Data center two-stage stochastic optimization method and system considering wind and light uncertainty

Abstract

The invention relates to the technical field of data center optimization, and particularly discloses a data center two-stage stochastic optimization method and system considering wind and light uncertainty, and the method comprises the steps: carrying out the modeling of a data center energy supply system, generating a wind and light output sample based on Monte Carlo simulation, enabling the generated random sample to meet wind power and photovoltaic output complementation through Corisky decomposition; introducing a first-order autoregression model, generating random scenes in combination with multivariate normal distribution, and obtaining a wind and light output curve in each scene; a two-stage stochastic optimization model containing computing power scheduling and multi-energy coordination is constructed by taking the minimum expected operation cost of a system as a target, the model is solved, and an optimal time sequence operation strategy of a computing power task is decided. According to the two-stage random optimization, the adjustment cost caused by prediction errors is covered with low risk premium, and effective support is provided for reliable operation of the data center under new energy output fluctuation and computing power load time sequence mismatch.

Description

Technical Field

[0001] This invention relates to the field of data center optimization technology, and in particular to a two-stage stochastic optimization method and system for data centers that takes into account wind and solar uncertainties. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As the core carrier of cloud computing technology, data centers are experiencing exponential growth in construction scale, and their enormous electricity demand has become a key bottleneck restricting the development of the energy and information industries. While introducing distributed wind and solar renewable energy to build integrated source-load-storage data centers is an effective path to achieve energy and information integration, the strong randomness of renewable energy output and the rigid characteristics of traditional data center operation modes create a significant spatiotemporal mismatch, leading to severe wind and solar curtailment and high operating costs. Therefore, exploring the internal flexibility of data centers and building a proactive response mechanism for source-load coordination are critical scientific issues that urgently need to be addressed.

[0004] Current research primarily focuses on two dimensions: complementary energy sources on the power supply side and flexibility on the load side. On the power supply side, existing literature emphasizes equipment-level physical flexibility and multi-type energy coupling planning. While integrated energy system planning and combined cooling, heating, and power (CCHP) architectures for data centers have improved energy efficiency through "heat-driven power generation" or CHP decoupling, and introduced technologies such as ice storage or hydrogen energy storage to mitigate fluctuations using peak-valley electricity price differences and long-term energy storage, these studies often treat data centers as given power boundaries or simple interruptible loads. This approach ignores the proactive adjustment potential of data centers as cyber-physical systems. Relying solely on source-side redundancy to address supply-demand mismatches makes it difficult to balance system economy and scheduling flexibility.

[0005] On the load side, data center loads exhibit good spatiotemporal schedulability. Spatially, load allocation strategies across different geographical locations leverage regional differences in electricity prices, temperature, and renewable energy output to guide load tasks to areas with lower energy costs. However, this strategy heavily relies on wide-area, high-bandwidth networks and faces challenges in coordinating supply and demand balance within individual data centers. Temporally, while coordinated scheduling based on integrated demand response and workload shifting can absorb renewable energy by delaying batch processing tasks, it still suffers from the following shortcomings in modeling computing load mechanisms and handling uncertainties related to renewable energy: (1) The granularity of computing load modeling is too coarse, lacking a characterization of load task backlog and transfer. Existing literature mostly adopts macroscopic power balance models, simplifying computing load into scalable power blocks or linear curves, ignoring the mechanism modeling of some computing tasks as flexible resources. In fact, the accumulation, transfer and processing of load tasks involves a process involving the coupling of arrival rate, service rate, queue length and latency. Although some studies have introduced task dependency models, their computational complexity is too high, making it difficult to directly adapt to long-term microgrid scheduling frameworks. Furthermore, due to the inability to accurately describe task backlog, scheduling strategies often oscillate between overly conservative and default risk.

[0006] (2) The generation of scenarios with uncertainties in new energy sources fails to accurately reflect the coupling characteristics of wind and solar power output. Although robust optimization and partial robust optimization can ensure the safety of the system, their decision-making logic based on extreme scenarios will lead to redundant equipment configuration. Although stochastic programming seeks a balance between economy and reliability, existing studies mostly assume that the prediction error follows an independent and identically distributed pattern, ignoring the time autocorrelation and complementarity of renewable energy output. Deviating from the above correlations will cause the set of scenarios used to deviate significantly from the actual operating conditions, which in turn leads to a decrease in the stability of the scheduling strategy when facing continuous extreme operating conditions. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a two-stage stochastic optimization method and system for data centers that considers the uncertainties of wind and solar power output. Targeting the stochastic fluctuations and temporal coupling characteristics of wind and solar power output, a scenario generation model based on first-order autoregression and Kolesky decomposition is constructed to accurately capture the temporal correlation and complementarity of wind and solar power. Considering the differences in load latency tolerance, a flexible computing power response model based on discrete-time task flows is established, quantifying the migration capability and backlog constraints of services with different latency tolerances in the time dimension through queue state equations. With the goal of minimizing the expected operating cost of the system, a two-stage stochastic optimization model incorporating computing power scheduling and multi-energy coordination is constructed to determine the optimal temporal operation strategy for computing power tasks.

[0008] In some implementations, the following technical solutions are adopted: A two-stage stochastic optimization method for data centers that takes into account wind and solar uncertainties includes: Modeling of data center power supply system, taking into account the differences in load latency tolerance, quantifying the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete time task flow, and introducing cross-time balance constraints of computing power tasks to ensure that all tasks are processed before the deadline, thus constructing a refined load response model. Based on Monte Carlo simulation, samples of wind and solar power output are generated. Koleski decomposition is used to ensure that the generated random samples satisfy the complementarity of wind and solar power output. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenarios, and wind and solar power output curves under each scenario are obtained. With the goal of minimizing the expected operating cost of the system, a two-stage stochastic optimization model including computing power scheduling and multi-energy coordination is constructed. The model is solved to determine the optimal time-series operation strategy for computing power tasks.

[0009] As a further solution, the queue state evolution equation based on discrete-time task flow satisfies: current t+1 The backlog of the load task queue at a given time is equal to the backlog of the load task queue at the previous time. t Time of the first k The sum of arrivals of each type of task, and t Time of the first k The difference in processing volume between similar tasks; At any given time, the backlog of each type of load task queue is not less than 0, and the backlog of the load task queue at the beginning and end of each type of load task is equal to 0.

[0010] As a further solution, a cross-time balance constraint for computing power tasks is introduced to ensure that all tasks are processed before the deadline, specifically: For any type of computing task that arrives at any time, the system must process them all before their maximum allowable latency expires; however, the processing time cannot exceed the total scheduling cycle.

[0011] As a further approach, the generated random samples are made to satisfy the complementarity of wind power and photovoltaic output through Kolesky decomposition, specifically: Assumption t The prediction errors for wind power and solar power at any given time follow a multivariate normal distribution, where the covariance matrix is... ∑ The correlation between wind and solar power output was described: ;in, The coupling coefficient for wind and solar power output; For matrix ∑ Perform Koleski decomposition to obtain the lower triangular matrix. L and satisfy L The product of its transpose is equal to ∑ ; Based on the lower triangular matrix L Generate correlation coefficients ρ spatial random disturbance term .

[0012] As a further approach, an error sequence with time autocorrelation is generated based on a first-order autoregressive model: ; in, σ i The standard deviation of the prediction error; ρ tem This is the time autocorrelation coefficient; In scenario s, the first i Energy-like energy in the present t The prediction error value at time; In scenario s, the first i The prediction error value of the energy type at the previous time step; s represents the index of the random scene. t Represents the time step. i Represents specific types of renewable energy. PV represents photovoltaic power, and WT represents wind power.

[0013] As a further approach, after generating random scenes, a fast forward selection method based on Wasserstein distance is used to reduce the number of scenes. This is achieved by quantifying and minimizing the probability distance between the reduced set and the original set, thus selecting the scenes that are not yet complete. N s A typical scenario and its reconstruction probability π s .

[0014] As a further solution, a two-stage stochastic optimization model is constructed, which includes computing power scheduling and multi-energy coordination. Specifically, the two-stage stochastic optimization model is constructed with the goal of minimizing the expected value of the overall operating cost of the system under the entire scenario set. In the first stage, the scheduling plan of computing power tasks is used as the decision variable. In the second stage, the stochasticity of wind and solar power output scenarios is adjusted in real time. The decision variables include gas turbine output, energy storage system charging and discharging status, and interaction power with the public power grid. The load task scheduling decision in the first stage affects the energy scheduling cost in the second stage by reshaping the computing power load task demand.

[0015] In other embodiments, the following technical solutions are adopted: A two-stage stochastic optimization system for data centers that takes into account wind and solar uncertainties includes: The system modeling module is used to model the data center power supply system. It considers the differences in load latency tolerance, quantifies the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete time task flow, and introduces the cross-time balance constraint of computing power tasks to ensure that all tasks are processed before the deadline, thus constructing a refined load response model. The scene generation module is used to generate wind and solar power output samples based on Monte Carlo simulation. The generated random samples are made to satisfy the complementarity of wind and solar power output through Koleski decomposition. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenes, and the wind and solar power output curves under each scene are obtained. The system optimization module is used to construct a two-stage stochastic optimization model that includes computing power scheduling and multi-energy coordination with the goal of minimizing the expected operating cost of the system. The module solves the model and determines the optimal timing strategy for computing power tasks.

[0016] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions adapted to be loaded and executed by the processor as described above for a two-stage stochastic optimization method for data centers taking into account wind and solar uncertainties.

[0017] In other embodiments, the following technical solutions are adopted: A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described two-stage stochastic optimization method for data centers taking into account wind and solar uncertainties.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the stochastic fluctuations and temporal coupling characteristics of wind and solar power output by constructing a scene generation model based on first-order autoregression and Kolesky decomposition, accurately capturing the temporal correlation and complementarity of wind and solar power. A discrete-time task flow model established for data center load quantifies the temporal elasticity of different computing tasks, effectively guiding data centers to shift from passive energy consumption to an active response mode where load follows energy source changes. Regarding the uncertainty of wind and solar power output, the two-stage stochastic optimization in this embodiment covers the adjustment costs caused by prediction errors with a low risk premium, providing effective support for the reliable operation of data centers under the conditions of fluctuating renewable energy output and temporal mismatch between computing load and renewable energy.

[0019] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] Figure 1 This is a flowchart of a two-stage stochastic optimization method for data centers that takes into account the uncertainties of wind and solar power in an embodiment of the present invention. Figure 2 This is a schematic diagram of the data center power supply system architecture in an embodiment of the present invention; Figure 3 This is a schematic diagram of the probability distribution of photovoltaic prediction errors in an embodiment of the present invention; Figure 4 This is a schematic diagram of the probability distribution of wind power prediction errors in an embodiment of the present invention; Figure 5 This is a schematic diagram of the joint probability distribution of wind and solar prediction errors in an embodiment of the present invention; Figure 6 This is a set of photovoltaic power output scenarios in the embodiments of the present invention; Figure 7 This is a set of wind power output scenarios in the embodiments of the present invention; Figure 8 This is a schematic diagram illustrating the changes in supply and demand response characteristics of the computational task in this embodiment of the invention; Figure 9 This is a schematic diagram illustrating the changes in queue backlog status in an embodiment of the present invention; Figure 10 This is a schematic diagram comparing the data center computing load operation characteristics before and after optimization in an embodiment of the present invention; Figure 11 This is a schematic diagram of power supply and demand balance in an embodiment of the present invention; Figure 12 This is a schematic diagram of cold energy supply and demand balance in an embodiment of the present invention; Figure 13 This is a schematic diagram of the state of charge of the energy storage device in an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Example 1 In one or more embodiments, a two-stage stochastic optimization method for data centers that takes into account wind and solar uncertainties is disclosed, combined with Figure 1 Specifically, it includes the following process: S101: Model the data center power supply system, taking into account the differences in load latency tolerance. Quantify the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete-time task flow. At the same time, introduce the cross-time balance constraint of computing power tasks to ensure that all tasks are processed before the deadline, and construct a refined load response model.

[0024] Specifically, the data center power supply system constructed in this embodiment is composed of coupled energy flow and information flow, mainly including three parts: energy supply layer, energy conversion and storage layer, and load and information layer. Its system structure is as follows: Figure 2 As shown. The energy supply layer includes the main grid interconnection lines, distributed photovoltaic and wind turbines; the energy conversion and storage layer includes gas turbines, electrochemical energy storage and ice storage refrigeration systems; the load side covers information technology (IT) equipment and cooling equipment.

[0025] The system in this embodiment aims to reduce the operating cost of the data center while meeting the computing load tasks through source-load-storage coordinated scheduling.

[0026] The specific process of modeling the flexibility of computing load is as follows: Data centers host tens of thousands of heterogeneous tasks, which exhibit significant differences in time. For example, online services such as web browsing and online transactions have extremely high real-time requirements and cannot be delayed; while batch processing tasks such as data backup and model training have a higher tolerance for latency. This flexibility in the timeline gives computing load an adjustment capability similar to "virtual energy storage".

[0027] To accurately characterize this feature, this embodiment introduces a queue state model based on discrete-time task flow and constructs a refined load response model.

[0028] The computing load is divided into 4 categories, denoted as [4 categories]. κ .definition t Time of the first k The arrival count of the task type is Processing capacity is The backlog of the load task queue is The queue state evolution equation is: (1) (2) (3) in, For the scheduling period, Represents a set of tasks. for t Time of the firstk The number of arrivals for each type of task. For decision variables, it represents t Time of the first k The processing volume of this type of task; As a system state variable, representing t Constant load task queue backlog , These represent the backlog of the load task queue at the initial and final times, respectively. That is, the current... t+1 The backlog at any given time depends on the backlog at the previous time step and the number of newly arrived tasks. and current processing volume .

[0029] (4) In the formula: D k,max For each load type, the maximum delay time is given at any given time. t All tasks arriving up to that moment must be completed. t+D k,max Processing must be completed before the specified time.

[0030] The above formula means that for any type of computing task that arrives at any time, the system must process them all before their maximum allowable latency expires, but the processing time cannot exceed the total scheduling cycle.

[0031] Indicates in The number of arrivals of the k-th type of task at time step. Indicates in The processing volume of the k-th type of task at time step. Represents the maximum delay time for the k-th type of load; for any time t, all tasks arriving up to that time must... Processing must be completed before the specified time. and These represent the time set of the entire scheduling cycle and the category set of all computing power tasks, respectively. It is a boundary protection mechanism that ensures that the deadline of a task will not exceed the total physical scheduling period T (e.g., 24 hours).

[0032] Unlike traditional average latency evaluation metrics, this embodiment introduces a cross-time period balance constraint for computing power tasks (Formula (4)) to ensure that all tasks are processed before the deadline, which is the core guarantee for the feasibility of the physical model.

[0033] Data centers consist of IT equipment and auxiliary equipment, the latter including cooling and lighting, among other things. IT equipment power consumption... P IT,tThis is related to server utilization, the number of servers, and other data devices. The IT equipment power consumption model is shown below: (5) In the formula: N server The number of servers in the data center; P base and P peak These represent the standby power consumption and peak power consumption of a single server, respectively; ω k and u cpu,t These are computing load and CPU utilization of data center servers, respectively. The vast majority of the electrical energy consumed by IT equipment is ultimately converted into heat and dissipated into the server room environment. In order to maintain the temperature of the data center, the cooling system must remove this heat in a timely manner.

[0034] Therefore, the heat load of the data center cooling system H DC,t Directly coupled with the power consumption of IT equipment: (6) in, This represents the total power consumption of all IT devices in the data center at time t. This represents the electro-thermal conversion coefficient of IT equipment. Since the vast majority of the electrical energy consumed by IT equipment during operation is dissipated as waste heat, this coefficient reflects the proportion of electrical energy converted into heat energy. In practical engineering calculations, it is usually approximated as... (That is, the heat generation power is equal to the power consumption). and These represent the lower and upper limits of CPU utilization for data center servers, respectively. In actual scheduling, to ensure system stability and reserve computing power redundancy, CPU utilization must be limited to a reasonable range to avoid server overload, crashes, or low energy efficiency due to idleness.

[0035] The modeling process for other key equipment is as follows: (1) Gas turbine model: The gas turbine, as the core controllable equipment of the system, generates electricity by burning natural gas. Simultaneously, to improve energy efficiency, the high-temperature flue gas it produces is converted into heat energy through a waste heat recovery device to drive an absorption chiller. The operating constraints are as follows: (7) In the formula: P ICE,t Power generation capacity; F gas,t This refers to natural gas consumption. L It is the lower heating value of natural gas; η ICE,e For power generation efficiency; H rec,t For the effective heat to be recovered; γ he The heat-to-electricity ratio of the gas turbine; η whr This refers to the waste heat recovery efficiency.

[0036] In addition, the operation of gas turbines is subject to constraints on upper and lower output limits and gradeability: (8) In the formula: and These represent the lower limit of safe operation and the upper limit of rated output of the gas turbine power generation, respectively. This represents the maximum uphill rate of the gas turbine, which is the maximum difference in power generation that the unit can be allowed to increase between two adjacent scheduling periods. This represents the maximum downhill ramp rate of the gas turbine, which is the maximum difference in power generation that the unit can be allowed to reduce between two adjacent scheduling periods.

[0037] (2) Cooling system model: To meet the enormous heat dissipation demands of the data center, the system employs a synergistic cooling strategy that combines waste heat cooling with electric cooling. The cooling load is shared by high-efficiency electric chillers and absorption chillers utilizing waste heat. (9) In the formula: P EC,t Electricity consumption for cooling; Q EC,t and Q ABS,t These are the cooling capacities of electric refrigeration and absorption refrigeration, respectively. C EC and C ABS These are the coefficients of performance (COPs) of electric chillers and absorption chillers, respectively. Absorption chillers utilize waste heat from gas turbines, resulting in lower energy efficiency but also lower cost. Their cooling capacity... Q ABS,t Limited by the recovery of heat. and These are the minimum cooling capacity lower limit and the maximum rated cooling capacity upper limit of the electric chiller, respectively. Indicates in t At any given time, the waste heat recovery device recovers effective heat power from the gas turbine exhaust.

[0038] (3) Energy storage system model: Data centers are equipped with electrochemical energy storage and ice storage to handle the time shift of electrical and cold energy, respectively. Although they use different storage media, their charging / cold state mechanisms are highly similar, and a general charging / cold state model is established as follows: (10) (11) in, S x,t for t The charged / cooled state at any given moment, η x,ch and η x,dis The charging and discharging efficiency of energy storage systems; P x,ch,t and P x,dis,t These are the charging and discharging power at time t, respectively; C x,cap This refers to the rated capacity of the energy storage. S x,max and S x,min These are the maximum and minimum capacity limits for the energy storage system; 0-1 variables. u x,ch / dis It is a charging and discharging indicator for the energy storage system; S x,0 and S x,T These refer to the charged / cooled states of the energy storage system at the beginning and end of a scheduling cycle, respectively.

[0039] S102: Based on Monte Carlo simulation, samples of wind and solar power output are generated. Koleski decomposition is used to make the generated random samples satisfy the complementarity of wind and solar power output. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenarios, and wind and solar power output curves under each scenario are obtained.

[0040] Specifically, this embodiment generates samples of wind and solar power output based on Monte Carlo simulation, assuming... t The prediction error for wind power and solar power at any given time is... ξ PV,t and ξ WT,t And it follows a multivariate normal distribution. The covariance matrix is... ∑ The correlation between wind and solar power output was described: (12) In the formula, ρ spatial The coupling coefficient for wind and solar power output.

[0041] To ensure that the generated random samples satisfy the complementary characteristics of wind power and photovoltaic output, the matrix... ∑ Perform Koleski decomposition to obtain the lower triangular matrix. L and satisfy L The product of its transpose is equal to ∑ .

[0042] like Z t Let be mutually independent random vectors that follow a standard normal distribution. LZ t That is, a random perturbation term that satisfies spatial correlation.

[0043] Based on the complementary characteristics of wind and solar power output, a first-order autoregressive model (AR(1)) is introduced, and random scenes are generated by combining multivariate normal distribution. Koleski decomposition is used to generate the correlation coefficient. ρ spatial random disturbance term Z s,t Subsequently, an error sequence with time autocorrelation is generated according to the AR(1) process: (13) in, σ i The standard deviation of the prediction error; ρ tem This is the time autocorrelation coefficient; In scenario s, the first i Energy-like energy in the present t The prediction error value at time; In scenario s, the first i The prediction error value of the energy type at the previous moment; s represents the index of the random scenario, which physically means a specific wind and solar power output fluctuation condition that the system may encounter in the future; t Represents the time step, and its physical meaning is a specific moment within the scheduling period (e.g., the tth hour). i Represents specific types of renewable energy. PV represents photovoltaic power, and WT represents wind power.

[0044] Finally, the generated first s The power output curves for each scenario are as follows: (14) In the formula: P pre,i,t It contributes to the initial forecasting of photovoltaic and wind power.

[0045] The large size of the initial scene set significantly increases the computational burden. Therefore, this embodiment employs a fast forward selection method based on Wasserstein distance to achieve scene reduction. This method filters scenes by quantizing and minimizing the probabilistic distance between the reduced set and the original set. N s A typical scenario and its reconstruction probability π s The reduced scene set can effectively cover various working conditions, from extreme weather to normal wind and solar power output, which is the key to ensuring that the subsequent optimization model can cope with the uncertainties of wind and solar power.

[0046] S103: With the goal of minimizing the expected operating cost of the system, construct a two-stage stochastic optimization model that includes computing power scheduling and multi-energy coordination, solve the model, and determine the optimal timing operation strategy for computing power tasks.

[0047] In the two-stage stochastic optimization model constructed in this embodiment, the decision variable in the first stage is the scheduling plan of the computing power tasks. Given that computing power task processing has a specific time window and involves service quality commitments, a unified and fixed scheduling scheme must be provided before uncertainties occur to ensure the continuity and integrity of task execution. The second stage involves real-time correction for the stochasticity of wind and solar power output scenarios. The decision variables include gas turbine output, energy storage system charging and discharging status, and interaction power with the public power grid. These variables represent the optimal response results for specific scenarios, aiming to achieve a dynamic balance between system economy and physical constraints under various operating conditions.

[0048] Specifically, within the two-stage stochastic programming framework, the optimization objective of the data center is not to seek a local optimum under a specific deterministic scenario, but rather to construct a robust load scheduling scheme to address the operational risks arising from random fluctuations in wind and solar power output. Specifically, the load scheduling decisions in the first stage influence the energy scheduling costs in the second stage by reshaping computing load demand. For example, shifting tasks from periods of high electricity prices or wind / solar scarcity to periods of low electricity prices or abundant wind and solar power can significantly reduce operating costs. Therefore, the objective function is expressed as minimizing the expected overall operating cost of the system across the entire set of scenarios: (15) In the formula: F Expected operating costs for the data center; N S The number of typical scenarios; π s For the first s The probability of each wind power and solar power output scenario occurring. For each specific scenario... s Its operating costs C op,sIt consists of three parts: the cost of purchasing electricity for interaction with the external power grid, the fuel consumption cost of the gas turbine, and the operation and maintenance cost of key equipment. (16) (17) In the formula: c grid,t for t The price at which the data center purchases electricity from the external power grid; P grid,t,s for t Moment Scene s The purchased power capacity. For data centers in the s Total electricity purchase cost in a typical scenario. For data centers in the s The total gas purchase cost generated by a gas turbine consuming natural gas in a typical scenario. For data centers in the s The total operation and maintenance cost of all energy conversion and storage devices (such as wind and solar turbines, gas turbines, chillers, and energy storage) in a typical scenario. This indicates the time step for optimized scheduling, which is set to 1 hour. This parameter is used to convert the real-time operating power of each time period into the actual energy consumed, in order to calculate the economic cost.

[0049] (18) In the formula: c gas The price per cubic meter of natural gas. P ICE,t,s For gas turbines in s Scene t Power generation at the current moment; η ICE,e For the power generation efficiency of gas turbines, L This refers to the calorific value of natural gas.

[0050] (19) In the formula: J For the collection of all devices, c om,j For equipment j The unit power operation and maintenance coefficient, P j,t,s This refers to the output of the corresponding equipment. The equipment set... J include c om,PV , c om,WT , c om,ICE ,c om,WHR , c om,ABS , c om,EC , c om,ES and c om,ISS These represent the operation and maintenance cost coefficients for photovoltaic, wind power, gas turbine, waste heat recovery device, absorption refrigeration, electric refrigeration, electric energy storage, and ice storage, respectively.

[0051] The specific constraints of the above two-stage stochastic optimization model are as follows: In the first stage, the scheduling of computing power tasks in the data center must meet the flexibility modeling and processing constraints of Equations (1) to (5) for the computing power load tasks in the data center.

[0052] In the second phase, energy devices in each scenario must meet physical constraints and real-time power balance requirements. Multi-energy flow power balance constraints are fundamental to maintaining stable data center operation and constant temperature. For electrical loads, the sum of power from the source side must equal the sum of power consumed by the load side; for cooling loads, the supplied cooling capacity must meet the needs of the data center and other cooling loads.

[0053] (1) Power balance constraints: (20) In the formula: P grid,t,s Electricity purchased from the power grid for data centers; P ICE,t,s It is the power generation capacity of the gas turbine; P PV,t,s and P WT,t,s These are the actual outputs of photovoltaic and wind power, respectively. P IT,t This refers to the total power consumption of the servers in the data center. For the first s In a typical scenario t Discharge power of the energy storage system at any given time. For the first s In a typical scenario t The charging power of the instantaneous energy storage system. For the first s In a typical scenario t The power consumption of the instantaneous refrigerator.

[0054] (2) Cold energy balance constraint: (twenty one) In the formula: Q ABS,t,sThis refers to the cooling capacity of the absorption chiller. Q ISS,ch,t,s and Q ISS,dis,t,s These are the cold storage capacity and cold release capacity of the ice storage system, respectively. Q EC,t,s The cooling capacity of the electric chiller. H DC,t,s This is to meet the cooling load requirements of the data center.

[0055] (3) Renewable energy output constraints: (twenty two) In the formula: P i,t,s For the actual output of wind or solar power, P pre,i,t,s This represents the maximum output at time t in scenario s.

[0056] (4) Output constraints of electric chillers: (twenty three) In the formula: Q EC,min and Q EC,min These are the lower and upper limits of electric refrigeration.

[0057] (5) Energy storage constraints: The constraints on the state of charge / cold and the operating charge / discharge power and capacity of the electric energy storage and ice storage system are shown in Equations (10) to (11).

[0058] The following is a computational example analysis of the method in this embodiment, and the computing power system built is as follows: 24-hour day-ahead optimization scheduling is implemented using a typical data center as an example, with a time step of 1 hour. The system configuration includes a 3.3MW photovoltaic array, a 6.6MW wind turbine, and a 6MW gas turbine. The grid electricity price follows a time-of-use pricing mechanism, with high-price periods of 1.2 yuan / kWh from 10:00-15:00 and 19:00-21:00, and low-price periods of 0.3 yuan / kWh for the remaining time periods. The computing load is divided into four categories based on latency sensitivity. The first category consists of rigid data center tasks with a maximum latency... D k,max The first category is 0 hours; the second to fourth categories are flexible tasks. D k,max The simulation durations were 2 hours, 4 hours, and 6 hours, respectively. Other relevant parameters of the model are shown in Table 1. The simulation platform used was MATLAB 2023b, which called the CPLEX 12.10 solver.

[0059] Table 1. Parameters related to the simulation system parameter numerical values Gas turbine power generation efficiency 0.44 Gas turbine heat production efficiency 0.47 Waste heat recovery efficiency 0.9 Energy efficiency ratio of electric chillers 3.5 Energy efficiency ratio of absorption chiller 0.7 Energy storage charging / discharging efficiency 0.95 Ice storage cold charging efficiency 0.67 Ice storage cold discharge efficiency 0.9 Standard deviation of photovoltaic prediction error 0.15 Standard deviation of wind power forecast error 0.20 Correlation coefficient between wind and solar power output -0.3 The random scene set generated by the first-order autoregressive model—the Kolesky decomposition method—effectively reproduces the coupling characteristics of wind and solar power output. For example... Figure 3 and Figure 4 The photovoltaic and wind power prediction errors shown in the figure generate error sequences that highly fit the normal distribution in terms of probability density, verifying the statistical significance of the probability model. Figure 5 The correlation between wind and solar power output was visually demonstrated using the confidence ellipse method. In the figure, the gray-filled area and the solid boundary represent the 95% and 68% confidence intervals of the initial scene set, respectively. The ellipse tilts downwards to the right, intuitively reflecting the negative correlation between wind and solar power output. The calculated correlation coefficient is -0.305, which highly matches the theoretical value, accurately reproducing the complementary mechanism of wind and solar power output. The 24-hour error data (red dots) of the 30 typical scenes after reduction not only cover the high-probability-density area inside the ellipse but also effectively preserve extreme scenes distributed at the confidence region boundaries and even outside.

[0060] 30 typical scenarios processed by the fast forward selection algorithm, such as Figure 6 and Figure 7 As shown, after reduction, the correlation coefficient of the scene set drifted to -0.421. This is because the reduction algorithm retains boundary scenes farther from the center in the probability space, which often correspond to extreme complementary conditions of strong light and weak wind or strong wind and weak light. This statistical enhancement makes the input scene set include a more severe risk of resource mismatch, forcing the subsequent stochastic programming model to reserve more adjustment capacity, effectively improving the robustness of the scheduling strategy. The reduced scene set not only fully covers the fluctuation envelope of the prediction curve, but also retains extreme windless and lightless conditions, improving the effectiveness of the subsequent robust optimization strategy under boundary conditions.

[0061] Figure 8 and Figure 9 The presentation demonstrated the scheduling of Category 4 loads in a data center under stochastic optimization. Category 4 tasks and loads with a maximum delay of 6 hours exhibited strong characteristics of high-price load accumulation and low-price load release. Specifically, during the peak electricity price period from 10:00 to 12:00, the system proactively reduced the number of load tasks to be processed. At this time, tasks entered a virtual buffer queue, resulting in a significant increase in the backlog of tasks. After 13:00, when photovoltaic output increased and electricity prices dropped, the system quickly released the backlog of tasks, realizing the transfer of loads along the time axis.

[0062] Due to the influence of load scheduling strategies, the data center computing power load energy consumption curve has been restructured. For example... Figure 10As shown, compared to the initial computing load distribution, the optimized computing load exhibits a significant "reverse electricity price" distribution. During peak hours, server utilization is reduced from 63% to around 37%, while it rebounds to 90% during nighttime periods with abundant wind power. This restructuring alleviates peak power supply pressure on the grid while significantly enhancing the impact of renewable energy consumption on reducing data center operating costs.

[0063] This paper analyzes the multi-energy coupling mechanism of a data center (electricity, cooling, and gas) under a typical scenario where solar power is normal and wind power generation is high at night. The power balance is as follows: Figure 11-13 As shown, the system operation exhibits distinct time-segmented collaborative characteristics, with 10:00-15:00 being the peak daytime photovoltaic (PV) generation period. During this time, the computing load and the PV curve remain highly synchronized, resulting in strong PV output (…). Figure 11 The green zone enables the system to prioritize the supply of photovoltaic power to IT equipment and high-efficiency electric chillers, while the gas turbine maintains low-power operation to supplement part of the power load and cooling load, realizing the active response of load following source.

[0064] During the peak electricity price period from 7:00 PM to 9:00 PM, facing the triple constraints of no solar power, weak wind power, and high electricity prices, the output of gas turbines increases. As the core power and cooling support for data centers, their high-temperature flue gas undertakes about 50% of the cooling load through absorption chillers. Figure 12 (Yellow area). This cascade utilization model based on combined cooling, heating and power (CCHP) effectively decouples cooling from electricity-driven constraints, minimizing the cost of purchasing electricity from the grid.

[0065] During the periods of 0:00–6:00 and 23:00–24:00, abundant wind power resources drive the system to activate a dual mode of valley filling using computing power and energy storage charging. For example... Figure 13 As shown, the electrochemical energy storage and ice storage device operates at full power to handle backlogged loads, converting excess wind power into chemical energy and cooling energy, thus achieving proactive response to wind and solar power fluctuations.

[0066] To quantify the value of the time flexibility of computing power load and the economic cost of the two-stage stochastic optimization strategy in dealing with wind and solar uncertainties, this embodiment constructs three comparative scenarios for empirical analysis: Scenario 1, a rigid scenario that does not consider wind and solar output uncertainty and ignores load flexibility; Scenario 2, a deterministic optimization scenario that considers load flexibility but ignores wind and solar uncertainties; and Scenario 3, a two-stage stochastic optimization scenario that takes into account both wind and solar uncertainties and load flexibility.

[0067] Table 2 presents the optimization results for three scenarios. Based on the deterministic perspective of Scenario 1, the time elasticity of computing load significantly improves the economics of data centers. Comparing the simulation results of Scenario 1 and Scenario 2, after introducing flexible scheduling of computing load, the operating cost of the data center decreased from RMB 37,628.16 to RMB 35,746.93, a reduction of approximately 5.0%. This economic benefit mainly stems from the flexible response mechanism of the computing load. The system directly avoids the proportion of load during high-price periods and reduces gas turbine fuel consumption by shifting delay-tolerant loads from high-price periods (RMB 1.2 / kWh) to periods of high photovoltaic power generation or low-price periods (RMB 0.3 / kWh). This time-dimensional load reconfiguration optimizes the operating conditions of multi-energy complementary data centers and releases the adjustment potential of flexible load in data centers.

[0068] Table 2 Optimization results for each scenario Scene Operating cost / yuan Electricity purchase cost / yuan Gas purchase cost / yuan Maintenance cost / yuan 1 37628.16 14129.52 18141.62 5357.01 2 35746.93 17982.81 12982.47 4781.67 3 36080.30 18197.23 13090.83 4792.24 In actual operation, the randomness of wind and solar power output inevitably introduces additional economic costs. Considering the randomness of wind and solar power output, the two-stage stochastic optimization results for Scenario 3 show that the expected operating cost of the data center is 36,080.3 yuan, a decrease of 4.1% compared to the baseline cost of Scenario 1, but a slight increase of 333.37 yuan compared to the ideal scenario of Scenario 2. This difference essentially quantifies the risk premium paid by the system to cope with the uncertainty risk of renewable energy. Deterministic optimization, by ignoring the distribution of prediction errors, may lead to insufficient reserves in real-time operation; while two-stage stochastic programming allows the system to reserve sufficient reserve capacity and energy storage charging and discharging margin during the day-ahead scheduling phase to adapt to power fluctuations under the 30 typical scenarios in the second stage. This increased cost is the risk premium paid by the system to cope with the volatility risk of renewable energy.

[0069] A comprehensive comparison of scenarios 1 and 3 shows that, even considering the additional costs arising from the uncertainty of wind and solar power integration, the collaborative optimization strategy proposed in this paper still reduces data center operating costs by 4.11%. This result demonstrates that, in this example, the economic surplus created by the system's utilization of computing power flexibility not only fully covers the uncertain costs of wind and solar power integration but also achieves a certain net benefit, validating the effectiveness of the proposed method.

[0070] In summary, to address the issues of high energy consumption and fluctuating renewable energy output in data centers, this embodiment constructs a two-stage stochastic optimization scheduling framework for data centers that considers the uncertainty of wind and solar power output and the flexibility of computing load. Simulation results demonstrate that the discrete-time task flow model established for data center load in this embodiment quantifies the temporal elasticity of different computing tasks, effectively guiding the data center from a passive energy consumption mode to an active response mode where load follows source changes. Simulation results show that the source-load-storage collaborative mechanism can fully utilize the flexibility of computing load and the output characteristics of photovoltaic and wind power, exhibiting good adaptability and economy under strict load constraints, verifying the value of the computing power-power collaborative mechanism in reducing data center operating costs. Regarding the uncertainty of wind and solar power output, the two-stage stochastic optimization in this embodiment covers the adjustment costs caused by prediction errors with a low risk premium, providing effective support for the reliable operation of data centers under conditions of fluctuating renewable energy output and time-series mismatch of computing load.

[0071] Example 2 In one or more embodiments, a two-stage stochastic optimization system for data centers that takes into account wind and solar uncertainties is disclosed, comprising: The system modeling module is used to model the data center power supply system. It considers the differences in load latency tolerance, quantifies the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete time task flow, and introduces the cross-time balance constraint of computing power tasks to ensure that all tasks are processed before the deadline, thus constructing a refined load response model. The scene generation module is used to generate wind and solar power output samples based on Monte Carlo simulation. The generated random samples are made to satisfy the complementarity of wind and solar power output through Koleski decomposition. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenes, and the wind and solar power output curves under each scene are obtained. The system optimization module is used to construct a two-stage stochastic optimization model that includes computing power scheduling and multi-energy coordination with the goal of minimizing the expected operating cost of the system. The module solves the model and determines the optimal timing strategy for computing power tasks.

[0072] The specific implementation methods of the above modules are completely consistent with those in Example 1, and will not be described in detail again.

[0073] Example 3 In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the processor to perform the two-stage stochastic optimization method for data centers taking into account wind and solar uncertainties as described in Embodiment 1.

[0074] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0075] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0076] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0077] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the two-stage stochastic optimization method for data centers taking into account wind and solar uncertainties as described in Embodiment 1.

[0078] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A two-stage stochastic optimization method for data centers considering wind and solar uncertainties, characterized in that, include: Modeling of data center power supply system, taking into account the differences in load latency tolerance, quantifying the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete time task flow, and introducing cross-time balance constraints of computing power tasks to ensure that all tasks are processed before the deadline, thus constructing a refined load response model. Based on Monte Carlo simulation, samples of wind and solar power output are generated. Koleski decomposition is used to ensure that the generated random samples satisfy the complementarity of wind and solar power output. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenarios, and wind and solar power output curves under each scenario are obtained. With the goal of minimizing the expected operating cost of the system, a two-stage stochastic optimization model including computing power scheduling and multi-energy coordination is constructed. The model is solved to determine the optimal time-series operation strategy for computing power tasks.

2. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 1, characterized in that, The queue state evolution equation based on discrete-time task flow satisfies: current t+1 The backlog of the load task queue at a given time is equal to the backlog of the load task queue at the previous time. t Time of the first k The sum of arrivals of each type of task, and t Time of the first k The difference in processing volume between similar tasks; At any given time, the backlog of each type of load task queue is not less than 0, and the backlog of the load task queue at the beginning and end of each type of load task is equal to 0.

3. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 1, characterized in that, A cross-time task balancing constraint is introduced to ensure that all tasks are processed before the deadline. Specifically: For any type of computing task that arrives at any time, the system must process them all before their maximum allowable latency expires; however, the processing time cannot exceed the total scheduling cycle.

4. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 1, characterized in that, The generated random samples are decomposed using Koleski decomposition to ensure that wind power and photovoltaic output are complementary, specifically: Assumption t The prediction errors for wind power and solar power at any given time follow a multivariate normal distribution, where the covariance matrix is... ∑ The correlation between wind and solar power output was described: ;in, The coupling coefficient for wind and solar power output; For matrix ∑ Perform Koleski decomposition to obtain the lower triangular matrix. L and satisfy L The product of its transpose is equal to ∑ ; Based on the lower triangular matrix L Generate correlation coefficients ρ spatial random disturbance term .

5. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 4, characterized in that, Generate an error sequence with time autocorrelation based on a first-order autoregressive model: ； in, σ i The standard deviation of the prediction error; ρ tem This is the time autocorrelation coefficient; In scenario s, the first i Energy-like energy in the present t The prediction error value at time; In scenario s, the first i The prediction error value of the energy type at the previous time step; s represents the index of the random scene. t Represents the time step. i Represents specific types of renewable energy. PV represents photovoltaic power, and WT represents wind power.

6. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 1, characterized in that, After generating random scenes, a fast forward selection method based on Wasserstein distance is used to reduce the number of scenes. By quantizing and minimizing the probability distance between the reduced set and the original set, scenes are selected. N s A typical scenario and its reconstruction probability π s .

7. The two-stage stochastic optimization method for data centers considering wind and solar uncertainties as described in claim 1, characterized in that, A two-stage stochastic optimization model incorporating computing power scheduling and multi-energy coordination is constructed. Specifically, the model aims to minimize the expected comprehensive operating cost of the system across the entire scenario set. In the first stage, the scheduling plan of computing power tasks is used as the decision variable. In the second stage, the model is adjusted in real time to account for the stochasticity of wind and solar power output scenarios. The decision variables include gas turbine output, energy storage system charging and discharging status, and interaction power with the public power grid. The load task scheduling decision in the first stage influences the energy scheduling cost in the second stage by reshaping the computing power load task demand.

8. A two-stage stochastic optimization system for data centers that takes into account wind and solar uncertainties, characterized in that, include: The system modeling module is used to model the data center power supply system. It considers the differences in load latency tolerance, quantifies the migration capability and backlog constraints of services with different latency tolerances in the time dimension through the queue state evolution equation based on discrete time task flow, and introduces the cross-time balance constraint of computing power tasks to ensure that all tasks are processed before the deadline, thus constructing a refined load response model. The scene generation module is used to generate wind and solar power output samples based on Monte Carlo simulation. The generated random samples are made to satisfy the complementarity of wind and solar power output through Koleski decomposition. A first-order autoregressive model is introduced and combined with multivariate normal distribution to generate random scenes, and the wind and solar power output curves under each scene are obtained. The system optimization module is used to construct a two-stage stochastic optimization model that includes computing power scheduling and multi-energy coordination with the goal of minimizing the expected operating cost of the system. The module solves the model and determines the optimal timing strategy for computing power tasks.

9. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-7, which is a two-stage stochastic optimization method for data centers that takes into account wind and solar uncertainties.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed as described in any one of claims 1-7, the two-stage stochastic optimization method for data centers taking into account wind and solar uncertainties.

Images