Research On Expressway Network Traffic State Prediction Method With Missing Data

As urban agglomerations and metropolitan areas have become the“structural potential”of China’s development and the“new wind direction”of promoting economic growth,the trade between cities and regions has become more and more flourishing.Then,the expressway is an important channel in these regions,which occurs in frequent traffic congestion,and poses challenges to traffic safety and efficient operations.Timely and accurate traffic state prediction can be used to obtain future traffic information and the evolution law of traffic state,which is the basis for the formulation and implementation of system traffic control and services,and is also a prerequisite for ensuring the safe and effective operation of the traffic system.With the development of information collection,navigation mobile communication technology and other technologies,traffic forecasting has gradually changed to a data-driven forecasting model.However,due to detector failure,network delay and network packet loss and other unpredictable factors,data loss is inevitable in the actual environment.The data-driven model builds a regression model for the time series of specific parameters,so the data quality seriously restricts its application in actual traffic forecasting.In addition,the inherent nature of co-evolution naturally exists between traffic state parameters.Hence,how to effectively use the high-dimensional spatio-temporal characteristics of the traffic state and the coupling law between parameters,and introduce these laws as prior knowledge for the data-driven prediction model is to improve the prediction accuracy and robustness.And it is also the key to realizing the traffic state prediction of the expressway network facing the lack of data.Therefore,taking the high-speed road network as the research subject,this paper considers the characteristics of the co-evolution law of traffic state parameters,and combines deep learning and tensor theory to propose the road network traffic state prediction method for complex missing situations.From the perspective of theory and engineering,the proposed method not only improves the understanding of the coevolution law among traffic state parameters,but also provides accurate data basis for expressway operation and management,which is beneficial to improve the traffic safety,efficient management and intelligent level of expressway network.The main research contents of this paper can be divided into the following work:First,a road network traffic data completion method for various loss situations is studied.Aiming at the problems that high-dimensional multi-modal features are difficult to obtain and complex types exist in traffic data completion,a CP-decomposition-based completion model introduced manifold regularisation(MCP)is proposed.MCP analyzes the low-dimensional spatio-temporal characteristics of road network traffic data under multi-mode,and introduces manifold learning to analyze the mapping relationship between highdimensional spatio-temporal characteristics of the complex road networks and low-dimensional representation mode.It introduces manifold learning to map the relationship between high-dimensional spatio-temporal features and low-dimensional representation patterns of traffic state,which learns and acquires complex spatiotemporal features in various dimensions of the constructed traffic data while completing traffic data.Experiments on different traffic datasets show that the algorithm can adapt to a variety of loss scenarios and significantly improve the performance of traffic data completion.In the intercity highway network data set,it is proved that the accuracy of the algorithm is improved by nearly 50% compared with the traditional completion method in the case of missing 10% of data,and even if the missing rate is as high as the extreme case(80%),the algorithm can also achieve relatively good completion effects.At the same time,through the visualization of the decomposition factor matrix,it can be proved that the proposed algorithm successfully realizes the feature rule learning of each mode of traffic data while completing.Second,a complete-prediction method of road network traffic state parameters with data loss is studied.Aiming at the influence of various loss types in traffic prediction,considering the spatio-temporal coupling pattern characteristics and the spatial characteristics of road network nodes,a tensor combined temporal similarity revisited graph convolutional gated recurrent unit(T-TRGCGR)is proposed.T-TRGCGR is consist of an attention mechanism,graph convolutional network and recurrent neural network,and it introduces a graph regularized based tensor completion algorithm for complex loss of state parameters.On the basis of considering the two-way structure of traffic network,the tensor completion model and deep learning prediction network architecture of T-TRGCGR preliminarily integrated by graph Laplace matrix are introduced,and considering time similarity of traffic data to optimize the input of traffic state parameter,so as to realize the completion-prediction of traffic state parameters with missing values.The experiment proves that T-TRGCGR can achieve high prediction accuracy of traffic data with complex loss conditions.Compared with other algorithms,the error reduces 0.40 ～ 3.60,and the prediction accuracy improves 2 ～ 10% in the5 min prediction task.At the same time,in the experimental part,the performance and interpretability of TTRGCGR are analyzed.In the visualization of the attention model,the evolution law of the highway network traffic state data and the working mechanism of the deep learning algorithm for the traffic state prediction task are revealed.Third,a multi-parameter collaborative prediction method of road network traffic state with data loss is mentioned.Aiming at the influence of traffic state parameter prediction which always ignores the multiparameter co-evolution mechanism of traffic state data,considering the challenge of multi-parameter highdimensional data structure to traditional RNN algorithm,a multi-parameters hybrid tensor DL networks(MHT-Net)is proposed.Constructing the multi-parameter tensor graph convolutional network(MTGCN-C)and multi-parameter tensor graph convolutional network considering traffic prior knowledge(MTGCN-F)to obtain the spatial characteristics and the synergistic mechanism between traffic state parameters.The tensor GRU model that integrates the linear tensor layer based on Tucker decomposition(TT-GRU)is designed to obtain the time dynamic pattern of the synergy between parameters,and at the same time realizes the tensorization of network data input,output and parameters as well as improves the calculation speed of the proposed algorithm.Based on the intercity highway network data set,the proposed MHT-Net is compared with other existing prediction models in different prediction intervals,and the error is reduced by 0.60 ～ 1.00 under the5 min experimental task,and the accuracy is increased by 30 ～ 40%.The experimental results in different loss rate tasks also show that the algorithm can adapt to various missing rates,and still maintain a relatively stable prediction effect when the loss is 90%.At the same time,the algorithm can well deal with the situation of unbalanced data.Even in the bad condition of adopting traffic flow data as MHT-Net input,the prediction accuracy of speed and occupancy can still obtain a high level,and the speed accuracy can reach 90.79 %.