Enhancing network resource management through machine learning for spatio-temporal beam-level traffic forecasting

Accurate forecasting of Downlink Throughput Volume (DLThpVol) at the beam level is essential for improving resource management in modern communication networks. This study addresses the challenges posed by complex, high-dimensional spatio-temporal traffic data, leveraging multivariate time series that include critical factors such as Physical Resource Block (PRB) utilization and user count. Recent benchmarks on traditional and deep learning models (e.g., iTransformer, PatchTST, DLinear) achieve Mean Absolute Errors (MAEs) ranging from 0.1967 to 0.2005 on short-term targets and up to 0.2352 on longer-term forecasts, but opportunities remain for improvement through domain-informed feature engineering. We propose a dual-pipeline Gradient Boosting Decision Tree (GBDT)-based framework for beam-level DLThpVol prediction that incorporates carefully engineered temporal and spatial features (e.g., PRB utilization dynamics, beam-level user clustering). Our models achieve MAEs of 0.1919 (short-term) and 0.2261 (long-term), outperforming several deep learning benchmarks by up to 11.4% on short-term forecasts. These results demonstrate that interpretable, feature-driven ensemble learners can provide competitive forecasting performance while maintaining computational efficiency. Although the work does not directly implement congestion-aware resource allocation, the improved forecast accuracy lays the foundation for future studies on predictive resource management, such as PRB provisioning and energy-efficient beam scheduling. Our findings highlight the importance of combining domain knowledge with interpretable machine learning for advancing spatio-temporal traffic forecasting in communication networks.