Research On Data-driven Link Travel Time Estimation And Prediction

Travel time prediction in advance is of great significance for users obtaining traffic status,making travel plan according to traffic state in advance and improving the quality of trip.As the successful application of deep learning theory in image classification,speech recognition,machine translation and text understanding,it is possible to introduce deep learning theory into traffic field and it provides a new solution for link travel time prediction incorporating floating car trajectory with deep learning model.Therefore,we first solved the problem of link travel time estimation in the condition of of data sparsity.On the base of above knowledge,we extracted traffic characteristics from big data of trajectory from floating car and designed a variety of deep learning model based on the TensorFlow platform.Consequently,we make further research on link travel time prediction,link travel time prediction based on spatiotemporal correlation and long time prediction,respectively.The specific research work and achievements include the following aspects:(1)In view of the problem of trajectory data sparsity,we researched link travel time estimation method in the condition of of data sparsity.Through analyzing the correlation of link traffic in time and space,we extracted spatiotemporal correlation of traffic characteristic between target and adjacent links and proposesd a link travel time estimation method based on spatial and temporal correlation;In addition,we also proposed a link travel time estimation method based on road network similarity by measuring the similarity of link between attribute information and spatial structure information.We research link travel time estimation method under the condition of missing trajectory data from two aspects of spatiotemporal correlation and similarity of road network.Finally,the effectiveness of our proposed method is verified by experiments.(2)We introduced deep learning theory into travel time prediction field and proposed a method of link travel time prediction based on deep neural network model.Through analyzing the correlation of road transportation,we designed time rransversal feature vector,period vertical feature vector and cycle feature vector,which reflectiong traffic characteristic.We designed different deep autoencode neural network model,deep LSTM neural network model and deep convolution neural network model based on constructed time feature matrix,period feature matrix,cycle feature matrix and fusion feature matrix.We researched the influence of the number of encoding layer,the number of LSTM layer and the number of convolution layer on the prediction precision.Finally,we validate the effectiveness of the model by real traffic data.The experimental results show that the deep LSTM neural network model has the best accuracy in the three kinds of model.(3)Considering the topological relationship among road network to road traffic characteristics,we extracted spatiotemporal characteristics of road traffic.We put forward link travel time prediction method based on spatiotemporal correlation of target and adjacent links.We expressed correlation of road traffic using spatial adjacency matrix and analyzed the spatial and temporal correlation of road traffic characteristics,.We designed spaiotemporal transversal feature vector,spaiotemporal vertical feature vector and spaiotemporal cycle feature vector,which reflectiong spaiotemporal traffic characteristic.We researched link travel time prediction method based on deep autoencoder neural network model,deep LSTM neural network and deep convolutional neural network model using spaiotemporal fusion feature vector.Finally,we validate the effectiveness of the model by real traffic data.(4)In view of the difference between long time prediction and short-term prediction of link travel time,we proposed long time prediction method based on deep neural network model.We designed spatiotemporal fusion feature vector suitable for long-term prediction by analyzing the relationship between the maximum number of lag period which has correlation and the number of predicted period.On the base of above knowledge,we researched long time prediction method based on deep autoencoder neural network model,deep LSTM neural network and deep convolutional neural network model.At last,we validate the effectiveness of long time prediction by real traffic data.