首页|期刊导航|综合智慧能源|考虑数据中心柔性任务的综合能源系统能-算协同深度强化学习方法

考虑数据中心柔性任务的综合能源系统能-算协同深度强化学习方法OA

Deep reinforcement learning method for energy-computing coordinated scheduling in integrated energy systems considering flexible data center tasks

中文摘要英文摘要

随着数字经济、人工智能和云计算快速发展,数据中心(DC)用电规模持续增长,如何挖掘其调节潜力并实现与综合能源系统(IES)的协同运行已成为提升能源利用效率和运行经济性的重要问题.为此,建立包含上级电网、光伏、热电联产机组、燃气锅炉、电锅炉、电储能及DC的联合动态经济调度模型.首先,考虑延时敏感型和延时容忍型的任务差异,构建基于先入先出原则的柔性任务队列模型,通过任务到达量、队列积压量、到期任务量和队列紧急度描述柔性任务的时序演化过程,并结合活动服务器数量、服务器利用率、平均响应时间及动态电压频率调节特性,建立任务执行量与DC用电功率之间的映射关系.其次,将综合能源系统-数据中心(IES-DC)联合调度问题构建为马尔可夫决策过程,采用深度确定性策略梯度(DDPG)算法学习能源设备运行与柔性任务处理的联合调度策略.最后,在24 h典型日场景下开展算例验证,并采用软演员-评论家(SAC)算法、近端策略优化(PPO)算法和粒子群优化(PSO)算法作为对比算法.算例结果表明,所提方法能够在保证电热功率平衡、电储能荷电状态恢复、柔性任务全部完成且无超期任务的条件下,引导部分延时容忍型任务由峰电价时段向平、谷电价时段转移.与全部任务到达后立即执行的场景相比,DDPG柔性调度使上级电网购电成本降低2.17%;在相同系统模型和固定测试场景下,SAC和PPO的日运行成本分别较DDPG高1.50%和12.79%,DDPG日运行成本仅较PSO高0.50%,但其训练完成后的24 h调度计算速度约为PSO的21 800倍.结果表明,DC柔性任务可作为IES的可调负荷参与能-算协同调度,DDPG在算例中能够兼顾任务时延约束与系统运行经济性,并表现出较好的综合调度性能.

With the rapid development of the digital economy,artificial intelligence,and cloud computing,the electricity consumption of data centers continues to increase.Exploiting their regulation potential and achieving coordinated operation with integrated energy systems have become important challenges for improving energy utilization efficiency and operational economy.To address these issues,a joint dynamic economic scheduling model was established incorporating the utility grid,photovoltaic generation,combined heat and power units,gas boilers,electric boilers,battery energy storage,and data centers.Considering the differences between delay-sensitive and delay-tolerant tasks,a flexible task queue model based on the first-in-first-out(FIFO)principle was developed.The temporal evolution of flexible tasks was characterized by task arrivals,queue backlog,due tasks,and queue urgency.By further incorporating the number of active servers,server utilization,average response time,and dynamic voltage and frequency scaling(DVFS)characteristics,a mapping relationship between task execution and data center power consumption was established.The IES-DC joint scheduling problem was formulated as a Markov decision process,and the deep deterministic policy gradient(DDPG)algorithm was employed to learn a joint scheduling policy for energy equipment operation and flexible task processing.Case studies were conducted under a 24 h typical-day scenario,with SAC,PPO,and particle swarm optimization(PSO)adopted as comparison methods.The case study results showed that the proposed method could shift part of the delay-tolerant tasks from peak-price periods to flat-and valley-price periods while maintaining electric and thermal power balance,restoring the state of charge(SOC)of the energy storage system,completing all flexible tasks,and avoiding overdue tasks.Compared with the scenario in which all tasks were executed immediately upon arrival,DDPG-based flexible scheduling reduced the electricity purchasing cost from the utility grid by 2.17%.Under the same system model and fixed test scenario,the daily operating costs of SAC and PPO were 1.50%and 12.79%higher than that of DDPG,respectively.The daily operating cost of DDPG was only 0.50%higher than that of PSO,but the computation time for 24 h scheduling after training was only about 1/21 800 of that required by PSO.The findings demonstrate that flexible data center tasks can participate in energy-computing coordinated scheduling as adjustable loads of the integrated energy system,and that DDPG can effectively balance task delay constraints and system operating economy while exhibiting good overall scheduling performance in the case studies.

曲鑫梅;钟俊杰;赵津津;孙康鸿;张振;曹一家

长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114长沙理工大学电气与信息工程学院,长沙 410114||长沙理工大学电网防灾减灾全国重点实验室,长沙 410114

能源科技

算电协同综合能源系统数据中心柔性任务动态经济调度深度强化学习深度确定性策略梯度

computing-electricity coordinationintegrated energy systemdata centerflexible taskdynamic economic dispatchdeep reinforcement learningdeep deterministic policy gradient

《综合智慧能源》 2026 (9)

56-67,12

国家自然科学基金项目(52407079)湖南省自然科学基金项目(2025JJ60372) National Natural Science Foundation of China(52407079)Natural Science Foundation of Hunan Province of China(2025JJ60372)

10.3969/j.issn.2097-0706.2026.09.005

评论