Research On Data Imputation Of Multi-Source Traffic Spatiotemporal Big Data

With the rapid development of internet technology and traffic informatization,the scale of traffic data is getting larger and larger.In an intelligent traffic system,complete and effective traffic data is of great significance to traffic management.However,when collecting traffic data in real life,due to the occurrence of some inevitable events(such as equipment damage,bad weather,etc.),data collection will be interrupted and some data will be missing,which reduces the effectiveness of the data set and restricts the development of intelligent transportation construction.Effectively imputing the missing values in traffic data set has important theoretical and practical research significance.However,the imputation of traffic data is very challenging.On the one hand,the changes of road traffic data over time are non-stationary,such as morning and evening peaks,holidays,etc.,will affect the trend of traffic data.In the approaching time,traffic data has a strong time dependence.At the same time,traffic data also shows significant long-term cycle correlation;On the other hand,the real-world traffic road network has a complex spatial structure,and there are spatial correlations between different road network nodes.In addition,the missing pattern of the data also has an impact on the imputation of missing values.In response to the above problems,this thesis proposes an end-to-end spatiotemporal imputation network(STIN)to impute the missing values of traffic data.First,for the significant periodic features of traffic data,a non-causal time convolutional network is used to capture the long-term periodic correlation of each road segment data,and then the encoder-decoder structure is constructed;For the spatial correlation of the traffic data,the bidirectional long and short-term memory network based on the spatial attention mechanism is used.This network captures the dynamic spatial correlation between each traffic section and entire traffic network at the same time;Aiming at the close-time correlation of traffic data,a long and short-term memory network based on time attention mechanism is used to capture the time correlation between data at different moments and data at current moment.At the same time,the spatiotemporal correlation attenuation factor matrix is constructed based on the missing information of the data,by integrate the missing pattern information of the data into missing value imputation,it effectively improves the imputation accuracy.When the degree of data missing is high,the information provided by a single data source is not enough to support the imputation of missing values.On the basis of STIN,this thesis proposes a traffic data imputation model based on multi-source data fusion(MSTIN),The attention mechanism is used to capture the correlation between different traffic data and target data,which improves the imputation effect of the model.Using real-world traffic data sets as experimental data,this thesis compares the data imputation performance of the proposed MSTIN model and other benchmark models through experiments.The results show that the MSTIN model proposed in this thesis can accurately restore missing data,and its performance is better than other algorithms.At the same time,the influence of different modules in the model on the imputation performance is analyzed,and the effectiveness of modeling for spatiotemporal correlation and fusion of multi-source data is verified.