Spatial-temporal Cellular Traffic Prediction: A Novel Method Based on Causality and Graph Attention Network

Abstract

Cellular traffic prediction is crucial for intelligent network operations, such as load-aware resource management and proactive network optimization. In this paper, to explicitly characterize the temporal dependence and spatial relationship of nonstationary real-world cellular traffic, we propose a novel prediction method. First, we decompose traffic data into three components which represent various cellular traffic patterns. Second, to capture the spatial relationship among base stations (BSs), we model each component as a directed causal graph by variable-lag transfer entropy (VLTE) based causal structure learning. Third, we design a deep learning model combining graph attention network (GAT) and gated recurrent unit (GRU) to predict each component. GRU is used to capture temporal dependence. GAT is trained to quantitatively analyze spatial relationship and aggregate spatial features. Finally, we integrate the prediction results of three components to obtain the cellular traffic prediction result. We conduct extensive experiments on real-world traffic data, and the results show that our proposed method outperforms other common methods.