Abstract: Considering the present studies on short-term traffic flow prediction algorithms focused on the lowdimension information features of traffic flow, which fails to meet the requirements of prediction accuracy, this paper introduces the high accuracy low-rank tensor completion theory to construct dynamic tensor models based on week, day, and period and designs a short-term traffic flow prediction algorithm, which combines the multi-dimensional temporal characteristics of traffic flow. The proposed prediction algorithm is verified using velocity data of the Dujiakan section of Beijing-Hong Kong-Macao Expressway. The results show that this algorithm can achieve good prediction results quickly based on fewer historical data. It can achieve effective prediction for the traffic flow of weekday and weekend, and the mean value of mean absolute error is approximately 3.6%; furthermore, the fluctuation of traffic flow is tracked in real time. For the flow with missing data, the prediction accuracy of the proposed algorithm would decrease with the increase in the ratio of missing data. However, compared with the three classical prediction algorithms, the proposed algorithm shows better prediction accuracy using the preprocessed missing data. 

Key words: traffic engineering, short-term traffic flow prediction, high accuracy low-rank tensor completion, tracking