An Adaptive Curriculum Event-Aware Federated Averaging Framework for Resource Demand Prediction in 5G Non-Standalone Networks

Abstract

The dynamic and heterogeneous 5G traffic complicates short-term resource-demand prediction. While federated learning (FL) enables privacy-preserving training by keeping raw data local. However, FedAvg and FedProx do not explicitly account for client traffic heterogeneity or sudden demand variations. To address these limitations, this paper proposes ACE-FedAvg, an Adaptive Curriculum Event-Aware Federated Averaging framework for next-hour downlink Physical Resource Block (DL PRB) utilization prediction in 5G Non-Standalone (NSA) networks. ACE-FedAvg jointly incorporates curriculum-based client selection, adaptive client importance estimation, and event-aware aggregation to prioritize informative clients and traffic conditions during federated training. The framework is evaluated using operational 5G NSA measurements from 127 Next Generation NodeBs (gNodeBs) collected over ten consecutive days, with the first nine days used for training and one unseen day for testing. ACE-FedAvg is compared with FedAvg and FedProx using multiple evaluation metrics. On the unseen test day, ACE-FedAvg achieves the lowest root mean square error (RMSE) of 0.654, peak RMSE of 1.715, top-5 error (1.501) and weighted RMSE of 0.773, compared with 0.670, 1.853, 1.733 and 0.799, respectively, for FedAvg. These results indicate improved overall and high-demand prediction performance for federated DL PRB utilization prediction under realistic 5G NSA network conditions.