Data-Driven Lagrangian Optimization for O-RAN Efficiency

Abstract

This letter presents a data-driven solution based on an energy-saving rApp that uses unsupervised learning to optimize cell switch-on/off decisions, enhancing the energy efficiency of Open Radio Access Network (O-RAN) networks while guaranteeing Quality of Service (QoS) in terms of outage probability. The proposed rApp employs a data-driven policy trained with a differentiable Lagrangian loss function obtained by employing Gaussian approximation, Lagrangian relaxation and the primal-dual method. Validated in an environment aligned with the O-RAN Alliance, the proposed solution achieved a power consumption reduction of approximately 50% during low-demand periods while strictly maintaining QoS requirements.