Data center computing power-energy collaborative optimization method considering task time sequence flexibility

Abstract

The invention provides a data center computing power-energy collaborative optimization method considering task time sequence flexibility, which is used for a server cluster bearing off-line batch processing load and optimizing computing power and energy collaborative optimization capability by mining time shifting potential of a computing power task. The method comprises the following steps: S1, aiming at an off-line batch processing computing power load of a computing power task, constructing a computing power task sequence decomposition model based on a directed acyclic graph (DAG), and describing a time sequence dependency relationship inside the off-line batch processing load; step S2, based on the step 1, setting a computing power task-energy consumption mapping model based on a Transform and a Kolmogov-Arnod neural network TKAN; s3, based on the step 2, setting a data center computing power-energy collaborative optimization model considering computing power task dependency characteristics; according to the method, through fine modeling of the computing power task dependency relationship and the energy consumption mapping relationship, the adjustment potential of the data center on the time scale is mined, and collaborative optimization of reducing the operation cost and improving new energy consumption and service quality is realized.

Description

Data Center Computing Power-Energy Co-optimization Method Considering Task Timing Flexibility Technical Field

[0001] This invention relates to the field of data center energy management and optimization scheduling technology, and in particular to a data center computing power-energy collaborative optimization method that takes into account task timing flexibility. Specifically, it relates to a data center computing power-energy collaborative optimization method that takes into account task timing flexibility for offline batch processing computing loads in cloud data center clusters. Background Technology

[0002] Existing research has explored strategies for optimizing data center energy efficiency, making significant contributions to reducing energy consumption. However, existing optimization methods are mostly aimed at online service loads that are real-time but have weak dependencies, lacking in-depth research on offline batch processing loads that are common in large cloud data center clusters, have complex topology dependencies, and are energy-intensive.

[0003] Specifically, most existing studies optimize computing loads as a freely movable whole. In reality, the computing tasks within a computing load have strict logical dependencies. This simplistic approach, ignoring these dependencies, often overestimates the adjustability potential of the computing load, making the optimization results difficult to implement in real-world production scenarios. Regarding energy consumption modeling, there is a lack of precise models that accurately characterize the nonlinear coupling relationships of multidimensional variables. Existing research mostly employs traditional mechanistic models or fundamental data-driven methods.

[0004] In reality, data center energy consumption comes not only from IT equipment but is also significantly affected by cooling systems. Fluctuations in computing load create complex "computing power-electricity-heat-cooling" coupling relationships. Existing methods struggle to accurately describe the nonlinear coupling between computing tasks, IT resource utilization, indoor and outdoor temperatures, and various types of energy consumption, and they also have limitations in handling long-term time-series dependencies, resulting in limited accuracy of energy consumption models.

[0005] Furthermore, in terms of multi-energy synergistic optimization, there is a lack of synergistic mechanisms that comprehensively consider the interaction between various types of energy consumption and service quality. Existing research often focuses only on optimizing electrical energy consumption, without fully considering the synergistic effects of cooling / heating energy consumption and multi-energy complementary equipment. It also lacks a comprehensive consideration that quantifies service quality as a cost or incentive mechanism and incorporates it into the optimization objective. Against this backdrop, how to accurately quantify task dependencies, accurately characterize the energy consumption relationship under the influence of ambient temperature, and achieve efficient synergistic optimization of computing power and various energy types are urgent problems to be solved. Summary of the Invention

[0006] This invention proposes a data center computing power-energy synergistic optimization method that takes into account the flexibility of task timing. By finely modeling the computing power task dependency relationship and energy consumption mapping relationship, it explores the adjustment potential of data centers on the time scale, and achieves synergistic optimization of reducing operating costs, improving the consumption of new energy and service quality.

[0007] The present invention adopts the following technical solution.

[0008] A data center computing power-energy synergistic optimization method that takes into account task timing flexibility is used for server clusters that carry offline batch processing loads. The method optimizes the synergistic optimization capability between data center computing power and energy by finely modeling the computing power task dependency relationship and energy consumption mapping relationship, and by mining the time-shifting potential of computing power tasks, thereby reducing operating costs. The method includes the following steps:

[0009] Step S1: For the offline batch processing computing load of computing power tasks, construct a computing power task sequence decomposition model based on a directed acyclic graph (DAG) to describe the temporal dependencies within the offline batch processing load.

[0010] Step S2: Based on Step 1, set up a computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold neural network TKAN.

[0011] Step S3: Based on step two, set up a data center computing power-energy collaborative optimization model that takes into account the computing power task dependency characteristics.

[0012] The data center is a public or private cloud data center server cluster containing multiple computing nodes. The server cluster includes multiple server nodes connected by a network. The computing power task is an offline batch processing computing power load with time-series dependencies. The offline batch processing computing power load includes AI training and big data processing. The computing power task load consists of multiple computing power tasks, which are distributed to different server nodes in the server cluster for execution.

[0013] Step S1, the process of constructing the computing power task sequence decomposition model specifically includes:

[0014] 1) Dependency Analysis Between Computing Tasks: This method targets offline batch processing computing loads running on server clusters. Unlike online requests with immediate responses, such data center computing loads are not sensitive to latency, allowing for flexible scheduling within a specific time frame. Specifically, based on latency tolerance and computing usage, most computing tasks constituting these loads exhibit latency tolerance and are adjustable over time. The method establishes the execution order of computing tasks by identifying the temporal dependencies between tasks—that is, some subsequent computing tasks need to read the results of preceding computing tasks, forming a unidirectional, loop-free structure—and establishing time constraints for the computing task sequence. Based on this dependency analysis, the execution time and relaxation time (the maximum amount of time a task can be delayed without affecting the total completion time) of each computing task can be derived. Subsequently, the data center sets the start time and acceptable deadline for each computing task based on the latest completion time, service quality, and operating costs of the computing load. Finally, the task scheduler allocates computing tasks to servers with sufficient computing resources based on the aforementioned time constraints and resource requirements. The servers then return the results to the user.

[0015] 2) Computing power task sequence decomposition modeling: Directed acyclic graphs (DAGs) are chosen as the modeling basis because their acyclicity avoids cyclic dependencies, and their topological ordering properties naturally adapt to task time constraints. This model transforms the high-dimensional, large-scale mixed-integer programming problem into a linear programming problem adapted to computing power-energy co-optimization. The specific modeling process is as follows:

[0016] Step 1: Extract the computing task characteristics from the server cluster data, including the name of the computing load. Task Name Task start time Task end time CPU usage of the task Task storage usage ;

[0017] Step 2: Construct the DAG model; first, integrate the data extracted in Step 1. and Subsequently, based on the DAG theoretical framework, a DAG model reflecting the dependency characteristics between computing power tasks was established to accurately characterize the topological order and execution priority between different task nodes, laying the foundation for quantifying the time adjustability of computing power tasks.

[0018] Step 3: Quantify the time adjustability of computing power tasks. The time adjustability of computing power tasks is related to their relaxation time and execution time; execution time The definition is as follows:

[0019] Formula 1;

[0020] In the formula, and respectively computing power tasks End time and start time; time adjustability The definition is as follows:

[0021]

[0022] Formula 2;

[0023] In the formula, This indicates that computing task i has no dependent tasks; Indicates task Tasks dependent on computing power This means that the computing power task needs to be completed. Only then can the computing power task begin to be executed. ; This indicates that each computing task in the computing task set I depends on a computing task. It should be noted that after this step is completed, the original computing load is broken down into several computing tasks according to the task execution order.

[0024] Step 4: Constructing Timing Constraints for Computing Tasks: To ensure that the execution of computing tasks falls between the start and end times, the timing constraints for computing task execution are defined as follows:

[0025] Formula 3;

[0026] In the formula, This represents the execution ratio of computing task i in scenario s during time period k. ; The start time of the time period to which the computing task belongs; scheduling period. Each day is divided into The duration of each time period is... =15min; In addition, to effectively evaluate the latency of computing tasks, the actual start time of the computing task is... and waiting time for:

[0027] Formula 4;

[0028] Formula 5;

[0029] The above process reflects the implementation steps from dataset to computational task sequence decomposition model. Formula 3 ensures that the computational tasks are executed within the time interval k between the start and end times, while also ensuring that the task execution time meets the specified conditions. As required, Formulas 4 and 5 define the actual start time and waiting time of computing tasks, laying the foundation for subsequent optimization of service quality.

[0030] In step S2, the computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold Neural Network (TKAN) specifically includes a conventional data center energy consumption model and a data-driven computing power-energy consumption mapping model.

[0031] A typical data center energy consumption model is as follows: Generally, data center energy consumption includes electrical energy consumption and cooling energy consumption. Based on the computing power and task processing methods, it is divided into fixed energy consumption and dynamic energy consumption. Fixed energy consumption is the basic requirement for maintaining data center operation, while the dynamic part is related to the usage of CPU and storage, and is also affected by indoor and outdoor temperatures. The total power consumption of the server... Mainly affected by CPU utilization With indoor air temperature The effect is shown in the following formula:

[0032] Formula 6;

[0033] Formula 7;

[0034] In the formula: and These are idle power and peak power, respectively. Here, represents the temperature fitting coefficient. The indoor temperature of a data center increases with server heat dissipation and needs to be maintained within the required range through a cooling system. The corresponding thermal dynamic process is represented by a thermodynamic equivalent thermal parameter model as follows:

[0035] Formula 8;

[0036] In the formula, The heat dissipation power of the data center is expressed as ; This refers to the cooling capacity; and These are the equivalent thermal fusion and equivalent thermal resistance of the data center, respectively; and These are the indoor and outdoor temperatures, respectively.

[0037] Because the dynamic component of energy consumption is affected by CPU, storage, and indoor and outdoor temperatures, exhibiting nonlinear characteristics, for clear analysis, the power consumption and cooling energy consumption of a data center are defined as follows:

[0038] Formula 9;

[0039] Formula 10;

[0040] In the formula, and These are the data center's power consumption and cooling energy consumption, respectively. and These are the fixed power consumption and fixed cooling energy consumption of the data center, respectively. and They are respectively affected by CPU resources Storage resources Indoor temperature and outdoor temperature The impact on dynamic power consumption and dynamic cooling energy consumption.

[0041] The data-driven computing power-energy consumption mapping model is characterized by a heterogeneous and nonlinear relationship between computing power tasks and energy consumption. Traditional mechanistic models, based on strong physical assumptions, struggle to precisely quantify the mapping relationship between various types of energy consumption and computing power task sequences. In contrast, data-driven methods do not rely on physical assumptions and, leveraging the learning characteristics of data, possess a powerful ability to approximate complex nonlinear functions. Therefore, this paper proposes the Transformer-Kanmogorov-Arnold Network (TKAN) computing power task-energy consumption mapping model, combining the efficient nonlinear modeling capabilities of the Kolmogorov-Arnold Network (KAN) with the complex data dependency extraction advantages of the Transformer.

[0042] Establish a computational power task-energy consumption mapping model for Transformer-Kolmogorov-Arnold Neural Network (TKAN), including the following steps;

[0043] Step 1: Data Processing and Normalization, specifically: collecting and processing computing power tasks and related data. Input data includes CPU usage. Storage and computing usage and indoor temperature and outdoor temperature The tag data includes dynamic power consumption and dynamic cooling energy consumption;

[0044] Step 2: Building and training TKAN, specifically: integrating the Transformer module and the KAN module, using the Transformer's multi-head attention mechanism to extract long-term temporal dependency features of computing power task sequences, and using the KAN network's learnable univariate activation function to approximate the heterogeneous nonlinear mapping relationship between computing power tasks and energy consumption.

[0045] Step 3: Model testing and validation, specifically: In order to effectively evaluate the generalization performance of the model, cross-validation is introduced to divide the dataset into training set and test set, and mean squared error is selected as the standard for evaluating the performance of the trained TKAN computing power task-energy consumption mapping model.

[0046] After training the TKAN computing power task-energy consumption mapping model, the dynamic energy consumption of computing power tasks was obtained through testing. This allowed for the calculation of the total energy consumption of the data center during each scheduling period, thus coupling computing power tasks with energy consumption. Based on this, the electrical and cooling energy consumption of the data center is expanded as follows:

[0047] Formula 11;

[0048] Formula 12;

[0049] In the formula, , The power consumption and cooling consumption of computing task i are denoted as .

[0050] In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, includes an objective function and constraints. The objective function is used to define the collaborative optimization model; specifically, the optimization model aims to minimize the total operating cost of the data center, covering operating costs. Carbon emission costs Service quality cost Service quality costs comprise two parts:

[0051] Delay penalty cost: When the waiting time of a computing task exceeds its execution time, a service quality loss cost is incurred.

[0052] Transfer Compensation: To compensate for the impact of scheduling changes on user service quality, the system pays transfer compensation to the data center. The compensation amount depends on the sufficiency of photovoltaic output and the electricity price level during the scheduling period, and is usually higher during peak photovoltaic output or off-peak electricity price periods. When the transfer compensation is greater than the delay penalty cost, the service quality is negative, which is an incentive; when the transfer compensation is less than the penalty cost, the service quality is positive, which is a cost. The specific expression is:

[0053] Formula 13;

[0054] Formula 14;

[0055] Formula 15;

[0056] Formula 16;

[0057]

[0058] Formula 17;

[0059] In the formula, , Let and represent the probability of occurrence of the s-th typical photovoltaic power output scenario obtained from historical data clustering, and the probability of occurrence of the typical computing task scenario, respectively. Assuming they are independent, their joint scenario probability is... ; , , , These represent the electricity purchase price, average electricity purchase price, gas purchase price, and carbon emission price, respectively. , These represent the carbon emission intensity coefficients for purchasing energy from the gas grid and the power grid, respectively. , These represent the service quality penalty factor and the transfer compensation factor, respectively. , These respectively indicate the degree of influence from photovoltaic power and electricity price guidance; , , , , These represent the purchased electricity, purchased gas, photovoltaic power output, dynamic energy consumption, and average photovoltaic power output at time k in the s-th scenario, respectively.

[0060] In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, has constraints that form the constraint condition model for constructing the collaborative optimization model. Specifically, the constraints include constraints on data center cogeneration units, electric chiller units, absorption chiller units, heat exchangers, electric energy storage, and power balance constraints; among which, the cogeneration unit constraints are:

[0061] Formula 18;

[0062] Formula 19;

[0063] In the formula, , These represent the electrical power and thermal power output of the combined heat and power generation at time k in the s-th scenario, respectively. , These represent the electrical conversion efficiency and thermal conversion efficiency of combined heat and power (CHP), respectively. Indicates the calorific value of the gas used in a combined heat and power (CHP) unit;

[0064] The constraints of the electric chiller unit are:

[0065] Formula 20;

[0066] In the formula, , These represent the power consumption and cooling capacity required for cooling at time k in the s-th scenario, respectively. For electric cooling efficiency;

[0067] The constraints of absorption chiller units are:

[0068] Formula 21;

[0069] In the formula, , These represent the heat required for absorption cooling at time k in the s-th scenario and its cooling capacity, respectively. This refers to the absorption cooling efficiency.

[0070] In step S3, the waste heat generated by the data center servers depends on the power consumption of the data center's computing tasks. The waste heat recovered through the heat exchanger is defined as follows:

[0071] Formula 22;

[0072] In the formula, Let K be the server heat recovered by the heat exchanger at time k in the s-th scenario. For waste heat recovery efficiency;

[0073] The constraints of electrical energy storage are:

[0074] Formula 23;

[0075] In the formula, For the electrical energy stored at time k in the s-th scenario; , , These are the self-discharge rate of energy storage, charging efficiency, and discharging efficiency, respectively. , These are the charging power and discharging power of the energy storage at time k-1 in the s-th scenario, respectively; , These are the minimum and maximum capacities of electrical energy storage, respectively. , These represent the charging state and the discharging state, respectively.

[0076] Since thermal energy storage and electrical energy storage have similar models, they will not be elaborated upon here.

[0077] The power balance constraint is:

[0078] Formula 24;

[0079] Formula 25;

[0080] Formula 26;

[0081] In the formula, , Let represent the data center's cold energy consumption and hot energy consumption at time k in the s-th scenario, respectively; , These represent the charging and releasing power of the thermal energy storage at time k in the s-th scenario, respectively.

[0082] The data center computing power-energy synergistic optimization model, which takes into account the computing power task dependency characteristics, is related to the computing power task sequence decomposition model and the computing power task-energy consumption coupling constraint. The computing power task sequence decomposition model is expressed by Equations 1 to 5, and the computing power task-energy consumption coupling constraint is expressed by Equations 11 and 12.

[0083] The computing power-energy co-optimization model, which takes into account the task-dependent characteristics of computing power, is expressed by the following formula:

[0084] Formula 27;

[0085] st (1-5)—(11-26) Formula 28.

[0086] When the optimization method is used, if the data center power supply includes photovoltaic power generation, then to address the uncertainty of the arrival time of computing tasks and photovoltaic output, a large number of random scenarios are first generated using Monte Carlo simulation, and then the hierarchical clustering algorithm is used to reduce the large number of scenarios to a typical N scenarios (e.g., N=10) to balance computing accuracy and efficiency; then the DAG constraints obtained in step S1 and the TKAN energy consumption prediction model trained in step S2 are embedded into the optimization model in step three.

[0087] In step one, the computing power task characteristics of Alibaba cluster data are extracted.

[0088] Solve the problem using the Gurobi commercial solver in a Python environment.

[0089] This invention provides a data center computing power-energy synergistic optimization method that considers task timing flexibility. The data center is a public cloud data center or a dedicated cloud data center cluster, containing multiple server nodes interconnected via a network. The computing power tasks are mainly offline batch processing tasks with time-dependent relationships, including but not limited to distributed computing tasks, artificial intelligence training tasks, or scientific computing tasks. These tasks typically exhibit a directed acyclic graph (DAG) structure. By finely modeling the computing power task dependencies and energy consumption mapping relationships, the method explores the adjustment potential of the data center on a time scale, achieving synergistic optimization that reduces operating costs, improves renewable energy consumption, and enhances service quality.

[0090] This invention addresses the challenges of surging energy consumption and high operating costs in data centers under high computing power demands. It proposes a data center computing power-energy collaborative optimization method that considers task timing flexibility. By constructing a computing power task sequence decomposition model based on a directed acyclic graph, this invention can accurately quantify the delay time and execution timing constraints of various computing power tasks. This effectively avoids the problem of scheduling results deviating from reality or incurring additional penalty costs due to overestimating the computing power adjustment potential, thus providing a more flexible energy consumption adjustment capability that better suits practical application scenarios.

[0091] This invention proposes a computing power task-energy consumption mapping model that integrates Transformer and Kolmogorov-Arnold Neural Network (TKAN). This model can quickly and accurately capture the complex nonlinear relationship between computing power task sequences and electricity and cooling energy consumption, taking into account indoor and outdoor temperature variations. This provides precise energy consumption boundary conditions for collaborative optimization. By constructing a computing power-energy collaborative optimization model that considers task timing flexibility, task timing can be flexibly adjusted to match fluctuations in new energy output and time-of-use electricity prices. This optimizes the energy consumption structure of data centers, achieving an effective balance between economy, low carbon emissions, and service quality, thereby effectively reducing the total operating cost of the system. Attached Figure Description

[0092] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0093] Figure 1 is a flowchart of the computing power-energy collaborative optimization method in an embodiment of the present invention;

[0094] Figure 2 is a schematic diagram of the computing power task dependency characteristic analysis and DAG model in an embodiment of the present invention;

[0095] Figure 3 is a flowchart of the computing power task-energy consumption mapping model based on TKAN in an embodiment of the present invention;

[0096] Figure 4 is a comparison chart of data center flexibility power consumption under different optimization models in the embodiments of the present invention;

[0097] Figure 5 is a comparison diagram of data center energy balance under different optimization models in an embodiment of the present invention. Detailed Implementation

[0098] As shown in Figure 1, a data center computing power-energy synergistic optimization method that takes into account task timing flexibility is used for server clusters that carry offline batch processing loads. By finely modeling the computing power task dependency relationship and energy consumption mapping relationship, and by mining the time-shifting potential of computing power tasks, the method optimizes the synergistic optimization capability between data center computing power and energy, thereby reducing operating costs. The method includes the following steps.

[0099] Step S1: For the offline batch processing computing load of computing power tasks, construct a computing power task sequence decomposition model based on a directed acyclic graph (DAG) to describe the temporal dependencies within the offline batch processing load.

[0100] Step S2: Based on Step 1, set up a computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold neural network TKAN.

[0101] Step S3: Based on step two, set up a data center computing power-energy collaborative optimization model that takes into account the computing power task dependency characteristics.

[0102] The data center is a public or private cloud data center server cluster containing multiple computing nodes. The server cluster includes multiple server nodes connected by a network. The computing power task is an offline batch processing computing power load with time-series dependencies. The offline batch processing computing power load includes AI training and big data processing. The computing power task load consists of multiple computing power tasks, which are distributed to different server nodes in the server cluster for execution.

[0103] Step S1, the process of constructing the computing power task sequence decomposition model specifically includes:

[0104] 1) Analysis of the dependency characteristics between computing power tasks, specifically:

[0105] This embodiment targets offline batch processing computing loads running on server clusters. Unlike online requests that require immediate responses, such data center computing loads are not sensitive to latency, allowing for flexible scheduling within a specific timeframe. Specifically, based on latency tolerance and computing usage, most of the computing tasks constituting these loads possess latency-tolerant characteristics and are adjustable over time. The execution order of computing tasks is established based on the temporal dependencies between tasks—that is, some subsequent computing tasks need to read the results of preceding tasks—forming a unidirectional, loop-free structure. This establishes time constraints on the computing task sequence. For example, in Figure 2, tasks 2 and 3 must be executed after task 1 is completed, while task 4 must be started after tasks 2 and 3 are completed.

[0106] Based on the results of this dependency analysis, the execution time and relaxation time (the maximum amount of time a task can be postponed without affecting the total completion time) of each computing task can be derived. Subsequently, the data center sets the start time and acceptable deadline of each computing task according to the latest completion time of the computing load, service quality, operating costs and other objectives. Finally, the task scheduler allocates the computing tasks to servers with sufficient computing resources for execution based on the above time constraints and resource requirements. The servers then return the results to the users.

[0107] 2) Computing power task sequence decomposition modeling: Directed acyclic graphs (DAGs) are chosen as the modeling basis because their acyclicity avoids cyclic dependencies, and their topological ordering properties naturally adapt to task time constraints. This model transforms the high-dimensional, large-scale mixed-integer programming problem into a linear programming problem adapted to computing power-energy co-optimization. The specific modeling process is as follows:

[0108] Step 1: Extract the computing task characteristics from the server cluster data, including the name of the computing load. Task Name Task start time Task end time CPU usage of the task Task storage usage ;

[0109] Step 2: Construct the DAG model; first, integrate the data extracted in Step 1. and Subsequently, based on the DAG theoretical framework, a DAG model reflecting the dependency characteristics between computing power tasks was established to accurately characterize the topological order and execution priority between different task nodes, laying the foundation for quantifying the time adjustability of computing power tasks.

[0110] Step 3: Quantify the time adjustability of computing power tasks. The time adjustability of computing power tasks is related to their relaxation time and execution time; execution time The definition is as follows:

[0111] Formula 1;

[0112] In the formula, and respectively computing power tasks End time and start time; time adjustability The definition is as follows:

[0113]

[0114] Formula 2;

[0115] In the formula, This indicates that computing task i has no dependent tasks; Indicates task Tasks dependent on computing power This means that the computing power task needs to be completed. Only then can the computing power task begin to be executed. ; This indicates that each computing task in the computing task set I depends on a computing task. It should be noted that after this step is completed, the original computing load is broken down into several computing tasks according to the task execution order.

[0116] Step 4: Constructing Timing Constraints for Computing Tasks: To ensure that the execution of computing tasks falls between the start and end times, the timing constraints for computing task execution are defined as follows:

[0117] Formula 3;

[0118] In the formula, This represents the execution ratio of computing task i in scenario s during time period k. ; The start time of the time period to which the computing task belongs; scheduling period. Each day is divided into The duration of each time period is... =15min; In addition, to effectively evaluate the latency of computing tasks, the actual start time of the computing task is... and waiting time for:

[0119] Formula 4;

[0120] Formula 5;

[0121] The above process reflects the implementation steps from dataset to computational task sequence decomposition model. Formula 3 ensures that the computational tasks are executed within the time interval k between the start and end times, while also ensuring that the task execution time meets the specified conditions. As required, Formulas 4 and 5 define the actual start time and waiting time of computing tasks, laying the foundation for subsequent optimization of service quality.

[0122] In step S2, the computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold Neural Network (TKAN) specifically includes a conventional data center energy consumption model and a data-driven computing power-energy consumption mapping model.

[0123] A typical data center energy consumption model is as follows: Generally, data center energy consumption includes electrical energy consumption and cooling energy consumption. Based on the computing power and task processing methods, it is divided into fixed energy consumption and dynamic energy consumption. Fixed energy consumption is the basic requirement for maintaining data center operation, while the dynamic part is related to the usage of CPU and storage, and is also affected by indoor and outdoor temperatures. The total power consumption of the server... Mainly affected by CPU utilization With indoor air temperature The effect is shown in the following formula:

[0124] Formula 6;

[0125] Formula 7;

[0126] In the formula: and These are idle power and peak power, respectively. Here, represents the temperature fitting coefficient. The indoor temperature of a data center increases with server heat dissipation and needs to be maintained within the required range through a cooling system. The corresponding thermal dynamic process is represented by a thermodynamic equivalent thermal parameter model as follows:

[0127] Formula 8;

[0128] In the formula, The heat dissipation power of the data center is expressed as ; This refers to the cooling capacity; and These are the equivalent thermal fusion and equivalent thermal resistance of the data center, respectively; and These are the indoor and outdoor temperatures, respectively.

[0129] Because the dynamic component of energy consumption is affected by CPU, storage, and indoor and outdoor temperatures, exhibiting nonlinear characteristics, for clear analysis, the power consumption and cooling energy consumption of a data center are defined as follows:

[0130] Formula 9;

[0131] Formula 10;

[0132] In the formula, and These are the data center's power consumption and cooling energy consumption, respectively. and These are the fixed power consumption and fixed cooling energy consumption of the data center, respectively. and They are respectively affected by CPU resources Storage resources Indoor temperature and outdoor temperature The impact on dynamic power consumption and dynamic cooling energy consumption.

[0133] The data-driven computing power-energy consumption mapping model is characterized by a heterogeneous and nonlinear relationship between computing power tasks and energy consumption. Traditional mechanistic models, based on strong physical assumptions, struggle to precisely quantify the mapping relationship between various types of energy consumption and computing power task sequences. In contrast, data-driven methods do not rely on physical assumptions and, leveraging the learning characteristics of data, possess a powerful ability to approximate complex nonlinear functions. Therefore, this paper proposes the Transformer-Kanmogorov-Arnold Network (TKAN) computing power task-energy consumption mapping model, combining the efficient nonlinear modeling capabilities of the Kolmogorov-Arnold Network (KAN) with the complex data dependency extraction advantages of the Transformer.

[0134] The flowchart is shown in Figure 3. The Transformer-KAN (TKAN) computing power task-energy consumption mapping model is established, including the following steps;

[0135] Step 1: Data Processing and Normalization, specifically: collecting and processing computing power tasks and related data. Input data includes CPU usage. Storage and computing usage and indoor temperature and outdoor temperature The tag data includes dynamic power consumption and dynamic cooling energy consumption;

[0136] Step 2: Building and training TKAN, specifically: integrating the Transformer module and the KAN module, using the Transformer's multi-head attention mechanism to extract long-term temporal dependency features of computing power task sequences, and using the KAN network's learnable univariate activation function to approximate the heterogeneous nonlinear mapping relationship between computing power tasks and energy consumption.

[0137] Step 3: Model testing and validation, specifically: In order to effectively evaluate the generalization performance of the model, cross-validation is introduced to divide the dataset into training set and test set, and mean squared error is selected as the standard for evaluating the performance of the trained TKAN computing power task-energy consumption mapping model.

[0138] After training the TKAN computing power task-energy consumption mapping model, the dynamic energy consumption of computing power tasks was obtained through testing. This allowed for the calculation of the total energy consumption of the data center during each scheduling period, thus coupling computing power tasks with energy consumption. Based on this, the electrical and cooling energy consumption of the data center is expanded as follows:

[0139] Formula 11;

[0140] Formula 12;

[0141] In the formula, , The power consumption and cooling consumption of computing task i are denoted as .

[0142] In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, includes an objective function and constraints. The objective function is used to define the collaborative optimization model; specifically, the optimization model aims to minimize the total operating cost of the data center, covering operating costs. Carbon emission costs Service quality cost Service quality costs comprise two parts:

[0143] Delay penalty cost: When the waiting time of a computing task exceeds its execution time, a service quality loss cost is incurred.

[0144] Transfer Compensation: To compensate for the impact of scheduling changes on user service quality, the system pays transfer compensation to the data center. The compensation amount depends on the sufficiency of photovoltaic output and the electricity price level during the scheduling period, and is usually higher during peak photovoltaic output or off-peak electricity price periods. When the transfer compensation is greater than the delay penalty cost, the service quality is negative, which is an incentive; when the transfer compensation is less than the penalty cost, the service quality is positive, which is a cost. The specific expression is:

[0145] Formula 13;

[0146] Formula 14;

[0147] Formula 15;

[0148] Formula 16;

[0149]

[0150] Formula 17;

[0151] In the formula, , Let and represent the probability of occurrence of the s-th typical photovoltaic power output scenario obtained from historical data clustering, and the probability of occurrence of the typical computing task scenario, respectively. Assuming they are independent, their joint scenario probability is... ; , , , These represent the electricity purchase price, average electricity purchase price, gas purchase price, and carbon emission price, respectively. , These represent the carbon emission intensity coefficients for purchasing energy from the gas grid and the power grid, respectively. , These represent the service quality penalty factor and the transfer compensation factor, respectively. , These respectively indicate the degree of influence from photovoltaic power and electricity price guidance; , , , , These represent the purchased electricity, purchased gas, photovoltaic power output, dynamic energy consumption, and average photovoltaic power output at time k in the s-th scenario, respectively.

[0152] In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, has constraints that form the constraint condition model for constructing the collaborative optimization model. Specifically, the constraints include constraints on data center cogeneration units, electric chiller units, absorption chiller units, heat exchangers, electric energy storage, and power balance constraints; among which, the cogeneration unit constraints are:

[0153] Formula 18;

[0154] Formula 19;

[0155] In the formula, , These represent the electrical power and thermal power output of the combined heat and power generation at time k in the s-th scenario, respectively. , These represent the electrical conversion efficiency and thermal conversion efficiency of combined heat and power (CHP), respectively. Indicates the calorific value of the gas used in a combined heat and power (CHP) unit;

[0156] The constraints of the electric chiller unit are:

[0157] Formula 20;

[0158] In the formula, , These represent the power consumption and cooling capacity required for cooling at time k in the s-th scenario, respectively. For electric cooling efficiency;

[0159] The constraints of absorption chiller units are:

[0160] Formula 21;

[0161] In the formula, , These represent the heat required for absorption cooling at time k in the s-th scenario and its cooling capacity, respectively. This refers to the absorption cooling efficiency.

[0162] In step S3, the waste heat generated by the data center servers depends on the power consumption of the data center's computing tasks. The waste heat recovered through the heat exchanger is defined as follows:

[0163] Formula 22;

[0164] In the formula, Let K be the server heat recovered by the heat exchanger at time k in the s-th scenario. For waste heat recovery efficiency;

[0165] The constraints of electrical energy storage are:

[0166] Formula 23;

[0167] In the formula, For the electrical energy stored at time k in the s-th scenario; , , These are the self-discharge rate of energy storage, charging efficiency, and discharging efficiency, respectively. , These are the charging power and discharging power of the energy storage at time k-1 in the s-th scenario, respectively; , These are the minimum and maximum capacities of electrical energy storage, respectively. , These represent the charging state and the discharging state, respectively.

[0168] Since thermal energy storage and electrical energy storage have similar models, they will not be elaborated upon here.

[0169] The power balance constraint is:

[0170] Formula 24;

[0171] Formula 25;

[0172] Formula 26;

[0173] In the formula, , Let represent the data center's cold energy consumption and hot energy consumption at time k in the s-th scenario, respectively; , These represent the charging and releasing power of the thermal energy storage at time k in the s-th scenario, respectively.

[0174] The data center computing power-energy synergistic optimization model, which takes into account the computing power task dependency characteristics, is related to the computing power task sequence decomposition model and the computing power task-energy consumption coupling constraint. The computing power task sequence decomposition model is expressed by Equations 1 to 5, and the computing power task-energy consumption coupling constraint is expressed by Equations 11 and 12.

[0175] The computing power-energy co-optimization model, which takes into account the task-dependent characteristics of computing power, is expressed by the following formula:

[0176] Formula 27;

[0177] st (1-5)—(11-26) Formula 28.

[0178] When the optimization method is used, if the data center power supply includes photovoltaic power generation, then to address the uncertainty of the arrival time of computing tasks and photovoltaic output, a large number of random scenarios are first generated using Monte Carlo simulation, and then the hierarchical clustering algorithm is used to reduce the large number of scenarios to a typical N scenarios (e.g., N=10) to balance computing accuracy and efficiency; then the DAG constraints obtained in step S1 and the TKAN energy consumption prediction model trained in step S2 are embedded into the optimization model in step three.

[0179] In step one, the computing power task characteristics of Alibaba cluster data are extracted.

[0180] Solve the problem using the Gurobi commercial solver in a Python environment.

[0181] Example 1:

[0182] This embodiment uses Alibaba's public cloud data center cluster tracking data released in 2018 as the baseline data. This dataset records detailed information about various offline batch processing loads and their internal tasks running in large-scale server clusters, exhibiting a typical DAG dependency structure, which conforms to the public cloud data center architecture and computing load type characteristics targeted by this invention. After preprocessing, it is used as the data source for computing tasks in simulation verification.

[0183] To verify the effectiveness of the proposed TKAN computing power task-energy consumption mapping model, it is compared with four existing methods: Transformer model, KAN model, and Backpropagation Neural Network (BPNN), and the mean absolute error (MAE) and mean square error (MSE) are used for evaluation. Considering the potential impact of indoor and outdoor temperatures on the computing power task-energy consumption mapping model, this study sets up three feature selection scenarios for the computing power task-energy consumption model to systematically evaluate the impact of the combination of input and output variables on the model's prediction accuracy: 1) Input vector is CPU and storage resources, output vector is power consumption (CPU & RAM-Power, CR-P); 2) Input vector is CPU and storage resources, output vector is power consumption and cooling energy consumption (CPU & RAM-Power, Cool, CR-PC); 3) Input vector is CPU, storage resources, indoor temperature, and outdoor temperature, output vector is power consumption and cooling energy consumption (CPU & RAM & Temperature-Power, Cool, CRT-PC).

[0184] Table 1 shows the results of 30 independent runs of the four computational task-energy mapping models under the three training strategies mentioned above. Experimental data shows that as the completeness of input features and the output dimension increase (from CR-P to CRT-PC), the prediction errors of all models decrease significantly. Taking the TKAN model as an example, its power consumption MAE decreased from 38.42 for CR-P to 36.61 for CRT-PC, and its cold energy consumption MAE decreased from 11.38 for CR-PC to 10.99 for CRT-PC. This trend reveals the differences in temperature sensitivity among different models: when the input features are expanded from CR-PC to CRT-PC, TKAN's cold energy consumption MAE decreases by 3.5%, which is better than KAN and SVR's 3.3% and 0.3%, respectively, confirming that TKAN has a stronger feature recognition ability for cold energy and related environmental factors. Comprehensive data analysis shows that the TKAN model improves the overall accuracy by an average of 4.24%. This is because the present invention combines the advantages of Transformer in extracting long-term time-series dependencies with the efficient nonlinear modeling capabilities of KAN, which can more accurately capture the complex nonlinear coupling relationship between computing power task sequences and multiple types of energy consumption (electricity, cooling) under temperature changes, thereby providing more accurate energy consumption boundary conditions for collaborative optimization.

[0185] Table 1. Impact of different training strategies on the mapping model (average of 30 runs)

[0186]

[0187] In this study's model, thermal energy consumption represents the conventional campus heating demand, unaffected by computing power tasks. Cooling energy consumption, however, follows a similar trend to electrical energy consumption and is influenced by the decomposition of computing power task sequences. Therefore, this section analyzes the impact of computing power task sequence decomposition on flexible electrical energy consumption. To this end, three different comparison strategies are set up to evaluate the optimization potential of the computing power task sequence decomposition model for the adjustable energy consumption of data centers: 1) an optimal scheduling model of data center without considering computing power task sequence decomposition (N-OMDC); 2) a collaborative optimization model of data center computing power-energy without considering DAG dependency (N-COMDC); and 3) a collaborative optimization model of data center computing power-energy considering DAG dependency, which is the model presented in this paper.

[0188] To effectively assess the power consumption for data center flexibility, flexibility requirement assessment metrics are set as follows:

[0189] Formula 29;

[0190] In the formula, This represents the initial energy consumption at time point k. The energy consumption after coordinated scheduling over time period k; For time period k, the flexible energy consumption is represented by the positive or negative sign, which indicates the direction of adjustment, representing upward or downward adjustment of energy consumption space, respectively.

[0191] Figure 4 shows a comparison of the flexible power consumption of data centers under the three models. The black curve represents the original power consumption curve of N-OMDC; the pink and gray areas represent the adjustable range of flexible power consumption for N-COMDC and the model presented in this paper, respectively. Compared to N-OMDC, N-COMDC significantly expands the schedulable range of computing tasks because it does not consider the dependency characteristics between computing tasks, i.e., the deadline for computing tasks is uniformly set to 24:00. This results in approximately 80% or more of the computing tasks being concentrated in periods of low electricity prices or high photovoltaic output, specifically 11:00-19:00 and 22:00-24:00. Therefore, N-COMDC exhibits a significant maximum upward adjustment capability (814.18 kW, equivalent to 137.7% of the baseline average power consumption) and a maximum downward adjustment capability (513.60 kW, equivalent to 86.9% of the baseline average power consumption). However, this model, by ignoring the dependencies between computing tasks and exaggerating the flexibility of computing tasks, limits its effectiveness and applicability in real-world scenarios. In contrast, the model presented in this paper exhibits varying degrees of energy consumption regulation potential throughout the day, with particularly significant regulation ranges during the 0:00-7:00 and 9:00-19:00 time periods. Its maximum upward and downward regulation capabilities are 141.89 kW (equivalent to 24% of the baseline average load) and 157.24 kW (equivalent to 26.6% of the baseline average energy consumption), respectively. This performance stems from the model's ability to optimize scheduling within the schedulable time frame of computing tasks, effectively promoting the time-based transfer of energy consumption. Furthermore, by introducing quality of service costs, the model in this paper achieves significant energy consumption regulation effects during the 0:00-7:00 and 9:00-19:00 time periods.

[0192] To analyze the impact of computing power task sequence decomposition on the synergistic optimization of data center computing power and energy, optimization scheduling analysis was conducted on the three comparative schemes mentioned above. Table 2 shows the optimization scheduling results of the three models. As shown in Table 2, the model in this paper has the lowest total operating cost, which is 6.7% and 5.8% lower than N-OMDC and N-COMDC, respectively. Specifically, compared with N-OMDC, the energy cost and carbon cost of the model in this paper are reduced by RMB 320.18 and RMB 94.17, respectively, while obtaining RMB 2424.31 in computing power service incentives. This advantage stems from the fact that, under the premise of satisfying DAG dependency constraints, the model in this paper schedules computing power to high photovoltaic or low electricity price periods through task time-shifting strategies, achieving synergistic benefits of energy cost optimization, green electricity consumption improvement, and service quality assurance. Although the model in this paper increases energy cost by RMB 768.86 and carbon cost by RMB 32.43 compared with N-COMDC, it also obtains RMB 2424.31 in computing power service incentives. While N-COMDC increases the flexibility of computing power scheduling through a unified task deadline strategy, it increases service quality costs by 1364.83 yuan due to violations of DAG dependency constraints, making it difficult to meet the actual needs of industrial scenarios. In summary, the computing power-energy co-optimization method based on DAG constraints achieves comprehensive optimization of economy and low carbon emissions while ensuring service quality by coordinating the dynamic matching of computing power tasks and energy systems.

[0193] Table 2 Optimization results of N-OMDC, N-COMDC, and the model presented in this paper.

[0194]

[0195] Figure 5 compares the power balance states of the N-OMDC and the optimized model presented in this paper. In the N-OMDC, the purchased electricity and gas volumes are 31990.31 kWh and 4386.12 m³, respectively; while in the model presented in this paper, the purchased electricity and gas volumes are 31644.12 kWh and 4426.12 m³, respectively. Compared to the N-OMDC, the purchased electricity volume in the model presented in this paper is reduced by 346.19 kWh, and the purchased gas volume is reduced by 40.0 m³, representing reductions of 1.08% and 0.91%, respectively. The change in energy consumption structure is mainly attributed to the following factors: during periods of sufficient photovoltaic output (10:00–15:00), the amount of computing power tasks is increased, thereby reducing dependence on the upstream power grid. At the same time, the energy consumption for electric cooling increases by 50.79 kWh during this period. Conversely, during peak grid electricity price periods (19:00–22:00), the energy consumption for electric cooling decreases by 85.12 kWh. This is mainly due to the implementation of the proactive computing task time shifting strategy. By shifting high-energy-consuming computing tasks to periods when photovoltaic output is abundant, not only is electricity procurement avoided during periods of high electricity prices, but the photovoltaic absorption rate is also improved, ultimately optimizing the total operating cost.

Claims

1. A data center computing power-energy co-optimization method that takes into account task timing flexibility, characterized in that: The method described is used for server clusters that support offline batch processing loads. It optimizes the synergistic optimization between data center computing power and energy by tapping the time-shifting potential of computing tasks. The process includes the following steps: Step S1: For the offline batch processing computing load of computing power tasks, construct a computing power task sequence decomposition model based on a directed acyclic graph (DAG) to describe the temporal dependencies within the offline batch processing load; Step S2: Based on Step 1, set a computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold neural network (TKAN); Step S3: Based on Step 2, set a data center computing power-energy co-optimization model that takes into account the computing power task dependency characteristics.

2. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 1, characterized in that: The data center is a public or private cloud data center server cluster containing multiple computing nodes. The server cluster includes multiple server nodes connected by a network. The computing power task is an offline batch processing computing power load with time-series dependencies. The offline batch processing computing power load includes AI training and big data processing. The computing power task load consists of multiple computing power tasks, which are distributed to different server nodes in the server cluster for execution.

3. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 1, characterized in that: In step S1, the process of constructing the computing power task sequence decomposition model specifically includes: 1) Analysis of the dependency characteristics between computing power tasks, specifically: based on the temporal dependency characteristics between tasks, that is, some subsequent computing power tasks need to read the results of previous computing power tasks, forming a unidirectional, loop-free structure, establishing the execution order between computing power tasks, establishing time constraints for the computing power task sequence, deriving the execution time and relaxation time of each computing power task, and then, based on the latest completion time of the computing power load, service quality, operating cost and other objectives, the data center sets the start time and acceptable deadline time of each computing power task. Finally, based on the above time constraints and resource requirements, the task scheduler allocates computing power tasks to servers with sufficient computing resources for execution, and the servers then return the results to the users; 2) Modeling the computing power task sequence decomposition: transforming the high-dimensional, large-scale load mixed integer programming problem into a linear programming problem adapted to computing power-energy co-optimization. The specific modeling process is as follows: Step 1: Extract the computing power task characteristics of the server cluster data, including the computing power load name. Task Name Task start time Task end time CPU usage of the task Task storage usage Step 2: Constructing the DAG model; First, integrate the data extracted in Step 1. and Subsequently, based on the DAG theoretical framework, a DAG model reflecting the dependency characteristics between computing power tasks is established to accurately characterize the topological order and execution priority between different task nodes; Step 3: Quantify the time adjustability of computing power tasks. The time adjustability of computing power tasks is related to their relaxation time and execution time; execution time The definition is as follows: Formula 1; where, and respectively computing power tasks End time and start time; time adjustability The definition is as follows: Formula 2; where, This indicates that computing task i has no dependent tasks; Indicates task Tasks dependent on computing power This means that the computing power task needs to be completed. Only then can the computing power task begin to be executed. ； This indicates that each computing task in the computing task set I depends on a computing task. Step 4: Constructing Timing Constraints for Computing Tasks: To ensure that the execution of computing tasks falls between the start and end times, the timing constraints for computing task execution are defined as follows: Formula 3; where, This represents the execution ratio of computing task i in scenario s during time period k. ； The start time of the time period to which the computing task belongs; scheduling period. To effectively assess the latency of computing tasks, the actual start time of the computing task is... and waiting time for: Official 4; Formula 5; Formula 3 ensures that the computing task is executed within the time interval k between the start and end times, and also ensures that the task execution time meets the requirements. As required, Formulas 4 and 5 define the actual start time and waiting time of the computing task.

4. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 3, characterized in that: In step S2, the computing power task-energy consumption mapping model based on Transformer and Kolmogorov-Arnold neural network specifically includes a conventional data center energy consumption model and a data-driven computing power-energy consumption mapping model.

5. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 4, characterized in that: A typical data center energy consumption model includes electrical energy consumption and cooling energy consumption. Based on the computing power and task processing methods, it is divided into fixed energy consumption and dynamic energy consumption. Fixed energy consumption is the basic requirement for maintaining data center operation, while the dynamic part is related to the usage of CPU and storage, and is also affected by indoor and outdoor temperatures. The total power consumption of the server... Mainly affected by CPU utilization With indoor air temperature The effect is shown in the following formula: Official 6; Formula 7; where: and These are idle power and peak power, respectively. The temperature fitting coefficient is used; the data center maintains the temperature within the required range through a cooling system, and the corresponding thermal dynamic process is represented by a thermodynamic equivalent thermal parameter model as follows: Formula 8; where, The heat dissipation power of the data center is expressed as ； This refers to the cooling capacity; and These are the equivalent thermal fusion and equivalent thermal resistance of the data center, respectively; and Indoor and outdoor temperatures, respectively; the power consumption and cooling energy consumption of a data center are defined as follows: Official 9; Formula 10; where, and These are the data center's power consumption and cooling energy consumption, respectively. and These are the fixed power consumption and fixed cooling energy consumption of the data center, respectively. and They are respectively affected by CPU resources Storage resources Indoor temperature and outdoor temperature The impact on dynamic power consumption and dynamic cooling energy consumption.

6. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 5, characterized in that: The data-driven computing power-energy consumption mapping model is as follows: There is a heterogeneous and nonlinear relationship between computing power tasks and energy consumption. A converter-Kolmogorov-Arnold neural network computing power task-energy consumption mapping model is established, including the following steps; Step 1: Data Processing and Normalization, specifically: collecting and processing computing power tasks and related data. Input data includes CPU usage. Storage and computing usage and indoor temperature and outdoor temperature The labeled data includes dynamic power consumption and dynamic cooling energy consumption; Step 2: TKAN construction and training, specifically: integrating the Transformer module and the KAN module, using the Transformer's multi-head attention mechanism to extract long-term temporal dependency features of the computing power task sequence, and using the KAN network's learnable univariate activation function to approximate the heterogeneous nonlinear mapping relationship between computing power tasks and energy consumption; Step 3: Model testing and validation, specifically: to effectively evaluate the model's generalization performance, cross-validation is introduced to divide the dataset into training and testing sets, and mean squared error is used as the standard for evaluating the performance of the trained TKAN computing power task-energy consumption mapping model; after completing the TKAN computing power task-energy consumption mapping model training, the dynamic energy consumption of computing power tasks is obtained through testing, and then the total energy consumption of the data center in each scheduling period is calculated, realizing the coupling of computing power tasks and energy consumption. Based on this, the power and cooling energy consumption of the data center are extended to the following form: Official 11; Formula 12; where, 、 The power consumption and cooling consumption of computing task i are denoted as .

7. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 6, characterized in that: In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, includes an objective function and constraints. The objective function is used to define the collaborative optimization model; specifically, the optimization model aims to minimize the total operating cost of the data center, covering operating costs. Carbon emission costs Service quality cost Service quality cost includes two parts: latency penalty cost: when the waiting time of a computing task exceeds its execution time, a service quality loss cost is incurred; transfer compensation: to compensate for the impact of changes in the scheduling plan on the user's service quality, the system pays transfer compensation to the data center. The compensation amount depends on the sufficiency of photovoltaic output and the electricity price level during the dispatch period. When the transfer compensation is greater than the delay penalty cost, the service quality is negative, which is an incentive; when the transfer compensation is less than the penalty cost, the service quality is positive, which is a cost. The specific expression is: Official 13; Official 14; Official 15; Official 16; Formula 17; where, 、 and represent the probability of occurrence of the s-th typical photovoltaic power output scenario and the probability of occurrence of the typical computing power task scenario, respectively, obtained by clustering historical data. Assuming the two are independent, their joint scenario probability is: ； 、 、 、 These represent the electricity purchase price, average electricity purchase price, gas purchase price, and carbon emission price, respectively. 、 These represent the carbon emission intensity coefficients for purchasing energy from the gas grid and the power grid, respectively. 、 These represent the service quality penalty factor and the transfer compensation factor, respectively. 、 These respectively indicate the degree of influence from photovoltaic power and electricity price guidance; 、 、 、 、 These represent the purchased electricity, purchased gas, photovoltaic power output, dynamic energy consumption, and average photovoltaic power output at time k in the s-th scenario, respectively.

8. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 7, characterized in that: In step S3, the data center computing power-energy collaborative optimization model, which takes into account the computing power task dependence characteristics, has constraints that form the constraint condition model for constructing the collaborative optimization model. Specifically, the constraints include constraints on data center cogeneration units, electric chiller units, absorption chiller units, heat exchangers, electric energy storage, and power balance constraints; among which, the cogeneration unit constraints are: Official 18; Formula 19; where, 、 These represent the electrical power and thermal power output of the combined heat and power generation at time k in the s-th scenario, respectively. 、 These represent the electrical conversion efficiency and thermal conversion efficiency of combined heat and power (CHP), respectively. This indicates the calorific value of the gas used in a combined heat and power (CHP) unit; the constraints for an electric chiller unit are: Formula 20; where, 、 These represent the power consumption and cooling capacity required for cooling at time k in the s-th scenario, respectively. For electric cooling efficiency; the constraints for absorption chiller units are: Formula 21; where, 、 These represent the heat required for absorption cooling at time k in the s-th scenario and its cooling capacity, respectively. This refers to the absorption cooling efficiency.

9. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 8, characterized in that: In step S3, the waste heat generated by the data center servers depends on the power consumption of the data center's computing tasks. The waste heat recovered through the heat exchanger is defined as follows: Formula 22; where, Let K be the server heat recovered by the heat exchanger at time k in the s-th scenario. For waste heat recovery efficiency; the constraints of electrical energy storage are: Formula 23; where, For the electrical energy stored at time k in the s-th scenario; 、 、 These are the self-discharge rate of energy storage, charging efficiency, and discharging efficiency, respectively. 、 These are the charging power and discharging power of the energy storage at time k-1 in the s-th scenario, respectively; 、 These are the minimum and maximum capacities of electrical energy storage, respectively. 、 These represent the charging and discharging states, respectively; the power balance constraint is: Official 24; Official 25; Formula 26; where, 、 Let represent the data center's cold energy consumption and hot energy consumption at time k in the s-th scenario, respectively; 、 Let represent the charging and discharging power of the thermal energy storage at time k in the s-th scenario, respectively. The data center computing power-energy collaborative optimization model, which considers the computing power task dependency characteristics, is related to the computing power task sequence decomposition model and the computing power task-energy consumption coupling constraints. The computing power task sequence decomposition model is expressed by Equations 1 to 5, the computing power task-energy consumption coupling constraints are expressed by Equations 11 and 12, and the computing power-energy collaborative optimization model, which considers the computing power task dependency characteristics, is expressed by the following formula: Formula 27; st (1-5)-(11-26) Formula 28.

10. The data center computing power-energy co-optimization method considering task timing flexibility according to claim 8, characterized in that: When the optimization method is used, if the power supply of the data center includes photovoltaic power generation, then to address the uncertainty of the arrival time of computing tasks and photovoltaic output, a large number of random scenarios are first generated by Monte Carlo simulation, and then the hierarchical clustering algorithm is used to reduce the large number of scenarios into N typical scenarios; then the DAG constraints obtained in step S1 and the TKAN energy consumption prediction model trained in step S2 are embedded into the optimization model in step three; in step one, the computing task features of Alibaba cluster data are extracted.

Images