A Hierarchical Geospatial-Temporal Fusion Transformer Framework for Interpretable Mobile Network Payload Forecasting

Abstract

With the escalating demand for mobile data, accurate forecasting of network traffic is paramount for resource management and Quality of Service (QoS). While Long Short-Term Memory (LSTM) networks have been applied, they often struggle with long-range dependencies and incorporating static metadata. More advanced spatio-temporal models often operate as “opaque model,” limiting their operational utility. This paper proposes GNTFT, a novel, hierarchical framework centered on the Temporal Fusion Transformer (TFT) that prioritizes both accuracy and interpretability. Our system first segments the network into behaviorally homogeneous clusters, then learns static, graph-based spatial embeddings for each sector using a Graph Neural Network (GNN). These embeddings are fused with historical KPI data within a specialized TFT model for each cluster. We conduct a rigorous benchmark against a traditional LSTM, Spatio-Temporal Graph Convolutional Network (STGCN), Diffusion Convolutional Recurrent Neural Network (DCRNN), and the state-of-the-art Graph WaveNet (GWN). The results demonstrate that while GWN achieves the highest absolute accuracy (MAE), our proposed Full GNTFT framework achieves superior performance across all metrics (MAE: 110.12, MAPE: 22.05%), compared to the baseline “TFT Only” model (MAE: 158.82, MAPE: 31.42%). Crucially, the GNTFT’s hierarchical architecture provides deep interpretability; by training specialized models for each behavioral cluster, we demonstrate that the predictive drivers of network traffic are highly heterogeneous. This work presents a transparent solution that bridges the gap between raw predictive power and the actionable, trustable intelligence that state-of-the-art models cannot provide.