Research on Congestion Adaptive Control in Power Data Centers Based on Reinforcement Learning and Bayesian Networks

Efficient congestion management is crucial for ensuring stable and optimized performance in power data centers. Traditional heuristic and rule-based congestion control methods often fail to adapt effectively to dynamic network conditions, resulting in high latency, packet loss, and reduced resource efficiency. To address these limitations, we propose a novel adaptive congestion-control framework integrating Reinforcement Learning (RL) with Bayesian Networks (BN), named RL-BNCC. The proposed model leverages deep reinforcement learning to dynamically optimize control policies in response to real-time variations in traffic load and resource utilization. Concurrently, Bayesian inference provides probabilistic modeling of network congestion uncertainties, enhancing the robustness and reliability of control decisions. Extensive experiments conducted on benchmark datasets GridSTATS, GridBench, FlowNet, and PowerLoad demonstrate the superiority of the proposed RL-BNCC framework. Specifically, RL-BNCC achieves an average throughput improvement of up to 12.5%, latency reduction up to 22.8%, packet loss rate decrease up to 21.3%, and energy efficiency enhancement up to 18.6% compared to state-of-the-art methods including DCTCP, HPCC, PCC, and conventional RL approaches. Ablation studies further validate the critical contributions of RL-based adaptive policy optimization, Bayesian uncertainty-aware decision-making, and real-time adaptive thresholding in achieving optimal congestion management. This integrated framework establishes a strong foundation for intelligent, adaptive, and scalable congestion control solutions in modern power data centers.