基于深度强化学习的基站动态开关研究
Research on Dynamic Base Station Switching Based on Deep Reinforcement Learning
王瑜 1范燕琳 1孙洋洋 1熊建胜 1蒋涛 1周莹 1韩志博 2李子怡 2王振乾3
作者信息
- 1. 中国联合网络通信集团有限公司智网创新中心,北京 100048
- 2. 北京邮电大学安全生产智能监控北京市重点实验室,北京 100876
- 3. 北京科技大学工业互联网研究院,北京 100083
- 折叠
摘要
随着5G移动互联网的快速发展,为了满足用户不断增长的流量需求,5G基站大规模部署,导致能耗急剧增加.针对以上问题,通过采用流量预测与深度强化学习相结合的方法,建立基站动态开关模型.具体而言,该模型通过密集卷积神经网络(Densely Connected Convolutional Networks,DenseNet)对基站流量进行预测;进一步地,基于精确的移动流量预测,将基站开关控制问题建模为一个马尔科夫决策过程(Markov Decision Process,MDP),然后通过强化学习方法进行求解.此外,强化学习的reward函数设计在优化基站开关成本时综合考虑了多方面的因素,包括能耗和用户服务质量(Quality of Service,QoS)下降成本,目标是在降低能耗的前提下,最小化长期的基站能量消耗.最终通过对真实数据集的大量实验验证,提出的模型与当前使用的基站常开策略相比,能够节约37％的能量消耗,且节能效果也优于传统启发式算法.
Abstract
With the rapid development of 5G mobile internet,the large-scale deployment of 5G base stations to meet the ever-grow-ing demand for data traffic has led to a sharp increase in energy consumption.To address this issue,a dynamic base station switching model is established by combining traffic prediction with deep reinforcement learning.Specifically,this model uses a densely connected convolutional neural network to predict base station traffic.Further,based on accurate mobile traffic predictions,the base station switc-hing control problem is modeled as a Markov Decision Process(MDP),and solved using reinforcement learning methods.Moreover,the reward function design in reinforcement learning comprehensively considers multiple factors in optimizing base station switching costs,including energy consumption and the cost of degrading Quality of Service(QoS).The objective is to minimize long-term base station energy consumption while reducing energy use,without compromising QoS.Extensive experiments on real datasets demonstrate that the proposed model can save 37％of energy consumption compared to the current keep-on strategy for base stations,and its ener-gy-saving performance also surpasses traditional heuristic algorithms.
关键词
基站/蜂窝网络/动态开关/流量预测/深度Q网络Key words
base station/cellular network/dynamic switch/traffic prediction/DQN引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据