facilities are constructed is slow,which results in the seriousness of the contradictions between supply and demand of traffic.At the same time,with wide deployment of Intelligent Transportation Systems(ITS),the data including pedestrian,vehicle and road network could be acquired by multiple sensing devices in real time,especially the popularization of portable GPS devices which bring a lot of convenience to transportation management and citizen trip.The massive GPS trajectory data contains abundant temporal and spatial information,which provides the opportunities for analyzing the congestion of urban trunk road,finding the features of citizen travel behaviour,and assisting transportation department to plan the effective policy.This paper applies the latest research achievement of the machine learning,including clustering analysis and deep neural network,to study and excavate the deep nonlinear feature of trajectory data.Therefore,this paper mainly research three aspects as follows.The first problem in this paper is the study of urban's trunk road congestion.Firstly,the spots clustering method is used to identify hot spots of urban congestion,the result shows that the second ring road in Xi'an is the most congested road.Then,by the aspects of urban vehicle growth and private car tail limit policy,this paper analyzes the spatial-temporal distribution characteristics of the road congestion of the Second Ring Road to provide the opinion to distinguish the congestion.The second problem is that predicts travel demand by taking taxi.Firstly,the massive taxi GPS data is combined with external effective factors,including holidays,weather condition and air quality index.Secondly,The prediction model named CNN-LSTM-ResNet(CLR)is proposed which based on the hybrid deep learning among Convolutional Neural Network(CNN),Long Short-Term Memory(LSTM)and residual network.CNN layer is used to extract the spatial characteristics of traffic flow in urban area in this model,and residual units are introduced to deepen the network layers,then LSTMlayer is used to study the proximity,periodicity and trend of GPS data.Finally,the above three components are fused by weights,and further fused with external factors.Comparing with traditional forecasting models,proposed model has higher predict precision.Therefore,by forecasting passenger's travel demand,transportation departments can dispatch vehicles in advance to balance supply and demand.The last problem is that estimates the passengers travelling time based on the location of origination-destination(OD)and relative departure time.First,aiming at the problem that the grid method can not distinguish the traffic flow of different altitude roads in the multistory overpass area,the hierarchical division method of grid-nested road network is used to present a fine-grained representation of taxi OD flows in this paper.The proposed method brings the traffic flow with road signs,thus refining the expression of the OD flow to the level of road network.Then,on the basis of verifying the positive correlation between travel time and travel frequency,a model named convLSTM-Conv2DTranposeSeparableConv2D(CLTS)based on hybrid deep neural network is designed to estimate the taxi passenger's travel time in urban areas.In the model,firstly,ConvLSTM layer is used to learn the spatial-temporal feature of OD flow.Then the Conv2 DTranspose layer is utilized to extract the sparse feature of OD flow in order to overcome the sparse problem of fine-grained OD flow.Finally,the SeparableConv2 D layer is applied to convolve the positive correlation between trip duration OD matrix and trip frequency OD matrix by the way of point by point.The real taxi trajectory data is used in this paper,the proposed model has the lower Root Mean Squared Error(RMSE)and Mean Absolute Error(MAE)than that of traditional forecasting model and the current deep neural network models.