计及算力与储冷灵活性的数据中心多能微网运行优化OA
Operational optimization of multi-energy microgrids for data centers considering flexibility of computing loads and cold energy storage
针对园区风光出力与数据中心连续算力-制冷需求之间的逐时时序错位,聚焦月最大需量约束下可延迟算力负荷与显热水储冷系统的互补调节机理,构建了包含园区风电、光伏、电网、数据中心IT负荷、冷却系统和显热水储冷系统的数据中心多能微网模型.算力侧将聚合计算需求拆分为不可转移部分与可转移部分到达量,并在服务时窗、逐批任务守恒、月末队列清零和中央处理器利用率上下限约束下选择执行时刻;制冷侧根据IT热负荷建立冷却液分配单元、设施水泵、计算机房空气处理机组和冷水机功率模型,以储冷量状态方程、充放冷互斥、容量及期末闭合约束描述储冷系统;电网侧显式考虑逐时购售电互斥、上网容量、蒙西分时电价和月最大需量电费.采用三级词典序目标依次最小化弃电量、净电费和无效调节量.算例以呼和浩特典型5月气象序列驱动风光出力与干球温度自由冷却模型,利用露点温度、湿球温度校核凝露边界;以Google 2019生产集群轨迹和Dell PowerEdge R740 SPECpower实测曲线构造1 MW等值IT负荷.主场景取20%柔性算力任务份额、4 h最大延迟及2h额定储冷时长的储冷系统(额定充放冷功率651.77 kW、容量1 303.53 kW·h),并设置基线、仅算力负荷时移/储冷、规则控制、完美预测联合优化和滚动优化对比场景.结果表明:相对基线场景,仅算力负荷时移使月最大需量和净电费分别降低15.33%和11.38%,仅储冷分别降低9.55%和8.95%,联合优化分别下降17.83%和16.13%,本地风光自消纳率提高1.66百分点.Shapley分解表明,算力负荷时移与储冷在联合削峰量中分别贡献66.2%和33.8%,在净电费节省中分别贡献57.6%和42.4%.交互效应分析表明,上网容量与售电价格共同变化时,其交互影响明显强于柔性算力任务比例与额定储冷时长或最大延迟共同变化所产生的影响.算力负荷时移主要承担跨小时能量重排,储冷主要提供峰值附近的短时冷量补偿,二者在统一的任务、热量、电功率和月需量约束下协同,能够同时改善本地风光消纳、削减月最大需量并降低净电费.
To address the hourly temporal mismatch between wind and photovoltaic(PV)power output in industrial parks and the continuous computing and cooling demand of a data center,the complementary regulation mechanism of deferrable computing loads and a sensible-heat water-based cold energy storage system under monthly maximum demand constraints was investigated.A data center multi-energy microgrid model incorporating park-level wind and PV generation,the power grid,data center IT loads,a cooling system,and a sensible-heat water-based cold energy storage system was developed.On the computing side,the aggregate computing demand was divided into a nonshiftable component and arrivals of a shiftable component,and execution times were selected subject to service-window constraints,batch-wise task-conservation constraints,end-of-month queue-clearance constraints,and lower and upper limits on central processing unit utilization.On the cooling side,power models for the coolant distribution unit,facility water pump,computer-room air-handling unit,and chiller were established based on the IT heating load.Cold energy storage system was described using a state equation for cold energy storage,mutually exclusive charging and discharging constraints,capacity constraints,and end-of-period closure constraints.On the grid side,hourly mutual exclusivity between electricity purchase and sale,grid-export capacity,the time-of-use tariff for western Inner Mongolia,and the monthly peak-demand charge were explicitly considered.A three-level lexicographic objective was used to sequentially minimize electricity curtailment,net electricity cost,and ineffective adjustment.The case study used a meteorological series in typical May for Hohhot to model wind and PV output and dry-bulb-temperature-based free cooling.Dew-point and wet-bulb temperatures were used to verify the condensation boundary.The Google 2019 production-cluster trace and the measured Dell PowerEdge R740 SPECpower curve were used to construct a 1 MW equivalent IT load.The main scenario assumed a 20%share of flexible computing tasks,a maximum delay of 4 h,and a cold energy storage system with a rated duration of 2 h(rated charging and discharging cooling power of 651.77 kW and capacity of 1 303.53 kW·h).Baseline,computing-load-shifting-only,cold-energy-storage-only,rule-based-control,perfect-prediction joint optimization,and rolling-optimization scenarios were also established for comparison.The results showed that compared with the baseline scenario,computing-load shifting alone reduced monthly peak demand and net electricity cost by 15.33%and 11.38%,respectively,and cold energy storage alone reduced these values by 9.55%and 8.95%,respectively.Joint optimization reduced them by 17.83%and 16.13%,respectively,while increasing the local self-consumption rate of wind and PV power by 1.66 percentage points.Shapley decomposition showed that computing-load shifting and cold energy storage contributed 66.2%and 33.8%,respectively,to joint peak reduction,and 57.6%and 42.4%,respectively,to net electricity cost savings.Interaction effect analysis showed that the joint variations in grid-export capacity and electricity selling price had a significantly stronger interactive effect than the combined variations in the flexible computing task share and either the rated cold energy storage duration or the maximum delay.Computing-load shifting mainly performs inter-hour energy allocation,whereas cold energy storage mainly provides short-term cooling compensation around peak periods.Under the unified constraints of computing tasks,thermal energy,electric power,and monthly demand,the two systems can operate synergistically to simultaneously improve local consumption of wind and PV power,reduce monthly maximum demand,and lower net electricity cost.
杜宇昊;严儒井;李孟兰
贵州大学 电气工程学院,贵阳 550025贵州大学 电气工程学院,贵阳 550025贵州江源电力建设有限公司,贵阳 550002
能源科技
数据中心需求响应算力负荷时移储冷可再生能源消纳多能微网算电协同
data centerdemand responsecomputing-load shiftingcold energy storagerenewable energy consumptionmulti-energy microgridcomputing-electricity coordination
《综合智慧能源》 2026 (8)
117-129,13
贵州省科技支撑计划项目(黔科合支撑[2025]一般021)国家自然科学基金项目(52406227)Guizhou Provincial Science and Technology Support Program(Qiankehe-Support[2025]General 021)National Natural Science Foundation of China(52406227)
评论