Research On Traffic Missing Data Recovery Method Based On Sparse Representation

With the improvement of socio-economic and the increase of urbanization rate,people’s travel demand increases rapidly,which highlights the problem of traffic congestion caused by the imbalance between the supply and demand of road network capacity.It is of great significance to speed up the improvement of intelligent transportation system,enhance the traffic management and control technology capacity of urban managers,improve people’s travel efficiency and the effective carrying capacity of existing road network to solve traffic congestion.At the same time,the improvement of the intelligent transportation system and the data collection ability make the massive traffic data be obtained.Due to various reasons in the collection,transmission and storage of traffic data,there is a problem of data missing,which makes it difficult to carry out the work of traffic flow prediction and intelligent guidance in the intelligent traffic control technology.Therefore,the recovery of missing traffic data has become an important research content in the field of transportation.In this paper,machine learning method is used to systematically study the recovery of missing traffic data.The main work is as follows:（1）From the perspective of difference and correlation,this article analyzes the traffic flow characteristics in the actual road network traffic flow data.Aiming at the problem of data missing,this article analyzes the causes of data missing and different models of data missing,and introduce three kinds of missing data generation models:complete random missing,random missing and mixed missing,which are used in the experimental data preprocessing,and provides a theoretical basis for the follow-up research on the recovery method of missing traffic data.（2）In this paper,the probabilistic principal component analysis,local least square regression and low-rank matrix completion model are studied,and the experiments are carried out on the real traffic data of Portland,USA.The experimental results show that when the missing rate is between10%²0%,the local linear component of non-linear data is used in the local least square regression,which is better than other methods by 0.87%¹4.24%.When the missing rate is increased to 30%⁵0%,the probabilistic principal component analysis shows better performance.（3）Based on the principle of sparse representation,each traffic data sample is represented as a sparse linear combination of other samples,and a method of recovering missing traffic data based on sparse representation is proposed.In view of the problem that the solution is too sparse due to the regularization of L₁-norm and too dense due to the regularization of L₂-norm,the elastic network regularization is used to integrate the advantages of both L₁-norm regularization and L₂-norm regularization,so that the solution is neither too sparse nor too dense.（4）In order to solve the limitations of linear SR-EN model in the recovery of missing traffic data,a nonlinear mapping method is proposed to map the traffic data into the high-dimensional feature space,so that the mapped traffic data samples are distributed in multiple linear subspaces.Aiming at the problem of high computational complexity in explicit nonlinear mapping,the traffic data is implicitly mapped to high-dimensional feature space by using kernel method.In order to solve the optimization problem of KSR-EN model,the monotone fast iterative threshold contraction algorithm and the projection gradient descent method based on the Armijo step rule are used to solve the problem alternately.Experiments were carried out on the simulated data and the real data of Portland City.The experimental results show that,compared with SR-EN model,the nonlinear KSR-EN model has better adaptability to the nonlinear data and higher recovery accuracy.（5）In view of the impact of different methods of traffic data recovery on traffic flow prediction,this article puts forward the analysis framework of the impact of traffic data recovery model on traffic flow prediction（TIAF-TMVR）.Based on LSSVR,LSTM and KNN,two simple methods to deal with missing data and four methods mentioned above are compared on real traffic data.The experimental results show that it is of practical significance to recover missing data.Finally,the data recovered by KSR-EN model shows good robustness.