Network-wide Multistep Traffic State Prediction And Dynamic Routing Considering Missing Data Problem

With the rapid development of information technology,the network-wide traffic detection data can precisely obtain the traffic status.However,it brings into the new challenges of complex missing patterns,high-dimensional joint distribution,and huge computation burden.Therefore,the approach to complete data processing and predict the future traffic state,so as to support the intelligent traffic management demands further exploration.Based on the real-world traffic datasets,this dissertation fulfills a series of researches on network-scale traffic data imputation,multistep traffic state prediction,and data-driven routing strategy.These contents logically go forward one by one,which bridge the gap between raw detection data and transportation service applications.The main contents and results are summarized as follows:(1)The customized bidirectional recurrent neural network is proposed for non-original missing data imputation.A new calculation unit is designed within the architecture,which utilizes the graph convolution operation and 1 × 1convolution module to extract the spatiotemporal dependencies in road networks.This approach outperforms other state-of-the-art methods by 6%-15%,and its computation efficiency is competent for online deployment.(2)The traffic speed data is integrated into the spatiotemporal matrix factorization framework for original missing traffic flow data imputation.The correlation coefficient of traffic speed between road segments constitutes the spatial constraint.The continuity of traffic flow constitutes the temporal constraint.The proposed approach is superior than other benchmark models by3%-20%.(3)The customized sequence-to-sequence learning model is proposed for multistep traffic state prediction.The graph convolution operation is incorporated into encoder and the periodic messages serve as the inputs for decoder.Then the attention mechanism is leveraged to fuse the encoder and decoder.Its prediction errors decrease by 5%-11% than other classic machine learning and deep learning methods.(4)The data organization scheme based on the road segment clustering is proposed for network-wide prediction.The dependencies between road segments are measured via the designed spatiotemporal similarity index.On this basis,the hierarchical clustering is executed.Then the datasets are organized according to the clustering results.The cluster-level prediction reduces the calculation time by about 70% than the link-wise and all-network scheme.(5)The data-driven routing method is proposed,which considers the future traffic state in the road network.The traffic state prediction results are model ed as time-varying factors and the empirical distribution of prediction errors are modeled as stochastic variable.Then the path with least expect travel time(LET)in this stochastic time-dependent network is defined.The algorithm for solving LET problem is further designed.Comparing to the mainstream routing strategies,this approach can save about 3%-8% travel time in peak hours.To sum up,this dissertation proposes a whole series of methods for the multistep traffic state prediction and dynamic routing.Six large-scale real-world traffic datasets are collected for the numerical experiments.In contrast with over thirty benchmark models,the effectiveness of proposed approaches is validated.