Data center energy-saving refrigeration method and device based on power consumption-load-pressure prediction model

Abstract

The invention provides a data center energy-saving refrigeration method and device based on a power consumption-load-pressure prediction model, and the method comprises the steps: deploying a heterogeneous sensor network in the whole domain of a data center, collecting the original data of rack-level power consumption, server load and environmental parameters, and processing the original data to obtain a standardized time series data set; and synchronously constructing a power consumption prediction sub-model, a load prediction sub-model and a pressure prediction sub-model based on the time sequence data set, realizing feature crossing by sharing a hidden layer, and predicting joint probability distribution of power consumption, load and temperature of each rack in a future adjustment period. According to the method, a coupling prediction framework of a power consumption-load-pressure three-sub-model sharing hidden layer is constructed, so that deep crossing and closed-loop constraint of multi-physical field features are realized; and the thermodynamic similarity and the physical space connectivity of the rack are organically fused, so that precise division and physical feasibility of the refrigeration area are ensured, the internal temperature standard deviation of the area is effectively reduced, and the refrigeration resource matching degree is improved.

Description

Energy-saving cooling method and device for data centers based on power consumption-load-pressure prediction model Technical Field

[0001] This invention relates to the field of intelligent cooling control technology for data centers, specifically to an energy-saving cooling method and device for data centers based on a power consumption-load-pressure prediction model. Background Technology

[0002] With the rapid development of the digital economy, data centers, as the core carriers of computing infrastructure, are experiencing exponential growth in scale and density. Statistics show that data center cooling systems account for 35%-40% of their total energy consumption, posing a serious challenge to traditional cooling control strategies based on fixed thresholds or simple feedback. Existing technologies suffer from the following main shortcomings: First, isolated and lagging prediction models. Current mainstream solutions mostly employ independent power consumption prediction, load prediction, or temperature field simulation models, lacking an effective information exchange mechanism between these models. Power consumption prediction relies solely on historical energy consumption curves, failing to consider the impact of dynamic changes in server load on heat dissipation requirements; temperature field simulations often use offline CFD (Computational Fluid Dynamics) modeling, which takes hours to compute and cannot respond in real time to minute-level fluctuations in IT load. This fragmented modeling leads to significant deviations between prediction results and actual conditions, leaving the cooling system in a "passive response" state, unable to cope with sudden load surges.

[0003] Secondly, the static and inefficient nature of cooling zones. Traditional data centers generally employ fixed physical partitioning (such as by row, column, or server room) cooling strategies, assuming that the thermodynamic behavior of racks within the same area is homogeneous. However, modern data centers handle mixed workloads, with significant differences in power density and airflow efficiency among different racks. Static partitioning results in both "overcooling" and "overheating," leading to severe waste of cooling resources. While some studies have proposed the concept of dynamic partitioning, they rely solely on simple clustering based on a single temperature parameter, failing to integrate multi-dimensional characteristics such as power consumption, load, and pressure, and neglecting to establish coupled constraints of spatial continuity and thermodynamic similarity. This results in partitioning outcomes that are physically infeasible or have unclear thermodynamic significance.

[0004] Third, the control strategies are simplistic and short-sighted. Existing energy-saving controls mostly focus on minimizing instantaneous energy consumption by increasing supply air temperature or reducing fan speed, but neglect the impact of temperature fluctuations on server reliability and the lifespan loss caused by frequent equipment start-ups and shutdowns. Some advanced solutions introduce model predictive control (MPC), but the objective function only includes an energy consumption term, and the constraints are limited to the physical limits of the equipment, lacking a quantitative assessment of long-term goals such as SLA (Service Level Agreement) default risk and equipment wear and tear costs. In addition, control commands mostly terminate at the regional level, without establishing a refined mapping relationship between regional policies and actuators such as terminal air conditioners, chillers, and airflow dampers, resulting in distorted policy implementation.

[0005] The aforementioned problems are intertwined, forming a vicious cycle of "inaccurate prediction—imprecise zoning—inadequate control—no reduction in energy consumption," which severely restricts further optimization of data center PUE (Power Usage Effectiveness). Therefore, there is an urgent need for a novel energy-saving cooling method that can deeply integrate multi-physics field coupled prediction, dynamic intelligent thermal zone partitioning, and multi-objective hierarchical optimization control. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a data center energy-saving cooling method and apparatus based on a power consumption-load-pressure prediction model, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a data center energy-saving cooling method based on a power consumption-load-pressure prediction model, comprising the following steps: S1: Deploying a heterogeneous sensor network across the entire data center to collect raw data on rack-level power consumption, server load, and environmental parameters, and processing the raw data to obtain a standardized time-series dataset; S2: Simultaneously constructing a power consumption prediction sub-model, a load prediction sub-model, and a pressure prediction sub-model based on the time-series dataset, achieving feature cross-referencing through a shared hidden layer, and predicting the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment cycle; S3: Dividing the data center into several cooling zones according to the joint probability distribution received within a preset period, and calculating a thermal fingerprint feature vector for each cooling zone, including average power consumption density, temperature non-uniformity, and airflow efficiency indicators; S4: Constructing an objective function with the minimum cost of a future adjustment cycle, constructing constraints based on the results of the zone division in S3 and the physical properties of the equipment, and solving based on the objective function and constraints to obtain the optimal control sequence; S5: Generating a regional pre-cooling strategy based on the optimal control sequence, decomposing the regional strategy into equipment-level control instructions, and finally issuing them to each execution unit.

[0008] Preferably, the step of predicting the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment cycle by achieving feature cross-multiplication through a shared hidden layer specifically includes: achieving feature cross-multiplication through a fully connected hidden layer with a dimension of 64 to form a unified prediction framework.

[0009] In the prediction framework, power consumption characteristics guide the location of heat sources in the pressure field simulation, load characteristics constrain the boundary conditions of temperature field evolution, and pressure characteristics feed back into the heat dissipation loss correction of power consumption prediction. Finally, the joint probability distribution of power consumption, load, and temperature of each rack in the next adjustment cycle is output.

[0010] Preferably, the specific process of dividing the data center into several cooling zones according to the joint probability distribution received according to the preset period includes: receiving the joint probability distribution of power consumption, load and temperature of each rack in a future adjustment period generated by the prediction model; extracting rack-level feature vectors based on the joint probability distribution and constructing a rack physical adjacency graph; and dividing the data center into several cooling zones using a spatiotemporal coupled density clustering algorithm; wherein racks in the same zone have similar thermodynamic behavior and are physically continuous.

[0011] Preferably, the spatiotemporal coupled density clustering algorithm adopts the improved DBSCAN algorithm, and the formula for measuring the similarity between racks is: ;in, Let i represent the feature vector of rack i. This represents the eigenvector of rack j. This represents the standard deviation of the feature vector across the entire sample. This represents the connectivity indicator function of the rack adjacency graph G.

[0012] Furthermore, the neighborhood radius of the DBSCAN algorithm is dynamically set to... ,in To address the regional average temperature uncertainty, adaptive prediction error is achieved.

[0013] Preferably, the calculation of the thermal fingerprint feature vector includes: average power consumption density: Temperature non-uniformity: Airflow efficiency index: ;in, Let represent the average power consumption density of the region, i represent the rack number index within region R, and R represent the set of cooling regions obtained through clustering. This represents the expected predicted power consumption of rack i. Represents the physical volume of region R. Indicates temperature non-uniformity. Represents the standard deviation function. This represents the set of predicted inlet temperatures for all racks within region R. Represents the mean function, This indicates an airflow efficiency index. This indicates the actual air supply volume. This indicates the theoretical required air volume.

[0014] Preferably, the specific steps of S4 include: constructing a multi-objective function that includes energy consumption cost, SLA default risk, and equipment wear and tear cost, with the objective of minimizing the total cost of a future adjustment cycle; wherein the calculation formula for the multi-objective function is: ;in, Indicates energy consumption cost, This indicates the risk of SLA default. Indicates equipment wear and tear costs. , and Represents the weighting coefficient, and Based on the region division results, thermal fingerprint characteristics, and equipment physical properties, a four-layer constraint system including a physical layer, an environmental layer, a regional layer, and a system layer is generated. Based on the model predictive control framework, a state transition equation is established and a predictive disturbance is introduced. An interior point solver is used to solve the optimization problem in a rolling manner to obtain the optimal control sequence for each cooling region.

[0015] Preferably, the state transition equation is constructed based on the joint probability distribution of S2: ;in, The system's state vector at the next moment, The state vector represents the inlet temperature of each rack, the cooling power of each zone, and the air volume of each zone. A represents the state matrix. B represents the control vector, which includes the zone supply air temperature setpoint and the zone fan speed ratio; B represents the control matrix. It describes the predicted and determined disturbances.

[0016] Preferably, the specific process of generating a regional precooling strategy based on the optimal control sequence includes: for high-load areas where thermal shock is predicted, the supply air temperature is gradually reduced ten minutes in advance for active precooling; for subcooled areas with continuous low load, the supply air temperature setpoint is increased and the air volume is reduced to tilt cooling resources toward hot spots.

[0017] Preferably, a data center energy-saving cooling device based on a power consumption-load-stress prediction model is provided to implement the aforementioned data center energy-saving cooling method based on the power consumption-load-stress prediction model. The device includes a data acquisition module, a probability density generation module, a region partitioning module, and a strategy generation module. The data acquisition module collects raw data on rack-level power consumption, server load, and environmental parameters through a heterogeneous sensor network, and processes the raw data to obtain a standardized time-series dataset. The probability density generation module simultaneously constructs a power consumption prediction sub-model, a load prediction sub-model, and a stress prediction sub-model based on the time-series dataset obtained by the data acquisition module. Feature cross-referencing is achieved through a shared hidden layer to predict the performance of each rack-level power consumption sub-model within a future adjustment cycle. The system generates a joint probability distribution of power consumption, load, and temperature. The region division module receives the joint probability distribution at a preset period and divides the data center into several cooling zones. For each cooling zone, it calculates a thermal fingerprint feature vector containing average power consumption density, temperature non-uniformity, and airflow efficiency. The strategy generation module constructs an objective function by minimizing the cost over a future adjustment cycle. Based on the region division results and equipment physical properties, it constructs constraints and solves the objective function and constraints to obtain the optimal control sequence. The instruction generation module also generates a region-level pre-cooling strategy based on the optimal control sequence, decomposes the region-level strategy into equipment-level control instructions, and finally distributes them to each execution unit.

[0018] This invention provides a data center energy-saving cooling method and device based on a power consumption-load-stress prediction model, which has the following beneficial effects: 1. This invention achieves deep cross-fertilization and mutual constraint of multi-physics field features by constructing a coupled prediction framework with a shared hidden layer for the three sub-models of power consumption, load, and stress. Power consumption prediction introduces load rate and temperature as external variables, effectively reducing prediction errors; load prediction uses time series decomposition and gradient boosting tree fusion to improve the sensitivity to capturing sudden fluctuations in business operations; stress prediction, based on a CFD reduced-order model, compresses the simulation time from hours to within 30 seconds, meeting real-time control requirements. The three sub-models achieve information feedback through a shared hidden layer, forming a closed loop of "power consumption locating heat sources—load constraining boundaries—stress correcting heat dissipation losses," and the final output joint probability distribution provides a high-confidence quantitative basis for subsequent decision-making.

[0019] 2. This invention organically integrates rack thermodynamic similarity with physical spatial connectivity. The exponential kernel function in the similarity measurement formula effectively captures the nonlinear relationships in the feature space, while the adjacency graph connectivity indicator function ensures the physical continuity of the partitioning results, avoiding invalid partitioning across channels and obstacles. The dynamic neighborhood radius adaptively predicts uncertainty, causing the partition boundaries to automatically adjust with the confidence level of the temperature field. This prevents the loss of control granularity caused by over-clustering and avoids the heat island effect caused by under-clustering. Compared with traditional static partitioning, this method effectively reduces the standard deviation of temperature within the region and improves the matching degree of refrigeration resources.

[0020] 3. The multi-objective function design of this invention breaks through the limitations of single energy consumption minimization, achieving Pareto optimality for energy cost, SLA default risk, and equipment wear cost through weighting coefficients. The energy cost term incorporates airflow efficiency indicators to dynamically adjust penalty weights, prioritizing optimization of inefficient areas; the SLA default risk term quantifies the probability of temperature exceeding limits based on joint probability distribution integration and assigns weights based on business importance, ensuring that the core rack temperature exceedance rate is <0.1%; the equipment wear cost term suppresses frequent actuator operations through linear penalties, reducing compressor start-stop times by more than 50% and effectively extending equipment lifespan. The four-layer constraint system fully covers everything from physical limits to system redundancy, ensuring the engineering feasibility of the optimization results. Attached Figure Description

[0021] Figure 1 is a flowchart of the main steps of the data center energy-saving cooling method based on the power consumption-load-pressure prediction model of the present invention; Figure 2 is a block diagram of the data center energy-saving cooling device based on the power consumption-load-pressure prediction model of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0024] As shown in Figure 1, this embodiment of the invention provides a data center energy-saving cooling method based on a power consumption-load-pressure prediction model, including the following steps: S1: Deploy a heterogeneous sensor network throughout the data center to collect raw data on rack-level power consumption, server load, and environmental parameters. After processing the raw data, a standardized time-series dataset is obtained. Specifically, each rack is equipped with a smart PDU to collect server power consumption data at a frequency of 1Hz. A lightweight agent program is deployed on the server motherboard to aggregate CPU, memory, and disk utilization every 5 seconds to generate a comprehensive load index. Temperature sensor grids are installed at the front and rear doors of the rack, hot and cold aisles, and floor air outlets, with a node spacing of ≤2 meters. Micro differential pressure sensors are also installed to measure airflow pressure differences. The environmental parameter sampling period is 10 seconds.

[0025] The collected raw data needs to be filtered out for jump data, the moving average method is used to smooth random noise, and missing values ​​are linearly imputed. Finally, the data is aligned with a unified timestamp and written into the time series database to form a standardized time series dataset.

[0026] S2: Based on the time series dataset, a power consumption prediction sub-model, a load prediction sub-model, and a stress prediction sub-model are constructed synchronously. Feature crossover is achieved through a shared hidden layer to predict the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment cycle. Specifically, based on the standardized time series dataset, a coupled prediction sub-model with three shared hidden layers is constructed synchronously.

[0027] The power consumption prediction sub-model employs a long short-term memory neural network, taking 72 hours of historical power consumption data, real-time load rate, ambient temperature, and service scheduling plan as input, and outputting a minute-by-minute power consumption prediction curve for the next hour. The model is incrementally trained every 24 hours. The load prediction sub-model integrates time series decomposition and gradient boosting trees, splitting the load into periodic, trend, and random components, which are predicted using the Prophet algorithm and XGBoost regression respectively, and then superimposed to generate the load distribution for the next two hours. The pressure prediction sub-model is based on a reduced-order computational fluid dynamics model, taking rack power consumption distribution, air conditioning supply parameters, and channel closure status as input, and rapidly simulating the evolution of temperature and pressure fields within 30 seconds. The three sub-models achieve feature cross-referencing through a fully connected hidden layer with a shared dimension of 64, forming a unified prediction framework.

[0028] Specifically, in the prediction framework, power consumption characteristics guide the location of heat sources in the pressure field simulation, load characteristics constrain the boundary conditions of temperature field evolution, pressure characteristics feed back into the heat dissipation loss correction of power consumption prediction, and finally output the joint probability distribution of power consumption, load and temperature of each rack in the next adjustment cycle.

[0029] S3: The data center is divided into several cooling zones according to the joint probability distribution received at a preset period, and a thermal fingerprint feature vector containing average power consumption density, temperature non-uniformity, and airflow efficiency is calculated for each cooling zone. The specific process of dividing the data center into several cooling zones according to the joint probability distribution received at a preset period includes: receiving the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment period generated by the prediction model; extracting rack-level feature vectors based on the joint probability distribution and constructing a rack physical adjacency graph; and dividing the data center into several cooling zones using a spatiotemporal coupled density clustering algorithm. Among these zones, racks within the same zone exhibit similar thermodynamic behavior and are physically spatially continuous.

[0030] The spatiotemporal coupled density clustering algorithm uses an improved DBSCAN algorithm, and the formula for measuring the similarity between racks is: ;in, Let i represent the feature vector of rack i. This represents the eigenvector of rack j. This represents the standard deviation of the feature vector across the entire sample. This represents the connectivity indicator function of the rack adjacency graph G.

[0031] Furthermore, the neighborhood radius of the DBSCAN algorithm is dynamically set to... ,in To address the regional average temperature uncertainty and achieve adaptive prediction error, the calculation of the thermal fingerprint feature vector includes: average power consumption density. Temperature non-uniformity: Airflow efficiency index: ;in, Let represent the average power consumption density of the region, i represent the rack number index within region R, and R represent the set of cooling regions obtained through clustering. This represents the expected predicted power consumption of rack i. Represents the physical volume of region R. Indicates temperature non-uniformity. Represents the standard deviation function. This represents the set of predicted inlet temperatures for all racks within region R. Represents the mean function, This indicates an airflow efficiency index. This indicates the actual air supply volume. This indicates the theoretical required air volume.

[0032] S4: Construct an objective function to minimize the cost of the next adjustment cycle. Based on the regional division results in S3 and the physical properties of the equipment, construct constraints. Solve the problem based on the objective function and constraints to obtain the optimal control sequence. Construct a multi-objective function that includes energy consumption cost, SLA default risk, and equipment wear cost, with the goal of minimizing the total cost of the next adjustment cycle. Based on the regional division results, thermal fingerprint characteristics, and equipment physical properties, generate a four-layer constraint system including physical layer, environmental layer, regional layer, and system layer. Based on the model predictive control framework, establish state transition equations and introduce predictive disturbances. Use an interior-point solver to solve the optimization problem in a rolling manner to obtain the optimal control sequence for each cooling zone.

[0033] The formula for calculating the multi-objective function is as follows: ;in, Indicates energy consumption cost, This indicates the risk of SLA default. Indicates equipment wear and tear costs. , and Represents the weighting coefficient, and .

[0034] Specifically, the objective function construction needs to comprehensively consider three types of costs: energy consumption cost is a quadratic function of the air conditioning power supply in each area, with coefficients dynamically adjusted based on the airflow efficiency index in the thermal fingerprint, and the lower the efficiency, the greater the penalty weight; SLA default risk is based on the probability integral of the temperature exceeding 27°C in the joint probability distribution, multiplied by the business importance weight (10 for core areas and 1 for ordinary areas); equipment wear and tear cost is a linear penalty for the number of compressor start-ups and shutdowns and the number of fan speed changes, encouraging smooth adjustments. The weighted sum of these three costs forms a single objective function, and the weight coefficients are calibrated quarterly using the analytic hierarchy process.

[0035] The constraint system is divided into four layers: physical layer constraints include the upper limit of cooling capacity of a single air conditioner (usually 400kW), supply air temperature range (18-25℃), and fan speed ratio (30%-100%); environmental layer constraints require that the inlet temperature of any rack not exceed 27℃ and the pressure difference between the hot and cold aisles not be less than 15Pa to ensure airflow organization; area layer constraints are based on the S3 division results, stipulating that the same area shall implement uniform supply air parameters, and the supply air temperature difference between different areas shall not exceed 5℃ to avoid air mixing; system layer constraints require that the total cooling power not exceed the UPS allocation capacity, and that N+1 redundant units must be kept in standby status.

[0036] The state transition equation is constructed based on the joint probability distribution of S2: ;in, The system's state vector at the next moment, The state vector represents the inlet temperature of each rack, the cooling power of each zone, and the air volume of each zone. A represents the state matrix. B represents the control vector, which includes the zone supply air temperature setpoint and the zone fan speed ratio; B represents the control matrix. Describe the predicted disturbances (such as load changes, outdoor temperature).

[0037] S5: Generate a regional pre-cooling strategy based on the optimal control sequence, decompose the regional strategy into device-level control instructions, and finally send them to each execution unit.

[0038] The specific process of generating a regional precooling strategy based on the optimal control sequence includes: for high-load areas where thermal shock is predicted, the supply air temperature is gradually reduced ten minutes in advance for active precooling; for subcooled areas with continuous low load, the supply air temperature setpoint is increased and the air volume is reduced to tilt cooling resources toward hot spots.

[0039] The regional strategy generation follows the "predictive compensation" principle: For regions predicted by S2 to experience a thermal shock within the next 10 minutes (load surge exceeding 30% and predicted temperature exceeding 25℃), the system initiates active pre-cooling 10 minutes in advance, gradually reducing the supply air temperature to the target value at a rate of 0.5℃ per minute to prevent the compressor from running at full frequency due to sudden high load; for regions identified by S3 as experiencing sustained low load and overcooling (average load rate <20% and temperature <22℃), the strategy increases the supply air temperature setpoint to 24℃ and reduces the fan speed in 5% increments, with the released cooling capacity being redistributed to hotspot regions through global optimization. Furthermore, the strategy generation module reads the latest control sequence from S4 every 5 minutes, performs feedforward compensation based on the actual temperature deviation of the current region, and outputs regional command packets.

[0040] Equipment-level command decomposition needs to consider actuator characteristics: mapping the zone supply air temperature to the chiller outlet water temperature setting (one chiller per zone) and the electric valve opening (PID parameters dynamically tuned based on zone heat capacity); decomposing airflow demand into speed commands for each precision air conditioning EC fan, using a consistency algorithm to ensure synchronized adjustment of multiple fans in the same zone; baffle opening commands are directly sent to the servo motors of the channel enclosure system. All commands are sent via the BACnet / IP protocol and cached in a local queue to prevent loss due to network jitter.

[0041] This application presents a data center energy-saving cooling method based on a power consumption-load-stress prediction model. By constructing a coupled prediction framework with a shared hidden layer for the three sub-models of power consumption, load, and stress, it significantly improves prediction accuracy and response speed, achieving deep cross-fertilization and closed-loop constraints of multi-physics features. It innovatively proposes a spatiotemporal coupled density clustering algorithm, organically integrating rack thermodynamic similarity with physical spatial connectivity to ensure accurate and physically feasible cooling zone division. Furthermore, through multi-objective function design to balance and optimize energy consumption costs, SLA default risks, and equipment wear costs, combined with a four-layer constraint system, it achieves dual guarantees of energy saving and service quality. This significantly reduces the overall PUE while ensuring the core rack temperature does not exceed the limit, effectively extending equipment lifespan.

[0042] As shown in Figure 2, this embodiment also provides a data center energy-saving cooling device based on a power consumption-load-stress prediction model, used to implement the above-mentioned data center energy-saving cooling method based on a power consumption-load-stress prediction model. It includes a data acquisition module, a probability density generation module, a region partitioning module, and a strategy generation module. The data acquisition module collects raw data on rack-level power consumption, server load, and environmental parameters through a heterogeneous sensor network, and processes the raw data to obtain a standardized time-series dataset. The probability density generation module synchronously constructs a power consumption prediction sub-model, a load prediction sub-model, and a stress prediction sub-model based on the time-series dataset obtained by the data acquisition module. Feature crossover is achieved through a shared hidden layer to predict a future adjustment... The system calculates the joint probability distribution of power consumption, load, and temperature for each rack within a given period. The region partitioning module receives the joint probability distribution and divides the data center into several cooling zones based on the preset period. For each cooling zone, it calculates a thermal fingerprint feature vector containing average power density, temperature non-uniformity, and airflow efficiency. The strategy generation module constructs an objective function by minimizing the cost of a future adjustment cycle, constructs constraints based on the region partitioning results and equipment physical properties, and solves the objective function and constraints to obtain the optimal control sequence. The instruction generation module also generates a region-level pre-cooling strategy based on the optimal control sequence, decomposes the region-level strategy into equipment-level control instructions, and finally issues them to each execution unit.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data center energy-saving cooling method based on a power consumption-load-pressure prediction model, characterized in that, Includes the following steps: S1: Deploy a heterogeneous sensor network across the entire data center to collect raw data on rack-level power consumption, server load, and environmental parameters. Process the raw data to obtain a standardized time-series dataset. S2: Based on the time-series dataset, simultaneously construct a power consumption prediction sub-model, a load prediction sub-model, and a stress prediction sub-model. A shared hidden layer enables feature cross-pollination, predicting the joint probability distribution of power consumption, load, and temperature for each rack within a future adjustment cycle. S3: Divide the data center into several cooling zones based on the joint probability distribution received within a preset cycle. Calculate a thermal fingerprint feature vector for each cooling zone, including average power density, temperature non-uniformity, and airflow efficiency. S4: Construct an objective function with the minimum cost for a future adjustment cycle. Based on the zone division results in S3 and the physical properties of the equipment, construct constraints. Solve the objective function and constraints to obtain the optimal control sequence. S5: Generate a zone-level pre-cooling strategy based on the optimal control sequence. Decompose the zone-level strategy into device-level control commands and finally distribute them to each execution unit.

2. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 1, characterized in that, The method of predicting the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment cycle by achieving feature cross-connection through a shared hidden layer specifically includes: achieving feature cross-connection through a shared fully connected hidden layer with a dimension of 64 to form a unified prediction framework; in the prediction framework, power consumption features guide the location of heat sources in the pressure field simulation, load features constrain the boundary conditions of the temperature field evolution, pressure features feed back into the heat dissipation loss correction of power consumption prediction, and finally outputs the joint probability distribution of power consumption, load, and temperature of each rack within a future adjustment cycle.

3. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 1, characterized in that, The specific process of dividing the data center into several cooling zones according to the joint probability distribution received according to the preset period includes: receiving the joint probability distribution of power consumption, load and temperature of each rack in a future adjustment period generated by the prediction model; extracting rack-level feature vectors based on the joint probability distribution and constructing a rack physical adjacency graph; and dividing the data center into several cooling zones using a spatiotemporal coupled density clustering algorithm; wherein racks in the same zone have similar thermodynamic behavior and are physically continuous.

4. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 3, characterized in that, The spatiotemporal coupled density clustering algorithm uses an improved DBSCAN algorithm, and the formula for measuring the similarity between racks is: ;in, Let i represent the feature vector of rack i. This represents the eigenvector of rack j. This represents the standard deviation of the feature vector across the entire sample. The connectivity indicator function represents the rack adjacency graph G; and the neighborhood radius of the DBSCAN algorithm is dynamically set to... ,in To address the regional average temperature uncertainty, adaptive prediction error is achieved.

5. A data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 3, characterized in that, The calculation of the thermal fingerprint feature vector includes: average power consumption density: Temperature non-uniformity: Airflow efficiency index: ;in, Let represent the average power consumption density of the region, i represent the rack number index within region R, and R represent the set of cooling regions obtained through clustering. This represents the expected predicted power consumption of rack i. Represents the physical volume of region R. Indicates temperature non-uniformity. Represents the standard deviation function. This represents the set of predicted inlet temperatures for all racks within region R. Represents the mean function, This indicates an airflow efficiency index. This indicates the actual air supply volume. This indicates the theoretical required air volume.

6. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 1, characterized in that, The specific steps of S4 include: constructing a multi-objective function that includes energy consumption cost, SLA default risk, and equipment wear and tear cost, with the goal of minimizing the total cost of a future adjustment cycle; wherein the calculation formula for the multi-objective function is: ;in, Indicates energy consumption cost, This indicates the risk of SLA default. Indicates equipment wear and tear costs. 、 and Represents the weighting coefficient, and Based on the region division results, thermal fingerprint characteristics, and equipment physical properties, a four-layer constraint system including a physical layer, an environmental layer, a regional layer, and a system layer is generated. Based on the model predictive control framework, a state transition equation is established and a predictive disturbance is introduced. An interior point method solver is used to solve the optimization problem in a rolling manner to obtain the optimal control sequence for each cooling region.

7. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 1, characterized in that, The state transition equation is constructed based on the joint probability distribution of S2: ;in, The system's state vector at the next moment, The state vector represents the inlet temperature of each rack, the cooling power of each zone, and the air volume of each zone. A represents the state matrix. B represents the control vector, which includes the zone supply air temperature setpoint and the zone fan speed ratio; B represents the control matrix. It describes the predicted and determined disturbances.

8. The data center energy-saving cooling method based on a power consumption-load-pressure prediction model according to claim 1, characterized in that, The specific process of generating a regional precooling strategy based on the optimal control sequence includes: for high-load areas where thermal shock is predicted, the supply air temperature is gradually reduced ten minutes in advance for active precooling; for subcooled areas with continuous low load, the supply air temperature setpoint is increased and the air volume is reduced to tilt cooling resources toward hot spots.

9. A data center energy-saving cooling device based on a power consumption-load-pressure prediction model, used to implement the data center energy-saving cooling method based on a power consumption-load-pressure prediction model as described in any one of claims 1-8, characterized in that, It includes a data acquisition module, a probability density generation module, a region partitioning module, and a strategy generation module. The data acquisition module collects raw data on rack-level power consumption, server load, and environmental parameters through a heterogeneous sensor network, and processes the raw data to obtain a standardized time-series dataset. The probability density generation module synchronously constructs power consumption prediction sub-models, load prediction sub-models, and pressure prediction sub-models based on the time-series dataset obtained by the data acquisition module. It achieves feature cross-referencing through a shared hidden layer to predict the joint probability distribution of power consumption, load, and temperature for each rack within a future adjustment cycle. The region partitioning module divides the data center into several cooling zones based on the joint probability distribution received within a preset period and calculates a thermal fingerprint feature vector for each cooling zone, including average power consumption density, temperature non-uniformity, and airflow efficiency. The strategy generation module constructs an objective function by minimizing the cost for a future adjustment cycle, constructs constraints based on the region partitioning results from the region partitioning module and the physical properties of the equipment, and solves the objective function and constraints to obtain the optimal control sequence. The instruction generation module is also used to generate a region-level pre-cooling strategy based on the optimal control sequence, decomposes the region-level strategy into device-level control instructions, and finally issues them to each execution unit.

Images