Missing Traffic Flow Data Imputation On Spatio-Temporal Generative Adversarial Network

With the development of machine learning and deep learning,the imputation accuracy of time-series missing data has also been greatly improved.Traffic flow data,as a new type of time-series data,the missing phenomenon is common and inevitable due to its own Spatio-Temporal features and the influence of complex external factors.How to better learn the latent distribution of the existing traffic data becomes the key to improve the accuracy of incomplete data imputation.The imputation accuracy of previous approaches,proposed to deal with missing traffic flow data,are proved to reduce performance under complex traffic conditions.This paper analyzes the characteristics of traffic flow data and the existing problems in imputation approaches,combines the idea of generating data from Generative Adversarial Network,and fuses the Spatio-Temporal correlation of traffic data,then proposes the Spatio-Temporal Generative Adversarial Network to impute missing traffic flow data.It improves the accuracy of data completion and simultaneously expands the application on missing time series traffic flow data of Generative Adversarial Network.This paper firstly analyzes the Spatio-Temporal features of traffic flow data,then proposes the imputation of missing traffic data based on the Spatio-Temporal Deep Convolutional Generative Adversarial Network(ST-DCGAN).Firstly the data was processed into a time-correlated multi-channel traffic matrix,then used the games between generator network and discriminator network to obtain the Spatio-Temporal distribution feature of traffic flow data,and finally used the trained model to fill the missing traffic data.In addition,considering the influence of external factors on the change of traffic flow data,this paper proposed the Spatio-Temporal Improved Generative Adversarial Network for missing traffic data imputation with external factors(ST-IMGAN-ext)based on the ST-DCGAN,which changes the input and output of the network,optimizes the network structure and objective function,and simultaneously considers the impact of temporal and spatial dependencies and external factors for the changes of traffic flow data.In order to measure the effectiveness of the Spatio-Temporal Generative Adversarial Network for missing traffic flow data imputation,experiments were performed on the real open source TaxiBJ GPS dataset.The experimental results were evaluated by Root MeanSquare Error(RMSE).Compared with the traditional imputation methods,this method,proposed in this paper,is proved to be more accurate and applicability.