Traffic Data Imputation Based on Self-Supervised Learning

In the regional highway network, there are numerous toll stations generating massive amounts of data on a daily basis. However, due to equipment and network issues, there may be delays in data transmission for some stations. In such cases, the transmitted data may not be sufficient to meet the requirements for real-time traffic flow prediction. To achieve real-time traffic data imputation and dynamic traffic flow prediction, this paper firstly proposed a method for data imputation of highway traffic flow data based on self-supervised learning, which adopts time series model based on attention mechanism (Seq2Seq-Att). Then the self-supervised learning method was used to train the model. Finally, the reliability of the method was verified by taking 80 toll stations in the highway network of Guangdong province as an example. The results show that the method in this paper can flexibly capture the missing pattern in traffic data and give a reasonable value according to the internal correlation of the data. This method is generally superior to other methods and has good performance under different missing rates. The overall MAPE is about 17.7% and the WMAPE is 12.8%. In the case of high missing rate, this method has obvious advantages over other methods. The results of traffic volume prediction indicate that the prediction accuracy of traffic flow prediction using the data completed by this method is close to the situation of using complete data.