WiP: TWINERGY - Towards Deployable Cell Shutdown with Network Digital Twins

Abstract

Cell shutdown is a critical mechanism for reducing mobile network energy consumption, but optimizing it remains difficult in practice. Live trial-and-error can overload cells and degrade service; currently deployed threshold-based policies are interpretable but hard to optimize at scale. This paper presents Twinergy, which searches for safe binary shutdown decisions in a high-fidelity network digital twin and maps them into thresholds for currently deployed threshold-based policies. Twinergy combines a data- and knowledge-driven digital twin, scalable multi-agent reinforcement learning (RL), and differentiable threshold mapping. Twinergy saves 26.0%/12.4% energy on real-world 4G/5G network data with 14,349 base stations. Live trials show up to 13.07%/12.20% energy-efficiency gains with no user-experience degradation.