HDTwin-E5G: AI-Driven Digital Twin and Hybrid Deep Learning Framework for Energy-Efficient 5G Networks

Energy-efficient operation of 5G radio access networks is increasingly constrained by the tension between aggressive base-station (BS) hibernation and stringent quality-of-service (QoS) requirements under highly time-varying traffic. Existing optimisation- and learning-based schemes often assume access to fine-grained traffic traces and lack a structured mechanism for translating coarse operational measurements into deployable hibernation policies with explicit trade-off calibration. To address these limitations, this paper proposesHDTwin-E5G, a hybrid digital-twin (DT) framework that integrates traffic data regeneration, prediction, optimisation-driven feasibility filtering, and reinforcement-learning (RL) refinement for energy-aware BS control. The framework first reconstructs minute-level traffic profiles from hourly measurements using a Gaussian Process Regression (GPR) regeneration module with smoothing, thereby enabling stable short-horizon forecasting. It then employs an MDP-based decision stage and a stochastic constraint-aware optimizer to ensure SLA compliance, followed by a DT-validated Double DQN agent that refines feasible actions through an energy-saving ratio (ESR)–shaped reward. Importantly,HDTwin-E5Gintroduces a DT-first reward-weight tuning procedure to select Pareto-efficient trade-offs and derive deployable presets across traffic regimes. Experimental evaluations across representative urban-dense and suburban deployments demonstrate that the proposed framework achieves 28.7% and 33.1% ESR, respectively, while maintaining stable latency (approximately 42 ms and 39 ms), highlighting the effectiveness of combining data regeneration with DT-guided hybrid control for practical, QoS-aware energy savings in 5G networks.