A real-time energy management method and system for data center

Abstract

The present invention discloses a real-time energy management method and system for a data center, which belongs to the field of electric energy management. The method includes: establishing a real-time energy management model of a data center in a current preset period, and reconstructing the model into a Markov decision process; optimizing the optimal real-time energy management strategy of the data center by solving the Bellman equation in each period, using the batch load, power storage and cold storage three-dimensional state variables in the queue to approximate the value function respectively, overcoming the problem of difficulty in solving the value function, combining the batch load approximate value function, power storage approximate value function and cold storage approximate value function in the queue obtained by offline training in advance, obtaining a three-dimensional approximate value function and substituting it into the Bellman equation in the Markov decision process, solving the Bellman equation, and obtaining a set of decision variables that are approximately globally optimal for managing the data center. Real-time energy management can be performed on a data center running in an uncertain environment.

Description

Technical Field

[0001] The present invention belongs to the field of electric energy management, and more specifically, relates to a real-time energy management method and system for a data center. Background Art

[0002] With the development of Internet+ and cloud computing, the energy consumption of data centers has increased rapidly, and the problems of high energy consumption and high electricity costs have become increasingly prominent. Energy management and optimization of data centers are important means for operators to enhance their market competitiveness. Most of the existing data center energy management methods are aimed at data centers operating in a certain environment. In fact, when a data center is running, there will be uncertain situations where the actual value of data load, new energy output, and grid electricity price are different from the predicted value. These uncertainties will cause the optimal decision obtained in a certain environment to no longer be the optimal decision in actual operation. Therefore, existing methods cannot perform real-time energy management for data centers operating in uncertain environments.

[0003] During daily operation, data centers need to make corresponding decisions based on real-time information of random factors, and the real-time changes of random factors will seriously affect the safety and economy of real-time operation of data centers. How to ensure the economic operation of data centers in real-time operation considering the uncertainty of data load, power grid electricity price, and new energy output is an urgent problem to be solved. Summary of the invention

[0004] In view of the defects of the prior art and the need for improvement, the present invention provides a real-time energy management method and system for a data center, which aims to solve the problem that the existing energy management methods are unable to perform real-time energy management on a data center operating in an uncertain environment.

[0005] To achieve the above-mentioned purpose, according to one aspect of the present invention, a real-time energy management method for a data center is provided, comprising: S1, based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in the current preset time period and the product of the corresponding coefficient, a real-time energy management model is established, wherein the data center comprises a power supply module, an energy consumption module, an electricity storage module, a cold storage module and an electric energy exchange module; S2, with the batch processing load executed by the energy consumption module, the charging and discharging power of the electricity storage module and the storage and discharge power of the cold storage module as a set of decision variables, the real-time energy management model is reconstructed into a Markov Decision process; S3, respectively approximate the batch load, power storage and cold storage in the state variable queue in the Markov decision process, combine the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained by pre-offline training to obtain a three-dimensional approximation function, and substitute the three-dimensional approximation function into the Bellman equation in the Markov decision process; S4, solve the Bellman equation to obtain a globally optimal set of decision variables, and manage the energy consumption module, power storage module and cold storage module based on the globally optimal set of decision variables.

[0006] Furthermore, before S1, it also includes: S0′, generating multiple training scenarios based on the day-ahead predicted value and error distribution information of random factors; S0″, taking the first training scenario as the initial scenario, solving the globally optimal decision variable set of the data center under the current training scenario, and using the value function of the current training scenario in the solving process as the initial value of the value function of the next training scenario; S0″′, repeatedly executing S0″ until the globally optimal decision variable set of the data center under the last training scenario is solved, and the value function of the last training scenario in the solving process is the value function obtained by pre-discrete training.

[0007] Furthermore, the real-time energy management model is:

[0008]

[0009] Where F is the total electricity cost of the data center in the current preset period T, α t and β t are the input energy cost coefficient and output energy cost coefficient of the data center at time t, respectively; Δt is the time between two adjacent moments; and are respectively the energy consumed and the energy generated by the data center at time t.

[0010] Furthermore, the energy consumption module includes IT equipment and refrigeration equipment, and the real-time energy management model satisfies IT equipment constraints, refrigeration equipment constraints, power storage module constraints and cold storage module constraints, wherein: the IT equipment constraints include: service level agreement constraints, data center capacity constraints, IT equipment quantity constraints, CPU utilization margin constraints and IT equipment total energy consumption constraints; the refrigeration equipment constraints include: refrigeration equipment total energy consumption constraints, chiller energy consumption constraints, CRAH unit energy consumption constraints and water pump energy consumption constraints; the power storage module constraints include: power storage charging and discharging state constraints, power storage charging and discharging power upper and lower limit constraints and power storage module capacity and minimum power level constraints; the cold storage module constraints include: cold storage and discharge state constraints, cold storage and discharge power upper and lower limit constraints, cold storage module capacity and minimum energy level constraints and chiller and cold storage module interaction constraints.

[0011] Furthermore, the Markov decision process is:

[0012]

[0013]

[0014]

[0015] Among them, M t is the Markov decision process, S t is the set of state variables, x t is the set of decision variables, W t is a collection of external information, is the state transfer equation, C t is the energy management objective function, Δt is the time between two adjacent moments, d t-Δt is the batch load executed at time t-Δt, Q t is the batch processing load in the queue at time t, Batt t and TS t are the storage capacity of the electric storage module and the cold storage capacity of the cold storage module at time t, respectively. t and b t are the interactive load and batch load arriving at time t, respectively. is the energy generated by the power module at time t, α t is the cost parameter of the electrical energy input to the data center, They are respectively t , b t , α t prediction error.

[0016] Furthermore, it is used to solve the state variable set S at time t+Δtt+Δt The state transition equation for:

[0017] S t+Δt (1) = x t (1)

[0018]

[0019]

[0020]

[0021]

[0022] Among them, k is the index, S t+Δt (k), S t (k), x t (k) W t+Δt (k) are S t+Δt , S t 、x t , W t+Δt The kth element in is the set of day-ahead forecast values ​​of random factors at time t+Δt, is the charge and discharge efficiency of the power storage module, is the cooling efficiency of the cooling module, They are respectively t , b t , α t The day-ahead forecast value.

[0023] Furthermore, the three-dimensional approximation function is:

[0024]

[0025]

[0026] in, is the set of state variables after the decision at time t, is the three-dimensional approximation function, They are the batch load approximation function, the power storage approximation function, and the cooling storage approximation function. are the storage capacity of the electric storage module, the cold storage capacity of the cold storage module, and the batch processing load in the queue after the decision at time t, respectively. a，k，t is the slope of the kth piecewise linear function at time t, k a is the number of segments of the ath piecewise linear function, α=1, 2, 3, ra，k is the length of the kth segment function mapped to the horizontal axis, is the set of state variables for value function approximation.

[0027] Furthermore, after obtaining the three-dimensional approximate value function in S3, the step further includes: solving the slope of the three-dimensional approximate value function by difference, and updating the slope by iteration:

[0028]

[0029]

[0030] in, is the slope sampling value of the kth segment of the ath piecewise linear function at the nth iteration t, is the ath state variable at time t of the nth iteration, is the change of the ath state variable at time t in the nth iteration, is a state variable and the change in state variables The approximate value function of the sum of is a state variable The approximate function of is the ath state variable at the nth iteration t-Δt after the decision, is the slope value of the kth segment of the ath piecewise linear function at the nth iteration t-Δt, θ n-1 is the slope update step size, is the slope value of the kth segment of the ath piecewise linear function at the n-1th iteration t-Δt, It is the slope sampling value of the kth segment of the ath piecewise linear function at time t of the nth iteration.

[0031] Furthermore, the energy management objective function is:

[0032] C t (S t , x t , W t )=G b，t (S t , x t , W t )-G s，t (S t , x t , W t )

[0033] The Bellman equation is:

[0034]

[0035] Among them, Ct (S t , x t , W t ) is the energy management objective function, G b，t (S t , x t , W t ) is the profit value of the data center output energy, G s，t (S t , x t , W t ) is the cost of input energy to the data center, V t (S t )and are the value functions before and after decision making, respectively, γ is the attenuation factor, is the state variable after decision making.

[0036] According to another aspect of the present invention, a real-time energy management system for a data center is provided, comprising: a model building module, for building a real-time energy management model based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in a current preset time period and the product of the corresponding coefficient, wherein the data center comprises a power supply module, an energy consumption module, an electricity storage module, a cold storage module and an electric energy exchange module; a reconstruction module, for reconstructing the real-time energy management model into a Markov decision process with the batch load executed by the energy consumption module, the charging and discharging power of the electricity storage module and the storage and discharge power of the cold storage module as a set of decision variables ; A processing module, used to respectively perform value function approximation on the batch load, power storage and cold storage in the state variable queue in the Markov decision process, combine the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained in advance through offline training to obtain a three-dimensional approximation function, and substitute the three-dimensional approximation function into the Bellman equation in the Markov decision process; a solving and management module, used to solve the Bellman equation to obtain a globally optimal set of decision variables, and manage the energy consumption module, power storage module and cold storage module based on the globally optimal set of decision variables.

[0037] In general, the above technical scheme conceived by the present invention can achieve the following beneficial effects: the original real-time energy management model is reconstructed by using Markov decision process modeling, the state variables involved in the coupling constraints between time periods are approximated by value functions, and the multi-period optimization problem is converted into a series of single-period optimization problems, thereby reducing the difficulty of solving real-time energy management problems in the data center; in addition, the real-time changes of new energy processing, grid power cost parameters, interactive loads and batch loads in uncertain environments in the power module are fully considered, and the approximate value function with excellent performance obtained through offline training enables the data center to obtain a real-time energy management strategy that is close to the global optimal when it is online, thereby ensuring the economy of the data center operation; it can cooperate with the coordinated operation of multiple energy storage to utilize electricity more economically and reduce the energy consumption cost of the data center. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a data center real-time energy management method provided by an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of a data center information flow and energy flow system provided by an embodiment of the present invention;

[0040] Figures 3A-3E Schematic diagrams of training scenarios for interactive charge day-ahead prediction values, batch load day-ahead prediction values, wind power day-ahead prediction values, photovoltaic day-ahead prediction values, and grid electricity price day-ahead prediction values ​​provided in embodiments of the present invention;

[0041] Figure 4 The best power supply plan obtained by the data center real-time energy management method provided by the embodiment of the present invention for the scenario shown in FIG3 ;

[0042] Figure 5 The optimal power consumption obtained by the data center real-time energy management method provided by the embodiment of the present invention for the scenario shown in FIG3 ;

[0043] Figure 6 The relationship between data load processing and grid electricity price obtained by the data center real-time energy management method provided by the embodiment of the present invention for the scenario shown in FIG3 ;

[0044] Figure 7 A block diagram of a data center real-time energy management system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0046] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects but not necessarily to describe a specific order or sequence.

[0047] Figure 1 A flow chart of a data center real-time energy management method provided by an embodiment of the present invention. Figure 1 , combined with Figure 2-Figure 6 , the real-time energy management method for a data center in this embodiment is described in detail. Figure 1 The method includes operations S1 to S4.

[0048] Operation S1, based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in the current preset time period and the product of the corresponding coefficient, establish a real-time energy management model, the data center includes a power module, an energy consumption module, a power storage module, a cold storage module and an electric energy exchange module.

[0049] See also Figure 2 In this embodiment, the data center includes a power module, an energy consumption module, a power storage module, a cold storage module, and an electric energy exchange module. The power module includes, for example, a wind farm, a photovoltaic panel, or a combination of the two; the energy consumption module includes IT equipment (such as Figure 2 Server clusters), refrigeration equipment (such as Figure 2 cooling system) and stray power consumption items (such as Figure 2 The power storage module is, for example, a battery; the cold storage module is, for example, a cold storage tank; and the power interaction module is used to realize power interaction between the data center and the external power grid.

[0050] In this embodiment, before executing the data center real-time energy management method, it is necessary to collect technical parameters of each component in the data center. Figure 2 The technical parameters of each component of the data center shown include:

[0051] IT equipment related technical parameters: number of servers The capacity of each server F, the capacity margin of the server σ, and the energy consumption of each server when idle p idle , the peak power consumption of each server p peak , batch load maximum queue delay time T max .

[0052] Refrigeration equipment related technical parameters: Refrigeration system chiller energy consumption empirical model parameters γ 1 , γ 2 , γ 3 , the performance coefficient EER of the chiller, the minimum operating power P of the chiller chiller min .

[0053] Related technical parameters of power storage module: initial battery capacity Batt 0 , the minimum allowable battery charge level Batt , Battery capacity Maximum battery charge and discharge power and Battery charge and discharge efficiency

[0054] Technical parameters of cold storage module: Initial cooling capacity level TS of cold storage tank 0 , the minimum allowable energy level of the cold storage tank TS , Cold storage tank capacity Maximum storage and release power of cold storage tank and Cold storage tank cooling efficiency

[0055] In this embodiment, the parameters of the system server and energy storage are shown in Table 1:

[0056] Table 1

[0057]

[0058] Considering factors such as data load scheduling, server sleep, coordinated operation of multiple types of energy storage, and interaction with the power grid, the established real-time energy management model is:

[0059]

[0060] Wherein, F is the total electricity cost of the data center in the current preset period T, is the revenue value of the data center’s output energy at time t, is the cost of energy input into the data center at time t, α t and β t are the input energy cost coefficient and output energy cost coefficient of the data center at time t, respectively; Δt is the time between two adjacent moments; and are the energy consumed and energy generated by the data center at time t respectively. + When the inner is positive, the value is [] + The value of the expression in the input, otherwise, the value is 0. Enter the energy cost coefficient α tFor example, the price of electricity purchased from the outside by the data center, output energy cost coefficient β t For example, the price of the power output of the data center. It should be noted that in this embodiment, the existing α t and β t , not for α t and β t Do any management control.

[0061] for Figure 2 The data centers shown, and They are:

[0062]

[0063]

[0064]

[0065] in, is the total energy consumption of IT equipment in the energy consumption module, is the total energy consumption of the refrigeration equipment in the energy consumption module, Charging power for the power storage module, is the cooling power of the cold storage module, P Mis is the power consumption of the stray power consumption item in the energy consumption module, is the output of the power module, The wind power output is is the photovoltaic output, The amount of electricity released by the power storage module. The cold energy released by the cold storage module.

[0066] The real-time energy management model satisfies IT equipment constraints, refrigeration equipment constraints, power storage module constraints, and cold storage module constraints. IT equipment constraints include: service level agreement constraints, data center capacity constraints, IT equipment quantity constraints, CPU utilization margin constraints, and IT equipment total energy consumption constraints. Refrigeration equipment constraints include: total refrigeration equipment energy consumption constraints, chiller energy consumption constraints, CRAH unit energy consumption constraints, and water pump energy consumption constraints. Power storage module constraints include: power storage charge and discharge state constraints, power storage charge and discharge power upper and lower limits constraints, and power storage module capacity and minimum power level constraints. Cold storage module constraints include: cold storage and discharge state constraints, cold storage and discharge power upper and lower limits constraints, cold storage module capacity and minimum energy level constraints, and interaction constraints between chillers and cold storage modules.

[0067] Specifically, for IT equipment, the batch processing load Q in the initial queue of period t is t for:

[0068]

[0069] Among them, B t D is the cumulative amount of batch load that arrives at the data center and is processed in period t and all previous periods. t-1 is the cumulative amount of batch load that arrives at the data center and is processed in period t-1 and all previous periods, b τ is the batch processing load arriving at the server at time τ, d τ is the batch processing load processed by the server at time τ.

[0070] The service level agreement constraints are:

[0071]

[0072] Among them, T max The maximum queue delay for batch workloads.

[0073] The data center capacity constraint is:

[0074] μ t =a t +d t

[0075] 0≤μ t ≤N

[0076] Among them, μ t is the IT demand in period t, a t is the interactive load arriving at the data center during period t, and N is the total capacity of the data center.

[0077] The number of IT equipment (servers) is limited to:

[0078]

[0079] Among them, m t The number of activated servers; M The minimum number of servers to be enabled; is the total number of servers in the data center.

[0080] The CPU utilization margin constraint is:

[0081]

[0082]

[0083]

[0084]

[0085] in, is the CPU utilization of server i, F is the capacity of each server, s t is the CPU utilization of the data center, and σ is the capacity margin of the server.

[0086] The total energy consumption constraint of IT equipment is:

[0087]

[0088]

[0089] in, is the energy consumption of activated server i, p idle is the energy consumption of the server when it is idle, p peak is the peak power consumption of the server, is the total energy consumption of IT equipment system.

[0090] The total energy consumption constraint of the refrigeration equipment is:

[0091]

[0092] in, is the total energy consumption of the refrigeration system, is the energy consumption of the chiller, is the energy consumption of CRAH unit, Energy consumption of the water pump.

[0093] The energy consumption constraint of the chiller is:

[0094]

[0095] Among them, P IT max is the peak power consumption of IT equipment system, parameter γ 1 =0.32,γ 2 =0.11,γ 3 =0.63, EER is the performance coefficient of the chiller.

[0096] The energy consumption constraint of CRAH unit is:

[0097]

[0098]

[0099] in, is the basic energy consumption of CRAH unit, is the heat dissipation energy consumption of the CRAH unit, f is the air flow rate, and the required air flow rate of the computer room is f = f max ·s t , η heat is the performance coefficient.

[0100] The energy consumption constraint of the water pump is:

[0101]

[0102] The charge and discharge state constraints of the power storage (battery) are:

[0103]

[0104] in, and are the charge and discharge states of the battery respectively.

[0105] The upper and lower limits of the storage charging and discharging power are:

[0106]

[0107]

[0108] in, P ch and P dis are the minimum charge and discharge power of the battery respectively; and They are the battery charging and discharging power in period t, and are the maximum charge and discharge power of the battery respectively.

[0109] The capacity of the power storage module and the minimum power level constraints are:

[0110]

[0111]

[0112] Among them, Batt t is the energy level of the battery at the beginning of period t, is the charge and discharge efficiency of the battery, Batt is the minimum energy level of the battery, is the capacity of the battery.

[0113] The cooling state constraints are:

[0114]

[0115] in, and They are respectively the storage and release states of the cold storage system.

[0116] The upper and lower limits of the cooling power are:

[0117]

[0118]

[0119] in, P st and P re are the minimum storage and discharge power of the cold storage tank, and They are the storage and release power of the cold storage tank in period t, and They are the maximum cooling power of the cold storage tank respectively.

[0120] The capacity of the cold storage module and the minimum energy level constraints are:

[0121]

[0122]

[0123] Among them, TS t is the energy level at the beginning of period t, is the cold storage efficiency of the cold storage tank, TS is the minimum energy level of the cold storage tank, is the capacity of the cold storage tank.

[0124] The interaction constraints between the chiller and the cold storage module are:

[0125]

[0126] Among them, P chiller min The chiller of the refrigeration system operates at minimum power.

[0127] Operation S2 reconstructs the real-time energy management model into a Markov decision process using the batch load executed by the energy consumption module, the charging and discharging power of the power storage module, and the storing and releasing power of the cooling storage module as a set of decision variables.

[0128] Refactoring the real-time energy management model of a data center into a Markov decision process:

[0129]

[0130] Among them, M t is a Markov decision process, S t The set of state variables that need to be known to solve the data center energy management problem at time t, x t is the set of decision variables obtained when solving the data center energy management problem at time t, W t is the set of uncertain external information at time t, Solve the state transition equation of the state variables at the next moment for time t, C t The energy management objective function for energy management of the data center at time t.

[0131] Specifically, solving the data center energy management problem at time t requires knowing the batch load d executed at time t-Δt in the previous period t-Δt , the batch processing load Q in the queue at time t t , the storage capacity of the power storage module at time t Batt t , the cold storage capacity TS of the cold storage module at time t t , the interactive load a arriving at time t t , the batch load arriving at time t b t , the energy generated by the power module at time t and the cost parameter α of the electricity input to the data center at time t t , so the state variable set S at time t t for:

[0132]

[0133] Solving the data center energy management problem at time t requires determining the batch load d executed by the energy consumption module in the current period. t , Charging power of the power storage module Discharge power of the energy storage module Cooling power of the cooling module And the cooling power of the cold storage module Therefore, the decision variable set x at time t is t for:

[0134]

[0135] Considering the uncertainty of data load arrival, renewable energy output and grid electricity price, the external information set W at time t is t and the set of day-ahead forecast values ​​of random factors for:

[0136]

[0137]

[0138] in, is the prediction error of interactive load, is the prediction error of batch load, is the prediction error of the energy produced by the power module, is the prediction error of the cost parameter of the electrical energy input to the data center, is the day-ahead forecast value of interactive load, is the day-ahead forecast value of batch processing load, is the day-ahead forecast value of the energy generated by the power module, is the day-ahead forecast of the cost parameter for the electrical energy input to the data center.

[0139] The transfer function can use the coupling relationship between time periods to solve the state variables at the next moment. Specifically, it is used to solve the state variable set S at time t+Δt t+Δt The state transition equation for:

[0140] S t+Δt (1) = x t (1)

[0141]

[0142]

[0143]

[0144]

[0145] Among them, k is the index, S t+Δt (k), S t (k), x t (k) W t+Δt (k) are S t+Δt , S t 、x t , W t+Δt The kth element in is the set of day-ahead forecast values ​​of random factors at time t+Δt, is the charge and discharge efficiency of the power storage module, is the cooling efficiency of the cooling module.

[0146] The energy management objective function of the data center energy management problem at time t is:

[0147] C t (S t , x t , W t )=G b，t (S t , x t , W t )-G s，t (S t , x t , W t )

[0148] Among them, C t (S t , x t , W t ) is the energy management objective function, G b，t (S t, x t , W t ) is the profit value of the data center output energy, G s，t (S t , x t , W t ) is the cost of inputting energy into the data center.

[0149] Operation S3 performs value function approximation on the batch load, power storage and cold storage in the state variable queue in the Markov decision process respectively, combines the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained in advance through offline training to obtain a three-dimensional approximation function, and substitutes the three-dimensional approximation function into the Bellman equation in the Markov decision process.

[0150] According to an embodiment of the present invention, before operation S1, offline training is required to obtain the batch load approximation function, the power storage approximation function and the cold storage approximation function in the queue, specifically including operation S0′-operation S0″′.

[0151] In operation S0', multiple training scenarios are generated based on the day-ahead predicted value of the random factor and the error distribution information. Specifically, for example, the multiple training scenarios are generated using a Monte Carlo method.

[0152] In operation S0″, the first training scenario is used as the initial scenario to solve the global optimal decision variable set of the data center under the current training scenario, and the value function of the current training scenario in the solving process is used as the initial value of the value function of the next training scenario.

[0153] In operation S0″′, operation S0″ is repeatedly performed until the globally optimal decision variable set of the data center in the last training scenario is solved. During the solving process, the value function of the last training scenario is the value function obtained by the pre-discrete training.

[0154] The batch load Q in the state variable queue with inter-period coupling constraints is respectively calculated using a monotonically increasing piecewise linear function. t 、Battery power Batt t And the cold storage tank TS t Approximate the value function and combine it into a three-dimensional approximate value function. The obtained three-dimensional approximate value function is:

[0155]

[0156]

[0157] in, is the set of state variables after the decision at time t, is a three-dimensional approximation function, They are the batch load approximation function, the power storage approximation function, and the cooling storage approximation function. are the storage capacity of the electric storage module, the cold storage capacity of the cold storage module, and the batch processing load in the queue after the decision at time t, respectively. a，k，t is the slope of the kth piecewise linear function at time t, k a is the number of segments of the ath piecewise linear function, α=1, 2, 3, r a，k is the length of the kth segment function mapped to the horizontal axis, is the set of state variables for value function approximation.

[0158] After the three-dimensional approximate value function is obtained in operation S3, the method further includes: solving the slope of the three-dimensional approximate value function by difference, and updating the slope by iteration:

[0159]

[0160]

[0161] in, is the slope sampling value of the kth segment of the ath piecewise linear function at the nth iteration t, is the ath state variable at time t of the nth iteration, is the change of the ath state variable at time t in the nth iteration, is a state variable and the change in state variables The approximate value function of the sum of is a state variable The approximate function of is the ath state variable at the nth iteration t-Δt after the decision, is the slope value of the kth segment of the ath piecewise linear function at the nth iteration t-Δt, θ n-1 is the slope update step size, is the slope value of the kth segment of the ath piecewise linear function at the n-1th iteration t-Δt, It is the slope sampling value of the kth segment of the ath piecewise linear function at time t of the nth iteration.

[0162] Specifically, in this embodiment, a leveling algorithm is used to ensure the monotonically increasing slope of the piecewise linear function, that is:

[0163]

[0164] Furthermore, the three-dimensional approximation function is substituted into the Bellman equation in the Markov decision process:

[0165]

[0166] Among them, V t (S t )and are the value functions before and after decision making, respectively, γ is the attenuation factor, is the state variable after decision making.

[0167] Operation S4 is to solve the Bellman equation to obtain a globally optimal set of decision variables, and manage the energy consumption module, the power storage module, and the cold storage module based on the globally optimal set of decision variables.

[0168] In this embodiment, through the value function obtained by offline training and the external real-time information obtained at the current moment, the data center can make a current operation decision, which has the characteristic of being close to the global optimum.

[0169] To further illustrate the data center real-time energy management method provided by this embodiment, 500 training scenarios are generated by the Monte Carlo method using the day-ahead prediction information and error distribution to train the three-dimensional approximate value function. The trained approximate value function will be used for online optimization. During online optimization, the Monte Carlo method is used to regenerate the test scenario to simulate the real-time information within the day. The training scenario of the interactive charge day-ahead prediction value is as follows: Figure 3A As shown in the figure, the training scenario of batch load day-ahead forecast value is as follows Figure 3B As shown in Figure 2, the training scenario for the wind power day-ahead forecast value is as follows: Figure 3C As shown in the figure, the training scenario of the photovoltaic day-ahead forecast value is as follows Figure 3D As shown in the figure, the training scenario of the day-ahead electricity price forecast is as follows Figure 3E Specifically, considering the two types of data, the prediction errors of load arrival, wind power output, photovoltaic output and grid electricity price all obey the normal distribution and satisfy W t ~N(0,0.1 2 ).

[0170] The simulation results are as follows: The total offline training time is 4209.2s, which is less than two hours. The approximate value function required for intraday online optimization can be obtained by solving it on the day before. During online optimization, the average calculation time for a single test scenario is 2.20s, which is suitable for real-time applications in actual projects. The optimal power supply plan for the data center is as follows: Figure 4 , the best power consumption is Figure 5 , the relationship between data load processing and grid electricity price is as follows Figure 6 . Combined Figure 4-Figure 6It can be seen that: when the electricity price is high, by delaying the processing of batch loads, the power consumed by the data center is reduced, the electricity purchased from the outside is reduced, and there may be battery discharge and cold storage tank cooling, which reduces the electricity cost; when the electricity price is low, the data center chooses to execute batch loads as much as possible, the power consumption increases, the electricity purchased from the outside increases, and the battery and cold storage tank are charged and stored. Therefore, the real-time energy management method for the data center provided by the present invention enables batch load scheduling to actively adapt to changes in electricity prices while coordinating the operation of multiple energy storage to more economically utilize electricity and reduce the energy consumption cost of the data center.

[0171] Figure 7 A block diagram of a data center real-time energy management system provided by an embodiment of the present invention. Figure 7 The data center real-time energy management system 700 includes a model building module 710 , a reconstruction module 720 , a processing module 730 , and a solution and management module 740 .

[0172] The model building module 710, for example, executes operation S1, which is used to establish a real-time energy management model based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in the current preset time period and the product of the corresponding coefficient. The data center includes a power supply module, an energy consumption module, a power storage module, a cold storage module and an electric energy exchange module.

[0173] The reconstruction module 720, for example, performs operation S2, which is used to reconstruct the real-time energy management model into a Markov decision process using the batch load executed by the energy consumption module, the charging and discharging power of the power storage module, and the storage and release power of the cold storage module as a set of decision variables.

[0174] The processing module 730, for example, executes operation S3, which is used to respectively approximate the batch load, power storage and cold storage in the state variable queue in the Markov decision process, and combines the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained in advance through offline training to obtain a three-dimensional approximation function, and substitutes the three-dimensional approximation function into the Bellman equation in the Markov decision process.

[0175] The solving and managing module 740, for example, performs operation S4 to solve the Bellman equation, obtain a globally optimal set of decision variables, and manage the energy consumption module, the power storage module, and the cold storage module based on the globally optimal set of decision variables.

[0176] The data center real-time energy management system 700 is used to execute the above Figure 1-Figure 6 The data center real-time energy management method in the embodiment shown. For details not covered in this embodiment, please refer to the aforementioned Figure 1-Figure 6 The real-time energy management method for a data center in the illustrated embodiment will not be described in detail here.

[0177] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time energy management method for a data center, characterized in that: include: S1, establishing a real-time energy management model based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in the current preset period and the product of the corresponding coefficient, wherein the data center includes a power module, an energy consumption module, a power storage module, a cold storage module and an electric energy exchange module; S2, reconstructing the real-time energy management model into a Markov decision process with the batch processing load executed by the energy consumption module, the charging and discharging power of the power storage module, and the storage and release power of the cold storage module as a set of decision variables; S3, respectively performing value function approximation on the batch load, power storage and cold storage in the state variable queue in the Markov decision process, combining the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained by offline training in advance to obtain a three-dimensional approximation function, and substituting the three-dimensional approximation function into the Bellman equation in the Markov decision process; S4, solving the Bellman equation to obtain a globally optimal set of decision variables, and managing the energy consumption module, the power storage module, and the cold storage module based on the globally optimal set of decision variables; Wherein, the three-dimensional approximation function is: in, is the set of state variables after the decision at time t, is the three-dimensional approximation function, They are the batch load approximation function, the power storage approximation function, and the cooling storage approximation function. are the storage capacity of the electric storage module, the cold storage capacity of the cold storage module, and the batch processing load in the queue after the decision at time t, respectively. a,k,t is the slope of the kth piecewise linear function at time t, k a is the number of segments of the ath piecewise linear function, a=1,2,3, r a,k is the length of the kth segment function mapped to the horizontal axis, is the set of state variables for value function approximation.

2. The data center real-time energy management method according to claim 1, characterized in that: The S1 also includes: S0′, generates multiple training scenarios based on the day-ahead forecast value and error distribution information of random factors; S0″, taking the first training scenario as the initial scenario, solving the globally optimal decision variable set of the data center under the current training scenario, and using the value function of the current training scenario in the solving process as the initial value of the value function of the next training scenario; S0″′, repeatedly execute S0″ until the globally optimal decision variable set of the data center in the last training scenario is solved, and the value function of the last training scenario in the solving process is the value function obtained by the pre-discrete training.

3. The data center real-time energy management method according to claim 1, characterized in that: The real-time energy management model is: Where F is the total electricity cost of the data center in the current preset period T, α t and β t are the input energy cost coefficient and output energy cost coefficient of the data center at time t, respectively; Δt is the time between two adjacent moments; and are respectively the energy consumed and the energy generated by the data center at time t.

4. The data center real-time energy management method according to any one of claims 1 to 3, characterized in that: The energy consumption module includes IT equipment and refrigeration equipment, and the real-time energy management model satisfies IT equipment constraints, refrigeration equipment constraints, power storage module constraints, and cold storage module constraints, wherein: The IT equipment constraints include: service level agreement constraints, data center capacity constraints, IT equipment quantity constraints, CPU utilization margin constraints and IT equipment total energy consumption constraints; The refrigeration equipment constraints include: total energy consumption constraints of refrigeration equipment, energy consumption constraints of chillers, energy consumption constraints of CRAH units and energy consumption constraints of water pumps; The power storage module constraints include: power storage charging and discharging state constraints, power storage charging and discharging power upper and lower limit constraints, and power storage module capacity and minimum power level constraints; The cold storage module constraints include: cold storage and discharge state constraints, cold storage and discharge power upper and lower limit constraints, cold storage module capacity and minimum energy level constraints, and cold water machine and cold storage module interaction constraints.

5. The data center real-time energy management method according to claim 1, characterized in that: The Markov decision process is: M t ={S t ,x t ,W t ,f t trans ,C t } Among them, M t is the Markov decision process, S t is the set of state variables, x t is the set of decision variables, W t is the external information set, f t trans is the state transfer equation, C t is the energy management objective function, Δt is the time between two adjacent moments, d t-Δt is the batch load executed at time t-Δt, Q t is the batch processing load in the queue at time t, Batt t and TS t are the storage capacity of the electric storage module and the cold storage capacity of the cold storage module at time t, respectively. t and b t are the interactive load and batch load arriving at time t, respectively. is the energy generated by the power module at time t, α t is the cost parameter of the electric energy input to the data center, a t 、b t , α t They are respectively t 、b t , α t prediction error.

6. The data center real-time energy management method according to claim 5, characterized in that: Used to solve the state variable set S at time t+Δt t+Δt The state transition equation f t trans (S t ,x t ,W t+Δt )for: S t+Δt (1)=x t (1) Among them, k is the index, S t+Δt (k), S t (k), x t (k) W t+Δt (k) are S t+Δt , S t 、x t , W t+Δt The kth element in is the set of day-ahead forecast values ​​of random factors at time t+Δt, is the charge and discharge efficiency of the power storage module, is the cooling efficiency of the cooling module, They are respectively t 、b t , α t The day-ahead forecast value.

7. The data center real-time energy management method according to claim 5, characterized in that: After obtaining the three-dimensional approximate value function in S3, the following further includes: The slope of the three-dimensional approximation function is solved by difference method, and the slope is updated by iteration method: in, is the slope sampling value of the kth segment of the ath piecewise linear function at the nth iteration t, is the ath state variable at time t of the nth iteration, is the change of the ath state variable at time t in the nth iteration, is a state variable and the change in state variables The approximate value function of the sum of is a state variable The approximate function of is the ath state variable at the nth iteration t-Δt after the decision, is the slope value of the kth segment of the ath piecewise linear function at the nth iteration t-Δt, θ n-1 is the slope update step size, is the slope value of the kth segment of the ath piecewise linear function at the n-1th iteration t-Δt, It is the slope sampling value of the kth segment of the ath piecewise linear function at time t of the nth iteration.

8. The data center real-time energy management method according to any one of claims 5 to 7, characterized in that: The energy management objective function is: C t (S t ,x t ,W t )=G b,t (S t ,x t ,W t )-G s,t (S t ,x t ,W t ) The Bellman equation is: Among them, C t (S t ,x t ,W t ) is the energy management objective function, G b,t (S t ,x t ,W t ) is the profit value of the data center output energy, G s,t (S t ,x t ,W t ) is the cost of input energy to the data center, V t (S t )and are the value functions before and after decision making, respectively, γ is the attenuation factor, is the state variable after decision making.

9. A real-time energy management system for a data center, characterized in that: include: A model building module, used to build a real-time energy management model based on the accumulation of the difference between the energy consumed and the energy generated by the data center at each moment in the current preset period and the product of the corresponding coefficient, wherein the data center includes a power supply module, an energy consumption module, a power storage module, a cold storage module and an electric energy exchange module; A reconstruction module, used to reconstruct the real-time energy management model into a Markov decision process by taking the batch processing load executed by the energy consumption module, the charging and discharging power of the power storage module, and the storage and release power of the cold storage module as a set of decision variables; A processing module, for respectively performing value function approximation on the batch load, power storage and cold storage in the state variable queue in the Markov decision process, combining the batch load approximation function, power storage approximation function and cold storage approximation function in the queue obtained by pre-offline training to obtain a three-dimensional approximation function, and substituting the three-dimensional approximation function into the Bellman equation in the Markov decision process; A solution and management module, used for solving the Bellman equation to obtain a globally optimal set of decision variables, and managing the energy consumption module, the power storage module and the cold storage module based on the globally optimal set of decision variables; Wherein, the three-dimensional approximation function is: in, is the set of state variables after the decision at time t, is the three-dimensional approximation function, They are the batch load approximation function, the power storage approximation function, and the cooling storage approximation function. are the storage capacity of the electric storage module, the cold storage capacity of the cold storage module, and the batch processing load in the queue after the decision at time t, respectively. a,k,t is the slope of the kth piecewise linear function at time t, k a is the number of segments of the ath piecewise linear function, a=1,2,3, r a,k is the length of the kth segment function mapped to the horizontal axis, is the set of state variables for value function approximation.

Images