Robust optimization method for low-carbon economy of data center cluster considering multiple types of loads

Abstract

The application discloses a kind of data center cluster low-carbon economy robust optimization method considering multiple types of load, belong to intelligent power grid and energy management technical field, solve the problem of high energy consumption, low efficiency existing in the prior art data center cluster management method.The method comprises: constructing the day-ahead optimization model and real-time optimization model of data center cluster considering multiple types of load;With the first time as the time interval, with the minimum total cost of data center cluster as the target day-ahead optimization model to construct objective function;Real-time optimization model is constructed with the second time as the time interval, including the upper model for DCCO decision and the lower model for each DCP decision;Solve the day-ahead optimization model, calculate the charge and discharge plan of shared energy storage system with the first time as the time interval;The shared energy storage system charge and discharge plan is transmitted to the real-time optimization model, and the real-time optimization model is solved to carry out real-time scheduling scheme for each DCP.

Description

Technical Field

[0001] This invention relates to the field of smart grid and energy management technology, and in particular to a low-carbon, economically robust optimization method for data center clusters that considers multiple types of loads. Background Technology

[0002] Traditional DCC (Data Center Cluster) energy management primarily focuses on optimizing single-energy flow (electricity), neglecting the coupling characteristics of multiple energy flows such as electricity, heat, cooling, and computing power. Furthermore, it fails to adequately tap into the flexible response potential of various user-side loads (temperature control equipment, interruptible tasks, shared energy storage systems, etc.), leading to low energy efficiency and rigid supply-demand matching. In addition, the increasing penetration of renewable energy and heightened volatility in the electricity market further amplify the multiple uncertainties in DCC operation (such as the randomness of wind and solar power output and the heterogeneity of demand-side behavior). Existing single-stage or deterministic optimization models struggle to simultaneously achieve economic efficiency, low carbon emissions, and robustness, necessitating breakthroughs in optimization bottlenecks involving multi-timescale collaboration and multi-stakeholder game dynamics.

[0003] DCC energy management involves several aspects: In terms of multi-energy flow joint sharing modeling, existing energy dispatch optimization research lacks in-depth modeling of the coupling and coordination mechanisms of multiple energy flows such as electricity, heat, and cooling. The spatiotemporal coupling characteristics of multiple energy flows have not been fully incorporated into the joint optimization framework, leading to prominent resource competition and efficiency losses in actual dispatching. In terms of uncertain modeling and robust optimization, its core assumptions rely on known or fixed probability distribution parameters (mean and variance of wind power forecasting errors), ignoring the fuzziness and time-varying nature of the probability distribution itself in real-world scenarios. Regarding the non-dynamic nature of P2P transaction pricing mechanisms and user behavior incentives, existing peer-to-peer (P2P) energy trading mechanisms mostly adopt static pricing rules (such as fixed profit-sharing ratios or cost allocation), failing to incorporate real-time market signals, user preferences, and energy supply and demand elasticity into a dynamic game framework. Regarding the convergence and computational efficiency of real-time games, the convergence and equilibrium uniqueness of the game are difficult to guarantee, especially when data center prosumers (DCPs) adopt distributed decision-making, which may lead to strategy conflicts and price fluctuations. Traditional centralized solution algorithms face computational latency issues, making it difficult to meet the timeliness requirements of real-time scheduling.

[0004] Therefore, designing more rational energy management optimization methods for data center clusters to reduce energy consumption and improve computing efficiency is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a low-carbon, economical, and robust optimization method for data center clusters that considers multiple types of loads, in order to solve the problems of high energy consumption and low efficiency in existing data center cluster management methods.

[0006] This invention discloses a low-carbon, economical, robust optimization method for data center clusters considering multiple types of loads, the method comprising:

[0007] Construct day-ahead optimization and real-time optimization models for data center clusters that consider multiple types of loads; wherein, the objective function of the day-ahead optimization model is constructed with the first time interval as the time interval and the goal of minimizing the total day-ahead operating cost of the data center cluster; the real-time optimization model is constructed with the second time interval as the time interval, including an upper-level model for DCCO decision and a lower-level model for each DCP decision;

[0008] Solve the day-ahead optimization model to calculate the charging and discharging schedule of the shared energy storage system with the first time interval as the time interval;

[0009] The charging and discharging plans of the shared energy storage system, with the first time interval as the time interval, are transmitted to the real-time optimization model. The real-time optimization model is solved to schedule each DCP in real time.

[0010] The first time is greater than the second time.

[0011] Based on the above solution, the present invention also makes the following improvements:

[0012] Furthermore, the objective function of the current optimization model is expressed as:

[0013] (1)

[0014] in, express Electricity purchase cost during different time periods This indicates that all gas-fired boilers are in Total fuel cost over a period of time Indicating shared energy storage systems The cost of charging and discharging during specific time periods; Value at risk under worst-case conditions Risk aversion coefficient; This indicates the scheduling cycle of the current optimized model.

[0015] further,

[0016] (2)

[0017] in, express Electricity price during the specified time period express Electricity purchase volume during the time period;

[0018] (3)

[0019] in, Indicates the first The gas-fired boilers in each DCP Gas consumption during a given time period express Gas and electricity prices during specific time periods This represents the total number of DCPs;

[0020] (4)

[0021] in, This represents the loss cost coefficient of a shared energy storage system. , These respectively represent the shared energy storage system in Charging power and discharging power during a given period.

[0022] Furthermore, the decision variables in the recently optimized model include , , , .

[0023] Furthermore, the constraints of the recently optimized model include power balance constraints, thermal balance constraints, SES operation constraints, and WCVaR risk constraints.

[0024] Furthermore, the power balance constraint is expressed as:

[0025] (5)

[0026] in, Indicating shared energy storage systems Net discharge power during the period ; Indicates the first RES in a DCP The power output of wind and solar power during a given period, i.e., the power generation capacity of renewable energy sources; Indicates the first The IT server in the DCP Power consumption during a given period; Indicates the first The cooling system in each DCP Power during a given time period; Indicates the first Applications in DCP Load power during a given time period.

[0027] Furthermore, the second time is 15 minutes, and the first time is 1 hour;

[0028] The objective function, which aims to maximize DCCO revenue, is constructed as follows:

[0029] (6)

[0030] in, Indicating shared energy storage systems Trading prices during the period Indicating shared energy storage systems Net discharge power during the period; This indicates the scheduling cycle of the real-time optimization model;

[0031] The shared energy storage system's charging and discharging schedule will be based on hourly intervals. , The data is transmitted to the upper-level model within the real-time optimization model for solving the problems for each hour. , ;

[0032] These are the decision variables for the upper-level model.

[0033] Furthermore, the upper-level model satisfies the IDR flexibility constraint.

[0034] Furthermore, each DCP makes independent decisions, constructing an objective function with the goal of minimizing the real-time cost of the current DCP; the... The objective function of a DCP is expressed as:

[0035] (7)

[0036] (8)

[0037] in, Indicates the first A DCP and a shared energy storage system in Trading power during a given time period; Indicates the first A DCP at Transaction power in the P2P market during a given period; This indicates that the P2P market is in Trading prices during a specific time period; Indicates the first A DCP in Adjustment costs for different time periods, such as task delay penalties or thermal comfort compensation;

[0038] The decision variables of the lower-level model include , and It is used for real-time scheduling of the corresponding DCP.

[0039] Furthermore, the lower-level model satisfies P2P transaction balance constraints, data load delay constraints, and thermal comfort constraints.

[0040] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0041] This invention addresses the challenges of high energy consumption, high carbon emissions, and spatiotemporal mismatch in energy supply and demand within data center clusters. It proposes a novel two-stage energy-sharing model integrating multi-load integrated demand response (IDR) to construct an economical, efficient, low-carbon, and resilient energy collaborative system. This model integrates shared energy storage (SES), a hydrogen-thermal combined heat and power system, and a P2P trading platform through a two-stage optimization architecture (day-ahead scheduling and real-time decision-making) and a multi-energy flow dynamic coupling mechanism, thereby overcoming the bottleneck of heterogeneous energy cross-temporal and spatiotemporal regulation. The innovative points achieved are as follows:

[0042] (1) In terms of multi-energy flow collaborative scheduling, a unified optimization framework for cross-energy flow spatiotemporal coupling is constructed. A dynamic optimization method based on "energy flow decoupling-cooperative matching" is proposed to break through the limitations of traditional single-energy independent scheduling. By constructing a multi-energy flow energy hub model, the physical constraints and time scale differences (thermal inertia and instantaneous response of electrical energy) of electricity, heat, cold and data flow are decoupled into hierarchical optimization problems to achieve dynamic balance across energy flows. A "virtual energy bus" mechanism is introduced to allow different energy sources to be flexibly replaced according to priority in a unified spatiotemporal dimension. For example, the cooling demand can be flexibly adjusted by utilizing the computing power load of data centers, or the waste heat recovery can be linked with the peak and valley electricity prices of the power grid, thereby tapping the global collaborative potential in complex coupling.

[0043] (2) In terms of real-time scheduling game theory, a distributed collaborative and fast convergence equilibrium mechanism was established, and a hybrid architecture of "master-slave game-distributed learning" was designed to resolve the conflict between game convergence and computational efficiency in large-scale real-time scheduling. To address the equilibrium uniqueness problem in Stackelberg games, a guidance strategy based on dynamic pricing at marginal cost was proposed. This strategy constrains the policy space of game participants through price signals, preventing the divergence of distributed decision-making conflicts. Simultaneously, combined with the improved ADMM algorithm, the global optimization was decomposed into multi-agent parallel computing tasks. By utilizing edge nodes to pre-train local policies and lightweight cloud aggregation, communication overhead was significantly reduced. This architecture ensures game equilibrium while achieving minute-level real-time response capability.

[0044] (3) A comprehensive demand response mechanism was established, realizing user behavior guidance and flexible resource coordination. An incentive-compatible and flexible adaptation demand response system was constructed to solve the dual problems of insufficient user participation and behavioral uncertainty. On the one hand, by linking dynamic internal electricity prices with external market design, user energy-saving benefits were linked with system economic goals, driving proactive response with price leverage; on the other hand, based on user behavior characteristics cluster analysis, load levels of "rigid-flexible-interruptible" were divided, and response strategies (task delay fault tolerance threshold, temperature control comfort elastic range) were designed differently, so that optimization goals and user preferences were dynamically adapted, transforming users from passive participants to active collaborators, releasing the flexible potential of demand-side resources.

[0045] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0046] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0047] Figure 1 A flowchart of a low-carbon, economical, robust optimization method for data center clusters considering multiple types of loads, provided in Embodiment 1 of the present invention;

[0048] Figure 2 This is a schematic diagram of the DCC energy sharing framework provided in an embodiment of the present invention;

[0049] Figure 3 The specific internal structure of the DCP provided in the embodiments of the present invention;

[0050] Figure 4 The decision-making process of the DCC framework provided in this embodiment of the invention adopts a two-stage collaborative optimization mechanism;

[0051] Figure 5 A flowchart of the optimization algorithm provided in an embodiment of the present invention. Detailed Implementation

[0052] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0053] A specific embodiment of the present invention discloses a low-carbon, economical, robust optimization method for data center clusters that considers multiple types of loads, the flowchart of which is shown below. Figure 1 As shown, the specific explanation is as follows.

[0054] Step S1: Construct a day-ahead optimization model and a real-time optimization model for a data center cluster that considers multiple types of loads; wherein, the objective function of the day-ahead optimization model is constructed with the first time interval as the time interval and the goal of minimizing the total day-ahead operating cost of the data center cluster; the real-time optimization model is constructed with the second time interval as the time interval, including an upper-level model for DCCO decision and a lower-level model for each DCP decision.

[0055] Step S2: Solve the day-ahead optimization model to calculate the charging and discharging schedule of the shared energy storage system with the first time interval as the time interval;

[0056] Step S3: Transmit the charging and discharging plans of the shared energy storage system at first time intervals to the real-time optimization model, solve the real-time optimization model, and perform real-time scheduling of each DCP;

[0057] The first time is greater than the second time.

[0058] The DCC energy sharing framework proposed in this embodiment is as follows: Figure 2 As shown, the system comprises a Demand Response Coordinating Operator (DCCO) and multiple Distributed Producers and Consumers (DCPs), achieving distributed energy sharing through a hierarchical coordination mechanism that integrates electricity, heat, and gas. The DCC energy sharing framework utilizes Shared Energy Storage Systems (SES) for dynamic regulation and energy buffering, and reduces direct interaction costs with the power grid and heating network through a peer-to-peer (P2P) trading mechanism, thereby improving the system's economy and efficiency. Furthermore, this framework achieves complementary and globally optimized allocation of electrical, thermal, and gas energy through multi-energy flow coupling. Figure 2 Among the producers and consumers, RES (Renewable Energy System) provides renewable energy to DCP, and DCP integrates these renewable energy sources to form an integrated "generation-storage-use" energy system, and obtains electricity and heat through P2P trading mechanism to reduce the direct interaction costs with the power grid and heating network.

[0059] As a core hub, the DCCO (Distributed Generation Control Center) plays a triple role: energy dispatch center, P2P transaction agent, and shared management platform. Firstly, it dynamically adjusts the imbalance of electricity in P2P transactions between DCPs through the SES (Supply, Energy, and Power) system. Its charging and discharging prices fall within the grid price range, mitigating grid fluctuations while ensuring economic viability. Furthermore, it dynamically trades with the grid based on the SES charging and discharging plan and the DCP's energy gap to maintain system balance. Secondly, as a heat P2P transaction intermediary, it smooths the imbalance of heat between DCPs through the heating network, achieving global optimization of heat energy allocation. Thirdly, it acts as an agent for DCPs to purchase natural gas from the natural gas network (GN), ensuring power balance for heating equipment such as gas-fired boilers. As a distributed generation and consumption unit, the DCP integrates renewable energy sources such as wind and solar power, IT equipment, heating and cooling load facilities (such as electric / absorption chillers and gas-fired boilers), and residential buildings, forming an integrated "generation-storage-use" energy system. Its electricity and heat are preferentially obtained through the P2P market coordinated by the DCCO, reducing direct interaction costs with the grid (PG) and heating network (HN). This framework innovatively integrates decentralized producers and consumers into a collaborative network. Through SES buffering, multi-energy flow coupling, and P2P trading mechanisms, it constructs a highly resilient and low-cost decentralized energy sharing ecosystem, providing a new path for the efficient operation of multi-energy systems.

[0060] Figure 3 The internal structure of the DCP is shown. Figure 3 In this system, external energy supply primarily comes from electricity and heat from the electricity-heat P2P trading market coordinated by the DCCO, as well as natural gas purchased from the natural gas network (GN). The data center cluster (DCC) proposed in this embodiment consists of a data center cluster operator (DCCO) and multiple data center prosumers (DCPs), and its multi-tiered architecture includes the following three layers:

[0061] (1) Physical Layer. Each DCP integrates distributed renewable energy (wind power, photovoltaic), IT server clusters, hybrid cooling systems (air cooling + liquid cooling), gas boilers, energy storage devices, and ancillary residential buildings (including electricity / heat loads), forming a quadruple coupled energy flow of electricity, heat, cooling, and data. Among them, the IT load is affected by the spatiotemporal distribution of data traffic, the heat load is related to the building's thermal inertia, and the cooling load is driven by the server's heat dissipation requirements. Shared Energy Storage (SES) is uniformly managed by the DCCO and connected to the cluster's internal power grid through bidirectional converters, providing energy storage and release services across time scales.

[0062] (2) Information Layer. Each DCP deploys edge computing nodes to achieve local renewable energy output forecasting, load sensing, and equipment monitoring; the DCCO builds a blockchain platform to record P2P energy transactions, SES charging and discharging, and demand response (IDR) data to ensure transparency and traceability. Through a two-stage optimization interface (based on the OPC-UA protocol), the vertical integration of day-ahead optimization plans and real-time adjustment instructions is achieved, supporting efficient cross-system communication.

[0063] (3) Market layer. DCC participates in external wholesale electricity market (day-ahead / real-time market) and heat market transactions as an aggregator entity; internally, it establishes a shared energy storage leasing market (DCCO pricing) and a P2P energy sharing market (dynamic supply and demand ratio pricing) to promote the optimal allocation of resources within the cluster. This architecture achieves deep integration of energy flow, information flow and value flow through multi-level collaboration of physical-information-market.

[0064] In terms of energy flow coordination, the electricity flow achieves clean supply by prioritizing the consumption of local renewable energy. Any shortfall is met through grid purchases or storage via shared energy storage systems (SES). Surplus electricity is then traded peer-to-peer (P2P) or stored in SES, forming a dynamically balanced energy network. The heat flow utilizes a combination of gas-fired boilers and server waste heat recovery systems to supply energy, meeting building heating needs while improving system efficiency through heat storage or energy conversion, demonstrating multi-energy complementarity. The cooling flow is based on a hybrid cooling system, dynamically adjusting the ratio of air cooling to liquid cooling according to server heat dissipation requirements and ambient temperature to achieve precise matching of cooling supply and equipment load.

[0065] In terms of time-scale coordination, during the day-ahead phase, SES charging and discharging plans, external market transaction volumes, and baseline operating strategies for data center power supply equipment (DCPs) are formulated based on forecast data, aiming for optimal global economics through multi-energy coupling optimization. During the real-time phase, ultra-short-term forecast data and actual operating conditions are combined to dynamically adjust P2P transaction prices, demand-side response (IDR) strategies, and SES charging and discharging power, enhancing the system's flexibility in responding to load fluctuations and energy price changes. This mechanism achieves coordinated control across multiple time scales through hierarchical optimization, ensuring the economic efficiency and reliability of the energy system.

[0066] Figure 4The decision-making process of the DCC framework adopts a two-stage collaborative optimization mechanism—day-ahead optimization and real-time optimization—to address uncertainties at different time scales through hierarchical modeling. In the day-ahead optimization stage, the focus is on resolving the time coupling issue of the State of Charge (SOC) of the shared energy storage system (SES). Due to the continuous nature of SOC (i.e., current charging and discharging decisions affect future capacity availability), optimizing it entirely in the real-time stage could lead to short-sighted strategies and capacity mismatch. Therefore, this stage uses a stochastic optimization model to develop globally optimal charging and discharging power curves for the SES: First, a scenario tree is constructed with the goal of minimizing SES operating costs, incorporating the probability distributions of renewable energy output (wind / solar) and load fluctuations; second, Weighted Conditional Value at Risk (WCVaR) is introduced as a robustness indicator to quantify expected losses in extreme scenarios. The risk aversion intensity is adjusted through weighting coefficients, thereby generating a day-ahead plan that balances economic efficiency and robustness. The output of this stage not only provides a baseline charging and discharging strategy for the SES but also lays the boundary conditions for subsequent real-time optimization. In the real-time optimization phase, a Stackelberg game model is constructed with the DCCO as the leader and the DCPs as followers to achieve multi-stakeholder collaborative decision-making in a dynamic market environment. In the upper-level model, the DCCO, as the market leader, dynamically adjusts the electricity sales / purchase price range of the SES based on the SES charging and discharging plan formulated in the day-ahead phase and real-time price signals from the external energy market (grid PG, heating network HN, natural gas GN). On the one hand, through an electricity-heat coupling pricing mechanism (such as ESDR electricity pricing and HSDR heat pricing based on the energy supply-demand ratio), the capacity constraints and market fluctuations of the SES are transformed into price incentive signals. On the other hand, combined with the heat price information of the heating network, electricity-heat P2P trading market rules are constructed to guide energy sharing behavior among DCPs. In the lower-level model, each DCP, as a price taker, optimizes two-dimensional decisions—energy consumption planning (such as IT equipment operating hours and cooling / heating equipment output) and integrated demand response (IDR) planning (such as load shifting and interruptible load calling)—with the goal of minimizing its own energy costs under the given P2P internal electricity and heat prices. Through iterative solutions of upper and lower level models (such as dual transformations based on KKT conditions or heuristic algorithms), the system eventually converges to a Stackelberg equilibrium: the Distributed Controller Platform (DCP) obtains the optimal scheduling scheme that satisfies individual economics, while the Distributed Controller Operational Cost (DCCO) determines the real-time electricity price of the Segregated Electricity Entity (SES) and ensures that the SES's State of Charge (SOC) state dynamically matches the day-ahead plan. This two-layer mechanism avoids the information overload problem of centralized optimization and enables autonomous coordination among distributed entities through price signals, significantly improving the system's adaptability to real-time fluctuations.

[0067] The following is an explanation of the current optimization model.

[0068] In the day-ahead optimization phase, the goal is to develop a global charge-discharge plan for the shared energy storage system (SES) by coordinating the multi-energy coupling relationships of electricity, heat, and data load within the data center cluster (DCC). This ensures system robustness while reserving flexibility for real-time operation. The core approach includes three aspects: First, to address the uncertainty of the probability distribution of wind and solar power output and load demand, the worst-case value at risk (WCVaR) model is used to quantify the risks of extreme scenarios, constructing risk avoidance constraints to enhance the robustness of the scheduling strategy. Second, based on the energy storage capacity limitations and charge-discharge efficiency characteristics of the SES, time coupling constraints are designed to ensure the continuity of energy scheduling between the day-ahead plan and the real-time phase, avoiding insufficient capacity or overcharging / discharging issues. In addition, by coordinating energy declaration information from the data center operator (DCCO) and multiple energy providers (DCPs), with the goal of optimizing the overall economics of the DCCC, the SES charge-discharge period and external market trading strategies are optimized, prioritizing the consumption of local renewable energy and reducing the cost of purchased electricity. The key design features of this stage are reflected in the triple coordination of "risk-time-entity": suppressing the impact of wind and solar fluctuations and load mutations through WCVaR risk constraints, ensuring the physical feasibility of the SES scheduling plan through time coupling constraints, and achieving a balance between global optimization and flexibility reservation through multi-entity coordination mechanisms, thereby laying the foundation for real-time dynamic adjustment.

[0069] (1) Current optimization model

[0070] Preferably, in the day-ahead optimization model, an objective function is constructed with an hourly time interval and the goal of minimizing the total day-ahead operating cost of the data center cluster. This objective function aims to minimize the total day-ahead cost of the DCC, including the cost of purchased electricity, gas costs, and energy storage losses, while reducing the financial risk caused by the uncertainty of wind and solar power output through WCVaR constraints. The objective function of the day-ahead optimization model is expressed as:

[0071] (1)

[0072] in, express Electricity purchase cost during different time periods This indicates that all gas-fired boilers are in Total fuel cost over a period of time Indicating shared energy storage systems Cost of charging and discharging during a specific period (in yuan); Value at risk under worst-case conditions is used to quantify the risk cost (in yuan) in extreme scenarios. The risk aversion coefficient (dimensionless) indicates that the larger the value, the more conservative the decision-making. This indicates the scheduling period of the day-ahead optimization model. In the day-ahead optimization model constructed in this embodiment, the scheduling period is defined as an hourly interval and a 24-hour day.

[0073] (2)

[0074] in, express Electricity price during the specified time period express Electricity purchase capacity during a given time period.

[0075] (3)

[0076] in, Indicates the first The gas-fired boilers in each DCP Gas consumption during a given time period express Gas and electricity prices during specific time periods This represents the total number of DCPs.

[0077] (4)

[0078] in, This represents the loss cost coefficient (yuan / kWh) of a shared energy storage system. , These respectively represent the shared energy storage system in Charging power and discharging power (kW) during the time period.

[0079] In the current optimization model, the decision variables include , , , .

[0080] In the specific implementation process, the decision variables in the current optimization model will be... , The data is transmitted to the real-time optimization model. It should be noted that because the day-ahead optimization model operates on an hourly time interval, the output of the day-ahead optimization model... , These are also parameters for each hour.

[0081] The constraints of the current optimization model include electrical energy balance constraints, thermal energy balance constraints, SES operational constraints, and WCVaR risk constraints, which are explained in detail below.

[0082] 1) Power balance constraints

[0083] (5)

[0084] in, Indicating shared energy storage systems Net discharge power (kW) during the period. ; Indicates the first RES in a DCP The power output of wind and solar power during a given time period (kW), i.e., the power generation capacity of renewable energy sources; Indicates the first The IT server in the DCP Power consumption (kW) during the time period; Indicates the first The cooling system in each DCP Power (kW) during the time period; Indicates the first Applications in DCP Load power (kW) for a given time period.

[0085] 2) Thermal energy balance constraint

[0086] (6)

[0087] in, Indicates the first The gas-fired boilers in each DCP Heat generated during the period (kWh); Indicates the first The cooling system in each DCP Waste heat recovery amount (kWh) during the period; Indicates the first The corresponding building in each DCP is Hot demand during certain periods.

[0088] (7)

[0089] in, This indicates the thermal efficiency of a gas-fired boiler.

[0090] 3) Operational constraints of shared energy storage systems

[0091] (8)

[0092] in, , For shared energy storage systems State variables of charging and discharging during a given period; For the rated capacity of SES, , These are charging efficiency and discharging efficiency, respectively. For shared energy storage systems State of charge during a given time period.

[0093] The state of charge of energy storage must meet the upper and lower limits of capacity and the minimum capacity requirement at the end of the cycle, as shown below:

[0094] (9)

[0095] in, , These represent the minimum and maximum values ​​of the charge capacity of the shared energy storage system, respectively. This indicates the state of charge of a shared energy storage system at the end of a cycle. This usually refers to the state of charge of the battery at the end of a specific time period (such as a day, a week, etc.). This indicates the minimum capacity requirement for a shared energy storage system at the end of the cycle.

[0096] 4) WCVaR Risk Constraints

[0097] Risk cost is defined using quantile optimization, with the following constraints:

[0098] (10)

[0099] Where F represents the set of probabilistic uncertainties in the output and load of wind and solar power, which includes all possible probabilistic distributions of wind and solar power output and load; It is an element in set F, that is, a specific probability distribution in the set, used to consider uncertainty in the calculation of risk constraints; Confidence Value at risk (in yuan); Represents the loss function; Indicates loss Greater than or equal to Under these conditions, loss The maximum expected value.

[0100] WCVaR risk constraints define risk cost through quantile optimization. Specifically, it considers the risk cost at a given confidence level. Below, the losses exceed The average loss under extreme conditions is considered. This method provides a more comprehensive risk assessment because it considers not only the maximum possible loss (VaR) but also the average loss under extreme conditions. By incorporating the WCVaR risk constraint into the objective function, the system's performance is ensured to be optimal while taking risk costs into account. As an additional constraint, it requires the system to meet certain risk levels during operation, thereby ensuring stable operation under different scenarios. This constraint, along with other constraints such as power balance, thermal balance, and energy storage operation, works in conjunction with the system's optimization process.

[0101] For the energy efficiency model of a hybrid cooling system, the hybrid cooling system in Cooling power consumption during the period Represented as:

[0102] (11)

[0103] in, , They represent Dynamic energy efficiency ratio and heat dissipation of IT equipment during different time periods; , They represent The chilled water temperature and ambient temperature during the period , These represent the efficiency coefficients of the air-cooled and liquid-cooled subsystems, respectively. Indicates the reference temperature.

[0104] (2) Real-time optimization model

[0105] In the real-time optimization model, a collaborative decision-making framework based on Stackelberg game theory is constructed: the leader, the data center cluster operator (DCCO), dynamically adjusts the charging and discharging price strategy of the shared energy storage system (SES) to maximize its own revenue while incentivizing prosumers (DCPs) to participate in energy sharing; while the followers, the DCPs, based on the SES price signal, synchronously optimize the interactive demand response (IDR) strategy and the peer-to-peer (P2P) energy trading plan to minimize real-time operating costs.

[0106] Based on this, this embodiment designs a dynamic pricing mechanism that links supply and demand—it dynamically adjusts P2P transaction prices based on the electricity supply-demand ratio (ESDR) and the heat supply-demand ratio (HSDR). When energy supply exceeds demand, prices are lowered to stimulate demand; conversely, prices are raised to suppress excessive trading, and joint bidding matching of electricity and waste heat is supported to promote the cross-category collaborative utilization of electricity and heat energy.

[0107] The model leverages a three-tiered refined IDR mechanism to enhance load flexibility: on the data side, it utilizes the spatiotemporal elasticity of server workloads to migrate non-real-time computing tasks to off-peak electricity periods or nearby data centers; on the heating side, it intelligently adjusts the operating time of the heating system by utilizing building thermal inertia and user thermal comfort tolerance range; and on the electricity consumption side, it optimizes differentiated load patterns based on a tiered response strategy for rigid loads, transferable loads, and adjustable loads (air conditioning), thereby systematically improving the energy supply and demand balance capability.

[0108] The real-time optimization model in this embodiment is implemented based on the Stackelberg game framework, consisting of an upper-level model for DCCO decision-making and a lower-level model for each DCP to make decisions separately.

[0109] 1) Upper-level model (DCCO decision)

[0110] For example, the upper-level model constructs an objective function with a 15-minute time interval and the goal of maximizing DCCO revenue, expressed as:

[0111] (12)

[0112] in, Indicating shared energy storage systems Trading prices during the period Indicating shared energy storage systems Net discharge power during the period; This represents the scheduling period of the real-time optimization model. The scheduling period is an integer multiple of 15 minutes. For example, a scheduling period of 4 hours can be set, and a scheduling period includes 16 time periods.

[0113] In the specific implementation process, the charging and discharging schedule of the shared energy storage system will be planned on an hourly basis. , The data is transmitted to the upper-level model within the real-time optimization model for solving the problems for each hour. , Based on this, the upper-level model is solved to obtain the results at 15-minute intervals. .

[0114] The upper-level model satisfies the IDR flexibility constraint. The IDR flexibility constraint is a crucial restriction in the objective function optimization process. It ensures the flexibility and stability of power dispatch schemes, thus contributing to the optimization of the objective function. In the objective function, whether maximizing DCCO revenue or minimizing the real-time cost of DCP, the impact of the IDR flexibility constraint must be considered. This helps balance the stability and economy of the power system during optimization. By limiting the range of power dispatch schemes, the IDR flexibility constraint ensures that the schemes meet the stability and reliability requirements of the power system, thereby contributing to the maximization of overall benefits.

[0115] 2) Lower-level model (DCPs decision-making)

[0116] Each DCP makes its own decision, constructing an objective function with the goal of minimizing the real-time cost of the current DCP; the... The objective function of a DCP is expressed as:

[0117] (13)

[0118] in, Indicates the first A DCP and a shared energy storage system in Trading power (kW) during the time period; Indicates the first A DCP at Transaction power (kW) in the P2P market during a given period; This indicates that the P2P market is in Trading prices during a specific time period; Indicates the first A DCP in Adjustment costs (in yuan) for different time periods, such as task delay penalties or thermal comfort compensation.

[0119] The upper-level model and the lower-level model satisfy the following:

[0120] (14)

[0121] The dynamic P2P pricing mechanism satisfies:

[0122] (15)

[0123] In the formula, This represents the price elasticity coefficient (dimensionless). express Electricity supply and demand ratio during a given period , Indicates the first A DCP in Load power during the time period; express The heat supply and demand ratio during a given period .

[0124] The decision variables of the lower-level model include , and It is used for real-time scheduling of the corresponding DCP.

[0125] The lower-level model satisfies P2P transaction balance constraints, data load latency constraints, and thermal comfort constraints.

[0126] 1) P2P transaction balance constraints

[0127] (16)

[0128] To achieve a balance in the total amount of electricity trading within the cluster, external intervention must be avoided.

[0129] 2) Data load delay constraints

[0130] The spatiotemporal transferability of data load is reflected through task delay constraints, which are as follows:

[0131] (17)

[0132] in, Indicates task exist Execution status of the time period A value of 1 indicates that the program runs, and a value of 0 indicates that it does not run. This indicates the task start time; the task will not begin execution before this time. The maximum allowable delay time is defined within which tasks can be delayed without violating constraints. Represents a set of tasks.

[0133] The purpose of this constraint is to ensure that tasks are completed within a specified time and to avoid performance degradation and reduced service quality due to delays.

[0134] 3) Thermal comfort constraints

[0135] (18)

[0136] In the formula: In order to be in The predicted average vote value for the time period (quantifying human thermal sensation); In order to be in Percentage of unsatisfactory expectations for a given period.

[0137] (19)

[0138] in, , For the task Data load latency penalty coefficient; , The penalty coefficient for violating thermal comfort standards; This refers to the actual execution time of the task.

[0139] The flowchart of the optimization algorithm is as follows Figure 5 As shown, the specific explanation is as follows: The above optimization process unfolds with dynamic game theory at its core: First, the Data Center Operator (DCCO) calculates and publishes a global price signal that includes P2P transaction prices and demand-side response (IDR) incentives based on the real-time status of the shared energy storage system (SES) and the energy application demands of the power supply equipment (DCPs). Subsequently, each DCP optimizes its local strategy based on the price signal, generating a scheduling scheme that balances cost and constraints by adjusting server load, starting and stopping gas boilers, and other IDR methods, combined with P2P transaction volume decisions. Finally, through multiple rounds of price-demand interaction iteration, the two sides of the game (DCCO and DCPs) reach an equilibrium state—at which point the DCCO cannot increase revenue through unilateral price adjustments, and the DCPs cannot further reduce costs through strategy adjustments, forming a stable Pareto optimal solution.

[0140] The core advantages of the above model are reflected in three aspects: First, in terms of two-stage complementarity, robust optimization and real-time dynamic game theory form a collaborative framework of "risk pre-control - flexible adjustment." The former avoids risks in extreme scenarios through WCVaR constraints, while the latter utilizes Stackelberg game theory to explore the flexibility potential of distributed resources. Second, in terms of multi-dimensional flexibility integration, the IDR mechanism achieves cross-temporal and spatial coordination of electricity, heat, and computing power through data load spatiotemporal migration, thermal comfort constraint optimization, and multi-type load hierarchical management. Meanwhile, the P2P trading mechanism opens up cross-entity flow channels for surplus electricity / heat, constructing a multi-energy complementary elastic network. Third, in terms of incentive-compatible design, the interest coordination mechanism based on master-slave game theory ensures both the global optimization authority of the DCCO as the system coordinator and grants DCPs autonomous decision-making power based on price signals, avoiding efficiency losses through strategy equilibrium.

[0141] This solution, through a triple collaborative mechanism of "risk-game-market," not only improves the economics of data center clusters but also effectively mitigates the dual impacts of renewable energy volatility and uneven spatial and temporal computing load, providing a closed-loop solution for low-carbon operation of data centers with a high proportion of renewable energy penetration.

[0142] (1) Call CPLEX / Gurobi to solve the day-ahead optimization model

[0143] That is, in the day-ahead phase, robust optimization and MILP hybrid solutions are used. Specifically, day-ahead optimization aims to minimize total operating costs (electricity purchase cost, fuel cost, wind curtailment penalty, etc.) and introduces WCVaR constraints to control risks under extreme scenarios. The uncertainty set mainly models wind and solar power output prediction errors and load fluctuations as interval uncertainty sets:

[0144] (20)

[0145] In the formula, Let represent the predicted values ​​of wind power generation, photovoltaic power generation, power system load, and surface heat system load in the uncertainty set, respectively. , These represent the upper and lower bounds of the power system load, respectively. , These represent the upper and lower limits of the load on the thermal system, respectively.

[0146] Robust equivalence model transformation is employed using the Column and Constraint Generation (C&CG) algorithm to decompose the problem into: a Master Problem, which determines the scheduling baseline for energy storage (SES) and hydrogen energy systems (H2); and Subproblems. The worst-case uncertainty scenario is searched to verify the robustness of the solution. The Mixed Integer Linear Programming (MILP) model is solved using CPLEX / Gurobi, with the specific solution steps as follows:

[0147] 1) Initialize the main problem and set initial feasible solutions (such as SES charge / discharge baseline and hydrogen production plan).

[0148] 2) Solve the subproblems to verify the feasibility of the current solution within the uncertain set. If there are scenarios that violate the constraints, generate a Benders cut and add it to the main problem.

[0149] 3) Iteratively update the main problem and subproblems until the difference between the objective functions of the main problem and subproblems is less than the threshold.

[0150] (2) Solve the real-time optimization model using distributed ADMM.

[0151] Upper Layer (DCCO): Aiming to maximize profits, it updates the transaction price of the shared energy storage system using a gradient descent method. And broadcast it to all DCPs.

[0152] (twenty one)

[0153] in, Indicates the first The transaction price of the shared energy storage system in the next iteration; Indicates the learning rate. The revenue function of DCCO is represented by .

[0154] The lower layer (DCPs) aims to minimize energy costs by adjusting the distribution of computational loads and energy demand, and solves the problem using the Distributed ADMM algorithm.

[0155] The advantages of the model constructed in this embodiment are as follows:

[0156] (1) Two-stage complementarity optimization

[0157] Recently, by introducing WCVaR constraints during the robust optimization phase, risks in extreme scenarios were effectively avoided, ensuring the stability of the system under uncertain conditions. This optimization strategy provides a solid foundation for subsequent real-time dynamic game theory, achieving "risk pre-control."

[0158] The real-time dynamic game phase utilizes the Stackelberg game model to explore the flexibility potential of distributed resources, enabling flexible system adjustment. Through multiple rounds of price-demand interaction iterations, the two sides in the game reach an equilibrium state, forming a stable Pareto optimal solution.

[0159] (2) Multidimensional flexible integration:

[0160] The IDR mechanism achieves cross-temporal and spatial coordination of electrical energy, thermal energy, and computing power through data load spatiotemporal migration, thermal comfort constraint optimization, and multi-type load hierarchical management. This multi-dimensional and flexible integration improves the overall efficiency and response speed of the system.

[0161] The P2P trading mechanism opens up cross-entity channels for the flow of surplus electricity / heat, constructing a resilient network of multi-energy complementarity. This mechanism promotes the optimal allocation and efficient utilization of resources, further enhancing the system's economic efficiency.

[0162] (3) Incentive-compatible design

[0163] Based on a master-slave game-theoretic mechanism for coordinating interests, this design safeguards the overall optimization authority of the DCCO as the system coordinator while granting DCPs autonomous decision-making power based on price signals. This approach avoids efficiency losses and achieves a win-win situation for both sides in the game.

[0164] Through strategy equilibrium, it is ensured that when the system reaches equilibrium, DCCO cannot increase revenue through unilateral price adjustments, and DCPs cannot further reduce costs through strategy adjustments, thus forming a stable Pareto optimal solution.

[0165] In summary, this embodiment proposes a two-stage energy sharing model for data center clusters. By integrating Integrated Demand Response (IDR), Shared Energy Storage (SES), and a P2P trading mechanism, it achieves efficient energy sharing and collaborative optimization. In terms of technical feasibility, the joint scheduling of the hydrogen thermal device and SES effectively mitigates the intermittency of wind and solar power output, improving system flexibility. Regarding economic advantages, the dynamic pricing mechanism combined with multi-energy flow optimization significantly reduces operating costs and promotes mutual benefit between operators and producers / consumers. In terms of extended value, this model can be extended to integrated energy scenarios such as industrial parks and transportation hubs, supporting cross-sectoral low-carbon transformation. Simulation results show that the model reduces data center operating costs while balancing user satisfaction and renewable energy absorption rates. Future research will further deepen multi-energy flow coupling dynamics modeling, server-level real-time control algorithms, and data privacy protection technologies to provide universal solutions for high-energy-consuming scenarios.

[0166] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0167] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A robust optimization method for low-carbon economy in data center clusters considering multiple types of loads, characterized in that, The method includes: A day-ahead optimization model and a real-time optimization model for a data center cluster considering multiple types of load are constructed. The objective function of the day-ahead optimization model is constructed with the first time interval as the time interval and the goal of minimizing the total day-ahead operating cost of the data center cluster. The real-time optimization model is constructed with the second time interval as the time interval, including an upper-level model for DCCO decision-making of the data center cluster operator and a lower-level model for DCP decision-making of each distributed production and consumption unit. The first time interval is longer than the second time interval. 1) The objective function of the current optimization model is expressed as: (1) in, express Electricity purchase cost during different time periods This indicates that all gas-fired boilers are in Total fuel cost over a period of time Indicating shared energy storage systems The cost of charging and discharging during specific time periods; Value at risk under worst-case conditions Risk aversion coefficient; This indicates the scheduling cycle of the current optimization model; (2) in, This represents the loss cost coefficient of a shared energy storage system. , These respectively represent the shared energy storage system in Charging power and discharging power during a given time period; The decision variables of the current optimization model include and The constraints of the current optimized model include WCVaR risk constraints. 2) The objective function of the upper-level model, which aims to maximize the DCCO revenue of data center cluster operators, is expressed as: (3) in, Indicating shared energy storage systems Trading prices during the period Indicating shared energy storage systems Net discharge power during the period; This indicates the scheduling cycle of the real-time optimization model; The decision variables of the upper-level model include ; 3) Each distributed production and consumption unit (DCP) makes independent decisions, and the objective function of the lower-level model for each DCP decision is constructed with the goal of minimizing the real-time cost of the current DCP; The objective function of a distributed producer-consumption unit (DCP) is expressed as: (4) in, Indicates the first Distributed Productivity Units (DCPs) and Shared Energy Storage Systems Trading power during a given time period; Indicates the first A distributed producer-consumption unit (DCP) at The transaction power in the peer-to-peer (P2P) market during a given period; This indicates that the peer-to-peer (P2P) market is in Trading prices during a specific time period; Indicates the first A distributed producer-consumption unit (DCP) in Adjustment costs during specific time periods; The decision variables of the lower-level model include , and ; Solving the day-ahead optimization model, we calculate the charging and discharging schedule of the shared energy storage system with the first time interval as the time interval. , ; The charging and discharging schedule of the shared energy storage system will be based on the first time interval. , The data is transmitted to the real-time optimization model, where the upper-level model is solved to obtain the values ​​at the second time interval. ; Based on the lower-level model and Solve the lower-level model to obtain the decision variables of the lower-level model, and perform real-time scheduling of the corresponding distributed production and consumption unit (DCP). This represents the total number of Distributed Producer-Consumer Units (DCPs).

2. The low-carbon, economical, robust optimization method for data center clusters considering multiple types of loads as described in claim 1, characterized in that, (5) in, express Electricity price during the specified time period express Electricity purchase volume during the time period; (6) in, Indicates the first Gas-fired boilers in a distributed production and consumption unit (DCP) Gas consumption during a given time period express Gas and electricity prices during certain time periods.

3. The robust optimization method for low-carbon economy of data center clusters considering multiple types of loads as described in claim 2, wherein the decision variables of the current-day optimization model include: and .

4. The robust optimization method for low-carbon economy of data center clusters considering multiple types of loads as described in claim 3, the constraints of the day-ahead optimization model include power balance constraints, thermal balance constraints, and shared energy storage system (SES) operation constraints.

5. The low-carbon economic robust optimization method for data center clusters considering multiple load types as described in claim 4, wherein the power balance constraint is expressed as: (7) in, Indicating shared energy storage systems Net discharge power during the period ; Indicates the first Renewable energy systems (RES) in distributed generation and consumption units (DCPs) The power output of wind and solar power during a given period, i.e., the power generation capacity of renewable energy sources; Indicates the first The IT server in a Distributed Producer-Consumer Unit (DCP) Power consumption during a given period; Indicates the first The cooling system in a distributed production and consumption unit (DCP) Power during a given time period; Indicates the first Application in Distributed Production and Consumption Units (DCPs) Load power during a given time period.

6. The low-carbon economic robust optimization method for data center clusters considering multiple types of loads as described in claim 5, wherein the second time is 15 minutes and the first time is 1 hour.

7. The low-carbon, economical, robust optimization method for data center clusters considering multiple types of loads as described in claim 6, characterized in that, The upper-level model satisfies the comprehensive demand response IDR flexibility constraint.

8. The low-carbon, economical, robust optimization method for data center clusters considering multiple types of loads as described in claim 7, characterized in that, The lower-level model satisfies the peer-to-peer (P2P) transaction balance constraint, data load delay constraint, and thermal comfort constraint.

Images