A cellular traffic prediction method based on diffusion convolutional GRU and multi-head attention mechanism

Amid the proliferation of big data, cellular traffic prediction has emerged as a critical component of intelligent communication networks. However, cellular traffic data exhibits complex spatial-temporal characteristics, posing significant challenges for timely and accurate prediction. Although numerous cellular traffic prediction methods have been proposed in recent years, many approaches fail to fully account for spatial-temporal dependencies, periodic patterns, and other influential factors, leading to suboptimal predictive performance. To address these limitations, a novel cellular network traffic prediction model, DCG-MAM (A Cellular Traffic Prediction Method Based on Diffusion Convolutional GRU and Multi-head Attention Mechanism), is proposed. The primary innovation of the proposed model lies in its well-designed multi-module framework, which effectively captures intricate spatial-temporal dependencies, periodic characteristics, and external influences, thus enabling more accurate traffic prediction. The proposed approach begins by partitioning cellular traffic data into three independent time segments, incorporating temporal feature metadata embedding to capture periodic traffic patterns. Subsequently, a combination of Diffusion Convolutional Gated Recurrent Units (DCGRU) and the Multi-head Attention mechanism is employed to extract spatial and temporal features. Furthermore, the external factors affecting cellular traffic are systematically incorporated into the model. Comprehensive experiments conducted on real-world datasets demonstrate that the DCG-MAM model significantly reduces prediction errors and achieves superior predictive accuracy.