Distributed resource planning method for data center integrated energy system in multi-element market environment

Abstract

This invention discloses a distributed resource planning method for a data center integrated energy system under a diversified market environment, belonging to the field of power system planning. The method constructs a two-layer optimization framework. The upper layer uses a multi-objective robust Bayesian optimization algorithm under input noise to optimize distributed resource capacity. The lower layer establishes a simulation model of the data center integrated energy system operation based on the diversified market benefits of energy storage. It incorporates Gaussian distributed noise to model uncertainty factors, trains a robust Gaussian process surrogate model, and selects sampling points for iterative optimization through a robust expectation hypervolume improvement function. Finally, it outputs a robust optimal configuration scheme through Pareto sorting. This invention solves the problems of poor robustness, imbalance in multi-objective optimization, single energy storage value assessment, and algorithm inefficiency in existing methods, achieving stable configuration schemes, balanced objectives, and accurate benefits, thereby improving system economy and renewable energy absorption rate.

Description

Technical Field

[0001] This invention belongs to the field of power system planning, specifically relating to a distributed resource planning method for a data center integrated energy system in a diversified market environment. Background Technology

[0002] With the deep integration of data centers and new energy sources, the operation of integrated energy systems for data centers exhibits significant coupling and uncertainty. On the one hand, data center loads have high power density and adjustability, and their operating strategies have a significant impact on the load balance and new energy absorption capacity of integrated energy systems. On the other hand, the randomness and volatility of renewable energy sources such as wind and solar power, sunlight, temperature, internal heat sources in data centers, and the internal heat dissipation coefficients of information technology equipment lead to a large amount of input uncertainty in system operation. Distributed resource allocation schemes directly affect the economics and stability of integrated energy systems for data centers.

[0003] Traditional distributed resource optimization not only fails to fully consider the impact of input disturbances, easily leading to configuration schemes that cannot meet system requirements or result in excessive redundancy and resource waste during actual operation, but also often neglects the energy storage benefit calculation mechanism in a diversified market environment, easily leading to low economic efficiency in energy storage planning results. Furthermore, the distributed resource allocation and scheduling of integrated energy systems in data centers exhibit hierarchical dependencies. Without establishing a reasonable upper and lower-level optimization framework, the overall system operating cost is often too high, or the renewable energy absorption rate decreases.

[0004] The existing technology has the following problems:

[0005] 1) Ignoring the impact of input uncertainty: Existing distributed resource planning mostly adopts deterministic or scenario enumeration methods, without systematically modeling new energy output, sunlight, outdoor temperature, etc., resulting in poor robustness of configuration schemes in actual operation and sensitivity to prediction bias.

[0006] 2) Single multi-objective optimization method: Traditional multi-objective optimization methods usually transform multiple objectives into a single objective by weighted summation. This method relies too much on subjective weight setting and is difficult to truly reflect the trade-off relationship between the system and the objectives.

[0007] 3) Lack of multi-market environment modeling and singular energy storage value assessment: Existing research mostly focuses on energy storage participation in a single market, such as the electricity market, and lacks synergistic optimization models in multi-market coupled environments such as electricity, capacity, and primary frequency regulation. This makes it impossible to accurately assess the comprehensive benefits and contributions of energy storage in different markets, and also fails to solve the problem of refined capacity allocation of energy storage.

[0008] 4) Limited robustness and convergence of the algorithm: Existing heuristic algorithms are inefficient in solving problems under high-dimensional uncertain conditions, are prone to getting trapped in local optima, and cannot provide an adaptive processing mechanism for input noise. Summary of the Invention

[0009] Based on the above shortcomings, this invention provides a distributed resource planning method for integrated energy systems in data centers under diverse market environments. This method solves the problem that existing methods do not fully consider the uncertainties of inputs such as new energy output and temperature, resulting in poor robustness. At the same time, it solves the problems that traditional multi-objective optimization relies on subjective weights, making it difficult to balance economic efficiency and new energy absorption rate, as well as the problems that existing algorithms have low efficiency and are prone to getting trapped in local optima under high-dimensional uncertain conditions.

[0010] The technical solution adopted in this invention is as follows: A distributed resource planning method for a data center integrated energy system in a diversified market environment, comprising the following steps:

[0011] S1. Construct a two-layer optimization framework for distributed resources in a data center integrated energy system: establish an upper-layer capacity configuration model and a lower-layer operation simulation model. The upper-layer decision variable is the capacity of various distributed resources, and the lower-layer model performs operation simulation based on a given capacity scheme. The two are coupled and iterated through configuration scheme and objective function value feedback.

[0012] S2. Algorithm Dataset Initialization: Determine the upper-level decision variables, namely the capacity of various distributed resources, generate an initial capacity configuration sample set by random or quasi-random sampling, and call the lower-level model to evaluate the multi-objective response value of each sample to construct the initial observation dataset.

[0013] S3. Constructing a probabilistic surrogate model considering noise input: Modeling uncertainties as Gaussian noise, including: new energy output, temperature, illumination, and heat source; training a Gaussian process model based on the initial sample set to predict the statistical distribution of the objective function under noise;

[0014] S4. Optimal point selection based on maximizing the hypervolume function: The robust expected hypervolume improvement function is used as the acquisition function. The next most promising capacity configuration sample is selected by maximizing this function. The selected sample is passed to the lower-level model for simulation to obtain the target value. The sample set and surrogate model are updated and repeated until convergence.

[0015] S5. Establish a lower-level operation simulation optimization model: In the lower-level model, based on the capacity configuration given by the upper level, the energy storage revenue of the power market, capacity market, and primary frequency regulation market is comprehensively considered to optimize the equipment output, load migration, and cooling power at each moment, and to establish a complete operation constraint system that includes system power balance, energy storage constraints, and temperature constraints.

[0016] S6. Algorithm Convergence Judgment and Optimal Solution Output: The convergence of the algorithm is judged based on the threshold of the robust expected hypervolume improvement function or the maximum number of iterations. The final sample set is Pareto sorted, and the set of optimal capacity configuration schemes that are robust to input noise is output.

[0017] Furthermore, in step S1, the upper-layer multi-objective robust Bayesian optimization algorithm under input noise and the lower-layer distributed resource collaborative optimization model of the integrated energy system of the data center form an interactive closed loop. The specific process is as follows: the upper layer first initializes the dataset, constructs a Gaussian process surrogate model, selects the next sampling point through the robust expectation hypervolume improvement function, and passes the capacity configuration scheme of the distributed resources corresponding to the sampling point to the lower-layer model; after the lower-layer model constructs typical scenarios and solves them, it feeds back the operating cost and new energy absorption rate in the objective function to the upper-layer model. The upper-layer model updates the dataset according to the feedback objective function value and retrains the Gaussian process surrogate model; if the convergence condition is not met, the Gaussian process surrogate model continues to be updated; if the convergence condition is met, the robust Pareto optimal solution set, i.e., the optimal configuration scheme of distributed resources, is output.

[0018] Furthermore, in step S2, the hyperparameters required for the multi-objective robust Bayesian optimization algorithm are input, including: the initial number of sample points. Maximum number of iterations Convergence threshold The covariance matrix used to describe the distribution of input noise. and the number of scenarios used for uncertainty modeling The distributed resources mentioned above include: wind power systems, photovoltaic systems, battery energy storage systems, supercapacitor energy storage systems, and data center cooling systems, and their configuration variables are shown in formula (1):

[0019]

[0020] In the formula: , , These represent the rated power of the cooling systems for wind farms, photovoltaic power plants, and data centers, respectively. , These are the rated power and capacity of the first type of battery energy storage used to ensure power supply and participate in the spot market; , These are the rated power and capacity of the second type of battery energy storage, which prioritizes fulfilling capacity market obligations while also considering power supply security and spot trading. , These are the rated power and capacity of the first type of supercapacitor used to ensure power supply and participate in the spot market; , These are the rated power and capacity of the second type of supercapacitor energy storage specifically designed for the frequency regulation market; a power value greater than 0 indicates discharging, and a power value less than 0 indicates charging.

[0021] Distributed resource capacity variables Constrained by upper and lower limits, its expression is shown in formula (2):

[0022]

[0023] In the formula: , They represent The lower and upper limits of the values;

[0024] For the initial sample points The distributed resource collaborative optimization model of the integrated energy system of the data center is used to evaluate the corresponding multiple target response values. This forms a multi-target observation pair; all input samples and their corresponding target values ​​are combined to form the initial observation dataset. As shown in formula (3):

[0025]

[0026] In the formula: A vector containing multiple objective function values , where the objective function The cost of minimizing the integrated energy system of the data center is shown in formula (4);

[0027]

[0028] In the formula: Let $\mathbf{ ...

[0029]

[0030] In the formula: It refers to the number of scenes, i.e., the number of samples with different noise or random input; This indicates that it is in the scene Annual operating costs The scene number, with a value range of 1, 2, ... In multi-objective optimization problems, As an expected value, it is often used in multiple objective functions to represent the statistical processing of scenario uncertainty;

[0031] It is the sum of the annual investment costs of various distributed resources, determined by upper-level decision variables. The distributed resource configuration scheme represented directly determines this, as shown in formula (6):

[0032]

[0033] In the formula: , , , , , , These represent the annual investment costs for wind power, photovoltaic power, the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, the second type of supercapacitor energy storage, and data center cooling systems, respectively.

[0034] Investment costs of wind power, solar power and data center cooling systems , , As shown in formula (7):

[0035]

[0036] In the formula: Indicates the type of equipment; , , These represent wind power, photovoltaic, and data center cooling systems, respectively. , , These represent the power cost coefficients for wind power, photovoltaic, and data center cooling systems, respectively. Represents the discount rate; , , These represent the lifespan of wind power, photovoltaic, and data center cooling systems, respectively.

[0037] Investment costs of two types of battery energy storage and two types of supercapacitor energy storage , , , As shown in formula (8):

[0038]

[0039] In the formula: Indicates the type of energy storage device; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. , , , These represent the power cost coefficients for four different energy storage methods. , , , These represent the capacity cost coefficients for four different energy storage methods. , , , These represent the lifespan of the four types of energy storage;

[0040] objective function To maximize the renewable energy absorption rate, as shown in formula (9):

[0041]

[0042] In the formula: The average absorption rate across all scenarios is shown in formula (10):

[0043]

[0044] Scene New energy consumption rate As in the formula As shown:

[0045]

[0046] In the formula: The actual amount of renewable energy consumed is shown in formula (12):

[0047]

[0048] In the formula: , They represent The output of photovoltaic power plants and wind farms at all times; This indicates the number of scheduling cycles in a year; Indicates the duration of the scheduling interval;

[0049] The total electricity generated from renewable energy sources is represented as shown in formula (13):

[0050]

[0051] Furthermore, in step S3, a Gaussian process is used as a surrogate model to statistically model the objective function. Bayesian optimization is used to predict the mean and variance of the objective function. Under the condition that there is noise in the input, the GP model is extended to a robust form to predict the mean and variance of the objective function under the input perturbation distribution, as shown in formula (14):

[0052]

[0053] In the formula: It is a mean function. It is the covariance function;

[0054] The input noise originates from new energy output, illumination, temperature, internal heat sources of the data center, and internal heat dissipation coefficients of information technology equipment. The set of input noise variables is shown in formula (15):

[0055]

[0056] In the formula: This indicates the number of scheduling operations within a scheduling cycle. and They represent The power output of photovoltaic power plants and the actual power output of wind farms at all times; express outdoor temperature at all times; express Temperature of adjacent areas to the data center at any given time; Indicates the wall number; Indicates wall Light intensity in the corresponding direction, ; This indicates the light intensity in the corresponding direction of the window; Indicates the internal heat source of the data center; This indicates the heat dissipation coefficient inside data center information technology equipment;

[0057] The system's actual input is subject to random disturbances The influence of this influence forms the observed input. As shown in formula (16):

[0058]

[0059] noise It follows a Gaussian distribution, as shown in formula (17):

[0060]

[0061] In the formula: The covariance matrix represents the noise and is used to characterize the statistical properties of the input disturbance.

[0062] Furthermore, in step S4, during the Bayesian optimization process, the acquisition function is used to evaluate the quality of the current sampling point and guide the selection of the next sampling point; a robust expectation hypervolume improvement function is adopted. As the acquisition function, its expression is shown in formula (18):

[0063]

[0064] In the formula: This is the current non-dominated solution set; The hypervolume index measures the size of the Pareto front coverage area of ​​a solution set; a larger value indicates a higher quality solution set. To determine the noise distribution The expected value indicates Based on input Noise sampling;

[0065] Improved by maximizing robust expectation hypervolume To select the optimal next sampling point Its expression is shown in formula (19):

[0066] .

[0067] Furthermore, in step S5, when the capacity configuration scheme is determined at the upper level, the lower-level model performs periodic operational simulation optimization of the data center's integrated energy system operation process. The lower-level optimization variables include the wind power output at each moment. Photovoltaic power output The first type of battery energy storage output The second type of battery energy storage output The first type of supercapacitor energy storage output The second type of supercapacitor energy storage output Data center IT equipment migration power matrix Data center cooling capacity Interaction power between the integrated energy system of the data center and the power distribution network ;

[0068] The optimization objective of the lower-level model is to minimize the operating cost of the integrated energy system of the data center while satisfying energy balance and equipment constraints. For the scene The operating costs of each distributed resource and the penalty costs of curtailing wind and solar power, minus the revenue from energy storage participating in the electricity, capacity, and primary frequency regulation markets, are expressed as formula (20):

[0069]

[0070] In the formula: , , , , These represent the operating costs of wind power, photovoltaic power, storage batteries, supercapacitors, and data center cooling systems, respectively. This indicates the penalty cost for abandoning wind and solar power. , , These represent the revenue from the electricity market, the revenue from the capacity market, and the revenue from the primary frequency regulation market for energy storage, respectively.

[0071] Operating costs of wind power As shown in formula (21):

[0072]

[0073] In the formula: This represents the fixed operation and maintenance cost coefficient for wind power.

[0074] Operating costs of photovoltaics As shown in formula (22):

[0075]

[0076] In the formula: This represents the fixed operation and maintenance cost coefficient for photovoltaic systems.

[0077] Operating costs of battery energy storage As shown in formula (23):

[0078]

[0079] In the formula: This represents the fixed operating and maintenance cost coefficient of the battery.

[0080] Operating costs of supercapacitor energy storage As shown in formula (24):

[0081]

[0082] In the formula: This represents the fixed operation and maintenance cost coefficient of a supercapacitor.

[0083] Data center operating costs As shown in formula (25):

[0084]

[0085] In the formula: express Electricity price at any time; express Energy consumption of data center information technology equipment at all times; express Energy consumption of the cooling system in a real-time data center; express Energy consumption of other equipment in the data center at any given time;

[0086] The punitive costs of abandoning wind and solar power As shown in formula (26):

[0087]

[0088] In the formula: This represents the penalty cost coefficient;

[0089] Electricity market revenue As shown in formula (27):

[0090]

[0091] In the formula: Energy storage adjusts the interaction power between the integrated energy system of a data center and the power distribution network through charging and discharging, in order to achieve arbitrage;

[0092] Energy storage capacity market revenue As shown in formula (28):

[0093]

[0094] In the formula: This indicates the market settlement price for capacity. The committed capacity for battery energy storage to participate in the capacity market is shown in formula (29):

[0095]

[0096] In the formula: It is the effective capacity factor. The number of hours of continuous discharge as specified by the market;

[0097] primary frequency regulation market revenue of energy storage As shown in formula (30):

[0098]

[0099] In the formula: This indicates the market clearing price for frequency modulation capacity; This indicates the unit price for mileage adjustment compensation. This is the frequency modulation mileage factor for the supercapacitor; The performance coefficient of a supercapacitor; It is the frequency regulation mileage of the second type of supercapacitor energy storage in the primary frequency regulation capacity market, and its constraint is shown in formula (38);

[0100] The constraints include:

[0101] 1) The power balance constraint of the integrated energy system of the data center is shown in formula (31):

[0102]

[0103] 2) Renewable energy output constraints, as shown in formulas (32) and (33):

[0104]

[0105]

[0106] 3) Energy storage-related constraints

[0107] The output power of energy storage is limited by the power of the energy storage configuration, as shown in formulas (34) and (35):

[0108]

[0109]

[0110] The state of charge at the beginning and end of the scheduling cycle must be equal, as shown in formula (36):

[0111]

[0112] The method for calculating the state of charge is shown in formula (37):

[0113]

[0114] In the formula: It refers to the types of energy storage devices; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. This indicates the self-discharge rate of the battery's energy storage. It refers to the charging and discharging efficiency of energy storage;

[0115] The second type of supercapacitor energy storage has a frequency regulation mileage in the primary frequency regulation capacity market, and its upper and lower limits are as shown in formula (38):

[0116]

[0117] In the formula: This is the frequency modulation capacity proportional coefficient;

[0118] 4) Data center constraints

[0119] 4.1) Power constraints for information technology equipment

[0120] This is a power transfer matrix with 100 rows and 100 columns. Among them, the most migrateable loads can be migrated. Dispatch intervals. Load of information technology equipment in the data center. The expression is shown in formula (39):

[0121]

[0122] In the formula: This indicates the real-time load that is not migrated from the data center; The adjusted data center delay load is expressed as shown in formula (40):

[0123]

[0124] Elements in the power transfer matrix This represents the load migration ratio, and its range is shown in formula (41):

[0125]

[0126] Since the total offline load remains unchanged before and after the migration, the migration must be completed before the deadline. Follow the row and constraint rules, as shown in formula (42):

[0127]

[0128] The selected offline workload can only be delayed and cannot be processed in advance. The maximum delay time for the data center's offline workload is... The scheduling interval is shown in formula (43):

[0129]

[0130] After adjustment The power of the data center in processing network load at any given moment is greater than the real-time load power at that moment, but less than the power of the data center in processing network load when it is running at full load, as shown in equation (44):

[0131]

[0132] In the formula: This indicates the power consumed by the data center when it is operating at full load to handle network load.

[0133] 4.2) Cooling power constraint:

[0134] Refrigeration system power Constrained by upper and lower limits, as shown in formula (45):

[0135]

[0136] In the formula: This is the upper limit of the operating power of the refrigeration system; the operating power of the refrigeration system The relationship between cooling capacity and refrigeration capacity is shown in formula (46):

[0137]

[0138] In the formula: The energy efficiency ratio of the air conditioning refrigeration system; for The cooling capacity of the air conditioning system at all times;

[0139] The dynamic model of the thermal inertia of the data center envelope includes one indoor air node and four wall nodes. The dynamic model of the thermal inertia of the data center envelope is established as shown in formulas (47)-(51):

[0140]

[0141]

[0142]

[0143]

[0144]

[0145] In the formula: express Time Wall temperature; Indicates wall The heat capacity; Indicates wall Thermal resistance; Indicates indoor temperature; For walls The heat absorption rate; For walls The area; The area of ​​the window; The heat capacity of the room; For window thermal resistance; The refractive index of the window;

[0146] To ensure the server operates normally, the indoor temperature... The upper and lower limits need to be met, as shown in formula (52):

[0147]

[0148] In the formula: , These represent the lower and upper limits of the indoor temperature set to ensure the normal operation of the server;

[0149] 5) Power purchase constraints of the distribution network

[0150] Power purchased by the distribution network Constrained by upper and lower limits, as shown in formula (53):

[0151]

[0152] In the formula: This represents the maximum value of electricity purchased by the distribution network.

[0153] Furthermore, in step S6, the next sampling point is obtained. Then, the point is evaluated in the model to obtain its multi-objective response results. , to new sample data Add to algorithm dataset Retrain the Gaussian process model, update the mean and variance estimates, and repeat steps S3 to S6 until the algorithm converges, i.e., the change in the robust expected hypervolume improvement function is less than the threshold. The algorithm may reach its maximum number of iterations. After convergence, Pareto ranking is performed using the multi-objective response results of all samples to obtain the robust optimal solution set under input noise conditions, which represents the optimal allocation scheme for distributed resources. .

[0154] This invention also provides a distributed resource planning system for a data center integrated energy system under a diversified market environment, used to implement the distributed resource planning method for a data center integrated energy system under a diversified market environment as described above. It includes a two-layer optimization framework module, a dataset initialization module, a probabilistic surrogate model construction module, a sampling point selection module, a lower-layer simulation optimization module, and a convergence determination and optimal solution output module. These modules work together to achieve robust optimized configuration of distributed resources, as detailed below:

[0155] The two-layer optimization framework module is used to construct a two-layer optimization framework for distributed resources in a data center integrated energy system. It includes an upper-layer capacity configuration unit and a lower-layer operation simulation unit. The upper-layer capacity configuration unit uses the capacity of various distributed resources as decision variables, while the lower-layer operation simulation unit performs operation simulations based on the capacity scheme given by the upper layer. The two are coupled and iteratively coupled through the feedback of the configuration scheme and the objective function value. Among them, the upper-layer capacity configuration unit integrates a multi-objective robust Bayesian optimization algorithm under input noise, and the lower-layer operation simulation unit integrates a collaborative optimization model for distributed resources in a data center integrated energy system, forming an interactive closed loop.

[0156] Dataset initialization module: used to determine the upper-level decision variables, namely the capacity of various distributed resources, generate an initial capacity configuration sample set by random or quasi-random sampling, call the lower-level simulation unit to evaluate the multi-objective response value of each sample, and construct the initial observation dataset; at the same time, input the hyperparameters required by the multi-objective robust Bayesian optimization algorithm, including the number of initial sample points, the maximum number of iterations, the convergence threshold, and the covariance matrix used to describe the distribution of input noise;

[0157] The probabilistic proxy model building module is used to model uncertainties as Gaussian noise, including new energy output, temperature, illumination, internal heat sources in data centers, and internal heat dissipation coefficients of information technology equipment; it trains a Gaussian process model based on the initial observation dataset, the Gaussian process model being robust, to predict the statistical distribution of the objective function under noise, and outputs the mean and variance of the objective function;

[0158] Sampling point selection module: It is used to select the next most promising capacity configuration sample by using the robust expected hypervolume improvement function as the acquisition function, and select the next capacity configuration sample by maximizing the function; the selected sample is passed to the lower-level running simulation unit, the target value is received from the lower-level feedback, the sample set and Gaussian process model are updated, and the above process is repeated until the convergence criterion is met.

[0159] The lower-level operation simulation optimization module is used to optimize equipment output, load migration, and cooling power at various times under the given capacity configuration of the upper layer, taking into account the energy market, capacity market, and primary frequency regulation market revenue of energy storage. It establishes a complete operation constraint system that includes system power balance constraints, renewable energy output constraints, energy storage-related constraints, data center constraints, and power purchase constraints of the distribution network. It calculates the operating cost, wind and solar curtailment penalty cost, and energy storage revenue in multiple markets under the scenario, and feeds back the objective function value to the upper-level capacity configuration unit.

[0160] Convergence determination and optimal solution output module: used to determine the convergence of the algorithm based on the threshold of the robust expected hypervolume improvement function or the maximum number of iterations; when the algorithm converges, the final sample set is Pareto sorted and the output is a set of optimal capacity configuration schemes for distributed resources that are robust to input noise.

[0161] Advantages and beneficial effects of the present invention:

[0162] 1) Effective handling of input noise: By simulating the uncertainty of new energy output, light, temperature, internal heat sources of data centers and internal heat dissipation coefficients of information technology equipment through input noise, the optimization results can maintain stable performance under various operating scenarios, significantly improving the practicality and robustness of the solution;

[0163] 2) Effective multi-objective optimization and trade-offs: By establishing an input noise model, the statistical distribution of the objective function under noise conditions is modeled using a Gaussian process. A robust expected hypervolume improvement function is adopted as the acquisition strategy to efficiently search for the Pareto optimal solution under uncertain input conditions. This algorithm can handle multiple objective functions simultaneously and perform natural trade-offs and optimizations among multiple objectives, achieving a dual balance between economic efficiency and renewable energy consumption rate.

[0164] 3) Establish a contribution assessment mechanism for energy storage in a diversified market environment: A decision-making framework for the collaborative optimization of energy storage participation in multiple markets such as electricity, capacity, and primary frequency regulation has been constructed, breaking through the limitations of traditional single-market assessment and realizing the contribution assessment and optimization planning of energy storage in a diversified market environment.

[0165] 4) Multi-objective robust Bayesian optimization algorithm with high efficiency and strong interpretability: The Bayesian optimization framework is superior to traditional heuristic algorithms in terms of sample utilization and search efficiency. At the same time, it intuitively reflects the uncertainty characteristics of each objective through probability distribution, which is convenient for engineering implementation and risk assessment.

[0166] In summary, this invention is applicable to integrated energy systems for data centers that include wind power, photovoltaics, batteries, supercapacitor energy storage, and cooling systems. In a multi-market coupled environment such as the electricity market, capacity market, and primary frequency regulation market, it achieves stable configuration schemes, balanced objectives, accurate returns, and efficient convergence through a two-layer robust optimization framework, noise modeling, multi-market collaborative evaluation, and Bayesian optimization algorithms. Attached Figure Description

[0167] Figure 1 It is the overall planning flowchart;

[0168] Figure 2 This is a flowchart of a multi-objective robust Bayesian optimization algorithm under input noise.

[0169] Figure 3 This is a dynamic model diagram of the thermal inertia of the data center building envelope; Detailed Implementation

[0170] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0171] Example 1

[0172] As attached Figure 1 As shown, a distributed resource planning method for a data center integrated energy system in a diversified market environment includes the following steps:

[0173] S1. Constructing a two-layer optimization framework for distributed resources in a data center integrated energy system: An upper-layer capacity configuration model and a lower-layer operation simulation model are established. The upper-layer decision variable is the capacity of various distributed resources, while the lower-layer model performs operation simulations based on a given capacity scheme. The two are coupled and iteratively coupled through feedback of configuration schemes and objective function values. The upper-layer multi-objective robust Bayesian optimization algorithm under input noise and the lower-layer collaborative optimization model for distributed resources in the data center integrated energy system form an interactive closed loop: The upper layer first initializes the dataset, constructs a Gaussian process surrogate model, selects the next sampling point through a robust expectation hypervolume improvement function, and passes the capacity configuration scheme of the distributed resources corresponding to the sampling point to the lower-layer model. After solving typical scenarios, the lower-layer model feeds back the operating cost and renewable energy absorption rate from the objective function to the upper-layer model. The upper-layer model updates the dataset based on the feedback objective function value and retrains the Gaussian process surrogate model. If the convergence condition is not met, the Gaussian process surrogate model continues to be updated; if the convergence condition is met, a robust Pareto optimal solution set, i.e., the optimal configuration scheme for distributed resources, is output.

[0174] S2. Algorithm Dataset Initialization: Determine the upper-level decision variables, namely the capacity of various distributed resources. Generate an initial capacity configuration sample set using random or quasi-random sampling. Then, call the lower-level model to evaluate the multi-objective response value of each sample to construct the initial observation dataset; (e.g.) Figure 2 As shown, steps S2-S6 constitute the complete flow of the multi-objective robust Bayesian optimization algorithm under input noise.

[0175] The hyperparameters required for the input multi-objective robust Bayesian optimization algorithm include: the initial number of sample points. Maximum number of iterations Convergence threshold The covariance matrix used to describe the distribution of input noise. and the number of scenarios used for uncertainty modeling .

[0176] Within the domain of design variables, random or quasi-random sampling methods (such as Sobol or Latin hypercube sampling) are used to select... initial sample points , representing the initial set of upper-level decision variables. Each sample point corresponds to a distributed resource capacity configuration scheme. The distributed resources mentioned in this scheme specifically include: wind power systems, photovoltaic systems, battery energy storage systems, supercapacitor energy storage systems, and data center cooling systems. Their configuration variables are shown in the formula. As shown:

[0177]

[0178] In the formula: , , These represent the rated power (kW) of the cooling system for wind farms, photovoltaic power plants, and data centers, respectively. and It is the first type of battery energy storage with rated power (kW) and capacity (kWh) used to ensure power supply and participate in the spot market; and It is the second type of battery energy storage with rated power (kW) and capacity (kWh) that prioritizes fulfilling capacity market obligations while taking into account power supply security and spot trading. and It is the first supercapacitor energy storage device with rated power (kW) and capacity (kWh) used to ensure power supply and participate in the spot market; and It is the second type of supercapacitor energy storage with rated power (kW) and capacity (kWh) specifically designed for participation in the frequency regulation market; a power greater than 0 indicates discharging, and a power less than 0 indicates charging.

[0179] Distributed resource capacity variables Subject to upper and lower limits, as shown in the formula As shown:

[0180]

[0181] In the formula: , They represent The lower and upper limits of the values.

[0182] For the initial sample points The distributed resource collaborative optimization model of the integrated energy system of the data center is used to evaluate the corresponding multiple target response values. This forms a multi-target observation pair. All input samples and their corresponding target values ​​are combined to form the initial observation dataset. , as in the formula As shown:

[0183]

[0184] In the formula: A vector containing multiple objective function values , where the objective function This represents minimizing the overall energy system cost of a data center, as shown in the formula. As shown:

[0185]

[0186] In the formula: This represents the expected operating cost across all scenarios, with the goal of minimizing the average operating cost across all possible scenarios, as shown in the formula. As shown:

[0187]

[0188] In the formula: It refers to the number of scenes, i.e., the number of samples with different noise or random input; This indicates that it is in the scene Annual operating costs (in yuan) The scene number, with a value range of 1, 2, ... In multi-objective optimization problems, The expectation value is typically used in multiple objective functions to represent the statistical treatment of scenario uncertainties. Robustness is achieved through the expectation operator, ensuring that the optimal solution possesses stability and economy under different new energy disturbance conditions.

[0189] This represents the sum of annual investment costs (in yuan) for all types of distributed resources. Investment costs are one-time and fixed, determined by higher-level decision variables. The distributed resource allocation scheme represented directly determines this, as shown in the formula. As shown:

[0190]

[0191] In the formula: , , , , , , The figures represent the annual investment costs (in yuan) for wind power, photovoltaic power, the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, the second type of supercapacitor energy storage, and the data center cooling system, respectively.

[0192] Investment costs of wind power, solar power and data center cooling systems , and As in the formula As shown:

[0193]

[0194] In the formula: Indicates the type of equipment; , , These represent wind power, photovoltaic, and data center cooling systems, respectively. , , These represent the power cost coefficients (yuan / kW) for wind power, photovoltaic, and data center cooling systems, respectively. Represents the discount rate; , , These represent the lifespan of wind power, photovoltaic, and data center cooling systems, respectively.

[0195] Investment costs of two types of battery energy storage and two types of supercapacitor energy storage , , , As in the formula As shown:

[0196]

[0197] In the formula: Indicates the type of energy storage device; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. , , , These represent the power cost coefficients (yuan / kW) for the four types of energy storage; , , , These represent the capacity cost coefficients (yuan / kWh) for four different energy storage methods. , , , These represent the lifespan of the four types of energy storage.

[0198] objective function To maximize the renewable energy absorption rate, as shown in the formula... As shown:

[0199]

[0200] In the formula: The average absorption rate across all scenarios is represented by the formula. As shown:

[0201]

[0202] Scene New energy consumption rate As in the formula As shown:

[0203]

[0204] In the formula: The actual amount of renewable energy electricity consumed (kWh) is represented by the formula. As shown:

[0205]

[0206] In the formula: and They represent The power output of photovoltaic power plants and wind farms at any given time (kW); This indicates the number of scheduling cycles in a year; Indicates the scheduling interval duration (h).

[0207] The total electricity generated from renewable energy sources (kWh) is expressed as follows: As shown:

[0208]

[0209] S3. Construction of a probabilistic surrogate model considering noisy input: Uncertainty factors are modeled as Gaussian noise, including: new energy output, temperature, illumination, and heat source; a Gaussian process model is trained based on the initial sample set to predict the statistical distribution of the objective function under noise; where, due to the high cost of evaluating the true objective function, this embodiment uses a Gaussian process (GP) as a surrogate model to statistically model the objective function, and predicts the mean and variance of the objective function through Bayesian optimization. Under the condition of noise in the input, the GP model is extended to a robust form to predict the mean and variance of the objective function under the input perturbation distribution, as shown in the formula. As shown:

[0210]

[0211] In the formula: It is a mean function. Let be the covariance function.

[0212] By learning from noisy sample data, the model can simultaneously estimate the expected response (i.e., the mean function) and uncertainty (i.e., the covariance function) under the input distribution, thus providing a reliable measure for robust optimization. Input noise refers to the uncertainty or fluctuation of input variables that affect the objective function, originating from renewable energy output, illumination, temperature, internal heat sources in data centers, and the internal heat dissipation coefficient of information technology equipment. The set of input noise variables is shown in the formula. As shown:

[0213]

[0214] In the formula: This indicates the number of scheduling operations within a scheduling cycle; and They represent The power output of the photovoltaic power station and the actual power output (kW) of the wind farm at any given time; express Outdoor temperature at any time (°C); express Temperature (°C) in the adjacent area of ​​the data center at any given time; Indicates the wall number; Indicates wall Light intensity (W / m²) in the corresponding direction, ; This indicates the light intensity in the corresponding direction of the window; This indicates the heat source (W) inside the data center; This indicates the heat dissipation coefficient inside data center information technology equipment.

[0215] The system's actual input is subject to random disturbances The influence of this influence forms the observed input. As shown in formula (16):

[0216]

[0217] noise It follows a Gaussian distribution, and its expression is shown in formula (17):

[0218]

[0219] In the formula: The covariance matrix represents the noise and is used to characterize the statistical properties of the input disturbance.

[0220] S4. Optimal Point Selection Based on Maximizing the Hypervolume Function: A robust expected hypervolume improvement function is used as the sampling function. By maximizing this function, the next most promising capacity configuration sample is selected. The selected sample is then passed to the lower-level model for simulation to obtain the target value. The sample set and surrogate model are updated, and this process is repeated until convergence. During the Bayesian optimization process, the sampling function is used to evaluate the quality of the current sampling point and guide the selection of the next sampling point. This embodiment uses a robust expected hypervolume improvement function. As an acquisition function, hypervolume measures the area covered by the solution set within the target space. The next most robust sampling point is selected by calculating the hypervolume gain under noisy conditions. The desired hypervolume gain optimizes the sampling process by considering multiple possible noise scenarios, thereby improving the model's robustness. The mathematical form of the robust desired hypervolume improvement function is shown in the formula... As shown:

[0221]

[0222] In the formula: This is the current non-dominated solution set; The hypervolume index measures the size of the Pareto front coverage area of ​​a solution set; a larger value indicates a higher quality solution set. To determine the noise distribution The expected value indicates Based on input Noise sampling;

[0223] Improved by maximizing robust expectation hypervolume To select the optimal next sampling point The determination method is as follows: As shown:

[0224]

[0225] S5. Establish a lower-level operation simulation optimization model: In the lower-level model, based on the capacity configuration given by the upper level, the energy storage revenue of the power market, capacity market, and primary frequency regulation market is comprehensively considered to optimize the equipment output, load migration, and cooling power at each moment, and to establish a complete operation constraint system that includes system power balance, energy storage constraints, and temperature constraints.

[0226] In this model, when the capacity configuration scheme is determined at the upper level, the lower-level model performs periodic operational simulations and optimizations on the data center's integrated energy system. The lower-level optimization variables include the wind power output at each moment. Photovoltaic power output The first type of battery energy storage output The second type of battery energy storage output The first type of supercapacitor energy storage output The second type of supercapacitor energy storage output Data center IT equipment migration power matrix Data center cooling capacity Interaction power between the integrated energy system of the data center and the power distribution network .

[0227] The optimization objective of the lower-level model is to minimize the operating cost of the data center's integrated energy system while satisfying energy balance and equipment constraints. Operating cost For the scene The operating costs of each distributed resource and the penalty costs for wind and solar curtailment, minus the revenue from energy storage participating in the electricity, capacity, and primary frequency regulation markets, can be expressed as the formula. As shown:

[0228]

[0229] In the formula: , , , , These figures represent the operating costs (in yuan) of wind power, photovoltaic power, storage batteries, supercapacitors, and data center cooling systems, respectively. This represents the penalty cost (in yuan) for abandoning wind and solar power. , , These represent the revenue from the energy market (RMB), capacity market (RMB), and primary frequency regulation market (RMB), respectively.

[0230] Operating costs of wind power As in the formula As shown:

[0231]

[0232] In the formula: This represents the fixed operation and maintenance cost coefficient for wind power (RMB / kW).

[0233] Operating costs of photovoltaics As in the formula As shown:

[0234]

[0235] In the formula: This represents the fixed operation and maintenance cost coefficient for photovoltaic systems (RMB / kW).

[0236] Operating costs of battery energy storage As in the formula As shown:

[0237]

[0238] In the formula: This represents the fixed operation and maintenance cost coefficient of the battery (yuan / kW).

[0239] Operating costs of supercapacitor energy storage As in the formula As shown:

[0240]

[0241] In the formula: This represents the fixed operation and maintenance cost coefficient (RMB / kW) for supercapacitors.

[0242] Data center operating costs As in the formula As shown:

[0243]

[0244] In the formula: express Electricity price per kWh (RMB / kWh); express Energy consumption (kW) of data center information technology equipment at any given time; express Energy consumption (kW) of the cooling system in the real-time data center; express The energy consumption (kW) of other equipment in a data center is relatively small and is usually considered a constant.

[0245] The punitive costs of abandoning wind and solar power As in the formula As shown:

[0246]

[0247] In the formula: This represents the penalty cost coefficient (yuan / kW).

[0248] Energy storage adjusts the interaction power between the integrated energy system of a data center and the power distribution network through charging and discharging. It is a key tool to help achieve arbitrage, and the returns from the electricity market. As in the formula As shown:

[0249]

[0250] In the formula: Energy storage adjusts the interaction power between the integrated energy system of a data center and the power distribution network through charging and discharging, in order to achieve arbitrage;

[0251] Energy storage capacity market revenue As in the formula As shown:

[0252]

[0253] In the formula: This indicates the market settlement price for capacity (RMB / kW·year); The committed capacity (kW) for battery energy storage to participate in the capacity market is expressed in the formula. As shown:

[0254]

[0255] In the formula: It is the effective capacity factor. The number of hours (h) of continuous discharge as specified by the market.

[0256] primary frequency regulation market revenue of energy storage As in the formula As shown:

[0257]

[0258] In the formula: This indicates the market clearing price for frequency regulation capacity (RMB / kWh); This indicates the unit price for mileage adjustment compensation (yuan / kW); This is the frequency modulation mileage factor for the supercapacitor; The performance coefficient of a supercapacitor; This refers to the frequency regulation mileage (kW) of the second type of supercapacitor energy storage in the primary frequency regulation capacity market, with constraints as shown in the formula. As shown.

[0259] The constraints include:

[0260] 1) Power balance constraints of the integrated energy system of the data center, as shown in the formula. As shown:

[0261]

[0262] 2) Renewable energy output constraints, as shown in the formula ,formula As shown:

[0263]

[0264]

[0265] 3) Energy storage-related constraints

[0266] Energy storage output is limited by the power of the energy storage configuration, as shown in the formula. ,formula As shown:

[0267]

[0268]

[0269] The state of charge at the beginning and end of the dispatch cycle must be equal to ensure the sustainable operation of the energy storage system, as shown in the formula. As shown:

[0270]

[0271] The method for calculating the state of charge is as follows: As shown:

[0272]

[0273] In the formula: It refers to the types of energy storage devices; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. This indicates the self-discharge rate of the battery's energy storage. It refers to the charging and discharging efficiency of energy storage.

[0274] This refers to the frequency regulation mileage (kW) of the second type of supercapacitor energy storage in the primary frequency regulation capacity market, with upper and lower limits as shown in the formula. As shown:

[0275]

[0276] In the formula: This is the frequency modulation capacity proportional coefficient.

[0277] 4) Data center constraints

[0278] 4.1) Power constraints for information technology equipment

[0279] This is a power transfer matrix with 100 rows and 100 columns. Among them, the most migrateable loads can be migrated. The load on the data center's IT equipment is calculated using the formula [formula missing]. As shown:

[0280]

[0281] In the formula: This represents the real-time load (kW) of the data center that is not portable. This represents the adjusted data center delayable load (kW), as shown in the formula. As shown:

[0282]

[0283] Elements in the power transfer matrix This represents the load migration ratio, and its range is shown in the formula. As shown:

[0284]

[0285] Since the total offline load remains unchanged before and after the migration, the migration must be completed before the deadline. Follow rows and constraints, as in formulas As shown:

[0286]

[0287] The selected offline workload can only be delayed and cannot be processed in advance. The maximum delay time for the data center's offline workload is [missing information]. Each scheduling interval, as shown in the formula As shown:

[0288]

[0289] After adjustment The power consumed by the data center to process network load at any given moment is greater than the real-time load power at that moment, but less than the power consumed by the data center to process network load when it is running at full load, as shown in the formula. As shown:

[0290]

[0291] In the formula: This indicates the power (kW) required for the data center to handle network load when operating at full capacity.

[0292] 4.2) Cooling power constraint:

[0293] Refrigeration system power Constrained by upper and lower limits, as in the formula As shown:

[0294]

[0295] In the formula: This is the upper limit of the operating power of the refrigeration system (kW); refrigeration system operating power The relationship with cooling capacity is shown in the formula. As shown:

[0296]

[0297] In the formula: The energy efficiency ratio of the air conditioning refrigeration system; for The cooling capacity (kW) of the air conditioning system at any given time.

[0298] like Figure 3 As shown, the thermal inertial dynamic model of the data center envelope includes one indoor air node and four wall nodes.

[0299] Establish a dynamic model of the thermal inertia of the data center building envelope, as shown in the formula. - As shown:

[0300]

[0301]

[0302]

[0303]

[0304]

[0305] In the formula: express Time Wall Temperature (°C); Indicates wall The heat capacity (J / °C); Indicates wall Thermal resistance (°C / W); Indicates indoor temperature (°C); For walls The heat absorption rate; For walls The area (m²); The area of ​​the window (m²); The room's heat capacity (J / °C); The thermal resistance of the window (°C / W); The refractive index of the window.

[0306] To ensure the server operates normally, the indoor temperature... It needs to meet the set upper and lower limits, such as the formula. As shown:

[0307]

[0308] In the formula: , These represent the upper and lower limits (°C) of indoor temperature set to ensure the normal operation of the server.

[0309] 5) Power purchase constraints of the distribution network

[0310] Power purchased by the distribution network Constrained by upper and lower limits, as in the formula As shown:

[0311]

[0312] In the formula: This represents the maximum power (kW) that can be purchased by the distribution network.

[0313] S6. Algorithm Convergence Determination and Optimal Solution Output: Obtain the Next Sampling Point Then, the point is evaluated in the model to obtain its multi-objective response results. , to new sample data Add to algorithm dataset Retrain the Gaussian process model, update the mean and variance estimates, and repeat steps S3 to S6 until the algorithm converges, i.e., the change in the robust expected hypervolume improvement function is less than the threshold. ( =0.5%) or reach the maximum number of iterations. After the algorithm converges, Pareto ranking is performed using the multi-objective response results of all samples to obtain the robust optimal solution set under input noise conditions, which is the optimal allocation scheme for distributed resources. The solution set exhibits strong stability and performance balance under different input perturbation conditions.

[0314] Example 2

[0315] This embodiment provides a distributed resource planning system for a data center integrated energy system in a diversified market environment. It implements the distributed resource planning method for a data center integrated energy system in a diversified market environment as described in Embodiment 1. The system includes a two-layer optimization framework module, a dataset initialization module, a probabilistic surrogate model construction module, a sampling point selection module, a lower-layer simulation optimization module, and a convergence determination and optimal solution output module. These modules work together to achieve robust optimization configuration of distributed resources, as detailed below:

[0316] The two-layer optimization framework module is used to construct a two-layer optimization framework for distributed resources in a data center integrated energy system. It includes an upper-layer capacity configuration unit and a lower-layer operation simulation unit. The upper-layer capacity configuration unit uses the capacity of various distributed resources as decision variables, while the lower-layer operation simulation unit performs operation simulations based on the capacity scheme given by the upper layer. The two are coupled and iteratively coupled through the feedback of the configuration scheme and the objective function value. Among them, the upper-layer capacity configuration unit integrates a multi-objective robust Bayesian optimization algorithm under input noise, and the lower-layer operation simulation unit integrates a collaborative optimization model for distributed resources in a data center integrated energy system, forming an interactive closed loop.

[0317] Dataset initialization module: used to determine the upper-level decision variables, namely the capacity of various distributed resources, generate an initial capacity configuration sample set by random or quasi-random sampling, call the lower-level simulation unit to evaluate the multi-objective response value of each sample, and construct the initial observation dataset; at the same time, input the hyperparameters required by the multi-objective robust Bayesian optimization algorithm, including the number of initial sample points, the maximum number of iterations, the convergence threshold, and the covariance matrix used to describe the distribution of input noise;

[0318] The probabilistic proxy model building module is used to model uncertainties as Gaussian noise, including new energy output, temperature, illumination, internal heat sources in data centers, and internal heat dissipation coefficients of information technology equipment; it trains a Gaussian process model based on the initial observation dataset, the Gaussian process model being robust, to predict the statistical distribution of the objective function under noise, and outputs the mean and variance of the objective function;

[0319] Sampling point selection module: Used to select the next most promising capacity configuration sample. Using the robust expected hypervolume improvement function as the acquisition function, the next capacity configuration sample is selected by maximizing this function; the selected sample is passed to the lower-level simulation unit, receiving the target value from the lower layer, updating the sample set and Gaussian process model, and repeating the above process until the convergence criterion is met;

[0320] The lower-level operation simulation optimization module is used to optimize equipment output, load migration, and cooling power at various times under the given capacity configuration of the upper layer, taking into account the energy market, capacity market, and primary frequency regulation market revenue of energy storage. It establishes a complete operation constraint system that includes system power balance constraints, renewable energy output constraints, energy storage-related constraints, data center constraints, and power purchase constraints of the distribution network. It calculates the operating cost, wind and solar curtailment penalty cost, and energy storage revenue in multiple markets under the scenario, and feeds back the objective function value to the upper-level capacity configuration unit.

[0321] Convergence determination and optimal solution output module: used to determine the convergence of the algorithm based on the threshold of the robust expected hypervolume improvement function or the maximum number of iterations; when the algorithm converges, the final sample set is Pareto sorted and the output is a set of optimal capacity configuration schemes for distributed resources that are robust to input noise.

Claims

1. A distributed resource planning method for a data center integrated energy system in a diversified market environment, characterized in that, Includes the following steps: S1. Construct a two-layer optimization framework for distributed resources in a data center integrated energy system: establish an upper-layer capacity configuration model and a lower-layer operation simulation model. The upper-layer decision variable is the capacity of various distributed resources, and the lower-layer model performs operation simulation based on a given capacity scheme. The two are coupled and iterated through configuration scheme and objective function value feedback. S2. Algorithm Dataset Initialization: Determine the upper-level decision variables, namely the capacity of various distributed resources, generate an initial capacity configuration sample set by random or quasi-random sampling, and call the lower-level model to evaluate the multi-objective response value of each sample to construct the initial observation dataset. S3. Constructing a probabilistic surrogate model considering noise input: Modeling uncertainties as Gaussian noise, including: new energy output, temperature, illumination, and heat source; training a Gaussian process model based on the initial sample set to predict the statistical distribution of the objective function under noise; S4. Optimal point selection based on maximizing the hypervolume function: The robust expected hypervolume improvement function is used as the acquisition function. The next most promising capacity configuration sample is selected by maximizing this function. The selected sample is passed to the lower-level model for simulation to obtain the target value. The sample set and surrogate model are updated and repeated until convergence. S5. Establish a lower-level operation simulation optimization model: In the lower-level model, based on the capacity configuration given by the upper level, the energy storage revenue of the power market, capacity market, and primary frequency regulation market is comprehensively considered to optimize the equipment output, load migration, and cooling power at each moment, and to establish a complete operation constraint system that includes system power balance, energy storage constraints, and temperature constraints. S6. Algorithm Convergence Judgment and Optimal Solution Output: The convergence of the algorithm is judged based on the threshold of the robust expected hypervolume improvement function or the maximum number of iterations. The final sample set is Pareto sorted, and the set of optimal capacity configuration schemes that are robust to input noise is output.

2. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S1, the upper-layer multi-objective robust Bayesian optimization algorithm under input noise and the lower-layer distributed resource collaborative optimization model of the integrated energy system of the data center form an interactive closed loop. The specific process is as follows: The upper layer first initializes the dataset, constructs a Gaussian process surrogate model, selects the next sampling point through the robust expectation hypervolume improvement function, and passes the capacity configuration scheme of the distributed resources corresponding to the sampling point to the lower-layer model; After the lower-layer model constructs typical scenarios and solves them, it feeds back the operating cost and new energy absorption rate in the objective function to the upper-layer model. The upper-layer model updates the dataset and retrains the Gaussian process surrogate model according to the feedback objective function value; If the convergence condition is not met, the Gaussian process surrogate model is updated again; If the convergence condition is met, the robust Pareto optimal solution set, i.e., the optimal configuration scheme of distributed resources, is output.

3. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S2, input the hyperparameters required for the multi-objective robust Bayesian optimization algorithm, including: the initial number of sample points. Maximum number of iterations Convergence threshold The covariance matrix used to describe the distribution of input noise. and the number of scenarios used for uncertainty modeling The distributed resources mentioned above include: wind power systems, photovoltaic systems, battery energy storage systems, supercapacitor energy storage systems, and data center cooling systems, and their configuration variables are shown in formula (1): In the formula: , , These represent the rated power of the cooling systems for wind farms, photovoltaic power plants, and data centers, respectively. , These are the rated power and capacity of the first type of battery energy storage used to ensure power supply and participate in the spot market; , These are the rated power and capacity of the second type of battery energy storage, which prioritizes fulfilling capacity market obligations while also considering power supply security and spot trading. , These are the rated power and capacity of the first type of supercapacitor used to ensure power supply and participate in the spot market; , These are the rated power and capacity of the second type of supercapacitor energy storage specifically designed for the frequency regulation market; a power value greater than 0 indicates discharging, and a power value less than 0 indicates charging. Distributed resource capacity variables Constrained by upper and lower limits, its expression is shown in formula (2): In the formula: , They represent The lower and upper limits of the values; For the initial sample points The distributed resource collaborative optimization model of the integrated energy system of the data center is used to evaluate the corresponding multiple target response values. This forms a multi-target observation pair; all input samples and their corresponding target values ​​are combined to form the initial observation dataset. As shown in formula (3): In the formula: A vector containing multiple objective function values , where the objective function The cost of minimizing the integrated energy system of the data center is shown in formula (4); In the formula: Let $\mathbf{ ... In the formula: It refers to the number of scenes, i.e., the number of samples with different noise or random input; This indicates that it is in the scene Annual operating costs The scene number, with a value range of 1, 2, ... In multi-objective optimization problems, As an expected value, it is often used in multiple objective functions to represent the statistical processing of scenario uncertainty; It is the sum of the annual investment costs of various distributed resources, determined by upper-level decision variables. The distributed resource configuration scheme represented directly determines this, as shown in formula (6): In the formula: , , , , , , These represent the annual investment costs for wind power, photovoltaic power, the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, the second type of supercapacitor energy storage, and data center cooling systems, respectively. Investment costs of wind power, solar power and data center cooling systems , , As shown in formula (7): In the formula: Indicates the type of equipment; , , These represent wind power, photovoltaic, and data center cooling systems, respectively. , , These represent the power cost coefficients for wind power, photovoltaic, and data center cooling systems, respectively. Represents the discount rate; , , These represent the lifespan of wind power, photovoltaic, and data center cooling systems, respectively. Investment costs of two types of battery energy storage and two types of supercapacitor energy storage , , , As shown in formula (8): In the formula: Indicates the type of energy storage device; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. , , , These represent the power cost coefficients for four different energy storage methods. , , , These represent the capacity cost coefficients for four different energy storage methods. , , , These represent the lifespan of the four types of energy storage; objective function To maximize the renewable energy absorption rate, as shown in formula (9): In the formula: The average absorption rate across all scenarios is shown in formula (10): Scene New energy consumption rate As shown in formula (11): In the formula: The actual amount of renewable energy consumed is shown in formula (12): In the formula: , They represent The output of photovoltaic power plants and wind farms at all times; This indicates the number of scheduling cycles in a year; Indicates the duration of the scheduling interval; The total electricity generated from renewable energy sources is represented as shown in formula (13): 。 4. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S3, a Gaussian process is used as a surrogate model to statistically model the objective function. Bayesian optimization is used to predict the mean and variance of the objective function. Under the condition that there is noise in the input, the GP model is extended to a robust form to predict the mean and variance of the objective function under the input perturbation distribution, as shown in formula (14): In the formula: It is a mean function. It is the covariance function; The input noise originates from new energy output, illumination, temperature, internal heat sources of the data center, and internal heat dissipation coefficients of information technology equipment. The set of input noise variables is shown in formula (15): In the formula: This indicates the number of scheduling operations within a scheduling cycle. and They represent The power output of photovoltaic power plants and the actual power output of wind farms at all times; express outdoor temperature at all times; express Temperature of adjacent areas to the data center at any given time; Indicates the wall number; Indicates wall Light intensity in the corresponding direction, ; This indicates the light intensity in the corresponding direction of the window; Indicates the internal heat source of the data center; This indicates the heat dissipation coefficient inside data center information technology equipment; The system's actual input is subject to random disturbances The influence of this influence forms the observed input. As shown in formula (16): noise It follows a Gaussian distribution, as shown in formula (17): In the formula: The covariance matrix represents the noise and is used to characterize the statistical properties of the input disturbance.

5. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S4, during the Bayesian optimization process, the acquisition function is used to evaluate the quality of the current sampling point and guide the selection of the next sampling point; a robust expectation hypervolume improvement function is adopted. As the acquisition function, its expression is shown in formula (18): In the formula: This is the current non-dominated solution set; The hypervolume index measures the size of the Pareto front coverage area of ​​a solution set; a larger value indicates a higher quality solution set. To determine the noise distribution The expected value indicates Based on input Noise sampling; Improved by maximizing robust expectation hypervolume To select the optimal next sampling point Its expression is shown in formula (19): 。 6. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S5, when the capacity configuration scheme is determined at the upper level, the lower-level model performs periodic operational simulation optimization of the data center's integrated energy system. The lower-level optimization variables include the wind power output at each moment. Photovoltaic power output The first type of battery energy storage output The second type of battery energy storage output The first type of supercapacitor energy storage output The second type of supercapacitor energy storage output Data center IT equipment migration power matrix Data center cooling capacity Interaction power between the integrated energy system of the data center and the power distribution network ; The optimization objective of the lower-level model is to minimize the operating cost of the integrated energy system of the data center while satisfying energy balance and equipment constraints. For the scene The operating costs of each distributed resource and the penalty costs of curtailing wind and solar power, minus the revenue from energy storage participating in the electricity, capacity, and primary frequency regulation markets, are expressed as formula (20): In the formula: , , , , These represent the operating costs of wind power, photovoltaic power, storage batteries, supercapacitors, and data center cooling systems, respectively. This indicates the penalty cost for abandoning wind and solar power. , , These represent the revenue from the electricity market, the revenue from the capacity market, and the revenue from the primary frequency regulation market for energy storage, respectively. Operating costs of wind power As shown in formula (21): In the formula: This represents the fixed operation and maintenance cost coefficient for wind power. Operating costs of photovoltaics As shown in formula (22): In the formula: This represents the fixed operation and maintenance cost coefficient for photovoltaic systems. Operating costs of battery energy storage As shown in formula (23): In the formula: This represents the fixed operating and maintenance cost coefficient of the battery. Operating costs of supercapacitor energy storage As shown in formula (24): In the formula: This represents the fixed operation and maintenance cost coefficient of a supercapacitor. Data center operating costs As shown in formula (25): In the formula: express Electricity price at any time; express Energy consumption of data center information technology equipment at all times; express Energy consumption of the cooling system in a real-time data center; express Energy consumption of other equipment in the data center at any given time; The punitive costs of abandoning wind and solar power As shown in formula (26): In the formula: This represents the penalty cost coefficient; Electricity market revenue As shown in formula (27): In the formula: Energy storage adjusts the interaction power between the integrated energy system of a data center and the power distribution network through charging and discharging, in order to achieve arbitrage; Energy storage capacity market revenue As shown in formula (28): In the formula: This indicates the market settlement price for capacity. The committed capacity for battery energy storage to participate in the capacity market is shown in formula (29): In the formula: It is the effective capacity factor. The number of hours of continuous discharge as specified by the market; primary frequency regulation market revenue of energy storage As shown in formula (30): In the formula: This indicates the market clearing price for frequency modulation capacity; This indicates the unit price for mileage adjustment compensation. This is the frequency modulation mileage factor for the supercapacitor; The performance coefficient of a supercapacitor; It is the frequency regulation mileage of the second type of supercapacitor energy storage in the primary frequency regulation capacity market, and its constraint is shown in formula (38); The constraints include: 1) The power balance constraint of the integrated energy system of the data center is shown in formula (31): 2) Renewable energy output constraints, as shown in formulas (32) and (33): 3) Energy storage-related constraints The output power of energy storage is limited by the power of the energy storage configuration, as shown in formulas (34) and (35): The state of charge at the beginning and end of the scheduling cycle must be equal, as shown in formula (36): The method for calculating the state of charge is shown in formula (37): In the formula: It refers to the types of energy storage devices; , , , These represent the first type of battery energy storage, the second type of battery energy storage, the first type of supercapacitor energy storage, and the second type of supercapacitor energy storage, respectively. This indicates the self-discharge rate of the battery's energy storage. It refers to the charging and discharging efficiency of energy storage; The second type of supercapacitor energy storage has a frequency regulation mileage in the primary frequency regulation capacity market, and its upper and lower limits are as shown in formula (38): In the formula: This is the frequency modulation capacity proportional coefficient; 4) Data center constraints 4.1) Power constraints for information technology equipment This is a power transfer matrix with 100 rows and 100 columns. Among them, the most migrateable loads can be migrated. Each scheduling interval, the load of information technology equipment in the data center. Expressions are like formulas As shown: In the formula: This indicates the real-time load that is not migrated from the data center; The adjusted data center delay load is expressed as shown in formula (40): Elements in the power transfer matrix This represents the load migration ratio, and its range is shown in formula (41): Since the total offline load remains unchanged before and after the migration, the migration must be completed before the deadline. Follow the row and constraint rules, as shown in formula (42): The selected offline workload can only be delayed and cannot be processed in advance. The maximum delay time for the data center's offline workload is... The scheduling interval is shown in formula (43): After adjustment The power of the data center in processing network load at any given moment is greater than the real-time load power at that moment, but less than the power of the data center in processing network load when it is running at full load, as shown in formula (44): In the formula: This indicates the power consumed by the data center when it is operating at full load to handle network load. 4.2) Cooling power constraint: Refrigeration system power Constrained by upper and lower limits, as shown in formula (45): In the formula: This is the upper limit of the operating power of the refrigeration system; the operating power of the refrigeration system The relationship with cooling capacity is shown in formula (46): In the formula: The energy efficiency ratio of the air conditioning refrigeration system; for The cooling capacity of the air conditioning system at all times; The dynamic model of the thermal inertia of the data center envelope includes one indoor air node and four wall nodes. The dynamic model of the thermal inertia of the data center envelope is established as shown in formulas (47)-(51): In the formula: express Time Wall temperature; Indicates wall The heat capacity; Indicates wall Thermal resistance; Indicates indoor temperature; For walls The heat absorption rate; For walls The area; The area of ​​the window; The heat capacity of the room; For window thermal resistance; The refractive index of the window; To ensure the server operates normally, the indoor temperature... The upper and lower limits need to be met, as shown in formula (52): In the formula: , These represent the lower and upper limits of the indoor temperature set to ensure the normal operation of the server; 5) Power purchase constraints of the distribution network Power purchased by the distribution network Constrained by upper and lower limits, as shown in formula (53): In the formula: This represents the maximum value of electricity purchased by the distribution network.

7. The distributed resource planning method for a data center integrated energy system in a diversified market environment according to claim 1, characterized in that, In step S6, the next sampling point is obtained. Then, the point is evaluated in the model to obtain its multi-objective response results. , to new sample data Add to algorithm dataset Retrain the Gaussian process model, update the mean and variance estimates, and repeat steps S3 to S6 until the algorithm converges, i.e., the change in the robust expected hypervolume improvement function is less than the threshold. Or it may reach the maximum number of iterations.

8. A distributed resource planning system for a data center integrated energy system in a diversified market environment, used to implement the distributed resource planning method for a data center integrated energy system in a diversified market environment as described in claims 1-7, characterized in that, It includes a two-layer optimization framework module, a dataset initialization module, a probabilistic surrogate model construction module, a sampling point selection module, a lower-level simulation optimization module, and a convergence determination and optimal solution output module. These modules work together to achieve robust optimization configuration of distributed resources, as detailed below: The two-layer optimization framework module is used to construct a two-layer optimization framework for distributed resources in a data center integrated energy system. It includes an upper-layer capacity configuration unit and a lower-layer operation simulation unit. The upper-layer capacity configuration unit uses the capacity of various distributed resources as decision variables, while the lower-layer operation simulation unit performs operation simulations based on the capacity scheme given by the upper layer. The two are coupled and iteratively coupled through the feedback of the configuration scheme and the objective function value. Among them, the upper-layer capacity configuration unit integrates a multi-objective robust Bayesian optimization algorithm under input noise, and the lower-layer operation simulation unit integrates a collaborative optimization model for distributed resources in a data center integrated energy system, forming an interactive closed loop. Dataset initialization module: used to determine the upper-level decision variables, namely the capacity of various distributed resources, generate an initial capacity configuration sample set by random or quasi-random sampling, call the lower-level simulation unit to evaluate the multi-objective response value of each sample, and construct the initial observation dataset; at the same time, input the hyperparameters required by the multi-objective robust Bayesian optimization algorithm, including the number of initial sample points, the maximum number of iterations, the convergence threshold, and the covariance matrix used to describe the distribution of input noise; The probabilistic proxy model building module is used to model uncertainties as Gaussian noise, including new energy output, temperature, illumination, internal heat sources in data centers, and internal heat dissipation coefficients of information technology equipment; it trains a Gaussian process model based on the initial observation dataset, the Gaussian process model being robust, to predict the statistical distribution of the objective function under noise, and outputs the mean and variance of the objective function; Sampling point selection module: It is used to select the next most promising capacity configuration sample by using the robust expected hypervolume improvement function as the acquisition function, and select the next capacity configuration sample by maximizing the function; the selected sample is passed to the lower-level running simulation unit, the target value is received from the lower-level feedback, the sample set and Gaussian process model are updated, and the above process is repeated until the convergence criterion is met. The lower-level operation simulation optimization module is used to optimize equipment output, load migration, and cooling power at various times under the given capacity configuration of the upper layer, taking into account the energy market, capacity market, and primary frequency regulation market revenue of energy storage. It establishes a complete operation constraint system that includes system power balance constraints, renewable energy output constraints, energy storage-related constraints, data center constraints, and power purchase constraints of the distribution network. It calculates the operating cost, wind and solar curtailment penalty cost, and energy storage revenue in multiple markets under the scenario, and feeds back the objective function value to the upper-level capacity configuration unit. Convergence determination and optimal solution output module: used to determine the convergence of the algorithm based on the threshold of the robust expected hypervolume improvement function or the maximum number of iterations; when the algorithm converges, the final sample set is Pareto sorted and the output is a set of optimal capacity configuration schemes for distributed resources that are robust to input noise.

Images