<![CDATA[Energy consumption in 5G-Advanced and 6G Radio Access Networks (RANs) is a critical operational challenge, as energy costs account for 15–40% of mobile network operators' expenditures and base stations account for approximately 73% of total network power consumption. Traditional rule-based energy management, relying on fixed sleep schedules and static thresholds, is inadequate for dynamic traffic patterns and stringent Quality of Service (quality of service) requirements in next-generation networks. This study develops an AI-driven RAN Energy Controller that integrates machine learning-based multi-horizon traffic forecasting, predictive load balancing, and safe reinforcement learning (RL) orchestration within a three-plane Data, Intelligence, and Control architecture aligned with O-RAN principles. Multi-state sleep scheduling is optimized using constrained integer programming, while continuous adjustments to antenna tilt and transmit power enable adaptive energy management. Safety is enforced through action masking and Lagrangian constraint mechanisms to prevent coverage violations. A 48-hour synthetic multi-cell RAN simulation is conducted under carrier-grade guardrails, including P5 RSRP ? –110 dBm, PRB utilization ? 90%, and BLER 0.5%. The proposed controller reduces energy consumption by 24.3% relative to an always-on baseline while maintaining 99.8% system uptime and a 99.7% handover success rate. Zero coverage violations are observed under forecast errors of up to ±30%. Compared with rule-based methods and static scheduling, the controller achieves higher energy savings of 24.3%, compared with 15.4% and 12.1%, respectively, while avoiding the reported quality of service degradation of 1.8% and 2.3%. Multi-cell validation yields an accuracy of 84.7%. The results demonstrate that the proposed architecture can combine predictive intelligence, constrained optimization, and safe RL to enable autonomous, QoS-aware energy optimization for carrier-grade O-RAN deployments]]>
