A Practical AI-Driven Strategy for Cell On/Off Switching under Adaptable QoS Constraints

Abstract

The rapid expansion of 5G networks has intensified concerns over their sustainability, as denser Radio Access Network (RAN) deployments have increased overall power consumption. Although numerous studies have examined energy-efficient cell on/off switching, few have focused on approaches capable of dynamically adapting to operator-defined Quality of Service (QoS) requirements. In this paper, we propose a Long Short Term Memory (LSTM)based strategy, trained using a dataset from a European Mobile Network Operator (MNO), that enforces both target throughput levels and outage-tolerance constraints. Unlike previous approaches, our model adapts to different QoS requirements by tuning a decision threshold at inference time, enabling operators to balance energy savings and service guarantees without retraining. Across an unseen week, the method attains 63 to 96 % of an oracle's energy savings while largely meeting operator-specified constraints. We also provide CO2 and OPEX estimates under representative scenarios to quantify potential operator benefits.