Research On Data-driven Trael Time Estimation And Prediction In Urban Road Network

Nowadays,the urban road transportation system in China experience increasing congestion with the advancement of urbanization,which threatens not only the transport efficiency but also the environment.To tackle these problems,knowledge about traffic condition is critically important at many levels of transportation planning and management.As an extremely important basic traffic information,travel time directly describes the traffic condition on road network and has drawn a great of attention in recent years.Accurate and reliable travel time in urban road network is an important foundation of the Intelligent Transportation System(ITS).It is potential to help improving the efficiency of the transportation system and thus reducing urban traffic congestion,and plays an extremely important role in many aspects of urban traffic planning and transportation management.However,due to the complexity of the urban transportation system,the travel time in the urban road network is impacted by the combination of internal and external factors,and thus is intrinsically uncertain.As a result,accurate travel time estimation and prediction in the urban signalized network is an extremely challenging subject.With the development of information technology and intelligent transportation systems,the collection methods of traffic data are becoming more and more abundant,and transportation has entered the era of big data from a period of lack of data.The ever-increasing massive data contains a large amount of useful information and knowledge,which provides many possible opportunities to solve complex traffic problems in urban road networks.How to find some approaches to mine the knowledge from massive data,so as to estimate and predict the travel time of urban road network accurately has been a hot and difficult point in transportation research.Aiming at alleviating traffic congestion by providing useful traffic knowledge,this thesis focuses on the research of data-driven approaches for travel time estimation and prediction in urban road network under the environment of traffic big data era,from the data mining perspective.Through the proposed approaches,travel time in urban road network can be estimated and predicted more accurately,which provides a better data support for improving the operational efficiency of transportation system and then helping to alleviate traffic congestion in urban road network.The work conducted in this thesis contributes to the development of ITS in both theory and practice.In accordance with the objectives,the main research work and innovations of this thesis are summarized as follows.(1)Massive and sparse GPS data processing methodTaxi-based floating car GPS data is usually large scale in amount and sparse,and the quality of data is vulnerable to various impact factors.In order to solve these problems,this paper studies the methods of data processing and travel time extraction from massive and sparse GPS data in depth,and proposed some improved approaches according to the existing drawbacks.Aiming at fixing the possible errors of the original GPS data,several kinds of data preprocessing measures are proposed from the perspectives of sampling time interval,instantaneous driving speed,vehicle dwell time and spatial position drift,through which the quality of GPS data can be effectively improved.In order to obtain travel time information from the massive and sparse GPS data,the travel time extraction processing method based on GPS data is studied in depth and improved,including GPS data map matching,sparse data path inference and path travel time allocation.In order to obtain better GPS data map matching results,a map matching method based on Hidden Markov Model HMM is proposed.On this basis,an improved Dijkstra shortest path search algorithm is proposed to derive the vehicle travel path between two adjacent sparse GPS data points.Afterwards,the Hellinga algorithm is chosen to allocate the path travel time between two GPS data points to all the road segments of the path.Through the proposed data processing methods,travel time of road segment can be effectively extracted based on the massive and sparse GPS trajectory,which provides a data basis for the subsequent travel time estimation and prediction tasks.(2)Network-level travel time estimation in urban road network using sparse GPS big dataThere exist many challenges for the analysis of travel time in urban road network based on GPS data,including data sparsity,traffic condition fluctuation,and effective and efficient network-level modeling.Aiming at addressing these challenges,this paper proposes a network-level travel time estimation in urban road network using sparse GPS big data,with the help of tensor-based modeling.The proposed approach is a data-driven spatial-temporal relationship model,consisting of four major components: map matching,travel time modeling,probabilistic travel condition clustering and travel time estimation based on tensor decomposition.It considers not only the spatial correlation of travel time on different road segments,but also the difference of travel time under different traffic conditions,and accounts for both the fine-grained temporal correlation of travel time in recent time slots and the coarse-grained temporal correlation of travel time between historical time slots and current time slots.Extensive experimental results of the real case study in the urban road network in Beijing,China,based on the big and sparse GPS trajectory,demonstrate that the proposed model has the ability to estimate not only the travel time of different road segments in the road network under different traffic conditions in the current time slots,but also the corresponding occurrence probability.Compared with the competing methods,the proposed approach is potential to estimate more travel times with higher accuracy.Besides,it is more robust to model parameters.(3)Tensor-based Bayesian probabilistic travel time estimation in urban road networkDue to the fluctuations in traffic demand and supply,stochastic arrivals at intersections and personal driving behaviors,travel time in urban road network is intrinsically uncertain and the travel time of different drivers are different.To solve these problems,this paper proposes a tensor-based Bayesian probabilistic model for travel time estimation in urban road network,based on the idea of probabilistic modeling.The proposed approach models the travel time of different drivers on different road segments in the road network in different time slots as a third-order tensor.Considering the uncertainty of the travel time on the road segment in the urban road network,each element in the tensor is modeled as a random variable following the lognormal distribution.Through the fully Bayesian treatment,the hyper-parameters of the model can be automatically tuned and thus the complexity of the model is automatically controlled,which helps to avoid the over-fitting problem when using large-scale but sparse data.The proposed approach is a context-aware spatial-temporal relationship model,which account for not only the spatial correlation of the travel time on different road segments,but also the difference between the travel times of different drivers,as well as the fine-grained temporal correlation of travel time between the recent time slots and the coarse-grained temporal correlation of travel time between the current time slots and the historical time slots.The actual case study based on massive GPS data in Beijing,China verifies that the proposed model is able to estimate the travel time in the urban road network with high precision without over-fitting.It is a robust model and the estimation accuracy is not sensitive to the initialization of the model parameters.(4)Deep learning-based travel time prediction in urban road network incorporating contextual informationTo solve the problems of limited model power for shallow architecture and lack of consideration of context information in existing travel time prediction studies,this paper proposes a deep learning-based travel time prediction model in urban road network incorporating contextual information.The proposed approach models the problem of travel time prediction with a data-driven machine learning paradigm,based on the extracted road-specific features,context-specific features,temporally correlated features and spatially correlated features.On this basis,a deep network is constructed to better capture the high complexity of travel time in urban road network,using sparse denoising auto-encoder as building blocks.In order to learn the deep network effectively,a layer-by-layer semi-supervised pre-training method is devised based on the idea of greedy learning.The proposed model is a deep learning-based model,which considers not only the influence of the characteristics of the road segment itself and the surrounding environment on travel time,but also the spatial correlation of travel time on different road segments and the temporal correlation of travel time in different time slots.It extracts and takes full advantage of the information contained in the large amount of data without label,and is able to learn the features used for travel time prediction layer by layer in an unsupervised manner.As a result,it is deep network with powerful ability to capture the complex nonlinear phenomena in urban transportation.The experimental results of the case study in the urban road network in Beijing,China,demonstrate that the proposed deep network has the ability to make full use of the useful information extracted from the massive data with noise,and then predict the travel time in urban road network with high accuracy.Compared with the competing methods,it exhibits certain advantages with better model robustness and scalability.