Machine Learning-Based Predictive Congestion Management and Dynamic Resource Allocation in 5G Networks.

Abstract

Congestion in fifth-generation (5G) wireless networks has evolved from a simple bandwidth limitation problem into a complex interaction among ultra-dense deployments, heterogeneous service classes, and highly dynamic user mobility patterns. Conventional reactive congestion management techniques, including threshold-based admission control and static scheduling policies, are inadequate for handling the millisecond-level resource allocation dynamics required in 5G networks. This study proposes the Predictive Congestion and Resource Orchestration System (PACROS), a machine learning-driven framework that integrates short-horizon traffic forecasting, uncertainty-aware resource pre-allocation, and reinforcement learning-based closed-loop remediation within a unified congestion management architecture. PACROS employs a temporal convolutional network (TCN) enhanced with multi-head attention to predict per-cell traffic load over a 500-ms forecasting horizon, while a constrained optimization module proactively allocates physical resource blocks (PRBs), modulation and coding scheme (MCS) levels, and buffer admission thresholds ahead of predicted demand surges. A proximal policy optimization (PPO)-based remediation agent dynamically adjusts network resources in real time to mitigate residual congestion caused by forecasting uncertainty. The framework was evaluated using an ns-3/5GLENA-based 5G Non-Standalone simulation environment comprising 36 gNodeBs, 900 user equipment nodes, and urban mobility-driven traffic traces with injected anomaly scenarios. Results show that PACROS reduced mean buffer occupancy by 39.4%, decreased 95th-percentile packet delay by 44.7%, lowered URLLC service-level agreement (SLA) violation rates by 58.8%, improved PRB utilization efficiency by 22.3%, and reduced handover-triggered service interruptions by 31.8% compared with a 3GPP-compliant proportional fair baseline scheduler. PACROS also achieved an in-deadline URLLC packet delivery rate of 96.1%, outperforming all baseline methods during both steady-state and anomalous traffic conditions. These findings demonstrate the effectiveness of predictive, uncertainty-aware, and closed-loop congestion management as a practical foundation for autonomous and SLA-aware 5G network operation.