ST-GAUTrans: Spatial-Temporal Gated Attention Transformer for Efficient Network Traffic Prediction

With the explosive growth of internet traffic and increasing complexity of network environments, accurate network traffic prediction has become pivotal for resource optimization and anomaly detection. Existing Transformer-based prediction methods suffer from high computational complexity and insufficient integration of multi-scale temporal features, resulting in high computational costs and inadequate prediction accuracy. To address these challenges, we propose the STGAUTrans model, which employs Gated Attention Unit (GAU) to efficiently model spatiotemporal characteristics of network traffic and introduces a Multi-scale Temporal Fusion (MSTF) module to adaptively integrate features across different temporal scales, significantly reducing computational complexity while improving prediction accuracy. The model comprises two key components: the Spatial Transformer Module (STM) for modeling spatial dependencies, and the Temporal Transformer Module (TTM) for extracting short-term trends and periodic variations, where the MSTF module within TTM adaptively integrates features across different temporal scales. Experimental validation conducted on the large-scale real-world cellular network traffic dataset from Milan demonstrates that the proposed model outperforms existing benchmark methods in both prediction accuracy and computational efficiency.