1.河海大学计算机与信息学院,江苏省南京市 211100;2.华能澜沧江水电股份有限公司,云南省昆明市 650214
Ceph系统被广泛应用于电力数据的存储,现有数据中心依赖人工经验对存储配置参数进行优化。但是,人工经验无法适应电网的动态变化,准确性低,无法满足激增的电网边缘快速存储与处理需求。针对以上问题,在电力数据存储系统中提出一种基于强化学习的数据负载感知自适应配置参数推荐方法。该方法基于强化学习的马尔可夫链蒙特卡洛采样算法配置参数样本集和相应的集群性能,使用分层建模方法构建性能预测模型,采用集群性能代价函数与负载相似度估算结合算法,提供负载的快速感知和配置参数的持续优化。实验结果表明,采样效率得到有效提升,预测精度优于现有预测模型,满足电力数据采集系统的运行要求,且寻优耗时也低于现有的黑盒参数调优方法。
国家重点研发计划资助项目(2018YFC0407105);华能集团总部科技项目(HNKJ19-H12);国网新源科技项目(SGTYHT/19-JS-217)。
屠子健(1998—),男,硕士研究生,主要研究方向:分布式计算和控制系统。E-mail:tuzijian_1018@163.com
毛莺池(1976—),女,通信作者,博士,教授,博士生导师,主要研究方向:分布式计算、并行处理、分布式数据管理、移动传感系统和物联网。E-mail:yingchimao@hhu.edu.cn
吴明波(1975—),男,硕士,主要研究方向:电力生产数据分析、电力生产设备状态评价与分析、智慧电厂研究与实践。E-mail:13888197035@163.com
1.School of Computer and Information, Hohai University, Nanjing 211100, China;2.Huaneng Lancang River Hydropower Co., Ltd., Kunming 650214, China
Ceph system has been widely applied in the storage of electric power data. The existing data center relies on the human experience to optimize the storage configuration parameters. However, the human experience is unable to adapt to the dynamic changes of the power grid, and the accuracy is insufficient. Therefore, the rapid storage and processing demands at the edge of the power grid cannot be satisfied. Targeting at the problems above, the data load-aware adaptive configuration parameter recommendation (DACR) method based on reinforcement learning is proposed in the electric power data storage system. This method allocates the parameter sample set and relevant cluster performance by Markov chain Monte Carlo sampling algorithm based on reinforcement learning, and builds the performance prediction model by the hierarchy modeling method. Also, the fast perception of load and continuous optimization of configuration parameters are provided by adopting the combined algorithm of cluster performance cost function and load similarity estimation. The experiment results show that the sampling efficiency is effectively improved and the prediction accuracy is better than that of existing prediction models. This method can satisfy the operation requirements of power data acquisition systems and its optimization time is shorter than that of the existing black box parameter tuning method.
[1] | 屠子健,毛莺池,吴明波,等.基于强化学习的电力数据存储系统参数自适应调优[J].电力系统自动化,2022,46(4):112-122. DOI:10.7500/AEPS20210311001. TU Zijian, MAO Yingchi, WU Mingbo, et al. Reinforcement Learning Based Parameter Adaptive Tuning for Electric Power Data Storage System[J]. Automation of Electric Power Systems, 2022, 46(4):112-122. DOI:10.7500/AEPS20210311001. |