Gated Temporal-Graph Attention Networks for Large-Scale Cellular Traffic Volume Prediction

With the exponential increase in data traffic, driven by the rapid growth of connected devices and high-bandwidth applications, precise prediction of network consumption becomes pivotal for optimizing telecommunications infrastructure. This challenge is intensified by the requirements of mobility prediction and Cellular Vehicle to Everything (C-V2X) applications, which necessitate high accuracy, low latency, and high reliability in network traffic prediction. In this paper, we propose a lightweight yet highly scalable Gated Temporal Graph Attention Network (Gated T-GAT) that forecasts network traffic volumes across complex environments, achieving state-of-the-art accuracy on critical short-horizon (15-minute) forecasts, an operationally critical window for telecom resource optimization. Our compact regression-based model, which integrates an MLP classifier with a Temporal-GAT, outperforms current multi-cell traffic models while eclipsing single-cell predictors. This work underscores the efficacy of coupling Temporal-GAT with an MLP classifier in managing large-scale network traffic, offering a scalable, efficient, and robust solution for high-demand forecasting scenarios.