Data center energy system energy-saving economic dispatching method based on prediction model

Abstract

The invention discloses an energy-saving economic dispatching method for a data center energy system based on a prediction model. The method comprises the following steps: establishing a digital twinborn model of the data center energy system; dividing the data center into a plurality of energy consumption partitions, and installing a temperature acquisition device in each energy consumption partition; after a cooling capacity demand prediction model of the data center under each time period partition is established, target combination parameters and target energy efficiency indexes PUE of the chilled water temperature and the air supply temperature of a refrigeration subsystem under each time period are obtained through calculation; establishing a data center energy efficiency index PUE prediction model, and establishing an upper-layer refrigeration subsystem optimization scheduling model by taking the minimum deviation between an energy efficiency index PUE prediction value and a target energy efficiency index PUE, the minimum deviation between actual combined parameters of chilled water temperature and air supply temperature and target combined parameters and the minimum cold capacity supply and demand deviation of each partition as targets; and establishing a lower-layer data center optimization scheduling model by taking the minimum power utilization operation cost of the system and the minimum load migration cost of the IT equipment as a target.

Description

Technical Field

[0001] The present invention belongs to the technical field of data centers, and in particular relates to an energy-saving and economical scheduling method for a data center energy system based on a prediction model. Background Art

[0002] With the widespread adoption of next-generation technologies like cloud computing, big data, and artificial intelligence, the explosive growth in data demand has made computing power a new key production factor. This means that the scale and growth of data center construction have entered a period of rapid development. Data centers, comprised of hundreds of IT devices, perform massive amounts of data transmission, storage, and computing tasks, supporting the stable operation of the internet.

[0003] Currently, with the continuous growth in the scale and number of data centers, data center equipment consumes more power. Furthermore, to protect IT equipment from damage due to excessive temperatures, which could threaten data security, data center cooling systems must operate 24 / 7. Many data centers typically set a lower temperature range to ensure safe operation of equipment, resulting in higher energy consumption for cooling equipment. However, the ambient temperature in a data center changes dynamically with IT equipment load and external factors, and the cooling capacity required by the data center also fluctuates accordingly. Therefore, how to regulate various devices based on the fluctuations in ambient temperature, improve the energy efficiency and economic benefits of the data center, reduce cooling energy consumption, and achieve a balance between cooling supply and demand are currently urgent issues that need to be addressed.

[0004] Based on the above technical problems, it is necessary to design a new energy-saving and economic scheduling method for data center energy systems based on prediction models. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology and provide an energy-saving and economic scheduling method for the data center energy system based on a prediction model. By establishing a digital twin model of the data center energy system, dividing energy consumption zones, predicting the cooling demand of the zones, predicting the energy efficiency index PUE of the data center, and establishing an upper-level refrigeration subsystem optimization scheduling model and a lower-level data center optimization scheduling model, the economic benefits of the data center can be improved, the operating energy consumption can be reduced, and the cooling supply and demand balance of each zone can be achieved.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] The present invention provides a method for energy-saving and economical scheduling of a data center energy system based on a prediction model, which includes:

[0008] S1. Establish a digital twin model of the data center energy system, including the power supply subsystem, refrigeration subsystem, and IT equipment subsystem; the refrigeration subsystem includes refrigeration equipment, cold storage equipment, and terminal precision air conditioning units;

[0009] S2. Monitor the energy consumption data of each device in the IT equipment subsystem. Based on the location layout of each device and the air supply circuit layout of the terminal precision air conditioning units, divide the data center into multiple energy consumption zones. Install a temperature collection device in each energy consumption zone to obtain real-time ambient temperature data for each energy consumption zone.

[0010] S3. Based on the ambient temperature data of each energy consumption zone and the operating and meteorological data of each device within the zone, a cooling demand forecast model for each time zone of the data center is established. The process mechanism of the refrigeration subsystem is analyzed using the digital twin model of the data center energy system. The target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem at each time zone, as well as the target energy efficiency index (PUE), are calculated.

[0011] S4. Establish a data center energy efficiency indicator (PUE) prediction model, taking into account the differences in the operating characteristics of each device in the refrigeration subsystem, the cold storage capacity, and the control mechanism of each refrigeration circuit. With the goal of minimizing the deviation between the predicted PUE value and the target PUE value, minimizing the deviation between the actual combination of chilled water temperature and supply air temperature and the target combination, and minimizing the deviation between the supply and demand of cooling capacity in each zone, establish an optimization scheduling model for the upper refrigeration subsystem.

[0012] S5. With the goal of minimizing the system power operation cost and IT equipment load migration cost, a lower-level data center optimization scheduling model is established. Through the interactive solution of the two-level optimization scheduling model, the optimal scheduling strategy of the system is obtained.

[0013] Furthermore, in S1, the establishment of the digital twin model of the data center energy system includes:

[0014] At the physical layer: construct the physical entity of the data center energy system based on the model, size, material and layout position of each device in the power supply subsystem, refrigeration subsystem and IT equipment subsystem; the power supply subsystem includes new energy generator sets, traditional conventional power supply units and municipal power grids; the refrigeration equipment in the refrigeration subsystem includes chillers, cooling water pumps, cooling towers and chilled water pumps, and there are multiple of them; the cold storage equipment in the refrigeration subsystem is connected to the refrigeration equipment to store excess cold and release it when cold is lost; the number of the terminal precision air-conditioning units is arranged according to the equipment scale of the IT equipment subsystem, and corresponding supply air circuits and return air circuits are provided; the IT equipment subsystem includes servers, storage devices and network equipment;

[0015] At the twin layer: Introduce the corresponding physical attribute information, behavior information, and rule information between devices into the physical entity of the data center energy system to establish a digital twin model of the data center energy system mechanism. The physical attribute information includes the rated operating parameters and performance parameters of the equipment; the behavior information includes the operating status and operation behavior of the equipment; the rule information includes the start and stop sequence of the equipment, the input and output coupling relationship between the equipment, and the power supply, cooling, and storage and release mechanisms.

[0016] At the data layer: The actual operating data, historical data, and simulation data of each device and circuit in the data center energy system are collected through the data interface to drive the mechanism digital twin model, forming a mechanism-modeled and data-driven digital twin model of the data center energy system;

[0017] At the transport layer: Use data transmission to perform bidirectional mapping between the physical entities of the data center energy system and the digital twin model data;

[0018] In the model training layer: the historical data of the digital twin model of the data center energy system is used as training data, and the machine learning algorithm optimized by the intelligent optimization algorithm is used for model training. The accuracy of the digital twin model is evaluated using the model's predicted values and actual values.

[0019] Furthermore, the S2 specifically includes:

[0020] The server in the IT equipment subsystem is taken as the key energy consumption device, and an IT equipment energy consumption model is established, which is expressed as:

[0021]

[0022] is the energy consumption of the i-th server in period t; P b is the energy consumption of the server in the CPU idle state; i is the energy consumption coefficient of the i-th server; f i,t is the CPU operating frequency of the i-th server in period t; η i,t is the workload of the i-th server in period t; μ i,t is the service rate of the i-th server in period t; P s is the energy consumption of the server in sleep state;

[0023] Real-time monitoring of IT equipment energy consumption data to obtain energy consumption trends. Data centers are divided into low-energy consumption zones, medium-energy consumption zones, and high-energy consumption zones based on energy consumption intervals and layout information, combining the location and layout of IT equipment and the supply and return air paths of terminal precision air conditioning units. Each zone is dynamically divided based on the energy consumption trends of IT equipment, and IT equipment in the same zone is cooled by the same group of terminal precision air conditioning units.

[0024] In each energy consumption zone, select representative locations based on the scale and space of IT equipment to install a corresponding number of temperature collection devices to monitor the ambient temperature data of each energy consumption zone.

[0025] Furthermore, in S3, based on the ambient temperature data of each energy consumption zone and the operating data and meteorological data of each device in the zone, a cooling demand prediction model for each time zone of the data center is established, including:

[0026] Obtain the server inlet temperature, server outlet temperature, server energy consumption, server fan speed, server workload, processor load, temperature, memory ratio, disk utilization, and device distribution density within each energy consumption zone. Combined with the ambient temperature data of each energy consumption zone, the external temperature and humidity of the data center, and considering the sudden interference of extreme events, equipment failures, and environmental fluctuations in the data center, establish a data sample that affects cooling demand.

[0027] After extracting important features that affect cooling demand from data samples, prompt words are constructed and the data is textualized using natural language. At the same time, a knowledge distillation mechanism is used to extract important knowledge information to guide cooling demand prediction.

[0028] After cross-modal alignment of the time series data and text data in the cooling demand data samples, they are input into the pre-trained large language model for fine-tuning training, so that the model can understand the data knowledge in the cooling demand scenario and establish a cooling demand prediction model for each time period partition of the data center.

[0029] Furthermore, in S3, the digital twin model of the data center energy system is used to analyze the process mechanism of the refrigeration subsystem, and the target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem in each time period and the target energy efficiency index PUE are calculated, including:

[0030] The digital twin model of the data center energy system is used to simulate the collaborative cooling process mechanism of the refrigeration equipment, cold storage equipment, and terminal precision air-conditioning units under different cooling demand conditions of the refrigeration subsystem to meet the cooling demand of each time zone.

[0031] Analyze the impact of adjusting the cooling water pump flow on the chilled water temperature and the supply air temperature, and set the optimal chilled water temperature and supply air temperature combination ranges for the cooling subsystem under the cooling demand of each zone at different times. Also analyze the impact of chilled water temperature on the operating status and energy consumption of the cooling subsystem, and the impact of supply air temperature on the operating status and energy consumption of the IT equipment subsystem, and set the energy efficiency indicator (PUE) range for the cooling subsystem under the chilled water temperature and supply air temperature combination ranges for each zone.

[0032] Setting the optimization objective function, including minimizing the energy consumption of the refrigeration subsystem, maximizing the energy efficiency ratio of the refrigeration equipment, and optimizing the ambient temperature of each zone;

[0033] An intelligent optimization algorithm is used to find the optimal target combination parameters of chilled water temperature and supply air temperature, as well as the target energy efficiency index PUE for each zone of the refrigeration subsystem at each time period.

[0034] Furthermore, in S4, a data center energy efficiency indicator PUE prediction model is established, including:

[0035] The model input data includes chilled water supply temperature, cooling water supply temperature, chilled water pump flow, cooling water pump flow, cooling tower fan control signal, cold storage capacity of cold storage equipment, ambient temperature data of each partition, terminal precision air conditioning unit control signal, supply air temperature, data center external air temperature and humidity, IT equipment subsystem operating parameters, load change rate, and the current data center energy efficiency index PUE; the model output data includes the data center energy efficiency index PUE at the next moment;

[0036] The attention mechanism in the Transformer algorithm is used to mine the dependencies at any position in the data sequence and perform weighted aggregation of information at different time steps. The GRU algorithm is used to capture short-term dynamic changes in data and adapt to local time series characteristics to establish a data center energy efficiency indicator PUE prediction model.

[0037] Furthermore, in S4, an optimization scheduling model for the upper refrigeration subsystem is established, including:

[0038] Calculate the differences in operating characteristics of each refrigeration equipment in the refrigeration subsystem under different operating parameters, including the differences in the COP of each chiller under different cooling capacity allocation results, and the differences in the efficiency of the cooling tower, chilled water pump, and cooling water pump under different weather conditions and refrigeration operating parameters;

[0039] Set up the operation mechanism of the cold storage device in the refrigeration subsystem: during the period of low electricity price of the power supply subsystem, the excess cooling energy produced by the refrigeration subsystem is stored in the cold storage device; when the cooling energy produced by the refrigeration subsystem exceeds the required cooling energy, the excess cooling energy is stored in the cold storage device; and during the peak cooling demand period or high electricity price period, the cooling energy is released through the cold storage device;

[0040] According to the operation mechanism of the cold storage equipment, predict the real-time cold storage capacity of the cold storage equipment;

[0041] Taking into account the differences in the operating characteristics of each refrigeration equipment under different operating conditions and the real-time cold storage capacity of the cold storage equipment, the different combined operating modes of each equipment in the refrigeration subsystem and the control mechanisms of the cooling water circuit, chilled water circuit, supply air circuit, and return air circuit under different combined operating modes are analyzed;

[0042] The goals are to minimize the deviation between the predicted energy efficiency index PUE and the target energy efficiency index PUE, minimize the deviation between the actual combination parameters of chilled water temperature and supply air temperature and the target combination parameters, and minimize the deviation between the supply and demand of cooling capacity in each zone. It is expressed as:

[0043]

[0044] f1 is the minimum deviation target of energy efficiency index PUE; T is the scheduling period; PUE pre,t 、PUE tar,t are the predicted value and target value of the energy efficiency index PUE in time period t; f2 is the minimum target deviation of the combined parameters of chilled water temperature and supply air temperature; N is the number of partitions; T chi,act,i,t 、T chi,tar,i,t are the actual temperature and target temperature of the chilled water in the ith partition during period t; T w,act,i,t 、T w,tar,i,t are the actual supply air temperature and target temperature of the ith partition during period t; f3 is the minimum target for the cooling capacity supply and demand deviation of each partition; Q i,sup,t , Q i,dem,t are the cooling supply and demand values of the ith partition in period t, respectively;

[0045] Set the operating constraints of the refrigeration subsystem, including at least: operating constraints of the chiller, cooling water pump, cooling tower, chilled water pump and precision air conditioning unit, cooling supply and demand balance constraints, and capacity constraints of the cold storage equipment;

[0046] According to the set refrigeration subsystem objectives and operation constraints, an upper-level refrigeration subsystem optimization scheduling model is established.

[0047] Furthermore, in S5, with the goal of minimizing the system power operation cost and IT equipment load migration cost, a lower-level data center optimization scheduling model is established, which is expressed as:

[0048]

[0049] C is the minimum target for system power operation cost and IT equipment load migration cost; M is the number of equipment in the cooling subsystem and IT equipment subsystem; P j,t is the operating power consumption of the jth device in period t; is the electricity price during period t; P lm,i,t is the load migration amount of the i-th partition in time period t; ε i,t is the unit load migration cost of the i-th partition in time period t;

[0050] The constraints of the lower-level data center optimization scheduling model are set to at least include: computing load constraints of each partition in the IT equipment subsystem, equipment operation constraints of the cooling subsystem, and power constraints.

[0051] Furthermore, in S5, the optimal scheduling strategy of the system is obtained by interactively solving the two-layer optimization scheduling model, including:

[0052] Solve the optimization scheduling model of the upper-level refrigeration subsystem and output the start-stop combination status and operating power of the refrigeration equipment in the refrigeration subsystem, the storage and discharge status and storage and discharge capacity of the cold storage equipment, and the control strategy of each branch circuit in the terminal precision air-conditioning unit;

[0053] Solve the optimization scheduling model of the lower-level data center and output the load migration amount of each energy consumption zone, the power load strategy of the cooling equipment and IT equipment subsystem in the cooling subsystem;

[0054] Through the interactive influence of the output strategies of the upper and lower optimization scheduling models, an iterative solution is performed to obtain the optimal scheduling strategy for the system.

[0055] Furthermore, after S5, it also includes: considering the hysteresis of the ambient temperature response of the IT equipment subsystem of the data center, selecting the working condition in which the ambient temperature changes after the refrigeration subsystem adjusts the chilled water temperature and the supply air temperature as the identification object, establishing an ambient temperature response lag time prediction model, and adjusting the refrigeration subsystem in advance according to the lag time.

[0056] The beneficial effects of the present invention are:

[0057] (1) The present invention establishes a digital twin model of the data center energy system, including the power supply subsystem, refrigeration subsystem, and IT equipment subsystem. It can map the actual data center energy system into a virtual space, facilitating the subsequent establishment of model predictions and simulation of optimized scheduling strategies.

[0058] (2) The present invention monitors the energy consumption data of each device in the IT equipment subsystem, and divides the data center into multiple energy consumption zones based on the location layout of each device and the air supply circuit layout of the terminal precision air-conditioning unit. A temperature acquisition device is installed in each energy consumption zone to obtain the ambient temperature data of each energy consumption zone in real time. This enables subsequent precise cooling adjustment based on the energy consumption zone, thereby saving energy consumption of the cooling equipment.

[0059] (3) The present invention establishes a cooling demand prediction model for each time period of the data center based on the ambient temperature data of each energy consumption zone and the operating data and meteorological data of each device in the zone, and then uses the digital twin model of the data center energy system to analyze the process mechanism of the refrigeration subsystem, and calculates the target combination parameters of the chilled water temperature and supply air temperature and the target energy efficiency index PUE of the refrigeration subsystem in each time period; by accurately predicting the cooling demand of each zone, the operation of the refrigeration subsystem can be subsequently optimized to maximize the balance between supply and demand of cooling capacity and avoid unnecessary energy consumption, and optimize the calculation of the target combination parameters of the chilled water temperature and supply air temperature and the target energy efficiency index PUE. The process mechanism analysis can be performed from the cold end to the refrigeration end of the next level to achieve the matching of cooling demand and temperature target combination parameters and PUE, further reduce system energy consumption, and improve overall energy efficiency;

[0060] (4) The present invention establishes a data center energy efficiency index PUE prediction model, and at the same time considers the operating characteristic differences of each device in the refrigeration subsystem, the cold storage capacity and the control mechanism of each refrigeration circuit, with the goal of minimizing the deviation between the energy efficiency index PUE prediction value and the target energy efficiency index PUE, minimizing the deviation between the actual combination parameters of the chilled water temperature and the supply air temperature and the target combination parameters, and minimizing the deviation between the supply and demand of the cooling capacity of each zone, to establish an upper-level refrigeration subsystem optimization scheduling model; with the goal of minimizing the system electricity operation cost and the IT equipment load migration cost, to establish a lower-level data center optimization scheduling model, and obtain the optimal scheduling strategy of the system through the interactive solution of the two-layer optimization scheduling model; on the one hand, by predicting the energy efficiency index PUE and establishing the upper-level optimization scheduling model, it is possible to discover parameter deviation problems in advance and take measures. Corresponding optimization measures are taken to ensure that the predicted value of the energy efficiency index PUE is close to the target energy efficiency PUE value, and the temperature combination parameters can meet the target parameters, ensuring the balance of cooling supply and demand in each zone, reducing the energy consumption of the refrigeration subsystem and improving the accuracy of equipment regulation. At the same time, it can give full play to the differences in the operating characteristics of each equipment to improve the energy efficiency of the equipment, and combine the cold storage capacity to achieve peak shaving and valley filling of cooling demand, thereby improving the flexibility and reliability of system scheduling; on the other hand, a lower-level optimization scheduling model is established with cost as the target to regulate the power load and reduce the operating cost of the data center as much as possible, and consider the mutual influence of the scheduling strategies of the two-layer scheduling model, conduct iterative interaction, obtain the optimal scheduling strategy, and ensure the accuracy of the scheduling strategy as a whole, improve the economic benefits of the data center and reduce operating energy consumption.

[0061] Other features and advantages will be described in the following description, and in part will become apparent from the description, or understood by practicing the invention. The purpose and other advantages of the invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 This is a flow chart of a method for energy-saving and economical scheduling of a data center energy system based on a prediction model according to the present invention;

[0065] Figure 2 This is a schematic diagram of the operating structure of the data center energy system of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] like Figure 1 As shown, this embodiment provides a method for energy-saving and economical scheduling of a data center energy system based on a prediction model, which includes:

[0068] S1. Establish a digital twin model of the data center energy system, including the power supply subsystem, refrigeration subsystem, and IT equipment subsystem; the refrigeration subsystem includes refrigeration equipment, cold storage equipment, and terminal precision air conditioning units;

[0069] S2. Monitor the energy consumption data of each device in the IT equipment subsystem. Based on the location layout of each device and the air supply circuit layout of the terminal precision air conditioning units, divide the data center into multiple energy consumption zones. Install a temperature collection device in each energy consumption zone to obtain real-time ambient temperature data for each energy consumption zone.

[0070] S3. Based on the ambient temperature data of each energy consumption zone and the operating and meteorological data of each device within the zone, a cooling demand forecast model for each time zone of the data center is established. The process mechanism of the refrigeration subsystem is analyzed using the digital twin model of the data center energy system. The target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem at each time zone, as well as the target energy efficiency index (PUE), are calculated.

[0071] S4. Establish a data center energy efficiency indicator (PUE) prediction model, taking into account the differences in the operating characteristics of each device in the refrigeration subsystem, the cold storage capacity, and the control mechanism of each refrigeration circuit. With the goal of minimizing the deviation between the predicted PUE value and the target PUE value, minimizing the deviation between the actual combination of chilled water temperature and supply air temperature and the target combination, and minimizing the deviation between the supply and demand of cooling capacity in each zone, establish an optimization scheduling model for the upper refrigeration subsystem.

[0072] S5. With the goal of minimizing the system power operation cost and IT equipment load migration cost, a lower-level data center optimization scheduling model is established. Through the interactive solution of the two-level optimization scheduling model, the optimal scheduling strategy of the system is obtained.

[0073] like Figure 2 As shown, the operating structure of the data center energy system includes: power supply subsystem, refrigeration subsystem and IT equipment subsystem. The power supply subsystem includes traditional generator sets, wind turbine generator sets, photovoltaic generator sets and municipal power grid, which are used to provide electricity to the refrigeration subsystem and IT equipment subsystem. There are multiple IT devices in the IT equipment subsystem, which are divided into multiple energy consumption zones. This figure only illustrates one energy consumption zone, which adopts the down-air supply and up-air return method. The refrigeration subsystem includes cooling towers, chillers, cooling water pumps, chilled water pumps, cold storage equipment and precision air conditioners. Among them, the chilled water pump pressurizes low-temperature chilled water and sends it to the cold water pipe of the precision air conditioner. The cold air is blown to the energy consumption zone through the fan coil to exchange heat. The hot air returns to the air and heats the chilled water. After the chilled water temperature rises, it returns to the chiller; the chiller is mainly composed of a condenser, an evaporator, a compressor and an expansion valve. After the low-temperature and low-pressure refrigerant enters the evaporator, it exchanges heat with the chilled water. The chilled water temperature drops and the refrigerant absorbs heat and turns into high-temperature and low-pressure hot steam. The delayed low-pressure gaseous refrigerant is pressurized by the compressor and enters the condenser. During the condensation process, the refrigerant releases heat energy and turns into a high-pressure liquid. After passing through the expansion valve, it becomes a low-temperature, low-pressure gas-liquid mixture and returns to the evaporator; the low-temperature cooling water heats up after absorbing the heat released by the refrigerant and enters the cooling tower, and discharges the heat into the atmosphere after heat exchange with the air; the cold storage equipment is used to store the excess cold produced by the chiller and release the stored cold when the chiller is insufficient. In actual application, a cold storage pump and a cold release pump will be provided to assist in the storage and discharge of cold.

[0074] This diagram only illustrates a certain chiller. During actual operation, the refrigeration subsystem usually includes multiple chillers with different performance parameters. The chilled water output by the chiller will be distributed to each water supply branch through the water distributor. These branches will then send the chilled water to the precision air-conditioning units in each partition. The air supply temperature will be adjusted by the air-conditioning units. The chilled water flowing back from each return branch will be collected by the water collector and then sent back to the chiller for further cooling.

[0075] In this embodiment, in S1, establishing a digital twin model of the data center energy system includes:

[0076] At the physical layer: construct the physical entity of the data center energy system based on the model, size, material and layout position of each device in the power supply subsystem, refrigeration subsystem and IT equipment subsystem; the power supply subsystem includes new energy generator sets, traditional conventional power supply units and municipal power grids; the refrigeration equipment in the refrigeration subsystem includes chillers, cooling water pumps, cooling towers and chilled water pumps, and there are multiple of them; the cold storage equipment in the refrigeration subsystem is connected to the refrigeration equipment to store excess cold and release it when cold is lost; the number of the terminal precision air-conditioning units is arranged according to the equipment scale of the IT equipment subsystem, and corresponding supply air circuits and return air circuits are provided; the IT equipment subsystem includes servers, storage devices and network equipment;

[0077] At the twin layer: Introduce the corresponding physical attribute information, behavior information, and rule information between devices into the physical entity of the data center energy system to establish a digital twin model of the data center energy system mechanism. The physical attribute information includes the rated operating parameters and performance parameters of the equipment; the behavior information includes the operating status and operation behavior of the equipment; the rule information includes the start and stop sequence of the equipment, the input and output coupling relationship between the equipment, and the power supply, cooling, and storage and release mechanisms.

[0078] At the data layer: The actual operating data, historical data, and simulation data of each device and circuit in the data center energy system are collected through the data interface to drive the mechanism digital twin model, forming a mechanism-modeled and data-driven digital twin model of the data center energy system;

[0079] At the transport layer: Use data transmission to perform bidirectional mapping between the physical entities of the data center energy system and the digital twin model data;

[0080] In the model training layer: the historical data of the digital twin model of the data center energy system is used as training data, and the machine learning algorithm optimized by the intelligent optimization algorithm is used for model training. The accuracy of the digital twin model is evaluated using the model's predicted values and actual values.

[0081] The efficiency of a cooling tower during operation is expressed as:

[0082]

[0083] NTU ct is the number of heat transfer units in the cooling tower; m is the heat capacity ratio of inlet and outlet water and air;

[0084] The total heat transfer capacity of the cooling tower is expressed as:

[0085] Q ct =(h a,o -h a,i )×(f CTA ×ρ air );

[0086] h a,o 、h a,i are the inlet and outlet enthalpies of the cooling tower air respectively; f CTA is the cooling tower air flow rate; ρ air is the air mass density; the cooling tower fan power is expressed as:

[0087]

[0088] f n is the speed of the nth fan; p max,n is the maximum power of the nth wind turbine; W is the number of wind turbines;

[0089] The operating COP and cooling capacity of the chiller are expressed as:

[0090] COP=COP rated ×COP ratio ;

[0091] Q C =Q C_rated ×C ratio ;

[0092] COP is the operating COP; COP rated is the rated COP; COP ratio is the ratio of operating COP to rated COP; Q C is the operating cooling capacity; Q C_rated is the rated cooling capacity; C ratio It is the ratio of operating cooling capacity to rated cooling capacity;

[0093] The load of the chiller is expressed as:

[0094] Q load =m chw ×p chw (T chw,in -T chw,set );

[0095] Q load is the chiller load; m chw is the mass flow rate of chilled water; p chw is the specific heat of chilled water at constant pressure; T chw,in is the chilled water inlet temperature; T chw,set Set the supply temperature for chilled water;

[0096] The power consumption of the chiller is expressed as:

[0097]

[0098] P c is the power consumption of the chiller; FFLP is the full load power fraction;

[0099] The power of cooling water pump and chilled water pump is expressed as:

[0100] P=P rated (a0+a1γ+a2γ 2 +a3γ 3 +…);

[0101] P rated is the rated power of the water pump; γ is the control signal, ranging from 0 to 1; a0, a1, a2, a3 and other polynomials are obtained through water pump parameter identification;

[0102] The terminal precision air conditioning unit achieves a balance between the set air outlet temperature and the average temperature of the liquid in the coil, and can adjust the supply air temperature to the set temperature through the coil valve; the model is expressed as:

[0103] Q f =m a (h a,in -h a,out )-m cond h cond ;

[0104] Q f is the air outlet temperature; m a is the return air mass flow rate; h a,in is the return air enthalpy, h a,out is the outlet air enthalpy; m cond is the mass flow rate of condensed water; h cond is the enthalpy of condensed water.

[0105] In this embodiment, S2 specifically includes:

[0106] The server in the IT equipment subsystem is taken as the key energy consumption device, and an IT equipment energy consumption model is established, which is expressed as:

[0107]

[0108] is the energy consumption of the i-th server in period t; P b is the energy consumption of the server in the CPU idle state; i is the energy consumption coefficient of the i-th server; f i,t is the CPU operating frequency of the i-th server in period t; η i,t is the workload of the i-th server in period t; μ i,tis the service rate of the i-th server in period t; P s is the energy consumption of the server in sleep state;

[0109] Real-time monitoring of IT equipment energy consumption data to obtain energy consumption trends. Data centers are divided into low-energy consumption zones, medium-energy consumption zones, and high-energy consumption zones based on energy consumption intervals and layout information, combining the location and layout of IT equipment and the supply and return air paths of terminal precision air conditioning units. Each zone is dynamically divided based on the energy consumption trends of IT equipment, and IT equipment in the same zone is cooled by the same group of terminal precision air conditioning units.

[0110] In each energy consumption zone, select representative locations based on the scale and space of IT equipment to install a corresponding number of temperature collection devices to monitor the ambient temperature data of each energy consumption zone.

[0111] It should be noted that the energy consumption zones are not fixed, but change dynamically according to the workload, number of servers, density, etc. Moreover, as a computing workload, the workload can be transferred in time and space. For example, some adjustable workloads can be postponed from peak load periods to low load periods, or part of the load can be migrated to other servers for processing.

[0112] In this embodiment, in S3, based on the ambient temperature data of each energy consumption zone and the operating data and meteorological data of each device in the zone, a cooling demand prediction model for each time zone of the data center is established, including:

[0113] Obtain the server inlet temperature, server outlet temperature, server energy consumption, server fan speed, server workload, processor load, temperature, memory ratio, disk utilization, and device distribution density within each energy consumption zone. Combined with the ambient temperature data of each energy consumption zone, the external temperature and humidity of the data center, and considering the sudden interference of extreme events, equipment failures, and environmental fluctuations in the data center, establish a data sample that affects cooling demand.

[0114] After extracting important features that affect cooling demand from data samples, prompt words are constructed and the data is textualized using natural language. At the same time, a knowledge distillation mechanism is used to extract important knowledge information to guide cooling demand prediction.

[0115] After cross-modal alignment of the time series data and text data in the cooling demand data samples, they are input into the pre-trained large language model for fine-tuning training, so that the model can understand the data knowledge in the cooling demand scenario and establish a cooling demand prediction model for each time period partition of the data center.

[0116] The cooling demand for data centers across different time periods and zones is subject to interference from operational data, external environmental data, and unexpected events. These factors exhibit complex spatiotemporal dependencies, making traditional models ill-equipped to handle the interference of real-time data updates and unexpected events. In contrast, the LLM large language model (LLM) boasts powerful multimodal data processing and deep reasoning capabilities, integrating heterogeneous data from multiple sources and capturing complex patterns within the data through adaptive learning. Through large-scale data training, the LLM effectively handles nonlinear relationships and spatiotemporal patterns, adapting to the dynamic changes in data center cooling demand. Furthermore, LLM possesses powerful computing capabilities, enabling efficient processing of large-scale data and generating real-time cooling demand forecasts. Through distributed computing and parallel processing, the LLM can meet cooling demand across different timescales and support cooling forecasting tasks for large-scale data centers, demonstrating excellent scalability and flexibility.

[0117] By introducing prompt learning and integrating it with the specific characteristics of the cooling demand forecasting task, a prompt template is designed to add additional context to the input time series. Task instructions and data statistics of the input sequence are provided in natural language, thereby enhancing the LLM's understanding of the input sequence and guiding the inference and prediction process. The primary task in prompt design is to accurately define the prediction objective and task requirements. Typically, the task objective involves predicting a value at a certain point in the future or analyzing trends over a period of time based on historical data. When designing prompts, this task objective must be clearly communicated to ensure the model correctly understands the required output format and prediction type. To further enhance the model's learning capabilities, prompt design can also include a description of the statistical characteristics of the time series data segment (such as the maximum, minimum, mean, and variance). These statistical parameters help the model better identify trends, volatility, and potential cyclical characteristics in the data, thereby improving its ability to capture time series patterns. This approach enables the model to not only predict specific values but also better understand the distribution patterns and changing trends of the data, providing a more accurate basis for future predictions.

[0118] Since the large language model is pre-trained on a large-scale dataset and has rich knowledge, there are significant modal differences between the textual knowledge and the running time series of the large language model in the data center. A knowledge distillation strategy is proposed to extract the main knowledge information to guide the cooling demand forecast. Usually, semantically similar words are clustered as synonyms. The cluster center is used to represent the surrounding words to reduce the number of word units.

[0119] Cross-modal alignment of time series and text data is crucial for multimodal learning. Time series data reflects the dynamics of data center energy systems, while text data provides a wealth of additional information. Cross-modal alignment allows for the organic fusion of these two modalities, enhancing data complementarity and improving the model's predictive accuracy and decision support capabilities. Common alignment methods include feature extraction and shared space mapping, associative learning based on self-attention mechanisms, multimodal joint training, and end-to-end learning. Feature extraction methods process time series and text data separately, extracting their respective features using deep learning models, and mapping them into a shared space. Self-attention mechanisms establish dynamic correlations between time series and text data, automatically adjusting attention weights to better capture the underlying relationships between the two modalities. Multi-task learning methods achieve collaborative learning by jointly optimizing time series and text tasks.

[0120] The large language model used is the LLaMA model, and the fine-tuning training method includes:

[0121] 1) Initialize model parameters before model training begins;

[0122] 2) During each training iteration, the training data is first received and the cooling demand forecast is calculated through forward propagation, which involves passing the input data through each layer of the model until the output layer;

[0123] 3) After the output layer outputs the cooling demand forecast, the model calculates the deviation between the forecast and the actual value, which is called loss. Here, the loss function is represented by cross entropy.

[0124] 4) The model calculates the gradient of the loss with respect to each parameter using the backpropagation algorithm, starting from the output layer and going backwards through each layer of the model, calculating the partial derivative of the loss with respect to each parameter;

[0125] 5) Use intelligent optimization algorithms to update model parameters to reduce losses;

[0126] 6) During training, periodically use the validation set to check the performance of the model.

[0127] In this embodiment, in S3, the process mechanism of the refrigeration subsystem is analyzed using the digital twin model of the data center energy system to calculate the target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem in each time period and the target energy efficiency index PUE, including:

[0128] The digital twin model of the data center energy system is used to simulate the collaborative cooling process mechanism of the refrigeration equipment, cold storage equipment, and terminal precision air-conditioning units under different cooling demand conditions of the refrigeration subsystem to meet the cooling demand of each time zone.

[0129] Analyze the impact of adjusting the cooling water pump flow on the chilled water temperature and the supply air temperature, and set the optimal chilled water temperature and supply air temperature combination ranges for the cooling subsystem under the cooling demand of each zone at different times. Also analyze the impact of chilled water temperature on the operating status and energy consumption of the cooling subsystem, and the impact of supply air temperature on the operating status and energy consumption of the IT equipment subsystem, and set the energy efficiency indicator (PUE) range for the cooling subsystem under the chilled water temperature and supply air temperature combination ranges for each zone.

[0130] Setting the optimization objective function, including minimizing the energy consumption of the refrigeration subsystem, maximizing the energy efficiency ratio of the refrigeration equipment, and optimizing the ambient temperature of each zone;

[0131] An intelligent optimization algorithm is used to find the optimal target combination parameters of chilled water temperature and supply air temperature, as well as the target energy efficiency index PUE for each zone of the refrigeration subsystem at each time period.

[0132] The data center energy efficiency indicator PUE refers to the ratio of the total power consumed by the data center to the power consumed by the IT equipment subsystem. It is used to evaluate the power efficiency of the data center and is expressed as:

[0133]

[0134] E IT Power consumption of IT equipment subsystem; E AC Power consumption for the refrigeration subsystem; E other Power consumption for other electrical equipment in the data center.

[0135] In this embodiment, in S4, establishing a data center energy efficiency indicator PUE prediction model includes:

[0136] The model input data includes chilled water supply temperature, cooling water supply temperature, chilled water pump flow, cooling water pump flow, cooling tower fan control signal, cold storage capacity of cold storage equipment, ambient temperature data of each partition, terminal precision air conditioning unit control signal, supply air temperature, data center external air temperature and humidity, IT equipment subsystem operating parameters, load change rate, and the current data center energy efficiency index PUE; the model output data includes the data center energy efficiency index PUE at the next moment;

[0137] The attention mechanism in the Transformer algorithm is used to mine the dependencies at any position in the data sequence and perform weighted aggregation of information at different time steps. The GRU algorithm is used to capture short-term dynamic changes in data and adapt to local time series characteristics to establish a data center energy efficiency indicator PUE prediction model.

[0138] It should be noted that the core of the Transformer algorithm is the multi-head attention mechanism. The calculation process of the self-attention mechanism includes:

[0139] Input transformation: The input sequence X is transformed through three different linear transformations to generate query (Q), key (K) and value (V) vectors, respectively, which are expressed as:

[0140]

[0141] W Q 、W K 、W V are the corresponding weight matrices respectively;

[0142] Attention score calculation: Calculate the attention score by performing the dot product operation of Q and K, and scale it to stabilize the gradient;

[0143] Weight distribution: Use the softmax function to convert the attention score into a probability distribution as the weight of each value vector;

[0144] Weighted summation: Multiply the weight by V to obtain the final attention output, which is expressed as:

[0145]

[0146] d k is the dimension of the key vector.

[0147] The multi-head attention mechanism maps the Q, K, and V vectors to multiple subspaces respectively, calculates attention independently in each subspace, and finally concatenates the results and changes them linearly, so that the model can simultaneously focus on different features and positions of the input data, which can be expressed as:

[0148] MultiHead(Q,K,V)=Concat(head1,head2,…,head N )W 0 ;

[0149] head i =Attention(QW i Q ,KW i K ,VW i V );

[0150] W 0 is the parameter matrix of the output projection; head i is the self-attention distribution of the i-th head; W i Q 、W i K and W i V is the linear projection parameter matrix of the i-th head;

[0151] The core of the GRU algorithm is to introduce a gating mechanism, including an update gate and a reset gate, which can selectively forget or update certain information. The calculation formula is expressed as:

[0152] r t =σ(W r h t-1 +U r x t );

[0153] z t =σ(W z x t +U z h t-1 );

[0154]

[0155] r t is the reset gate; σ is the sigmoid activation function; x t is the input at time t; z t is the update gate; h t-1 is the hidden state at the previous moment; is the candidate hidden state; tanh is the tanh activation function; h t is the output at time t; W r 、U r 、W z 、U z 、W h 、U h is the weight matrix.

[0156] In actual applications, the input data of the data center energy efficiency indicator PUE prediction model is fed into both the Transformer and GRU networks for processing. The outputs of the two networks are then concatenated along the depth dimension through a concatenation module and finally fed into a fully connected layer to output the final PUE prediction result. The attention mechanism in the Transformer algorithm enables weighted processing of different positions in the data sequence, achieving better data sequence modeling capabilities. This allows the prediction model to focus more on the important information of the input features and to jointly attend to information from different representation subspaces at different positions. The GRU algorithm also enables the network to selectively forget or update some information, which can greatly improve the performance of the prediction model.

[0157] In this embodiment, in S4, an optimization scheduling model for the upper refrigeration subsystem is established, including:

[0158] Calculate the differences in operating characteristics of each refrigeration equipment in the refrigeration subsystem under different operating parameters, including the differences in the COP of each chiller under different cooling capacity allocation results, and the differences in the efficiency of the cooling tower, chilled water pump, and cooling water pump under different weather conditions and refrigeration operating parameters;

[0159] Set up the operation mechanism of the cold storage device in the refrigeration subsystem: during the period of low electricity price of the power supply subsystem, the excess cooling energy produced by the refrigeration subsystem is stored in the cold storage device; when the cooling energy produced by the refrigeration subsystem exceeds the required cooling energy, the excess cooling energy is stored in the cold storage device; and during the peak cooling demand period or high electricity price period, the cooling energy is released through the cold storage device;

[0160] According to the operation mechanism of the cold storage equipment, predict the real-time cold storage capacity of the cold storage equipment;

[0161] Taking into account the differences in the operating characteristics of each refrigeration equipment under different operating conditions and the real-time cold storage capacity of the cold storage equipment, the different combined operating modes of each equipment in the refrigeration subsystem and the control mechanisms of the cooling water circuit, chilled water circuit, supply air circuit, and return air circuit under different combined operating modes are analyzed;

[0162] The goals are to minimize the deviation between the predicted energy efficiency index PUE and the target energy efficiency index PUE, minimize the deviation between the actual combination parameters of chilled water temperature and supply air temperature and the target combination parameters, and minimize the deviation between the supply and demand of cooling capacity in each zone. It is expressed as:

[0163]

[0164] f1 is the minimum deviation target of energy efficiency index PUE; T is the scheduling period; PUE pre,t 、PUE tar,t are the predicted value and target value of the energy efficiency index PUE in time period t; f2 is the minimum target deviation of the combined parameters of chilled water temperature and supply air temperature; N is the number of partitions; T chi,act,i,t 、T chi,tar,i,t are the actual temperature and target temperature of the chilled water in the ith partition during period t; T w,act,i,t 、T w,tar,i,t are the actual supply air temperature and target temperature of the ith partition during period t; f3 is the minimum target for the cooling capacity supply and demand deviation of each partition; Q i,sup,t , Q i,dem,t are the cooling supply and demand values of the ith partition in period t, respectively;

[0165] Set the operating constraints of the refrigeration subsystem, including at least: operating constraints of the chiller, cooling water pump, cooling tower, chilled water pump and precision air conditioning unit, cooling supply and demand balance constraints, and capacity constraints of the cold storage equipment;

[0166] According to the set refrigeration subsystem objectives and operation constraints, an upper-level refrigeration subsystem optimization scheduling model is established.

[0167] In this embodiment, in S5, with the goal of minimizing the system power operation cost and IT equipment load migration cost, a lower-level data center optimization scheduling model is established, which is expressed as:

[0168]

[0169] C is the minimum target for system power operation cost and IT equipment load migration cost; M is the number of equipment in the cooling subsystem and IT equipment subsystem; P j,t is the operating power consumption of the jth device in period t; is the electricity price during period t; P lm,i,t is the load migration amount of the i-th partition in time period t; ε i,t is the unit load migration cost of the i-th partition in time period t;

[0170] The constraints of the lower-level data center optimization scheduling model are set to at least include: computing load constraints of each partition in the IT equipment subsystem, equipment operation constraints of the cooling subsystem, and power constraints.

[0171] In this embodiment, in S5, the optimal scheduling strategy of the system is obtained by interactively solving the two-layer optimization scheduling model, including:

[0172] Solve the optimization scheduling model of the upper-level refrigeration subsystem and output the start-stop combination status and operating power of the refrigeration equipment in the refrigeration subsystem, the storage and discharge status and storage and discharge capacity of the cold storage equipment, and the control strategy of each branch circuit in the terminal precision air-conditioning unit;

[0173] Solve the optimization scheduling model of the lower-level data center and output the load migration amount of each energy consumption zone, the power load strategy of the cooling equipment and IT equipment subsystem in the cooling subsystem;

[0174] Through the interactive influence of the output strategies of the upper and lower optimization scheduling models, an iterative solution is performed to obtain the optimal scheduling strategy for the system.

[0175] In this embodiment, after S5, the process further includes: considering the hysteresis of the ambient temperature response of the IT equipment subsystem of the data center, selecting the working condition in which the ambient temperature changes after the refrigeration subsystem regulates the chilled water temperature and the supply air temperature as the identification object, establishing an ambient temperature response lag time prediction model, and regulating the refrigeration subsystem in advance according to the lag time.

[0176] In actual applications, when establishing an ambient temperature response lag time prediction model, the ambient temperature response lag time identification algorithm process is first established, and then the ambient temperature response lag time prediction model is established using a machine learning algorithm. Specifically, it includes:

[0177] The identification object is the condition where the ambient temperature changes after the refrigeration subsystem controls the chilled water temperature and the supply air temperature.

[0178] The time window is divided into two equally spaced time windows and the time window is gradually slid forward. The average difference in ambient temperature between the two time windows is calculated. If the difference exceeds the set threshold, the moment is considered as the operating condition change point. Otherwise, the time window is continuously slid forward until the operating condition change point is known or the time window has slid to the cutoff time point.

[0179] Based on the operating condition change point and time window, the chilled water temperature and supply air temperature time series, as well as the ambient temperature time series, are obtained from the start of the operating condition change to the end of the time window.

[0180] The ambient temperature time series is moved forward step by step to obtain a new ambient temperature series and construct an ambient temperature response time lag matrix;

[0181] Calculate the correlation coefficient between the chilled water temperature and supply air temperature time series and the ambient temperature response time lag matrix, and take the lag time corresponding to the maximum correlation coefficient as the ambient temperature response lag time under this working condition;

[0182] The original operation data of the data center cooling system is obtained and substituted into the above-mentioned time identification algorithm process to perform ambient temperature lag identification, and establish the relationship between the lag time and the characteristics of each operation data; the correlation analysis method is used to perform correlation analysis to obtain the correlation coefficient between each operation data and the ambient temperature response lag time; the operation data features with high correlation coefficients are fused with the original operation data as a training sample set, and after model training and learning using a machine learning algorithm, an ambient temperature response lag time prediction model is established.

[0183] In addition, the functional modules in each embodiment of the present invention can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part. If the function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk, or an optical disk.

[0184] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A data center energy system energy-saving and economic scheduling method based on a prediction model, characterized in that: It includes: S1. Establish a digital twin model of the data center energy system, including the power supply subsystem, refrigeration subsystem, and IT equipment subsystem; the refrigeration subsystem includes refrigeration equipment, cold storage equipment, and terminal precision air conditioning units; S2. Monitor the energy consumption data of each device in the IT equipment subsystem. Based on the location layout of each device and the air supply circuit layout of the terminal precision air conditioning units, divide the data center into multiple energy consumption zones. Install a temperature collection device in each energy consumption zone to obtain real-time ambient temperature data for each energy consumption zone. S3. Based on the ambient temperature data of each energy consumption zone and the operating and meteorological data of each device within the zone, a cooling demand forecast model for each time zone of the data center is established. The process mechanism of the refrigeration subsystem is analyzed using the digital twin model of the data center energy system. The target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem at each time zone, as well as the target energy efficiency index (PUE), are calculated. S4. Establish a data center energy efficiency indicator (PUE) prediction model, taking into account the differences in the operating characteristics of each device in the refrigeration subsystem, the cold storage capacity, and the control mechanism of each refrigeration circuit. With the goal of minimizing the deviation between the predicted PUE value and the target PUE value, minimizing the deviation between the actual combination of chilled water temperature and supply air temperature and the target combination, and minimizing the deviation between the supply and demand of cooling capacity in each zone, establish an optimization scheduling model for the upper refrigeration subsystem. S5. With the goal of minimizing the system power operation cost and IT equipment load migration cost, a lower-level data center optimization scheduling model is established. Through the interactive solution of the two-level optimization scheduling model, the optimal scheduling strategy of the system is obtained.

2. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S1, the establishment of a digital twin model of the data center energy system includes: At the physical layer: construct the physical entity of the data center energy system based on the model, size, material and layout position of each device in the power supply subsystem, refrigeration subsystem and IT equipment subsystem; the power supply subsystem includes new energy generator sets, traditional conventional power supply units and municipal power grids; the refrigeration equipment in the refrigeration subsystem includes chillers, cooling water pumps, cooling towers and chilled water pumps, and there are multiple of them; the cold storage equipment in the refrigeration subsystem is connected to the refrigeration equipment to store excess cold and release it when cold is lost; the number of the terminal precision air-conditioning units is arranged according to the equipment scale of the IT equipment subsystem, and corresponding supply air circuits and return air circuits are provided; the IT equipment subsystem includes servers, storage devices and network equipment; At the twin layer: Introduce the corresponding physical attribute information, behavior information, and rule information between devices into the physical entity of the data center energy system to establish a digital twin model of the data center energy system mechanism. The physical attribute information includes the rated operating parameters and performance parameters of the equipment; the behavior information includes the operating status and operation behavior of the equipment; the rule information includes the start and stop sequence of the equipment, the input and output coupling relationship between the equipment, and the power supply, cooling, and storage and release mechanisms. At the data layer: The actual operating data, historical data, and simulation data of each device and circuit in the data center energy system are collected through the data interface to drive the mechanism digital twin model, forming a mechanism-modeled and data-driven digital twin model of the data center energy system; At the transport layer: Use data transmission to perform bidirectional mapping between the physical entities of the data center energy system and the digital twin model data; In the model training layer: the historical data of the digital twin model of the data center energy system is used as training data, and the machine learning algorithm optimized by the intelligent optimization algorithm is used for model training. The accuracy of the digital twin model is evaluated using the model's predicted values and actual values.

3. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: The S2 specifically includes: The server in the IT equipment subsystem is taken as the key energy consumption device, and an IT equipment energy consumption model is established, which is expressed as: is the energy consumption of the i-th server in period t; P b is the energy consumption of the server in the CPU idle state; i is the energy consumption coefficient of the i-th server; f i,t is the CPU operating frequency of the i-th server in period t; η i,t is the workload of the i-th server in period t; μ i,t is the service rate of the i-th server in period t; P s is the energy consumption of the server in sleep state; Real-time monitoring of IT equipment energy consumption data to obtain energy consumption trends. Data centers are divided into low-energy consumption zones, medium-energy consumption zones, and high-energy consumption zones based on energy consumption intervals and layout information, combining the location and layout of IT equipment and the supply and return air paths of terminal precision air conditioning units. Each zone is dynamically divided based on the energy consumption trends of IT equipment, and IT equipment in the same zone is cooled by the same group of terminal precision air conditioning units. In each energy consumption zone, select representative locations based on the scale and space of IT equipment to install a corresponding number of temperature collection devices to monitor the ambient temperature data of each energy consumption zone.

4. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S3, based on the ambient temperature data of each energy consumption zone and the operating data and meteorological data of each device in the zone, a cooling demand prediction model for each time zone of the data center is established, including: Obtain the server inlet temperature, server outlet temperature, server energy consumption, server fan speed, server workload, processor load, temperature, memory ratio, disk utilization, and device distribution density within each energy consumption zone. Combined with the ambient temperature data of each energy consumption zone, the external temperature and humidity of the data center, and considering the sudden interference of extreme events, equipment failures, and environmental fluctuations in the data center, establish a data sample that affects cooling demand. After extracting important features that affect cooling demand from data samples, prompt words are constructed and the data is textualized using natural language. At the same time, a knowledge distillation mechanism is used to extract important knowledge information to guide cooling demand prediction. After cross-modal alignment of the time series data and text data in the cooling demand data samples, they are input into the pre-trained large language model for fine-tuning training, so that the model can understand the data knowledge in the cooling demand scenario and establish a cooling demand prediction model for each time period partition of the data center.

5. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S3, the digital twin model of the data center energy system is used to analyze the process mechanism of the refrigeration subsystem, and the target combination parameters of the chilled water temperature and supply air temperature of the refrigeration subsystem in each time period and the target energy efficiency index PUE are calculated, including: The digital twin model of the data center energy system is used to simulate the collaborative cooling process mechanism of the refrigeration equipment, cold storage equipment, and terminal precision air-conditioning units under different cooling demand conditions of the refrigeration subsystem to meet the cooling demand of each time zone. Analyze the impact of adjusting the cooling water pump flow on the chilled water temperature and the supply air temperature, and set the optimal chilled water temperature and supply air temperature combination ranges for the cooling subsystem under the cooling demand of each zone at different times. Also analyze the impact of chilled water temperature on the operating status and energy consumption of the cooling subsystem, and the impact of supply air temperature on the operating status and energy consumption of the IT equipment subsystem, and set the energy efficiency indicator (PUE) range for the cooling subsystem under the chilled water temperature and supply air temperature combination ranges for each zone. Setting the optimization objective function, including minimizing the energy consumption of the refrigeration subsystem, maximizing the energy efficiency ratio of the refrigeration equipment, and optimizing the ambient temperature of each zone; An intelligent optimization algorithm is used to find the optimal target combination parameters of chilled water temperature and supply air temperature, as well as the target energy efficiency index PUE for each zone of the refrigeration subsystem at each time period.

6. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S4, a data center energy efficiency indicator PUE prediction model is established, including: The model input data includes chilled water supply temperature, cooling water supply temperature, chilled water pump flow, cooling water pump flow, cooling tower fan control signal, cold storage capacity of cold storage equipment, ambient temperature data of each partition, terminal precision air conditioning unit control signal, supply air temperature, data center external air temperature and humidity, IT equipment subsystem operating parameters, load change rate, and the current data center energy efficiency index PUE; the model output data includes the data center energy efficiency index PUE at the next moment; The attention mechanism in the Transformer algorithm is used to mine the dependencies at any position in the data sequence and perform weighted aggregation of information at different time steps. The GRU algorithm is used to capture short-term dynamic changes in data and adapt to local time series characteristics to establish a data center energy efficiency indicator PUE prediction model.

7. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S4, an optimization scheduling model for the upper refrigeration subsystem is established, including: Calculate the differences in operating characteristics of each refrigeration equipment in the refrigeration subsystem under different operating parameters, including the differences in the COP of each chiller under different cooling capacity allocation results, and the differences in the efficiency of the cooling tower, chilled water pump, and cooling water pump under different weather conditions and refrigeration operating parameters; Set up the operation mechanism of the cold storage device in the refrigeration subsystem: during the period of low electricity price of the power supply subsystem, the excess cooling energy produced by the refrigeration subsystem is stored in the cold storage device; when the cooling energy produced by the refrigeration subsystem exceeds the required cooling energy, the excess cooling energy is stored in the cold storage device; and during the peak cooling demand period or high electricity price period, the cooling energy is released through the cold storage device; According to the operation mechanism of the cold storage equipment, predict the real-time cold storage capacity of the cold storage equipment; Taking into account the differences in the operating characteristics of each refrigeration equipment under different operating conditions and the real-time cold storage capacity of the cold storage equipment, the different combined operating modes of each equipment in the refrigeration subsystem and the control mechanisms of the cooling water circuit, chilled water circuit, supply air circuit, and return air circuit under different combined operating modes are analyzed; The goals are to minimize the deviation between the predicted energy efficiency index PUE and the target energy efficiency index PUE, minimize the deviation between the actual combination parameters of chilled water temperature and supply air temperature and the target combination parameters, and minimize the deviation between the supply and demand of cooling capacity in each zone. It is expressed as: f1 is the minimum deviation target of energy efficiency index PUE; T is the scheduling period; PUE pre,t 、PUE tar,t are the predicted value and target value of the energy efficiency index PUE in time period t; f2 is the minimum target deviation of the combined parameters of chilled water temperature and supply air temperature; N is the number of partitions; T chi,act,i,t 、T chi,tar,i,t are the actual temperature and target temperature of the chilled water in the ith partition during period t; T w,act,i,t 、T w,tar,i,t are the actual supply air temperature and target temperature of the ith partition during period t; f3 is the minimum target for the cooling capacity supply and demand deviation of each partition; Q i,sup,t , Q i,dem,t are the cooling supply and demand values of the ith partition in period t, respectively; Set the operating constraints of the refrigeration subsystem, including at least: operating constraints of the chiller, cooling water pump, cooling tower, chilled water pump and precision air conditioning unit, cooling supply and demand balance constraints, and capacity constraints of the cold storage equipment; According to the set refrigeration subsystem objectives and operation constraints, an upper-level refrigeration subsystem optimization scheduling model is established.

8. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S5, with the goal of minimizing the system power operation cost and IT equipment load migration cost, an optimization scheduling model for the lower-level data center is established, which is expressed as: C is the minimum target for system power operation cost and IT equipment load migration cost; M is the number of equipment in the cooling subsystem and IT equipment subsystem; P j,t is the operating power consumption of the jth device in period t; is the electricity price during period t; P lm,i,t is the load migration amount of the i-th partition in period t; ε i,t is the unit load migration cost of the i-th partition in time period t; The constraints of the lower-level data center optimization scheduling model are set to at least include: computing load constraints of each partition in the IT equipment subsystem, equipment operation constraints of the cooling subsystem, and power constraints.

9. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: In S5, the optimal scheduling strategy of the system is obtained through interactive solution of the two-layer optimization scheduling model, including: Solve the optimization scheduling model of the upper-level refrigeration subsystem and output the start-stop combination status and operating power of the refrigeration equipment in the refrigeration subsystem, the storage and discharge status and storage and discharge capacity of the cold storage equipment, and the control strategy of each branch circuit in the terminal precision air-conditioning unit; Solve the optimization scheduling model of the lower-level data center and output the load migration amount of each energy consumption zone, the power load strategy of the cooling equipment and IT equipment subsystem in the cooling subsystem; Through the interactive influence of the output strategies of the upper and lower optimization scheduling models, an iterative solution is performed to obtain the optimal scheduling strategy for the system.

10. The energy-saving and economical dispatching method for a data center energy system according to claim 1, characterized in that: After S5, the process further includes: considering the hysteresis of the ambient temperature response of the IT equipment subsystem in the data center, selecting the working condition in which the ambient temperature changes after the refrigeration subsystem regulates the chilled water temperature and the supply air temperature as the identification object, establishing an ambient temperature response lag time prediction model, and regulating the refrigeration subsystem in advance based on the lag time.

Images