Drift-Aware Federated Orchestration for 5G Traffic Forecasting: A System-Level Simulation Study

Traffic conditions in modern cellular networks vary considerably over time and across locations, making efficient coordination increasingly difficult for distributed learning systems. Existing distributed traffic forecasting methods have largely focused on improving prediction models or optimizing individual coordination mechanisms, whereas the coordination layer itself is commonly kept static under non-stationary traffic conditions. To address this issue, we develop a closed-loop adaptive orchestration framework in which a lightweight composite volatility index (CVI) continuously monitors traffic dynamics and jointly regulates client participation, aggregation, and participation fairness. The framework is evaluated using both controlled simulations and three independent real-world cellular traffic traces collected from two different measurement methodologies. Across all real-world traces, the proposed fairness-aware selection strategy consistently achieves the maximum participation fairness while maintaining prediction accuracy comparable to existing methods and reducing communication overhead by 12–23%. Controlled synthetic experiments further demonstrate that CVI-driven adaptive participation significantly outperforms fixed-budget coordination under the same communication budget ([Formula: see text]). Overall, the results show that lightweight closed-loop orchestration can substantially improve coordination robustness and fairness in non-stationary cellular environments without modifying the underlying prediction models.