Intelligent and Sustainable Control for Energy-Efficient Open Radio Access Networks

The rapid densification of next-generation wireless networks, particularly in the evolution toward Sixth Generation (6G), has significantly increased the energy consumption and carbon footprint of Radio Access Network (RAN), where Radio Units (RUs) represent the dominant source of power usage. This thesis investigates intelligent control mechanisms for RU operation management in Open Radio Access Network (O-RAN) systems, targeting three key challenges: energy efficiency, scalability, and sustainability.
First, an O-RAN-compliant control framework is developed through the design of intelligent xApps deployed in the Near Real Time RIC (Near-RT RIC). Using a commercial RAN Intelligent Controller (RIC) simulator, the proposed approach achieves up to 50% reduction in power consumption by dynamically adapting RU operational states while maintaining Quality of Service (QoS).
Building on this foundation, the RU sleep control problem is formulated as a Markov Decision Process (MDP), and a Deep Reinforcement Learning (DRL) solution based on Twin Delayed Deep Deterministic Policy Gradient (TD3) is proposed. By leveraging continuous action spaces, the proposed approach avoids the exponential growth of action combinations and achieves over 50% energy savings compared to the always-on baseline, outperforming Deep Q-Learning Network (DQN)-based methods by up to 6% while improving convergence stability.
To address scalability, a federated DRL framework aligned with the hierarchical O-RAN architecture is introduced, where local agents operate within Near-RT RICs and a global model is coordinated by the Non Real Time RIC (Non-RT RIC). This approach achieves up to 43.75% faster convergence and reduces training energy consumption by 37.4%, while maintaining network-wide energy savings above 50%.
Finally, this thesis extends the optimization paradigm to sustainability by incorporating carbon awareness into RU control. A carbon-aware DRLframework is developed to account for heterogeneous energy sources, enabling differentiated activation strategies. Results show up to 20% reduction in carbon emissions compared to heuristic baselines, with joint energy–carbon optimization achieving approximately 12%–28% lower emissions than energy-only approaches.