Spatial-Temporal Attention-Convolution Network for Citywide Cellular Traffic Prediction

• Published in: IEEE communications letters Vol. 24; no. 11; pp. 2532 - 2536
• Main Authors: Zhao, Nan, Ye, Zhiyang, Pei, Yiyang, Liang, Ying-Chang, Niyato, Dusit
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.11.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Cellular communication; Cellular traffic prediction; Convolution; Correlation; deep learning; Electronic mail; Laplace equations; Machine learning; Modules; Prediction methods; Resource management; Resource utilization; spatial-temporal attention-convolution network; spatial-temporal correlation; Traffic information; Traffic models; Urban areas
• ISSN: 1089-7798; 1558-2558
• DOI: 10.1109/LCOMM.2020.3012279

Cellular traffic prediction plays an important role in network management and resource utilization. However, due to the high nonlinearity and dynamic spatial-temporal correlation, it is challenging to obtain the traffic prediction accurately. In this letter, a spatial-temporal attention-convolution network is proposed to predict the citywide cellular traffic. Considering the temporal correlation of the cellular traffic, the traffic data is modeled by the hourly, daily and weekly traffic components independently. In each traffic component, the spatial-temporal attention module is designed to capture the dynamic spatial-temporal correlation of cellular traffic; the spatial-temporal convolution module utilizes the graph convolution and standard convolution to obtain the spatial and temporal features simultaneously. By considering the external factors, the prediction of final mobile traffic is obtained. Experiment results indicate that the proposed approach outperforms the existing prediction methods on the real-world cellular traffic datasets.