A low-carbon and high-efficiency scheduling method and system for a data center based on calculation of carbon three-flow coupling optimization

Abstract

This application discloses a low-carbon and efficient data center scheduling method and system based on the coupling optimization of computing, power, and carbon flows. By collecting computing power flow, power flow, and carbon flow data and establishing a three-flow coupling model, a multi-objective optimization scheduling model is constructed with the goals of maximizing green electricity consumption, minimizing peak-valley load differences, minimizing carbon emissions, minimizing operating costs, and maximizing computing power reliability. A hybrid solution is achieved using an improved NSGA-III algorithm combined with deep reinforcement learning, generating coordinated scheduling instructions for computing power, power, and carbon emissions. This realizes integrated optimization of data center computing power scheduling, green electricity consumption, grid peak shaving and valley filling, and low-carbon operation. The system consists of a three-flow perception layer, a coupled optimization decision layer, a collaborative execution layer, and a visualization and control layer. It can significantly improve green electricity utilization, reduce peak-valley differences and carbon emissions, while simultaneously reducing electricity costs, and is suitable for various low-carbon and efficient data center scheduling scenarios.

Description

Technical Field

[0001] This application relates to the field of data center energy management and intelligent scheduling technology, specifically to a low-carbon and efficient scheduling method and system for data centers based on the optimization of the coupling of computing, electricity, carbon, and other three flows. Background Technology

[0002] With the rapid development of cloud computing, big data and artificial intelligence industries, the scale and computing power demand of data centers are growing exponentially, and their power consumption and carbon emissions have become the core bottlenecks restricting the sustainable development of the industry.

[0003] Current data center scheduling models suffer from a disconnect between computing power, electricity, and carbon emissions, lacking a unified coupled optimization mechanism that prevents the global optimal allocation of resources. Furthermore, due to the random fluctuations in green electricity output from wind and solar power and the spatiotemporal mismatch with data center loads, the capacity for green electricity absorption is generally insufficient, and the adjustment potential of IT loads, energy storage, and cooling systems has not been fully explored to achieve coordinated peak shaving and valley filling of the power grid. In addition, traditional scheduling is often limited to optimizing single objectives such as electricity costs or computing efficiency, making it difficult to simultaneously address multi-dimensional needs such as low carbon emissions, economy, and reliability.

[0004] Although existing technologies involve computing-electricity coordination or carbon-electricity scheduling, none of them have established a multi-objective optimization model with deep coupling of three flows, lack an integrated control mechanism for green electricity consumption and peak shaving and valley filling, and cannot meet the requirements of real-time adaptive scheduling. Summary of the Invention

[0005] This application provides a low-carbon and efficient scheduling method and system for data centers based on the optimization of the coupling of computing, electricity, carbon, and other three flows. This addresses the problems in existing technologies, such as the lack of a multi-objective optimization model with deep coupling of the three flows, the lack of an integrated control mechanism for green electricity consumption and peak shaving and valley filling, and the inability to meet the requirements of real-time adaptive scheduling.

[0006] This application provides a low-carbon and efficient scheduling method for data centers based on the optimization of the coupling of computing, power, and carbon flows, including step S1: collecting computing power flow data, power flow data, and carbon flow data of the data center, performing normalization, noise reduction, and spatiotemporal alignment preprocessing, and constructing a computing, power, and carbon flow database;

[0007] Step S2: Construct a computing power-electricity coupling model and an electricity-carbon emission coupling model, and establish a multi-dimensional constraint set that includes power balance constraints, computing power service level agreement constraints, energy storage charging and discharging constraints, green electricity consumption constraints, carbon emission upper limit constraints, and grid peak-valley interaction constraints, to form a computing power-carbon three-flow coupling model;

[0008] Step S3: Construct a five-dimensional multi-objective optimization scheduling model with the objectives of maximizing green electricity consumption rate, minimizing load peak-valley difference, minimizing carbon emissions, minimizing operating costs, and maximizing computing power service level;

[0009] Step S4: The multi-objective optimization scheduling model is solved by using a hybrid optimization strategy that combines the improved NSGA-III algorithm with DRL to generate the Pareto optimal scheduling solution set;

[0010] Step S5: Determine the weights of each objective based on the analytic hierarchy process, select the globally optimal scheduling scheme from the Pareto optimal scheduling solution set, and decompose it into computing power scheduling instructions, power scheduling instructions, and carbon control instructions for coordinated control and execution in the data center.

[0011] Preferably, in step S2, the computing power-electricity coupling model is expressed as follows: ,in, Real-time power consumption of IT equipment For computing load rate, For server energy efficiency, This refers to the intake air temperature.

[0012] Preferably, the power-carbon emission coupling model in step S2 is expressed as follows: ,in, For carbon emission rate, , , These represent the charging power for mains power, green electricity, and energy storage, respectively. For real-time grid carbon factor, This represents the carbon emission coefficient for energy storage.

[0013] Preferably, the improved NSGA-III algorithm in step S4 includes:

[0014] Introduce a uniformly distributed reference point mechanism;

[0015] The target space adaptive normalization method is adopted to adaptively scale the values ​​of five objective functions, namely green energy absorption rate, load peak-valley difference, carbon emissions, operating costs and computing power SLA, by calculating ideal points and extreme points, thereby eliminating the difference in dimensions and magnitudes.

[0016] A reference point-based association selection strategy is adopted to prioritize the retention of individuals associated with sparse reference points, thereby maintaining population diversity.

[0017] Integrated constraint-compatible optimization uses constraint dominance rules to handle infeasible solutions that violate operational boundary conditions.

[0018] Preferably, the DRL in step S4 adopts a near-end strategy optimization algorithm, forming a hybrid optimization strategy with the improved NSGA-III algorithm.

[0019] Preferably, in step S5, the specific steps for determining the weights of each objective using the analytic hierarchy process include:

[0020] A three-tiered structure is constructed, comprising an objective layer, a criterion layer, and a solution layer, wherein the criterion layer corresponds to the five optimization objectives described above.

[0021] The judgment matrix of the criterion layer is constructed based on the Saaty 1-9 scaling method;

[0022] Calculate the largest eigenvalue and the corresponding eigenvector of the judgment matrix, and obtain the weight coefficients of each optimization objective after normalization.

[0023] Preferably, after step S1, a time-series prediction step based on a long short-term memory network is also included.

[0024] The present invention also provides an efficient scheduling system for implementing the method, including a computing, power, and carbon emission sensing layer: comprising a computing power monitoring unit, a power monitoring unit, and a carbon emission monitoring unit;

[0025] Coupled optimization decision layer: includes a three-flow coupled modeling module for constructing a computing-electricity-carbon correlation model, a multi-objective scheduling module that integrates the improved NSGA-III and DRL hybrid optimization algorithms, and a prediction module for predicting future scheduling parameters;

[0026] Collaborative Execution Layer: Includes computing power scheduling unit, power control unit, and refrigeration linkage unit for coordinated regulation of the refrigeration system;

[0027] Visualized control layer: Includes a three-flow monitoring panel and alarm module for displaying system operating status, scheduling effect, carbon footprint and energy consumption reports.

[0028] Preferably, the smart meter in the power monitoring unit and the carbon emission metering module in the carbon emission monitoring unit are integrated into a single carbon meter, which is used to collect both power data and carbon emission data.

[0029] Compared with existing technologies, the low-carbon and high-efficiency scheduling method and system for data centers based on computation, power, carbon, and three-flow coupling optimization provided in this application has at least the following beneficial effects:

[0030] 1. This invention breaks down information silos by using a unified computing power-electricity-carbon emission coupling model, achieving coordinated optimization of the three flows, thereby increasing the comprehensive utilization rate of resources by more than 25%, and improving the utilization rate of computing power resources and energy storage.

[0031] 2. By utilizing LSTM time-series prediction and multi-objective scheduling, the green electricity utilization rate of data centers is significantly increased from about 60% in the traditional model to over 90%, and the wind and solar curtailment rate is reduced to below 5%.

[0032] 3. The load peak-valley difference is reduced, allowing for deeper participation in grid peak-shaving ancillary services and increased annual revenue.

[0033] 4. Annual carbon emissions are reduced by more than 35%, and operating electricity costs are reduced by 15%-20%. In a simulation of a 10MW data center, the annual carbon reduction reaches 12,000 tons, and the annual electricity cost is saved by approximately 2.8 million yuan.

[0034] 5. By improving the non-dominated sorting and reference point mechanism of NSGA-III, the Pareto front is ensured to be uniformly distributed in the five-dimensional target space, avoiding solution set aggregation or front collapse. Users can flexibly choose the scheduling scheme according to actual needs.

[0035] In summary, by collecting data on computing power flow, power flow, and carbon flow and establishing a coupled model of these three flows, a multi-objective optimization scheduling model is constructed. This model aims to maximize green electricity consumption, minimize peak-valley load differences, minimize carbon emissions, minimize operating costs, and maximize computing power reliability. A hybrid solution is achieved using an improved NSGA-III algorithm combined with deep reinforcement learning, generating coordinated scheduling instructions for computing power, power, and carbon emissions. This enables integrated optimization of data center computing power scheduling, green electricity consumption, grid peak shaving and valley filling, and low-carbon operation. The system consists of a three-flow perception layer, a coupled optimization decision layer, a collaborative execution layer, and a visualization and control layer. It can significantly improve green electricity utilization, reduce peak-valley differences and carbon emissions, while simultaneously lowering electricity costs. It is suitable for various low-carbon and efficient scheduling scenarios in data centers. Attached Figure Description

[0036] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0037] Figure 1 This invention provides an architecture diagram of a computational, carbon, three-flow coupling optimization scheduling system.

[0038] Figure 2 Flowchart of the multi-objective optimization scheduling method provided by this invention;

[0039] Figure 3 The flowchart of the improved NSGA-III algorithm provided by this invention; Detailed Implementation

[0040] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0041] Example 1:

[0042] A low-carbon and efficient scheduling method for data centers based on the optimization of computation, power, carbon, and three-stream coupling includes the following steps:

[0043] Step S1: Collect computing power flow data, power flow data and carbon flow data from the data center, perform normalization, noise reduction and spatiotemporal alignment preprocessing, and construct a computing power carbon flow database;

[0044] 1.1 Collect computing power flow data: Collect IT equipment computing power load, task type tags, server running status, number of queued tasks, and expected computing power requirements through server BMC or virtualization platform API.

[0045] 1.2. Power flow data collection: Collect mains power through smart meters; collect green power output curves through the communication interface of photovoltaic / wind power inverters; collect SOC and charging / discharging power through energy storage BMS; collect PUE and cooling system power consumption through data center infrastructure monitoring system.

[0046] 1.3. Collect carbon flow data: Obtain hourly marginal carbon factors of the power grid from the regional power grid carbon emission factor database; confirm the zero-carbon attribute of green electricity based on green certificates or direct power purchase agreements; obtain equipment carbon emission coefficients with reference to the provincial greenhouse gas inventory compilation guidelines; and set monthly / daily quota constraints based on the company's annual carbon quota.

[0047] 1.4 Preprocessing:

[0048] Normalization is performed by mapping each parameter to the [0,1] interval using the Min-Max method;

[0049] Denoising is achieved by using a sliding median filter for transient spike noise.

[0050] Spatiotemporal alignment: All data are uniformly resampled to a 1-hour interval to construct a computational carbon three-flow database.

[0051] Preferably, after step S1, a time-series prediction step based on a long short-term memory network is also included. The LSTM time-series prediction module is used to mine the time-series dependency features of historical three-stream data to achieve high-precision prediction of computing load, green power output, grid electricity price, and carbon factor in the next 24 hours, providing forward-looking boundary conditions for NSGA-III global optimization and reducing scheduling deviations caused by random fluctuations.

[0052] In actual use, the input features are: historical 72 hours, 12-dimensional features (computing load rate, proportion of latency-sensitive tasks, photovoltaic output, wind power output, grid electricity price, energy storage SOC, cooling power consumption, air intake temperature, total server power consumption, grid carbon factor, external temperature, and hourly code).

[0053] Network structure: Input layer (72×12) → First LSTM layer (128 units, used to extract shallow temporal features) → Second LSTM layer (64 units, used to extract deep long-term dependency features) → Dropout layer (inactivation rate 0.2) → Output layer (4 neurons, corresponding to the computing power load, total green electricity output, time-of-use electricity price, and grid carbon factor in the next 24 hours).

[0054] Training parameters: Optimizer Adam, learning rate 0.001; loss function MSE; batch size 32; maximum iterations 100 epochs; early stopping strategy, training stops if the validation set loss does not decrease for 10 consecutive epochs.

[0055] Predictive application: Roll out the prediction of data for the next 24 hours every hour, and use it as the input boundary for optimizing the model in step S3.

[0056] Step S2: Construct a computing power-electricity coupling model and an electricity-carbon emission coupling model. Establish a multi-dimensional constraint set that includes power balance constraints, computing power service level agreement constraints, energy storage charging and discharging constraints, green electricity consumption constraints, carbon emission ceiling constraints, and grid peak-valley interaction constraints, forming a computing-electricity-carbon three-flow coupling model; establish the mapping relationship between computing power load and electricity consumption. ,in, Real-time power consumption of IT equipment For computing load rate, For server energy efficiency, The inlet air temperature is used as the reference point; an electro-carbon conversion model is constructed based on carbon emission flow theory. ,in, For carbon emission rate, , , These represent the charging power for mains power, green electricity, and energy storage, respectively. For real-time grid carbon factor, This represents the carbon emission coefficient for energy storage.

[0057] Step S3: Construct a five-dimensional multi-objective optimization scheduling model with the objectives of maximizing green electricity consumption rate, minimizing load peak-valley difference, minimizing carbon emissions, minimizing operating costs, and maximizing computing power service level;

[0058]

[0059] The total scheduling duration (e.g., 24 hours). This is the scheduling time step.

[0060] The constraints include:

[0061] Computing power constraints: Delayed task transfer constraints

[0062] Electrical constraints: Energy storage SOC ∈ [20%, 80%] For energy storage discharge power, Power for charging energy storage.

[0063] Carbon constraints:

[0064] Grid constraints: peak load limits and valley load limits.

[0065] Step S4: The multi-objective optimization scheduling model is solved using a hybrid optimization strategy combining the improved NSGA-III algorithm and DRL to generate the Pareto optimal scheduling solution set. The improved non-dominated sorting genetic algorithm (NSGA-III) is a third-generation high-dimensional multi-objective evolutionary optimization algorithm designed for optimization problems with three or more objective dimensions. It relies on the reference point guidance mechanism, non-dominated hierarchical strategy, and adaptive normalization method to take into account the convergence, uniformity of distribution, and diversity of optimization solutions. It is suitable for multi-objective scheduling optimization scenarios under the coupling of computing, electricity, and carbon in data centers and can efficiently solve the globally optimal scheduling solution set under multiple conflict objectives.

[0066] Compared to traditional algorithms, the main improvements of the improved NSGA-III algorithm include:

[0067] 4.1 Introducing a uniform reference point mechanism: Abandoning the drawbacks of random optimization in traditional algorithms, fixed reference points are pre-generated in the target space according to uniform distribution rules. These reference points serve as optimization guides, allowing the iteratively generated scheduling solutions to uniformly approximate the entire Pareto front. This avoids the problems of solution aggregation, uneven distribution, and front collapse under high-dimensional objectives, ensuring that various optimization objectives can be balanced and taken into account, and adapting to multiple dimensions of objectives such as green electricity consumption, carbon emissions, electricity costs, peak shaving and valley filling, and computing power reliability.

[0068] 4.2 Adaptive normalization of the target space: To address the problem of inconsistent dimensions and large differences in numerical magnitude among the optimization targets of computation, electricity, and carbon, the target space is adaptively scaled and normalized by calculating ideal points and extreme points. This eliminates optimization bias caused by differences in dimensions and magnitudes, ensuring fair optimization of each target at the same scale.

[0069] 4.3 Reference point-based association selection strategy: Adopt the association matching rule between individuals and reference points to associate each optimized individual with the nearest reference point, give priority to retaining individuals corresponding to sparse reference points, accurately maintain the uniform distribution of the solution set, avoid missing high-quality scheduling solutions, and ensure that the final Pareto optimal solution set is fully covered and evenly distributed, which facilitates the selection of suitable scheduling schemes according to actual operating conditions.

[0070] 4.4 Constraint Compatibility Optimization: Adapts to multiple coupled constraints such as data center computing power load constraints, power balance constraints, energy storage SOC upper and lower limit constraints, power grid peak shaving and valley filling constraints, and carbon emission quota constraints. It processes infeasible solutions through constraint dominance rules, eliminates scheduling schemes that violate operating boundary conditions, and ensures that the output optimized solution conforms to the actual engineering operation requirements and is feasible for implementation.

[0071] Table 1 Comparison of NSGA-III with other multi-objective optimization algorithms

[0072]

[0073] This invention employs a hybrid optimization strategy of "improved NSGA-III + DRL," combining the advantages of both algorithms to further enhance the optimization efficiency and real-time performance of multi-objective scheduling coupled with computing, power, and carbon. Specifically, the improved NSGA-III algorithm serves as the core of global optimization, responsible for finding the Pareto optimal scheduling solution set under multi-objective conflicts, laying the foundation for global optimization. Deep reinforcement learning (DRL) serves as the core of local optimization and real-time adaptation, utilizing its powerful environmental awareness and adaptive adjustment capabilities to perform secondary optimization on the Pareto solution set output by NSGA-III, while adapting to the real-time fluctuations of computing, power, and carbon parameters, achieving a dual optimization effect of "global optimality + real-time adaptation."

[0074] Compared to the single NSGA-III algorithm, this hybrid strategy has three core advantages: First, it compensates for the lack of real-time response capability of NSGA-III, and can quickly adapt to the dynamic fluctuations of computing, power, carbon, and other parameters, ensuring the real-time performance of the scheduling strategy; second, through local fine-tuning of DRL, it further improves the adaptability and optimization accuracy of the scheduling solution, making the scheduling scheme more in line with actual operating conditions; third, it achieves the synergy of global optimization and local adaptation, ensuring both long-term global optimality and meeting short-term real-time control requirements, perfectly adapting to the dynamic scheduling scenario of computing, power, carbon, and other parameters coupled in data centers.

[0075] Step S5: Determine the weights of each objective based on the analytic hierarchy process, select the globally optimal scheduling scheme from the Pareto optimal scheduling solution set, and decompose it into computing power scheduling instructions, power scheduling instructions, and carbon control instructions for coordinated control and execution in the data center.

[0076] For the Pareto optimal solution set output by the improved NSGA-III algorithm, the Analytic Hierarchy Process (AHP) is used to quantify the relative importance of the five optimization objectives, determine the scientific weight coefficients, and select the globally optimal scheduling scheme that adapts to the actual operation requirements of the data center from the non-dominated solution set. The complete implementation steps are as follows:

[0077] 5.1 Constructing a three-level hierarchical structure of AHP

[0078] Based on the core requirements of computational, electrical, carbon, and three-flow coupled scheduling, a three-level structure of target layer, criterion layer, and scheme layer is constructed, with the hierarchical relationship completely corresponding to the optimization objective of this patent:

[0079] 1. Objective Layer: Screening of Optimal Scheduling Schemes for Multi-Objective Coupling of Computing, Power, Carbon, and Flow in Data Centers

[0080] 2. Criterion Layer: Corresponding to the 5 core optimization objectives of this patent, in order of their original numbers:

[0081] B1: Maximize the green electricity consumption rate (corresponding to the original optimization objective F1)

[0082] B2: Minimize the peak-to-valley load difference (corresponding to the original optimization objective F2)

[0083] B3: Minimize carbon emissions (corresponding to the original optimization objective F3)

[0084] B4: Minimize operating costs (corresponding to the original optimization objective F4)

[0085] B5: Maximize the Service Level Agreement (SLA) for computing power (corresponding to the original optimization target F5)

[0086] 3. Scheme layer: Improve the Pareto optimal solution set output by the NSGA-III algorithm to include all alternative scheduling schemes.

[0087] 5.2 Constructing the criterion-level judgment matrix

[0088] 5.2.1 Using the internationally accepted Saaty 1-9 scale method, the relative importance of the five criteria is compared pairwise to construct a 5th order judgment matrix A=(aij)5×5, where aij represents the importance scale of criterion Bi relative to criterion Bj.

[0089] Table 2 Scale Definition

[0090]

[0091] Construct a judgment matrix (the scale can be dynamically adjusted according to the data center operation priority):

[0092]

[0093] Matrix judgment logic explanation: Reliability based on computing power With carbon emission control as the core premise Green electricity consumption With dual carbon as the core objective, and operating costs Based on operation, peak-valley difference regulation To achieve the goal of grid-friendly collaboration, it aligns with the industry-standard priority of low-carbon scheduling for data centers.

[0094] 5.3 Calculate the target weight vector

[0095] The root method is used to calculate the largest eigenvalue of the judgment matrix and its corresponding weight vector. The calculation steps are as follows:

[0096] 5.3.1 Calculate the geometric mean of all elements in each row of the judgment matrix:

[0097]

[0098] 5.3.2. Normalize the geometric mean to obtain the weight vector for each objective:

[0099]

[0100] in, For the first The weight coefficients of each optimization objective satisfy... .

[0101] Substituting these values ​​into the judgment matrix yields the final weights for the five optimization objectives.

[0102] Table 3 Final Weight Results

[0103]

[0104] 5.4 Consistency Check

[0105] To avoid logical contradictions in the judgment matrix, a consistency check is required to ensure the weight results are scientifically valid. The check steps are as follows:

[0106] 5.4.1 Calculate the largest eigenvalue of the judgment matrix. :

[0107]

[0108] in, To determine the first result after multiplying the matrix and the weight vector Each element.

[0109] 5.4.2 Calculate the consistency index :

[0110]

[0111] in, To determine the order of a matrix.

[0112] Find the average random consistency index (Industry-standard fixed values), corresponding to a 5th order matrix .

[0113] 5.4.3 Calculate the consistency ratio Determine the validity of the matrix:

[0114]

[0115] when If the judgment matrix passes the consistency check, the weight result is valid; otherwise, the scale of the judgment matrix needs to be adjusted and recalculated.

[0116] Substituting the data, we get: , , The judgment matrix passed the consistency test, and the weight results are scientific and effective, and can be used to screen for the optimal solution.

[0117] 5.5 Pareto Optimal Solution Selection

[0118] Based on the above weighting results, the Pareto optimal solution set output by NSGA-III is comprehensively scored to select the globally optimal scheduling scheme. The steps are as follows:

[0119] Target value normalization: Eliminate differences in the dimensions and magnitudes of various targets, and use different normalization formulas to distinguish between maximization and minimization targets.

[0120] Maximize objectives (green energy consumption rate, computing power SLA):

[0121]

[0122] Minimize objectives (peak-to-valley difference, carbon emissions, operating costs):

[0123]

[0124] in, For the first The first alternative option One target original value, , For the Pareto solution set, the first The maximum and minimum values ​​of each target. This is the normalized value, with a range of [0,1].

[0125] Calculate the overall score: Combine the weight vector to calculate the overall score for each alternative. :

[0126]

[0127] Selecting the optimal solution: Overall score The best alternative is the globally optimal scheduling scheme that takes into account all five optimization objectives. After decomposition, it generates executable collaborative scheduling instructions for computing power, electricity, and carbon emissions.

[0128] Example 2:

[0129] A low-carbon and efficient data center scheduling system based on the optimization of computing, power, and carbon emission coupling includes a computing, power, and carbon emission sensing layer, comprising a computing power monitoring unit, a power monitoring unit, and a carbon emission monitoring unit.

[0130] The computing power monitoring unit includes collecting computing power load and task information from servers, virtual machines, and containers;

[0131] The power monitoring unit includes smart meters, green electricity monitoring, energy storage BMS, and grid interface terminals;

[0132] The carbon emission monitoring unit includes carbon factor acquisition and carbon emission measurement modules.

[0133] The smart meter in the power monitoring unit and the carbon emission metering module in the carbon emission monitoring unit are integrated by the electric carbon meter, which can meet the needs of power monitoring and perform carbon emission monitoring.

[0134] Coupled optimization decision layer: includes a three-flow coupled modeling module for constructing a computing-electricity-carbon correlation model, a multi-objective scheduling module that integrates the improved NSGA-III and DRL hybrid optimization algorithms, and a prediction module for predicting future scheduling parameters;

[0135] Collaborative Execution Layer: Includes a computing power scheduling unit, a power control unit, and a cooling linkage unit for coordinated regulation of the cooling system; the computing power scheduling unit includes task scheduling, load balancing, and virtual machine migration controller; the power control unit includes energy storage converter PCS, distributed power access unit (controlling photovoltaic output), wind power controller, and mains power switching device; the cooling linkage unit includes intelligent regulation of air conditioning, chiller, and liquid cooling system.

[0136] Visualized control layer: Includes a three-flow monitoring panel and alarm module for displaying system operating status, scheduling effect, carbon footprint and energy consumption reports.

[0137] Specifically, when the data center is configured with: 10MW IT load, 2MW photovoltaic + 1.5MW wind power, 3MWh lithium battery energy storage, and a computing-based carbon scheduling system (integrating an improved NSGA-III and PPO hybrid optimization algorithm), the results are as follows:

[0138] 1. Off-peak season (23:00-7:00, low price, low carbon)

[0139] Full green power generation + energy storage charging + IT full-load operation to absorb surplus green power

[0140] Shift latency-tolerant tasks to off-peak periods to maximize the use of low-cost green electricity.

[0141] 2. Standard section (7:00-17:00, at the same price)

[0142] Prioritize green electricity, use energy storage for backup, and allocate computing power on demand.

[0143] Maintain stable load and manage carbon emissions in real time.

[0144] 3. Peak period (17:00-23:00, high price, high carbon)

[0145] Energy storage and discharge for power supply reduces the cost of purchasing electricity from the grid.

[0146] Delayed tasks are executed, non-core loads are reduced, and peak shaving is carried out in the power grid.

[0147] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and incorporate common knowledge or customary techniques in the art disclosed herein. The specification and examples are to be considered exemplary only, and the true scope of this application is indicated by the claims.

[0148] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The embodiments of this application described above do not constitute a limitation on the scope of protection of this application.

Claims

1. A low-carbon and efficient scheduling method for data centers based on computation, power, carbon, and three-stream coupling optimization, characterized in that, include: Step S1: Collect computing power flow data, power flow data and carbon flow data from the data center, perform normalization, noise reduction and spatiotemporal alignment preprocessing, and construct a computing power carbon flow database; Step S2: Construct a computing power-electricity coupling model and an electricity-carbon emission coupling model, and establish a multi-dimensional constraint set that includes power balance constraints, computing power service level agreement constraints, energy storage charging and discharging constraints, green electricity consumption constraints, carbon emission upper limit constraints, and grid peak-valley interaction constraints, to form a computing power-carbon three-flow coupling model; Step S3: Construct a five-dimensional multi-objective optimization scheduling model with the objectives of maximizing green electricity consumption rate, minimizing load peak-valley difference, minimizing carbon emissions, minimizing operating costs, and maximizing computing power service level; Step S4: The multi-objective optimization scheduling model is solved by using a hybrid optimization strategy that combines the improved NSGA-III algorithm with DRL to generate the Pareto optimal scheduling solution set; Step S5: Determine the weights of each objective based on the analytic hierarchy process, select the globally optimal scheduling scheme from the Pareto optimal scheduling solution set, and decompose it into computing power scheduling instructions, power scheduling instructions, and carbon control instructions for coordinated control and execution in the data center.

2. The method according to claim 1, characterized in that, In step S2, the computing power-electricity coupling model is expressed as follows: ,in, Real-time power consumption of IT equipment For computing load rate, For server energy efficiency, This refers to the intake air temperature.

3. The method according to claim 1, characterized in that, The power-carbon emission coupling model in step S2 is expressed as follows: , in, For carbon emission rate, , , These represent the charging power for mains power, green electricity, and energy storage, respectively. For real-time grid carbon factor, This represents the carbon emission coefficient for energy storage.

4. The method according to claim 1, characterized in that, The improved NSGA-III algorithm in step S4 includes: Introduce a uniformly distributed reference point mechanism; The target space adaptive normalization method is adopted to adaptively scale the values ​​of five objective functions, namely green energy absorption rate, load peak-valley difference, carbon emissions, operating costs and computing power SLA, by calculating ideal points and extreme points, thereby eliminating the difference in dimensions and magnitudes. A reference point-based association selection strategy is adopted to prioritize the retention of individuals associated with sparse reference points, thereby maintaining population diversity. Integrated constraint-compatible optimization uses constraint dominance rules to handle infeasible solutions that violate operational boundary conditions.

5. The method according to claim 4, characterized in that, The DRL in step S4 adopts a near-end strategy optimization algorithm, which forms a hybrid optimization strategy with the improved NSGA-III algorithm.

6. The method according to claim 1, characterized in that, In step S5, the specific steps for determining the weights of each objective using the analytic hierarchy process include: A three-tiered structure is constructed, comprising an objective layer, a criterion layer, and a solution layer, wherein the criterion layer corresponds to the five optimization objectives described above. The judgment matrix of the criterion layer is constructed based on the Saaty 1-9 scaling method; Calculate the largest eigenvalue and the corresponding eigenvector of the judgment matrix, and obtain the weight coefficients of each optimization objective after normalization.

7. The method according to claim 1, characterized in that, Following step S1, a time-series prediction step based on a long short-term memory network is also included.

8. A low-carbon, high-efficiency scheduling system for data centers based on computation, power consumption, carbon, and three-stream coupling optimization, used to execute the method according to any one of claims 1 to 7, characterized in that, The system includes: The computing, power, and carbon emission sensing layer includes a computing power monitoring unit, a power monitoring unit, and a carbon emission monitoring unit. Coupled optimization decision layer: includes a three-flow coupled modeling module for constructing a computing-electricity-carbon correlation model, a multi-objective scheduling module that integrates the improved NSGA-III and DRL hybrid optimization algorithms, and a prediction module for predicting future scheduling parameters; Collaborative Execution Layer: Includes computing power scheduling unit, power control unit, and refrigeration linkage unit for coordinated regulation of the refrigeration system; Visualized control layer: Includes a three-flow monitoring panel and alarm module for displaying system operating status, scheduling effect, carbon footprint and energy consumption reports.

9. The system according to claim 8, characterized in that, The smart meter in the power monitoring unit and the carbon emission metering module in the carbon emission monitoring unit are integrated into a single carbon meter, which is used to collect both power data and carbon emission data.

Images