Green data resource low-carbon scheduling management system based on load prediction

Abstract

This invention discloses a green data resource low-carbon scheduling and management system based on load prediction, specifically relating to the field of data center energy-saving technology. It includes a carbon-load joint prediction module, a task carbon-aware scheduling module, and a cooling-computing collaborative low-carbon control module. The carbon-load joint prediction module uses an attention-based multi-task LSTM network to predict load and carbon intensity in parallel. It obtains the predicted distribution through multiple forward propagations using Dropout, takes the lower confidence bound as the risk-adjusted carbon intensity value, calculates the marginal carbon emissions per unit task, and marks the low-carbon efficiency window. The task carbon-aware scheduling module performs differentiated processing based on task latency tolerance and user low-carbon scheduling flags. For latency-tolerant tasks, it uses dynamic programming to find the optimal starting scheduling time for carbon emissions within the low-carbon efficiency window. This invention achieves end-to-end low-carbon collaboration in prediction, scheduling, and cooling, effectively reducing carbon emissions from data centers.

Description

Technical Field

[0001] This invention relates to the field of data center energy-saving technology, and more specifically, to a green data resource low-carbon scheduling and management system based on load forecasting. Background Technology

[0002] With the rapid development of big data and cloud computing, the energy consumption and carbon emissions of data centers are becoming increasingly prominent issues. Traditional data center resource scheduling mainly focuses on computing performance and service quality, typically employing strategies such as shortest task priority and load balancing, with little consideration for the spatiotemporal differences in electricity carbon emission intensity. In recent years, some research has begun to introduce carbon-aware scheduling, using predicted data of grid carbon emission factors to migrate computing tasks to periods or regions with lower carbon intensity. However, most existing carbon-aware scheduling methods use deterministic carbon intensity prediction, ignoring the decision-making risks brought about by prediction uncertainty. This can easily lead to errors in prediction, resulting in actual carbon emissions increasing instead of decreasing. Furthermore, load forecasting and carbon intensity forecasting are usually performed independently, failing to fully utilize the coupling relationship between the two, resulting in limited prediction accuracy.

[0003] At the task scheduling level, existing solutions often employ a uniform carbon-aware strategy for all tasks, failing to differentiate based on task latency sensitivity and user's low-carbon scheduling intentions. This can lead to latency-sensitive tasks potentially violating service level agreements (SLAs) while waiting for low-carbon windows. For latency-tolerant tasks, existing scheduling algorithms mostly use threshold-based heuristics, making it difficult to find the globally optimal starting scheduling moment within a limited time window. Furthermore, task scheduling and cooling system control are isolated. Cooling systems typically respond passively based on real-time temperature, failing to utilize load forecasting information for pre-cooling or cold storage operations. During high-carbon periods, they still consume significant amounts of electricity to maintain cooling, lacking the ability to achieve full-cycle collaborative optimization between computational load and the cooling system.

[0004] To address the aforementioned issues, this invention proposes a green data center resource low-carbon scheduling and management system based on load forecasting. Through the organic coordination of three modules—carbon-load joint forecasting, task carbon-aware scheduling, and cooling-computing collaborative control—it achieves end-to-end low-carbon operation of the data center, from forecasting to scheduling to cooling. This system not only improves the accuracy of load and carbon intensity forecasts but also reduces decision-making risks arising from forecast uncertainty through risk adjustment strategies. Furthermore, it adopts differentiated scheduling strategies based on task type and utilizes model predictive control to achieve collaborative optimization of the cooling system and computing load, thereby effectively reducing the overall carbon emissions of the data center. Summary of the Invention

[0005] To overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide a resource low-carbon scheduling and management system based on load prediction in green data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The resource low-carbon scheduling and management system based on load forecasting and green data includes the following modules: The carbon-load joint prediction module is used to collect historical time series data, and predict the load value and carbon intensity value of multiple future time steps in parallel through an attention-based multi-task long short-term memory encoding and decoding network. Based on multiple random forward propagations, the mean of load prediction and the mean and standard deviation of carbon intensity prediction are obtained. The confidence lower bound of carbon intensity prediction is taken as the risk-adjusted carbon intensity value. The expected marginal carbon emissions per unit task are calculated, and the time window in which the marginal carbon emissions are lower than 30% of the long-term mean is marked as a low-carbon and high-efficiency window. The task carbon-aware scheduling module is used to receive computing task requests submitted by users, classify tasks based on their carbon sensitivity according to the maximum tolerable completion delay and task type label, obtain low-carbon and high-efficiency windows from the carbon-load joint prediction module for delay-tolerant tasks, and use dynamic programming to solve for the starting scheduling time that minimizes the marginal carbon emissions of the task throughout its execution period. For delay-sensitive tasks, the module queries the instantaneous carbon intensity of the power area where each node is located in real time and prioritizes scheduling to the node with the lowest carbon intensity and abundant resources. The cooling-computing collaborative low-carbon control module is used to establish an equivalent thermal dynamic model of the data center. It receives the outputs of the carbon-load joint prediction module and the task carbon perception scheduling module. With the integral of the product of the total energy consumption of the cooling system and the carbon intensity and the temperature violation penalty term as the target, it uses model prediction control to continuously optimize the air conditioning supply temperature and cooling power. It executes pre-cooling commands during the low-carbon and high-efficiency window period and reduces the cooling power to release cooling capacity during the high-carbon window period.

[0007] Specifically, the carbon-charge joint prediction module performs the following steps: The system continuously collects multi-source historical time series data through data interfaces, including computing load, carbon emission factor, outdoor temperature and humidity, and average server power consumption. It then performs timestamp alignment and fills in missing values ​​to construct a multi-dimensional sample set. A multi-task long short-term memory encoding and decoding network based on attention mechanism is constructed. The normalized historical load sequence and carbon intensity sequence are concatenated as the encoder input, and the decoder outputs the load prediction value and carbon intensity prediction value for multiple future time steps in parallel. A carbon intensity trend direction consistency regularization term is added to the loss function to ensure the reliability of the prediction of the carbon intensity rise and fall trend. Use the trained model to perform online rolling predictions; The Dropout layer is enabled during each inference, and multiple random forward propagations are performed consecutively to obtain multiple sets of future prediction sequences. The mean and standard deviation of the load prediction and the mean and standard deviation of the carbon intensity prediction are calculated at each time step, and the lower confidence bound of the carbon intensity prediction is taken as the risk-adjusted carbon intensity value.

[0008] Specifically, the carbon-charge joint prediction module is also used for: Based on the risk-adjusted carbon intensity, the average server power consumption per unit load, and the time step, the expected marginal carbon emissions per unit task are calculated; a time window in which the expected marginal carbon emissions per unit task are 30% lower than the long-term average is marked as a low-carbon and high-efficiency window, and the duration of the window is marked for use by the task carbon awareness scheduling module. When searching for the optimal start time, the task carbon sensing and scheduling module prioritizes candidate time periods that fall within the low-carbon and high-efficiency window. If multiple windows meet the constraints, the module selects the start time with the lowest marginal carbon emissions.

[0009] Specifically, the task carbon sensing scheduling module performs the following steps: Receive computing task requests submitted by users, extract resource request amount, maximum tolerable completion delay, task type label and whether the user allows low-carbon scheduling. If the user does not specify, low-carbon scheduling is allowed by default. Tasks are classified by carbon sensitivity based on preset rules, with the priority rule being: if the user explicitly allows low-carbon scheduling, the task can be given priority for carbon-aware scheduling, but its maximum tolerable delay must not be violated. For delay-tolerant tasks, the sequence of future low-carbon and high-efficiency windows is obtained from the carbon-load joint prediction module. With the latest completion time of the task as a constraint, dynamic programming is used to solve for the starting scheduling time that minimizes the marginal carbon emissions during the entire execution period of the task, and the computing node group with the lowest carbon intensity within the window is matched. For latency-sensitive tasks, the instantaneous carbon intensity of the power area where each node is located is queried in real time. Priority is given to scheduling to the node with the lowest current carbon intensity and sufficient resources. At the same time, a dynamic voltage and frequency adjustment strategy is assigned to the task, and the CPU frequency is set to the lowest level to meet the performance requirements without violating the latency constraints, so as to reduce energy consumption.

[0010] Specifically, the specific rules for carbon sensitivity classification are as follows: If the maximum tolerable latency for a task is less than 5 minutes or the type is online transaction or real-time inference, it is marked as latency-sensitive. For tasks with a maximum tolerable latency between 5 and 15 minutes, if the task type is labeled as batch data processing, backup, or offline analysis, it is marked as latency-tolerant; if it is an interactive query or a non-critical real-time service, it is marked as latency-sensitive. If the tolerance for latency is greater than 15 minutes and the type is batch processing, data backup, or model training, it is marked as latency-tolerant. Users with low-carbon permission flags will be given priority in enabling carbon-aware scheduling under the same conditions.

[0011] Specifically, the cooling-computation coordinated low-carbon control module performs the following steps: A first-order equivalent thermal dynamic model of the data center was established, and the model parameters were identified online using the recursive least squares method. Receive the future load sequence and carbon intensity sequence provided by the carbon-load joint prediction module, and the power consumption distribution prediction of computing nodes in each future time period provided by the task carbon sensing and scheduling module, and construct a thermal state prediction sequence for multiple future control cycles; The integral of the product of the total energy consumption of the cooling system and the carbon intensity of the power grid is used as the carbon emission cost function. A temperature violation penalty term is added. Under the constraint that the rack inlet air temperature does not exceed the safety threshold, the model predictive control rolling optimization is used to solve the air conditioning supply air temperature setpoint and cooling power output for multiple future cycles. If we are currently in a low-carbon, high-efficiency window, we will execute a pre-cooling command to increase the cooling capacity in advance with an air supply temperature lower than the preset normal set point. When it is predicted that we will soon enter a high-carbon window, we will gradually reduce the cooling power and increase the air supply temperature within the safe temperature range, using the stored cooling capacity to offset the server's heat generation. The optimized air supply temperature curve and cooling power command are sent to the computer room air conditioning group control unit for execution, and feedback correction is performed based on the latest predicted and measured temperature deviations.

[0012] Specifically, in the cooling-computation collaborative low-carbon control module: The power consumption for cooling is expressed as the ratio of cooling power to real-time energy efficiency ratio. The real-time energy efficiency ratio is dynamically estimated using a quadratic regression model affected by supply air temperature and outdoor temperature. The regression coefficients are obtained by least squares fitting of historical operating data and are updated periodically.

[0013] Specifically, the objective function of the model predictive control is: ;in To optimize the target value, To predict the total number of control cycles, To control the periodic index, The electrical power consumed for cooling is determined by the cooling power and the energy efficiency ratio. The carbon intensity prediction value is taken from the risk-adjusted carbon intensity output by the carbon-charge joint prediction module. , For time step, For the first The rack inlet air temperature for each control cycle For temperature safety limits, The temperature violation penalty coefficient; Constraints include equality constraints after discretization of the thermal dynamics model, Cooling power upper and lower limits; each rolling optimization step only executes the first control variable.

[0014] The technical effects and advantages of this invention are as follows: This invention employs an attention-based multi-task LSTM network to predict load and carbon intensity in parallel through a carbon-load joint prediction module, and introduces a carbon intensity trend direction consistency regularization term, significantly improving the reliability of carbon intensity rise and fall trend prediction. Simultaneously, it utilizes Dropout multiple forward propagations to obtain the predicted distribution, taking the confidence lower bound as the risk-adjusted carbon intensity value, effectively avoiding the decision-making risk of mistakenly entering a high-carbon window due to underestimation. The task carbon-aware scheduling module differentiates tasks based on their latency sensitivity and user low-carbon scheduling flags. For latency-tolerant tasks, dynamic programming is used to solve for the optimal starting scheduling time for carbon emissions within the low-carbon efficiency window. For latency-sensitive tasks, real-time scheduling to the node with the lowest carbon intensity is supplemented by dynamic voltage and frequency adjustment, maximizing carbon emission reduction benefits while meeting service quality requirements.

[0015] This invention further establishes a data center thermal dynamics model through a cooling-computing collaborative low-carbon control module. Using the integral of the product of cooling power consumption and carbon intensity as the target, it employs model predictive control to continuously optimize air conditioning supply air temperature and cooling power, and introduces a dynamic regression model for cooling energy efficiency ratio to achieve adaptive and accurate metering of cooling energy consumption under various operating conditions. This module automatically pre-cools and stores cooling capacity during low-carbon windows and releases cooling capacity during high-carbon windows to reduce cooling power consumption, thereby organically coordinating computing load scheduling and cooling system control to minimize data center carbon emissions throughout the entire lifecycle. Compared with existing technologies, this invention simultaneously improves prediction accuracy, low-carbon scheduling, and cooling synergy, resulting in significant energy-saving and emission-reduction effects. Attached Figure Description

[0016] Figure 1 This is a system block diagram of the present invention. Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, the resource low-carbon scheduling and management system module based on load forecasting green data is as follows: The carbon-load joint prediction module uses an attention-based multi-task LSTM network to perform rolling predictions of future load and grid carbon intensity, and utilizes Dropout multiple forward propagations to obtain the mean and standard deviation of the predictions. A risk adjustment strategy (confidence lower bound) is further employed to obtain a conservative carbon intensity estimate, which is used to calculate the marginal carbon emissions per unit task and to mark low-carbon, high-efficiency windows.

[0018] Includes the following steps: Historical time series data is continuously collected through data interfaces, including the computing load of the data center at the second level, the carbon emission factor per minute from the real-time carbon emission interface of the power grid, outdoor temperature and humidity, and average power consumption of the server. The multi-source data is timestamped and missing values ​​are processed by linear interpolation and forward filling to construct a multi-dimensional sample set. A multi-task long short-term memory (LSTM) encoder-decoder network based on an attention mechanism is constructed. The normalized historical load sequence and carbon intensity sequence are concatenated as the encoder input, and the decoder outputs the predicted load and carbon intensity values ​​for the next H time steps in parallel. The loss function is defined as the weighted sum of the mean square error of the load prediction and the mean square error of the carbon intensity prediction, with an added carbon intensity trend direction consistency regularization term to ensure the reliability of the carbon intensity rise and fall trend prediction. This regularization term is defined as: comparing the sign of the carbon intensity difference between adjacent time steps in the predicted sequence with the sign of the actual carbon intensity difference; if the signs are inconsistent, a penalty is applied. Specifically… ,in This is a regularization term (penalty value) for the consistency of carbon intensity trend direction. The total number of time steps to be predicted. For time step index, It is a sign function (returns +1, 0, or -1). , For true carbon intensity at time and The value of +1, , To predict carbon intensity at time and The value of +1, This is used as an indicator function, and this term is added to the total loss as a weighting factor to ensure the accuracy of the model's prediction of the direction of carbon intensity rise and fall.

[0019] The trained model is used for online rolling prediction, predicting the load curve and grid carbon intensity curve for the next hour every 5 minutes; After obtaining deterministic point predictions of load and carbon intensity, the trained attention-based multi-task LSTM encoder-decoder network is subjected to a Dropout layer during each inference iteration, performing K consecutive random forward propagations to obtain K sets of prediction sequences for the next H steps. Let the load of the k-th prediction be denoted as... Carbon strength is , Then, for each time step, the predicted mean and standard deviation are calculated: ; ; in, For a moment The average load forecast This represents the total number of random forward propagations in Dropout. For the first Index of the next random propagation, For the first The time obtained from the prediction Load value, For a moment The standard deviation of load forecast; For a moment The predicted mean of carbon intensity, For the first The time obtained from the prediction Carbon strength value, For a moment The predicted standard deviation of carbon intensity.

[0020] Considering the low-carbon conservatism of scheduling decisions, the lower confidence bound of carbon intensity prediction is taken as the risk-adjusted carbon intensity value. This is to reduce the probability of mistakenly entering a high-carbon window due to underestimation: ;in This is the risk-adjusted carbon intensity value. This represents the predicted mean of carbon intensity. This is the risk aversion coefficient; taking a negative value yields the lower bound, which is recommended. =−1.645 (corresponding to the 95% one-sided confidence lower bound). Load forecasting uses the mean. Maintain accurate estimates of resource needs.

[0021] Finally, the adjusted carbon strength and load average Proceed to the next step: Expected marginal carbon emissions per unit of task The calculation formula uses Replace the original carbon intensity prediction value; the process is as follows: Based on the predicted load and carbon intensity, the expected marginal carbon emissions per unit task are calculated for each future time window. The future time windows are divided into equal-length periods, and for each period, the carbon intensity is taken. Expected marginal carbon emissions per unit of task Depend on Calculation, where For time step, The average server power consumption per unit load is calculated from historical statistics. This metric reflects the additional carbon emissions resulting from performing a standard unit task during that period.

[0022] The time window in which the marginal carbon emissions are 30% lower than the long-term average is marked as the low-carbon and high-efficiency window, and the sustainable duration of the window is marked for the task carbon perception and scheduling module to call. When the task carbon perception and scheduling module is looking for the optimal start time, it will give priority to the candidate time period that falls within the low-carbon and high-efficiency window. If multiple windows meet the constraints, the start time with the minimum marginal carbon emissions will be selected.

[0023] The task carbon-aware scheduling module classifies tasks based on their latency tolerance and user low-carbon scheduling flags. For latency-tolerant tasks, dynamic programming is used to find the optimal starting scheduling time for carbon emissions within the low-carbon and high-efficiency window. For latency-sensitive tasks, the module selects the computing node with the lowest current carbon intensity in real time and uses dynamic voltage and frequency adjustment to reduce energy consumption.

[0024] Includes the following steps: Receive computing task requests submitted by users, extract the resource request amount, maximum tolerable completion delay, task type label and whether the user allows low-carbon scheduling from the task description. If the user does not specify, low-carbon scheduling is allowed by default. Tasks are classified by carbon sensitivity based on preset rules: Priority rule: If the user explicitly allows low-carbon scheduling (which is allowed by default), the task can be given priority for carbon-aware scheduling, but must not violate its maximum tolerable delay.

[0025] If the maximum tolerable latency of a task is less than 5 minutes or the type is online transaction or real-time inference, it is marked as latency-sensitive (high carbon sensitivity). For tasks with a maximum tolerable latency between 5 and 15 minutes: If the task type label is batch data processing, backup, or offline analysis, then mark it as latency-tolerant (carbon-sensing scheduling is possible). If it is an interactive query or a non-critical real-time service, it is marked as latency-sensitive (scheduled according to the shortest latency strategy). If the tolerable latency is greater than 15 minutes and the type is batch processing, data backup, or model training, it is marked as latency-tolerant. This classification ensures that no task category is missed, and users with low-carbon permission flags are given priority in enabling carbon-aware scheduling under the same conditions.

[0026] For delay-tolerant tasks, a future low-carbon, high-efficiency window sequence is obtained from the carbon-load joint prediction module. Using the latest completion time of the task as a constraint, dynamic programming is employed within the candidate window to find the starting scheduling time that minimizes the marginal carbon emissions throughout the task's execution period. The required execution time of the task is discretized into fixed time steps, and states are defined. From time Begin executing the remaining portion of the task until completion, reaching the minimum cumulative marginal carbon emissions, and the state transitions to... ,in This represents the minimum cumulative marginal carbon emissions from the start of the remaining portion of the task at time t until completion. The minimum cumulative marginal carbon emissions starting from the next moment. This represents the percentage of work completed in a single step. The marginal carbon emissions per unit task at this moment are provided for the carbon-load joint prediction module. The boundary condition is that the task must be completed before its latest completion time D. The optimal start time and the corresponding low-carbon window are obtained by recursively tracing from D to the current moment. And match the computation node group with the lowest carbon intensity within that window; For latency-sensitive tasks, the instantaneous carbon intensity of the power area where each node is located is queried in real time. Priority is given to scheduling to the node with the lowest current carbon intensity and sufficient resources. At the same time, a dynamic voltage and frequency adjustment strategy is assigned to the task, and the CPU frequency is set to the lowest level to meet the performance requirements without violating the latency constraints, so as to reduce energy consumption. The scheduling decisions for all tasks are output as a task-node-start time binding table, which is then sent to the resource manager for execution. The scheduling time and node power consumption information are also synchronized to the cooling coordination module.

[0027] The cooling-computing collaborative low-carbon control module establishes a thermal dynamic model of the data center. Combining future load and carbon intensity predictions, it uses model predictive control to continuously optimize air conditioning supply air temperature and cooling power, aiming to minimize the integral of the product of cooling power consumption and carbon intensity. By pre-cooling commands, it stores cooling capacity before the high-carbon window and releases it during high-carbon periods, achieving full-cycle low-carbon collaboration between cooling and computing load.

[0028] Includes the following steps: Establish a first-order equivalent thermal dynamic model of the data center. The model expression is as follows: ,in The rack inlet air temperature, The air conditioning supply temperature, Outdoor temperature This represents the total power consumption of the server. For the equivalent heat capacity of the computer room, For the thermal conductivity with the outside world, For air supply volume, The air heat capacity coefficient is used to determine the cooling capacity, which is determined by the supply air temperature difference and air volume. air density, Specific heat capacity of air. Model parameters. and Online identification using recursive least squares method; The system receives the future load sequence and carbon intensity sequence provided by the carbon-load joint prediction module, as well as the power consumption distribution prediction of computing nodes in each future time period provided by the task carbon awareness scheduling module, and constructs the thermal state prediction sequence for the next H control cycles. The integral of the product of the total energy consumption of the cooling system and the carbon intensity of the power grid is used as the carbon emission cost function. The electrical power consumed by cooling is included. Expressed as cooling power Compared with real-time energy efficiency ratio The ratio, i.e. Energy efficiency ratio Use supply air temperature and outdoor temperature Dynamically estimate the impact using a quadratic regression model: ; where the coefficient to The data was obtained by least-squares fitting of historical operating data from chillers or precision air conditioners under different operating conditions, and was updated regularly with seasonal changes. Substituting the expression into the carbon emission cost function The term makes the objective function optimize the supply air temperature. and cooling power At that time, it can automatically weigh the impact of changes in energy efficiency ratio on actual power consumption and carbon emissions: when the outdoor temperature is low or the supply air temperature is set high, With a larger temperature range, less electricity is needed to produce the same cooling capacity. Model predictive control will tend to moderately increase the supply air temperature or utilize natural cooling sources under such conditions, thereby more realistically minimizing carbon emissions throughout the entire cycle. The original temperature penalty term and constraints remain unchanged, and the optimization solution process uses sequential quadratic programming or interior-point methods to handle this nonlinear objective. This addition upgrades the carbon emission accounting of cooling control from a fixed efficiency assumption to precise measurement based on adaptive operating conditions.

[0029] Incorporating a temperature violation penalty term, and under the constraint that the rack inlet air temperature does not exceed a safe threshold, model predictive control is used to perform rolling optimization to solve for the air conditioning supply air temperature setpoint and cooling power output for the next H cycles. The optimization problem is in the form that the decision variable is the supply air temperature. and cooling power Minimize the target ,in To optimize the target value (carbon emission cost + temperature penalty). To predict the total number of control cycles, The electrical power consumed for cooling is determined by the cooling power and the energy efficiency ratio. The carbon intensity prediction value is taken from the risk-adjusted carbon intensity output by the carbon-charge joint prediction module. To ensure a conservative response of cooling scheduling during high-carbon periods, For temperature safety limits, The temperature violation penalty coefficient; constraints include equality constraints after discretization of the thermal dynamics model, Cooling power upper and lower limits. Each rolling optimization step only executes the first control variable.

[0030] If the current period is a low-carbon and high-efficiency window, the solver will automatically generate a pre-cooling command to increase the cooling capacity in advance with the supply air temperature 2-3℃ lower than the conventional set point, so as to reduce the thermal mass temperature of the computer room and store the cooling capacity in the building envelope and rack. When the high-carbon window is predicted to be approaching, the solver gradually reduces the cooling power and increases the supply air temperature within the safe temperature range, using the stored cooling capacity to offset the server's heat generation, thus significantly reducing cooling power consumption during the high-carbon period. The optimized air supply temperature curve and cooling power command are sent to the computer room air conditioning group control unit for execution. Feedback correction is performed every 15 minutes based on the latest predicted and measured temperature deviation, realizing full-cycle low-carbon collaborative scheduling of cooling and computing load.

[0031] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0032] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0033] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0034] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0035] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0036] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0037] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0038] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A resource low-carbon scheduling and management system based on load forecasting and green data, characterized in that, Includes the following modules: The carbon-load joint prediction module is used to collect historical time series data, and predict the load value and carbon intensity value of multiple future time steps in parallel through an attention-based multi-task long short-term memory encoding and decoding network. Based on multiple random forward propagations, the mean of load prediction and the mean and standard deviation of carbon intensity prediction are obtained. The confidence lower bound of carbon intensity prediction is taken as the risk-adjusted carbon intensity value. The expected marginal carbon emissions per unit task are calculated, and the time window in which the marginal carbon emissions are lower than 30% of the long-term mean is marked as a low-carbon and high-efficiency window. The task carbon-aware scheduling module is used to receive computing task requests submitted by users, classify tasks based on their carbon sensitivity according to the maximum tolerable completion delay and task type label, obtain low-carbon and high-efficiency windows from the carbon-load joint prediction module for delay-tolerant tasks, and use dynamic programming to solve for the starting scheduling time that minimizes the marginal carbon emissions of the task throughout its execution period. For delay-sensitive tasks, the module queries the instantaneous carbon intensity of the power area where each node is located in real time and prioritizes scheduling to the node with the lowest carbon intensity and abundant resources. The cooling-computing collaborative low-carbon control module is used to establish an equivalent thermal dynamic model of the data center. It receives the outputs of the carbon-load joint prediction module and the task carbon perception scheduling module. With the integral of the product of the total energy consumption of the cooling system and the carbon intensity and the temperature violation penalty term as the target, it uses model prediction control to continuously optimize the air conditioning supply temperature and cooling power. It executes pre-cooling commands during the low-carbon and high-efficiency window period and reduces the cooling power to release cooling capacity during the high-carbon window period.

2. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 1, characterized in that, The carbon-charge joint prediction module specifically performs the following steps: The system continuously collects multi-source historical time series data through data interfaces, including computing load, carbon emission factor, outdoor temperature and humidity, and average server power consumption. It then performs timestamp alignment and fills in missing values ​​to construct a multi-dimensional sample set. A multi-task long short-term memory encoding and decoding network based on attention mechanism is constructed. The normalized historical load sequence and carbon intensity sequence are concatenated as the encoder input, and the decoder outputs the load prediction value and carbon intensity prediction value for multiple future time steps in parallel. A carbon intensity trend direction consistency regularization term is added to the loss function to ensure the reliability of the prediction of the carbon intensity rise and fall trend. Use the trained model to perform online rolling predictions; The Dropout layer is enabled during each inference, and multiple random forward propagations are performed consecutively to obtain multiple sets of future prediction sequences. The mean and standard deviation of the load prediction and the mean and standard deviation of the carbon intensity prediction are calculated at each time step, and the lower confidence bound of the carbon intensity prediction is taken as the risk-adjusted carbon intensity value.

3. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 2, characterized in that, The carbon-charge joint prediction module is also used for: Based on the risk-adjusted carbon intensity, the average server power consumption per unit load, and the time step, the expected marginal carbon emissions per unit task are calculated; a time window in which the expected marginal carbon emissions per unit task are 30% lower than the long-term average is marked as a low-carbon and high-efficiency window, and the duration of the window is marked for use by the task carbon awareness scheduling module. When searching for the optimal start time, the task carbon sensing and scheduling module prioritizes candidate time periods that fall within the low-carbon and high-efficiency window. If multiple windows meet the constraints, the module selects the start time with the lowest marginal carbon emissions.

4. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 1, characterized in that, The task carbon sensing scheduling module specifically performs the following steps: Receive computing task requests submitted by users, extract resource request amount, maximum tolerable completion delay, task type label and whether the user allows low-carbon scheduling. If the user does not specify, low-carbon scheduling is allowed by default. Tasks are classified by carbon sensitivity based on preset rules, with the priority rule being: if the user explicitly allows low-carbon scheduling, the task can be given priority for carbon-aware scheduling, but its maximum tolerable delay must not be violated. For delay-tolerant tasks, the sequence of future low-carbon and high-efficiency windows is obtained from the carbon-load joint prediction module. With the latest completion time of the task as a constraint, dynamic programming is used to solve for the starting scheduling time that minimizes the marginal carbon emissions during the entire execution period of the task, and the computing node group with the lowest carbon intensity within the window is matched. For latency-sensitive tasks, the instantaneous carbon intensity of the power area where each node is located is queried in real time. Priority is given to scheduling to the node with the lowest current carbon intensity and sufficient resources. At the same time, a dynamic voltage and frequency adjustment strategy is assigned to the task, and the CPU frequency is set to the lowest level to meet the performance requirements without violating the latency constraints, so as to reduce energy consumption.

5. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 4, characterized in that, The specific rules for carbon sensitivity classification are as follows: If the maximum tolerable latency for a task is less than 5 minutes or the type is online transaction or real-time inference, it is marked as latency-sensitive. For tasks with a maximum tolerable latency between 5 and 15 minutes, if the task type is labeled as batch data processing, backup, or offline analysis, it is marked as latency-tolerant; if it is an interactive query or a non-critical real-time service, it is marked as latency-sensitive. If the tolerance for latency is greater than 15 minutes and the type is batch processing, data backup, or model training, it is marked as latency-tolerant. Users with low-carbon permission flags will be given priority in enabling carbon-aware scheduling under the same conditions.

6. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 1, characterized in that, The cooling-computation coordinated low-carbon control module specifically performs the following steps: A first-order equivalent thermal dynamic model of the data center was established, and the model parameters were identified online using the recursive least squares method. Receive the future load sequence and carbon intensity sequence provided by the carbon-load joint prediction module, and the power consumption distribution prediction of computing nodes in each future time period provided by the task carbon sensing and scheduling module, and construct a thermal state prediction sequence for multiple future control cycles; The integral of the product of the total energy consumption of the cooling system and the carbon intensity of the power grid is used as the carbon emission cost function. A temperature violation penalty term is added. Under the constraint that the rack inlet air temperature does not exceed the safety threshold, the model predictive control rolling optimization is used to solve the air conditioning supply air temperature setpoint and cooling power output for multiple future cycles. If we are currently in a low-carbon, high-efficiency window, we will execute a pre-cooling command to increase the cooling capacity in advance with an air supply temperature lower than the preset normal set point. When it is predicted that we will soon enter a high-carbon window, we will gradually reduce the cooling power and increase the air supply temperature within the safe temperature range, using the stored cooling capacity to offset the server's heat generation. The optimized air supply temperature curve and cooling power command are sent to the computer room air conditioning group control unit for execution, and feedback correction is performed based on the latest predicted and measured temperature deviations.

7. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 6, characterized in that, In the cooling-computation coordinated low-carbon control module: The power consumption for cooling is expressed as the ratio of cooling power to real-time energy efficiency ratio. The real-time energy efficiency ratio is dynamically estimated using a quadratic regression model affected by supply air temperature and outdoor temperature. The regression coefficients are obtained by least squares fitting of historical operating data and are updated periodically.

8. The resource low-carbon scheduling and management system based on load prediction in green data according to claim 6, characterized in that, The objective function of the model predictive control is: ； in To optimize the target value, To predict the total number of control cycles, To control the periodic index, The electrical power consumed for cooling is determined by the cooling power and the energy efficiency ratio. The carbon intensity prediction value is taken from the risk-adjusted carbon intensity output by the carbon-charge joint prediction module. , For time step, For the first The rack inlet air temperature for each control cycle For temperature safety limits, The penalty coefficient for temperature violation; Constraints include equality constraints after discretization of the thermal dynamics model, Cooling power upper and lower limits; each rolling optimization step only executes the first control variable.

Images