RL-FLOW: A Rectified Flow-Enhanced Reinforcement Learning Algorithm for Base Station Energy-saving Control

The exponentially increasing mobile traffic volume is driving a surge in energy consumption across global cellular networks. While next-generation technologies such as 5 G enhance network throughput and reduce latency, they concurrently mandate higher power demands for denser base station deployments and advanced processing units. This intensifying energy challenge positions efficiency optimization as a critical priority in the architectural design of next-generation mobile networks. In response to these challenges, we present a novel algorithm RLFLOW where Rectified Flow’s generative process is modeled as the action output in reinforcement learning (RL), leveraging ODE solving for accelerated training and inference. Experimental results demonstrate that, compared to conventional DDPM-based RL, our method reduces training time from 81 hours to 12 hours (7 x acceleration) and inference latency from 1 hour and 5 minutes to 5 minutes ($13 x$ speedup), while maintaining near-identical task performance, thereby achieving significant computational efficiency gains without compromising task performance.