Research On Travel Time Estimation Method Based On Deep Learning

With the soaring car ownership and the explosive growth of travel demand,the gap between road supply and demand is widening rapidly,which not only brings enormous pressure to traffic managers,but also causes great inconvenience to pedestrians.Real-time and accurate road network travel time can provide necessary information for traffic managers to allocate traffic resources and travelers to make informed travel plans.Therefore,the travel time estimation of the road network is essential for traffic management and route planning.In recent years,with the rapid development of smart city construction,GPS and other positioning equipment are widely dispatched,which provides rich data for travel time estimation.Meanwhile,data analysis technology represented by deep learning has also made considerable progress.In this context,with GPS trajectory data as the data source,this paper aims to employ deep learning technology to carry out in-depth research on the travel time estimation of the urban road network.Vehicle GPS trajectory data not only provide traffic information in a wide area for the study of travel time estimation,but also pose great challenges to the efficient usage of this huge and sparse data source.In addition,due to the complexity and relevance of road network traffic conditions,high-performance models are needed in the process of travel time modeling,and the spatialtemporal correlation should be fully considered at a network level.Therefore,based on GPS trajectory data,the following problems should be solved in the task of travel time estimation: big data processing;data missing caused by data sparseness;high-performance model;spatiotemporal correlation at a network level;implementation of proposed solutions.To tackle the abovementioned problems,with vehicle GPS trajectory data,this paper estimates road network travel time with specific solutions,verifies them by combining the numerical experiment with real-world traffic data,and develops an end-to-end software application system.The main contents and contributions of this paper are as follows:(1)Imputation of travel time with Generative Adversarial NetworksDespite their huge volume,GPS trajectory data are still sparse for a complex and large urban traffic network.Some remote links may have few or no data available,which will lead to the lack of travel time information in the road network.Therefore,this paper proposes a Travel Time Imputation Generative Adversarial Network(TTI-GAN)to provide complete data for related research on travel time estimation.Based on the spatialtemporal correlation of travel time at a network level,TTI-GAN can generate accurate travel time data for data-missing links by modeling travel time distributions of data-rich links.In order to analyze the spatialtemporal correlation of travel times,this paper utilizes Skip-Gram neural network model to encode the structure information and spatialtemporal evolution information of the road network.Then the low-dimensional semantic vectors generated by the Skip-Gram model serve as the input of the TTI-GAN model.In addition,to model link travel time distribution effectively,this paper introduces a Dynamic Clustering with Univariable Wasserstein Distance(DC-UWD)algorithm to cluster link travel time distributions.With the consideration of the mean,standard deviation and shape of the probability distribution,DC-UWD can accurately classify link travel time distributions into appropriate clusters.With the clustering labels,TTI-GAN model can effectively learn the heterogeneity of link travel times to improve its performance.(2)Prediction of average link travel time based on multitask learningBased on the travel time data supported by TTI-GAN,this paper presents an Multi-Task Learning Temporal Convolutional Neural Network(MTL-TCNN)to predict link travel time with the consideration of spatialtemporal correlation at a network level.In the MTL-TCNN model,the Temporal Convolution Neural Network(TCNN)can accurately and efficiently model time series data by combining dilated causal convolutions with residual mappings.In addition,this paper proposes a Spatiotemporal Dynamic Time Warping(ST-DTW)algorithm to select the most informative features for MTL-TCNN by quantifying the correlation among prediction tasks.(3)Estimation of route travel time distribution with modified Information Maximizing Generative Adversarial NetworksAverage travel time could provide direct traffic information for traffic management and route planning,but it cannot reflect the diversity of traffic conditions.Therefore,taking travel time data provided by TTI-GAN as the data source,this paper proposes an estimation method of travel time distribution,and constructs Trip Information Maximizing Generative Adversarial Networks(T-InfoGAN)by improving InfoGAN.To estimate travel time distribution,the focus of this study is not only to analyze the link travel times,but also to characterize the transition states among links.In this paper,a Dynamic Clustering with Multivariable Wasserstein Distance(DC-MWD)algorithm is introduced to study the transition of traffic states.Then the Cluster labels are treated as the input to help the T-InfoGAN model accurately learn the transition rule of traffic states and improve the performance of the model.In addition,without assuming the probability distribution type of travel times,T-InfoGAN model has great generalization and flexibility in estimating travel time distribution.(4)Development of travel time estimation systemIn this paper,with GPS trajectory data,an end-to-end software application system,named Travel Time Estimation System(TTES),is developed for the road network travel time estimation.In the way of mixed language programming,the system employs C#,Python and R to provide users with various functions: map display,GPS trajectory data load and display,vector map download,map matching,travel time extraction,data imputation,travel time prediction,travel time distribution estimation,route planning,traffic condition prediction,and so on.TTES implements the proposed travel time estimation solutions and validates the effectiveness and feasibility of these solutions through real-world data from different cities.