STGNN-AMD: Spatio-Temporal Graph Neural Network with Adaptive Multi-Scale Decomposition for Mobile Network Traffic Prediction

Abstract

In next-generation communication networks, adaptive resource orchestration, traffic engineering, and quality-ofservice (QoS) assurance are increasingly important, and accurate traffic prediction (TP) is a key enabling factor because it supports proactive and demand-aware resource allocation. Recently, graph neural networks (GNNs) have achieved strong performance on traffic prediction by modeling dependencies among base stations and spatio-temporal dynamics. However, existing methods relying on fixed decomposition rules and predefined periods such as daily cycles struggle to capture complex dynamics, leading to pattern interference and compromised robustness. To address these issues, we propose a Spatio-Temporal Graph Neural Network with Adaptive Multi-Scale Decomposition (STGNN-AMD). Specifically, the model first adaptively adjusts decomposition granularity to separate trend and residual components under redundancy-reducing constraints. Subsequently, it jointly leverages a dual-stream gated temporal convolution module, graph convolution, and attention mechanisms to characterize complementary spatiotemporal dynamics among base stations. The experimental results and ablation studies show that STGNN-AMD consistently outperforms baselines in MAE and RMSE, providing a robust and generalizable TP solution that can serve as a predictive foundation for demand-aware resource orchestration in future network operations.