Data center cluster two-stage energy sharing method for multi-element load comprehensive demand response

Abstract

According to the data center cluster two-stage energy sharing method based on the multi-element load comprehensive demand response, the first stage is a day-ahead optimization stage, uncertainty probability distribution of wind power, photovoltaic power generation and loads is fully considered in the stage, and a shared energy storage (SES) optimization scheduling method based on a conditional worst-at-risk value (WCVaR) is constructed. The second stage is a real-time optimization stage: firstly, designing a novel point-to-point (P2P) transaction mechanism based on an electricity and heat supply-demand ratio so as to realize joint sharing of electricity and heat between DCPs; then, a refined IDR model is provided, the model considers the space-time transferability of data loads and the flexibility, the thermal hysteresis effect and the thermal comfort of the household appliances, and indexes such as the efficiency improvement ratio are introduced to evaluate the IDR model; and finally, constructing a benefit function of the DCCO and the DCPs.

Description

Technical Field

[0001] The present invention relates to the field of two-stage energy sharing of data center clusters, and in particular to a two-stage energy sharing method for data center clusters responding to multi-load integrated demands. Background Art

[0002] In recent years, the model of shared energy storage (SES) providing services to multiple stakeholders has become a cost-effective alternative to the solution of configuring independent distributed energy storage (DES) for each user. This approach takes advantage of economies of scale and reduces installation costs. The concept of shared energy storage is an innovative business model that opens up new paths for achieving economic benefits. In the present invention, SES is designed to promote energy sharing within a data center cluster (DCC). In order to cope with the imbalance in power distribution between different data centers, the charging and discharging strategies of SES need to be optimized to achieve efficient energy redistribution in the form of an intermediary.

[0003] Research on uncertainty has been extensive, with robust optimization (RO) and stochastic optimization (SO) being the two most widely used approaches. RO solves optimization problems under worst-case scenarios based on an uncertain set of random variables. A rule-based RO method and energy management scheme are used to coordinate energy flows between microgrid components. A two-stage RO model was constructed for a multi-energy-carrier municipal waste-to-energy system, aiming to minimize operating costs while accounting for uncertainties in heat, electricity, and hydrogen. Although RO has been proven effective, practical applications often involve problems with uncertainties characterized by probability distributions. SO utilizes this probabilistic information, with the conditional value-at-risk (CVaR) method being a widely recognized risk assessment tool. For example, the CVaR method has been used to address the risk of energy storage scheduling errors caused by uncertainty during the day-ahead period. It has also been used to achieve an optimal balance between expected returns and risks under wind power fluctuations. A simple and efficient risk assessment method based on CVaR theory has been proposed. However, this method generally assumes that random variables have fixed probability distributions. In real-world situations, accurately determining the probability distributions of random factors such as wind power, photovoltaic power generation, and various loads presents significant challenges. The worst-case value-at-risk (WCVaR) approach is used to assess the impact of uncertainty in wind power, load, and price distributions on potential worst-case profits. However, combining multiple probability distributions may not fully reflect the potential variations of all uncertain parameters. Therefore, when optimizing SES at the day-ahead stage, in-depth research is needed to determine how the WCVaR approach can adequately cover all possible probability distributions and assess operating cost risks to ensure the development of appropriate operational plans.

[0004] Optimizing scheduling and energy management in data centers integrating renewable energy (RES) has become a key area of research. Numerous methods and strategies have been proposed for designing energy management solutions. Among these, the Stackelberg game modeling energy transactions between service providers and users uses a leader-follower structure. The interaction between shared energy storage (SES) and multiple users is described as a Stackelberg game, and a new energy storage sharing scheme is proposed. Based on the Stackelberg game approach, a multi-photovoltaic prosumer energy sharing framework is constructed, coordinating the consumption behavior of energy storage devices and users through internal pricing. A Stackelberg game model is developed for multi-building prosumers and operators in zero-carbon communities, achieving energy sharing. However, these studies have not fully considered peer-to-peer (P2P) energy sharing among multiple prosumers, failing to achieve complete sharing. P2P energy sharing refers to the direct exchange of energy between prosumers or between prosumers and consumers, which can reduce associated transaction costs. Many scholars have devoted themselves to studying P2P sharing mechanisms, introducing theories such as auction models and multi-agent models. A P2P framework for microgrids was constructed using both differentiated and unified k-double auction mechanisms. Furthermore, a multi-agent deep reinforcement learning framework was introduced to explore the challenges of P2P energy trading between residential, commercial, and industrial microgrids. These approaches rely on the microgrid's electricity market, requiring buyers within the market to participate in structured bidding auctions. However, in some countries, electricity markets for demand-side participants have yet to be established, and energy policies focus solely on regulating feed-in tariffs for renewable energy.

[0005] Demand response programs have become a popular strategy in energy management, offering an alternative means of balancing supply and demand resources. Demand response has been implemented in microgrids by shifting specific loads from peak to off-peak hours or reducing load consumption. A multi-objective economic dispatch model has been developed that considers the orderly charging and discharging behavior of electric vehicles and the participation of shiftable loads. Research on demand response strategies focused on electricity loads has shown that the proactive participation of these loads can significantly improve the economics and operational efficiency of energy systems. However, these approaches overlook complex demand response characteristics, such as the flexibility of household appliances. In contrast, price-based demand response strategies have been developed for industrial, commercial, and residential microgrids, considering both electricity and heat demand. The application of demand response programs to the optimal scheduling of a "cooling-heating-electricity-gas" microgrid has been explored. Simple and effective methods and strategies have been proposed for heat load demand response, but these methods do not account for thermal hysteresis and thermal comfort, where users can tolerate slight differences between indoor and desired temperatures. Data centers, in addition to conventional loads such as electricity and heat, also contain data loads. They have great potential for participating in demand response. Inspired by traditional demand response solutions, a demand response model that considers delay-tolerant data loads to ensure reliable data center operation has been proposed. This model considers data loads in data centers as a unique flexible load, mitigating the volatility of renewable energy. This model provides important insights for data center demand response solutions. However, these studies only address a single type of data load, fail to consider the spatial and temporal transferability of data loads, and fail to fully exploit the potential flexibility of data loads. Therefore, further improvement is urgently needed for data center-based demand response solutions. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a two-stage energy sharing method for data center clusters with multi-load integrated demand response.

[0007] To achieve the above object, the technical solution of the present invention is:

[0008] The two-stage energy sharing method for data center clusters with multi-load integrated demand response is as follows:

[0009] Step 1 DCC decision

[0010] Step 1.1 System Framework

[0011] like Figure 1As shown in Figure 2, the proposed DCC energy sharing framework consists of a DCCO and multiple DCPs. Leveraging cyber-physical system integration and innovative energy operation models, the DCCO serves as an energy hub and a key shared management platform. It hosts a SES, responsible for energy sharing and information exchange, and serves as the SES's dispatch center. The DCCO acts as an agent for P2P electricity-heat transactions between DCPs. Based on the electricity-heat P2P transaction mechanism, the DCCO contracts with DCPs and provides P2P transaction services, ensuring energy and payment balances within the DCPs.

[0012] Based on this sharing framework, rational DCPs no longer interact directly with the power grid (PG) or the heat network (HN). Instead, they actively participate in the P2P electricity-heat sharing mechanism between DCPs through the DCCO. When coordinating P2P electricity transactions between DCPs, the DCCO owns the SES, whose electricity purchase and sale prices are between those of the grid. Therefore, any imbalance in electricity generated by P2P transactions between DCPs is fully offset by the DCCO. Based on the SES's charge and discharge schedules and the imbalance in electricity between DCPs, the DCCO will also trade with the grid to achieve its own dynamic electricity balance. When coordinating P2P heat transactions between DCPs, the heat imbalance generated by P2P transactions between DCPs is directly absorbed by the HN. Under the DCCO's agency, DCPs can purchase natural gas from the external natural gas network (GN) to achieve their own gas-to-electricity balance.

[0013] A data center is a prosumer that includes wind power generation, photovoltaic power generation, data load, conventional electric load (REL), and conventional heat load (RHL). Each DCP primarily consists of distributed renewable energy (wind turbines and photovoltaic panels), IT equipment, cooling systems (electric chillers (EC) and double-effect absorption chillers (DAC)), heating equipment (gas boilers (GB)), and residential buildings. Figure 2 The specific structure of the DCP is shown. It is worth noting that external energy includes electricity and heat obtained from the electricity-heat P2P trading market under the coordination of the DCCO, as well as natural gas purchased from the GN.

[0014] Step 1.2 Model of the energy component in DCC

[0015] Step 1.2.1 IT Equipment

[0016] Generally speaking, IT equipment includes servers (accounting for 80% of IT equipment power consumption) and storage and communication equipment (accounting for 20% of IT equipment power consumption). It is related to the number of servers running in real time and can be described by formula (1). When a server in DCP i does not process network tasks, its power remains at the idle power P iser-idle However, when it is in operation, it consumes peak power P i ser-peak .

[0017]

[0018] Assuming all servers within the same DCP are homogeneous, runtime tasks are evenly distributed while ensuring that the number of occupied servers is minimized, as described by the M / M / 1 queue theory:

[0019]

[0020] Combining equations (1) and (2), we can obtain equation (3):

[0021]

[0022] Power consumption of storage and communication devices It can be expressed as follows:

[0023]

[0024] Step 1.2.2 Hybrid Refrigeration System

[0025] The heat generated by IT equipment is described by formula (3). The heat generated by the server Double-effect absorption chillers (DACs) can be used to dissipate heat generated by communication storage devices. An electric refrigerator (EC) can be used for dissipation.

[0026]

[0027] From the above analysis, we can see that the total power consumed by data load is and thermal power It can be expressed as:

[0028]

[0029] Step 1.2.3 Shared Electricity Storage

[0030] The present invention selects lithium iron phosphate battery (LB) as SES, and its basic mathematical model can be expressed as formula (7).

[0031]

[0032] The cyclic aging of LB is determined by the temperature and depth of discharge during the charge and discharge process. The present invention uses a life decay model to calculate the decay of SES;

[0033] Step 1.2.4 Gas Boiler

[0034] GB provides high-temperature steam for DAC and meets the normal heat load of DCC. The model of GB can be expressed as follows:

[0035]

[0036] Step 1.3 Decision-making process

[0037] DCC decision-making process Figure 3 As shown in the figure, the system includes both a day-ahead optimization phase and a real-time optimization phase. Notably, the BES's SOC is time-coupled, making it more logical to determine its charge and discharge schedule through day-ahead optimization. In the day-ahead phase, a DCCO day-ahead stochastic optimization scheduling model based on the worst-case value at risk (VaR) is constructed to account for the uncertainties of renewable energy and load, aiming to obtain the BES's charge and discharge schedule. In the real-time optimization phase, the DCCO collects external market electricity, heat, and gas prices and uses the BES's charge and discharge schedule as real-time data for energy sharing optimization. Based on a reasonable internal electricity price set by the DCCO, the DCCO attracts DCPs to participate in P2P transactions of electricity and heat and load demand response. DCPs determine their own optimized energy consumption values based on the internal price and complete the P2P transaction. Ultimately, through iterative updates of internal prices and energy consumption, the game equilibrium between the DCCO and DCPs is achieved.

[0038] Step 2: Random optimization phase

[0039] Step 2.1 Objective Function

[0040] In DCC energy sharing, DCCO plays a leading role in the decision-making process. The day-ahead optimization goal of DCCO is to minimize the energy and carbon transaction costs with the outside world, as follows:

[0041]

[0042] Where, The unit price of electricity purchased from the grid; Purchase power from the grid for DCCO; The unit price of electricity sold to the grid; To sell electricity to the grid; The unit price of gas purchased from the gas network; Indicates the amount of gas purchased from the gas network; To buy hot unit price from hot online; To buy heat from the heat network; is the unit price of carbon trading; is the carbon trading volume of DCC.

[0043]

[0044] Where, is the total carbon emissions of DCC; Carbon emission quota for DCC; is the carbon emission coefficient of electricity purchased from the grid; The carbon emission coefficient for gas boilers consuming gas for heating is 0.359; Heating power for gas boiler; is the carbon emission factor of the electricity consumed; is the carbon emission factor of heat energy consumption; η buy Carbon emission rights for purchasing electricity from the grid; η re is the carbon emission rights of renewable energy; he Carbon emission rights for heat production by gas boilers.

[0045] Step 2.2 Constraints

[0046] The constraints in the day-ahead optimization phase include energy balance constraints, grid transaction constraints, gas transaction constraints, and heat transaction constraints. The remaining constraints include equations (1), (5), (7), and (8), which will not be repeated here.

[0047]

[0048] Step 3: Real-time optimization based on master-slave game

[0049] Step 3.1 P2P transaction mechanism of DCPs

[0050] When designing the P2P energy transaction price, three criteria must be considered: 1) In order to encourage more DCs prosumers to participate in energy sharing first, the P2P internal electricity price must be limited to Internal heat prices must be limited to 2) The economic balance of DCs energy transactions must be maintained; 3) The balance between buyers and sellers must be ensured. Considering the inverse relationship between energy prices and energy supply and demand ratio in economics, this paper establishes an internal transaction price for electricity based on the energy supply and demand ratio. and heat supply-demand ratio It can be expressed as:

[0051]

[0052] Here we need to formulate the P2P internal transaction electricity and heat prices. Here we take the internal electricity price as an example to introduce it in detail. The reference value of the internal electricity price can be described as an inverse proportional function:

[0053]

[0054] when When there is no DCP for electricity sales, all purchased electricity needs to be purchased from shared energy storage. Buy. When When the power supply and demand in DCPs are balanced, the P2P internal power purchase price and power sales price are equal to when When there is no DCP that needs to purchase electricity, the electricity that all DCPs need to sell needs to be The price is sold to shared electric energy storage. Therefore, the following equation can be obtained:

[0055]

[0056] will a e , b e and c e Substituting into formula (13) we can get V t e,ref In order to meet the economic equilibrium criterion, V t e,ref The model is refined and a complete internal electricity price model is derived.

[0057] (1) If DCPs need to purchase electricity from SES. Therefore, it is necessary to compensate SES for the price difference:

[0058]

[0059] In the interests of fairness, the difference should be divided equally among the DCPs:

[0060]

[0061] Then, when The internal electricity price at can be obtained as:

[0062]

[0063] (2) If DCPs need to sell electricity to shared energy storage. Therefore, the price difference is:

[0064]

[0065] In summary, the internal electricity price is deduced as follows:

[0066]

[0067] Similarly, the internal heat price can be derived as:

[0068]

[0069] Step 3.2 Demand Response Model of DCP

[0070] Step 3.2.1 Data Load Demand Response Model

[0071] Considering the temporal and spatial transferability of data payload, data payload can be divided into interactive data payload (IDL), batch data payload (BDL) and fixed data payload (FDL).

[0072] (1) IDL

[0073] IDL, also known as spatially portable data load, refers to data loads with fixed request processing times but can be flexibly migrated between DCPs in different geographical locations as needed. This includes data loads such as audio-visual entertainment, web browsing, and audio and video calls that do not have strict requirements on computing power or quality.

[0074]

[0075] Where, is the interactive load in DCi; is the initial interactive load of DCi; is the proportion of interactive load; is the interactive load transferred from DCi to DCi'; is the interactive load transferred from DCPi to DCPi' at time t; is the bandwidth of the communication channel; w L is the amount of work; is the propagation delay; A 0-1 variable that transfers workload from DCi to DCi';

[0076] When DCP i transfers IDL to DCP i', DCP i needs to pay compensation to DCP i'. and heat coefficient As shown in formula (24). The unit profit V of the migrated data load t d,in and outgoing data load V t d,out The unit compensation cost is shown in formula (25):

[0077]

[0078] (2)BDL

[0079] Also known as time-shiftable load, it refers to a data load with a fixed workload but flexible time within a certain time range, including big data offline analysis. Based on the transferability and delay tolerance of batch processing loads, they are divided into four categories: and The maximum forward and backward shift times of the four types of interactive loads are Δt,…,jΔt,…, and JΔt, respectively, and their proportion in each period is 1 / 4.

[0080]

[0081] Where, is the proportion of batch processing load; is the initial total batch load; is the initial value of the jth batch load; J is the number of types of batch load; is the BDL transferred from time t' to time t.

[0082] (3) Rigid load RDL

[0083] Rigid workloads are critical data loads that have strict requirements on both data processing time and quality, such as real-time dynamic query of trains and industrial process control. Rigid workloads lack operational adjustability and cannot participate in DR resource scheduling.

[0084]

[0085] According to formulas (23)-(27), the data load of DCP i after IDR can be obtained, as shown in formula (28):

[0086]

[0087] Step 3.2.2 Electricity load demand response model

[0088] According to the power load of the staff dormitory in the data center microgrid, the power load participating in demand response can be divided into four categories: transferable non-interruptible power load, transferable interruptible power load, curtailable power load and rigid load.

[0089] (1) Transferable non-interruptible load PEL

[0090] The washing machines, dishwashers, dryers, kettles, vacuum cleaners, disinfection cabinets and other equipment in the staff dormitory can be used every day. sta ,t end ] period, these devices cannot be interrupted once they start working, and they are divided into K types according to their continuous working time. The working duration of the kth type of equipment is T k =kΔt. For simplicity, it is assumed that the load size of the kth device is 1 / K of the transferable and uninterruptible load.

[0091]

[0092] Where y t,k is the operating status of the kth device, 1 means it is running; P k is the rated power of the kth device.

[0093] (2) Transferable and interruptible load TEL

[0094] Equipment such as electric vehicles can work intermittently within 24 hours, which can be expressed as:

[0095]

[0096] Where, It is the transferable and interruptible load after demand response; It is the transferable and interruptible load before demand response; is the electrical power transferred from t' to t.

[0097] (3) Rigid load CEL

[0098] Lighting, cooking equipment and other equipment in the staff dormitory are rigid loads and cannot participate in demand response. The total power load after demand response can be expressed as:

[0099]

[0100] Step 3.2.3 Heat load demand response model

[0101] During winter heating, users’ perception of heating temperature is vague, so the predicted mean value (PMV) is used. To describe the user's comfort experience of indoor temperature changes. The PMV range recommended by ISO 7730 is [-0.5, 0.5]. It is worth noting that the indoor temperature must comply with the local limit T of the heating zone. lim The initial THL before IDR is assumed to be the THL when PMV is optimal and constant equal to zero, that is,

[0102]

[0103] Among them, R clo Represents the thermal resistance of clothing; T com Represents the average temperature under comfortable conditions; M met Represents the energy metabolism rate of the human body; represents the indoor temperature of the residential building. Considering the dynamic change process of indoor temperature, the transient heat balance equation of THL and indoor temperature is shown in Equation (33).

[0104] Considering the dynamic change of room temperature, taking the indoor temperature as the state variable and the heat demand of the staff dormitory building as the heat load, the transient heat balance equation of heat demand and temperature is established:

[0105]

[0106] Where: is the outdoor temperature; K i is the heat transfer coefficient; S iis the surface area of the residential building; c air is the specific heat capacity of air; ρ air is the air density of the residential building; V i is the volume of the residential building. Since the scheduling time interval Δt is 1 hour, formula (33) can be simplified to the following form:

[0107]

[0108] (2)FHL

[0109] represents the FHL after DCP i participates in IDR, which can be expressed as:

[0110]

[0111] Step 3.3DCCO benefit function

[0112] DCCO is the agent for P2P transactions between DCPs. DCPs must provide service fees to DCCO. To ensure the profitability of DCPs participating in P2P transactions, the service fee unit price is determined by external prices and internal P2P transaction prices. DCCO's main profits include: DCPs' service fees Electricity trading income SES's residual value recovery income The main costs include: SES life decay cost Daily operation and maintenance costs of SES and carbon trading costs

[0113] (1) Service fee income

[0114]

[0115] (2) Electricity trading income

[0116]

[0117] Where, Profits from trading between DCCO and the grid; The transaction income between DCCO and DCPs; Purchase power from the grid for DCCO; Indicates the power sold by DCCO to the grid; is the power purchased by DCPs from DCCO; Represents the electric power sold by DCPs to DCCO.

[0118] (2) SEES lifespan degradation cost

[0119]

[0120] Where, and are the rated power and rated capacity of BES respectively; C p and C e are the power unit price and capacity unit price of BES respectively; r is the discount rate; δ is the annual operating days; T is the daily operating hours.

[0121] (3) SEES daily operation and maintenance costs

[0122]

[0123] Where, ζ ope is the operation and maintenance factor.

[0124] (4) SEES residual value recovery income

[0125]

[0126] Where, ζ re is the residual value recovery factor.

[0127] (5) Carbon trading costs

[0128]

[0129] Step 3.4 DCPs benefit function

[0130] Each DCP can perform demand response and P2P energy trading to achieve DCP's own energy balance and economic operation. The utility of each DCP includes: service fee cost IDL transfer costs Electricity transaction costs P2P heat transaction costs Gas transaction costs Carbon trading costs Demand response costs The utility and cost of the above DCP can be expressed as:

[0131] (1) Service fee cost

[0132]

[0133] (2) IDL transfer costs

[0134]

[0135] (3) Electricity transaction costs

[0136]

[0137] (4) Heat energy transaction costs

[0138]

[0139] (5) Gas transaction costs

[0140]

[0141] (6) Carbon trading costs

[0142]

[0143] (7) Demand response costs

[0144]

[0145] Where, β i,l , α i,l is the setting parameter of the comfort loss function of the lth load in DCP i, D i,l is the response of the lth load in DCP i.

[0146] Step 3.5 Real-time optimization scheduling model

[0147] As the profit-making entity in the regional market, DCCO should balance the conflicting interests between the two when formulating the transaction electricity price between DCCO and DCPs. DCCO and DCPs trade electricity for their respective economic goals. The upper-level operator sets the internal electricity price for the day, and the lower-level DCP performs its own optimized scheduling based on the electricity price and participates in demand response and P2P transactions. Conversely, the lower-level DCPs' energy trading strategies with the outside world will be uploaded to the upper-level model. After DCCO calculates its own benefits, it will re-formulate the electricity price accordingly until the optimal electricity price is found. The upper and lower-level models are continuously iterated until the optimal SES electricity price and the optimal scheduling strategy for DCPs are found;

[0148] In summary, there is a conflict of interest between DCCO and DCPs, and the decision-making of both parties is sequential, which conforms to the dynamic game scenario of Stackelberg hierarchy and can be described as follows:

[0149]

[0150] In the formula, DCCO is the upper-level leader and DCPs are the lower-level followers. It is the strategy set formulated by the upper-level leaders, namely the internal electricity price and hydrogen price set. is the strategy set of the lower followers. O is the objective function of DCCO, which is calculated by formula (51). Similarly, L D is the objective function of DCP i, which is calculated by formula (53).

[0151] First, the DCCO, as the leader of the game, should formulate its own strategy. Then, each DCP should formulate its own strategy to respond to the leader's strategy as a follower. Ultimately, based on the mutual influence of the DCCO and all DCP strategies, a Stackelberg game equilibrium can be reached. At this point, the Stackelberg game equilibrium solution is Need to meet:

[0152]

[0153] Step 3.5.1 Leader Model

[0154] (1) Objective function

[0155] DCCO optimizes the electricity sales and purchase prices with SES to maximize its own interests. The optimization objectives are:

[0156]

[0157] (2) Constraints

[0158]

[0159] Step 3.5.2 Follower DCP i Model

[0160] (1) Objective function

[0161] DCP i optimizes IDR strategy and P2P transaction strategy to minimize its own operating costs. The optimization goals are:

[0162]

[0163] (2) Constraints

[0164]

[0165] Step 4 Solution

[0166] Step 4.1: Optimize before

[0167] Step 4.1.1 Typical scenario generation

[0168] Considering the uncertainty and strong correlation of total wind power, total photovoltaic power, total data load, total conventional electric load, and total conventional thermal load in DCC, the correlation of the five variables is characterized by the Frank-Copula function based on historical data. The five-variable Frank-Copula function is in the form of:

[0169]

[0170] Where C(·) is the Copula function; are the probability distribution functions of wind power, photovoltaic power, data load, conventional electric load, and conventional thermal load, respectively. is a general term for the above probability distribution functions. N is the number of random variables. Subsequently, Latin hypercube sampling and K-means clustering techniques are applied to generate and reduce scenarios, resulting in S relevant representative scenarios.

[0171] Step 4.1.2 Reconstruction and solution of day-ahead optimization based on WCVaR

[0172] The uncertainty of wind power, photovoltaic power, and multi-energy loads in DCPs can lead to operational cost risks. In actual projects, it is difficult to determine the exact probability distribution information of the random variables wind power, photovoltaic power, and load. However, WCVaR can calculate the maximum risk loss when only partial probability distribution information of the random variables is known, and obtain the conditional value at risk under the worst-case scenario. Therefore, the WCVaR method is used to evaluate the day-ahead scheduling cost risk of DCCOs. First, the expression for CvaR is given as:

[0173]

[0174] Where,

[0175] [f(x, y)-β] + = max{0, f(x, y) - β}; β is the upper limit of the conditional value at risk; α is the confidence level; x is the decision variable; y is the random variable, and f(x, y) is the loss function, i.e., Equation (9). ψ(y) represents the continuous probability density function of y. The corresponding format of CvaR under discrete probability distribution is:

[0176]

[0177] Where p s is the probability of scenario s; Equation (60) only considers a specified probability distribution of random variables, but in actual situations, the probability distribution of random variables is also uncertain. Therefore, the CvaR method is extended to the WCVaR method, which can be expressed as:

[0178]

[0179] Assume that the reference probability of each scenario is , and the probability obtained based on scenario generation and reduction technology, whose credibility is still unknown. In order to ensure that the scenario probability belongs to a certain probability distribution, the box-type probability distribution uncertainty set U of wind power, photovoltaic power and load should satisfy the following form:

[0180]

[0181] Where p is the vector of scene probabilities; p 0is the reference probability distribution of the scene; ε is the disturbance variable; e is the unit vector; ε and are the lower and upper limits of the given disturbance variable, respectively; an auxiliary variable u is introduced to transform (58) into the following form:

[0182]

[0183] Among them, the auxiliary variable u should satisfy the following constraints:

[0184]

[0185] In addition, the double-layer “min-max” format (60) is difficult to solve and should be further transformed into a single-layer format by introducing auxiliary functions and duality theory, as follows:

[0186]

[0187] where g * (u) is the optimal value of the following optimization model:

[0188]

[0189] The optimal value in Equation (64) can be obtained by solving the dual model according to the Lagrangian duality theory. δ, λ, and z are dual variables. The dual model can be expressed as:

[0190]

[0191] From the above analysis, we can see that the reconstructed day-ahead stochastic optimization scheduling model can be expressed as follows:

[0192]

[0193] Where ρ1 and ρ2 are the weight coefficients of the sub-goals.

[0194] Step 4.2 Real-time master-slave game

[0195] The SES electricity price set by the DCCO not only affects the DCPs' energy trading strategies but also indirectly influences their IDR and operational decisions, leading to a nonlinear objective function for the DCCO. Therefore, the Stackelberg game model is transformed into a large-scale, multivariable mixed-integer nonlinear optimization problem. This problem can be solved using a two-layer optimization approach, where the upper layer uses the particle swarm optimization (PSO) algorithm to determine the SES's optimal electricity price. The lower layer utilizes the CPLEX solver to determine the DCPs' IDR and P2P trading strategies. By exchanging only the SES's electricity price signals and the DCPs' operational strategy signals between the two optimization layers, the complexity of the Stackelberg game is reduced, thereby improving the accuracy and efficiency of the solution process.

[0196] Compared with the prior art, the present invention has the following beneficial effects:

[0197] This paper proposes a two-stage energy sharing method for data center clusters with integrated demand response for multiple loads. It designs a DCC framework that includes DCCOs and DCPs and proposes a Stackelberg game-based DCC two-stage energy sharing model that considers peer-to-peer transactions and an IDR mechanism between DCPs. The main contributions of this paper are as follows:

[0198] 1) To fully utilize the complementary characteristics of energy consumption of different DCPs, a novel two-stage energy sharing model is proposed based on shared electric energy storage. A DCC framework consisting of DCCO and DCPs is designed. The DCCO serves as both the dispatch center of the SES and the coordinator of sharing activities.

[0199] 2) In the day-ahead phase, a SES day-ahead optimal scheduling model based on WCVaR is constructed, taking into account the uncertainty of multiple loads and the temporal coupling characteristics of the SES energy state. In the real-time phase, a real-time energy sharing framework based on Stackelberg game is proposed, in which the SES electricity pricing mechanism of the upper-level DCCO is mutually constrained by the point-to-point trading strategy and the multi-load IDR strategy between the lower-level DCPs.

[0200] 3) Considering the joint sharing of electricity and heat, a novel dynamic P2P internal pricing mechanism for electricity and heat based on the energy supply and demand ratio was developed. Unlike previous studies, this pricing mechanism is strictly controlled within the external price range to enhance the enthusiasm of data center prosumers (DCPs) to participate in P2P transactions.

[0201] 4) Considering the spatiotemporal transferability of data loads, the flexibility of household appliances, heating delay and comfort of residential users, a novel refined IDR mechanism for multi-loads is proposed for DCPs to improve the operational flexibility of DCPs. BRIEF DESCRIPTION OF THE DRAWINGS

[0202] Figure 1 Energy sharing framework diagram;

[0203] Figure 2 The specific structure within the DCP;

[0204] Figure 3 DCC decision-making process;

[0205] Figure 4 Flowchart of the two-stage energy sharing model;

[0206] Figure 5 Typical daily data curve: Figure 5 (a) Wind power of DCPs; Figure 5 (b) PV power of DCPs; Figure 5 (c) REL of DCPs; Figure 5 (d) RHL of DCPs; Figure 5 (e) Data payload of DCPs; Figure 5 (f) electricity prices and outside temperature;

[0207] Figure 6 Five typical scenarios: Figure 6 (a) Total wind power of DCPs; Figure 6 (b) Total PV power of DCPs; Figure 6 (c) Total data payload of DCPs; Figure 6 (d) total REL of DCPs; Figure 6 (e) total RHL of DCPs;

[0208] Figure 7 Optimized scheduling results today: Figure 7 (a) Independent energy storage dispatch results for DCP 1; Figure 7 (b) Independent energy storage dispatch results of DCP 2; Figure 7 (c) Independent energy storage dispatch results of DCP 3; Figure 7 (d) Shared energy storage dispatch results of DCPs;

[0209] Figure 8 Variation of WCVaR with disturbance range;

[0210] Figure 9 Electrical balance diagram of DCPs: Figure 9 (a) DCP 1; Figure 9 (b) DCP 2; Figure 9 (c) DCP 3;

[0211] Figure 10 Thermal balance diagram of DCPs: Figure 10 (a) DCP 1; Figure 10 (b) DCP 2; Figure 10 (c) DCP 3;

[0212] Figure 11 Gas balance diagram of DCPs: Figure 11 (a) DCP 1; Figure 11 (b) DCP 2; Figure 11 (c) DCP 3;

[0213] Figure 12 The transaction price of SES set by DCCO is: Figure 12 (a) Case 3; Figure 12 (b) Case 4; Figure 12 (c) Case 5; Figure 12(d) Case 6;

[0214] Figure 13 P2P transaction electricity price: Figure 13 (a) P2P transaction electricity prices; Figure 13 (b) The change curve of ESDR in the prior art; Figure 13 (c) The change curve of ESDR in the present invention;

[0215] Figure 14 P2P transaction hot price: Figure 14 (a) P2P transaction hot price; Figure 14 (b) Price curve with HSDR;

[0216] Figure 15 The IDR result of the data payload in Example 6 is: Figure 15 (a) IDL transmission results of DCP 1; Figure 15 (b) IDL transmission results of DCP 2; Figure 15 (c) IDL transmission results of DCP 3; Figure 15 (d) BDL of DCP 1 before and after IDR; Figure 15 (e) BDL of DCP2 before and after IDR; Figure 15 (f) BDL of DCP 3 before and after IDR;

[0217] Figure 16 . IDR results of REL in Case 6: Figure 16 (a) PEL of DCP before IDR; Figure 16 (b) PEL of DCP after IDR; Figure 16 (c) TEL of DCP before and after IDR;

[0218] Figure 17 THL of DCP before and after .IDR: Figure 17 (1)DCP 1; Figure 17 (b) DCP 2; Figure 17 (c) DCP 3;

[0219] Figure 18 Residential indoor temperature and PMV in DCP: Figure 18 (a) DCP 1; Figure 18 (b) DCP 2; Figure 18 (c)DCP 3. DETAILED DESCRIPTION

[0220] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0221] like Figure 1-18As shown in FIG, a two-stage energy sharing method for a data center cluster with multi-load integrated demand response is proposed. The specific steps are as follows:

[0222] 1.1 System Framework

[0223] like Figure 1 As shown in Figure 2, the proposed DCC energy sharing framework consists of a DCCO and multiple DCPs. Leveraging cyber-physical system integration and innovative energy operation models, the DCCO serves as an energy hub and a key shared management platform. It hosts a SES, responsible for energy sharing and information exchange, and serves as the SES's dispatch center. The DCCO acts as an agent for P2P electricity-heat transactions between DCPs. Based on the electricity-heat P2P transaction mechanism, the DCCO contracts with DCPs and provides P2P transaction services, ensuring energy and payment balances within the DCPs.

[0224] Based on this sharing framework, rational DCPs no longer interact directly with the power grid (PG) or the heat network (HN). Instead, they actively participate in the P2P electricity-heat sharing mechanism between DCPs through the DCCO. When coordinating P2P electricity transactions between DCPs, the DCCO owns the SES, whose electricity purchase and sale prices are between those of the grid. Therefore, any imbalance in electricity generated by P2P transactions between DCPs is fully offset by the DCCO. Based on the SES's charge and discharge schedules and the imbalance in electricity between DCPs, the DCCO will also trade with the grid to achieve its own dynamic electricity balance. When coordinating P2P heat transactions between DCPs, the heat imbalance generated by P2P transactions between DCPs is directly absorbed by the HN. Under the DCCO's agency, DCPs can purchase natural gas from the external natural gas network (GN) to achieve their own gas-to-electricity balance.

[0225] A data center is a prosumer that includes wind power generation, photovoltaic power generation, data load, conventional electric load (REL), and conventional heat load (RHL). Each DCP primarily consists of distributed renewable energy (wind turbines and photovoltaic panels), IT equipment, cooling systems (electric chillers (EC) and double-effect absorption chillers (DAC)), heating equipment (gas boilers (GB)), and residential buildings. Figure 2 The specific structure of the DCP is shown. It is worth noting that external energy includes electricity and heat obtained from the electricity-heat P2P trading market under the coordination of the DCCO, as well as natural gas purchased from the GN.

[0226] 1.2 Model of energy components in DCC

[0227] 1.2.1 IT equipment

[0228] Generally speaking, IT equipment includes servers (accounting for 80% of IT equipment power consumption) and storage and communication equipment (accounting for 20% of IT equipment power consumption). It is related to the number of servers running in real time and can be described by formula (1). When a server in DCP i does not process network tasks, its power remains at the idle power P i ser-idle However, when it is in operation, it consumes peak power P i ser-peak .

[0229]

[0230] Assuming all servers within the same DCP are homogeneous, runtime tasks are evenly distributed while ensuring that the number of occupied servers is minimized, as described by the M / M / 1 queue theory:

[0231]

[0232] Combining equations (1) and (2), we can obtain equation (3):

[0233]

[0234] Power consumption of storage and communication devices It can be expressed as follows:

[0235]

[0236] 1.2.2 Hybrid Refrigeration System

[0237] The heat generated by IT equipment is described by formula (3). The heat generated by the server Double-effect absorption chillers (DACs) can be used to dissipate heat generated by communication storage devices. An electric refrigerator (EC) can be used for dissipation.

[0238]

[0239] From the above analysis, we can see that the total power consumed by data load is and thermal power It can be expressed as:

[0240]

[0241] 1.2.3 Shared Electric Energy Storage

[0242] The present invention selects lithium iron phosphate battery (LB) as SES, and its basic mathematical model can be expressed as formula (7).

[0243]

[0244] The cyclic aging of LB is determined by the temperature and depth of discharge during the charge and discharge process. The present invention uses a life decay model to calculate the decay of SES;

[0245] Step 1.2.4 Gas Boiler

[0246] GB provides high-temperature steam for DAC and meets the normal heat load of DCC. The model of GB can be expressed as follows:

[0247]

[0248] 1.3 Decision-making process

[0249] DCC decision-making process Figure 3 As shown in the figure, the system includes both a day-ahead optimization phase and a real-time optimization phase. Notably, the BES's SOC is time-coupled, making it more logical to determine its charge and discharge schedule through day-ahead optimization. In the day-ahead phase, a DCCO day-ahead stochastic optimization scheduling model based on the worst-case value at risk (VaR) is constructed to account for the uncertainties of renewable energy and load, aiming to obtain the BES's charge and discharge schedule. In the real-time optimization phase, the DCCO collects external market electricity, heat, and gas prices and uses the BES's charge and discharge schedule as real-time data for energy sharing optimization. Based on a reasonable internal electricity price set by the DCCO, the DCCO attracts DCPs to participate in P2P transactions of electricity and heat and load demand response. DCPs determine their own optimized energy consumption values based on the internal price and complete the P2P transaction. Ultimately, through iterative updates of internal prices and energy consumption, the game equilibrium between the DCCO and DCPs is achieved.

[0250] 2 days ago random optimization phase

[0251] 2.1 Objective Function

[0252] In DCC energy sharing, DCCO plays a leading role in the decision-making process. The day-ahead optimization goal of DCCO is to minimize the energy and carbon transaction costs with the outside world, as follows:

[0253]

[0254] Where, The unit price of electricity purchased from the grid; Purchase power from the grid for DCCO; The unit price of electricity sold to the grid; To sell electricity to the grid; The unit price of gas purchased from the gas network; Indicates the amount of gas purchased from the gas network; To buy hot unit price from hot online; To buy heat from the heat network; is the unit price of carbon trading; is the carbon trading volume of DCC.

[0255]

[0256] Where, is the total carbon emissions of DCC; Carbon emission quota for DCC; is the carbon emission coefficient of electricity purchased from the grid; The carbon emission coefficient for gas boilers consuming gas for heating is 0.359; Heating power for gas boiler; is the carbon emission factor of the electricity consumed; is the carbon emission factor of heat energy consumption; η buy Carbon emission rights for purchasing electricity from the grid; η re is the carbon emission rights of renewable energy; he Carbon emission rights for heat production by gas boilers.

[0257] 2.2 Constraints

[0258] The constraints in the day-ahead optimization phase include energy balance constraints, grid transaction constraints, gas transaction constraints, and heat transaction constraints. The remaining constraints include equations (1), (5), (7), and (8), which will not be repeated here.

[0259]

[0260] 3 Real-time optimization stage based on master-slave game

[0261] 3.1 P2P Transaction Mechanism of DCPs

[0262] When designing the P2P energy transaction price, three criteria must be considered: 1) In order to encourage more DCs prosumers to participate in energy sharing first, the P2P internal electricity price must be limited to Internal heat prices must be limited to 2) The economic balance of DCs energy transactions must be maintained; 3) The balance between buyers and sellers must be ensured. Considering the inverse relationship between energy prices and energy supply and demand ratio in economics, this paper establishes an internal transaction price for electricity based on the energy supply and demand ratio. and heat supply-demand ratio It can be expressed as:

[0263]

[0264] Here we need to formulate the P2P internal transaction electricity and heat prices. Here we take the internal electricity price as an example to introduce it in detail. The reference value of the internal electricity price can be described as an inverse proportional function:

[0265]

[0266] when When there is no DCP for electricity sales, all purchased electricity needs to be purchased from shared energy storage. Buy. When When the power supply and demand in DCPs are balanced, the P2P internal power purchase price and power sales price are equal to when When there is no DCP that needs to purchase electricity, the electricity that all DCPs need to sell needs to be The price is sold to shared electric energy storage. Therefore, the following equation can be obtained:

[0267]

[0268] will a e , b e and c e Substituting into formula (13) we can get V t e,ref In order to meet the economic equilibrium criterion, V t e,ref The model is refined and a complete internal electricity price model is derived.

[0269] (1) If DCPs need to purchase electricity from SES. Therefore, it is necessary to compensate SES for the price difference:

[0270]

[0271] In the interests of fairness, the difference should be divided equally among the DCPs:

[0272]

[0273] Then, when The internal electricity price at can be obtained as:

[0274]

[0275] (2) If DCPs need to sell electricity to shared energy storage. Therefore, the price difference is:

[0276]

[0277] In summary, the internal electricity price is deduced as follows:

[0278]

[0279] Similarly, the internal heat price can be derived as:

[0280]

[0281]

[0282] 3.2DCP Demand Response Model

[0283] 3.2.1 Data Load Demand Response Model

[0284] Considering the temporal and spatial transferability of data payload, data payload can be divided into interactive data payload (IDL), batch data payload (BDL) and fixed data payload (FDL).

[0285] (1) IDL

[0286] IDL, also known as spatially portable data load, refers to data loads with fixed request processing times but can be flexibly migrated between DCPs in different geographical locations as needed. This includes data loads such as audio-visual entertainment, web browsing, and audio and video calls that do not have strict requirements on computing power or quality.

[0287]

[0288] Where, is the interactive load in DCi; is the initial interactive load of DCi; is the proportion of interactive load; is the interactive load transferred from DCi to DCi'; is the interactive load transferred from DCPi to DCPi' at time t; is the bandwidth of the communication channel; w L is the amount of work; is the propagation delay; A 0-1 variable that transfers workload from DCi to DCi';

[0289] When DCP i transfers IDL to DCP i', DCP i needs to pay compensation to DCP i'. and heat coefficient As shown in formula (24). The unit profit V of the migrated data load t d,in and outgoing data load V t d,out The unit compensation cost is shown in formula (25):

[0290]

[0291] (2)BDL

[0292] Also known as time-shiftable load, it refers to a data load with a fixed workload but flexible time within a certain time range, including big data offline analysis. Based on the transferability and delay tolerance of batch processing loads, they are divided into four categories: and The maximum forward and backward shift times of the four types of interactive loads are Δt,…,jΔt,…, and JΔt, respectively, and their proportion in each period is 1 / 4.

[0293]

[0294] Where, is the proportion of batch processing load; is the initial total batch load; is the initial value of the jth batch load; J is the number of types of batch load; is the BDL transferred from time t' to time t.

[0295] (3) Rigid load RDL

[0296] Rigid workloads are critical data loads that have strict requirements on both data processing time and quality, such as real-time dynamic query of trains and industrial process control. Rigid workloads lack operational adjustability and cannot participate in DR resource scheduling.

[0297]

[0298] According to formulas (23)-(27), the data load of DCP i after IDR can be obtained, as shown in formula (28):

[0299]

[0300] 3.2.2 Electricity load demand response model

[0301] According to the power load of the staff dormitory in the data center microgrid, the power load participating in demand response can be divided into four categories: transferable non-interruptible power load, transferable interruptible power load, curtailable power load and rigid load.

[0302] (1) Transferable non-interruptible load PEL

[0303] The washing machines, dishwashers, dryers, kettles, vacuum cleaners, disinfection cabinets and other equipment in the staff dormitory can be used every day. sta ,t end ] period, these devices cannot be interrupted once they start working, and they are divided into K types according to their continuous working time. The working duration of the kth type of equipment is Tk =kΔt. For simplicity, it is assumed that the load size of the kth device is 1 / K of the transferable and uninterruptible load.

[0304]

[0305] Where y t,k is the operating status of the kth device, 1 means it is running; P k is the rated power of the kth device.

[0306] (2) Transferable and interruptible load TEL

[0307] Equipment such as electric vehicles can work intermittently within 24 hours, which can be expressed as:

[0308]

[0309] Where, It is the transferable and interruptible load after demand response; It is the transferable and interruptible load before demand response; is the electrical power transferred from t' to t.

[0310] (3) Rigid load CEL

[0311] Lighting, cooking equipment and other equipment in the staff dormitory are rigid loads and cannot participate in demand response. The total power load after demand response can be expressed as:

[0312]

[0313] 3.2.3 Heat load demand response model

[0314] During winter heating, users’ perception of heating temperature is vague, so the predicted mean value (PMV) is used. To describe the user's comfort experience of indoor temperature changes. The PMV range recommended by ISO 7730 is [-0.5, 0.5]. It is worth noting that the indoor temperature must comply with the local limit T of the heating zone. lim The initial THL before IDR is assumed to be the THL when PMV is optimal and constant equal to zero, that is,

[0315]

[0316] Among them, R clo Represents the thermal resistance of clothing; T com Represents the average temperature under comfortable conditions; M met Represents the energy metabolism rate of the human body; represents the indoor temperature of the residential building. Considering the dynamic change process of indoor temperature, the transient heat balance equation of THL and indoor temperature is shown in Equation (33).

[0317] Considering the dynamic change of room temperature, taking the indoor temperature as the state variable and the heat demand of the staff dormitory building as the heat load, the transient heat balance equation of heat demand and temperature is established:

[0318]

[0319] Where: is the outdoor temperature; K i is the heat transfer coefficient; S i is the surface area of the residential building; c air is the specific heat capacity of air; ρ air is the air density of the residential building; V i is the volume of the residential building. Since the scheduling time interval Δt is 1 hour, formula (33) can be simplified to the following form:

[0320]

[0321] (2)FHL

[0322] represents the FHL after DCP i participates in IDR, which can be expressed as:

[0323]

[0324] 3.3DCCO Benefit Function

[0325] DCCO is the agent for P2P transactions between DCPs. DCPs must provide service fees to DCCO. To ensure the profitability of DCPs participating in P2P transactions, the service fee unit price is determined by external prices and internal P2P transaction prices. DCCO's main profits include: DCPs' service fees Electricity trading income SES's residual value recovery income The main costs include: SES life decay cost Daily operation and maintenance costs of SES and carbon trading costs

[0326] (1) Service fee income

[0327]

[0328] (2) Electricity trading income

[0329]

[0330] Where, Profits from trading between DCCO and the grid; The transaction income between DCCO and DCPs; Purchase power from the grid for DCCO; Indicates the power sold by DCCO to the grid; is the power purchased by DCPs from DCCO; Represents the electric power sold by DCPs to DCCO.

[0331] (2) SEES lifespan degradation cost

[0332]

[0333] Where, and are the rated power and rated capacity of BES respectively; C p and C e are the power unit price and capacity unit price of BES respectively; r is the discount rate; δ is the annual operating days; T is the daily operating hours.

[0334] (3) SEES daily operation and maintenance costs

[0335]

[0336] Where, ζ ope is the operation and maintenance factor.

[0337] (4) SEES residual value recovery income

[0338]

[0339] Where, ζ re is the residual value recovery factor.

[0340] (5) Carbon trading costs

[0341]

[0342] 3.4 DCPs benefit function

[0343] Each DCP can perform demand response and P2P energy trading to achieve DCP's own energy balance and economic operation. The utility of each DCP includes: service fee cost IDL transfer costs Electricity transaction costs P2P heat transaction costs Gas transaction costs Carbon trading costs Demand response costs The utility and cost of the above DCP can be expressed as:

[0344] (1) Service fee cost

[0345]

[0346] (2) IDL transfer costs

[0347]

[0348] (3) Electricity transaction costs

[0349]

[0350] (4) Heat energy transaction costs

[0351]

[0352] (5) Gas transaction costs

[0353]

[0354] (6) Carbon trading costs

[0355]

[0356] (7) Demand response costs

[0357]

[0358] Where, β i,l , α i,l is the setting parameter of the comfort loss function of the lth load in DCP i, D i,l is the response quantity of the lth load in DCP i.

[0359] 3.5 Real-time Optimization Scheduling Model

[0360] As the profit-making entity in the regional market, DCCO should balance the conflicting interests between the two when formulating the transaction electricity price between DCCO and DCPs. DCCO and DCPs trade electricity for their respective economic goals. The upper-level operator sets the internal electricity price for the day, and the lower-level DCP performs its own optimized scheduling based on the electricity price and participates in demand response and P2P transactions. Conversely, the lower-level DCPs' energy trading strategies with the outside world will be uploaded to the upper-level model. After DCCO calculates its own benefits, it will re-formulate the electricity price accordingly until the optimal electricity price is found. The upper and lower-level models are continuously iterated until the optimal SES electricity price and the optimal scheduling strategy for DCPs are found;

[0361] In summary, there is a conflict of interest between DCCO and DCPs, and the decision-making of both parties is sequential, which conforms to the dynamic game scenario of Stackelberg hierarchy and can be described as follows:

[0362]

[0363] In the formula, DCCO is the upper-level leader and DCPs are the lower-level followers. It is the strategy set formulated by the upper-level leaders, namely the internal electricity price and hydrogen price set. is the strategy set of the lower followers. O is the objective function of DCCO, which is calculated by formula (51). Similarly, L D is the objective function of DCP i, which is calculated by formula (53).

[0364] First, the DCCO, as the leader of the game, should formulate its own strategy. Then, each DCP should formulate its own strategy to respond to the leader's strategy as a follower. Ultimately, based on the mutual influence of the DCCO and all DCP strategies, a Stackelberg game equilibrium can be reached. At this point, the Stackelberg game equilibrium solution is Need to meet:

[0365]

[0366] 3.5.1 Leader Model

[0367] (1) Objective function

[0368] DCCO optimizes the electricity sales and purchase prices with SES to maximize its own interests. The optimization objectives are:

[0369]

[0370] (2) Constraints

[0371]

[0372] 3.5.2 Follower DCP i Model

[0373] (1) Objective function

[0374] DCP i optimizes IDR strategy and P2P transaction strategy to minimize its own operating costs. The optimization goals are:

[0375]

[0376] (2) Constraints

[0377]

[0378] 4 Solution method

[0379] Optimization before 4.1

[0380] 4.1.1 Typical scenario generation

[0381] Considering the uncertainty and strong correlation of total wind power, total photovoltaic power, total data load, total conventional electric load, and total conventional thermal load in DCC, the correlation of the five variables is characterized by the Frank-Copula function based on historical data. The five-variable Frank-Copula function is in the form of:

[0382]

[0383] Where C(·) is the Copula function; are the probability distribution functions of wind power, photovoltaic power, data load, conventional electric load, and conventional thermal load, respectively. is a general term for the above probability distribution functions. N is the number of random variables. Subsequently, Latin hypercube sampling and K-means clustering techniques are applied to generate and reduce scenarios, resulting in S relevant representative scenarios.

[0384] 4.1.2 Day-Ahead Optimization Reconstruction and Solution Based on WCVaR

[0385] The uncertainty of wind power, photovoltaic power, and multi-energy loads in DCPs can lead to operational cost risks. In actual projects, it is difficult to determine the exact probability distribution information of the random variables wind power, photovoltaic power, and load. However, WCVaR can calculate the maximum risk loss when only partial probability distribution information of the random variables is known, and obtain the conditional value at risk under the worst-case scenario. Therefore, the WCVaR method is used to evaluate the day-ahead scheduling cost risk of DCCOs. First, the expression for CvaR is given as:

[0386]

[0387] Where,

[0388] [f(x, y)-β] + = max{0, f(x, y) - β}; β is the upper limit of the conditional value at risk; α is the confidence level; x is the decision variable; y is the random variable, and f(x, y) is the loss function, i.e., Equation (9). ψ(y) represents the continuous probability density function of y. The corresponding format of CvaR under discrete probability distribution is:

[0389]

[0390] Where p s is the probability of scenario s; Equation (60) only considers a specified probability distribution of random variables, but in actual situations, the probability distribution of random variables is also uncertain. Therefore, the CvaR method is extended to the WCVaR method, which can be expressed as:

[0391]

[0392] Assume that the reference probability of each scenario is , and the probability obtained based on scenario generation and reduction technology, whose credibility is still unknown. In order to ensure that the scenario probability belongs to a certain probability distribution, the box-type probability distribution uncertainty set U of wind power, photovoltaic power and load should satisfy the following form:

[0393]

[0394] Where p is the vector of scene probabilities; p 0 is the reference probability distribution of the scene; ε is the disturbance variable; e is the unit vector; ε and are the lower and upper limits of the given disturbance variable, respectively; an auxiliary variable u is introduced to transform (58) into the following form:

[0395]

[0396] Among them, the auxiliary variable u should satisfy the following constraints:

[0397]

[0398] In addition, the double-layer “min-max” format (60) is difficult to solve and should be further transformed into a single-layer format by introducing auxiliary functions and duality theory, as follows:

[0399]

[0400] where g * (u) is the optimal value of the following optimization model:

[0401]

[0402] The optimal value in Equation (64) can be obtained by solving the dual model according to the Lagrangian duality theory. δ, λ, and z are dual variables. The dual model can be expressed as:

[0403]

[0404] From the above analysis, we can see that the reconstructed day-ahead stochastic optimization scheduling model can be expressed as follows:

[0405]

[0406] Where ρ1 and ρ2 are the weight coefficients of the sub-goals.

[0407] 4.2 Real-time Master-Slave Game

[0408] The SES electricity price set by the DCCO not only affects the DCPs' energy trading strategies but also indirectly influences their IDR and operational decisions, leading to a nonlinear objective function for the DCCO. Therefore, the Stackelberg game model is transformed into a large-scale, multivariable mixed-integer nonlinear optimization problem. This problem can be solved using a two-layer optimization approach, where the upper layer uses the particle swarm optimization (PSO) algorithm to determine the SES's optimal electricity price. The lower layer utilizes the CPLEX solver to determine the DCPs' IDR and P2P trading strategies. By exchanging only the SES's electricity price signals and the DCPs' operational strategy signals between the two optimization layers, the complexity of the Stackelberg game is reduced, thereby improving the accuracy and efficiency of the solution process.

[0409] 4.3 Solution Flowchart

[0410] In summary, the CPLEX solver is used to solve the day-ahead optimization model, while the real-time optimization model uses a combination of PSO and CPLEX. Figure 4 As shown in the flowchart shown.

[0411] 5 Case Analysis

[0412] 5.1 Basic Data

[0413] To evaluate the practicality and efficiency of the proposed P2P energy trading mechanism and IDR mechanism for electricity and heat, we conducted a simulation using real data collected from Karamay, Xinjiang, China. Each DCP has different resource and load attributes. Figure 5 Data for a typical day is shown, including wind and PV generation, data load, relative energy efficiency (REL), relative humidity (RHL), electricity price, and temperature. The rated capacity and power rating of the SES are set to 2500 kWh / 400 kW. Each of the three DCPs includes 100 residential users, and their equipment parameters are shown in Table 7. The energy-consuming equipment used by these users is centrally controlled, and other parameters are shown in Table 8.

[0414] In order to verify the effectiveness of the proposed two-stage energy sharing model, we designed several cases for comparative analysis, as shown in Table 1.

[0415] Table 1. Settings for multiple comparison cases

[0416]

[0417] Optimization phase before 5.2:

[0418] 5.2.1 Typical scenario generation results

[0419] In the day-ahead optimization phase of Case 6, the Frank-copula function was used to describe the correlation between the five variables of wind power, photovoltaic power generation, data load, REL, and RHL of the DCPs, and 500 solutions were randomly generated. Then, the K-means method was used to reduce these 500 solutions and obtain five typical solutions, such as Figure 6 As shown, the probabilities are 0.158, 0.190, 0.226, 0.242 and 0.184 respectively.

[0420] 5.2.2 Analysis of Optimized Scheduling Results

[0421] Because there is no energy storage, day-ahead optimization is not required in Case 1. Case 2, which includes DES, requires day-ahead optimization of the DES in all three DCPs. The day-ahead optimization results are shown in Table 2. In Case 2, the DES parameters for each DCP are set to 1000 kWh / 200 kW. The total operating cost and WCVaR are RMB 27,556.31 and RMB 43,908.92, respectively, significantly exceeding the costs associated with Cases 3-6. This demonstrates that SES is more economical than DES.

[0422] Table 2 Optimization results before

[0423]

[0424] Figure 7 (a), (b) and (c) show the optimal scheduling results of DES for each DCP in Case 2. Figure 7 (d) shows the optimization results of SES in cases 3 to 6. In order to evaluate the effectiveness of the strategy, this paper proposes and applies the power conversion frequency SF, SOC variation range VR and average charge and discharge capacity AC.

[0425]

[0426] Among them, L swi (t) indicates the sign of energy storage power switching at time t, and a value of 1 indicates that the charge and discharge state has been switched.

[0427] SF1 DES , and SF SES The values of are 5, 7, 1 and 5 respectively, from Figure 7 As can be seen intuitively, the number of power switching times of the SES is much smaller than the sum of the power switching times of the three DESs. The operating cycles of the four energy storage devices are 20, 23, 9, and 19, respectively, indicating that the SES has a low operating frequency and low operating losses. Figure 7 (a), (b), (c) and (d) and VR SESThe values are all 0.8. Comparison shows that the SOC of both the DES and SES remains within the specified range. During specific time periods, such as 1:00-2:00 and 6:00-7:00, the three DESs exhibit opposing power flows and different charge and discharge states. This discrepancy results in a certain degree of energy storage resource waste; however, the SES demonstrates its ability to effectively address this issue. and AC SES The values of are 0.2117, 0.1696, 0.2492, and 0.1315, respectively, which indicates that SES has stronger charge and discharge capabilities.

[0428] 5.2.3 Effectiveness Analysis of the WCaR Method

[0429] Taking Case 6 as an example, we analyze the impact of confidence on WCVaR and compare it with CVaR. The results are shown in Table 3. As confidence increases, WCVaR also increases, as confidence reflects the DCCO's risk appetite and conservatism. When the confidence level increases from 0.80 to 0.95, the DCCO's risk appetite gradually shifts to risk aversion, and the optimal dispatch strategy becomes more conservative. At the same confidence level, WCVaR is higher than CVaR because the latter quantifies risk based on the deterministic probability distribution of random variables and scenario expectations. Furthermore, WCVaR represents CVaR under the worst-case probability distribution, making it more adaptable to drastic fluctuations in wind and photovoltaic power generation and load, and exhibiting greater robustness.

[0430] Table 3 The impact of confidence on CVaR and WCVaR

[0431]

[0432]

[0433] In the day-ahead optimization phase, the SES charging / discharging plan obtained using the CVaR and WCVaR methods will be introduced into the real-time optimization phase, thereby affecting the real-time optimization results. In order to verify the effectiveness of the WCVaR method, we randomly generated 100 real-time scenarios at a confidence level of 0.8. Under Case 6, Table 4 lists the average optimization scheduling results of CVaR and WCVaR in the 100 sets of scenarios corresponding to the real-time phase. It can be seen that although WCVaR is higher than CVaR in the day-ahead phase, the real-time optimization results corresponding to CVaR are slightly worse than those corresponding to WCVaR. Among them, L O The CVaR is 286.43 yuan lower than the WCVaR. D The CVaR is 1946.89 yuan higher than the WCVaR, which once again proves the effectiveness of WCVaR.

[0434] Table 4 Average optimized scheduling results of CVaR and WCVaR in real-time phase.

[0435]

[0436] To disturb For example, its value range is 0.001-0.004, and the change of WCVaR with the disturbance range is shown in Figure 8 As can be seen, as ρ increases, the value of WCVaR also increases. This is because the larger the ρ value, the larger the uncertainty range of the random variable, the worse the worst-case scenario, the more conservative the resulting scheduling plan, and the greater the cost risk.

[0437] 5.3 Real-time optimization stage:

[0438] 5.3.1 Analysis of real-time optimization scheduling results

[0439] The real-time optimized scheduling results for Cases 1-6 are shown in Table 5. Case 1 is the baseline case. Without energy storage, P2P trading, or the IDR mechanism, the daily operating costs of the DCPs reached CNY 6,668.36, CNY 8,035.76, and CNY 10,368.59, respectively. Comparing Cases 2 and 3 with Case 1 shows that the addition of energy storage makes the DCPs more economical and generates CNY 1,032.87 in revenue for the DCCO. Comparing Case 5 with Case 3 shows that the addition of the energy mutual assistance mechanism reduces the daily operating costs of the three DCPs by CNY 856.35, CNY 791.28, and CNY 586.54, respectively. Comparing Case 5 with Case 6, we can see that the addition of the IDR mechanism brought 2781.04 CNY in revenue to the DCCO and reduced the daily operating costs of DCs by 11.18%, 21.65%, and 32.87%, respectively. This proves that the model proposed in this paper achieves mutual benefit and win-win results among multiple entities.

[0440] Table 5 Real-time optimization scheduling results of cases 1-6

[0441]

[0442] 5.3.2 Energy balance analysis

[0443] Taking Case 6 as an example, the energy balance state of DCPs is analyzed. The electrical balance diagram of DCPs is as follows: Figure 9 As shown in the figure, due to the incentive of P2P electricity transaction price, DCPs no longer trade directly with DCCO and the power grid, but directly participate in P2P transactions. Moreover, when participating in P2P transactions, DCPs are in the power selling and power purchasing states at different times. For example, DCP1 participates in P2P transactions and is in the power purchasing state from 1:00 to 9:00, 15:00 and 24:00, and is in the power selling state at other times. The thermal balance diagram of DCPs is shown in the figure. Figure 10As shown in Figure 2, the heat load in DCPs includes residential heat load and DAC heat consumption power. The heat energy mainly comes from GB heat production. Due to the incentive of P2P heat price, DCPs participate in P2P transactions and do not trade directly with the heat network. DCP1 has 19 periods in the heat selling state, DCP2 and DCP3 have 20 and 16 periods in the heat purchasing state, respectively. The gas balance diagram of DCPs is shown in Figure 2. Figure 11 As shown, it can be seen that the power of gas purchased by DCPs from the gas grid is always equal to the gas power consumed by GB.

[0444] 5.3.3 Analysis of Internal Transaction Prices: Master-Slave Game

[0445] Figure 12 The figure shows the transaction price set by DCCO for SES in Case 3-6. It can be seen that SES’ electricity price is strictly locked in and After participating in P2P power trading, the DCP's unbalanced electricity will be traded with the DCCO's SES instead of directly with the PG. This can effectively reduce the electricity exchange between the DCP and PG, improving the sustainability of the DCC.

[0446] 5.3.4 Analysis of energy mutual assistance results

[0447] (1) Electricity P2P trading mechanism

[0448] The electricity P2P trading mechanism proposed in this paper is applicable to cases 4-6. Taking case 6 as an example, Figure 13 (a) shows the electricity price of P2P transactions in this case. To evaluate the effectiveness of the proposed P2P transaction mechanism, we compare it with the state-of-the-art methods.

[0449] Figure 13 (b) is the curve of P2P transaction price changing with SDR in the existing technology. and When hour, and and Inversely proportional relationship, as Increase, and There will be a downward trend, but will be smaller than when hour, and Will not follow SDR t changes with the change of This will weaken the incentive effect of the SDR pricing mechanism on internal transactions. At this time, each DCP can conduct P2P transactions and interact directly with the power grid, which will increase the amount of interaction with the power grid.

[0450] Figure 13 (c) The P2P transaction price in the present invention changes with the SDR curve, the price is and Changes between. and The value of depends not only on the price mechanism of SES based on the master-slave game, but also on the real-time electricity price of the power grid. and hour, and and They show an inverse proportional relationship, with Increase, and There will be a downward trend, but will be smaller than The pricing mechanism proposed in this paper can strictly lock the P2P electricity price within It can maximize the incentives and promotion of energy sharing among DCPs and reduce the amount of interaction with the power grid.

[0451] (2) Thermal Energy P2P Trading Mechanism

[0452] Case 5 and Case 6 both include heat energy mutual aid mechanisms and use the heat energy P2P transaction mechanism proposed in this article. Figure 14 (a) shows the P2P transaction price of thermal energy in Case 6. Figure 14 (b) is the curve of P2P price changing with SDR. It can be seen that when and hour, and and They show an inverse proportional relationship, with Increase, and There will be a downward trend, but will be smaller than Since the P2P heat energy transaction price is controlled at the heat network price ( and ), the heat mutual assistance mechanism can encourage DCPs to trade heat energy with each other rather than directly interact with the heat network, which can not only improve the economy of DCPs, but also reduce the heat exchange with the heat network and enhance the self-sufficiency of DCC.

[0453] 5.3.5 Comprehensive DR Results Analysis

[0454] In order to quantitatively analyze the proposed IDR mechanism, the efficiency improvement ratio, average load transfer rate and IDR satisfaction index are proposed as follows:

[0455] (1) Efficiency improvement ratio:

[0456] According to the electricity price set by DCCO for trading with DCPs, DCPs participate in demand response and P2P trading to improve their own operation economic efficiency. To quantitatively describe the positive impact of demand response program on the economic efficiency of DCP i, the efficiency improvement ratio is defined as

[0457]

[0458] Where, and are the operating costs of DCP i before and after participating in DR, respectively.

[0459] (2) Average load transfer rate:

[0460] In order to quantitatively analyze the degree of load transfer in different scenarios, the average load transfer rate is defined as:

[0461]

[0462] Where, E i,t,l and E i,t,l,adj They represent the values of DCPi before and after the lth load participates in IDR at time t, and L represents the number of load types.

[0463] (3) Energy satisfaction:

[0464] To describe the willingness of DCPs to participate in the IDR mechanism, energy satisfaction is defined as:

[0465]

[0466] Where, β i ' ,l and β i ” ,l are the preference constant coefficients of DCPi for the lth load.

[0467] Taking Case 6 as an example, the IDR mechanism's performance indicators are shown in Table 6. The efficiency improvements for the DCPs were 11.18%, 21.65%, and 32.87%, respectively, demonstrating that the IDR mechanism effectively improves the economic efficiency of DCPs. The average load transfer rates for the three DCPs were 3.45%, 3.77%, and 3.66%, respectively, indicating that only a small portion of the load in each DCP participated in the IDR program. The energy satisfaction of DCPs after participating in DR was higher than before DR, demonstrating that the IDR mechanism can account for the subjective nature of DCPs and encourage them to implement more satisfactory dispatch plans.

[0468] Table 6 Indicators of IDR mechanism

[0469]

[0470] (1) Data load DR

[0471] Figure 5 The data load DR results for Case 6 are shown. Figure × shows the spatially transferable load before and after DR. It can be seen that DC1 transfers data load to DC2 and DC3 for 6 periods each, DC2 transfers data load to DC1 and DC3 for 8 and 9 periods respectively, and DC3 transfers data load to DC1 and DC2 for 5 and 6 periods respectively. The figure shows that data load is not transferred to and from DCP i at the same time. At 1:00, DC1 and DC2 transfer 0.98×10 5 and 2.16×10 5 This is because DC3 has excess energy to process more data tasks, and the transfer cost to DC1 and DC2 is less than the cost of purchasing energy from the outside. At 22:00, DC3 transferred 0.26 and 2.92 data tasks to DC1 and DC2 respectively. This is because DC3 cannot handle all tasks, and by transferring tasks to other DCs, it reduces its own operating costs. Figure 15 is the time-transferable load before and after participating in DR. By analogy with formula (71), the transfer rates of the time-transferable load of DCPs can be calculated to be 11.88%, 22.78%, and 19.11%, respectively.

[0472] (2) Electric load DR analysis

[0473] Figure 16 The DR results of the transferable and non-interruptible loads of DCPs are shown. It is worth noting that the DR results of the three DCPs are exactly the same, and it can be found that flexible electrical appliances participate in flexible scheduling, and the operating hours are concentrated in the peak wind and solar power periods. Figure 16 The DR results of DCPs with interruptible and transferable loads are shown, and it can be seen that the TEL curves after IDR of the three DCPs are almost the same.

[0474] (3) Heat load DR analysis

[0475] Figure 17 Shows the THL before and after IDR, Figure 18 The PMV for the residential indoor temperature and DCP is shown. It can be seen that the PMV trend is consistent with the trend for the room temperature and THL. The slight change in THL is because the PMV is limited between -0.5 and 0.5 to meet the comfort level of the residential users, resulting in a small change in the temperature control load.

[0476] 6 Conclusion

[0477] This paper introduces a DCC framework, including DCCO and DCP, supported by a SES. It also proposes a two-stage energy sharing model for DCC that integrates P2P trading and an IDR mechanism. In the first stage, a day-ahead optimization model for the SES is proposed based on WCVaR. In the second stage, an energy sharing strategy based on the Steckelberg game is developed, which incorporates P2P trading for electricity and heat and an enhanced IDR mechanism for various loads. Numerical simulations validate the effectiveness of the proposed model, leading to the following key conclusions:

[0478] 1. The developed DCC framework and energy sharing strategy fully leveraged the complementary nature of DCP's energy consumption, reducing DCP's total daily operating expenses by 39.34%. Furthermore, it generated 2,781.04 yuan in revenue per day for DCCO, fostering a mutually beneficial relationship between DCCO and DCP.

[0479] 2. Compared with traditional CVaR, the WCVaR method applied at the day-ahead stage of this invention provides a more balanced risk assessment, effectively preventing extreme optimism or pessimism in CVaR results. Furthermore, the confidence level should be consistent with the risk appetite for operating costs. When the confidence level is increased from 80% to 95%, the DCCO CVaR increases by 16.94%.

[0480] 3. Taking into account the combined power and heat supply, the P2P transaction electricity price is strictly locked between the SES electricity price, and the P2P transaction heat price is strictly locked between the HN heat price, so as to maximize the DCP's willingness to participate in P2P transactions.

[0481] 4. The proposed improved multi-load IDR mechanism can effectively achieve more economical and flexible operation of DCP. After implementing the IDR mechanism, the EIR of DCP is as high as 11.18%, 21.65% and 32.87% respectively, and the SEC of DCP is also improved.

[0482] Table 7 Parameters of user equipment

[0483]

[0484] Table 8 Other parameters during the optimization process

[0485]

[0486]

[0487] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that are not conceived through creative work should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined in the claims.

Claims

1. A two-stage energy sharing method for data center clusters with multi-load integrated demand response. The specific steps are as follows: Step 1 DCC decision Step 1.1 System Framework The DCC energy sharing framework consists of a DCCO and multiple DCPs; Relying on cyber-physical system integration and innovation in energy operation models, the DCCO plays the role of an energy hub and serves as a key shared management platform. It owns a SES, is responsible for energy sharing and information exchange, and serves as the SES's dispatch center. The DCCO acts as the agent for P2P electricity-heat transactions between DCPs. Based on the electricity-heat P2P transaction mechanism, the DCCO signs contracts with DCPs and is responsible for providing P2P transaction services to ensure their energy and payment balances. Based on this sharing framework, rational DCPs no longer interact directly with the power grid (PG) or the heat network (HN). Instead, they are happy to participate in the P2P electricity-heat sharing mechanism between DCPs through the DCCO. When coordinating P2P electricity transactions between DCPs, the DCCO owns the SES, and the SES's electricity purchase and sale prices are between those of the grid. Therefore, the DCCO fully offsets the power imbalance generated by P2P transactions between DCPs. Based on the SES's charging and discharging schedules and the power imbalance between DCPs, the DCCO will also trade with the grid to achieve its own dynamic power balance. When coordinating P2P heat transactions between DCPs, the heat imbalance generated by P2P transactions between DCPs will be directly absorbed by the HN. Under the agency of the DCCO, DCPs can purchase natural gas from the external natural gas pipeline network (GN) to achieve their own gas-to-electricity balance. A data center is a prosumer that includes wind power generation, photovoltaic power generation, data load, conventional electric load (REL), and conventional heat load (RHL). Each DCP primarily includes distributed renewable energy (wind turbines and photovoltaic panels), IT equipment, cooling systems—electric chillers (ECs) and double-effect absorption chillers (DACs), heating equipment—gas boilers (GBs), and residential buildings. External energy sources include electricity and heat obtained from the electricity-heat P2P trading market coordinated by the DCCO, as well as natural gas purchased from the GN. Step 1.2 Model of energy components in DCC; Step 1.3 Decision-making process The DCC decision-making process consists of a day-ahead optimization phase and a real-time optimization phase. Since the BES's SOC is time-coupled, it is more logical to determine the BES's charge and discharge schedule through day-ahead optimization. In the day-ahead phase, a DCCO day-ahead stochastic optimization scheduling model based on the worst-case value-at-risk (VaR) is constructed to determine the BES's charge and discharge schedule, taking into account the uncertainty of renewable energy and load. In the real-time optimization phase, the DCCO collects external market electricity, heat, and gas prices and uses the BES's charge and discharge schedule as real-time data for energy sharing optimization. Based on a reasonable internal electricity price set by the DCCO, the DCCO attracts DCPs to participate in P2P transactions of electricity and heat and load demand response. DCPs determine their own energy consumption optimization values based on the internal price and complete the P2P transactions. Ultimately, through iterative updates of internal prices and energy consumption, the game equilibrium between the DCCO and DCPs is achieved. Step 2: Random optimization phase Step 2.1 Objective Function In DCC energy sharing, DCCO plays a leading role in the decision-making process. The day-ahead optimization goal of DCCO is to minimize the energy and carbon transaction costs with the outside world, as follows: Where, The unit price of electricity purchased from the grid; Purchase power from the grid for DCCO; The unit price of electricity sold to the grid; To sell electricity to the grid; The unit price of gas purchased from the gas network; Indicates the amount of gas purchased from the gas network; To buy hot unit price from hot online; To buy heat from the heat network; is the unit price of carbon trading; is the carbon trading volume of DCC; Where, is the total carbon emissions of DCC; Carbon emission quota for DCC; is the carbon emission coefficient of electricity purchased from the grid; The carbon emission coefficient for gas boilers consuming gas for heating is 0.359; Heating power for gas boiler; is the carbon emission factor of the electricity consumed; is the carbon emission factor of heat energy consumption; η buy Carbon emission rights for purchasing electricity from the grid; η re is the carbon emission rights of renewable energy; he Carbon emission credits for heat production by gas boilers; Step 2.2 Constraints The constraints in the day-ahead optimization phase include energy balance constraints, trading constraints with the power grid, trading constraints with the gas grid, and trading constraints with the heat grid. The remaining constraints include equations (1), (5), (7), and (8), which will not be repeated here. Step 3: Real-time optimization based on master-slave game Step 3.1 P2P transaction mechanism of DCPs When designing the P2P energy transaction price, three criteria must be considered: 1) In order to encourage more DCs prosumers to participate in energy sharing first, the P2P internal electricity price must be limited to Internal heat prices must be limited to 2) the economic balance of DCs’ energy transactions must be maintained; 3) A balance must be ensured between buyers and sellers; considering the inverse relationship between energy prices and the energy supply-demand ratio in economics, this paper establishes an internal electricity transaction price based on the energy supply-demand ratio; The power supply-demand ratio and heat supply-demand ratio It can be expressed as: Here we need to formulate the P2P internal transaction electricity and heat prices. Here we take the internal electricity price as an example to introduce it in detail. The reference value of the internal electricity price can be described as an inverse proportional function: when When there is no DCP for electricity sales, all purchased electricity needs to be purchased from shared energy storage. buy; when When the power supply and demand in DCPs are balanced, the P2P internal power purchase price and power sales price are equal to when When there is no DCP that needs to purchase electricity, the electricity that all DCPs need to sell needs to be The price is sold to the shared electric energy storage; therefore, the following equation can be obtained: will a e , b e and c e Substituting into formula (13) we can get In order to meet the economic equilibrium principle, Refine and derive a complete internal electricity price model; (1) If DCPs need to purchase electricity from SES; therefore, it is necessary to compensate SES for the price difference: In the interests of fairness, the difference should be divided equally among the DCPs: Then, when The internal electricity price at can be obtained as: (2) If DCPs need to sell electricity to shared energy storage; therefore, the price difference is: In summary, the internal electricity price is deduced as follows: Similarly, the internal heat price can be derived as: Step 3.2 Demand Response Model of DCP Step 3.3DCCO benefit function DCCO is the agent for P2P transactions between DCPs. DCPs must provide service fees to DCCO. To ensure the profitability of DCPs participating in P2P transactions, the unit price of service fees is determined by external prices and internal P2P transaction prices. DCCO's main profits include: DCPs' service fees Electricity trading income SES's residual value recovery income The main costs include: SES life decay cost Daily operation and maintenance costs of SES and carbon trading costs (1) Service fee income (2) Electricity trading income Where, Profits from trading between DCCO and the grid; The transaction income between DCCO and DCPs; Purchase power from the grid for DCCO; Indicates the power sold by DCCO to the grid; is the power purchased by DCPs from DCCO; represents the electric power sold by DCPs to DCCO; (2) SEES lifespan degradation cost Where, and are the rated power and rated capacity of BES respectively; C p and C e are the power unit price and capacity unit price of BES respectively; r is the discount rate; δ is the number of operating days per year; T is the number of daily operating hours; (3) SEES daily operation and maintenance costs Where, ζ ope is the operation and maintenance factor; (4) SEES residual value recovery income Where, ζ re is the residual value recovery factor; (5) Carbon trading costs Step 3.4 DCPs benefit function Each DCP can perform demand response and P2P energy trading to achieve its own energy balance and economic operation; the utility of each DCP includes: service fee cost IDL transfer costs Electricity transaction costs P2P heat transaction costs Gas transaction costs Carbon trading costs Demand response costs The utility and cost of the above DCP can be expressed as: (1) Service fee cost (2) IDL transfer costs (3) Electricity transaction costs (4) Heat energy transaction costs (5) Gas transaction costs (6) Carbon trading costs (7) Demand response costs Where, β i,l , α i,l is the setting parameter of the comfort loss function of the lth load in DCP i, D i,l is the response of the lth load in DCP i; Step 3.5 Real-time optimization scheduling model As the profit-making entity in the regional market, the DCCO must balance the conflicting interests between it and DCPs when setting transaction prices. DCCOs and DCPs trade electricity based on their respective economic objectives. The upper-level operator sets the daily internal electricity price, and the lower-level DCP optimizes its scheduling based on the price and participates in demand response and peer-to-peer (P2P) trading. Conversely, the lower-level DCPs' external energy trading strategies are uploaded to the upper-level model. After calculating its own benefits, the DCCO will re-set the electricity price accordingly until the optimal price is found. The upper and lower-level models are continuously iterated until the optimal SES price and DCPs' optimal scheduling strategy are found. In summary, there is a conflict of interest between DCCO and DCPs, and the decision-making of both parties is sequential, which conforms to the dynamic game scenario of Stackelberg hierarchy and can be described as follows: In the formula, DCCO is the upper-level leader, and DCPs are the lower-level followers; It is the strategy set formulated by the upper-level leaders, i.e., the internal electricity price and hydrogen price set; is the strategy set of the lower followers; L O is the objective function of DCCO, which is calculated by formula (51); similarly, L D is the objective function of DCPi, calculated by formula (53); First, as the leader of the game, the DCCO should formulate its own strategy; then, each DCP should formulate its own strategy to respond to the leader's strategy as a follower; finally, based on the mutual influence of the DCCO and all DCP strategies, the Stackelberg game equilibrium can be reached; at this point, the Stackelberg game equilibrium solution Need to meet: Step 3.5.1 Leader Model (1) Objective function DCCO optimizes the electricity sales and purchase prices with SES to maximize its own interests. The optimization objectives are: (2) Constraints Step 3.5.2 Follower DCP i Model (1) Objective function DCP i optimizes IDR strategy and P2P transaction strategy to minimize its own operating costs. The optimization goals are: (2) Constraints Step 4 Solution Step 4.1: Optimize before Step 4.1.1 Typical scenario generation Considering the uncertainty and strong correlation of total wind power, total photovoltaic power, total data load, total conventional electric load, and total conventional thermal load in DCC, the correlation of the five variables is characterized by the Frank-Copula function based on historical data. The form of the five-variable Frank-Copula function is: Where C(·) is the Copula function; are the probability distribution functions of wind power, photovoltaic power, data load, conventional electric load, and conventional thermal load, respectively. is a general term for the above probability distribution functions; N is the number of random variables; Subsequently, Latin hypercube sampling and K-means clustering techniques are applied to generate and reduce scenarios, thereby forming S relevant typical scenarios; Step 4.1.2 Reconstruction and solution of day-ahead optimization based on WCVaR The uncertainty of wind power, photovoltaic power, and multi-energy loads in DCPs can lead to operational cost risks. In actual projects, it is difficult to determine the exact probability distribution information of the random variables wind power, photovoltaic power, and load. However, WCVaR can calculate the maximum risk loss when only partial probability distribution information of the random variables is known, and obtain the conditional value at risk under the worst-case scenario. Therefore, the WCVaR method is used to assess the day-ahead scheduling cost risk of DCCOs. First, the expression for CvaR is given as: Where, [f(x, y)-β]+=max{0, f(x, y)-β}; β is the upper limit of the conditional value at risk; α is the confidence level; x is the decision variable; y is the random variable, f(x, y) is the loss function, that is, formula (9); ψ(y) represents the continuous probability density function of y; the corresponding format of CvaR under discrete probability distribution is: Where p s is the probability of scenario s; Equation (60) only considers a specified probability distribution of random variables, but in actual situations, the probability distribution of random variables is also uncertain. Therefore, the CvaR method is extended to the WCVaR method, which can be expressed as: Assume that the reference probability of each scenario is , and the probability obtained based on scenario generation and reduction technology, whose credibility is still unknown; in order to ensure that the scenario probability belongs to a certain probability distribution, the box-type probability distribution uncertainty set U of wind power, photovoltaic power and load should satisfy the following form: Where p is the vector of scene probabilities; p 0 is the reference probability distribution of the scene; ε is the disturbance variable; e is the unit vector; ε and are the lower and upper limits of the given disturbance variable, respectively; an auxiliary variable u is introduced to transform (58) into the following form: Among them, the auxiliary variable u should satisfy the following constraints: In addition, the double-layer "min-max" format (60) is difficult to solve and should be further transformed into a single-layer format by introducing auxiliary functions and duality theory, as follows: where g * (u) is the optimal value of the following optimization model: The optimal value in Equation (64) can be obtained by solving the dual model according to the Lagrangian duality theory. δ, λ, and z are dual variables. The dual model can be expressed as: From the above analysis, we can see that the reconstructed day-ahead stochastic optimization scheduling model can be expressed as follows: Where ρ1 and ρ2 are the weight coefficients of the sub-goals; Step 4.2 Real-time master-slave game The SES electricity price set by the DCCO not only affects the energy trading strategy of DCPs, but also indirectly affects their IDR and operational decisions, resulting in a nonlinear objective function for the DCCO. Therefore, the Stackelberg game model is transformed into a large-scale, multivariable mixed-integer nonlinear optimization problem. This problem can be solved using a two-layer optimization approach, where the upper-layer optimization uses the particle swarm optimization (PSO) algorithm to determine the optimal electricity price of the SES, while the lower-layer optimization utilizes the CPLEX solver to determine the IDR and P2P trading strategy of the DCPs. By exchanging only the SES electricity price signal and the DCP operational strategy signal between the two optimization layers, the complexity of the Stackelberg game is reduced, thereby improving the accuracy and efficiency of the solution process.

2. The method according to claim 1, wherein: The model of the energy component in step 1.2 DCC is as follows: Step 1.2.1 IT Equipment IT equipment includes servers, storage, and communication equipment; server power consumption It is related to the number of servers running in real time and is described by formula (1); when a server in DCP i does not process network tasks, its power remains at the idle power P i ser-idle ; However, when it is in operation, it consumes peak power P i ser-peak ; Assuming all servers within the same DCP are homogeneous, runtime tasks are evenly distributed while ensuring that the number of occupied servers is minimized, as described by the M / M / 1 queue theory: Combining equations (1) and (2), we can obtain equation (3): Power consumption of storage and communication devices It can be expressed as follows: Step 1.2.2 Hybrid Refrigeration System The heat generated by IT equipment is described by formula (3); the heat generated by the server Double-effect absorption chillers (DACs) can be used to dissipate heat generated by communication storage devices. An electric refrigerator (EC) can be used to dissipate the heat; From the above analysis, we can see that the total power consumed by data load is and thermal power It can be expressed as: Step 1.2.3 Shared Electricity Storage The present invention selects lithium iron phosphate battery (LB) as SES, and its basic mathematical model can be expressed as formula (7); The cycle aging of LB is determined by the temperature and discharge depth during the charge and discharge process; the present invention uses a life decay model to calculate the decay of SES; Step 1.2.4 Gas Boiler GB provides high-temperature steam for DAC and meets the normal heat load of DCC. The model of GB can be expressed as follows:

3. The method according to claim 1, wherein: Step 3.2 The demand response model of DCP is as follows: Step 3.2.1 Data Load Demand Response Model Considering the temporal and spatial transferability of data loads, data loads can be divided into interactive data loads (IDL), batch data loads (BDL) and fixed data loads (FDL); (1) IDL IDL, also known as spatially portable data load, refers to data loads with fixed request processing times but can be flexibly migrated between DCPs in different geographical locations as needed. This includes data loads such as audio-visual entertainment, web browsing, and audio and video calls that do not have strict requirements on computing power or quality. Where, is the interactive load in DCi; is the initial interactive load of DCi; is the proportion of interactive load; is the interactive load transferred from DCi to DCi'; is the interactive load transferred from DCPi to DCPi' at time t; is the bandwidth of the communication channel; w L is the amount of work; is the propagation delay; A 0-1 variable that transfers workload from DCi to DCi'; When DCP i transfers IDL to DCP i', DCP i needs to pay compensation to DCP i'; define the power consumption coefficient and heat coefficient As shown in formula (24), the unit profit V of the migrated data load is t d,in and outgoing data load V t d,out The unit compensation cost is shown in formula (25): (2)BDL Also known as time-shiftable load, it refers to a data load with a fixed workload but flexible time within a certain time range, including big data offline analysis. Based on the transferability and delay tolerance of batch processing loads, it can be divided into four categories: and The maximum forward and backward shift times of the four types of interactive loads are Δt,…,jΔt,…, and JΔt, respectively, and their proportion in each period is 1 / 4. Where, is the proportion of batch processing load; is the initial total batch load; is the initial value of the j-th batch load; J is the number of types of batch loads; is the BDL transferred to time t at time t'; (3) Rigid load RDL Rigid workloads are important data loads that have strict requirements on both data processing time and quality, such as real-time dynamic query of trains and industrial process control. Rigid workloads lack operational adjustability and cannot participate in scheduling operations as DR resources. According to formulas (23)-(27), the data load of DCP i after IDR can be obtained, as shown in formula (28): Step 3.2.2 Electricity load demand response model According to the power load of the staff dormitory in the data center microgrid, the power load participating in demand response can be divided into four categories: transferable non-interruptible load, transferable interruptible load, curtailable load and rigid load; (1) Transferable non-interruptible load PEL The washing machines, dishwashers, dryers, kettles, vacuum cleaners, disinfection cabinets and other equipment in the staff dormitory can be used every day. sta ,t end ] period, these devices cannot be interrupted once they start working, and the devices are divided into K types according to their continuous working time; the working duration of the kth type of device is T k =kΔt, for simplicity, assume that the load size of the kth equipment is 1 / K of the transferable and uninterruptible load; Where y t,k is the operating status of the kth device, 1 means it is running; P k is the rated power of the kth device; (2) Transferable and interruptible load TEL Equipment such as electric vehicles can work intermittently within 24 hours, which can be expressed as: Where, It is the transferable and interruptible load after demand response; It is the transferable and interruptible load before demand response; is the electric power transferred from t' to t; (3) Rigid load CEL Lighting, cooking equipment and other equipment in the staff dormitory are rigid loads and cannot participate in demand response. The total power load after demand response can be expressed as: Step 3.2.3 Heat load demand response model During winter heating, users’ perception of heating temperature is vague, so the predicted mean value (PMV) is used. To describe the user's comfort experience of indoor temperature changes; the PMV range recommended by ISO 7730 is [-0.5, 0.5]; it is worth noting that the indoor temperature must comply with the local limit T of the heating zone. lim The initial THL before IDR is assumed to be the THL when PMV is optimal and constant equal to zero, that is, Among them, R clo Represents the thermal resistance of clothing; T com Represents the average temperature under comfortable conditions; M met Represents the energy metabolism rate of the human body; represents the indoor temperature of the residential building; considering the dynamic change process of the indoor temperature, the transient heat balance equation of THL and indoor temperature is shown in formula (33); Considering the dynamic change of room temperature, taking the indoor temperature as the state variable and the heat demand of the staff dormitory building as the heat load, the transient heat balance equation of heat demand and temperature is established: Where: is the outdoor temperature; K i is the heat transfer coefficient; S i is the surface area of the residential building; c air is the specific heat capacity of air; ρ air is the air density of the residential building; V i is the volume of the residential building; since the scheduling time interval Δt is 1 hour, formula (33) can be simplified to the following form: (2)FHL represents the FHL after DCP i participates in IDR, which can be expressed as:

Images