2026 International Conference on Networks Computers and Communications (ICNCCom 2026)

Energy management in wireless edge networks requires joint decisions on node sleep states and processing frequency while maintaining quality of service under bursty traffic. This paper proposes a user-centric controller that combines dynamic power management (DPM), dynamic voltage and frequency scaling (DVFS), one-step demand prediction, and Double Deep Q-Learning (Double-DQN). The controller observes current and predicted load, queue state, previous configuration, and a user-derived QoS sensitivity signal, then selects one of 15 joint active-node and voltage-frequency actions. A reproducible discrete-time simulation with 20 edge nodes was evaluated over 30 independent traffic realizations. The ridge demand predictor achieved an R-squared value of 0.965 and a normalized root-mean-square error of 7.13%. The proposed controller reduced mean network power by 32.5% relative to an always-on policy and by 11.0% relative to a reactive DPM-DVFS policy. It maintained 99.98% throughput, zero delay-target violations, and a 95th-percentile queueing delay of 9.33 ms for a 20 ms target. Removing user feedback reduced power by less than 1% but increased the delay-target violation rate to 5.47%. The results demonstrate that user-aware predictive control can achieve a favorable power-QoS tradeoff, while also identifying the limits that must be addressed before deployment in a physical radio access network.
 

deep reinforcement learning, dynamic power management, dynamic voltage and frequency scaling, energy-efficient networks, user-in-the-loop control.