Method for scheduling workloads of data center in electricity market environment, and device

Abstract

Provided in embodiments of the present disclosure are a method for scheduling workloads of a data center in an electricity market environment, and a device. In the embodiments of the present disclosure, a multi-stage random optimization solution that allows a data center to participate in a multi-level electricity market is provided. The types and data dependency relationship of workloads are taken into consideration, and a load scheduling model, in a time dimension, for the workloads is constructed; on the basis of electricity source information of the data center, a multi-stage electricity cost optimization model is constructed, so as to comprehensively consider the exogenic uncertainty of an electricity price and carbon emissions and the endogenic uncertainty of the workloads; on the basis of the multi-stage electricity cost optimization model, multi-stage optimization solving is performed on the load scheduling model, so as to obtain a load scheduling scheme and an electricity purchase strategy. Therefore, a more accurate and more reliable estimation is made concerning the electricity flexibility of the data center, such that the electricity cost of the data center can be reduced, support is provided for the reduction of carbon emissions, and the consumption of renewable energy sources is encouraged, thereby achieving the energy use flexibility of the data center in the multi-level electricity market environment.

Description

Method and device for scheduling data center workload in power market environment Technical Field

[0001] The present disclosure relates to the field of data center technology, and more particularly to a method and device for scheduling data center workloads in an electricity market environment. Background Art

[0002] Cloud computing is typically implemented in Internet Data Centers (IDCs). In the digital economy, demand for cloud computing services continues to grow, and the scale and energy consumption of these critical infrastructures are also increasing. Reducing data center electricity costs while maintaining acceptable service quality for customers is a hot research topic.

[0003] Several studies have leveraged data center load flexibility to achieve the goal of reducing electricity costs through load management. For example, in studies aimed at reducing data center operating costs, some have optimized server configuration at the server level, taking into account load time constraints to achieve a trade-off between cost, performance, and reliability. Others have proposed a stochastic optimization method to solve the load allocation problem, aiming to reduce load-level electricity costs. Furthermore, others have studied the rationality of electricity prices in the power market to reduce data center electricity costs.

[0004] However, the above research still has some unresolved issues, such as the relatively simple modeling of the electricity market and the failure to consider the data dependencies between loads, which leads to inaccurate estimates of the electricity costs of data centers, resulting in suboptimal load scheduling, an inability to fully save the electricity costs of data centers, and hindering the effective participation of data centers in the electricity market.

[0005] Summary of the Invention

[0006] Various aspects of the present disclosure provide a method and apparatus for scheduling data center workloads in an electricity market environment, to solve the data center's load scheduling problem, reduce the data center's electricity costs, and improve the data center's effective participation in a multi-level electricity market.

[0007] An embodiment of the present disclosure provides a method for scheduling data center workloads in an electricity market environment, which is applied to a data center operating in a multi-level electricity market, where electricity transactions occur on multiple time scales. The method includes: obtaining information about multiple delay-tolerant workloads that the data center needs to process, where the information about the multiple workloads includes type information about each workload and data dependencies between the multiple workloads; constructing a load scheduling model for the multiple workloads in a time dimension based on the type information about each workload and the data dependencies between the multiple workloads, where the load scheduling model reflects the temporal uncertainty of the workloads; constructing a power cost optimization model for multiple stages corresponding to the data center based on power source information of the data center to optimize the power cost of the data center on the multiple time scales; performing multiple-stage optimization solutions on the load scheduling model based on the power cost optimization model for the multiple stages to obtain scheduling schemes for the multiple workloads on the multiple time scales; and scheduling the multiple workloads on the multiple time scales based on the scheduling schemes for the multiple workloads on the multiple time scales to reduce the power cost of the data center.

[0008] The embodiment of the present disclosure also provides a scheduling device for data center workloads in an electricity market environment, which is applied to a data center operating in a multi-level electricity market, where electricity transactions occur on multiple time scales. The device includes: an acquisition module, which is configured to acquire information about multiple workloads that need to be processed by the data center and are delay-tolerant, wherein the information about the multiple workloads includes type information of each workload and data dependencies between the multiple workloads; a first construction module, which is configured to construct a load scheduling model for the multiple workloads in the time dimension based on the type information of the multiple workloads and the data dependencies between the multiple workloads. The load scheduling model Reflecting the time uncertainty of the workload; the second construction module is configured to construct a power cost optimization model for multiple stages corresponding to the data center based on the power source information of the data center, so as to optimize the power cost of the data center on the multiple time scales; the optimization solution module is configured to perform multiple-stage optimization solutions on the load scheduling model based on the power cost optimization model of the multiple stages, so as to obtain scheduling schemes for the multiple workloads on multiple time scales; the scheduling module is configured to schedule the multiple workloads on multiple time scales respectively according to the scheduling schemes for the multiple workloads on multiple time scales, so as to reduce the power cost of the data center.

[0009] An embodiment of the present disclosure also provides an electronic device, comprising: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program stored in the memory, so as to implement the steps in the data center workload scheduling method provided by the embodiment of the present disclosure.

[0010] The embodiments of the present disclosure also provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement the steps in the method for scheduling data center workloads provided by the embodiments of the present disclosure.

[0011] In an embodiment of the present disclosure, a multi-stage optimization solution for a data center participating in a multi-level electricity market is provided. In this solution, a load scheduling model for the workload in the time dimension is constructed by taking into account the type of workload and data dependency. Based on the power source information of the data center, a power cost optimization model for the multiple stages corresponding to the data center is constructed. Based on the power cost optimization model for the multiple stages, the load scheduling model is optimized and solved in multiple stages to obtain a scheduling scheme for the workload on multiple time scales. Based on this, the workload is scheduled on multiple time scales. Through a reasonable workload scheduling scheme, the power cost of the data center can be reduced, thereby achieving energy flexibility of the data center in a multi-level electricity market environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The drawings described herein are used to provide a further understanding of the present disclosure and constitute a part of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure. In the drawings:

[0013] FIG1 is a flow chart of a method for scheduling workloads in a data center according to an embodiment of the present disclosure;

[0014] FIG2 is a directed acyclic graph showing data dependencies between workloads according to an embodiment of the present disclosure;

[0015] FIG3 is a schematic diagram of a process for constructing a load scheduling model according to an embodiment of the present disclosure;

[0016] FIG4a is a schematic diagram of the discrete relationship between the corresponding time scales of the three-level electricity market provided by an embodiment of the present disclosure;

[0017] FIG4 b is a schematic diagram of a process for constructing a global electricity cost function according to an embodiment of the present disclosure;

[0018] FIG5 is a schematic diagram of the structure of a data center workload scheduling device provided by an embodiment of the present disclosure;

[0019] FIG6 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the present disclosure more clear, the technical solutions of the present disclosure will be clearly and completely described below in conjunction with the specific embodiments of the present disclosure and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present disclosure.

[0021] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0022] In the digital economy, demand for cloud computing services continues to grow, and the scale and energy consumption of data centers, a critical infrastructure, are also increasing. Technically, data centers have the potential to provide flexible loads to the power system while maintaining acceptable cloud computing service quality for customers. This is both cost-effective for both data centers and the power system and helps reduce carbon emissions. Data centers can provide flexible loads to the power system because most workloads are latency-tolerant, meaning they can tolerate a certain delay in execution. This delay can range from seconds to minutes to hours. These latency-tolerant workloads are referred to as latency-tolerant workloads. Alternatively, latency-tolerant workloads can include, but are not limited to, batch processing workloads such as data collection and machine learning. Batch processing workloads typically require significant computing power and energy consumption. Analysis of historical energy consumption data for batch workloads indicates that batch processing can account for approximately 70% of a data center's total electricity consumption.

[0023] On the basis that the data center is mainly based on delay-tolerant workloads (such as batch processing workloads), the data center has the ability to flexibly schedule workloads in the power market and obtain power cost benefits accordingly. Furthermore, when the power system providing power to the data center includes renewable energy, it can also promote the power system's consumption of renewable energy. In view of this, the embodiment of the present disclosure aims to utilize the inherent flexibility of delay-tolerant workloads in execution time to provide a multi-stage optimization solution for data centers to participate in multi-level power markets. In this solution, considering the type of workload and data dependency, a load scheduling model for the workload in the time dimension is constructed; based on the power source information of the data center, a power cost optimization model for multiple stages corresponding to the data center is constructed; based on the power cost optimization model for multiple stages, the load scheduling model is optimized and solved in multiple stages to obtain a scheduling scheme for the workload on multiple time scales. Based on this, the workload is scheduled on multiple time scales. Through a reasonable workload scheduling scheme, the power cost of the data center can be reduced, thereby achieving energy flexibility of the data center in a multi-level power market environment and improving the effective participation of the data center in the multi-level power market.

[0024] It is explained here that the solution provided by the embodiment of the present disclosure is applied to a data center operating in a multi-level electricity market. In the embodiment of the present disclosure, the number of levels of the multi-level electricity market is not limited. For example, it can be a two-level electricity market, a three-level electricity market, a four-level electricity market, etc. In this embodiment, electricity transactions in the multi-level electricity market occur on multiple time scales, and as the level of the electricity market increases, the corresponding time scale becomes smaller. Taking the three-level electricity market as an example, the three-level electricity market includes the first-level electricity market, the second-level electricity market, and the third-level electricity market, wherein the first-level electricity market corresponds to the first time scale, the second-level electricity market corresponds to the second time scale, and the third-level electricity market corresponds to the third time scale. The first time scale is greater than the second time scale, and the second time scale is greater than the third time scale. Further optionally, the first time scale is Q1 times the second time scale, and the second time scale is Q2 times the third time scale, where Q1 and Q2 are positive integers.

[0025] Furthermore, in the embodiment of the present disclosure, a three-tiered electricity market structure is adopted, in which electricity transactions occur on three time scales. These three time scales include 1 hour, 15 minutes and 5 minutes. Accordingly, the three-tiered electricity market includes a day-ahead market with a 1-hour interval, an intraday market with transactions every 15 minutes, and a real-time market with transactions every 5 minutes. It should be noted that the three time scales at which electricity transactions occur in the three-tiered electricity market are not limited to the examples given here, and may depend on the current and future evolution of the electricity market. In the day-ahead market, market participants bid for hourly energy contracts on the operating day. In the intraday market, market participants adjust the amount of electricity purchased based on forecast errors and deviations from expected energy consumption. In the real-time market, participants will clear their contracts based on the real-time electricity price based on the contracted electricity volume established in the previous market.

[0026] In addition, the three-level electricity market corresponds to three time scales. Market participants can participate in electricity purchase at three time scales. These three time scales can also be called the three stages of electricity purchase, namely the 1-hour stage, the 15-minute stage and the 5-minute stage.

[0027] It should be noted that in some embodiments of the present disclosure, a three-level electricity market with a day-ahead market with a 1-hour interval, an intraday market with a 15-minute transaction interval, and a real-time market with a 5-minute transaction interval will be used as an example for explanation, but it is not limited to this. The solution provided by the embodiments of the present disclosure is also applicable to other types of electricity markets.

[0028] The technical solutions provided by various embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0029] FIG1 is a flow chart of a method for scheduling data center workloads in a power market environment according to an embodiment of the present disclosure. The method is applied to data centers operating in a multi-level power market where power transactions occur on multiple time scales. As shown in FIG1 , the method includes:

[0030] 101. Obtain information about multiple delay-tolerant workloads that the data center needs to process, where the information about the multiple workloads includes type information of each workload and data dependencies between the multiple workloads.

[0031] 102. Based on the type information of each workload and the data dependency relationship between multiple workloads, a load scheduling model for multiple workloads in the time dimension is constructed. The load scheduling model reflects the time uncertainty of the workloads.

[0032] 103. Based on the power source information of the data center, construct a multi-stage power cost optimization model corresponding to the data center to optimize the power cost of the data center at multiple time scales;

[0033] 104. Based on the multi-stage power cost optimization model, the load scheduling model is optimized and solved in multiple stages to obtain scheduling solutions for multiple workloads on multiple time scales;

[0034] 105. According to a scheduling scheme for multiple workloads at multiple time scales, multiple workloads are scheduled separately at multiple time scales to reduce the power cost of a data center.

[0035] In the embodiments of the present disclosure, workloads are computing tasks submitted by customers to data centers, with different characteristics and requirements. Among them, delay-tolerant workloads are workloads with soft deadlines and intensive computing resource requirements, and usually have looser execution time constraints. For example, batch workloads are a common delay-tolerant workload. The slack time of a delay-tolerant workload can be a few seconds, a few minutes, or even a few hours. The slack time refers to the time that the workload is allowed to be delayed in execution if the workload is processed before the deadline. In this case, scheduling delay-tolerant workloads can be based on meeting resource requirements without having to pursue a fast response, which allows for greater time flexibility in scheduling delay-tolerant workloads. After the customer submits the delay-tolerant workload, the delay-tolerant workload will be assigned to a specific server in the data center to run. Among them, when the delay-tolerant workload is assigned to a specific server in the data center to run can save the data center's electricity costs to the greatest extent, which is the key issue to be addressed in the embodiments of the present disclosure.

[0036] In the embodiments of the present disclosure, the load granularity of the delay-tolerant workload is not limited, and it can be a computing task of any granularity that needs to be processed by the data center. In some application scenarios, computing tasks of larger granularity can be called jobs. By analyzing the computing logic of the job, it is divided into smaller sub-loads that can be executed sequentially or in parallel according to the computing logic of the job, so as to be better executed in multiple steps. A job can be divided into several tasks, and a task can be divided into several instances, thus forming a hierarchical structure of "job-task-instance". In these application scenarios, delay-tolerant workloads can be various jobs at the job granularity, various tasks at the task granularity, or various instances at the instance granularity, and there is no limitation on this.

[0037] In this embodiment, there is a certain data dependency relationship between the delay-tolerant workloads. The data dependency relationship between the delay-tolerant workloads can be described as a directed acyclic graph (DAG), which can be expressed as in represents the sub-load, is the workload j and Data dependency between workloads. According to the relationship of DAG, a workload can have one or more preceding or succeeding workloads. The preceding workloads and succeeding workloads with data dependencies can be called parent workloads and child workloads, respectively. Child workloads can only start after the execution of all parent workloads is completed. As shown in Figure 2, there are 8 workloads, namely workloads 1-8, among which the child loads of workload 1 include workloads 2, 3 and 5, the child loads of workload 2 include workloads 4 and 6, the child loads of workload 3 include workload 4, the child loads of workload 4 include workloads 6 and 7, the child loads of workload 5 include 8, the child loads of workload 6 include 8, and the child loads of workload 7 include 8.

[0038] In this embodiment, latency-tolerant workloads not only have data dependencies but also have different types. Specifically, based on whether the latency-tolerant workloads exhibit periodicity, they can be divided into periodic workloads and non-periodic workloads. For example, batch workloads can be divided into periodic batch workloads and non-periodic batch workloads.

[0039] In this embodiment, when a customer submits each workload to a data center, information about each workload can be obtained, and the workload information includes at least workload type information and data dependencies between workloads. It should be noted that customers can submit workloads to a data center at different times, that is, the time at which different workloads are submitted to a data center can be the same or different. After obtaining the workload information, a load scheduling model for multiple workloads in the time dimension can be constructed based on the type information of each workload included in the information and the data dependencies between multiple workloads. In this embodiment, the load scheduling model is a model to be solved or optimized, and the quantity to be solved in the model is the scheduling time of the workload. When workloads are scheduled at different times, different scheduling schemes can be generated. In addition, in the disclosed embodiment, the load scheduling model can reflect the inherent time uncertainty of the workload. The inherent time uncertainty of the workload can also be called inherent time flexibility, which mainly refers to the uncertainty of the arrival time and deadline of the workload.

[0040] Optionally, as shown in FIG3 , an optional implementation method for constructing a load scheduling model includes the following steps:

[0041] 1021. Determine a scheduling time window for each workload based on the type information of each workload and the data dependency relationship between the multiple workloads. The scheduling time window includes a start time and an end time of the workload.

[0042] 1022. According to the scheduling time window of each workload, the scheduling status of each workload in the time slot is used as a variable to construct the objective function and constraint conditions corresponding to each workload to obtain the load scheduling model of each workload in the time dimension; wherein, the time slot is a time gap on any time scale, and the objective function represents the possible scheduling method of the workload in each time slot included in its scheduling time window.

[0043] In the disclosed embodiment, considering that the time-related parameters of the workload may be affected by real-time performance and resource allocation, to simplify the problem, in the disclosed embodiment, the arrival time and deadline of the workload are considered uncertain, while the execution duration of the workload is considered a static parameter.

[0044] In addition, in the embodiment of the present disclosure, the arrival time of the workload is considered to be random and conforms to a certain probability distribution. In an optional embodiment, in order to model the randomness of the arrival time of the workload, it is assumed that it conforms to a normal distribution around the predicted start time; that is, the distribution of the arrival time of the workload is modeled as The mean is express The start time of workload j, with standard deviation σ j Calculations can be performed based on historical data. Whether it's a periodic or aperiodic workload, the above method can be used to determine its start time. Once the start time of each workload is predicted, the deadline issue for periodic workloads is also resolved accordingly. Considering that the execution of aperiodic workloads is highly affected by priority levels, the disclosed embodiment can apply the property that aperiodic workloads conform to a normal distribution to their deadlines.

[0045] Based on the above, in step 1021, optional implementations of determining the scheduling time window for each workload based on the type information of each workload and the data dependencies between the multiple workloads include:

[0046] Regarding periodic workloads: Periodic workloads involve computing tasks that run periodically, such as daily work reports required by a company's marketing department. The execution of periodic workloads is stored and repeatedly triggered for processing on demand. Therefore, the deadlines for periodic workloads are constrained by the specified completion time required by the customer. Periodic workloads are assigned the same schedule and execution order across different periods. In this case, even if the parent periodic workload completes before its deadline, the subsequent periodic workloads, as child workloads, will still have to wait for their scheduled start time. This demonstrates the potential for periodic workloads to have significant time flexibility in production.

[0047] For periodic workloads, we can first predict the start time of the periodic workload based on the probability distribution function that the periodic workload's start time conforms to, such as the normal distribution. Taking the normal distribution as an example, we can obtain that the start time of the periodic workload is the mean of the normal distribution. Once the data dependencies and start times of the periodic workloads are known, we can use the following formula (1) to determine the deadline of each periodic workload.

[0048] In formula (1), represents the deadline of workload j; child(j) represents the set of sub-workloads of workload j; Represents workload Start time of workload is any sub-load in the sub-load set of workload j, Represents a collection of periodic workloads.

[0049] Formula (1) indicates that the deadline of a periodic workload as a parent workload is the earliest start time of its child workloads. That is, for any periodic workload, the earliest start time among the start times of other periodic workloads that have data dependencies on it can be used as the deadline of the periodic workload. When the start time and deadline of each periodic workload are obtained, the discrete scheduling time window of periodic workloads with complex DAG hierarchical structure is

[0050] Continuing with Figure 2, Figure 2 not only shows the data dependency relationships between workloads (also referred to as parent-child relationships), but also shows the start, end, and deadline times of workloads 1-8 in a timeline format. The parent-child relationships between workloads 1-8 and the start, end, and deadline times of each workload are shown in Table 1 below:

[0051] Table 1

[0052] Regarding non-periodic workloads: non-periodic workloads are dynamically generated by customers at any time, they exhibit highly random patterns, and their DAG structures are unknown before they actually reach the production environment. Therefore, the above-mentioned way of calculating the deadline of periodic workloads cannot be applied to non-periodic workloads. Compared with periodic workloads, non-periodic workloads usually have lower priorities and can tolerate longer execution delays. In the disclosed embodiment, the acceptable delay of each non-periodic workload can be estimated by its execution time, because a longer execution time usually reflects a heavier computing load and leads to a higher tolerance for processing delays, and vice versa. In addition, given the strong correlation between the acceptable delay of a workload and the priority of the workload, in this embodiment, the priority of the non-periodic workload is also considered when calculating the deadline of the non-periodic workload.

[0053] In the disclosed embodiments, for a non-periodic workload, the start time of the non-periodic workload can first be predicted based on the probability distribution function that the start time of the non-periodic workload conforms to. Taking a normal distribution as an example, the start time of the non-periodic workload can be calculated as the mean of the normal distribution. Next, the deadline for the non-periodic workload can be generated based on the workload's acceptable latency, priority, execution duration, and a preset deadline factor.

[0054] Alternatively, the acceptable delay for the non-periodic workload can be estimated using equations (2)-(3), and the deadline for the non-periodic workload can be determined using equation (4).

[0055] In formulas (2)-(4), represents a collection of aperiodic workloads; represents the start time of the aperiodic workload j, represents the end time of the non-periodic workload j, represents the deadline of aperiodic workload j, represents the execution time of the non-periodic workload j, Denotes the acceptable delay of non-periodic workload j, which is proportional to the workload priority and its execution duration, and multiplied by the deadline factor α. Specifically, I j is the priority level assigned to aperiodic workload j, I j The value of the priority sequence Priority, where "1" represents the highest priority. Indicates the lowest priority. The acceptable delay for aperiodic workloads is related to the deadline factor. A higher deadline factor indicates a greater acceptable delay for aperiodic workloads. Consequently, the acceptable delay allows aperiodic workloads to provide greater power flexibility. Optionally, the deadline factor α can be an empirical value, and its value is not restricted.

[0056] After obtaining the start time and deadline of each workload, the scheduling time window of each workload can be obtained, which means that each workload needs to be scheduled and completed within its scheduling time window. On this basis, in step 1022, the objective function and constraints corresponding to each workload can be constructed according to the scheduling time window of each workload, with the scheduling status of each workload in the time slot as a variable, to obtain the load scheduling model of each workload in the time dimension. It can be understood that the load scheduling model of each workload in the time dimension includes an objective function and corresponding constraints, wherein the objective function expresses the possible scheduling method of the workload in the time slot included in its scheduling time window, that is, it describes how to schedule the workload in its scheduling time window, and how to schedule is a quantity to be determined or optimized. Constraints are conditions that need to be met to optimize and solve the objective function.

[0057] In this embodiment, workload scheduling within a data center adheres to three key principles: First, the data center must determine when to execute workloads based on their heterogeneous timing constraints and sequencing; second, workload execution must be uninterruptible, as internal workload execution delays are negligible; and third, the IT power load introduced by the workloads must be considered as flexible capacity participating in a multi-tiered electricity market.

[0058] Based on the above principles, for any workload, we can use the workload's scheduling status within a time slot as a variable to construct an equation that indicates the workload must be completed within the L time slots contained in the workload's scheduling time window. This equation serves as the objective function, with L being an unknown quantity. Then, using the workload's scheduling status within a time slot as a variable, we construct a first constraint that indicates the workload's scheduling time must be no later than the workload's deadline shifted forward by L time slots, and no earlier than the workload's start time. Furthermore, using the workload's scheduling status within a time slot as a variable, we construct a second constraint that indicates the workload must be completed before the workload's deadline. Ultimately, we obtain a load scheduling model for this workload in the time dimension.

[0059] Specifically, workload j can be used as an example to illustrate the process of constructing the load scheduling model corresponding to each workload. Specifically, the objective function can be constructed according to formula (5), and the constraints can be constructed according to formulas (6)-(7).

[0060] In formula (5), A binary variable representing the scheduling status of workload j in time slot k, with a value of 0 or 1. When the value is 1, it means that workload j is scheduled in time slot k. When the value is 0, it means that workload j is not scheduled in time slot k. Formula (5) shows that workload j needs The specific time slots are not determined and need to be determined by optimization. Here, in order to distinguish the number of time slots required for different workloads, the symbol with the subscript j is used. It can also be represented by the above symbol L, which is not limited to this. NT represents the time range of the time slot value, and time slot k belongs to NT.

[0061] Formula (6) indicates that the start time of workload j is no later than and no earlier than To ensure that there is enough time to execute workload j. Formula (7) indicates that workload j must be completed before its deadline. The combination of formula (6) and formula (7) ensures that the execution of workload j will not be interrupted.

[0062] It should be noted that the time granularity of time slots is not limited here. In the disclosed embodiments, workload scheduling in a data center is combined with the data center's electricity cost in a multi-tiered electricity market. The workload scheduling method affects the data center's electricity cost. Therefore, the scheduling of workloads in the data center can be optimized with the goal of minimizing the data center's electricity cost in the multi-tiered electricity market, resulting in a more reasonable workload scheduling method. The workload scheduling method specifies the specific time slots within which each workload is scheduled. Based on this, when combined with a multi-tiered electricity market, the time granularity of time slots can correspond to the multiple time scales used for electricity trading in the multi-tiered electricity market, and can be time slots within any time scale. Taking a three-tiered electricity market consisting of day-ahead, midday, and real-time as an example, time slots can be hours within a 1-hour time scale (each hour is a time slot within the 1-hour time scale), 15 minutes within a 15-minute time scale (each 15 minutes is a time slot within the 15-minute time scale), or 5 minutes within a 5-minute time scale (each 5 minutes is a time slot within the 5-minute time scale).

[0063] Based on the above analysis, after obtaining the load scheduling model of each workload in the time dimension, it is also possible to construct a power consumption model of the workload in each time slot with the time slot as the granularity. This power consumption model represents the power consumption (or power load) required by each workload scheduled in a time slot. Specifically, the energy consumption of each workload can be predicted based on the computing resource usage of each workload; generally speaking, the more computing resources are used, the greater the energy consumption of the workload, and vice versa; then, combining the load scheduling model of each workload in the time dimension constructed above and the power consumption of each workload, it is possible to construct a power consumption model of the workload in a time slot with the time slot as the granularity.

[0064] Specifically, the power consumption model within a time slot can be constructed using formula (8). Formula (8) represents the power consumption required by all workloads being executed within time slot k.

[0065] In formula (8), represents the power consumption of workload j, A binary variable indicating that workload j is scheduled in time slot k; NJ represents the set of workloads scheduled in time slot k.

[0066] The sum of the power consumption of these workloads across all time slots is provided by the data center's power system. In the disclosed embodiments, all systems capable of providing power to the data center are collectively referred to as the data center's power system. The data center's power system includes at least a multi-tiered electricity market, from which the data center can purchase electricity resources. Electricity prices in multi-tiered electricity markets fluctuate dynamically and exhibit randomness or temporal uncertainty. Therefore, the workload scheduling method directly affects the cost of electricity purchased from the multi-tiered electricity market. To reduce the cost of electricity consumed by the data center from the multi-tiered electricity market, workloads can be scheduled to time slots with lower electricity prices, as much as possible, based on the workload's acceptable latency. This allows the data center's power cost and workload scheduling to be jointly optimized, resulting in a more reasonable workload scheduling solution with the goal of reducing the data center's power cost. This solution can also provide a more optimal power procurement solution for the data center's power purchases from the multi-tiered electricity market.

[0067] To this end, the following will introduce the energy management framework of the data center:

[0068] In the disclosed embodiments, data centers participate in multi-tiered electricity markets and purchase electricity at different timescales based on the multiple timescales of electricity trading in these markets. For example, using a three-tiered electricity market consisting of day-ahead, mid-day, and real-time, data centers can bid for hourly energy contracts for the operating day in the day-ahead market; adjust purchased electricity in the intraday market based on forecast errors and deviations from expected energy usage; and clear electricity at the real-time price based on previously established contracts in the real-time market.

[0069] As shown in Figure 4a, it shows the discretization relationship of the three-level power market including day-ahead, mid-day and real-time on the time scale. An operating day consists of 24 hours, that is, the time intervals of the first time scale are represented as One hour consists of four 15-minute intervals, i.e., the time intervals of the second time scale are expressed as A 15-minute interval consists of three 5-minute intervals, i.e., the time intervals of the third time scale are expressed as Among them, the first time scale corresponds to the first stage, the second time scale corresponds to the second stage, and the third time scale corresponds to the third stage.

[0070] In the disclosed embodiments, the data center's power source information includes, but is not limited to, electricity purchased from a multi-tiered power market. Data centers can deploy both conventional generators and renewable energy sources, so the data center's power source information can also include electricity generated by conventional generators and electricity generated by renewable energy sources. Furthermore, the disclosed technical solutions encourage the use of renewable energy to achieve low-carbon data center operations.

[0071] Among them, in addition to the uncertainty of electricity prices in the multi-level electricity market (that is, electricity prices will fluctuate over time), renewable energy includes but is not limited to solar energy, wind energy, tidal energy, etc., and the power generation of these renewable energy sources also has the characteristics of changing over time. In the embodiment of the present disclosure, the power generation of renewable energy can be modeled as a Gaussian random process. In other words, the power generation of renewable energy at the kth time slot in the sth scenario conforms to the Gaussian random process, and its power generation can be expressed by the formula Given. Among them, represents the power generation of renewable energy at the kth time slot in the sth scenario; and represents the mean and variance within the kth time slot. The sth scenario is any one of the K scenarios, which are multiple possible scenarios determined based on the uncertainty of the power source at multiple time scales. Each scenario occurs with a certain probability, called the probability of occurrence of each scenario, and K is a positive integer. A description of the K scenarios can be found in the subsequent embodiments of this disclosure and will not be detailed here.

[0072] In this embodiment, based on the clear information about the power source of the data center, in order to facilitate the optimization of the power cost of the data center, based on the number of levels in the multi-level power market, a power cost optimization model corresponding to multiple stages of the data center can be constructed according to the power source information of the data center, so as to optimize the power cost of the data center at multiple time scales corresponding to the multi-level power market. In the embodiment of the present disclosure, the power cost optimization model for multiple stages of the data center reflects the external uncertainty of various power sources, such as the fluctuation of electricity prices in the multi-level power market, the randomness of the power generation of renewable energy, etc. The power cost optimization model for multiple stages represents the power cost management model of the data center at multiple time scales, and the power cost at each stage or time scale is to be optimized.

[0073] In an optional embodiment, constructing a multi-stage power cost optimization model corresponding to a data center based on the power source information of the data center includes: constructing a global power cost function corresponding to the data center based on the power source information of the data center; constructing multi-stage constraints based on the constraint information of various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model constructed above to multiple time scales; and using the global power cost function and the multi-stage constraints as the multi-stage power cost optimization model corresponding to the data center. The global power cost function reflects the cost required for the data center to obtain power from various power sources; and the multi-stage constraints are used to constrain the conditions that various power sources should meet in each stage.

[0074] In the various embodiments of this disclosure, the types and quantities of a data center's power sources are not limited. Different types and quantities of power sources will result in different cost information items included in the global power cost function. In some application scenarios, a data center's power supply sources include electricity purchased from a multi-tiered power market, electricity generated by conventional generators, and electricity generated by renewable energy. The following embodiments will use this as an example to illustrate the construction of a global power cost function and the constraints at multiple stages.

[0075] In the above application scenario, the global electricity cost function includes: the electricity procurement cost of the data center from the multi-level electricity market, the penalty cost for abandoning renewable energy, and the expected operating cost of the data center's own power supply. Among them, the expected operating cost of the data center's own power supply includes the power generation cost and startup and shutdown costs of conventional generators. The electricity procurement cost is settled based on the electricity price in the multi-level electricity market. In the embodiment of the present disclosure, the use of renewable energy is encouraged. When the electricity generated by renewable energy is not fully used, a certain penalty will be imposed. The situation where the electricity generated by renewable energy is not fully used is referred to as abandoning renewable energy. Based on this, the penalty cost at least includes the penalty for abandoning renewable energy. Furthermore, the penalty cost can also include the penalty caused by the deviation between the contracted electricity consumption and the actual electricity consumption in the electricity market at each level, such as the penalty for over-purchasing electricity or the penalty for under-purchasing electricity, so as to improve the rationality of the data center's electricity procurement in the electricity market at each level.

[0076] In the disclosed embodiment, K scenarios are pre-generated, taking into account the uncertainty of various power sources at multiple time scales. Therefore, when constructing a global power cost function for a data center based on its power source information, the power cost function for the data center under these K scenarios can be constructed based on the power source information. The power cost functions for the data center under these K scenarios are then weighted and summed based on the probabilities of their occurrence to form the global power cost function for the data center.

[0077] Among them, by dividing the electricity cost function in each scenario from the perspective of electricity cost type, it can be obtained that the electricity cost function in each scenario includes: the electricity procurement cost of the data center purchasing electricity from the multi-level electricity market in this scenario, the penalty cost of abandoning renewable energy in this scenario, and the expected operating cost of the data center's self-powered power supply in this scenario.

[0078] Furthermore, by dividing the electricity cost function in each scenario from the perspective of multiple time scales, it can be obtained that the electricity cost function in each scenario includes at least: the expected operating cost of the data center's self-power supply at the maximum time scale and the electricity procurement cost of purchasing electricity in the multi-level electricity market, the adjustment cost of the expected operating cost of self-power supply at non-minimum and non-maximum time scales and the electricity procurement cost of purchasing electricity in the multi-level electricity market, as well as the penalty cost of abandoning renewable energy at the minimum time scale and the penalty cost caused by the deviation between actual electricity consumption and contracted electricity consumption in electricity markets at all levels.

[0079] In order to better understand the global electricity cost function, the electricity cost function in each scenario, and the constraint functions of multiple stages in the embodiments of the present disclosure, in the following embodiments, a multi-level electricity market is taken as an example of a three-level electricity market. The three-level electricity market corresponds to three time scales, which include the first time scale, the second time scale, and the third time scale from large to small. The discretization relationship of the three time scales is shown in Figure 4a. In addition, the three-level electricity market can include a day-ahead market, an intraday market, and a real-time market. In addition, in the following embodiments, the power source of the data center includes electricity purchased from the multi-level electricity market, electricity generated by conventional generators, and electricity generated by renewable energy.

[0080] Specifically, as shown in FIG4b , the process of constructing the global electricity cost function includes the following steps:

[0081] 1031. For each scenario, construct an electricity cost function for any time slot on the first time scale based on the power generation cost of the self-powered generator in the data center, the power generation and startup cost in any time slot on the first time scale, and the electricity price and power purchase amount in any time slot on the first time scale;

[0082] 1032. For each scenario, construct an electricity cost function for any time slot on the second time scale based on the power generation cost of the self-powered generator in the data center and the power generation adjustment amount in any time slot on the second time scale, as well as the electricity price and power purchase amount in any time slot on the second time scale;

[0083] 1033. For each scenario, construct an electricity cost function for any time slot on the third time scale based on the electricity price, the underbidden electricity purchase amount, the overbid penalty coefficient, and the overbid electricity purchase amount in any time slot on the third time scale, as well as the amount of renewable energy abandoned and the penalty coefficient for abandoned renewable energy in any time slot on the third time scale;

[0084] 1034. For each scenario, generate an electricity cost function for the scenario based on the electricity cost function for any time interval at the first time scale, the electricity cost function for any time interval at the second time scale, and the electricity cost function for any time interval at the third time scale;

[0085] 1035. Perform a weighted summation of the power cost functions under the K scenarios according to the probability of occurrence of the K scenarios, and use the sum as the global power cost function corresponding to the data center.

[0086] In the above electricity cost function, the amount of electricity purchased and the amount of power generated by the generator in any time interval at the first time scale, as well as the amount of power generation adjustment and the amount of electricity purchased in any time interval at the second time scale are variables to be optimized, and these variables form the electricity procurement plan.

[0087] The global electricity cost function constructed based on the above construction process can be expressed using formula (9).

[0088] In formula (9), ω represents the scene, N Ω represents a scene set. In the embodiment of the present disclosure, it is assumed that the scene set includes K scenes; π ω represents the probability of the scene ω appearing; NT da represents the set of time intervals in the first time scale, t represents the tth time interval in the first time scale, and NT id represents the set of time slots in the second time scale, m represents the mth time slot in the second time scale; NT rt represents a set of time slots at the third time scale, and n represents the nth time slot at the third time scale.

[0089] In formula (9), i represents the generator i of the data center, NI is the set of generators in the data center; represents the power generation cost of generator i, and its unit can be electricity price unit / power unit (such as kW); It represents the power generation (or output) of generator i at hour t under scenario s. Its unit can be power unit, such as kW; It represents the startup cost of generator i in the first time scale (i.e., in the day-ahead market), and its unit can be the electricity price unit. In formula (9), represents the expected operating cost of generator i generating electricity at hour t in the day-ahead market under scenario s.

[0090] In formula (9), represents the electricity price in the day-ahead market at hour t under scenario s; represents the amount of electricity purchased by the data center in the day-ahead market at hour t. In formula (9), represents the electricity procurement cost of the data center in the day-ahead market at hour t under scenario s.

[0091] In formula (9), represents the power generation adjustment of generator i in the mth 15-minute time slot in the tth hour under scenario s, and its unit is power unit, such as kW; represents the electricity price of the mth 15-minute time slot in the intraday market in the tth hour under scenario s, represents the amount of electricity purchased in the mth 15-minute time slot in the tth hour in the intraday market under scenario s. In formula (9), represents the expected operating cost of generator i incurred by adjusting power generation in the mth 15-minute time slot in the tth hour in the intraday market under scenario s; represents the electricity procurement cost of the data center in the mth 15-minute time slot in the tth hour in the intraday market under scenario s.

[0092] In formula (9), represents the electricity price of the nth 5-minute time slot in the mth 15-minute time slot in the tth hour in the intraday market under scenario s; represents the amount of electricity purchased that is underbid in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s, referred to as underbid electricity purchased; rt Indicates the penalty coefficient for over-bidding; The value represents the amount of electricity purchased that was bid too high in the nth 5-minute time slot of the mth 15-minute time slot in hour t under scenario s, referred to as the overbid. Whether the bid is overpriced or underpriced is determined by comparing the electricity price in the power purchase contract with the real-time electricity price in the nth 5-minute time slot. If the price is higher than the real-time electricity price, this amount is considered overbid. If the price is lower than the real-time electricity price, this amount is considered underbid.

[0093] In formula (9), p pen represents the penalty coefficient for abandoning renewable energy generation; represents the amount of renewable energy that is abandoned in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s.

[0094] This embodiment aims to minimize the sum of the expected operating costs, electricity procurement costs, and penalty costs of data centers in a multi-level electricity market structure while satisfying certain constraints, thereby helping to reduce carbon emissions and promoting the power system's absorption of renewable energy.

[0095] In this embodiment, in addition to constructing a global power cost function, constraints for multiple phases must also be constructed. For any given time scale, based on the constraints of the various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model to that time scale, power source constraints and power balance constraints are constructed for the corresponding phase of that time scale. Power source constraints include at least one of generator operation constraints, renewable energy constraints, and power market transmission constraints. Power source constraints vary across different phases.

[0096] In the case of a three-tiered multi-level electricity market, the multiple time scales include, from largest to smallest, a first time scale, a second time scale, and a third time scale. Accordingly, the multiple stages include a first stage, a second stage, and a third stage corresponding to the three time scales. Based on this, the multi-stage constraints include constraints for the first stage, constraints for the second stage, and constraints for the third stage. These are described below:

[0097] Constraints for the first stage:

[0098] On the one hand, based on the power consumption information obtained by mapping the global power consumption model to any time interval on the first time scale and the power consumed by the data center during any time interval on the first time scale, equations with equivalent relationships are constructed as the power balance constraints for the first stage. On the other hand, based on the power generation, ramp rate upper and lower limits, and startup cost of the data center's self-powered generators during any time interval on the first time scale, power constraint formulas, start-stop constraint formulas, and start-stop cost constraint formulas are constructed as the generator operation constraints for the first stage. Furthermore, based on the power generation and power generation upper and lower limits of renewable energy in the data center during any time interval on the first time scale, renewable energy constraints for the first stage are constructed. The power generation of renewable energy conforms to a Gaussian distribution. Based on the power purchase amount and transmission line capacity upper and lower limits of the data center during any time interval on the first time scale, power market transmission constraints for the first stage are constructed. The power consumption information obtained by mapping the global power consumption model to any time interval on the first time scale refers to the power consumption information obtained by replacing the timestamps in the global power consumption model with the power consumption information obtained for each time interval on the first time scale.

[0099] Specifically, the constraints for the first stage can be constructed using equations (10)-(17). Equation (10) represents the power balance constraints in the day-ahead market. Equations (11)-(14) describe the generator operation constraints, including power constraints, start-stop constraints, and start-stop costs. Equations (15)-(16) represent the power constraints of renewable energy (i.e., renewable energy constraints). Equation (17) describes the transmission constraints of multiple power markets, i.e., power market transmission constraints. Eqs.(5)-(8),where k=t. Formula (18)

[0100] In formulas (10)-(17), represents the electricity generation of renewable energy at hour t under scenario s; represents the amount of electricity purchased at hour t in the day-ahead market; represents a binary variable indicating the start / stop status of generator i at hour t, and its value is 0 or 1; when the value is 0, it means that generator i is not started, and when the value is 1, it means that generator i is started; represents the power generation of generator i at hour t; It represents the power load of the data center at the tth hour in the s scenario, that is, the power consumption information generated at the tth hour of the first time scale according to the global power consumption model.

[0101] In formulas (10)-(17), represents the lower limit of power generation of generator i, Represents the upper limit of power generation of generator i; RD i ,RU i Indicates the ramp rate limit of generator i, one is the lower limit value and the other is the upper limit value; represents the start-up and shutdown cost of generator i; represents the lower limit of renewable energy generation at hour t under scenario s, represents the upper limit of renewable energy generation at hour t under scenario s; represents the wind power generation at hour t, Indicates the photovoltaic power generation or solar power generation at hour t. In addition, P g It represents the lower limit of transmission line capacity in the electricity market; Represents the upper limit of transmission line capacity in the electricity market.

[0102] Furthermore, under the constraints of formulas (10)-(17), the goal is to minimize formula (9) and apply the load scheduling model and global power consumption model in formulas (5)-(8) to optimize the workload in each 1-hour time slot of the first time scale to obtain a scheduling solution on the first time scale. Formula (18) represents the application of formulas (5)-(8) in the optimization process.

[0103] The constraints of the second stage are:

[0104] On the one hand, based on the power consumption information obtained by mapping the global power consumption model to any time slot in the second time scale and the amount of electricity obtained by the data center in any time slot in the second time scale, an equation with an equivalent relationship can be constructed as the power balance constraint condition for the second stage. On the other hand, based on the power generation adjustment amount and the upper and lower limits of the ramp rate of the generators that supply power to the data center in any time slot in the second time scale, as well as the power generation in the time slot corresponding to the second time scale in the first time scale, the power generation constraint formula and the power generation adjustment constraint formula of the generator are constructed as the generator operation constraint condition for the second stage; and based on the power generation amount and upper and lower limits of the power generation of renewable energy in the data center in any time slot in the second time scale, the renewable energy constraint condition for the second stage is constructed; based on the power purchase amount and the upper and lower limits of the transmission line capacity of the data center in any time slot in the second time scale, the power market transmission constraint condition for the second stage is constructed. Among them, the power consumption information obtained by mapping the global power consumption model to any time slot in the second time scale refers to replacing the timestamp in the global power consumption model with the power consumption information obtained in each time slot in the second time scale.

[0105] In detail, the constraints of the second stage can be constructed using formulas (19)-(24). In the formulas of the second stage, m is introduced in the subscript of the variable identifier, indicating the mth time interval in the tth hour, where m∈NT id :{1, 2, 3, 4}. Variables with different subscripts in the same time slot m share the same information in the first stage, which satisfies the non-predictable constraints in this multi-stage problem. Among them, formula (19) represents the power balance constraint in the second stage, allowing adjustments to planned output and power demand. Formula (20) shows the limit of the upward / downward ramp rate, that is, the power generation adjustment constraint. The power generation output is limited by formulas (21) and (22), that is, the power generation constraint formula of the generator. Renewable energy generation and electricity purchased from the multi-level power market should meet technical limitations, such as the renewable energy constraints shown in formulas (23) and (24). It can be a positive or negative number, indicating that the data center will increase or decrease the amount of electricity it purchases based on the intraday forecast results. Eqs.(5)-(8),where k=(t,m). Formula (25)

[0106] In the above formulas (19)-(24), represents the electricity generation of renewable energy in the mth 15-minute time slot in the tth hour under scenario s; represents the amount of electricity purchased in the mth 15-minute time slot in the tth hour in the intraday market under scenario s; represents the power generation of generator i in the mth 15-minute time slot in the tth hour under scenario s; It represents the data center power load in the mth 15-minute time slot in the tth hour under scenario s, that is, the power consumption information generated by mapping the global power consumption model to the mth 15-minute time slot in the tth hour of the second time scale.

[0107] In the above formulas (19)-(24), represents the power generation adjustment of generator i in the mth 15-minute time slot in the tth hour under scenario s; represents the power generation increase of generator i in the mth 15-minute time slot in the tth hour under scenario s; represents the power generation reduction of generator i in the mth 15-minute time slot in the tth hour under scenario s; A binary variable representing the scheduling status of workload j in the mth 15-minute time slot in the tth hour, with a value of 0 or 1; when the value is 1, it means it is scheduled; when the value is 0, it means it is not scheduled. In addition, represents the lower limit of renewable energy generation in the mth 15-minute time slot in the tth hour under scenario s, represents the upper limit of renewable energy generation in the mth 15-minute time slot in hour t under scenario s. For the meaning of symbols not explained here, refer to the relevant descriptions in the preceding or subsequent formulas.

[0108] In addition, under the constraints of formulas (19)-(24), the load scheduling model and global power consumption model in formulas (5)-(8) can be used to optimize the workload in each 15-minute time slot of each hour in the second time scale under the minimum of formula (9) to obtain a scheduling solution on the second time scale. Formula (25) represents the application of formulas (5)-(8) in the optimization solution process. As shown in formula (25), based on the scheduling results of the first time scale, the data center is allowed to redistribute the workload on the second time scale to reduce the mismatch between power supply and demand caused by prediction errors and uncertainties.

[0109] Constraints for the third stage:

[0110] On the one hand, based on the power consumption information obtained by mapping the global power consumption model to any time slot on the third time scale and the power consumed by the data center during any time slot on the third time scale, an equation with an equivalent relationship is constructed as the power balance constraint condition for the third stage. On the other hand, based on the power generation and upper and lower limits of the power generation of renewable energy in the data center during any time slot on the third time scale, the renewable energy constraint condition for the third stage is constructed. Furthermore, based on the overpriced and underpriced power purchases by the data center during any time slot on the third time scale, the power purchases during the time slot corresponding to the third time scale on the second time scale, and the upper and lower limits of the transmission line capacity, the power market transmission constraint condition for the third stage is constructed. Based on the power generation and upper limit of the power generation of renewable energy in the data center during any time slot on the third time scale, the renewable energy abandoned by the data center during any time slot on the third time scale is obtained. The power consumption information obtained by mapping the global power consumption model to any time slot on the third time scale refers to replacing the time slots in the global power consumption model with the power consumption information obtained from each time slot on the third time scale.

[0111] In detail, the constraints of the third stage can be constructed using formulas (26)-(33). In the subscript of the variables or parameters in the formula of the third stage, n is introduced, where n∈NT rt :{1, 2, 3}. Therefore, the subscript n represents the value of the nth 5-minute interval in the mth 15-minute time slot in the tth hour. Wherein, formula (26) describes the power balance constraint in the third stage, where the power generation of the generators in the data center is determined by the decisions made in the intraday stage because they cannot adjust the power output within 5 minutes. Formulas (27)-(29) calculate the power purchased from the real-time market. Renewable energy generation and transmission in the multi-level power market are restricted in the third stage, as shown in formulas (30) and (31) by the renewable energy constraints and power market transmission constraints. Formula (32) gives the calculation method for the abandoned renewable energy generation, and the abandoned renewable energy (such as abandoned wind power) is represented as a penalty cost in the global power cost function. Eqs.(5)-(8),where k=(t,m,n). Formula (34)

[0112] In the above formulas (26)-(33), represents the electricity generation of renewable energy in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s; represents the amount of electricity purchased in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour of the electricity market under scenario s; represents the power load of the data center in the nth 5-minute time slot of the mth 15-minute time slot of the tth hour in scenario s, that is, the power consumption information generated by mapping the global power consumption model to the nth 5-minute time slot of the mth 15-minute time slot of the tth hour in the third time scale; It represents the amount of electricity purchased with too low a bid in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s; It represents the amount of electricity purchased with an overpriced bid in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s; A binary variable indicating whether the bid in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour in scenario s is too high or too low. Its value is 0 or 1. When the value is 0, it means the bid is too high, and when the value is 1, it means the bid is too low. represents the amount of renewable energy available that is abandoned in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s; A binary variable representing the scheduling status of workload j in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour, with a value of 0 or 1; when the value is 1, it means it is scheduled; when the value is 0, it means it is not scheduled. In addition, represents the lower limit of renewable energy generation in the nth 5-minute time slot of the mth 15-minute time slot in the tth hour under scenario s, represents the upper limit of renewable energy generation in the nth 5-minute time slot of the mth 15-minute time slot in hour t under scenario s. For the meaning of symbols not explained here, refer to the relevant descriptions in the preceding or subsequent formulas.

[0113] In addition, under the constraints of formulas (26)-(33), the load scheduling model and global power consumption model in formulas (5)-(8) can be used to optimize the workload in each 5-minute time slot of each 15-minute period of each hour in the third time scale, with the goal of minimizing formula (9) to obtain a scheduling solution on the third time scale. Formula (34) represents the application of formulas (5)-(8) in the optimization solution process. As shown in formula (34), based on the scheduling results of the second time scale, the data center is allowed to redistribute the workload on the third time scale to reduce the mismatch between power supply and demand caused by prediction errors and uncertainties.

[0114] After obtaining the global power cost function and the constraints of multiple stages, the global power cost function and the constraints of multiple stages can be combined with the load scheduling model, and the load scheduling model can be optimized and solved in multiple stages to obtain scheduling solutions for multiple workloads on multiple time scales. Specifically, the time slots in the objective function and constraints of the load scheduling model in the time dimension of each workload can be mapped to various time gaps on multiple time scales to obtain additional constraints on multiple stages; based on the constraints and additional constraints of multiple stages, the global power cost function can be optimized and solved in multiple stages with the goal of minimizing the global power cost to obtain scheduling solutions for multiple workloads on multiple time scales. The combined optimization process can be found in formulas (18), (25) and (34), which will not be repeated here.

[0115] Furthermore, the multi-stage optimization solution process in the embodiments of the present disclosure can employ stochastic optimization models, robust optimization models, distributed robust optimization models, and the like. Robust optimization and distributed robust optimization models provide robust optimization results based on the uncertainty set, but are less economically efficient than stochastic optimization models. Preferably, a stochastic optimization model is employed to obtain a multi-stage stochastic optimization model.

[0116] In the above embodiments of the present disclosure, external uncertainties such as electricity prices and renewable energy in the power system and internal uncertainties such as arrival time and deadline of workloads are taken into consideration at the same time, which can achieve coordination of uncertainties in a larger range and provide data centers with better power procurement strategies and generator operation plans in multi-level power markets, thereby obtaining more economical data center operation strategies and enhancing the service competitiveness of data centers.

[0117] About the generation and reduction of scenes:

[0118] In the embodiment of the present disclosure, a scenario generation method is used to represent the reality of the uncertainty in various power sources. For example, Monte Carlo simulation can be used to generate different scenarios for each uncertain parameter. Taking the three-level power market as an example, in the first stage, the simulation generates N ω1 scenarios, each containing 24 hours of uncertainty. ω1 Each of the scenes can generate N ω2 For each second-stage scenario, generate N ω3 Therefore, a total of N ω =N ω1 ×N ω2 ×N ω3That is, for each of the three-level power market, scenarios are generated to obtain N ω Scenes; N ω1 is the number of scenarios corresponding to the first-level electricity market (i.e., the number of scenarios in the first stage), N ω2 Yes N ω1 The number of scenarios that each scenario in the second-level electricity market corresponds to (that is, the number of second-stage scenarios that each first-stage scenario corresponds to), N ω3 Yes N ω2 The number of scenarios that each scenario in the third-level electricity market corresponds to (that is, the number of scenarios in the third stage that each scenario in the second stage corresponds to). ω1 、N ω2 、N ω3 and N ω All are positive integers.

[0119] In order to reduce the computational complexity, a scene reduction technique is used to reduce the original scene set to a smaller subset while keeping the basic features within an acceptable approximation range. ω The scenes are clustered to obtain K scenes, and the probability of each scene in the K scenes is calculated based on the way each random point is assigned to the nearest cluster centroid.

[0120] Alternatively, a k-medoids clustering algorithm can be used to cluster all scenarios into several clusters, and the probability of the selected scenario is calculated by assigning each random point to the nearest centroid. The center of each cluster is then selected as a representative scenario for analysis, and each scenario has its own probability of occurrence. The effectiveness of the clustering can be tested by using the squared error and the sum of squares for error (SSE) metrics to test the representativeness and distribution of the centroid.

[0121] In order to more fully illustrate the effect of the technical solution disclosed herein, a scenario test was conducted on the solution. The following is a description of the scenario test process and the effect.

[0122] (1) Scenario example setting

[0123] In this scenario calculation, electricity market price and renewable energy generation data were collected from sample information systems or websites accessible through the tertiary electricity market. To match the dispatch granularity of the second and third phases, a finer-grained linear resampling of the wind and solar power generation datasets with a 1-hour resolution was performed. Using Latin hypercube sampling as a scenario generation technique, multiple scenarios were considered in each phase. Each scenario included wind power, photovoltaic power generation, and electricity prices. After applying the K-medoids technique to reduce the number of scenarios and evaluating the results using the SSE index, five scenarios were selected with time granularities of 1 hour, 15 minutes, and 5 minutes. Across the day-ahead, intraday, and real-time phases, electricity market prices fluctuated in all five scenarios. The output power of photovoltaic power generation also fluctuated in all five scenarios, as did the output power of wind power generation. In this scenario calculation example, the following assumptions can be made: electricity prices in the tertiary power market peak between 6:00 PM and 10:00 PM, and intraday and real-time markets are more volatile than the day-ahead market. Accordingly, solar panels (i.e., photovoltaic power generation) primarily generate electricity between 10:00 AM and 4:00 PM, with a maximum output of approximately 45-50 kilowatts. Wind turbines operate continuously throughout the day, reaching a peak power output of approximately 60 kilowatts around 3:00 AM and 11:00 AM. Furthermore, the probability of scenarios S1-S5 occurring in each of the three phases varies, and the present disclosure does not focus on specific probability values.

[0124] In this scenario, workload arrival and deadline times are considered discrete parameters and estimated based on their uncertainty modeling. To obtain the workload data required for this example, a method is used to extract periodic workloads from the data center's production log files and aperiodic workloads from publicly available data. The ratio of periodic and aperiodic workloads is set to 1:1. To account for workload arrival uncertainty and reduce computational burden, computational tasks are randomly sampled from the original dataset. Furthermore, to maintain the original distribution of workload arrivals within a day, proportional sampling is performed within each time interval. The estimation of workload arrival and deadline times is described above and will not be further elaborated here. In this scenario, the scheduling time window for periodic and aperiodic workloads includes the workload's original execution market and a workload's acceptable slack time window. Leveraging the workload's acceptable slack time window, workloads can be scheduled to execute during time slots with low electricity prices or relatively abundant renewable energy.

[0125] In this scenario, the power mapping model was developed using a data-driven approach, employing a back-propagation neural network (BPNN) to estimate the power consumption and resource utilization of each workload. Based on the above assumptions, the power flexibility provided by periodic workloads is primarily concentrated in the morning (from midnight to 10:00), while the power flexibility of aperiodic workloads exhibits a different trend, with greater flexibility during the midday period (from 10:00 to 16:00). This is because most periodic workloads typically need to be completed before morning, so they are executed early in the morning or late at night. Aperiodic workloads, on the other hand, are typically manually submitted by users and are therefore statistically distributed throughout the day. In this scenario, the power flexibility of aperiodic workloads is generally lower than that of periodic workloads. Notably, the power flexibility of aperiodic workloads can be quantified based on the deadline factor. A higher deadline factor indicates greater power flexibility for aperiodic workloads.

[0126] (2) Execution of the example

[0127] According to the solution provided by the above-mentioned embodiments of the present disclosure, a load scheduling model is constructed by utilizing the inherent uncertainty of the arrival time and deadline of the two workloads; a multi-stage electricity cost optimization model is constructed by utilizing the external uncertainty (or randomness) of electricity prices, renewable energy, etc. in the power source; and the load scheduling model is then optimized and solved in multiple stages based on the multi-stage electricity cost optimization model to obtain scheduling plans for the workloads on multiple time scales. At the same time, a better power procurement plan for the data center is obtained, which helps to reduce carbon emissions.

[0128] (3) Example results

[0129] In this example, a data center is powered by a three-tier electricity market, wind turbines, and solar power. Day-ahead scheduling decisions are made based on forecasted data and uncertainties across the three phases. The power purchase price in the day-ahead market serves as a reference for intraday and real-time scheduling, and determines the start and stop status of conventional generators.

[0130] Based on the above forecast, the data center anticipates lower electricity prices in the day-ahead market, while prices in the intraday and real-time markets fluctuate significantly and are generally higher. Because wind turbines can partially meet electricity demand during peak load periods (6:00 AM to 10:00 AM), the data center decides to purchase electricity from the day-ahead market to optimize overall costs. However, between 5:00 AM and 7:00 AM, prices in both the intraday and real-time markets are likely to fall, as wind and solar generation do not significantly increase during this period. Therefore, the data center needs to purchase electricity from both the intraday and real-time markets to meet the power needs of its workloads during this period.

[0131] In this scenario, the original predicted power demand curves of the data center in three phases were compared with the power demand curves optimized using the method provided by the embodiment of the present disclosure. It can be concluded that within each phase, the workload was rescheduled from the peak period to the low-price period, and the maximum load was reduced by about 50 kilowatts (about 15% of the total load), which occurred around 6 o'clock. This optimization is driven by the need to purchase electricity from the electricity market because during this period, when photovoltaic power generation is unavailable, wind power and conventional generators can partially meet the power demand. However, electricity prices are at their peak during this period, so the data center optimizes workload scheduling to periods with lower electricity prices to reduce costs.

[0132] (4) Sensitivity analysis

[0133] The solution proposed in the embodiments of the present disclosure is called a multi-stage stochastic programming (MSSP) scheme. The MSSP scheme is compared with three other schemes: a scheme that does not consider workload uncertainty, which can be expressed as MSSP-WL; a traditional two-stage stochastic programming scheme, which can be expressed as TSSP; and a TSSP scheme that does not consider workload uncertainty, which can be expressed as TSSP-WL. In the TSSP and TSSP-WL schemes, only the day-ahead and real-time phases of scheduling are performed.

[0134] A comparative analysis of the electricity costs of these schemes reveals that the MSSP scheme achieves costs 33.2% and 46.9% lower than those of the TSSP and TSSP-WL schemes, respectively. This is primarily because the TSSP scheme ignores the second-stage rescheduling, which adjusts the schedule based on more accurate forecasts of system uncertainties. In contrast, the MSSP scheme has a stronger ability to cope with system uncertainties and enables data centers to have temporal flexibility at different timescales. Furthermore, it is noteworthy that the curtailment rate of renewable energy is higher in the TSSP and TSSP-WL schemes than in the proposed MSSP scheme. When comparing the MSSP scheme with the MSSP-WL scheme, it is found that each component cost is lower, resulting in a total cost reduction of 5.6%. Similarly, the TSSP scheme achieves a 20.4% reduction in total cost compared to the TSSP-WL scheme. This finding demonstrates that considering workload uncertainty in data center load management strategies has a positive impact on reducing electricity costs. Furthermore, this scenario also considers data dependencies between workloads, which enables a more accurate and reliable estimation of data center power load flexibility.

[0135] Furthermore, in the MSSP scheme, different degrees of flexibility can be introduced for aperiodic tasks through the deadline factor to influence the results. In this example, the impact of deadline factors ranging from 2 to 20 on the total power flexibility and corresponding cost of aperiodic workloads is evaluated, and it is found that higher deadline factors can lead to lower overall costs.

[0136] It should be noted that the execution entity of each step of the method provided in the above embodiment can be the same device, or the method can be executed by different devices. For example, the execution entity of steps 101 to 103 can be device A; for another example, the execution entity of steps 101 and 102 can be device A, and the execution entity of step 103 can be device B; and so on.

[0137] In addition, some of the processes described in the above embodiments and the accompanying drawings include multiple operations that appear in a specific order, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0138] FIG5 is a schematic diagram of the structure of a data center workload scheduling device provided by an embodiment of the present disclosure. The device can be applied to data centers operating in a multi-level electricity market where electricity transactions occur on multiple time scales. As shown in FIG5 , the device includes:

[0139] An acquisition module 11 is configured to acquire information about multiple delay-tolerant workloads that the data center needs to process, the information about the multiple workloads including type information of each workload and data dependencies between the multiple workloads;

[0140] A first building module 12 is configured to build a load scheduling model for multiple workloads in a time dimension based on type information of each workload and data dependencies between the multiple workloads, wherein the load scheduling model reflects the time uncertainty of the workloads;

[0141] The second construction module 13 is configured to construct a power cost optimization model corresponding to multiple stages of the data center based on the power source information of the data center, so as to optimize the power cost of the data center at multiple time scales;

[0142] The optimization solution module 14 is configured to perform a multi-stage optimization solution on the load scheduling model based on the multi-stage power cost optimization model to obtain a scheduling solution for multiple workloads on multiple time scales;

[0143] The scheduling module 15 is configured to schedule the multiple workloads on multiple time scales respectively according to the scheduling schemes of the multiple workloads on multiple time scales, so as to reduce the power cost of the data center.

[0144] In an optional embodiment, the first construction module 12 is specifically configured to: determine the scheduling time window of each workload based on the type information of each workload and the data dependency relationship between multiple workloads, and the scheduling time window includes the start time and end time of the workload; based on the scheduling time window of each workload, with the scheduling status of each workload in the time slot as a variable, construct the objective function and constraint conditions corresponding to each workload to obtain a load scheduling model for each workload in the time dimension; wherein the time slot is a time gap at any time scale, and the objective function represents the possible scheduling method of the workload at each time slot included in its scheduling time window.

[0145] When determining the scheduling time window of each workload based on the type information of each workload and the data dependency relationship between multiple workloads, the first building module 12 is specifically configured as follows:

[0146] For periodic workloads, the start time of the periodic workload is predicted based on the probability distribution function that the periodic workload's start time conforms to. Based on the data dependencies between periodic workloads, for each periodic workload, the earliest start time of other periodic workloads with which it has data dependencies is used as its deadline.

[0147] For non-periodic workloads, the start time of the non-periodic workload is predicted based on the probability distribution function that the start time of the non-periodic workload conforms to; the deadline of the non-periodic workload is generated based on the acceptable delay, priority, execution time and preset deadline factor of the non-periodic workload.

[0148] The first construction module 12 is specifically configured to construct the objective function and constraint conditions corresponding to each workload based on the scheduling time window of each workload and the scheduling status of each workload in the time slot as a variable to obtain the load scheduling model of each workload in the time dimension:

[0149] For any workload, using the workload's scheduling state within a time slot as a variable, construct an equation representing the workload's need to be completed within L time slots included in the workload's scheduling time window as the objective function, where L is an unknown quantity.

[0150] Using the scheduling state of the workload in the time slot as a variable, construct a first constraint condition that indicates that the scheduling time of the workload is no later than the time corresponding to the workload's deadline shifted forward by L time slots, and no earlier than the workload's start time;

[0151] A second constraint condition is constructed, which indicates that the workload must be completed before the deadline of the workload, using the scheduling state of the workload in the time slot as a variable.

[0152] In an optional embodiment, the first construction module 12 is further configured to: predict the energy consumption of each workload based on the computing resource usage of each workload; and construct a power consumption model of the workload within a time slot with the time slot as the granularity based on the load scheduling model of each workload in the time dimension and the power consumption of each workload to obtain a global power consumption model.

[0153] In an optional embodiment, the second construction module 13 is specifically configured to: construct a global electricity cost function corresponding to the data center based on the power source information of the data center, the global electricity cost function including the power procurement cost of the data center from the multi-level power market, the penalty cost of abandoning renewable energy and the expected operating cost of the data center's self-power supply; construct constraints for multiple stages based on the constraint information of various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model to multiple time scales; and use the global power cost function and the constraints for multiple stages as a multi-stage power cost optimization model corresponding to the data center.

[0154] In an optional embodiment, when the second constructing module 13 constructs the global power cost function corresponding to the data center according to the power source information of the data center, it is specifically configured as follows:

[0155] Based on the power source information of the data center, constructing a power cost function of the data center under K scenarios; performing a weighted summation of the power cost functions under the K scenarios according to the probability of occurrence of the K scenarios to serve as a global power cost function corresponding to the data center; the K scenarios are multiple scenarios determined based on the uncertainty of the power source at multiple time scales, and K is a positive integer;

[0156] Among them, the electricity cost function in each scenario includes at least: the expected operating cost of the data center's self-power supply at the maximum time scale and the electricity procurement cost of purchasing electricity in the multi-level electricity market, the adjustment cost of the expected operating cost of self-power supply at non-minimum and non-maximum time scales and the electricity procurement cost of purchasing electricity in the multi-level electricity market, as well as the penalty cost of abandoning renewable energy at the minimum time scale and the penalty cost caused by the deviation between actual electricity consumption and contracted electricity consumption in electricity markets at all levels.

[0157] In an optional embodiment, the multi-level electricity market is a three-level electricity market, and the multiple time scales include, from largest to smallest, a first time scale, a second time scale, and a third time scale. Based on this, when constructing the electricity cost function of the data center under K scenarios, the second construction module 13 is specifically set to:

[0158] For each scenario, construct an electricity cost function for any time slot in the first time scale based on the power generation cost of the self-powered generator in the data center, the power generation and startup cost in any time slot in the first time scale, and the electricity price and power purchase amount in any time slot in the first time scale;

[0159] Constructing an electricity cost function for any time slot on the second time scale based on the power generation cost of the self-powered generator in the data center and the power generation adjustment amount in any time slot on the second time scale, as well as the electricity price and the power purchase amount in any time slot on the second time scale;

[0160] Based on the electricity price, the underbidden purchase quantity, the overbid penalty coefficient and the overbid purchase quantity in any time slot in the third time scale, and the amount of renewable energy abandoned and the penalty coefficient for abandoned renewable energy in any time slot in the third time scale, a power cost function for any time slot in the third time scale is constructed;

[0161] generating an electricity cost function for the scenario according to the electricity cost function for any time interval at the first time scale, the electricity cost function for any time interval at the second time scale, and the electricity cost function for any time interval at the third time scale;

[0162] The power purchase amount and power generation amount in any time interval under the first time scale, and the power generation adjustment amount and power purchase amount in any time interval under the second time scale are variables to be optimized.

[0163] In an optional embodiment, a scenario construction module is further included, which is configured to generate scenarios for the uncertain parameters in each of the three-level power markets to obtain N ω scenes; among them, Nω =N ω1 ×N ω2 ×N ω3 , N ω1 is the number of scenarios corresponding to the first-level electricity market, N ω2 Yes N ω1 The number of scenarios that each scenario in the second-level electricity market corresponds to, N ω3 Yes N ω2 The number of scenarios corresponding to each scenario in the third-level electricity market;

[0164] For the N ω The scenes are clustered to obtain K scenes, and the probability of each scene in the K scenes is calculated by assigning each random point to the nearest cluster centroid, N ω1 、N ω2 、N ω3 and N ω All are positive integers.

[0165] In an optional embodiment, when constructing the constraints of multiple stages, the second construction module 13 is specifically set to: for any time scale, based on the constraint information of various power sources adopted by the data center and the power consumption information obtained by mapping the global power consumption model to the any time scale, construct the power source constraints and power balance constraints of the corresponding stage of the any time scale; wherein the power source constraints include at least one of generator operation constraints, renewable energy constraints and power market transmission constraints.

[0166] In an optional embodiment, the multi-level electricity market is a three-level electricity market, and the multiple time scales include, from large to small, a first time scale, a second time scale, and a third time scale, and the multiple stages include a first stage, a second stage, and a third stage corresponding to the three time scales in sequence.

[0167] For the first stage corresponding to the first time scale, the second construction module 13 is specifically set to: construct an equation with an equivalent relationship as the power balance constraint condition of the first stage based on the power consumption information obtained by mapping the global power consumption model to any time slot under the first time scale and the power obtained by the data center in any time slot under the first time scale; construct the power constraint formula, start-stop constraint formula and start-stop cost constraint formula of the generator as the generator operation constraint condition of the first stage based on the power generation, upper and lower limits of ramp rate and startup cost of the generator powered by the data center itself in any time slot under the first time scale; construct the renewable energy constraint condition of the first stage based on the power generation and upper and lower limits of power generation of renewable energy in the data center in any time slot under the first time scale; the power generation of the renewable energy conforms to the Gaussian distribution; construct the power market transmission constraint condition of the first stage based on the power purchase amount and upper and lower limits of transmission line capacity of the data center in any time slot under the first time scale.

[0168] For the second stage corresponding to the second time scale, the second construction module 13 is specifically set to: construct an equation with an equivalent relationship as the power balance constraint condition of the second stage based on the power consumption information obtained by mapping the global power consumption model to any time slot under the second time scale and the power obtained by the data center in any time slot under the second time scale; construct the generator's power constraint formula and power generation adjustment constraint formula as the generator operation constraint condition of the second stage based on the power generation adjustment amount and the upper and lower limits of the ramp rate of the generator powered by the data center itself in any time slot under the second time scale, and the power generation in the time slot corresponding to the second time scale under the first time scale; construct the renewable energy constraint condition of the second stage based on the power generation and the upper and lower limits of the power generation of renewable energy in the data center in any time slot under the second time scale; construct the power market transmission constraint condition of the second stage based on the power purchase amount and the upper and lower limits of the transmission line capacity of the data center in any time slot under the second time scale.

[0169] For the third stage corresponding to the third time scale, the second construction module 13 is specifically set to: construct an equation with an equivalent relationship as the power balance constraint condition of the third stage based on the power consumption information obtained by mapping the global power consumption model to any time slot under the third time scale and the power obtained by the data center in any time slot under the third time scale; construct the renewable energy constraint condition of the third stage based on the power generation and the upper and lower limits of the power generation of the renewable energy in the data center in any time slot under the third time scale; construct the power market transmission constraint condition of the third stage based on the over-quoted power purchase amount, the under-quoted power purchase amount, the power purchase amount in the time slot corresponding to the third time scale under the second time scale, and the upper and lower limits of the transmission line capacity; obtain the renewable energy abandoned by the data center in any time slot under the third time scale based on the power generation and the upper limit of the power generation of the renewable energy in the data center in any time slot under the third time scale.

[0170] In an optional embodiment, the optimization solution module 14 is specifically configured to: map the objective function and time slots in the constraints of the load scheduling model of each workload in the time dimension to the time gaps on the multiple time scales to obtain the additional constraints on the multiple stages; and perform optimization solutions on the global power cost function in multiple stages with the goal of minimizing the global power cost according to the constraints and additional constraints of the multiple stages to obtain scheduling schemes for the multiple workloads on multiple time scales.

[0171] The device shown in FIG5 can execute the method shown in FIG1, and its implementation principle and technical effect are not described in detail. The specific manner in which each module and unit performs the operation of the device shown in FIG5 in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0172] FIG6 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure. As shown in FIG6 , the electronic device includes: a memory 11a and a processor 12a;

[0173] The memory 11a is used to store computer programs and can be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, contact data, phone book data, messages, pictures, videos, etc.

[0174] The memory 11a can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0175] The processor 12a is coupled to the memory 11a, and is configured to execute the computer program in the memory 11a, so as to perform the steps in the various method embodiments provided in the above embodiments of the present disclosure.

[0176] Furthermore, as shown in Figure 6, the electronic device also includes: a communication component 13a, a display 14a, a power supply component 15a, an audio component 16a and other components. Figure 6 only schematically shows some components, which does not mean that the electronic device only includes the components shown in Figure 6. In addition, the components in the dotted box in Figure 6 are optional components, not mandatory components, and the specific product form of the electronic device may depend on the product form. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone or an IOT (Internet of things) device, or it can be a server-side device such as a conventional server, a cloud server or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it may include the components in the dotted box in Figure 6; if the electronic device of this embodiment is implemented as a server-side device such as a conventional server, a cloud server or a server array, it may not include the components in the dotted box in Figure 6.

[0177] The detailed implementation process of the processor executing each action can be found in the relevant description in the aforementioned method embodiment or device embodiment, and will not be repeated here.

[0178] Accordingly, an embodiment of the present disclosure further provides a computer-readable storage medium storing a computer program, which can implement the steps in the above method embodiment when executed.

[0179] Accordingly, an embodiment of the present disclosure further provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the steps in the above method embodiment.

[0180] The above-mentioned memory can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0181] The above-mentioned communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0182] The above-mentioned display includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor can not only sense the boundary of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0183] The power supply assembly provides power to various components of the device in which the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power supply assembly is located.

[0184] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as call mode, recording mode, and voice recognition mode, the microphone is configured to receive external audio signals. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0185] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to magnetic disk storage, compact disc read-only memory (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0186] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0187] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0188] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0189] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interfaces, network interfaces, and memory.

[0190] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0191] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0192] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0193] The above are merely examples of the present disclosure and are not intended to limit the present disclosure. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure are intended to be included within the scope of the claims of the present disclosure.

Claims

1. A method for scheduling data center workloads in an electricity market environment, applied to a data center operating in a multi-level electricity market, where electricity transactions in the multi-level electricity market occur on multiple time scales, the method comprising: Acquire information of multiple delay-tolerant workloads that the data center needs to process, wherein the information of the multiple workloads includes type information of each workload and data dependencies between the multiple workloads; According to the type information of each workload and the data dependency relationship between the multiple workloads, a load scheduling model of the multiple workloads in a time dimension is constructed, wherein the load scheduling model reflects the time uncertainty of the workload; According to the power source information of the data center, constructing a power cost optimization model corresponding to multiple stages of the data center to optimize the power cost of the data center on the multiple time scales; According to the power cost optimization models of the multiple stages, the load scheduling model is optimized and solved in multiple stages to obtain scheduling schemes for the multiple workloads on multiple time scales; According to the scheduling schemes of the multiple workloads on multiple time scales, the multiple workloads are scheduled respectively on multiple time scales to reduce the electricity cost of the data center.

2. The method according to claim 1, wherein: According to the type information of each workload and the data dependency relationship between the multiple workloads, a load scheduling model of the multiple workloads in the time dimension is constructed, including: Determine a scheduling time window for each workload according to the type information of each workload and the data dependency relationship between the multiple workloads, wherein the scheduling time window includes a start time and a deadline of the workload; According to the scheduling time window of each workload, the scheduling state of each workload in the time slot is used as a variable to construct the objective function and constraint conditions corresponding to each workload, so as to obtain the load scheduling model of each workload in the time dimension; The time slot is a time gap at any time scale, and the objective function represents a possible scheduling method of the workload at each time slot included in its scheduling time window.

3. The method according to claim 2, wherein: According to the type information of each workload and the data dependency relationship between multiple workloads, the scheduling time window of each workload is determined, including: For periodic workloads, the start time of the periodic workload is predicted according to the probability distribution function that the start time of the periodic workload conforms to; according to the data dependency relationship between the periodic workloads, for each periodic workload, the earliest start time of other periodic workloads with which it has data dependency is used as its deadline; For non-periodic workloads, the start time of the non-periodic workload is predicted based on the probability distribution function that the start time of the non-periodic workload conforms to; the deadline of the non-periodic workload is generated based on the priority, execution duration and preset deadline factor of the non-periodic workload.

4. The method according to claim 2, wherein: According to the scheduling time window of each workload, the scheduling state of each workload in the time slot is used as a variable to construct the objective function and constraint conditions corresponding to each workload, so as to obtain the load scheduling model of each workload in the time dimension, including: For any workload, taking the scheduling state of the workload in the time slot as a variable, construct an equation representing that the workload needs to be completed within L time slots included in the scheduling time window of the workload as the objective function, where L is an unknown quantity; Taking the scheduling state of the workload in the time slot as a variable, construct a first constraint condition that represents that the scheduling time of the workload is no later than the time corresponding to the L time slots forward of the deadline of the workload and no earlier than the start time of the workload; A second constraint condition characterizing that the workload must be completed before the deadline of the workload is constructed by taking the scheduling state of the workload in the time slot as a variable.

5. The method according to any one of claims 2 to 4, wherein: Also includes: Predict the energy consumption of each workload based on the computing resource usage of each workload; According to the load scheduling model of each workload in the time dimension and the power consumption of each workload, a power consumption model of the workload in a time slot is constructed with the time slot as the granularity to obtain a global power consumption model.

6. The method according to claim 5, wherein: According to the power source information of the data center, a power cost optimization model for multiple stages corresponding to the data center is constructed, including: According to the power source information of the data center, construct a global power cost function corresponding to the data center, wherein the global power cost function includes the power procurement cost of the data center purchasing power from the multi-level power market, the penalty cost of abandoning renewable energy, and the expected operating cost of the data center supplying power by itself; Constructing constraint conditions at multiple stages according to constraint information of various power sources used by the data center and power consumption information obtained by mapping the global power consumption model to multiple time scales; The global electricity cost function and the constraints of multiple stages are used as a multi-stage electricity cost optimization model corresponding to the data center.

7. The method according to claim 6, wherein: According to the power source information of the data center, a global power cost function corresponding to the data center is constructed, including: According to the power source information of the data center, construct the power cost function of the data center under K scenarios; perform weighted summation of the power cost functions under the K scenarios according to the probability of occurrence of the K scenarios, as the global power cost function corresponding to the data center; the K scenarios are multiple scenarios determined according to the uncertainty of the power source in multiple time scales, and K is a positive integer; Among them, the electricity cost function under each scenario includes at least: the expected operating cost of the data center for self-power supply at the maximum time scale and the electricity procurement cost of purchasing electricity in the multi-level electricity market, the adjustment cost of the expected operating cost of self-power supply at non-minimum and non-maximum time scales and the electricity procurement cost of purchasing electricity in the multi-level electricity market, as well as the penalty cost of abandoning renewable energy at the minimum time scale and the penalty cost caused by the deviation between actual electricity consumption and contracted electricity consumption in electricity markets at all levels.

8. The method according to claim 7, wherein: The multi-level electricity market is a three-level electricity market, and the multiple time scales include, from large to small, a first time scale, a second time scale, and a third time scale; Then construct the electricity cost function of the data center under K scenarios, including: For each scenario, constructing an electricity cost function for any time slot in the first time scale according to the power generation cost of the self-powered generator in the data center, the power generation and startup cost in any time slot in the first time scale, and the electricity price and power purchase amount in any time slot in the first time scale; Constructing an electricity cost function for any time slot on the second time scale according to the power generation cost of the self-powered generator in the data center and the power generation adjustment amount in any time slot on the second time scale, as well as the electricity price and the power purchase amount in any time slot on the second time scale; According to the electricity price in any time slot in the third time scale, the purchase amount of electricity with too low a bid, the penalty coefficient of too high a bid and the purchase amount of electricity with too high a bid, as well as the amount of renewable energy abandoned in any time slot in the third time scale and the penalty coefficient of abandoned renewable energy, a power cost function in any time slot in the third time scale is constructed; According to the power cost function of any time slot in the first time scale, the power cost function of any time slot in the second time scale The power cost function of a time slot and the power cost function of any time slot in a third time scale are used to generate the power cost function in the scenario; The power purchase amount and power generation amount in any time interval under the first time scale, and the power generation adjustment amount and power purchase amount in any time interval under the second time scale are variables to be optimized.

9. The method according to claim 8, wherein: Also includes: Scenario generation is performed for the uncertain parameters in each of the three-level power market to obtain N ω scenes; among them, N ω =N ω1 ×N ω2 ×N ω3 , N ω1 is the number of scenarios corresponding to the first-level electricity market, N ω2 YesN ω1 The number of scenarios that each scenario in the second-level electricity market corresponds to, N ω3 YesN ω2 The number of scenarios corresponding to each scenario in the third-level electricity market; For the N ω The scenes are clustered to obtain K scenes, and the probability of each scene in the K scenes is calculated by assigning each random point to the nearest cluster centroid, N ω1 、N ω2 、N ω3 and N ω All are positive integers.

10. The method according to claim 6, wherein: According to the constraint information of various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model to multiple time scales, constraint conditions of multiple stages are constructed, including: For any time scale, based on the constraint information of various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model to the any time scale, construct power source constraint conditions and power balance constraint conditions for the corresponding stage of the any time scale; The power source constraint condition includes at least one of a generator operation constraint condition, a renewable energy constraint condition and a power market transmission constraint condition.

11. The method according to claim 10, wherein: The multi-level electricity market is a three-level electricity market, the multiple time scales include, from large to small, a first time scale, a second time scale, and a third time scale, and the multiple stages include a first stage, a second stage, and a third stage corresponding to the three time scales in sequence; For any time scale, according to the constraint information of various power sources used by the data center and the power consumption information obtained by mapping the global power consumption model to the any time scale, the power source constraint condition and the power balance constraint condition of the corresponding stage of the any time scale are constructed, including: For the first stage corresponding to the first time scale, the power consumption information obtained by mapping the global power consumption model to any time interval under the first time scale and the power consumption information of the data center under the first time scale The amount of electricity obtained in any time interval is used to construct an equation with an equivalent relationship as the power balance constraint condition in the first stage; According to the power generation, upper and lower limits of ramp rate and startup cost of the generator powered by the data center in any time interval at the first time scale, a power constraint formula, a start-stop constraint formula and a start-stop cost constraint formula of the generator are constructed as the generator operation constraint conditions of the first stage; According to the power generation and upper and lower limits of the power generation of the renewable energy in the data center in any time interval at the first time scale, the renewable energy constraint condition of the first stage is constructed; the power generation of the renewable energy conforms to the Gaussian distribution; According to the power purchase amount and the upper and lower limits of the transmission line capacity of the data center in any time interval at the first time scale, the power market transmission constraint conditions of the first stage are constructed.

12. The method according to claim 11, wherein: Also includes: For the second stage corresponding to the second time scale, according to the power consumption information obtained by mapping the global power consumption model to any time slot under the second time scale and the amount of power obtained by the data center in any time slot under the second time scale, an equation with an equivalent relationship is constructed as a power balance constraint condition for the second stage; According to the power generation adjustment amount and the upper and lower limits of the ramp rate of the generator powered by the data center in any time slot at the second time scale, and the power generation in the time slot corresponding to the second time scale at the first time scale, a power generation constraint formula and a power generation adjustment constraint formula of the generator are constructed as the generator operation constraint conditions of the second stage; According to the power generation and upper and lower limits of the power generation of the renewable energy in the data center in any time interval at the second time scale, construct the renewable energy constraint conditions of the second stage; According to the power purchase amount and the upper and lower limits of the transmission line capacity of the data center in any time interval at the second time scale, the power market transmission constraint conditions of the second stage are constructed.

13. The method according to claim 12, wherein: Also includes: For the third stage corresponding to the third time scale, according to the power consumption information obtained by mapping the global power consumption model to any time interval under the third time scale and the amount of electricity obtained by the data center in any time interval under the third time scale, an equation with an equivalent relationship is constructed as a power balance constraint condition for the third stage; According to the power generation of the renewable energy in the data center in any time interval under the third time scale and upper and lower limits of power generation to construct the renewable energy constraints of the third stage; According to the overly high purchase amount of the data center in any time slot under the third time scale, the underlying purchase amount, the purchase amount in the time slot corresponding to the third time scale under the second time scale, and the upper and lower limits of the transmission line capacity, the third stage power market transmission constraint conditions are constructed; According to the power generation and the upper limit of the power generation of the renewable energy in the data center in any time slot in the third time scale, the renewable energy abandoned by the data center in any time slot in the third time scale is obtained.

14. The method according to any one of claims 6 to 13, wherein: According to the power cost optimization models of the multiple stages, the load scheduling model is optimized and solved in multiple stages to obtain scheduling schemes for the multiple workloads on multiple time scales, including: Mapping the time slots in the objective function and constraint conditions in the load scheduling model of each workload in the time dimension to each time slot on the multiple time scales to obtain additional constraint conditions on the multiple stages; According to the constraints and additional constraints of the multiple stages, the global power cost function is optimized and solved in multiple stages with the goal of minimizing the global power cost, so as to obtain scheduling schemes for the multiple workloads on multiple time scales.

15. An electronic device, comprising: Memory and processor; The memory is used to store computer programs; The processor is coupled to the memory and is configured to execute a computer program stored in the memory to implement the steps in the method according to any one of claims 1 to 14.

16. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 14.

Images