A Multi-Data Center Computing Power-Power Coordinated Scheduling and Demand Response Capacity Optimization Method

Abstract

This invention provides a method for optimizing the capacity of demand response applications by coordinating computing power and power dispatch across multiple data centers. First, multi-dimensional historical time-series data of the data centers are collected. Then, machine learning algorithms are used to predict the workload of each data center at each time period of the future day within the future dispatch cycle, as well as the electricity market price of each data center's location. Next, a multi-data center computing power-power collaborative model considering uncertainty and spatiotemporal coupling is constructed. This collaborative model includes a joint optimization framework for computing power dispatch and power dispatch, with the objective of maximizing actual profit. The established collaborative model is then transformed into a mixed-integer linear programming problem and solved to obtain the optimal application capacity for each data center participating in demand response, as well as the workload of spatial and temporal migration, the charging and discharging power of the energy storage system, and the renewable energy generation. Each data center is then dispatched accordingly, achieving comprehensive optimization of the demand response application capacity across multiple data centers.

Description

Technical Field

[0001] This invention belongs to the field of data center energy management technology, and specifically relates to a method for multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization. Background Technology

[0002] As the core of computing infrastructure, data centers are increasingly facing challenges related to high energy consumption, high carbon emissions, and insufficient coordination between computing power and power resources. Traditional data centers typically operate with independent power scheduling and computing power allocation: power scheduling primarily adjusts load based on peak-valley electricity prices and grid supply and demand, while computing power scheduling focuses on local optimization of task latency and resource utilization. Existing research largely fails to adequately consider the differences in their operational mechanisms, making it difficult to achieve globally optimal allocation of computing power and power resources at the system level.

[0003] With the widespread adoption of demand response mechanisms across various industries, data centers, with their adjustable load characteristics, have demonstrated strong potential for participation. However, currently, most data centers participate passively on a per-data-center basis in demand response, lacking coordinated planning for cross-data center capacity requests. This often results in mismatches, with some data centers submitting redundant requests while others have insufficient capacity. Existing research largely ignores core requirements such as dynamic cross-center capacity adjustment and cross-provincial computing power transfer, and a collaborative optimization model covering the entire chain has not yet been established. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method. By using a spatiotemporally coupled multi-data center computing power-power collaborative model oriented towards uncertainty, the method achieves overall optimization of the demand response reporting capacity of multiple data centers, avoids capacity redundancy or insufficiency, significantly reduces operating costs, and alleviates the pressure on the power grid during peak periods.

[0005] To achieve the above objectives, the solution of this invention is: to provide a multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method, comprising the following steps:

[0006] S1. Data Acquisition and Prediction: Collect multi-dimensional historical time-series data from data centers, and then use machine learning algorithms to predict the workload of each data center at each time period of the day within the future scheduling cycle. and the electricity market prices in the locations of each data center ;

[0007] S2. Constructing a Collaborative Optimization Model: Based on the predicted data from step S1, and considering the temporal dynamics and spatial heterogeneity of data center computing workloads and geographical distribution, combined with the spatiotemporal characteristics of regional power grid supply, a multi-data center computing power-power collaborative model considering uncertainties and oriented towards spatiotemporal coupling is constructed. This collaborative model includes a joint optimization framework for the computing power scheduling model and the power scheduling model, with the objective of maximizing actual profit. Actual profit consists of demand response revenue, electricity purchase cost, and penalty cost, calculated using the following formula:

[0008]

[0009] In the formula, C represents the actual profit of the data center; This indicates the benefits that data centers gain from participating in demand response; This represents the unit price of electricity purchased by data center i at time t; This represents the power consumption of data center i at time t; , , This indicates the time-shifting, space-shifting, and penalty cost for unprocessed workloads with priority p. This indicates that data center i is transferred from time t. The number of tasks with priority p processed at any given time; This represents the amount of tasks with priority p that are transferred from data center i to data center j at time t; This represents the amount of tasks with priority p that data center i has not processed at time t;

[0010] S2.1 Constructing a computing power scheduling model: The computing power scheduling model is used to manage workloads of different priorities, establish scheduling constraints for their time transition and spatial migration, and output the task volume. , , ;

[0011] S2.2 Constructing a power dispatch model: The power dispatch model is used to manage grid power purchases, energy storage systems, local renewable energy sources, and data center equipment energy consumption, and establishes power supply and demand balance constraints for data centers, outputting grid power purchases. and the actual response load of data centers participating in demand response. ;

[0012] S2.3, Constructing a calculation model for the actual revenue R of demand response: The formula for calculating the actual revenue R obtained by the data center from participating in demand response is as follows:

[0013]

[0014] In the formula, This indicates the actual response load of the data center participating in demand response. Indicates response time; This represents the price subsidy coefficient; Indicates the response speed coefficient; This represents the subsidized unit price for data center i;

[0015] Considering the uncertainty of model predictions, demand response reporting capacity To retain a certain margin, the calculation formula is as follows:

[0016]

[0017] In the formula, Indicates the adjustment factor for the declared capacity;

[0018] S3. Model Solving and Strategy Execution: The collaborative model established in step S2 is transformed into a mixed-integer linear programming problem and solved to obtain the optimal requested capacity for each data center participating in demand response. The data center generates detailed operational instructions, including computing power scheduling strategies, energy storage system charging and discharging strategies, and renewable energy allocation strategies, based on the workload of spatial and temporal migration, the charging and discharging power of the energy storage system, and the renewable energy generation. Each data center then schedules its operations according to these instructions to participate in demand response.

[0019] Furthermore, in step S2.1, constructing the computing power scheduling model includes the following steps:

[0020] S2.1.1 Divide the workload into latency-sensitive tasks and latency-tolerant tasks according to business type, and assign different priorities.

[0021] The task priority set for data center workloads is priority={0, 1, 2}. } where 0 represents the highest priority, corresponding to latency-sensitive tasks that require real-time processing and only undergo space scheduling. Lower priority tasks correspond to latency-tolerant tasks, allowing time scheduling within a specified time window, while also supporting space scheduling. To avoid overloading computing resources, the processing capacity of each priority task is constrained. Considering the flexibility of low-priority tasks, a slight exceedance of the limit is allowed within the specified timeframe, expressed by the formula:

[0022]

[0023] In the formula, Indicates that data center i is in The number of tasks with a priority of p at any given time; This indicates the maximum utilization rate of the data center servers; This indicates the processing speed of a single server in the data center. Indicates that data center i is in The number of servers that are constantly running; Indicates the number of data centers;

[0024] S2.1.2 Establishing a time scheduling model

[0025] A maximum allowable transfer time Tmax is set, allowing latency-tolerant tasks to be transferred back and forth within the set maximum latency time Tmax, while satisfying the constraint that the total number of tasks that can be delayed in the data center cannot exceed the sum of the currently received tasks and the transferred tasks. This can be expressed by the formula:

[0026]

[0027] In the formula, This indicates that data center i is transferred from time t. The amount of tasks with a priority of 0 that are processed at any given time; This indicates that data center i is transferred from time t. The number of tasks with a priority of p at any given time; Let p represent the number of tasks with priority p received by data center i from time t, where p ≠ 0; Let p represent the amount of tasks with priority p that are transferred from data center j to data center i at time t, where p ≠ 0;

[0028] S2.1.3 Establishing a spatial scheduling model

[0029] Configure a multi-datacenter set as ={1, 2, Let N be the number of data centers, allowing tasks of all priorities to be located in the same set of data centers. The task transfer allocation is performed across geographical locations within the system, while satisfying the task transfer-out conservation constraint that the amount of tasks that can be transferred out cannot exceed the sum of the currently received tasks and the amount of tasks transferred in. This can be expressed by the following formula:

[0030]

[0031] In the formula, express The amount of tasks with priority p that are transferred from data center i to data center j at any given time. express The amount of tasks with priority p that are transferred from data center j to data center i at any given time;

[0032] S2.1.4, Satisfying the constraint of total task conservation before and after data center scheduling, expressed by the formula:

[0033]

[0034] In the formula, , These represent data center i in The amount of tasks with priority p that are transferred via time and space scheduling at any given moment; Indicates that data center i is in The number of tasks with a priority of p at any given time; This indicates that data center i is transferred from time t. The number of tasks with a priority of p at any given time; Indicates that data center i is in The number of tasks with priority p that are not processed at any given time; Indicates that data center i is in The amount of tasks received at any given time; Indicates data center i from Time shift The workload at any given moment.

[0035] Furthermore, in step S2.2, constructing the power dispatch model includes the following steps:

[0036] S2.2.1 Establishing an energy storage system model

[0037] The capacity of an energy storage system is determined by its capacity and charging / discharging power at the previous moment. The energy storage system model is as follows:

[0038]

[0039] In the formula, This represents the capacity of the energy storage system of data center i at time t; , These represent the charging and discharging efficiencies of the energy storage system, respectively. , These represent the charging and discharging power of the energy storage system, respectively. , These represent the minimum and maximum values ​​of the state of charge of the energy storage system, respectively. Rated capacity of the energy storage system for data center i; , These represent the maximum charging and discharging power of the energy storage system, respectively.

[0040] S2.2.2 Establishing a wind power generation system model

[0041] The power output of wind power generation is determined by the rated capacity of the wind power generation system and the wind power coefficient. The wind power coefficient is related to wind speed, rotor diameter, and wind energy utilization efficiency. The wind power generation system model is as follows:

[0042]

[0043] In the formula, This represents the wind power generation capacity of data center i at time t; This represents the rated capacity of the wind power generation system of data center i, which is a known quantity. This represents the wind power coefficient of data center i at time t;

[0044] S2.2.3 Establishing a photovoltaic power generation system model

[0045] Photovoltaic power generation is related to nominal power, total solar irradiance, and overall utilization efficiency. The photovoltaic power generation system model is as follows:

[0046]

[0047] In the formula, This represents the photovoltaic power generation of data center i at time t; This represents the overall utilization efficiency of the solar energy system; this is a known quantity. This represents the nominal power of the photovoltaic power generation system of data center i; This represents the total solar radiation actually received by data center i at time t; This represents the irradiance under standard test conditions, expressed as 1 kW / m².

[0048] S2.2.4. Power usage efficiency is used as the energy efficiency evaluation index.

[0049] To measure the energy efficiency of data centers, Power Usage Effectiveness (PUE) is used as the energy efficiency evaluation index, defined as the ratio of the total energy consumption of the data center to the energy consumption of server equipment, expressed by the formula:

[0050]

[0051] In the formula, Indicates power efficiency; This represents the total energy consumption of data center i at time t; This represents the energy consumption of the server equipment in data center i at time t;

[0052] S2.2.5 Establish an equipment energy consumption model

[0053] Server equipment energy consumption is related to its utilization rate and power consumption. The energy consumption of data center server equipment is as follows:

[0054]

[0055] In the formula, , These represent the power consumption of a single server device in data center i under full load and standby conditions, respectively. This represents the number of fully loaded servers in data center i at time t; This indicates the total number of servers configured in data center i;

[0056] S2.2.6 Establish a total energy consumption and power balance model

[0057] The power supply and demand of data centers need to be balanced in real time, which can be expressed by the formula:

[0058]

[0059] In the formula, This represents the power consumption of data center i at time t; This represents the photovoltaic power generation of data center i at time t; This represents the total energy consumption of data center i at time t; This represents the wind power generation capacity of data center i at time t; , These represent the charging and discharging power of the energy storage system, respectively.

[0060] S2.2.7 Establish a demand response load calculation model

[0061] When a data center participates in demand response, the actual response load Δ The calculation formula is:

[0062]

[0063] In the formula, This represents the actual response load of data center i at time t; This represents the baseline electrical load of data center i at time t.

[0064] Furthermore, in step S2, to improve the robustness of the collaborative model, the following steps are also included:

[0065] S2.4 Introducing Conditional Value at Risk (CVaR) to generate risk-resistant scheduling strategies

[0066] Considering the predicted workload and electricity market prices To account for uncertainty, a prediction error is set, and a scene set S is generated using the Monte Carlo simulation method. The number of scenes s in the set is greater than 500. The formula for calculating the actual value under scene s is expressed as:

[0067]

[0068] In the formula, This represents the unit price of electricity purchased by data center i at time t under scenario s; This represents a random error sample of the electricity purchase price per unit at time t for data center i in scenario s. This represents the amount of tasks with priority p received by data center i from time t in scenario s. This represents the random error sample of the amount of tasks with priority p received by data center i from time t in scenario s.

[0069] Assign probabilities to the scenes in the scene set S. This can be expressed as a formula:

[0070]

[0071] In the formula, This represents the probability under electricity price scenario s; This represents the probability under workload scenario s;

[0072] The constraints of CVaR are expressed by the following formula:

[0073]

[0074] In the formula, The loss function is defined as follows: , Indicates excess loss. Indicates the maximum acceptable loss;

[0075] After introducing CVaR, a balance needs to be struck between profit and high loss risk. The objective function of the coordination model is then modified as follows:

[0076]

[0077] In the formula, Represents the probability of scenario s; This represents the actual profit in scenario s; Indicates the risk aversion coefficient; Indicates the maximum acceptable loss; Indicates the confidence level; This indicates excess loss.

[0078] Furthermore, in step S2.3, the ratio S of the actual response load to the invited response quantity is... DR The calculation formula is:

[0079]

[0080] Based on the electricity demand response implementation plan for a region, determine the parameters ξ and S. DR , , The specific values, where the price subsidy coefficient ξ and S DR The relevant information is expressed by the formula:

[0081]

[0082] The response rate coefficient ν is related to ΔT, and can be expressed by the formula:

[0083]

[0084] When local power grids officially issue demand response invitations, the model parameters are further revised based on the published demand response methods, scale, time period, regional scope, and incentive price information. Then, in the subsequent step S3, a more accurate output value is obtained through model solving, and the data center fills in the response information accordingly.

[0085] Furthermore, in step S1, the multi-dimensional historical time-series data of the data center includes server operating status, workload data, and electricity market price data. The server operating status includes power, CPU utilization, and server processing speed in standby and full-load states.

[0086] Furthermore, in step S2.3, the reporting capacity correction factor k is set to 0.85-0.95.

[0087] Furthermore, in step S3, after importing historical data and assigning values ​​to each parameter, the coordination model is transformed into a mixed-integer linear programming problem. The Gurobi solver is used to solve the model, obtaining the optimal capacity declaration value for the data center to participate in demand response. The workload of spatial and temporal migration for each data center and The charging and discharging power of the energy storage system and Wind power generation and photovoltaic power generation .

[0088] Furthermore, in step S1, LSTM is used for model training. During model prediction, historical data from the past 30-60 days is used as input, and workload and electricity price data are used as prediction targets. Considering that the data center's scheduling period is a single day, each day is divided into 96 time slots, and the duration of each time slot is Δt = 15 minutes, i.e., T = {1, 2, ...} ,96}.

[0089] Furthermore, in step S1, after collecting multi-dimensional historical time-series data from the data center, the original historical data is preprocessed by filling in missing values, removing outliers, and standardizing the data before being used to make predictions using a machine learning model.

[0090] After adopting the above solution, the beneficial effects of the present invention are as follows:

[0091] 1. This application uses a prediction and optimization coordination model to dynamically determine the optimal demand response declaration capacity for each data center and generate specific operating strategies accordingly. This effectively avoids the blindness of declaration capacity and also avoids capacity redundancy or insufficiency, significantly improving demand response benefits and grid regulation effects.

[0092] 2. This application establishes a joint optimization framework for computing power scheduling and power scheduling, which closely integrates the temporal and spatial scheduling of computing tasks with the temporal and spatial allocation of power resources, overcoming the limitations of the traditional independent scheduling mode and achieving the optimal overall energy efficiency and economy of the system.

[0093] 3. This application fully explores the adjustment potential of latency-tolerant tasks by establishing a computing power scheduling model that includes time transfer and spatial migration, realizes the complementarity of computing power resources and load shaping across data centers and time periods, and smooths out local power peak pressure.

[0094] 4. This application incorporates wind power generation, photovoltaic power generation and energy storage system models into the optimization framework, guiding data centers to prioritize the use of local renewable energy during operation, thereby reducing dependence on external power grids and carbon emissions.

[0095] 5. This application quantifies and controls the financial risks caused by the uncertainty of electricity prices and workload forecasts by introducing the Conditional Value at Risk (CVaR) theory and Monte Carlo simulation, generating a more robust and risk-resistant scheduling strategy and ensuring the stable operation of the data center in a complex market environment. Attached Figure Description

[0096] Figure 1 This is a flowchart of the method of the present invention;

[0097] Figure 2 This refers to the electricity market prices for each data center location predicted by this embodiment of the invention for each time period of the day in the future.

[0098] Figure 3 This refers to the predicted workload of each data center for each time period of a future day, as per an embodiment of the present invention.

[0099] Figure 4 The optimal capacity declaration value and actual response load for each data center participating in demand response are obtained from the solutions of this embodiment of the invention.

[0100] Figures 5-7 This refers to the spatial migration task volume of each data center obtained from the solutions of this embodiment of the invention;

[0101] Figures 8-10 The time delay task amount for each data center obtained from the embodiments of the present invention;

[0102] Figure 11The final processing workload of each data center obtained from the solutions of this embodiment of the invention;

[0103] Figure 12 The charging and discharging power, wind power generation power, and photovoltaic power generation power of the energy storage system of each data center are obtained by solving the solution in the embodiments of the present invention. Detailed Implementation

[0104] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. The scope value described in this invention includes two endpoint values.

[0105] like Figure 1 As shown, this application provides a method for multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization, including the following steps:

[0106] S1. Data Acquisition and Prediction: Collect multi-dimensional historical time-series data from the data center, including server operating status, workload data, and electricity market prices. Server operating status includes power consumption, CPU utilization, and server processing speed under standby and full load conditions. The collected raw historical data is then preprocessed. Preprocessing steps include imputing missing values, removing outliers, and standardization. Specifically, linear interpolation can be used to imput missing values, the 3σ (3 times standard deviation) criterion can be used to remove outliers, and Min-Max normalization can be performed for standardization. Of course, the preprocessing methods used are not limited to these and other preprocessing methods can also be used.

[0107] After preprocessing, machine learning algorithms are used to predict the workload of each data center at different times of the day during the future scheduling cycle. and the electricity market prices in the locations of each data center For the data, LSTM (Lateral Recurrent Mechanism) can be used for model training, or other prediction models can be used. This embodiment uses the LSTM model as an example. When making predictions, historical data from the past 30-60 days is used as input, and workload and electricity price data are used as prediction targets. Considering that the data center's scheduling period is a single day, each day is divided into 96 time periods, and the duration of each time period is Δt = 15 minutes, i.e., T = {1, 2, ..., ...} Setting a time interval of 15 minutes ensures sufficient data volume without causing excessive computational load.

[0108] S2. Constructing a Collaborative Optimization Model: Based on the predicted data from step S1, and considering the temporal dynamics and spatial heterogeneity of data center computing workloads and geographical distribution, combined with the spatiotemporal characteristics of regional power grid supply, a multi-data center computing power-power collaborative model considering uncertainties and oriented towards spatiotemporal coupling is constructed. This collaborative model includes a joint optimization framework for the computing power scheduling model and the power scheduling model, with the objective of maximizing actual profit. Actual profit consists of demand response revenue, electricity purchase cost, and penalty cost, calculated using the following formula:

[0109]

[0110] In the formula, C represents the actual profit of the data center; This represents the revenue that the data center gains from participating in demand response, and its calculation formula is described in step S2.3 below; The unit price of electricity purchased by data center i at time t is obtained by prediction in step S1; This represents the power consumption of data center i at time t, output by the power dispatch model; , , This indicates the time-shifting, space-shifting, and penalty cost for unprocessed workloads with priority p, which can be set based on historical data. This indicates that data center i is transferred from time t. The number of tasks with priority p processed at any given time; This represents the amount of tasks with priority p that are transferred from data center i to data center j at time t; This represents the number of tasks with priority p that have not been processed in data center i at time t. , , All outputs are generated by the computing power scheduling module.

[0111] In the collaborative model, the specific steps for constructing the computing power scheduling model, the power scheduling model, and calculating the actual demand response benefit R are as follows:

[0112] S2.1 Constructing a computing power scheduling model: The computing power scheduling model is used to manage workloads of different priorities, establish scheduling constraints for their time transition and spatial migration, and output the task volume. , , Specifically, it includes the following steps:

[0113] S2.1.1 Divide the workload into latency-sensitive tasks and latency-tolerant tasks according to business type, and assign different priorities.

[0114] The task priority set for data center workloads is priority={0, 1, 2}.} where 0 represents the highest priority, corresponding to latency-sensitive tasks that require real-time processing and only undergo space scheduling. Lower priority tasks correspond to latency-tolerant tasks, allowing time scheduling within a specified time window, while also supporting space scheduling. To avoid overloading computing resources, the processing capacity of each priority task is constrained. Considering the flexibility of low-priority tasks, a slight exceedance of the limit is allowed within the specified timeframe, expressed by the formula:

[0115]

[0116] In the formula, Indicates that data center i is in The number of tasks with a priority of p at any given time; This indicates the maximum utilization rate of the data center servers; This indicates the processing speed of a single server in the data center. Indicates that data center i is in The number of servers that are constantly running; This indicates the number of data centers.

[0117] S2.1.2 Establishing a time scheduling model

[0118] A maximum allowable transfer time Tmax is set, allowing latency-tolerant tasks to be transferred back and forth within the set maximum latency time Tmax, while satisfying the constraint that the total number of tasks that can be delayed in the data center cannot exceed the sum of the currently received tasks and the transferred tasks. This can be expressed by the formula:

[0119]

[0120] In the formula, This indicates that data center i is transferred from time t. The amount of tasks with a priority of 0 that are processed at any given time; This indicates that data center i is transferred from time t. The number of tasks with a priority of p at any given time; This represents the amount of tasks with priority p (p≠0) received by data center i from time t, which is predicted in step S1; Let t represent the amount of tasks with priority p (p≠0) that are transferred from data center j to data center i at time t.

[0121] S2.1.3 Establishing a spatial scheduling model

[0122] Workloads can be distributed and transferred among data centers in different geographical locations to achieve resource complementarity. This paper defines a multi-data center set as... ={1, 2, Let N be the number of data centers, allowing tasks of all priorities to be located in the same set of data centers. The task transfer allocation is performed across geographical locations within the system, while satisfying the task transfer-out conservation constraint that the amount of tasks that can be transferred out cannot exceed the sum of the currently received tasks and the amount of tasks transferred in. This can be expressed by the following formula:

[0123]

[0124] In the formula, express The amount of tasks with priority p that are transferred from data center i to data center j at any given time. express The amount of tasks with priority p that are transferred from data center j to data center i at any given time;

[0125] S2.1.4, Satisfying the constraint of total task conservation before and after data center scheduling, expressed by the formula:

[0126]

[0127] In the formula, , These represent data center i in The amount of tasks with priority p that are transferred via time and space scheduling at any given moment; Indicates that data center i is in The number of tasks with a priority of p at any given time; This indicates that data center i is transferred from time t. The number of tasks with a priority of p at any given time; Indicates that data center i is in The number of tasks with priority p that are not processed at any given time; Indicates that data center i is in The amount of tasks received at any given time; Indicates data center i from Time shift The workload at any given moment.

[0128] S2.2 Constructing a power dispatch model: The power dispatch model is used to manage grid power purchases, energy storage systems, local renewable energy sources, and data center equipment energy consumption. The renewable energy sources include wind power and photovoltaic power generation. It also establishes power supply and demand balance constraints for the data center and outputs the grid power purchase capacity. The actual response load of data centers participating in demand response In addition to the charging and discharging power of the energy storage system, the local renewable energy power generation, and the energy consumption of data center equipment, the specific steps include:

[0129] S2.2.1 Establishing an energy storage system model

[0130] The capacity of an energy storage system is determined by factors such as the capacity at the previous moment and the charging and discharging power. The energy storage system model is as follows:

[0131]

[0132] In the formula, This represents the capacity of the energy storage system of data center i at time t; , These represent the charging and discharging efficiencies of the energy storage system, respectively. , These represent the charging and discharging power of the energy storage system, respectively. , These represent the minimum and maximum values ​​of the state of charge of the energy storage system, respectively. Rated capacity of the energy storage system for data center i; , These represent the maximum charging and discharging power of the energy storage system, respectively.

[0133] S2.2.2 Establishing a wind power generation system model

[0134] The power output of wind power generation is determined by the rated capacity of the wind power generation system and the wind power coefficient. The wind power coefficient is related to parameters such as wind speed, rotor diameter, and wind energy utilization efficiency. The wind power generation system model is as follows:

[0135]

[0136] In the formula, This represents the wind power generation capacity of data center i at time t; This represents the rated capacity of the wind power generation system of data center i; This represents the wind power coefficient of data center i at time t.

[0137] S2.2.3 Establishing a photovoltaic power generation system model

[0138] Photovoltaic power generation is related to parameters such as nominal power, total solar irradiance, and comprehensive utilization efficiency. The photovoltaic power generation system model is as follows:

[0139]

[0140] In the formula, This represents the photovoltaic power generation of data center i at time t; Indicates the overall utilization efficiency of the solar energy system; This represents the nominal power of the photovoltaic power generation system of data center i; This represents the total solar radiation actually received by data center i at time t; This represents the irradiance under standard test conditions, expressed as 1 kW / m².

[0141] S2.2.4. Power usage efficiency is used as the energy efficiency evaluation index.

[0142] To measure the energy efficiency of data centers, Power Usage Effectiveness (PUE) is used as the energy efficiency evaluation indicator. It is defined as the ratio of the total energy consumption of the data center to the energy consumption of server equipment, expressed by the formula:

[0143]

[0144] In the formula, Indicates power efficiency; This represents the total energy consumption of data center i at time t; This represents the energy consumption of the server equipment in data center i at time t.

[0145] S2.2.5 Establish an equipment energy consumption model

[0146] Server equipment energy consumption is related to its utilization rate and power consumption. The energy consumption of data center server equipment is as follows:

[0147]

[0148] In the formula, , These represent the power consumption of a single server device in data center i under full load and standby conditions, respectively, obtained from historical data collection; This represents the number of fully loaded servers in data center i at time t; This indicates the total number of servers configured in data center i.

[0149] S2.2.6 Establish a total energy consumption and power balance model

[0150] The power supply and demand of data centers need to be balanced in real time, which can be expressed by the formula:

[0151]

[0152] In the formula, This represents the power consumption of data center i at time t; This represents the photovoltaic power generation of data center i at time t; This represents the total energy consumption of data center i at time t; This represents the wind power generation capacity of data center i at time t; , These represent the charging and discharging power of the energy storage system, respectively.

[0153] S2.2.7 Establish a demand response load calculation model

[0154] When a data center participates in demand response, the actual response load Δ The calculation formula is:

[0155]

[0156] In the formula, This represents the actual response load of data center i at time t; This represents the baseline electrical load of data center i at time t, obtained based on historical data.

[0157] S2.3, Constructing a calculation model for the actual revenue R of demand response: The formula for calculating the actual revenue R obtained by the data center from participating in demand response is as follows:

[0158]

[0159] In the formula, This indicates the actual response load of the data center participating in demand response. Indicates response time; This represents the price subsidy coefficient; Indicates the response speed coefficient; This represents the subsidized unit price for data center i;

[0160] The ratio of actual response load to invited response volume (S) DR The calculation formula is:

[0161]

[0162] Parameters ξ, S DR , , The specific values ​​are obtained with reference to the electricity demand response implementation plans of each region. This embodiment uses the demand response implementation plan of Fujian Province as an example for modeling. Among them, the price subsidy coefficient ξ and S DR The relevant information is expressed by the formula:

[0163]

[0164] The response rate coefficient ν is related to ΔT, and can be expressed by the formula:

[0165]

[0166] It should be noted that the collaborative model in this application is modeled based on the demand response implementation plan in Fujian Province. To calculate the demand response declaration capacity value and scheduling strategy for other regions, the model parameters can be further modified when the power grid in each region officially issues a demand response invitation, taking into account the published demand response method, scale, time period, regional scope, incentive price, and other information. Then, in the subsequent step S3, a more accurate output value can be obtained by solving the model, and the data center can fill in the response information accordingly.

[0167] To ensure that data centers achieve high returns in demand response, the demand response requested capacity... The actual response load Δ can be calculated based on the model. To formulate, taking into account the uncertainty of model predictions, the demand response declaration capacity. To retain a certain margin, the calculation formula is as follows:

[0168]

[0169] In the formula, This represents the reporting capacity correction factor, ranging from 0.85 to 0.95.

[0170] To improve the robustness of the collaborative model and enable it to adapt to more scenarios, step S2 also includes the following steps:

[0171] S2.4 Introducing Conditional Value at Risk (CVaR) to generate risk-resistant scheduling strategies

[0172] Considering the predicted workload and electricity market prices To account for uncertainty, a prediction error is set, and a scene set S is generated using the Monte Carlo simulation method. The number of scenes s in the set is greater than 500. The formula for calculating the actual value under scene s is expressed as:

[0173]

[0174] In the formula, This represents the unit price of electricity purchased by data center i at time t under scenario s; This represents a random error sample of the electricity purchase price per unit at time t for data center i in scenario s. This represents the amount of tasks with priority p received by data center i from time t in scenario s. This represents the random error sample of the amount of tasks with priority p received by data center i from time t in scenario s.

[0175] Assign probabilities to the scenes in the scene set S. This can be expressed as a formula:

[0176]

[0177] In the formula, This represents the probability under electricity price scenario s; This represents the probability under workload scenario s.

[0178] The constraints of CVaR are expressed by the following formula:

[0179]

[0180] In the formula, The loss function is defined as follows: , Indicates excess loss. Indicates the maximum acceptable loss;

[0181] After introducing CVaR, a balance needs to be struck between profit and high loss risk. The objective function of the coordination model is then modified as follows:

[0182]

[0183] In the formula, Represents the probability of scenario s; This represents the actual profit in scenario s; Indicates the risk aversion coefficient; Indicates the maximum acceptable loss; Indicates the confidence level; This indicates excess loss.

[0184] S3. Model Solving and Strategy Execution: After importing historical data and assigning parameter values, the collaborative model established in step S2 is transformed into a mixed-integer linear programming problem and solved. Specifically, the Gurobi solver can be used to solve the problem, obtaining the optimal requested capacity for each data center participating in demand response. And the workload of spatial and temporal migration. and The charging and discharging power of the energy storage system and Wind power generation and photovoltaic power generation This generates detailed operational instructions, including computing power scheduling strategies, energy storage system charging and discharging strategies, and renewable energy allocation strategies. Each data center then schedules its operations according to these instructions to participate in demand response, ultimately maximizing profits under controllable risks.

[0185] The following is a further illustration through a specific embodiment:

[0186] This embodiment takes a data center equipped with wind power generation, photovoltaic power generation and energy storage systems and two data centers equipped with only energy storage systems as the research objects. The configuration parameters of the three data centers are shown in Table 1.

[0187] Table 1 - Data Center Configuration Parameters

[0188]

[0189] Based on historical data, the electricity market price for each time period of the day in the locations of each data center can be predicted. like Figure 2 As shown, the workload for each time period of the future day. like Figure 3 As shown.

[0190] By constructing a collaborative optimization model and solving it using the Gurobi solver, the optimal capacity declaration values ​​for each data center participating in demand response were obtained. and actual response load Δ like Figure 4 As shown, the results of computing power scheduling and power scheduling are as follows:

[0191] Computing power scheduling results: Spatial migration task volume of each data center like Figures 5-7 As shown, the workload of the delayed task like Figure 8-10 As shown, the final data center processing workload like Figure 11 As shown.

[0192] The power dispatch results are: the charging and discharging power of the energy storage system. , Wind power generation capacity and photovoltaic power generation like Figure 12 As shown.

[0193] In this embodiment, after optimization using the collaborative model, the total actual profit of the three data centers is 35,810.67 yuan. However, in the traditional calculation without collaborative model optimization, because it does not participate in demand response or workload transfer, it can only calculate the total electricity purchase cost of each data center. The actual profit before optimization was -31,172.26 yuan, which is 31,172.26 yuan. Therefore, after the optimization of the collaborative model in this application, the actual profit increased by 214.88%, which significantly improved the revenue of each data center and ensured the stable operation of the data center in a complex market environment.

[0194] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0195] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization, characterized in that, Includes the following steps: S1. Data Acquisition and Prediction: Collect multi-dimensional historical time-series data from data centers, and then use machine learning algorithms to predict the workload of each data center at each time period of the day within the future scheduling cycle. and the electricity market prices in the locations of each data center ; S2. Constructing a Collaborative Optimization Model: Based on the predicted data from step S1, and considering the temporal dynamics and spatial heterogeneity of data center computing workloads and geographical distribution, combined with the spatiotemporal characteristics of regional power grid supply, a multi-data center computing power-power collaborative model considering uncertainties and oriented towards spatiotemporal coupling is constructed. This collaborative model includes a joint optimization framework for the computing power scheduling model and the power scheduling model, with the objective of maximizing actual profit. Actual profit consists of demand response revenue, electricity purchase cost, and penalty cost, calculated using the following formula: In the formula, C represents the actual profit of the data center; This indicates the benefits that data centers gain from participating in demand response; This represents the unit price of electricity purchased by data center i at time t; This represents the power consumption of data center i at time t; , , This indicates the time-shifting, space-shifting, and penalty cost for unprocessed workloads with priority p. This indicates that data center i is transferred from time t. The number of tasks with priority p processed at any given time; This represents the amount of tasks with priority p that are transferred from data center i to data center j at time t; This represents the amount of tasks with priority p that data center i has not processed at time t; S2.1 Constructing a computing power scheduling model: The computing power scheduling model is used to manage workloads of different priorities, establish scheduling constraints for their time transition and spatial migration, and output the task volume. , , ; S2.2 Constructing a power dispatch model: The power dispatch model is used to manage grid power purchases, energy storage systems, local renewable energy sources, and data center equipment energy consumption, and establishes power supply and demand balance constraints for data centers, outputting grid power purchases. and the actual response load of data centers participating in demand response. ; S2.3, Constructing a calculation model for the actual revenue R of demand response: The formula for calculating the actual revenue R obtained by the data center from participating in demand response is as follows: In the formula, This indicates the actual response load of the data center participating in demand response. Indicates response time; This represents the price subsidy coefficient; Indicates the response speed coefficient; This represents the subsidized unit price for data center i; Considering the uncertainty of model predictions, demand response reporting capacity To retain a certain margin, the calculation formula is as follows: In the formula, Indicates the adjustment factor for the declared capacity; S3. Model Solving and Strategy Execution: The collaborative model established in step S2 is transformed into a mixed-integer linear programming problem and solved to obtain the optimal requested capacity for each data center participating in demand response. The data center generates detailed operational instructions, including computing power scheduling strategies, energy storage system charging and discharging strategies, and renewable energy allocation strategies, based on the workload of spatial and temporal migration, the charging and discharging power of the energy storage system, and the renewable energy generation. Each data center then schedules its operations according to these instructions to participate in demand response.

2. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S2.1, constructing the computing power scheduling model includes the following steps: S2.1.1 Divide the workload into latency-sensitive tasks and latency-tolerant tasks according to business type, and assign different priorities. The task priority set for data center workloads is priority={0, 1, 2}. } where 0 represents the highest priority, corresponding to latency-sensitive tasks that require real-time processing and only undergo space scheduling. Lower priority tasks correspond to latency-tolerant tasks, allowing time scheduling within a specified time window, while also supporting space scheduling. To avoid overloading computing resources, the processing capacity of each priority task is constrained. Considering the flexibility of low-priority tasks, a slight exceedance of the limit is allowed within the specified timeframe, expressed by the formula: In the formula, Indicates that data center i is in The number of tasks with a priority of p are processed at any given time. This indicates the maximum utilization rate of the data center servers; This indicates the processing speed of a single server in the data center. Indicates that data center i is in The number of servers that are constantly running; Indicates the number of data centers; S2.1.2 Establishing a time scheduling model A maximum allowable transfer time Tmax is set, allowing latency-tolerant tasks to be transferred back and forth within the set maximum latency time Tmax, while satisfying the constraint that the total number of tasks that can be delayed in the data center cannot exceed the sum of the currently received tasks and the transferred tasks. This can be expressed by the formula: In the formula, This indicates that data center i is transferred from time t. The amount of tasks with a priority of 0 that are processed at any given time; This indicates that data center i is transferred from time t. The number of tasks with a priority of p are processed at any given time. Let p represent the amount of tasks with priority p received by data center i from time t, where p ≠ 0; Let p represent the amount of tasks with priority p that are transferred from data center j to data center i at time t, where p ≠ 0; S2.1.3 Establishing a spatial scheduling model Configure a multi-datacenter set as ={1, 2, Let N be the number of data centers, allowing tasks of all priorities to be located in the same set of data centers. The task transfer allocation is performed across geographical locations within the system, while satisfying the task transfer-out conservation constraint that the amount of tasks that can be transferred out cannot exceed the sum of the currently received tasks and the amount of tasks transferred in. This can be expressed by the following formula: In the formula, express The amount of tasks with priority p that are transferred from data center i to data center j at any given time. express The amount of tasks with priority p that are transferred from data center j to data center i at any given time; S2.1.4, Satisfying the constraint of total task conservation before and after data center scheduling, expressed by the formula: In the formula, , These represent data center i in The amount of tasks with priority p that are transferred via time and space scheduling at any given moment; Indicates that data center i is in The number of tasks with a priority of p are processed at any given time. This indicates that data center i is transferred from time t. The number of tasks with a priority of p are processed at any given time. Indicates that data center i is in The number of tasks with priority p that are not processed at any given time; Indicates that data center i is in The amount of tasks received at any given time; Indicates data center i from Time shift The workload at any given moment.

3. The method for multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization as described in claim 1, characterized in that: In step S2.2, constructing the power dispatch model includes the following steps: S2.2.1 Establishing an energy storage system model The capacity of an energy storage system is determined by its capacity and charging / discharging power at the previous moment. The energy storage system model is as follows: In the formula, This represents the capacity of the energy storage system of data center i at time t; , These represent the charging and discharging efficiencies of the energy storage system, respectively. , These represent the charging and discharging power of the energy storage system, respectively. , These represent the minimum and maximum values ​​of the state of charge of the energy storage system, respectively. Rated capacity of the energy storage system for data center i; , These represent the maximum charging and discharging power of the energy storage system, respectively. S2.2.2 Establishing a wind power generation system model The power output of wind power generation is determined by the rated capacity of the wind power generation system and the wind power coefficient. The wind power coefficient is related to wind speed, rotor diameter, and wind energy utilization efficiency. The wind power generation system model is as follows: In the formula, This represents the wind power generation capacity of data center i at time t; This represents the rated capacity of the wind power generation system of data center i, which is a known quantity. This represents the wind power coefficient of data center i at time t; S2.2.3 Establishing a photovoltaic power generation system model Photovoltaic power generation is related to nominal power, total solar irradiance, and overall utilization efficiency. The photovoltaic power generation system model is as follows: In the formula, This represents the photovoltaic power generation of data center i at time t; This represents the overall utilization efficiency of the solar energy system; this is a known quantity. This represents the nominal power of the photovoltaic power generation system of data center i; This represents the total solar radiation actually received by data center i at time t; This represents the irradiance under standard test conditions, expressed as 1 kW / m². S2.2.

4. Power usage efficiency is used as the energy efficiency evaluation index. To measure the energy efficiency of data centers, Power Usage Effectiveness (PUE) is used as the energy efficiency evaluation index, defined as the ratio of the total energy consumption of the data center to the energy consumption of server equipment, expressed by the formula: In the formula, Indicates power efficiency; This represents the total energy consumption of data center i at time t; This represents the energy consumption of the server equipment in data center i at time t; S2.2.5 Establish an equipment energy consumption model Server equipment energy consumption is related to its utilization rate and power consumption. The energy consumption of data center server equipment is as follows: In the formula, , These represent the power consumption of a single server device in data center i under full load and standby conditions, respectively. This represents the number of fully loaded servers in data center i at time t; This indicates the total number of servers configured in data center i; S2.2.6 Establish a total energy consumption and power balance model The power supply and demand of data centers need to be balanced in real time, which can be expressed by the formula: In the formula, This represents the power consumption of data center i at time t; This represents the photovoltaic power generation of data center i at time t; This represents the total energy consumption of data center i at time t; This represents the wind power generation capacity of data center i at time t; , These represent the charging and discharging power of the energy storage system, respectively. S2.2.7 Establish a demand response load calculation model When a data center participates in demand response, the actual response load Δ The calculation formula is: In the formula, This represents the actual response load of data center i at time t; This represents the baseline electrical load of data center i at time t.

4. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S2, to improve the robustness of the collaborative model, the following steps are also included: S2.4 Introducing Conditional Value at Risk (CVaR) to generate risk-resistant scheduling strategies Considering the predicted workload and electricity market prices To account for uncertainty, a prediction error is set, and a scene set S is generated using the Monte Carlo simulation method. The number of scenes s in the set is greater than 500. The formula for calculating the actual value under scene s is expressed as: In the formula, This represents the unit price of electricity purchased by data center i at time t under scenario s; This represents a random error sample of the electricity purchase price per unit at time t for data center i in scenario s. This represents the amount of tasks with priority p received by data center i from time t in scenario s. This represents the random error sample of the amount of tasks with priority p received by data center i from time t in scenario s. Assign probabilities to the scenes in the scene set S. This can be expressed as a formula: In the formula, This represents the probability under electricity price scenario s; This represents the probability under workload scenario s; The constraints of CVaR are expressed by the following formula: In the formula, The loss function is defined as follows: , Indicates excess loss. Indicates the maximum acceptable loss; After introducing CVaR, a balance needs to be struck between profit and high loss risk. The objective function of the coordination model is then modified as follows: In the formula, Represents the probability of scenario s; This represents the actual profit in scenario s; Indicates the risk aversion coefficient; Indicates the maximum acceptable loss; Indicates the confidence level; This indicates excess loss.

5. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S2.3, the ratio of the actual response load to the invited response is S. DR The calculation formula is: Based on the electricity demand response implementation plan for a region, determine the parameters ξ and S. DR , , The specific values, where the price subsidy coefficient ξ and S DR The relevant information is expressed by the formula: The response rate coefficient ν is related to ΔT, and can be expressed by the formula: When local power grids officially issue demand response invitations, the model parameters are further revised based on the published demand response methods, scale, time period, regional scope, and incentive price information. Then, in the subsequent step S3, a more accurate output value is obtained through model solving, and the data center fills in the response information accordingly.

6. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S1, the multi-dimensional historical time-series data of the data center includes server operating status, workload data, and electricity market price data. The server operating status includes power, CPU utilization, and server processing speed in standby and full-load states.

7. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S2.3, the reporting capacity correction factor k is set to 0.85-0.

95.

8. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S3, after importing historical data and assigning values ​​to each parameter, the coordination model is transformed into a mixed-integer linear programming problem. The Gurobi solver is used to solve the model, obtaining the optimal capacity declaration value for data centers to participate in demand response. The workload of spatial and temporal migration for each data center and The charging and discharging power of the energy storage system and Wind power generation and photovoltaic power generation .

9. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S1, LSTM is used for model training. During model prediction, historical data from the past 30-60 days is used as input, and workload and electricity price data are used as prediction targets. Considering that the data center's scheduling period is a single day, each day is divided into 96 time slots, and the duration of each time slot is Δt = 15 minutes, i.e., T = {1, 2, ...} ,96}.

10. The multi-data center computing power-power collaborative scheduling and demand response reporting capacity optimization method as described in claim 1, characterized in that: In step S1, after collecting multi-dimensional historical time-series data from the data center, the original historical data is preprocessed by filling in missing values, removing outliers, and standardizing the data before being used to make predictions using a machine learning model.

Images