Search for: All records

IntroductionBeam-level traffic forecasting plays a vital role in the optimization of 5G networks by enabling proactive resource allocation and congestion control. However, the task is complicated by inherent data sparsity and the presence of multi-scale temporal dynamics, making accurate predictions difficult to achieve using conventional models. MethodsTo address these challenges, we propose a Gated Recurrent Unit (GRU)-based Multi-Task Learning (MTL) framework, enhanced by a weighted ensemble approach. We systematically evaluate the performance of six forecasting models—Linear Regression, DLinear, XGBoost, Echo State Network (ESN), Long Short-Term Memory (LSTM), and GRU-MTL—across three input sequence lengths (168-h, 24-h, and 8-h) using real-world beam-level data from the ITU AI for Good initiative. ResultsExperimental findings reveal that the GRU-MTL model significantly outperforms traditional baselines, achieving a Mean Absolute Error (MAE) of 0.2136 on 168-h sequences compared to LSTM’s 0.3223. Long sequences (168-h) reduce MAE by 56% relative to short 8-h windows, effectively mitigating the effects of sparsity. Furthermore, an ensemble of top-performing models (MTL, XGBoost, and Linear Regression) yields additional gains, reducing MAE to 0.2105—a 1.45% improvement over MTL alone. DiscussionThese results highlight the importance of long-term temporal context and model diversity for robust traffic prediction in sparse environments. The proposed framework offers practical guidelines: 168-h forecasting windows are optimal for weekly planning, and model ensembling enhances generalization across varying beam activity levels. This study contributes a scalable and accurate solution for spatio-temporal traffic forecasting in next-generation wireless networks.