Data center electricity and computing adjustment method and system based on electricity and heat computing collaborative interaction
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-26
Smart Images
Figure CN122285290A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of artificial intelligence applied to energy conservation and environmental protection technology, and relates to a data center computer regulation method and system, specifically a data center computer regulation method and system based on the synergistic interaction of electricity, heat and computing power. Background Technology
[0002] In recent years, the development of the digital economy and the application of artificial intelligence have placed enormous energy demands on infrastructure such as data centers, which bear these computing tasks, leading to high carbon emissions and cost issues. Data center infrastructure accounts for 2% of global electricity consumption and is increasing at least 15% annually. It is projected that by 2030, the national renewable energy capacity for data centers will reach 200-300 million kilowatts. However, most existing data centers still rely primarily on traditional energy sources, highlighting their energy consumption and carbon emission problems.
[0003] The exploration of flexible characteristics such as power, computing power, and heat is crucial for achieving dynamic balance between power supply and demand and efficient operation in computer computing. Among these, latency-tolerant computing tasks, in particular, can be flexibly adjusted based on power supply characteristics due to their long execution time. By considering the spatiotemporal adjustment characteristics of load, economical and low-carbon operation of data centers can be achieved. Furthermore, research on the coordinated operation of computing and heat in data centers can effectively reduce energy consumption and costs. However, the above rarely considers the flexible adjustment of power, computing power, and heat, lacks quantitative modeling and characterization of the flexible characteristics of computing tasks, and lacks research on the economic operation of data centers through the coordinated coupling of flexible characteristics of different energy flows.
[0004] Traditional computerized computing methods only handle demand response / energy storage on the power side or DVFS / resource allocation on the computing side, neglecting the coupling between cooling-temperature dynamics and power consumption. While this can optimize data center power consumption and economy to some extent, excessively high server power density can lead to a surge in cooling system power consumption. Therefore, it is necessary to reuse this energy to improve data center energy efficiency and economic operation. Furthermore, batch processing tasks are often treated coarsely in existing research, lacking more detailed characterization. Summary of the Invention
[0005] To address the aforementioned technical issues, this invention provides a data center power regulation method and system based on the coordinated interaction of electricity, heat, and computing power. Building upon traditional power optimization, it introduces a flexible quantification model for batch processing tasks to characterize the flexible adjustment characteristics of computing power tasks arriving at different times. It also considers the response time of interactive tasks as a hard constraint, thereby satisfying the power consumption experience of data center users and meeting SLA agreements.
[0006] The technical solution adopted by the method of the present invention is: a data center computer adjustment method based on the coordinated interaction of electrothermal computing power, comprising: Data center servers are classified into three categories: interactive task servers (Category A), batch task servers (Category B), and hybrid servers (Category H). Establish a service level agreement response time constraint model for interactive tasks, where interactive tasks are processed immediately upon arrival, and the response time does not exceed the maximum response time T. delay ; Establish a flexible transition matrix model for batch processing tasks by introducing a batch processing task coefficient transition matrix. Quantify the batch processing computing power tasks arriving at different times within the maximum flexibly adjustable time window D max The proportion of data migrated to other processing times; Establish a data center cooling and power supply coupling model, construct the relationship between indoor temperature dynamic changes and cooling power, heat generation power and heat dissipation power based on the thermodynamic balance equation, and set the feasible range of indoor temperature. With the overall economic operation of the data center as the goal, a two-stage robust optimization objective function is constructed, which includes equipment investment cost, electricity purchase cost, renewable energy curtailment penalty and carbon dioxide emission penalty. The solution yields task allocation and migration scheme, server operating frequency and number of servers in operation, cooling power or temperature trajectory, energy storage charging and discharging strategy and electricity purchase decision, so as to achieve dynamic supply and demand balance of the data center.
[0007] Preferably, the service level agreement response time constraint model for the interactive task is based on the M / M / 1 queuing theory model, where the task arrival rate at time t is [value missing]. and response time They are respectively: ; ; in, This represents the number of active Class A servers at time t, either active or in a working state. This represents the interactive computing task request allocated to the Class A server at time t. This indicates the service rate of Class A servers at their rated operating frequency. This represents the operating frequency of the Class A server at time t; The response time for interactive tasks on the H-type server is: ; in, This indicates the service rate of a Class H server at its rated operating frequency. This represents the operating frequency of the H-class server at time t; This represents the number of active H-class servers at time t, which are either active or in a working state. This represents the interactive computing task request assigned to the H-class server at time t.
[0008] Preferably, the flexible migration matrix model for the batch processing task includes: for t≤24 The batch processing task arriving at time D uses the first transition matrix. This indicates the proportion of processes that are deferred to a later time; for t>24 The batch processing task arriving at time D uses the second transition matrix. This represents the proportion processed at the start of the next adjustment cycle; the two transition matrices satisfy normalization constraints. ; ; ; in, , Both are 24 A matrix of 24 This represents the proportion of batch processing tasks that arrive at time t and are delayed until time t+D.
[0009] Preferably, the thermodynamic balance equation for the data center cooling and power supply coupling model is: ; ; , ; in, , and These represent the data center's density, heat capacity, and volume, respectively. Let be the indoor temperature of the data center at time t. , and The heat output of the data center at time t is respectively The cooling power generated by the refrigeration system and the power lost through heat conduction; COP is the coefficient of performance for refrigeration. For cooling power; k loss S is the heat transfer coefficient. DC The area for heat exchange between the data center and the outside environment, T amb The ambient temperature; The power consumed by the entire data center is .
[0010] As a preferred option, an electrochemical energy storage system model is also included, whose power balance equation is: ; in, The heat output of the data center at time t. For data center cooling power consumption, These are the power purchased from the power grid, the power supplied by wind turbines, and the power supplied by photovoltaics, respectively. and and represent the charging and discharging power of the energy storage system at time t; The constraints that the electrochemical energy storage system must satisfy are: ; in, Let t be the energy storage capacity of the battery energy storage system. and Let be the charging and discharging power of the energy storage system at time t. and These are the charge and discharge efficiencies, respectively. For the optimal capacity of the energy storage system, and These represent the energy storage capacity at the start and end times of the energy storage system, respectively. These represent the rated capacity of the energy storage system and the real-time capacity at time t, respectively.
[0011] Preferably, the two-stage robust optimization objective function is: ; Where F1 is the objective function for the first stage of optimization, and F2 is the objective function for the second stage of optimization; , ; in, For equipment investment costs, The unit capacity investment cost of equipment i, For optimized capacity of device i, Let r be the annual interest rate and I be the number of devices. ; Where, p k The feasible domain for typical scenarios (aggregated scenarios after clustering historical wind turbine and solar power output scenarios into K major categories, which belong to typical scenarios) The specific value in, For time-of-use electricity pricing, Let t be the power purchased at time t. Penalties for curtailing electricity from renewable energy sources Let t be the amount of renewable energy curtailed. Penalties for unit carbon dioxide emissions. Let t be the power purchased at time t.
[0012] As a preferred option, the constraints to be met for different types of services include: (1) The utilization constraint for any type of server is ; Indicates the maximum server utilization; (2) The number of any type of server activated at any given time ; Indicates the maximum number of k-type servers that can be installed; (3) The response time constraints for interactive tasks in A and H are as follows: ;T delay Maximum response time; (4) The maximum latency constraint for batch processing services is: ; This represents the maximum flexible migration window for batch processing tasks. (5) The normalized relative frequency constraint for any server is: , ; (6) The supply air temperature of the refrigeration system is constrained to be ; This indicates the minimum and maximum values of the temperature range specified for the data center.
[0013] As a preferred option, for simultaneous output from wind turbines and photovoltaic systems, the following capacity constraints must be met: ; in, , This indicates the installed capacity of photovoltaic and wind turbines.
[0014] The technical solution adopted by the system of this invention is: a data center computer control system based on electrothermal computing power collaborative interaction, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the data center computer regulation method based on electrothermal computing power collaborative interaction.
[0015] The present invention also provides a data center computer regulation product based on the coordinated interaction of electrothermal computing power, including computer program instructions, which, when the computer program instructions are run on a computer, cause the computer to execute the data center computer regulation method based on the coordinated interaction of electrothermal computing power.
[0016] Compared with the prior art, the beneficial effects of the present invention include: (1) This invention establishes an interactive task SLA response time constraint model and a batch processing task maximum delay window model, and divides the server into three categories: interactive, batch processing, and hybrid, so as to ensure that interactive tasks can be processed immediately without being preempted by batch processing tasks, while fully utilizing the flexible adjustment characteristics of batch processing tasks. (2) This invention introduces a batch processing flexible migration matrix to quantify the flexible adjustment characteristics of computing power tasks in the collaborative operation of computer computing and thermal computing. Through comparative analysis, the importance of the flexible adjustment characteristics of computing power tasks in the entire data economy operation can be obtained. (3) Under the above constraints, the present invention solves the task allocation / migration, server frequency and number of machines in operation, cooling power or temperature trajectory, energy storage charging and discharging and power purchase decision, so that the data center can realize dynamic supply and demand balance of computer computing. Attached Figure Description
[0017] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.
[0018] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram representing the daily computing power task request arrival rate in the experiment of this embodiment of the invention; Figure 3 These are power scheduling optimization diagrams under four scenarios in the experiments of this embodiment of the invention; where (a)-(d) are power scheduling optimization diagrams for S1-S4, respectively. Figure 4 The diagram shows the batch processing task transfer coefficients in the experiment of the embodiment of the present invention. Figures (a) and (b) respectively show the batch processing task migration and allocation schemes under the flexible characteristics of scenario S1, which only considers the computing power flexibility characteristics, and scenario S4, which simultaneously considers the power, computing power and heat characteristics. Figure 5 The figures (a)-(d) show the number of servers of different types started up under four scenarios in the experiment of the embodiment of the present invention. Figures (a)-(d) show the number of servers of the three types started up at different times under scenarios S1-S4. Detailed Implementation
[0019] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0020] First, the terminology used in this embodiment will be explained.
[0021] Batch processing tasks: Also known as batch jobs or batch tasks, these are user-submitted tasks that are queued by the scheduler and run in the background as a "complete unit of work." They typically do not require immediate results and can be processed with a delay, such as offline inference tasks and logs. Interactive tasks (also known as interactive workloads or interactive jobs) are workloads that directly interact with users and require immediate responses, such as web page requests, search queries, recommendation returns, online inference, and RPC calls. Their core characteristic is their high sensitivity to response latency (especially tail latency such as p95 / p99), typically with a clear latency target or upper limit; therefore, scheduling and resource management prioritize ensuring their latency and stability.
[0022] Service Level Agreement (SLA) is a service quality commitment agreed upon between a service provider and a customer / business partner. It typically includes clauses regarding latency, throughput, error rate, response time, support response time limits, and breach of contract compensation.
[0023] Dynamic Voltage and Frequency Scaling (DVFS) is a power management and performance scaling technology that dynamically adjusts the operating voltage and clock frequency of a processor (or GPU, etc.) based on load changes to reduce power consumption and heat generation while meeting performance requirements.
[0024] Please see Figure 1 This embodiment provides a data center computer adjustment method based on the collaborative interaction of electrothermal computing power, which includes the following steps: 1. Data center power consumption modeling; Typically, data center computing power task servers are divided into three categories: Type A: interactive task servers, which are dedicated to handling interactive tasks; Type B: batch processing task servers, which are dedicated to handling batch processing tasks; and Type H: hybrid servers, which execute both types of computing power tasks.
[0025] Models for interactive computing tasks and batch processing tasks were constructed respectively.
[0026] 1.1 Power consumption model for Class A servers; Due to the real-time processing nature of interactive tasks, assuming that the total number of interactive processing task requests allocated to the data center by the front-end server at time t is , the following formula holds true: ; in, and These represent the interactive computing task requests allocated to Class A servers and Class H servers at time t, respectively.
[0027] Since interactive task requests are processed immediately upon arrival, according to the SLA (Service Level Agreement) / QoS (Quality of Service) agreement, interactive tasks need to be processed within the maximum response time T. delay To ensure user satisfaction and comfort, the computing power task is distributed among different types of servers in this embodiment, and the task is evenly distributed across each active or working server. According to the M / M / 1 queuing theory model, the task arrival rate and response time at time t for server A are as follows: ; ; in, This represents the number of active Class A servers at time t, either active or in a working state. This indicates the service rate of Class A servers at their rated operating frequency. This represents the operating frequency of the Class A server at time t.
[0028] 1.2 Power Consumption Model for Class B Servers Class B servers are used to process batch processing tasks. Because these tasks are less time-sensitive, their time dimension can be flexibly adjusted according to network requirements. They also exhibit a certain periodicity. If we consider the total batch processing computing power arriving at a certain time t to be... The maximum adjustable time window is D. max At the same time, a batch processing task coefficient transition matrix is introduced. : ; Indicates 24 A matrix of 24, where This represents the proportion of batch processing tasks that arrive at time t and are delayed until time t+D.
[0029] To ensure that all computing tasks arriving at any given time have the same flexible adjustment time window, for batch processing tasks arriving at time t>24-D, since the time period is 24 hours, these tasks generally migrate towards the start of the next day. Because batch processing tasks have a certain periodicity, this paper argues that computing tasks arriving at time t>24-D can migrate towards the start of the adjustment period. Therefore, the transfer coefficient for this part of the migrated computing tasks is shown in the following formula: ; 24 A 24-dimensional matrix.
[0030] The two coefficient matrices must satisfy the following relationship: ; Combining the time transition coefficients of the two batch processing computing power tasks, the total computing power tasks processed at each time step in both Class A and Class H servers can be represented by the following matrix: ; The first part represents the total amount of computing power tasks processed at a certain time t, including the portion of tasks adjusted from previous times. The second part represents the portion of tasks whose adjustment time window exceeds 24 hours from the previous day. Because it has periodic changes, it is assumed to be the portion of tasks that can be adjusted back to the initial time of an adjustment cycle.
[0031] Since batch processing tasks do not need to be processed immediately upon arrival, their power consumption is generally based on the processing volume at each specific moment. Similarly, for Class B servers, the total task volume allocated at each moment can be expressed as: .
[0032] 1.3 H-class server power consumption model; Since the H-type server is a hybrid server type used to handle both interactive and batch processing tasks, based on the arrival rate of interactive tasks and the real-time processing volume of batch processing tasks, the total computing power tasks processed by the H-type server at any given time are: ; Generally, when interactive tasks and batch processing tasks arrive simultaneously, interactive tasks, which require real-time processing, are processed first on the H server during mixed task processing. Therefore, the response time for interactive tasks on the H server is: ; in, This indicates the service rate of a Class H server at its rated operating frequency. This represents the operating frequency of the H-class server at time t; This represents the number of active H-class servers at time t, which are either active or in a working state. This represents the interactive computing task request assigned to the H-class server at time t.
[0033] In power consumption modeling for computing tasks using Dynamic Voltage and Frequency Scaling (DVFS) technology, the main components are static power and dynamic computing power. Dynamic computing power is modeled based on the CPU's operating frequency and utilization rate within the server. Therefore, at time t, assuming uniform distribution of computing resources, the power consumption of each A, B, and H class server and the total power consumption are expressed as follows: ; Among them, C i,1 and C i,2 These are the server physical parameters, u t,i Server utilization can be expressed as follows: ; ; ; in, , , These represent the task arrival rate of Class A servers at time t, the task arrival rate of Class B servers with respect to latency tolerance load at time t, and the total computing power tasks processed by Class H servers at time t, respectively. Based on the above formula, the power consumed by the data center IT equipment at time t can be calculated as follows: ; 2. Construction of data center cooling and power supply model; The cooling capacity of a data center is adjusted by regulating the indoor temperature entering the building, thus leveraging its flexible thermal energy characteristics. The article treats the data center as a whole, and based on the thermodynamic balance equation, the following formula can be derived: ; in, , and These represent the data center's density, heat capacity, and volume, respectively. Let be the indoor temperature of the data center at time t. , and These represent the heat generation power of the data center at time t, the cooling power generated by the cooling system, and the power lost through heat conduction, respectively. Since the power consumption of CPU processing tasks in data centers is generally dissipated as heat, the heat generation power of servers in the data center is: ; In addition, data centers use dedicated cooling systems to cool server racks. The coefficient of performance (COP) is used to define the ratio of the cooling power generated by the cooling system to the electrical power consumed. Therefore, the cooling power generated by the cooling system can be used to obtain the electrical power consumed by the cooling system at time t: ; ; Where, k loss Where S is the heat transfer coefficient, S is the area of the data center exchanging heat with the outside environment, and T is the heat transfer coefficient. amb The ambient temperature. Power of the refrigeration system; Most of the power consumption in a data center is consumed by servers and cooling systems. The total power consumption of the entire data center can be approximated as follows:
[0034] Based on the above formula, a coupling relationship between the data center power consumption model and the cooling model is established. The power consumption of the cooling system is used to regulate the heat generated indoors by the servers processing computing tasks, maintaining the indoor temperature of the computer room within a certain range. At the same time, because the flexible thermal adjustment characteristic of the data center is that the indoor temperature of the data center can be changed by adjusting the cooling power, and the indoor temperature of the data center only needs to be maintained within a certain range, the cooling power of the cooling system can be adjusted flexibly according to the indoor temperature range of the data center. Generally, the indoor temperature range of the data center is set between 15 and 30 degrees Celsius.
[0035] In one implementation, the data center is primarily powered by the power grid, wind turbines, and photovoltaic systems. Simultaneously, an electrochemical energy storage system ensures flexible power adjustment for the data center. When the data center responds to the grid's time-of-use pricing, and the power consumption of wind turbines and photovoltaic systems is insufficient or excessive, the additional power is supplied by the energy storage system. Alternatively, the energy storage system's response power can be adjusted according to the data center's economic operation, thus achieving flexible power adjustment for the data center and ensuring a balance between power supply and demand. The power balance equation is as follows: ; in, The heat output of the data center at time t. , These represent the power of the cooling system, the power purchased by the grid, the output of the wind turbine, and the output of the photovoltaic system at time t, respectively. and and represent the charging and discharging power of the energy storage system at time t; The constraints that the electrochemical energy storage system must satisfy are: ; in, Let t be the energy storage capacity of the battery energy storage system. and Let be the charging and discharging power of the energy storage system at time t. For the optimal capacity of the energy storage system, and These represent charge and discharge efficiencies, and These represent the energy storage capacity at the start and end times of the energy storage system, respectively. These represent the energy storage capacity and the real-time capacity at time t, respectively.
[0036] By combining wind turbines, photovoltaics, and grid power supply, and utilizing the flexible characteristics of battery energy storage systems, data centers can be flexibly controlled to ensure stable and economical operation, thereby improving their overall economic efficiency and green, low-carbon characteristics.
[0037] 3. Objective function and constraints; With the overall economic operation as the overall goal, the aim is to achieve overall optimization of computer-aided collaboration. The overall objective function is shown in the following formula: ; Where F1 is the objective function for the first stage of optimization, and F2 is the objective function for the second stage of optimization.
[0038] In one implementation, the first-stage objective is to plan the capacity of equipment such as wind turbines and photovoltaic systems before operation. Through optimal capacity optimization based on worst-case typical scenario distributions, the capacity optimization results can cover operating conditions under all typical scenarios. This is specifically expressed as shown in the following formula: ; ; in, For equipment investment costs, The unit capacity investment cost of equipment i, For optimized capacity of device i, Let r be the total lifecycle of device i, and r be the annual interest rate.
[0039]
[0040] Where, p k The probabilistic feasible region for typical scenarios (aggregated scenarios after clustering historical wind turbine and solar power output scenarios into K major categories, i.e., typical scenarios) The specific value in, For time-of-use electricity pricing, Power purchased for the power grid Penalties for curtailing electricity from renewable energy sources For the abandoned power of new energy sources, Penalties for unit carbon dioxide emissions. Let t be the power purchased by the power grid at time t.
[0041] In one implementation, the objective function of the second stage means that, for any set of typical scenario probabilities in the feasible region, the worst probability distribution of the typical scenario is first solved, and then the optimal real-time running variable of each typical scenario under the worst probability distribution is solved to minimize its objective function.
[0042] The overall objective function expression for the computer-computer-thermal collaborative coupling economic operation scheduling using a two-stage robust optimization framework is shown in the following formula: ; The constraints that different types of services must satisfy are as follows, among which the inequality constraints satisfied by the equation are as follows, and the relative frequency f is set to four levels.
[0043]
[0044]
[0045]
[0046]
[0047]
[0048]
[0049] The first formula represents the utilization constraint for any type of server. The first formula represents the upper limit of server utilization; the second formula represents the number of three different servers active at any given time. This indicates the maximum number of servers installed in A. The first represents the maximum number of B and H type servers installed; the second represents the response time constraints for interactive tasks in A and H; the third represents the maximum latency constraint for batch processing services; the fourth represents the normalized relative frequency constraints for various server types; and the sixth represents the cooling system supply air temperature constraints. This indicates the minimum and maximum values of the temperature range specified for the data center.
[0050] In one implementation, power is simultaneously supplied to both wind turbines and photovoltaic systems, which requires meeting their capacity constraints. ; in, , This indicates the installed capacity of photovoltaic and wind turbines.
[0051] This embodiment aims at the overall economical operation of the data center. It constructs a two-stage robust optimization objective function that includes equipment investment costs, electricity purchase costs, renewable energy curtailment penalties, and carbon dioxide emission penalties (F1 represents investment costs, and F2 contains the costs of electricity purchase, curtailment, and carbon dioxide emission penalties, respectively). Solving this function yields the task allocation and migration scheme (modeled as follows: coefficient matrix expression). By modeling it, the flexible adjustment characteristics of computing power tasks can be quantified during the optimization process. That is, it can be clearly determined when computing power tasks arriving at any given time are executed and the number of tasks executed, thereby obtaining a task allocation and migration scheme after flexible adjustment of computing power tasks (see the results in Appendix 4 for details). Server operating frequency and number of servers started (in the data center power consumption modeling process constructed using DVFS technology) The expression, by modeling server power consumption under different scenarios and optimizing it using the power balance equation and the F1 and F2 objective functions, yields the number of servers powered on and the server power consumption under different scenarios, as well as the operating frequency of each type of server. See the appendix for details. Figure 5 The results shown), cooling power or temperature trajectory (i.e., the temperature balance equation constraint formula and cooling power) Energy storage charging and discharging strategies and power purchase decisions (i.e., electrochemical energy storage constraint equations and...) and This enables dynamic supply and demand balance in computer computing.
[0052] The invention will be further illustrated below through specific experiments.
[0053] This experiment uses HPE ProLiant DL380 Gen10 servers, with a total of 40,000 servers of three types. There are 10,000 Class A servers, divided into 2,000 racks, with 5 servers in each rack. The other two types of servers are divided into 300 racks, with 5 servers in each rack.
[0054] This experimental study considers the optimized operation of the power-computing-heat collaborative interaction under uncertain new energy scenarios, exploring the flexible adjustment characteristics of power, computing power, and heat. The following scenario optimization settings are used: S0 represents the reference scenario, which does not consider the flexible adjustment characteristics of computing power; that is, whether it is an interactive task or a batch processing task, it is processed immediately upon arrival, without considering the operation of the energy storage system or maintaining a fixed value for the intake air temperature of the data center cooling system; S1 represents the scenario considering only the flexible adjustment characteristics of the computing power side, that is, batch processing tasks can be flexibly processed with time delay; S2 represents the scenario considering only the flexible adjustment characteristics of the power side, that is, the energy storage system can flexibly adjust the dynamic balance of supply and demand; S3 represents the scenario considering only the flexible adjustment characteristics of the heat side, that is, the temperature can be flexibly adjusted within the feasible range; S4 is the scenario designed in this invention, representing the scenario considering the flexible adjustment characteristics during the power-computing-heat collaborative operation, as shown in Table 1.
[0055] Table 1. Computer-to-thermal coupling scenario settings
[0056] This experiment uses interactive requests processed through web searches conducted by users in Hong Kong using Google Chrome as the source of the interactive task dataset. Based on this search record, the total interactive task volume for the entire day of November 3, 2025, is used as a representative day to study the arrival rate of interactive task requests. The arrival rate curves for batch processing tasks and interactive tasks are shown below. Figure 2 As shown.
[0057] Table 2 shows the optimal capacity optimization results under different scenarios. It can be seen that the optimal photovoltaic capacity remains relatively stable, while the wind turbine capacity fluctuates significantly across different scenarios due to its lower daily investment compared to photovoltaic. Furthermore, in scenario S1, the flexibility of computing power scheduling allows for flexible scheduling of computing tasks based on the power output of both wind and photovoltaic turbines, resulting in relatively low power capacities for both. In scenario S2, which relies solely on energy storage for flexible adjustment, the inability to adjust computing tasks necessitates the energy storage system to coordinate power supply based on the real-time arrival rates of both computing tasks, leading to a surge in energy storage capacity. In scenario S4 of this experiment, the optimal economic operation of the computing system is ensured by rationally leveraging the positive adjustment characteristics of electricity, computing power, and heat.
[0058] Table 2. Optimal capacity optimization results under different scenarios (unit: MW)
[0059] Table 3 illustrates the economic characteristics of capacity planning and operation phases under different scenarios. It can be seen that the best overall economic performance is achieved in scenario S4 of this experiment, through reasonable adjustment of the flexibility of the three energy fields. Compared to not considering flexible adjustment characteristics, this reduces costs by nearly half. Secondly, it can be seen that in S1, relying solely on the flexible adjustment of computing power can reduce costs by 48.5%, indicating that leveraging the flexible adjustment characteristics of computing power is an important measure for the economical operation of data centers. Simultaneously, it can be seen that since computing power characteristic adjustment does not require additional power supply equipment, its investment ratio is relatively low, fully demonstrating that by scheduling computing power migration, it is possible to adapt to changes in electricity supply and achieve a balance between supply and demand. Furthermore, in S2, due to the rigid demand of computing power tasks and the high investment ratio of energy storage equipment, although energy storage is used to adjust the balance between power supply and demand, the economic performance only decreases by 22%. S3 represents flexible thermal energy regulation. However, by simply adjusting the flexible temperature range of the data center, i.e., adjusting the power of the cooling equipment, the impact on economic optimization is small, with only a 2% economic optimization effect. This is because the power adjustment of the cooling equipment has a small impact on the overall data center operating capacity and has a limited effect. Therefore, it needs to be combined with other objective functions to fully explore the potential of flexible temperature regulation.
[0060] Table 3. Economic Comparison of Different Operating Scenarios for Electricity, Computing, and Heating (Unit: Yuan)
[0061] Please see Figure 3 The diagram shows the power scheduling timing for four different scenarios (S0-S4). It can be seen that the power distribution differs significantly across the different scenarios. Figure 3 (a) In scenario S1, the computing tasks can be flexibly migrated, with most batch processing tasks migrating to times 1-7 and 23-24, as electricity costs are low during these periods, ensuring the economic efficiency of the supply and demand interaction in computing. Secondly, migration also occurs during times of abundant photovoltaic power generation, as these periods do not rely on grid power, ensuring a dynamic balance between supply and demand in computing while minimizing photovoltaic curtailment. This demonstrates that leveraging the flexible adjustment characteristics of computing tasks is crucial for the economic efficiency of computing. In comparison... Figure 3 (d) In the S4 scenario of the present invention, the power scheduling profile curve is similar to that of S1. It also uses time-of-use pricing and wind turbine photovoltaic power generation to flexibly adjust the computing power task. Furthermore, due to the presence of energy storage and cooling systems, it can maximize the optimization of system operation economy. Figure 3(b) In the S2 example, due to its large energy storage capacity, it can compensate for the supply and demand balance of electricity and computing power at different times. However, at certain times, despite the high electricity price, the system will still use grid power. This is mainly because when the computing power workload is too large and the wind turbine and photovoltaic power is insufficient, the cost of increasing the energy storage capacity is higher than the cost of purchasing electricity directly from the grid. Therefore, the optimization result will actively choose grid power. In addition, Figure 3 (c) In S3 shown, due to the limited thermal flexibility, the power supply and demand matching process of computer computing needs to rely on the power grid to achieve balance. Furthermore, due to the rigid demand of computing power tasks, the power curtailment of wind turbines and photovoltaics may occur at certain times, which will have a significant impact on economic costs.
[0062] Please see Figure 4 Figure 1 shows the batch processing task transfer coefficient diagram in the experiment of the embodiment of the present invention. Since the comparison study of scenarios S2 and S3 does not include the flexible migration characteristics of computing power tasks, the computing power tasks are processed as soon as they arrive in this scenario, and the computing power flexible adjustment potential is 0. Therefore, there is no computing power task allocation scheme. Figures (a) and (b) respectively show the batch processing task migration and allocation scheme under scenario S1, which only considers the flexible characteristics of computing power, and scenario S4, which simultaneously considers the three scenarios of power, computing power, and heat. The horizontal axis is the task processing time, and the vertical axis is the task arrival time. It can be seen that the batch processing computing power tasks that arrive at any time are processed at which times, that is, the flexible characteristics of processing at any time are quantified.
[0063] This invention can transform computing power flexibility from a concept into an optimizable variable, and can explicitly quantify the proportion of time delayed for processing.
[0064] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0065] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.
Claims
1. A data center electric and computing adjustment method based on electric heating computing power interaction, characterized in that, include: Data center servers are classified into three categories: interactive task servers (Category A), batch task servers (Category B), and hybrid servers (Category H). A service level agreement response time constraint model is established for interactive tasks, where an interactive task is processed as soon as it arrives and the response time does not exceed a maximum response time T delay ; A flexible migration matrix model of batch processing tasks is established, and a batch processing task coefficient transfer matrix is introduced , and the proportion of batch processing computing tasks arriving at different times within the maximum flexible adjustment time window D max is migrated to other times for processing. Establish a data center cooling and power supply coupling model, construct the relationship between indoor temperature dynamic changes and cooling power, heat generation power and heat dissipation power based on the thermodynamic balance equation, and set the feasible range of indoor temperature. With the overall economic operation of the data center as the goal, a two-stage robust optimization objective function is constructed, which includes equipment investment cost, electricity purchase cost, renewable energy curtailment penalty and carbon dioxide emission penalty. The solution yields task allocation and migration scheme, server operating frequency and number of servers in operation, cooling power or temperature trajectory, energy storage charging and discharging strategy and electricity purchase decision, so as to achieve dynamic supply and demand balance of the data center.
2. The data center computer adjustment method based on electrothermal computing power synergistic interaction according to claim 1, characterized in that: The service level agreement response time constraint model for the interactive task is based on the M / M / 1 queuing theory model, and the task arrival rate at time t is [value missing] in Class A servers. and response time They are respectively: ； ； in, This represents the number of active Class A servers at time t, either active or in a working state. This represents the interactive computing task request allocated to the Class A server at time t. This indicates the service rate of Class A servers at their rated operating frequency. This represents the operating frequency of the Class A server at time t; The response time for interactive tasks on the H-type server is: ； in, This indicates the service rate of a Class H server at its rated operating frequency. This represents the operating frequency of the H-class server at time t; This represents the number of active H-class servers at time t, which are either active or in a working state. This represents the interactive computing task request assigned to the H-class server at time t.
3. The data center computer adjustment method based on electrothermal computing power synergistic interaction according to claim 1, characterized in that: The flexible migration matrix model for the batch processing task includes: for t≤24 The batch processing task arriving at time D uses the first transition matrix. This indicates the proportion of processes that are deferred to a later time; for t>24 The batch processing task arriving at time D uses the second transition matrix. This represents the proportion processed at the start of the next adjustment cycle; the two transition matrices satisfy normalization constraints. ； ； ； in, , Both are 24 A matrix of 24 This represents the proportion of batch processing tasks that arrive at time t and are delayed until time t+D.
4. The data center computer adjustment method based on electrothermal computing power synergistic interaction according to claim 1, characterized in that: The thermodynamic balance equation for the data center cooling and power supply coupling model is as follows: ； ； ， ； in, , and These represent the data center's density, heat capacity, and volume, respectively. Let be the indoor temperature of the data center at time t. , and The heat output of the data center at time t is respectively The cooling power generated by the refrigeration system and the power lost through heat conduction; COP is the coefficient of performance for refrigeration. Data center cooling power consumption; k loss S is the heat transfer coefficient. DC The area for heat exchange between the data center and the outside environment, T amb The ambient temperature; The power consumed by the entire data center is .
5. The data center computer adjustment method based on electrothermal computing power collaborative interaction according to claim 1, characterized in that: It also includes an electrochemical energy storage system model, whose power balance equation is: ； in, The heat output of the data center at time t. For data center cooling power consumption, These are the power purchased from the power grid, the power supplied by wind turbines, and the power supplied by photovoltaics, respectively. and and represent the charging and discharging power of the energy storage system at time t; The constraints that the electrochemical energy storage system must satisfy are: ； in, Let t be the energy storage capacity of the battery energy storage system. and Let be the charging and discharging power of the energy storage system at time t. and These are the charge and discharge efficiencies, respectively. For the optimal capacity of the energy storage system, and These represent the energy storage capacity at the start and end times of the energy storage system, respectively. These are the rated energy storage capacity and the real-time energy storage capacity at time t, respectively.
6. The data center computer adjustment method based on electrothermal computing power synergistic interaction according to claim 1, characterized in that, The objective function for the two-stage robust optimization is: ； Where F1 is the objective function for the first stage of optimization, and F2 is the objective function for the second stage of optimization; ， ； in, For equipment investment costs, The unit capacity investment cost of equipment i, For optimized capacity of device i, Let r be the annual interest rate, and I be the number of devices. Device i includes wind turbines, photovoltaics, and energy storage devices. ； Where, p k For typical scenarios, the probabilistic feasible region The specific distribution in For time-of-use electricity pricing, Let be the grid power at time t. Penalties for curtailing electricity from renewable energy sources This refers to the power that has been abandoned. Penalties for unit carbon dioxide emissions. Let t be the grid power at time t; the aggregated scenario after clustering historical wind turbine and photovoltaic power output scenarios into K major categories is the typical scenario.
7. The data center computer adjustment method based on electrothermal computing power synergistic interaction according to claim 6, characterized in that, The constraints that must be met for different types of services include: (1) The utilization constraint for any type of server is ; This indicates the maximum server utilization value; (2) The number of any type of server activated at any given time ; Indicates the maximum number of k-type servers that can be installed; (3) The response time constraints for interactive tasks in A and H are as follows: ;T delay Maximum response time; (4) The maximum latency constraint for batch processing services is: ; This indicates the maximum time window for flexible batch migration; (5) The normalized relative frequency constraint for any server is: , ; (6) The supply air temperature of the refrigeration system is constrained to be ; This indicates the minimum and maximum temperatures specified by the data center.
8. The data center computer adjustment method based on electrothermal computing power collaborative interaction according to any one of claims 1-7, characterized in that: Simultaneously, for the output of wind turbines and photovoltaic systems, the capacity constraints that need to be met include: ； in, , This indicates the capacity of photovoltaic and wind turbines.
9. A data center computer control system based on electrothermal computing power synergistic interaction, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the data center computer regulation method based on electrothermal computing power collaborative interaction as described in any one of claims 1 to 8.
10. A data center computer control product based on electrothermal computing power collaborative interaction, comprising computer program instructions, characterized in that: When the computer program instructions are executed on the computer, the computer performs the data center computer adjustment method based on the collaborative interaction of electrothermal computing power as described in any one of claims 1 to 8.