Network Traffic Overload Prediction with Temporal Graph Attention Convolutional Networks

With the rapid explosion of network traffic volume, the prediction of traffic overload in the cellular network, which is defined as whether the traffic of a base station exceeds a predefined threshold, has become a crucial research problem regarding energy utilization and resource allocation. Most existing methods primarily model the dynamic patterns of traffic time series and compare the results with the predefined thresholds to predict traffic overload, taking into account a large amount of small-scale redundant data. To focus on the changes near the thresholds, i.e., the traffic burst circumstances, this paper adopts the soft-attention mechanism to fuse the threshold-based discrete and continuous time series characteristics to predict traffic overload. In addition, to capture the spatial correlations of the most related neighbors, we employ a Dual-KNN mechanism to select neighboring base stations and leverage the Graph Attention Network (GAT) to capture the spatial dependencies. Furthermore, we deploy a gated Temporal Convolutional Network (TCN) to model the temporal dependencies of the network traffic. Extensive experiments demonstrate that our proposed method effectively forecasts the traffic overload and outperforms the state-of-the-art algorithms by 4.07%.